Use run-time CPU feature detection for Neon DotProd SAD4D

Arm Neon DotProd implementations of vpx_sad*4d currently need to be
enabled at compile time since they're guarded by ifdef feature
macros. Now that run-time feature detection has been enabled for Arm
platforms, expose these implementations with distinct *neon_dotprod
names in separate files and wire them up to the build system and
rtcd.pl. Also add new test cases for the new DotProd functions.

Change-Id: Ie99ee0b03ec488626f52c3f13e4111fe26cc5619

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 3f9c020..3530e66 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -1346,6 +1346,21 @@ const SadMxNx4Param x4d_neon_tests[] = {
 };
 INSTANTIATE_TEST_SUITE_P(NEON, SADx4Test, ::testing::ValuesIn(x4d_neon_tests));
 
+#if HAVE_NEON_DOTPROD
+const SadMxNx4Param x4d_neon_dotprod_tests[] = {
+  SadMxNx4Param(64, 64, &vpx_sad64x64x4d_neon_dotprod),
+  SadMxNx4Param(64, 32, &vpx_sad64x32x4d_neon_dotprod),
+  SadMxNx4Param(32, 64, &vpx_sad32x64x4d_neon_dotprod),
+  SadMxNx4Param(32, 32, &vpx_sad32x32x4d_neon_dotprod),
+  SadMxNx4Param(32, 16, &vpx_sad32x16x4d_neon_dotprod),
+  SadMxNx4Param(16, 32, &vpx_sad16x32x4d_neon_dotprod),
+  SadMxNx4Param(16, 16, &vpx_sad16x16x4d_neon_dotprod),
+  SadMxNx4Param(16, 8, &vpx_sad16x8x4d_neon_dotprod),
+};
+INSTANTIATE_TEST_SUITE_P(NEON_DOTPROD, SADx4Test,
+                         ::testing::ValuesIn(x4d_neon_dotprod_tests));
+#endif  // HAVE_NEON_DOTPROD
+
 const SadSkipMxNx4Param skip_x4d_neon_tests[] = {
   SadSkipMxNx4Param(64, 64, &vpx_sad_skip_64x64x4d_neon),
   SadSkipMxNx4Param(64, 32, &vpx_sad_skip_64x32x4d_neon),
@@ -1401,6 +1416,21 @@ const SadSkipMxNx4Param skip_x4d_neon_tests[] = {
 };
 INSTANTIATE_TEST_SUITE_P(NEON, SADSkipx4Test,
                          ::testing::ValuesIn(skip_x4d_neon_tests));
+
+#if HAVE_NEONE_DOTPROD
+const SadSkipMxNx4Param skip_x4d_neon_dotprod_tests[] = {
+  SadSkipMxNx4Param(64, 64, &vpx_sad_skip_64x64x4d_neon_dotprod),
+  SadSkipMxNx4Param(64, 32, &vpx_sad_skip_64x32x4d_neon_dotprod),
+  SadSkipMxNx4Param(32, 64, &vpx_sad_skip_32x64x4d_neon_dotprod),
+  SadSkipMxNx4Param(32, 32, &vpx_sad_skip_32x32x4d_neon_dotprod),
+  SadSkipMxNx4Param(32, 16, &vpx_sad_skip_32x16x4d_neon_dotprod),
+  SadSkipMxNx4Param(16, 32, &vpx_sad_skip_16x32x4d_neon_dotprod),
+  SadSkipMxNx4Param(16, 16, &vpx_sad_skip_16x16x4d_neon_dotprod),
+  SadSkipMxNx4Param(16, 8, &vpx_sad_skip_16x8x4d_neon_dotprod),
+};
+INSTANTIATE_TEST_SUITE_P(NEON_DOTPROD, SADSkipx4Test,
+                         ::testing::ValuesIn(skip_x4d_neon_dotprod_tests));
+#endif  // HAVE_NEON_DOTPROD
 #endif  // HAVE_NEON
 
 //------------------------------------------------------------------------------

diff --git a/vpx_dsp/arm/sad4d_neon.c b/vpx_dsp/arm/sad4d_neon.c
index 3a548d0..713eec7 100644 (file)
--- a/vpx_dsp/arm/sad4d_neon.c
+++ b/vpx_dsp/arm/sad4d_neon.c
@@ -17,120 +17,6 @@
 #include "vpx_dsp/arm/mem_neon.h"
 #include "vpx_dsp/arm/sum_neon.h"
 
-#if defined(__ARM_FEATURE_DOTPROD)
-
-static INLINE void sad16_neon(uint8x16_t src, uint8x16_t ref,
-                              uint32x4_t *const sad_sum) {
-  uint8x16_t abs_diff = vabdq_u8(src, ref);
-  *sad_sum = vdotq_u32(*sad_sum, abs_diff, vdupq_n_u8(1));
-}
-
-static INLINE void sad64xhx4d_neon(const uint8_t *src, int src_stride,
-                                   const uint8_t *const ref[4], int ref_stride,
-                                   uint32_t res[4], int h) {
-  uint32x4_t sum_lo[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
-                           vdupq_n_u32(0) };
-  uint32x4_t sum_hi[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
-                           vdupq_n_u32(0) };
-  uint32x4_t sum[4];
-
-  int i = 0;
-  do {
-    uint8x16_t s0, s1, s2, s3;
-
-    s0 = vld1q_u8(src + i * src_stride);
-    sad16_neon(s0, vld1q_u8(ref[0] + i * ref_stride), &sum_lo[0]);
-    sad16_neon(s0, vld1q_u8(ref[1] + i * ref_stride), &sum_lo[1]);
-    sad16_neon(s0, vld1q_u8(ref[2] + i * ref_stride), &sum_lo[2]);
-    sad16_neon(s0, vld1q_u8(ref[3] + i * ref_stride), &sum_lo[3]);
-
-    s1 = vld1q_u8(src + i * src_stride + 16);
-    sad16_neon(s1, vld1q_u8(ref[0] + i * ref_stride + 16), &sum_hi[0]);
-    sad16_neon(s1, vld1q_u8(ref[1] + i * ref_stride + 16), &sum_hi[1]);
-    sad16_neon(s1, vld1q_u8(ref[2] + i * ref_stride + 16), &sum_hi[2]);
-    sad16_neon(s1, vld1q_u8(ref[3] + i * ref_stride + 16), &sum_hi[3]);
-
-    s2 = vld1q_u8(src + i * src_stride + 32);
-    sad16_neon(s2, vld1q_u8(ref[0] + i * ref_stride + 32), &sum_lo[0]);
-    sad16_neon(s2, vld1q_u8(ref[1] + i * ref_stride + 32), &sum_lo[1]);
-    sad16_neon(s2, vld1q_u8(ref[2] + i * ref_stride + 32), &sum_lo[2]);
-    sad16_neon(s2, vld1q_u8(ref[3] + i * ref_stride + 32), &sum_lo[3]);
-
-    s3 = vld1q_u8(src + i * src_stride + 48);
-    sad16_neon(s3, vld1q_u8(ref[0] + i * ref_stride + 48), &sum_hi[0]);
-    sad16_neon(s3, vld1q_u8(ref[1] + i * ref_stride + 48), &sum_hi[1]);
-    sad16_neon(s3, vld1q_u8(ref[2] + i * ref_stride + 48), &sum_hi[2]);
-    sad16_neon(s3, vld1q_u8(ref[3] + i * ref_stride + 48), &sum_hi[3]);
-
-    i++;
-  } while (i < h);
-
-  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
-  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
-  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
-  sum[3] = vaddq_u32(sum_lo[3], sum_hi[3]);
-
-  vst1q_u32(res, horizontal_add_4d_uint32x4(sum));
-}
-
-static INLINE void sad32xhx4d_neon(const uint8_t *src, int src_stride,
-                                   const uint8_t *const ref[4], int ref_stride,
-                                   uint32_t res[4], int h) {
-  uint32x4_t sum_lo[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
-                           vdupq_n_u32(0) };
-  uint32x4_t sum_hi[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
-                           vdupq_n_u32(0) };
-  uint32x4_t sum[4];
-
-  int i = 0;
-  do {
-    uint8x16_t s0, s1;
-
-    s0 = vld1q_u8(src + i * src_stride);
-    sad16_neon(s0, vld1q_u8(ref[0] + i * ref_stride), &sum_lo[0]);
-    sad16_neon(s0, vld1q_u8(ref[1] + i * ref_stride), &sum_lo[1]);
-    sad16_neon(s0, vld1q_u8(ref[2] + i * ref_stride), &sum_lo[2]);
-    sad16_neon(s0, vld1q_u8(ref[3] + i * ref_stride), &sum_lo[3]);
-
-    s1 = vld1q_u8(src + i * src_stride + 16);
-    sad16_neon(s1, vld1q_u8(ref[0] + i * ref_stride + 16), &sum_hi[0]);
-    sad16_neon(s1, vld1q_u8(ref[1] + i * ref_stride + 16), &sum_hi[1]);
-    sad16_neon(s1, vld1q_u8(ref[2] + i * ref_stride + 16), &sum_hi[2]);
-    sad16_neon(s1, vld1q_u8(ref[3] + i * ref_stride + 16), &sum_hi[3]);
-
-    i++;
-  } while (i < h);
-
-  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
-  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
-  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
-  sum[3] = vaddq_u32(sum_lo[3], sum_hi[3]);
-
-  vst1q_u32(res, horizontal_add_4d_uint32x4(sum));
-}
-
-static INLINE void sad16xhx4d_neon(const uint8_t *src, int src_stride,
-                                   const uint8_t *const ref[4], int ref_stride,
-                                   uint32_t res[4], int h) {
-  uint32x4_t sum[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
-                        vdupq_n_u32(0) };
-
-  int i = 0;
-  do {
-    const uint8x16_t s = vld1q_u8(src + i * src_stride);
-    sad16_neon(s, vld1q_u8(ref[0] + i * ref_stride), &sum[0]);
-    sad16_neon(s, vld1q_u8(ref[1] + i * ref_stride), &sum[1]);
-    sad16_neon(s, vld1q_u8(ref[2] + i * ref_stride), &sum[2]);
-    sad16_neon(s, vld1q_u8(ref[3] + i * ref_stride), &sum[3]);
-
-    i++;
-  } while (i < h);
-
-  vst1q_u32(res, horizontal_add_4d_uint32x4(sum));
-}
-
-#else  // !defined(__ARM_FEATURE_DOTPROD))
-
 static INLINE void sad16_neon(uint8x16_t src, uint8x16_t ref,
                               uint16x8_t *const sad_sum) {
   uint8x16_t abs_diff = vabdq_u8(src, ref);
@@ -229,8 +115,6 @@ static INLINE void sad16xhx4d_neon(const uint8_t *src, int src_stride,
   vst1q_u32(res, horizontal_add_4d_uint16x8(sum));
 }
 
-#endif  // defined(__ARM_FEATURE_DOTPROD)
-
 static INLINE void sad8_neon(uint8x8_t src, uint8x8_t ref,
                              uint16x8_t *const sad_sum) {
   uint8x8_t abs_diff = vabd_u8(src, ref);

diff --git a/vpx_dsp/arm/sad4d_neon_dotprod.c b/vpx_dsp/arm/sad4d_neon_dotprod.c
new file mode 100644 (file)
index 0000000..933fc48
--- /dev/null
+++ b/vpx_dsp/arm/sad4d_neon_dotprod.c
@@ -0,0 +1,176 @@
+/*
+ *  Copyright (c) 2021 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/arm/sum_neon.h"
+
+static INLINE void sad16_neon(uint8x16_t src, uint8x16_t ref,
+                              uint32x4_t *const sad_sum) {
+  uint8x16_t abs_diff = vabdq_u8(src, ref);
+  *sad_sum = vdotq_u32(*sad_sum, abs_diff, vdupq_n_u8(1));
+}
+
+static INLINE void sad64xhx4d_neon_dotprod(const uint8_t *src, int src_stride,
+                                           const uint8_t *const ref[4],
+                                           int ref_stride, uint32_t res[4],
+                                           int h) {
+  uint32x4_t sum_lo[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
+                           vdupq_n_u32(0) };
+  uint32x4_t sum_hi[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
+                           vdupq_n_u32(0) };
+  uint32x4_t sum[4];
+
+  int i = 0;
+  do {
+    uint8x16_t s0, s1, s2, s3;
+
+    s0 = vld1q_u8(src + i * src_stride);
+    sad16_neon(s0, vld1q_u8(ref[0] + i * ref_stride), &sum_lo[0]);
+    sad16_neon(s0, vld1q_u8(ref[1] + i * ref_stride), &sum_lo[1]);
+    sad16_neon(s0, vld1q_u8(ref[2] + i * ref_stride), &sum_lo[2]);
+    sad16_neon(s0, vld1q_u8(ref[3] + i * ref_stride), &sum_lo[3]);
+
+    s1 = vld1q_u8(src + i * src_stride + 16);
+    sad16_neon(s1, vld1q_u8(ref[0] + i * ref_stride + 16), &sum_hi[0]);
+    sad16_neon(s1, vld1q_u8(ref[1] + i * ref_stride + 16), &sum_hi[1]);
+    sad16_neon(s1, vld1q_u8(ref[2] + i * ref_stride + 16), &sum_hi[2]);
+    sad16_neon(s1, vld1q_u8(ref[3] + i * ref_stride + 16), &sum_hi[3]);
+
+    s2 = vld1q_u8(src + i * src_stride + 32);
+    sad16_neon(s2, vld1q_u8(ref[0] + i * ref_stride + 32), &sum_lo[0]);
+    sad16_neon(s2, vld1q_u8(ref[1] + i * ref_stride + 32), &sum_lo[1]);
+    sad16_neon(s2, vld1q_u8(ref[2] + i * ref_stride + 32), &sum_lo[2]);
+    sad16_neon(s2, vld1q_u8(ref[3] + i * ref_stride + 32), &sum_lo[3]);
+
+    s3 = vld1q_u8(src + i * src_stride + 48);
+    sad16_neon(s3, vld1q_u8(ref[0] + i * ref_stride + 48), &sum_hi[0]);
+    sad16_neon(s3, vld1q_u8(ref[1] + i * ref_stride + 48), &sum_hi[1]);
+    sad16_neon(s3, vld1q_u8(ref[2] + i * ref_stride + 48), &sum_hi[2]);
+    sad16_neon(s3, vld1q_u8(ref[3] + i * ref_stride + 48), &sum_hi[3]);
+
+  } while (++i < h);
+
+  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
+  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
+  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
+  sum[3] = vaddq_u32(sum_lo[3], sum_hi[3]);
+
+  vst1q_u32(res, horizontal_add_4d_uint32x4(sum));
+}
+
+static INLINE void sad32xhx4d_neon_dotprod(const uint8_t *src, int src_stride,
+                                           const uint8_t *const ref[4],
+                                           int ref_stride, uint32_t res[4],
+                                           int h) {
+  uint32x4_t sum_lo[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
+                           vdupq_n_u32(0) };
+  uint32x4_t sum_hi[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
+                           vdupq_n_u32(0) };
+  uint32x4_t sum[4];
+
+  int i = 0;
+  do {
+    uint8x16_t s0, s1;
+
+    s0 = vld1q_u8(src + i * src_stride);
+    sad16_neon(s0, vld1q_u8(ref[0] + i * ref_stride), &sum_lo[0]);
+    sad16_neon(s0, vld1q_u8(ref[1] + i * ref_stride), &sum_lo[1]);
+    sad16_neon(s0, vld1q_u8(ref[2] + i * ref_stride), &sum_lo[2]);
+    sad16_neon(s0, vld1q_u8(ref[3] + i * ref_stride), &sum_lo[3]);
+
+    s1 = vld1q_u8(src + i * src_stride + 16);
+    sad16_neon(s1, vld1q_u8(ref[0] + i * ref_stride + 16), &sum_hi[0]);
+    sad16_neon(s1, vld1q_u8(ref[1] + i * ref_stride + 16), &sum_hi[1]);
+    sad16_neon(s1, vld1q_u8(ref[2] + i * ref_stride + 16), &sum_hi[2]);
+    sad16_neon(s1, vld1q_u8(ref[3] + i * ref_stride + 16), &sum_hi[3]);
+
+  } while (++i < h);
+
+  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
+  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
+  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
+  sum[3] = vaddq_u32(sum_lo[3], sum_hi[3]);
+
+  vst1q_u32(res, horizontal_add_4d_uint32x4(sum));
+}
+
+static INLINE void sad16xhx4d_neon_dotprod(const uint8_t *src, int src_stride,
+                                           const uint8_t *const ref[4],
+                                           int ref_stride, uint32_t res[4],
+                                           int h) {
+  uint32x4_t sum[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
+                        vdupq_n_u32(0) };
+
+  int i = 0;
+  do {
+    const uint8x16_t s = vld1q_u8(src + i * src_stride);
+    sad16_neon(s, vld1q_u8(ref[0] + i * ref_stride), &sum[0]);
+    sad16_neon(s, vld1q_u8(ref[1] + i * ref_stride), &sum[1]);
+    sad16_neon(s, vld1q_u8(ref[2] + i * ref_stride), &sum[2]);
+    sad16_neon(s, vld1q_u8(ref[3] + i * ref_stride), &sum[3]);
+
+  } while (++i < h);
+
+  vst1q_u32(res, horizontal_add_4d_uint32x4(sum));
+}
+
+#define SAD_WXH_4D_NEON_DOTPROD(w, h)                                      \
+  void vpx_sad##w##x##h##x4d_neon_dotprod(                                 \
+      const uint8_t *src_ptr, int src_stride,                              \
+      const uint8_t *const ref_array[4], int ref_stride,                   \
+      uint32_t sad_array[4]) {                                             \
+    sad##w##xhx4d_neon_dotprod(src_ptr, src_stride, ref_array, ref_stride, \
+                               sad_array, (h));                            \
+  }
+
+SAD_WXH_4D_NEON_DOTPROD(16, 8)
+SAD_WXH_4D_NEON_DOTPROD(16, 16)
+SAD_WXH_4D_NEON_DOTPROD(16, 32)
+
+SAD_WXH_4D_NEON_DOTPROD(32, 16)
+SAD_WXH_4D_NEON_DOTPROD(32, 32)
+SAD_WXH_4D_NEON_DOTPROD(32, 64)
+
+SAD_WXH_4D_NEON_DOTPROD(64, 32)
+SAD_WXH_4D_NEON_DOTPROD(64, 64)
+
+#undef SAD_WXH_4D_NEON_DOTPROD
+
+#define SAD_SKIP_WXH_4D_NEON_DOTPROD(w, h)                             \
+  void vpx_sad_skip_##w##x##h##x4d_neon_dotprod(                       \
+      const uint8_t *src_ptr, int src_stride,                          \
+      const uint8_t *const ref_array[4], int ref_stride,               \
+      uint32_t sad_array[4]) {                                         \
+    sad##w##xhx4d_neon_dotprod(src_ptr, 2 * src_stride, ref_array,     \
+                               2 * ref_stride, sad_array, ((h) >> 1)); \
+    sad_array[0] <<= 1;                                                \
+    sad_array[1] <<= 1;                                                \
+    sad_array[2] <<= 1;                                                \
+    sad_array[3] <<= 1;                                                \
+  }
+
+SAD_SKIP_WXH_4D_NEON_DOTPROD(16, 8)
+SAD_SKIP_WXH_4D_NEON_DOTPROD(16, 16)
+SAD_SKIP_WXH_4D_NEON_DOTPROD(16, 32)
+
+SAD_SKIP_WXH_4D_NEON_DOTPROD(32, 16)
+SAD_SKIP_WXH_4D_NEON_DOTPROD(32, 32)
+SAD_SKIP_WXH_4D_NEON_DOTPROD(32, 64)
+
+SAD_SKIP_WXH_4D_NEON_DOTPROD(64, 32)
+SAD_SKIP_WXH_4D_NEON_DOTPROD(64, 64)
+
+#undef SAD_SKIP_WXH_4D_NEON_DOTPROD

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index d789353..feb48ee 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -373,6 +373,7 @@ DSP_SRCS-$(HAVE_SSE2)   += x86/sum_squares_sse2.c
 DSP_SRCS-$(HAVE_MSA)    += mips/sum_squares_msa.c
 
 DSP_SRCS-$(HAVE_NEON)   += arm/sad4d_neon.c
+DSP_SRCS-$(HAVE_NEON_DOTPROD) += arm/sad4d_neon_dotprod.c
 DSP_SRCS-$(HAVE_NEON)   += arm/sad_neon.c
 DSP_SRCS-$(HAVE_NEON_DOTPROD) += arm/sad_neon_dotprod.c
 DSP_SRCS-$(HAVE_NEON)   += arm/subtract_neon.c

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 8383bdd..ff97e68 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -928,28 +928,28 @@ specialize qw/vpx_sad4x4_avg neon msa sse2 mmi/;
 # Multi-block SAD, comparing a reference to N independent blocks
 #
 add_proto qw/void vpx_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad64x64x4d avx512 avx2 neon msa sse2 vsx mmi lsx/;
+specialize qw/vpx_sad64x64x4d avx512 avx2 neon neon_dotprod msa sse2 vsx mmi lsx/;
 
 add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad64x32x4d neon msa sse2 vsx mmi lsx/;
+specialize qw/vpx_sad64x32x4d neon neon_dotprod msa sse2 vsx mmi lsx/;
 
 add_proto qw/void vpx_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad32x64x4d neon msa sse2 vsx mmi lsx/;
+specialize qw/vpx_sad32x64x4d neon neon_dotprod msa sse2 vsx mmi lsx/;
 
 add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad32x32x4d avx2 neon msa sse2 vsx mmi lsx/;
+specialize qw/vpx_sad32x32x4d avx2 neon neon_dotprod msa sse2 vsx mmi lsx/;
 
 add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad32x16x4d neon msa sse2 vsx mmi/;
+specialize qw/vpx_sad32x16x4d neon neon_dotprod msa sse2 vsx mmi/;
 
 add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad16x32x4d neon msa sse2 vsx mmi/;
+specialize qw/vpx_sad16x32x4d neon neon_dotprod msa sse2 vsx mmi/;
 
 add_proto qw/void vpx_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad16x16x4d neon msa sse2 vsx mmi lsx/;
+specialize qw/vpx_sad16x16x4d neon neon_dotprod msa sse2 vsx mmi lsx/;
 
 add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad16x8x4d neon msa sse2 vsx mmi/;
+specialize qw/vpx_sad16x8x4d neon neon_dotprod msa sse2 vsx mmi/;
 
 add_proto qw/void vpx_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
 specialize qw/vpx_sad8x16x4d neon msa sse2 mmi/;
@@ -967,28 +967,28 @@ add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const
 specialize qw/vpx_sad4x4x4d neon msa sse2 mmi/;
 
 add_proto qw/void vpx_sad_skip_64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_64x64x4d neon avx2 sse2/;
+specialize qw/vpx_sad_skip_64x64x4d neon neon_dotprod avx2 sse2/;
 
 add_proto qw/void vpx_sad_skip_64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_64x32x4d neon avx2 sse2/;
+specialize qw/vpx_sad_skip_64x32x4d neon neon_dotprod avx2 sse2/;
 
 add_proto qw/void vpx_sad_skip_32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_32x64x4d neon avx2 sse2/;
+specialize qw/vpx_sad_skip_32x64x4d neon neon_dotprod avx2 sse2/;
 
 add_proto qw/void vpx_sad_skip_32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_32x32x4d neon avx2 sse2/;
+specialize qw/vpx_sad_skip_32x32x4d neon neon_dotprod avx2 sse2/;
 
 add_proto qw/void vpx_sad_skip_32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_32x16x4d neon avx2 sse2/;
+specialize qw/vpx_sad_skip_32x16x4d neon neon_dotprod avx2 sse2/;
 
 add_proto qw/void vpx_sad_skip_16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_16x32x4d neon sse2/;
+specialize qw/vpx_sad_skip_16x32x4d neon neon_dotprod sse2/;
 
 add_proto qw/void vpx_sad_skip_16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_16x16x4d neon sse2/;
+specialize qw/vpx_sad_skip_16x16x4d neon neon_dotprod sse2/;
 
 add_proto qw/void vpx_sad_skip_16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
-specialize qw/vpx_sad_skip_16x8x4d neon sse2/;
+specialize qw/vpx_sad_skip_16x8x4d neon neon_dotprod sse2/;
 
 add_proto qw/void vpx_sad_skip_8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]";
 specialize qw/vpx_sad_skip_8x16x4d neon sse2/;