diff options
author | Marat Dukhan <maratek@google.com> | 2020-04-24 01:46:00 -0700 |
---|---|---|
committer | XNNPACK Team <xnnpack-github-robot@google.com> | 2020-04-24 01:46:35 -0700 |
commit | f5425ea6472c231ec56e7718ac9524ec5fe569e9 (patch) | |
tree | 2aeb59f1a55305e0082214fb6fc4c8c54803b13b | |
parent | ca2ba702af21eb2febe06498d84c8ba45fbaecc0 (diff) | |
download | XNNPACK-f5425ea6472c231ec56e7718ac9524ec5fe569e9.tar.gz |
Additional NEON/NEONFMA DWCONV microkernels
Faster 2x2 and 5x5 depthwise convolution
PiperOrigin-RevId: 308216337
23 files changed, 6993 insertions, 5 deletions
diff --git a/BUILD.bazel b/BUILD.bazel index e451acc13..71bf4232c 100644 --- a/BUILD.bazel +++ b/BUILD.bazel @@ -505,10 +505,18 @@ NEON_UKERNELS = [ "src/f32-avgpool/9x-minmax-neon-c4.c", "src/f32-clamp/gen/neon-x4.c", "src/f32-clamp/gen/neon-x8.c", + "src/f32-dwconv/gen/up4x4-minmax-neon.c", + "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", + "src/f32-dwconv/gen/up8x4-minmax-neon.c", + "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", "src/f32-dwconv/gen/up4x9-minmax-neon.c", "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", "src/f32-dwconv/gen/up8x9-minmax-neon.c", "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", + "src/f32-dwconv/gen/up4x25-minmax-neon.c", + "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", + "src/f32-dwconv/gen/up8x25-minmax-neon.c", + "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", "src/f32-gavgpool-spchw/neon-x4.c", "src/f32-gavgpool/7p7x-minmax-neon-c4.c", "src/f32-gavgpool/7x-minmax-neon-c4.c", @@ -687,10 +695,18 @@ NEONFMA_UKERNELS = [ "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", + "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", + "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", + "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", + "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", + "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", + "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", + "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", + "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", diff --git a/CMakeLists.txt b/CMakeLists.txt index 8e1d3b797..c17617b92 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -501,10 +501,18 @@ SET(XNNPACK_NEON_MICROKERNEL_SRCS src/f32-avgpool/9x-minmax-neon-c4.c src/f32-clamp/gen/neon-x4.c src/f32-clamp/gen/neon-x8.c + src/f32-dwconv/gen/up4x4-minmax-neon.c + src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c + src/f32-dwconv/gen/up8x4-minmax-neon.c + src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c src/f32-dwconv/gen/up4x9-minmax-neon.c src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c src/f32-dwconv/gen/up8x9-minmax-neon.c src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c + src/f32-dwconv/gen/up4x25-minmax-neon.c + src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c + src/f32-dwconv/gen/up8x25-minmax-neon.c + src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c src/f32-gavgpool-spchw/neon-x4.c src/f32-gavgpool/7p7x-minmax-neon-c4.c src/f32-gavgpool/7x-minmax-neon-c4.c @@ -682,10 +690,18 @@ SET(XNNPACK_NEONFMA_MICROKERNEL_SRCS src/f32-igemm/gen/4x8s4-minmax-neonfma.c src/f32-igemm/gen/6x8s4-minmax-neonfma.c src/f32-igemm/gen/8x8s4-minmax-neonfma.c + src/f32-dwconv/gen/up4x4-minmax-neonfma.c + src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c + src/f32-dwconv/gen/up8x4-minmax-neonfma.c + src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c src/f32-dwconv/gen/up4x9-minmax-neonfma.c src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c src/f32-dwconv/gen/up8x9-minmax-neonfma.c src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c + src/f32-dwconv/gen/up4x25-minmax-neonfma.c + src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c + src/f32-dwconv/gen/up8x25-minmax-neonfma.c + src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c diff --git a/scripts/generate-f32-dwconv.sh b/scripts/generate-f32-dwconv.sh index 79d5b1784..2e1055c52 100755 --- a/scripts/generate-f32-dwconv.sh +++ b/scripts/generate-f32-dwconv.sh @@ -68,6 +68,16 @@ tools/xngen src/f32-dwconv/up-scalar.c.in -D CHANNEL_TILE=2 -D KERNEL_TILE=25 -D tools/xngen src/f32-dwconv/up-scalar.c.in -D CHANNEL_TILE=2 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -D WASM=1 -D ACTIVATION=MINMAX -o src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c ################################### ARM NEON ################################## +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -D FMA=0 -o src/f32-dwconv/gen/up4x4-minmax-neon.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -D FMA=0 -o src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -D FMA=0 -o src/f32-dwconv/gen/up8x4-minmax-neon.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -D FMA=0 -o src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c + +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -D FMA=1 -o src/f32-dwconv/gen/up4x4-minmax-neonfma.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -D FMA=1 -o src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -D FMA=1 -o src/f32-dwconv/gen/up8x4-minmax-neonfma.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -D FMA=1 -o src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c + tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -D FMA=0 -o src/f32-dwconv/gen/up4x9-minmax-neon.c tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -D FMA=0 -o src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -D FMA=0 -o src/f32-dwconv/gen/up8x9-minmax-neon.c @@ -78,6 +88,16 @@ tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=9 -D AC tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -D FMA=1 -o src/f32-dwconv/gen/up8x9-minmax-neonfma.c tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -D FMA=1 -o src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -D FMA=0 -o src/f32-dwconv/gen/up4x25-minmax-neon.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -D FMA=0 -o src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -D FMA=0 -o src/f32-dwconv/gen/up8x25-minmax-neon.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -D FMA=0 -o src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c + +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -D FMA=1 -o src/f32-dwconv/gen/up4x25-minmax-neonfma.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -D FMA=1 -o src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -D FMA=1 -o src/f32-dwconv/gen/up8x25-minmax-neonfma.c +tools/xngen src/f32-dwconv/up-neon.c.in -D CHANNEL_TILE=8 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -D FMA=1 -o src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c + ################################### x86 SSE ################################### tools/xngen src/f32-dwconv/up-sse.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f32-dwconv/gen/up4x4-minmax-sse.c tools/xngen src/f32-dwconv/up-sse.c.in -D CHANNEL_TILE=4 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c diff --git a/src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c b/src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c new file mode 100644 index 000000000..d8aad5abc --- /dev/null +++ b/src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c @@ -0,0 +1,321 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x25-minmax-neon.c b/src/f32-dwconv/gen/up4x25-minmax-neon.c new file mode 100644 index 000000000..972204efd --- /dev/null +++ b/src/f32-dwconv/gen/up4x25-minmax-neon.c @@ -0,0 +1,317 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x25__neon( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c b/src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c new file mode 100644 index 000000000..b1f30e96d --- /dev/null +++ b/src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c @@ -0,0 +1,321 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x25-minmax-neonfma.c b/src/f32-dwconv/gen/up4x25-minmax-neonfma.c new file mode 100644 index 000000000..3181d809f --- /dev/null +++ b/src/f32-dwconv/gen/up4x25-minmax-neonfma.c @@ -0,0 +1,317 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c b/src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c new file mode 100644 index 000000000..1bdd7fe2f --- /dev/null +++ b/src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c @@ -0,0 +1,111 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x4-minmax-neon.c b/src/f32-dwconv/gen/up4x4-minmax-neon.c new file mode 100644 index 000000000..a07dc6c09 --- /dev/null +++ b/src/f32-dwconv/gen/up4x4-minmax-neon.c @@ -0,0 +1,107 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x4__neon( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c b/src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c new file mode 100644 index 000000000..67140864a --- /dev/null +++ b/src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c @@ -0,0 +1,111 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up4x4-minmax-neonfma.c b/src/f32-dwconv/gen/up4x4-minmax-neonfma.c new file mode 100644 index 000000000..ee9d59472 --- /dev/null +++ b/src/f32-dwconv/gen/up4x4-minmax-neonfma.c @@ -0,0 +1,107 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c b/src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c new file mode 100644 index 000000000..e807c38ad --- /dev/null +++ b/src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c @@ -0,0 +1,513 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi3x4567, vk3x4567); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk5x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi5x4567, vk5x4567); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk6x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi6x4567, vk6x4567); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk7x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi7x4567, vk7x4567); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk8x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vi9x4567 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk9x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi9x4567, vk9x4567); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vi10x4567 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk10x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi10x4567, vk10x4567); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vi11x4567 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk11x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi11x4567, vk11x4567); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vi12x4567 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk12x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi12x4567, vk12x4567); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vi13x4567 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk13x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi13x4567, vk13x4567); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vi14x4567 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk14x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi14x4567, vk14x4567); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vi15x4567 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk15x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi15x4567, vk15x4567); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vi16x4567 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk16x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi16x4567, vk16x4567); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vi17x4567 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk17x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi17x4567, vk17x4567); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vi18x4567 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk18x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi18x4567, vk18x4567); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vi19x4567 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk19x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi19x4567, vk19x4567); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk20x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi20x4567, vk20x4567); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vi21x4567 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk21x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi21x4567, vk21x4567); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vi22x4567 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk22x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi22x4567, vk22x4567); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vi23x4567 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk23x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi23x4567, vk23x4567); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vi24x4567 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk24x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi24x4567, vk24x4567); + + // Add up all accumulators to vacc01234567p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w + 36); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w + 44); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w + 52); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w + 60); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w + 68); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w + 76); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w + 84); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w + 92); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w + 100); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w + 108); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w + 116); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w + 124); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w + 132); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w + 140); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w + 148); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w + 156); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w + 164); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w + 172); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w + 180); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w + 188); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w + 196); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w + 40); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w + 48); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w + 56); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w + 64); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w + 72); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w + 80); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w + 88); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w + 96); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w + 104); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w + 112); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w + 120); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w + 128); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w + 136); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w + 144); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w + 152); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w + 160); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w + 168); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w + 176); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w + 184); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w + 192); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w + 200); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x25-minmax-neon.c b/src/f32-dwconv/gen/up8x25-minmax-neon.c new file mode 100644 index 000000000..901481188 --- /dev/null +++ b/src/f32-dwconv/gen/up8x25-minmax-neon.c @@ -0,0 +1,506 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x25__neon( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi3x4567, vk3x4567); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk5x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi5x4567, vk5x4567); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk6x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi6x4567, vk6x4567); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk7x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi7x4567, vk7x4567); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk8x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vi9x4567 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk9x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi9x4567, vk9x4567); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vi10x4567 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk10x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi10x4567, vk10x4567); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vi11x4567 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk11x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi11x4567, vk11x4567); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vi12x4567 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk12x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi12x4567, vk12x4567); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vi13x4567 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk13x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi13x4567, vk13x4567); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vi14x4567 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk14x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi14x4567, vk14x4567); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vi15x4567 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk15x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi15x4567, vk15x4567); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vi16x4567 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk16x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi16x4567, vk16x4567); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vi17x4567 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk17x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi17x4567, vk17x4567); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vi18x4567 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk18x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi18x4567, vk18x4567); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vi19x4567 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk19x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi19x4567, vk19x4567); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk20x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi20x4567, vk20x4567); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vi21x4567 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk21x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi21x4567, vk21x4567); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vi22x4567 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk22x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi22x4567, vk22x4567); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vi23x4567 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk23x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi23x4567, vk23x4567); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vi24x4567 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk24x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi24x4567, vk24x4567); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w + 36); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w + 44); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w + 52); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w + 60); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w + 68); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w + 76); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w + 84); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w + 92); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w + 100); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w + 108); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w + 116); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w + 124); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w + 132); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w + 140); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w + 148); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w + 156); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w + 164); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w + 172); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w + 180); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w + 188); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w + 196); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w + 40); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w + 48); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w + 56); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w + 64); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w + 72); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w + 80); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w + 88); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w + 96); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w + 104); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w + 112); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w + 120); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w + 128); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w + 136); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w + 144); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w + 152); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w + 160); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w + 168); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w + 176); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w + 184); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w + 192); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w + 200); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c b/src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c new file mode 100644 index 000000000..c18dc390f --- /dev/null +++ b/src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c @@ -0,0 +1,513 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi3x4567, vk3x4567); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk5x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi5x4567, vk5x4567); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk6x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi6x4567, vk6x4567); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk7x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi7x4567, vk7x4567); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk8x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi8x4567, vk8x4567); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vi9x4567 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk9x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi9x4567, vk9x4567); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vi10x4567 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk10x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi10x4567, vk10x4567); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vi11x4567 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk11x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi11x4567, vk11x4567); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vi12x4567 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk12x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi12x4567, vk12x4567); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vi13x4567 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk13x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi13x4567, vk13x4567); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vi14x4567 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk14x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi14x4567, vk14x4567); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vi15x4567 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk15x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi15x4567, vk15x4567); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vi16x4567 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk16x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi16x4567, vk16x4567); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vi17x4567 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk17x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi17x4567, vk17x4567); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vi18x4567 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk18x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi18x4567, vk18x4567); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vi19x4567 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk19x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi19x4567, vk19x4567); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk20x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi20x4567, vk20x4567); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vi21x4567 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk21x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi21x4567, vk21x4567); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vi22x4567 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk22x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi22x4567, vk22x4567); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vi23x4567 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk23x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi23x4567, vk23x4567); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vi24x4567 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk24x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi24x4567, vk24x4567); + + // Add up all accumulators to vacc01234567p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w + 36); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w + 44); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w + 52); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w + 60); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w + 68); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w + 76); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w + 84); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w + 92); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w + 100); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w + 108); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w + 116); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w + 124); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w + 132); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w + 140); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w + 148); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w + 156); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w + 164); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w + 172); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w + 180); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w + 188); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w + 196); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w + 40); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w + 48); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w + 56); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w + 64); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w + 72); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w + 80); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w + 88); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w + 96); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w + 104); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w + 112); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w + 120); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w + 128); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w + 136); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w + 144); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w + 152); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w + 160); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w + 168); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w + 176); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w + 184); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w + 192); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w + 200); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x25-minmax-neonfma.c b/src/f32-dwconv/gen/up8x25-minmax-neonfma.c new file mode 100644 index 000000000..dcd781e1a --- /dev/null +++ b/src/f32-dwconv/gen/up8x25-minmax-neonfma.c @@ -0,0 +1,506 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + const float* i4 = input[4]; + assert(i4 != NULL); + const float* i5 = input[5]; + assert(i5 != NULL); + const float* i6 = input[6]; + assert(i6 != NULL); + const float* i7 = input[7]; + assert(i7 != NULL); + const float* i8 = input[8]; + assert(i8 != NULL); + const float* i9 = input[9]; + assert(i9 != NULL); + const float* i10 = input[10]; + assert(i10 != NULL); + const float* i11 = input[11]; + assert(i11 != NULL); + const float* i12 = input[12]; + assert(i12 != NULL); + const float* i13 = input[13]; + assert(i13 != NULL); + const float* i14 = input[14]; + assert(i14 != NULL); + const float* i15 = input[15]; + assert(i15 != NULL); + const float* i16 = input[16]; + assert(i16 != NULL); + const float* i17 = input[17]; + assert(i17 != NULL); + const float* i18 = input[18]; + assert(i18 != NULL); + const float* i19 = input[19]; + assert(i19 != NULL); + const float* i20 = input[20]; + assert(i20 != NULL); + const float* i21 = input[21]; + assert(i21 != NULL); + const float* i22 = input[22]; + assert(i22 != NULL); + const float* i23 = input[23]; + assert(i23 != NULL); + const float* i24 = input[24]; + assert(i24 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi3x4567, vk3x4567); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk5x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi5x4567, vk5x4567); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk6x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi6x4567, vk6x4567); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk7x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi7x4567, vk7x4567); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk8x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi8x4567, vk8x4567); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vi9x4567 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk9x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi9x4567, vk9x4567); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vi10x4567 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk10x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi10x4567, vk10x4567); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vi11x4567 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk11x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi11x4567, vk11x4567); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vi12x4567 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk12x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi12x4567, vk12x4567); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vi13x4567 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk13x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi13x4567, vk13x4567); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vi14x4567 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk14x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi14x4567, vk14x4567); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vi15x4567 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk15x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi15x4567, vk15x4567); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vi16x4567 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk16x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi16x4567, vk16x4567); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vi17x4567 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk17x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi17x4567, vk17x4567); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vi18x4567 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk18x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi18x4567, vk18x4567); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vi19x4567 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk19x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi19x4567, vk19x4567); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vi20x4567 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk20x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi20x4567, vk20x4567); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vi21x4567 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk21x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi21x4567, vk21x4567); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vi22x4567 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk22x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi22x4567, vk22x4567); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vi23x4567 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk23x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi23x4567, vk23x4567); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vi24x4567 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk24x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi24x4567, vk24x4567); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4; + const float32x4_t vk4x0123 = vld1q_f32(w + 36); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4; + const float32x4_t vk5x0123 = vld1q_f32(w + 44); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4; + const float32x4_t vk6x0123 = vld1q_f32(w + 52); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; + const float32x4_t vk7x0123 = vld1q_f32(w + 60); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; + const float32x4_t vk8x0123 = vld1q_f32(w + 68); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); i9 += 4; + const float32x4_t vk9x0123 = vld1q_f32(w + 76); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); i10 += 4; + const float32x4_t vk10x0123 = vld1q_f32(w + 84); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); i11 += 4; + const float32x4_t vk11x0123 = vld1q_f32(w + 92); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); i12 += 4; + const float32x4_t vk12x0123 = vld1q_f32(w + 100); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); i13 += 4; + const float32x4_t vk13x0123 = vld1q_f32(w + 108); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); i14 += 4; + const float32x4_t vk14x0123 = vld1q_f32(w + 116); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); i15 += 4; + const float32x4_t vk15x0123 = vld1q_f32(w + 124); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; + const float32x4_t vk16x0123 = vld1q_f32(w + 132); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); i17 += 4; + const float32x4_t vk17x0123 = vld1q_f32(w + 140); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); i18 += 4; + const float32x4_t vk18x0123 = vld1q_f32(w + 148); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); i19 += 4; + const float32x4_t vk19x0123 = vld1q_f32(w + 156); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); i20 += 4; + const float32x4_t vk20x0123 = vld1q_f32(w + 164); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); i21 += 4; + const float32x4_t vk21x0123 = vld1q_f32(w + 172); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); i22 += 4; + const float32x4_t vk22x0123 = vld1q_f32(w + 180); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); i23 += 4; + const float32x4_t vk23x0123 = vld1q_f32(w + 188); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); i24 += 4; + const float32x4_t vk24x0123 = vld1q_f32(w + 196); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + const float32x4_t vi4x0123 = vld1q_f32(i4); + const float32x4_t vk4x0123 = vld1q_f32(w + 40); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); + + const float32x4_t vi5x0123 = vld1q_f32(i5); + const float32x4_t vk5x0123 = vld1q_f32(w + 48); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); + + const float32x4_t vi6x0123 = vld1q_f32(i6); + const float32x4_t vk6x0123 = vld1q_f32(w + 56); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); + + const float32x4_t vi7x0123 = vld1q_f32(i7); + const float32x4_t vk7x0123 = vld1q_f32(w + 64); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); + + const float32x4_t vi8x0123 = vld1q_f32(i8); + const float32x4_t vk8x0123 = vld1q_f32(w + 72); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); + + const float32x4_t vi9x0123 = vld1q_f32(i9); + const float32x4_t vk9x0123 = vld1q_f32(w + 80); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi9x0123, vk9x0123); + + const float32x4_t vi10x0123 = vld1q_f32(i10); + const float32x4_t vk10x0123 = vld1q_f32(w + 88); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); + + const float32x4_t vi11x0123 = vld1q_f32(i11); + const float32x4_t vk11x0123 = vld1q_f32(w + 96); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); + + const float32x4_t vi12x0123 = vld1q_f32(i12); + const float32x4_t vk12x0123 = vld1q_f32(w + 104); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); + + const float32x4_t vi13x0123 = vld1q_f32(i13); + const float32x4_t vk13x0123 = vld1q_f32(w + 112); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi13x0123, vk13x0123); + + const float32x4_t vi14x0123 = vld1q_f32(i14); + const float32x4_t vk14x0123 = vld1q_f32(w + 120); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi14x0123, vk14x0123); + + const float32x4_t vi15x0123 = vld1q_f32(i15); + const float32x4_t vk15x0123 = vld1q_f32(w + 128); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi15x0123, vk15x0123); + + const float32x4_t vi16x0123 = vld1q_f32(i16); + const float32x4_t vk16x0123 = vld1q_f32(w + 136); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); + + const float32x4_t vi17x0123 = vld1q_f32(i17); + const float32x4_t vk17x0123 = vld1q_f32(w + 144); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); + + const float32x4_t vi18x0123 = vld1q_f32(i18); + const float32x4_t vk18x0123 = vld1q_f32(w + 152); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi18x0123, vk18x0123); + + const float32x4_t vi19x0123 = vld1q_f32(i19); + const float32x4_t vk19x0123 = vld1q_f32(w + 160); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi19x0123, vk19x0123); + + const float32x4_t vi20x0123 = vld1q_f32(i20); + const float32x4_t vk20x0123 = vld1q_f32(w + 168); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi20x0123, vk20x0123); + + const float32x4_t vi21x0123 = vld1q_f32(i21); + const float32x4_t vk21x0123 = vld1q_f32(w + 176); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi21x0123, vk21x0123); + + const float32x4_t vi22x0123 = vld1q_f32(i22); + const float32x4_t vk22x0123 = vld1q_f32(w + 184); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); + + const float32x4_t vi23x0123 = vld1q_f32(i23); + const float32x4_t vk23x0123 = vld1q_f32(w + 192); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi23x0123, vk23x0123); + + const float32x4_t vi24x0123 = vld1q_f32(i24); + const float32x4_t vk24x0123 = vld1q_f32(w + 200); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi24x0123, vk24x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c b/src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c new file mode 100644 index 000000000..42b53b0b1 --- /dev/null +++ b/src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c @@ -0,0 +1,156 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + vacc4567p1 = vmlaq_f32(vacc4567p1, vi3x4567, vk3x4567); + + // Add up all accumulators to vacc01234567p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x4-minmax-neon.c b/src/f32-dwconv/gen/up8x4-minmax-neon.c new file mode 100644 index 000000000..504b2e6a5 --- /dev/null +++ b/src/f32-dwconv/gen/up8x4-minmax-neon.c @@ -0,0 +1,149 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x4__neon( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + vacc4567p0 = vmlaq_f32(vacc4567p0, vi3x4567, vk3x4567); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p0 = vmlaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c b/src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c new file mode 100644 index 000000000..2a8d8c054 --- /dev/null +++ b/src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c @@ -0,0 +1,156 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + vacc4567p1 = vfmaq_f32(vacc4567p1, vi3x4567, vk3x4567); + + // Add up all accumulators to vacc01234567p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); + + // Add up all accumulators to vacc0123p0 + vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/f32-dwconv/gen/up8x4-minmax-neonfma.c b/src/f32-dwconv/gen/up8x4-minmax-neonfma.c new file mode 100644 index 000000000..657ddd23e --- /dev/null +++ b/src/f32-dwconv/gen/up8x4-minmax-neonfma.c @@ -0,0 +1,149 @@ +// Auto-generated file. Do not edit! +// Template: src/f32-dwconv/up-neon.c.in +// Generator: tools/xngen +// +// Copyright 2019 Google LLC +// +// This source code is licensed under the BSD-style license found in the +// LICENSE file in the root directory of this source tree. + +#include <assert.h> + +#include <arm_neon.h> + +#include <xnnpack/dwconv.h> + + +void xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma( + size_t channels, + size_t output_width, + const float** input, + const float* weights, + float* output, + size_t input_stride, + size_t output_increment, + const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) +{ + assert(channels != 0); + assert(output_width != 0); + + const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); + const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); + do { + const float* i0 = input[0]; + assert(i0 != NULL); + const float* i1 = input[1]; + assert(i1 != NULL); + const float* i2 = input[2]; + assert(i2 != NULL); + const float* i3 = input[3]; + assert(i3 != NULL); + input = (const float**) ((uintptr_t) input + input_stride); + + size_t c = channels; + const float* w = weights; + for (; c >= 8; c -= 8) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + float32x4_t vacc4567p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi0x4567, vk0x4567); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk1x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi1x4567, vk1x4567); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; + const float32x4_t vk3x4567 = vld1q_f32(w); w += 4; + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + vacc4567p0 = vfmaq_f32(vacc4567p0, vi3x4567, vk3x4567); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + vacc4567 = vminq_f32(vacc4567, vmax); + + vst1q_f32(output, vacc0123); output += 4; + vst1q_f32(output, vacc4567); output += 4; + } + for (; c >= 4; c -= 4) { + float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; + + + const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; + const float32x4_t vk0x0123 = vld1q_f32(w + 4); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4; + const float32x4_t vk1x0123 = vld1q_f32(w + 12); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4; + const float32x4_t vk2x0123 = vld1q_f32(w + 20); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4; + const float32x4_t vk3x0123 = vld1q_f32(w + 28); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + vst1q_f32(output, vacc0123); output += 4; + } + if XNN_UNLIKELY(c != 0) { + float32x4_t vacc0123p0 = vld1q_f32(w); + + + const float32x4_t vi0x0123 = vld1q_f32(i0); + const float32x4_t vk0x0123 = vld1q_f32(w + 8); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); + + const float32x4_t vi1x0123 = vld1q_f32(i1); + const float32x4_t vk1x0123 = vld1q_f32(w + 16); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); + + const float32x4_t vi2x0123 = vld1q_f32(i2); + const float32x4_t vk2x0123 = vld1q_f32(w + 24); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123); + + const float32x4_t vi3x0123 = vld1q_f32(i3); + const float32x4_t vk3x0123 = vld1q_f32(w + 32); + vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123); + + + float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin); + vacc0123 = vminq_f32(vacc0123, vmax); + + float32x2_t vacc01 = vget_low_f32(vacc0123); + if (c & 2) { + vst1_f32(output, vacc01); output += 2; + vacc01 = vget_high_f32(vacc0123); + } + if (c & 1) { + vst1_lane_f32(output, vacc01, 0); output += 1; + } + } + + output = (float*) ((uintptr_t) output + output_increment); + } while (--output_width != 0); +} diff --git a/src/init.c b/src/init.c index 8c044ac56..b9b80040c 100644 --- a/src/init.c +++ b/src/init.c @@ -236,7 +236,7 @@ static void init(void) { xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; - xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x4__psimd; + xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x4__neon; xnn_params.f32.dwconv[0].channel_tile = 4, xnn_params.f32.dwconv[0].primary_tile = 4, @@ -244,7 +244,7 @@ static void init(void) { xnn_params.f32.dwconv[1].channel_tile = 4; xnn_params.f32.dwconv[1].primary_tile = 9; - xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__psimd; + xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2; xnn_params.f32.dwconv[2].channel_tile = 4; xnn_params.f32.dwconv[2].primary_tile = 25; @@ -562,8 +562,8 @@ static void init(void) { xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; - xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x4__psimd; - xnn_params.f32.dwconv[0].channel_tile = 4; + xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma; + xnn_params.f32.dwconv[0].channel_tile = 8; xnn_params.f32.dwconv[0].primary_tile = 4; #if XNN_PLATFORM_IOS @@ -594,7 +594,7 @@ static void init(void) { } #endif // XNN_PLATFORM_IOS - xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__psimd; + xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2; xnn_params.f32.dwconv[2].channel_tile = 4; xnn_params.f32.dwconv[2].primary_tile = 25; diff --git a/src/xnnpack/dwconv.h b/src/xnnpack/dwconv.h index a42a25a3b..d6eab5fc4 100644 --- a/src/xnnpack/dwconv.h +++ b/src/xnnpack/dwconv.h @@ -40,6 +40,14 @@ extern "C" { size_t output_increment, \ const union xnn_f32_minmax_params* params); +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neon) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neon) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__psimd) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__psimd_acc2) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__psimd) @@ -92,6 +100,14 @@ DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neon) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neon) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma) +DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__psimd) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__psimd_acc2) DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__psimd) diff --git a/test/f32-dwconv-minmax.cc b/test/f32-dwconv-minmax.cc index 7758a7e90..cc9cff281 100644 --- a/test/f32-dwconv-minmax.cc +++ b/test/f32-dwconv-minmax.cc @@ -348,6 +348,642 @@ #if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_eq_4) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_div_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_lt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_gt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_eq_4) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_div_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_lt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_gt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEONFMA_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_eq_8) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_div_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_lt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_gt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_eq_8) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_div_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_lt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_gt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEONFMA_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 TEST(F32_DWCONV_MINMAX_UP4X9__NEONFMA, c_eq_4) { TEST_REQUIRES_ARM_NEON_FMA; DWConvMicrokernelTester() @@ -984,6 +1620,1278 @@ #if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_eq_4) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_div_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_lt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_gt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_eq_4) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_div_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_lt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_gt_4) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEONFMA_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_eq_8) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_div_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_lt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_gt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_eq_8) { + TEST_REQUIRES_ARM_NEON_FMA; + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_div_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_lt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_gt_8) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEONFMA_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON_FMA; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_eq_4) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_div_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_lt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_gt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_eq_4) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_div_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_lt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_gt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X25__NEON_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_eq_8) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_div_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_lt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_gt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_eq_8) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_div_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_lt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_gt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 25; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X25__NEON_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(25) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 TEST(F32_DWCONV_MINMAX_UP4X9__NEON, c_eq_4) { TEST_REQUIRES_ARM_NEON; DWConvMicrokernelTester() @@ -1619,6 +3527,642 @@ #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_eq_4) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_div_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_lt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_gt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_eq_4) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_div_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_div_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_div_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 8; channels < 64; channels += 12) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_lt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 4; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_gt_4) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_gt_4_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, c_gt_4_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 5; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(4) + .width(5) + .output_stride(23) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP4X4__NEON_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 20; channels += 3) { + DWConvMicrokernelTester() + .cr(4) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_eq_8) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_div_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_lt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_gt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + +#if XNN_ARCH_ARM || XNN_ARCH_ARM64 + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_eq_8) { + TEST_REQUIRES_ARM_NEON; + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_div_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_div_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_div_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 16; channels < 128; channels += 24) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_lt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 1; channels < 8; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_gt_8) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_gt_8_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, c_gt_8_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (uint32_t channels = 9; channels < 16; channels++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, multipixel) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, multipixel_with_step) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + for (size_t step = 2; step <= 4; step++) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .step(step) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, multipixel_with_output_stride) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(8) + .width(5) + .output_stride(43) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, multipixel_with_qmin) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmin(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } + + TEST(F32_DWCONV_MINMAX_UP8X4__NEON_ACC2, multipixel_with_qmax) { + TEST_REQUIRES_ARM_NEON; + for (size_t channels = 1; channels <= 40; channels += 7) { + DWConvMicrokernelTester() + .cr(8) + .kr(4) + .channels(channels) + .width(3) + .qmax(128) + .Test(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2); + } + } +#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 + + #if XNN_ARCH_X86 || XNN_ARCH_X86_64 TEST(F32_DWCONV_MINMAX_UP4X25__SSE, c_eq_4) { TEST_REQUIRES_X86_SSE; diff --git a/test/f32-dwconv-minmax.yaml b/test/f32-dwconv-minmax.yaml index 78036f55e..6a07f582c 100644 --- a/test/f32-dwconv-minmax.yaml +++ b/test/f32-dwconv-minmax.yaml @@ -6,14 +6,30 @@ - name: xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma_cortex_a55 pipelined: true assembly: true +- name: xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma +- name: xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma +- name: xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2 - name: xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma - name: xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2 - name: xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma - name: xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma +- name: xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma +- name: xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up4x25__neon +- name: xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up8x25__neon +- name: xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2 - name: xnn_f32_dwconv_minmax_ukernel_up4x9__neon - name: xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2 - name: xnn_f32_dwconv_minmax_ukernel_up8x9__neon - name: xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up4x4__neon +- name: xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2 +- name: xnn_f32_dwconv_minmax_ukernel_up8x4__neon +- name: xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2 - name: xnn_f32_dwconv_minmax_ukernel_up4x25__sse - name: xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2 - name: xnn_f32_dwconv_minmax_ukernel_up8x25__sse |