Merge "Add Neon implementation of vpx_hadamard_32x32" into main

3 files changed, 46 insertions, 3 deletions

diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index 10b1e79c1..f904e814a 100644
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -264,7 +264,8 @@ INSTANTIATE_TEST_SUITE_P(
 INSTANTIATE_TEST_SUITE_P(
     NEON, HadamardLowbdTest,
     ::testing::Values(HadamardFuncWithSize(&vpx_hadamard_8x8_neon, 8),
-                      HadamardFuncWithSize(&vpx_hadamard_16x16_neon, 16)));
+                      HadamardFuncWithSize(&vpx_hadamard_16x16_neon, 16),
+                      HadamardFuncWithSize(&vpx_hadamard_32x32_neon, 32)));
 #endif  // HAVE_NEON
 
 // TODO(jingning): Remove highbitdepth flag when the SIMD functions are
diff --git a/vpx_dsp/arm/hadamard_neon.c b/vpx_dsp/arm/hadamard_neon.c
index 523a63c6f..f6b6d7e3c 100644
--- a/vpx_dsp/arm/hadamard_neon.c
+++ b/vpx_dsp/arm/hadamard_neon.c
@@ -114,3 +114,45 @@ void vpx_hadamard_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride,
     coeff += 8;
   }
 }
+
+void vpx_hadamard_32x32_neon(const int16_t *src_diff, ptrdiff_t src_stride,
+                             tran_low_t *coeff) {
+  int i;
+
+  /* Rearrange 32x32 to 16x64 and remove stride.
+   * Top left first. */
+  vpx_hadamard_16x16_neon(src_diff + 0 + 0 * src_stride, src_stride, coeff + 0);
+  /* Top right. */
+  vpx_hadamard_16x16_neon(src_diff + 16 + 0 * src_stride, src_stride,
+                          coeff + 256);
+  /* Bottom left. */
+  vpx_hadamard_16x16_neon(src_diff + 0 + 16 * src_stride, src_stride,
+                          coeff + 512);
+  /* Bottom right. */
+  vpx_hadamard_16x16_neon(src_diff + 16 + 16 * src_stride, src_stride,
+                          coeff + 768);
+
+  for (i = 0; i < 256; i += 8) {
+    const int16x8_t a0 = load_tran_low_to_s16q(coeff + 0);
+    const int16x8_t a1 = load_tran_low_to_s16q(coeff + 256);
+    const int16x8_t a2 = load_tran_low_to_s16q(coeff + 512);
+    const int16x8_t a3 = load_tran_low_to_s16q(coeff + 768);
+
+    const int16x8_t b0 = vhaddq_s16(a0, a1);
+    const int16x8_t b1 = vhsubq_s16(a0, a1);
+    const int16x8_t b2 = vhaddq_s16(a2, a3);
+    const int16x8_t b3 = vhsubq_s16(a2, a3);
+
+    const int16x8_t c0 = vhaddq_s16(b0, b2);
+    const int16x8_t c1 = vhaddq_s16(b1, b3);
+    const int16x8_t c2 = vhsubq_s16(b0, b2);
+    const int16x8_t c3 = vhsubq_s16(b1, b3);
+
+    store_s16q_to_tran_low(coeff + 0, c0);
+    store_s16q_to_tran_low(coeff + 256, c1);
+    store_s16q_to_tran_low(coeff + 512, c2);
+    store_s16q_to_tran_low(coeff + 768, c3);
+
+    coeff += 8;
+  }
+}
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index d55ab67ce..51f5ebedd 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -799,7 +799,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     specialize qw/vpx_hadamard_16x16 avx2 sse2 neon vsx lsx/;
 
     add_proto qw/void vpx_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
-    specialize qw/vpx_hadamard_32x32 sse2 avx2/;
+    specialize qw/vpx_hadamard_32x32 sse2 avx2 neon/;
 
     add_proto qw/void vpx_highbd_hadamard_8x8/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
     specialize qw/vpx_highbd_hadamard_8x8 avx2/;
@@ -823,7 +823,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     specialize qw/vpx_hadamard_16x16 avx2 sse2 neon msa vsx lsx/;
 
     add_proto qw/void vpx_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff";
-    specialize qw/vpx_hadamard_32x32 sse2 avx2/;
+    specialize qw/vpx_hadamard_32x32 sse2 avx2 neon/;
 
     add_proto qw/int vpx_satd/, "const int16_t *coeff, int length";
     specialize qw/vpx_satd avx2 sse2 neon msa/;