Add Arm NEON implementation of h2v1_downsample

Adds an Arm NEON intrinsics implementation of h2v1_downsample. This
intrinsics implementation is used to generate code for AArch32 and
AArch64; previously there was only an AArch64 NEON assembly
implementation.

Removes the AArch64 NEON assembly implementation from jsimd_neon.S.

Bug: 922430
Change-Id: I58a83635959fa1dd4923cecdef64792ebb57b54f

5 files changed, 141 insertions, 76 deletions

diff --git a/BUILD.gn b/BUILD.gn
index 4c45afc..940bf9e 100644
--- a/BUILD.gn
+++ b/BUILD.gn
@@ -155,6 +155,7 @@ static_library("simd") {
     sources = [
       "simd/arm/arm/jsimd.c",
       "simd/arm/arm/jsimd_neon.S",
+      "simd/arm/common/jcsample-neon.c",
       "simd/arm/common/jdcolor-neon.c",
       "simd/arm/common/jdmerge-neon.c",
       "simd/arm/common/jdsample-neon.c",
@@ -168,6 +169,7 @@ static_library("simd") {
     sources = [
       "simd/arm/arm64/jsimd.c",
       "simd/arm/arm64/jsimd_neon.S",
+      "simd/arm/common/jcsample-neon.c",
       "simd/arm/common/jdcolor-neon.c",
       "simd/arm/common/jdmerge-neon.c",
       "simd/arm/common/jdsample-neon.c",
diff --git a/README.chromium b/README.chromium
index f70cb89..38a63b9 100644
--- a/README.chromium
+++ b/README.chromium
@@ -71,6 +71,7 @@ following changes which are not merged to upstream:
   - Implement slow IDCT using Arm NEON intrinsics
   - Precompute DCT block output pointers in IDCT functions
   - Implement fast IDCT using Arm NEON intrinsics
+  - Add Arm NEON implementation of h2v1_downsample
 * Patches to enable running the upstream unit tests through gtest.
   The upstream unit tests are defined here under the section 'TESTS':
   https://github.com/libjpeg-turbo/libjpeg-turbo/blob/master/CMakeLists.txt
diff --git a/simd/arm/arm/jsimd.c b/simd/arm/arm/jsimd.c
index ad176dc..2c5f4b9 100644
--- a/simd/arm/arm/jsimd.c
+++ b/simd/arm/arm/jsimd.c
@@ -304,6 +304,19 @@ jsimd_can_h2v2_downsample(void)
 GLOBAL(int)
 jsimd_can_h2v1_downsample(void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
   return 0;
 }
 
@@ -317,6 +330,9 @@ GLOBAL(void)
 jsimd_h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
                       JSAMPARRAY input_data, JSAMPARRAY output_data)
 {
+  jsimd_h2v1_downsample_neon(cinfo->image_width, cinfo->max_v_samp_factor,
+                             compptr->v_samp_factor, compptr->width_in_blocks,
+                             input_data, output_data);
 }
 
 GLOBAL(int)
diff --git a/simd/arm/arm64/jsimd_neon.S b/simd/arm/arm64/jsimd_neon.S
index 020353f..0adea9d 100644
--- a/simd/arm/arm64/jsimd_neon.S
+++ b/simd/arm/arm64/jsimd_neon.S
@@ -1118,82 +1118,6 @@ asm_function jsimd_quantize_neon
 
 /*
  * Downsample pixel values of a single component.
- * This version handles the common case of 2:1 horizontal and 1:1 vertical,
- * without smoothing.
- *
- * GLOBAL(void)
- * jsimd_h2v1_downsample_neon(JDIMENSION image_width, int max_v_samp_factor,
- *                            JDIMENSION v_samp_factor,
- *                            JDIMENSION width_in_blocks,
- *                            JSAMPARRAY input_data, JSAMPARRAY output_data);
- */
-
-asm_function jsimd_h2v1_downsample_neon
-    IMAGE_WIDTH     .req x0
-    MAX_V_SAMP      .req x1
-    V_SAMP          .req x2
-    BLOCK_WIDTH     .req x3
-    INPUT_DATA      .req x4
-    OUTPUT_DATA     .req x5
-    OUTPTR          .req x9
-    INPTR           .req x10
-    TMP1            .req x11
-    TMP2            .req x12
-    TMP3            .req x13
-    TMPDUP          .req w15
-
-    mov             TMPDUP, #0x10000
-    lsl             TMP2, BLOCK_WIDTH, #4
-    sub             TMP2, TMP2, IMAGE_WIDTH
-    get_symbol_loc  TMP3, Ljsimd_h2_downsample_neon_consts
-    add             TMP3, TMP3, TMP2, lsl #4
-    dup             v16.4s, TMPDUP
-    ld1             {v18.16b}, [TMP3]
-
-1:  /* row loop */
-    ldr             INPTR, [INPUT_DATA], #8
-    ldr             OUTPTR, [OUTPUT_DATA], #8
-    subs            TMP1, BLOCK_WIDTH, #1
-    b.eq            3f
-2:  /* columns */
-    ld1             {v0.16b}, [INPTR], #16
-    mov             v4.16b, v16.16b
-    subs            TMP1, TMP1, #1
-    uadalp          v4.8h, v0.16b
-    shrn            v6.8b, v4.8h, #1
-    st1             {v6.8b}, [OUTPTR], #8
-    b.ne            2b
-3:  /* last columns */
-    ld1             {v0.16b}, [INPTR]
-    mov             v4.16b, v16.16b
-    subs            V_SAMP, V_SAMP, #1
-    /* expand right */
-    tbl             v2.16b, {v0.16b}, v18.16b
-    uadalp          v4.8h, v2.16b
-    shrn            v6.8b, v4.8h, #1
-    st1             {v6.8b}, [OUTPTR], #8
-    b.ne            1b
-
-    br              x30
-
-    .unreq          IMAGE_WIDTH
-    .unreq          MAX_V_SAMP
-    .unreq          V_SAMP
-    .unreq          BLOCK_WIDTH
-    .unreq          INPUT_DATA
-    .unreq          OUTPUT_DATA
-    .unreq          OUTPTR
-    .unreq          INPTR
-    .unreq          TMP1
-    .unreq          TMP2
-    .unreq          TMP3
-    .unreq          TMPDUP
-
-
-/*****************************************************************************/
-
-/*
- * Downsample pixel values of a single component.
  * This version handles the common case of 2:1 horizontal and 2:1 vertical,
  * without smoothing.
  *
diff --git a/simd/arm/common/jcsample-neon.c b/simd/arm/common/jcsample-neon.c
new file mode 100644
index 0000000..2b110ea
--- /dev/null
+++ b/simd/arm/common/jcsample-neon.c
@@ -0,0 +1,122 @@
+/*
+ * jcsample-neon.c - downsampling (Arm NEON)
+ *
+ * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
+#include "../../jsimd.h"
+
+#include <arm_neon.h>
+
+
+static const uint8_t jsimd_h2_downsample_consts[] = {
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 0 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 1 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0E,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 2 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0D, 0x0D,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 3 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0C, 0x0C, 0x0C,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 4 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0B, 0x0B, 0x0B, 0x0B,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 5 */
+  0x08, 0x09, 0x0A, 0x0A, 0x0A, 0x0A, 0x0A, 0x0A,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 6 */
+  0x08, 0x09, 0x09, 0x09, 0x09, 0x09, 0x09, 0x09,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 7 */
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 8 */
+  0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x06,   /* Pad 9 */
+  0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x05, 0x05,   /* Pad 10 */
+  0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x04, 0x04, 0x04,   /* Pad 11 */
+  0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
+  0x00, 0x01, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03,   /* Pad 12 */
+  0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03,
+  0x00, 0x01, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,   /* Pad 13 */
+  0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+  0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,   /* Pad 14 */
+  0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,   /* Pad 15 */
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
+};
+
+
+/*
+ * Downsample pixel values of a single chroma component i.e. Cb, Cr.
+ * This version handles the common case of 2:1 horizontal and 1:1 vertical,
+ * without smoothing.
+ */
+
+void jsimd_h2v1_downsample_neon(JDIMENSION image_width,
+                                int max_v_samp_factor,
+                                JDIMENSION v_samp_factor,
+                                JDIMENSION width_in_blocks,
+                                JSAMPARRAY input_data,
+                                JSAMPARRAY output_data)
+{
+  JSAMPROW inptr, outptr;
+  /* Load expansion mask to pad remaining elements of last DCT block. */
+  const int mask_offset = 16 * ((width_in_blocks * 2 * DCTSIZE) - image_width);
+  const uint8x16_t expand_mask = vld1q_u8(
+                                    &jsimd_h2_downsample_consts[mask_offset]);
+  /* Load bias pattern alternating every pixel. */
+  const uint16x8_t bias = { 0, 1, 0, 1, 0, 1, 0, 1 };
+
+  for (unsigned outrow = 0; outrow < v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+
+    /* Downsample all but the last DCT block of pixels. */
+    for (unsigned i = 0; i < width_in_blocks - 1; i++) {
+      uint8x16_t pixels = vld1q_u8(inptr + i * 2 * DCTSIZE);
+      /* Add adjacent pixel values, widen to 16-bit and add bias. */
+      uint16x8_t samples_u16 = vpadalq_u8(bias, pixels);
+      /* Divide total by 2 and narrow to 8-bit. */
+      uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 1);
+      /* Store samples to memory. */
+      vst1_u8(outptr + i * DCTSIZE, samples_u8);
+    }
+
+    /* Load pixels in last DCT block into a table. */
+    uint8x16_t pixels = vld1q_u8(inptr + (width_in_blocks - 1) * 2 * DCTSIZE);
+#if defined(__aarch64__)
+    /* Pad the empty elements with the value of the last pixel. */
+    pixels = vqtbl1q_u8(pixels, expand_mask);
+#else
+    uint8x8x2_t table = { vget_low_u8(pixels), vget_high_u8(pixels) };
+    pixels = vcombine_u8(vtbl2_u8(table, vget_low_u8(expand_mask)),
+                         vtbl2_u8(table, vget_high_u8(expand_mask)));
+#endif
+    /* Add adjacent pixel values, widen to 16-bit and add bias. */
+    uint16x8_t samples_u16 = vpadalq_u8(bias, pixels);
+    /* Divide total by 2, narrow to 8-bit and store. */
+    uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 1);
+    vst1_u8(outptr + (width_in_blocks - 1) * DCTSIZE, samples_u8);
+  }
+}