bfly4m1 remove multiplies by 1 and 0.

PiperOrigin-RevId: 469254069

11 files changed, 83 insertions, 104 deletions

diff --git a/BUILD.bazel b/BUILD.bazel
index d081abdb4..5e9dbbc16 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -618,7 +618,7 @@ PROD_SCALAR_RISCV_MICROKERNEL_SRCS = [
 ]
 
 ALL_SCALAR_MICROKERNEL_SRCS = [
-    "src/cs16-bfly4/gen/scalar-m1-x1.c",
+    "src/cs16-bfly4/scalar-m1.c",
     "src/cs16-bfly4/gen/scalar-x1.c",
     "src/cs16-bfly4/gen/scalar-x2.c",
     "src/cs16-bfly4/gen/scalar-x3.c",
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 903c48cf5..a3f336a9d 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -606,7 +606,7 @@ SET(PROD_SCALAR_RISCV_MICROKERNEL_SRCS
   src/x32-zip/x4-scalar.c)
 
 SET(ALL_SCALAR_MICROKERNEL_SRCS
-  src/cs16-bfly4/gen/scalar-m1-x1.c
+  src/cs16-bfly4/scalar-m1.c
   src/cs16-bfly4/gen/scalar-x1.c
   src/cs16-bfly4/gen/scalar-x2.c
   src/cs16-bfly4/gen/scalar-x3.c
diff --git a/bench/cs16-bfly4.cc b/bench/cs16-bfly4.cc
index 43b25150b..e4c46d2d9 100644
--- a/bench/cs16-bfly4.cc
+++ b/bench/cs16-bfly4.cc
@@ -74,7 +74,7 @@ static void BenchmarkM1KernelSize(benchmark::internal::Benchmark* b)
   b->Args({1024, 1, 256});
 }
 
-BENCHMARK_CAPTURE(cs16_bfly4, cs16_scalar_m1_x1, xnn_cs16_bfly4m1_ukernel__scalar_x1)->Apply(BenchmarkM1KernelSize)->UseRealTime();
+BENCHMARK_CAPTURE(cs16_bfly4, cs16_scalar_m1, xnn_cs16_bfly4m1_ukernel__scalar)->Apply(BenchmarkM1KernelSize)->UseRealTime();
 BENCHMARK_CAPTURE(cs16_bfly4, cs16_scalar_x1, xnn_cs16_bfly4_ukernel__scalar_x1)->Apply(BenchmarkKernelSize)->UseRealTime();
 BENCHMARK_CAPTURE(cs16_bfly4, cs16_scalar_x2, xnn_cs16_bfly4_ukernel__scalar_x2)->Apply(BenchmarkKernelSize)->UseRealTime();
 BENCHMARK_CAPTURE(cs16_bfly4, cs16_scalar_x3, xnn_cs16_bfly4_ukernel__scalar_x3)->Apply(BenchmarkKernelSize)->UseRealTime();
diff --git a/scripts/generate-cs16-bfly4.sh b/scripts/generate-cs16-bfly4.sh
index 185c0fd13..992f61162 100755
--- a/scripts/generate-cs16-bfly4.sh
+++ b/scripts/generate-cs16-bfly4.sh
@@ -10,9 +10,6 @@ tools/xngen src/cs16-bfly4/scalar.c.in -D SAMPLE_TILE=2 -D M=0 -o src/cs16-bfly4
 tools/xngen src/cs16-bfly4/scalar.c.in -D SAMPLE_TILE=3 -D M=0 -o src/cs16-bfly4/gen/scalar-x3.c &
 tools/xngen src/cs16-bfly4/scalar.c.in -D SAMPLE_TILE=4 -D M=0 -o src/cs16-bfly4/gen/scalar-x4.c &
 
-tools/xngen src/cs16-bfly4/scalar.c.in -D SAMPLE_TILE=1 -D M=1 -o src/cs16-bfly4/gen/scalar-m1-x1.c &
-
-
 ################################## Unit tests #################################
 tools/generate-bfly4-test.py --spec test/cs16-bfly4.yaml --output test/cs16-bfly4.cc &
 
diff --git a/src/cs16-bfly4/gen/scalar-m1-x1.c b/src/cs16-bfly4/gen/scalar-m1-x1.c
deleted file mode 100644
index 09c47ff40..000000000
--- a/src/cs16-bfly4/gen/scalar-m1-x1.c
+++ /dev/null
@@ -1,91 +0,0 @@
-// Auto-generated file. Do not edit!
-//   Template: src/cs16-bfly4/scalar.c.in
-//   Generator: tools/xngen
-//
-// Copyright 2022 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 <stddef.h>
-#include <stdint.h>
-
-#include <xnnpack/math.h>
-#include <xnnpack/fft.h>
-
-
-void xnn_cs16_bfly4m1_ukernel__scalar_x1(
-    size_t samples,
-    int16_t* data,
-    const size_t stride,
-    const int16_t* twiddle) {
-
-  int16_t* out0 = data;
-
-  assert(samples == 1);
-  assert(data != NULL);
-  assert(stride != 0);
-  assert(twiddle != NULL);
-
-
-  if XNN_UNLIKELY(samples != 0) {
-    do {
-      int32_t vout0r = (int32_t) out0[0];
-      int32_t vout0i = (int32_t) out0[1];
-      int32_t vout1r = (int32_t) out0[2];
-      int32_t vout1i = (int32_t) out0[3];
-      int32_t vout2r = (int32_t) out0[4];
-      int32_t vout2i = (int32_t) out0[5];
-      int32_t vout3r = (int32_t) out0[6];
-      int32_t vout3i = (int32_t) out0[7];
-
-
-      // Note 32767 / 4 = 8191.  Should be 8192.
-      vout0r = math_asr_s32(vout0r * 8191 + 16384, 15);
-      vout0i = math_asr_s32(vout0i * 8191 + 16384, 15);
-      vout1r = math_asr_s32(vout1r * 8191 + 16384, 15);
-      vout1i = math_asr_s32(vout1i * 8191 + 16384, 15);
-      vout2r = math_asr_s32(vout2r * 8191 + 16384, 15);
-      vout2i = math_asr_s32(vout2i * 8191 + 16384, 15);
-      vout3r = math_asr_s32(vout3r * 8191 + 16384, 15);
-      vout3i = math_asr_s32(vout3i * 8191 + 16384, 15);
-
-      // Note 32767 should be 32768 representing a multiply by 1.
-      const int32_t vtmp0r = math_asr_s32(vout1r * 32767 + 16384, 15);
-      const int32_t vtmp0i = math_asr_s32(vout1i * 32767 + 16384, 15);
-      const int32_t vtmp1r = math_asr_s32(vout2r * 32767 + 16384, 15);
-      const int32_t vtmp1i = math_asr_s32(vout2i * 32767 + 16384, 15);
-      const int32_t vtmp2r = math_asr_s32(vout3r * 32767 + 16384, 15);
-      const int32_t vtmp2i = math_asr_s32(vout3i * 32767 + 16384, 15);
-
-      const int32_t vtmp5r = vout0r - vtmp1r;
-      const int32_t vtmp5i = vout0i - vtmp1i;
-      vout0r  += vtmp1r;
-      vout0i  += vtmp1i;
-      const int32_t vtmp3r = vtmp0r + vtmp2r;
-      const int32_t vtmp3i = vtmp0i + vtmp2i;
-      const int32_t vtmp4r = vtmp0r - vtmp2r;
-      const int32_t vtmp4i = vtmp0i - vtmp2i;
-      vout2r = vout0r - vtmp3r;
-      vout2i = vout0i - vtmp3i;
-
-      vout0r += vtmp3r;
-      vout0i += vtmp3i;
-
-      vout1r = vtmp5r + vtmp4i;
-      vout1i = vtmp5i - vtmp4r;
-      vout3r = vtmp5r - vtmp4i;
-      vout3i = vtmp5i + vtmp4r;
-
-      out0[0] = (int16_t) vout0r;
-      out0[1] = (int16_t) vout0i;
-      out0[2] = (int16_t) vout1r;
-      out0[3] = (int16_t) vout1i;
-      out0[4] = (int16_t) vout2r;
-      out0[5] = (int16_t) vout2i;
-      out0[6] = (int16_t) vout3r;
-      out0[7] = (int16_t) vout3i;
-    } while(--samples != 0);
-  }
-}
diff --git a/src/cs16-bfly4/scalar-m1.c b/src/cs16-bfly4/scalar-m1.c
new file mode 100644
index 000000000..d0f2af509
--- /dev/null
+++ b/src/cs16-bfly4/scalar-m1.c
@@ -0,0 +1,71 @@
+// Copyright 2022 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 <stddef.h>
+#include <stdint.h>
+
+#include <xnnpack/math.h>
+#include <xnnpack/fft.h>
+
+
+void xnn_cs16_bfly4m1_ukernel__scalar(
+    size_t samples,
+    int16_t* data,
+    const size_t stride,
+    const int16_t* twiddle) {
+
+  assert(samples == 1);
+  assert(data != NULL);
+  assert(stride != 0);
+  assert(twiddle != NULL);
+
+  int32_t vout0r = (int32_t) data[0];
+  int32_t vout0i = (int32_t) data[1];
+  int32_t vout1r = (int32_t) data[2];
+  int32_t vout1i = (int32_t) data[3];
+  int32_t vout2r = (int32_t) data[4];
+  int32_t vout2i = (int32_t) data[5];
+  int32_t vout3r = (int32_t) data[6];
+  int32_t vout3i = (int32_t) data[7];
+
+  // Note 32767 / 4 = 8191.  Should be 8192.
+  vout0r = math_asr_s32(vout0r * 8191 + 16384, 15);
+  vout0i = math_asr_s32(vout0i * 8191 + 16384, 15);
+  vout1r = math_asr_s32(vout1r * 8191 + 16384, 15);
+  vout1i = math_asr_s32(vout1i * 8191 + 16384, 15);
+  vout2r = math_asr_s32(vout2r * 8191 + 16384, 15);
+  vout2i = math_asr_s32(vout2i * 8191 + 16384, 15);
+  vout3r = math_asr_s32(vout3r * 8191 + 16384, 15);
+  vout3i = math_asr_s32(vout3i * 8191 + 16384, 15);
+
+  const int32_t vtmp5r = vout0r - vout2r;
+  const int32_t vtmp5i = vout0i - vout2i;
+  vout0r  += vout2r;
+  vout0i  += vout2i;
+  const int32_t vtmp3r = vout1r + vout3r;
+  const int32_t vtmp3i = vout1i + vout3i;
+  const int32_t vtmp4r = vout1r - vout3r;
+  const int32_t vtmp4i = vout1i - vout3i;
+  vout2r = vout0r - vtmp3r;
+  vout2i = vout0i - vtmp3i;
+
+  vout0r += vtmp3r;
+  vout0i += vtmp3i;
+
+  vout1r = vtmp5r + vtmp4i;
+  vout1i = vtmp5i - vtmp4r;
+  vout3r = vtmp5r - vtmp4i;
+  vout3i = vtmp5i + vtmp4r;
+
+  data[0] = (int16_t) vout0r;
+  data[1] = (int16_t) vout0i;
+  data[2] = (int16_t) vout1r;
+  data[3] = (int16_t) vout1i;
+  data[4] = (int16_t) vout2r;
+  data[5] = (int16_t) vout2i;
+  data[6] = (int16_t) vout3r;
+  data[7] = (int16_t) vout3i;
+}
diff --git a/src/xnnpack/fft.h b/src/xnnpack/fft.h
index 50bb7a2e9..9aa0afd79 100644
--- a/src/xnnpack/fft.h
+++ b/src/xnnpack/fft.h
@@ -26,7 +26,7 @@ DECLARE_CS16_BFLY4_UKERNEL_FUNCTION(xnn_cs16_bfly4_ukernel__scalar_x1)
 DECLARE_CS16_BFLY4_UKERNEL_FUNCTION(xnn_cs16_bfly4_ukernel__scalar_x2)
 DECLARE_CS16_BFLY4_UKERNEL_FUNCTION(xnn_cs16_bfly4_ukernel__scalar_x3)
 DECLARE_CS16_BFLY4_UKERNEL_FUNCTION(xnn_cs16_bfly4_ukernel__scalar_x4)
-DECLARE_CS16_BFLY4_UKERNEL_FUNCTION(xnn_cs16_bfly4m1_ukernel__scalar_x1)
+DECLARE_CS16_BFLY4_UKERNEL_FUNCTION(xnn_cs16_bfly4m1_ukernel__scalar)
 
 #define DECLARE_CS16_FFTR_UKERNEL_FUNCTION(fn_name) \
   XNN_INTERNAL void fn_name(                        \
diff --git a/test/bfly4-microkernel-tester.h b/test/bfly4-microkernel-tester.h
index 32bf9bb77..ae3182b4a 100644
--- a/test/bfly4-microkernel-tester.h
+++ b/test/bfly4-microkernel-tester.h
@@ -215,7 +215,7 @@ class BFly4MicrokernelTester {
     std::random_device random_device;
     auto rng = std::mt19937(random_device());
     auto i16rng = std::bind(std::uniform_int_distribution<int16_t>(), std::ref(rng));
-    const size_t fft_size = samples() * stride() * 4;  // 4 for bfly4.
+    const size_t fft_size = samples() == 1 ? 1 : (samples() * stride()) * 4;  // 4 for bfly4.
 
     // 256 complex numbers = fft_size * 2 = 512
     std::vector<int16_t> y(fft_size * 2 + XNN_EXTRA_BYTES / sizeof(int16_t));
diff --git a/test/cs16-bfly4.cc b/test/cs16-bfly4.cc
index 4815f35f8..a3217ec78 100644
--- a/test/cs16-bfly4.cc
+++ b/test/cs16-bfly4.cc
@@ -133,9 +133,9 @@ TEST(CS16_BFLY4__SCALAR_X4, samples_eq_64) {
 }
 
 
-TEST(CS16_BFLY4M1__SCALAR_X1, samples_eq_1) {
+TEST(CS16_BFLY4M1__SCALAR, samples_eq_1) {
   BFly4MicrokernelTester()
     .samples(1)
     .stride(64)
-    .Test(xnn_cs16_bfly4m1_ukernel__scalar_x1);
+    .Test(xnn_cs16_bfly4m1_ukernel__scalar);
 }
diff --git a/test/cs16-bfly4.yaml b/test/cs16-bfly4.yaml
index a8d2f129b..a9ddc8e58 100644
--- a/test/cs16-bfly4.yaml
+++ b/test/cs16-bfly4.yaml
@@ -9,4 +9,4 @@
 - name: xnn_cs16_bfly4_ukernel__scalar_x2
 - name: xnn_cs16_bfly4_ukernel__scalar_x3
 - name: xnn_cs16_bfly4_ukernel__scalar_x4
-- name: xnn_cs16_bfly4m1_ukernel__scalar_x1
+- name: xnn_cs16_bfly4m1_ukernel__scalar
diff --git a/tools/generate-bfly4-test.py b/tools/generate-bfly4-test.py
index 3160bbf6a..1bd2a31a7 100755
--- a/tools/generate-bfly4-test.py
+++ b/tools/generate-bfly4-test.py
@@ -28,11 +28,13 @@ parser.set_defaults(defines=list())
 
 def split_ukernel_name(name):
   m = 0
-  match = re.fullmatch(r"xnn_cs16_bfly4(m(\d+))?_ukernel__(.+)_x(\d+)", name)
+  samples_tile = 1
+  match = re.fullmatch(r"xnn_cs16_bfly4(m(\d+))?_ukernel__(.+)(_x(\d+))?", name)
   assert match is not None
   if match.group(2):
     m = int(match.group(2))
-  samples_tile = int(match.group(4))
+  if match.group(5):
+    samples_tile = int(match.group(5))
 
   arch, isa = xnncommon.parse_target_name(target_name=match.group(3))
   return m, samples_tile, arch, isa