Add Arm NEON implementation of h2v2_upsample

Adds an Arm NEON intrinsics implementation of h2v2_upsample. This is
new code, there was no previous NEON assembly implementation for
either AArch32 or AArch64.

Bug: 922430
Change-Id: Iaed1202e0c77e4d9e6b5dbd522d2fb0c851878b2

4 files changed, 71 insertions, 0 deletions

diff --git a/simd/arm/arm/jsimd.c b/simd/arm/arm/jsimd.c
index 541cdfed..4bf39319 100644
--- a/simd/arm/arm/jsimd.c
+++ b/simd/arm/arm/jsimd.c
@@ -322,6 +322,17 @@ jsimd_h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
 GLOBAL(int)
 jsimd_can_h2v2_upsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
@@ -346,6 +357,8 @@ GLOBAL(void)
 jsimd_h2v2_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
                     JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
 {
+  jsimd_h2v2_upsample_neon(cinfo->max_v_samp_factor, cinfo->output_width,
+                           input_data, output_data_ptr);
 }
 
 GLOBAL(void)
diff --git a/simd/arm/arm64/jsimd.c b/simd/arm/arm64/jsimd.c
index 5481dad7..7ce66470 100644
--- a/simd/arm/arm64/jsimd.c
+++ b/simd/arm/arm64/jsimd.c
@@ -397,6 +397,17 @@ jsimd_h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
 GLOBAL(int)
 jsimd_can_h2v2_upsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
@@ -421,6 +432,8 @@ GLOBAL(void)
 jsimd_h2v2_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
                     JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
 {
+  jsimd_h2v2_upsample_neon(cinfo->max_v_samp_factor, cinfo->output_width,
+                           input_data, output_data_ptr);
 }
 
 GLOBAL(void)
diff --git a/simd/arm/common/jdsample-neon.c b/simd/arm/common/jdsample-neon.c
index 8f4fd7c5..e4f5129d 100644
--- a/simd/arm/common/jdsample-neon.c
+++ b/simd/arm/common/jdsample-neon.c
@@ -513,3 +513,45 @@ void jsimd_h2v1_upsample_neon(int max_v_samp_factor,
     }
   }
 }
+
+
+/*
+ * The diagram below shows the operation of h2v2 (simple) upsampling. Each
+ * sample in the row is duplicated to form two output pixel channel values.
+ * This horizontally-upsampled row is then also duplicated.
+ *
+ *                                  p0    p1    p2    p3
+ *       +-----+-----+           +-----+-----+-----+-----+
+ *       |  s0 |  s1 |    ->     |  s0 |  s0 |  s1 |  s1 |
+ *       +-----+-----+           +-----+-----+-----+-----+
+ *                               |  s0 |  s0 |  s1 |  s1 |
+ *                               +-----+-----+-----+-----+
+ */
+
+void jsimd_h2v2_upsample_neon(int max_v_samp_factor,
+                              JDIMENSION output_width,
+                              JSAMPARRAY input_data,
+                              JSAMPARRAY *output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr0, outptr1;
+
+  for (int inrow = 0, outrow = 0; outrow < max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr0 = output_data[outrow++];
+    outptr1 = output_data[outrow++];
+
+    for (unsigned colctr = 0; 2 * colctr < output_width; colctr += 16) {
+      uint8x16_t samples = vld1q_u8(inptr + colctr);
+      /* Duplicate the samples - the store interleaves them to produce the */
+      /* pattern in the diagram above. */
+      uint8x16x2_t output_pixels = { samples, samples };
+      /* Store pixel values to memory for both output rows. */
+      /* Due to the way sample buffers are allocated, we don't need to worry */
+      /* about tail cases when output_width is not a multiple of 32. */
+      /* See "Creation of 2-D sample arrays" in jmemmgr.c for details. */
+      vst2q_u8(outptr0 + 2 * colctr, output_pixels);
+      vst2q_u8(outptr1 + 2 * colctr, output_pixels);
+    }
+  }
+}
diff --git a/simd/jsimd.h b/simd/jsimd.h
index 48726034..9b5781c2 100644
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -565,6 +565,9 @@ EXTERN(void) jsimd_h2v2_upsample_avx2
 EXTERN(void) jsimd_h2v1_upsample_neon
   (int max_v_samp_factor, JDIMENSION output_width, JSAMPARRAY input_data,
    JSAMPARRAY *output_data_ptr);
+EXTERN(void) jsimd_h2v2_upsample_neon
+  (int max_v_samp_factor, JDIMENSION output_width, JSAMPARRAY input_data,
+   JSAMPARRAY *output_data_ptr);
 
 EXTERN(void) jsimd_h2v1_upsample_dspr2
   (int max_v_samp_factor, JDIMENSION output_width, JSAMPARRAY input_data,