From: Scott LaVarnway <slavarnway@google.com>
Date: Thu, 19 Oct 2017 00:12:57 +0000 (-0700)
Subject: vpx: [x86] add vpx_hadamard_16x16_avx2()
X-Git-Tag: v1.7.0~92^2
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=55c126a5d72900816aded997e6ebc9098283a5c2;p=libvpx

vpx: [x86] add vpx_hadamard_16x16_avx2()

This version is ~1.91x faster than the sse2 version.  When
highbitdepth is enabled, it is ~1.74x.

Change-Id: I2b0e92ede9f55c6259ca07bf1f8c8a5d0d0955bd
---

diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index a55b15ad0..eacd84635 100644
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -268,6 +268,11 @@ INSTANTIATE_TEST_CASE_P(SSE2, Hadamard16x16Test,
                         ::testing::Values(&vpx_hadamard_16x16_sse2));
 #endif  // HAVE_SSE2
 
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(AVX2, Hadamard16x16Test,
+                        ::testing::Values(&vpx_hadamard_16x16_avx2));
+#endif  // HAVE_AVX2
+
 #if HAVE_VSX
 INSTANTIATE_TEST_CASE_P(VSX, Hadamard16x16Test,
                         ::testing::Values(&vpx_hadamard_16x16_vsx));
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index 2d071b96a..ace174570 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -289,6 +289,7 @@ endif
 # avg
 DSP_SRCS-yes           += avg.c
 DSP_SRCS-$(HAVE_SSE2)  += x86/avg_intrin_sse2.c
+DSP_SRCS-$(HAVE_AVX2)  += x86/avg_intrin_avx2.c
 DSP_SRCS-$(HAVE_NEON)  += arm/avg_neon.c
 DSP_SRCS-$(HAVE_NEON)  += arm/hadamard_neon.c
 DSP_SRCS-$(HAVE_MSA)   += mips/avg_msa.c
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 474f50519..16b1f235a 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -769,7 +769,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     specialize qw/vpx_hadamard_8x8 sse2 neon vsx/, "$ssse3_x86_64";
 
     add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, int src_stride, tran_low_t *coeff";
-    specialize qw/vpx_hadamard_16x16 sse2 neon vsx/;
+    specialize qw/vpx_hadamard_16x16 avx2 sse2 neon vsx/;
 
     add_proto qw/int vpx_satd/, "const tran_low_t *coeff, int length";
     specialize qw/vpx_satd sse2 neon/;
@@ -778,7 +778,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     specialize qw/vpx_hadamard_8x8 sse2 neon msa vsx/, "$ssse3_x86_64";
 
     add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
-    specialize qw/vpx_hadamard_16x16 sse2 neon msa vsx/;
+    specialize qw/vpx_hadamard_16x16 avx2 sse2 neon msa vsx/;
 
     add_proto qw/int vpx_satd/, "const int16_t *coeff, int length";
     specialize qw/vpx_satd sse2 neon msa/;
diff --git a/vpx_dsp/x86/avg_intrin_avx2.c b/vpx_dsp/x86/avg_intrin_avx2.c
new file mode 100644
index 000000000..2ded29c59
--- /dev/null
+++ b/vpx_dsp/x86/avg_intrin_avx2.c
@@ -0,0 +1,182 @@
+/*
+ *  Copyright (c) 2017 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 <immintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/x86/bitdepth_conversion_avx2.h"
+#include "vpx_ports/mem.h"
+
+static void hadamard_col8x2_avx2(__m256i *in, int iter) {
+  __m256i a0 = in[0];
+  __m256i a1 = in[1];
+  __m256i a2 = in[2];
+  __m256i a3 = in[3];
+  __m256i a4 = in[4];
+  __m256i a5 = in[5];
+  __m256i a6 = in[6];
+  __m256i a7 = in[7];
+
+  __m256i b0 = _mm256_add_epi16(a0, a1);
+  __m256i b1 = _mm256_sub_epi16(a0, a1);
+  __m256i b2 = _mm256_add_epi16(a2, a3);
+  __m256i b3 = _mm256_sub_epi16(a2, a3);
+  __m256i b4 = _mm256_add_epi16(a4, a5);
+  __m256i b5 = _mm256_sub_epi16(a4, a5);
+  __m256i b6 = _mm256_add_epi16(a6, a7);
+  __m256i b7 = _mm256_sub_epi16(a6, a7);
+
+  a0 = _mm256_add_epi16(b0, b2);
+  a1 = _mm256_add_epi16(b1, b3);
+  a2 = _mm256_sub_epi16(b0, b2);
+  a3 = _mm256_sub_epi16(b1, b3);
+  a4 = _mm256_add_epi16(b4, b6);
+  a5 = _mm256_add_epi16(b5, b7);
+  a6 = _mm256_sub_epi16(b4, b6);
+  a7 = _mm256_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm256_add_epi16(a0, a4);
+    b7 = _mm256_add_epi16(a1, a5);
+    b3 = _mm256_add_epi16(a2, a6);
+    b4 = _mm256_add_epi16(a3, a7);
+    b2 = _mm256_sub_epi16(a0, a4);
+    b6 = _mm256_sub_epi16(a1, a5);
+    b1 = _mm256_sub_epi16(a2, a6);
+    b5 = _mm256_sub_epi16(a3, a7);
+
+    a0 = _mm256_unpacklo_epi16(b0, b1);
+    a1 = _mm256_unpacklo_epi16(b2, b3);
+    a2 = _mm256_unpackhi_epi16(b0, b1);
+    a3 = _mm256_unpackhi_epi16(b2, b3);
+    a4 = _mm256_unpacklo_epi16(b4, b5);
+    a5 = _mm256_unpacklo_epi16(b6, b7);
+    a6 = _mm256_unpackhi_epi16(b4, b5);
+    a7 = _mm256_unpackhi_epi16(b6, b7);
+
+    b0 = _mm256_unpacklo_epi32(a0, a1);
+    b1 = _mm256_unpacklo_epi32(a4, a5);
+    b2 = _mm256_unpackhi_epi32(a0, a1);
+    b3 = _mm256_unpackhi_epi32(a4, a5);
+    b4 = _mm256_unpacklo_epi32(a2, a3);
+    b5 = _mm256_unpacklo_epi32(a6, a7);
+    b6 = _mm256_unpackhi_epi32(a2, a3);
+    b7 = _mm256_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm256_unpacklo_epi64(b0, b1);
+    in[1] = _mm256_unpackhi_epi64(b0, b1);
+    in[2] = _mm256_unpacklo_epi64(b2, b3);
+    in[3] = _mm256_unpackhi_epi64(b2, b3);
+    in[4] = _mm256_unpacklo_epi64(b4, b5);
+    in[5] = _mm256_unpackhi_epi64(b4, b5);
+    in[6] = _mm256_unpacklo_epi64(b6, b7);
+    in[7] = _mm256_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm256_add_epi16(a0, a4);
+    in[7] = _mm256_add_epi16(a1, a5);
+    in[3] = _mm256_add_epi16(a2, a6);
+    in[4] = _mm256_add_epi16(a3, a7);
+    in[2] = _mm256_sub_epi16(a0, a4);
+    in[6] = _mm256_sub_epi16(a1, a5);
+    in[1] = _mm256_sub_epi16(a2, a6);
+    in[5] = _mm256_sub_epi16(a3, a7);
+  }
+}
+
+static void hadamard_8x8x2_avx2(int16_t const *src_diff, int src_stride,
+                                tran_low_t *coeff) {
+  __m256i src[8];
+  src[0] = _mm256_loadu_si256((const __m256i *)src_diff);
+  src[1] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[2] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[3] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[4] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[5] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[6] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[7] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+
+  hadamard_col8x2_avx2(src, 0);
+  hadamard_col8x2_avx2(src, 1);
+
+  store_tran_low(_mm256_castsi256_si128(src[0]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[1]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[2]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[3]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[4]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[5]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[6]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[7]), coeff);
+  coeff += 8;
+
+  src[0] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[0], 1));
+  src[1] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[1], 1));
+  src[2] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[2], 1));
+  src[3] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[3], 1));
+  src[4] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[4], 1));
+  src[5] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[5], 1));
+  src[6] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[6], 1));
+  src[7] = _mm256_castsi128_si256(_mm256_extractf128_si256(src[7], 1));
+
+  store_tran_low(_mm256_castsi256_si128(src[0]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[1]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[2]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[3]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[4]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[5]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[6]), coeff);
+  coeff += 8;
+  store_tran_low(_mm256_castsi256_si128(src[7]), coeff);
+}
+
+void vpx_hadamard_16x16_avx2(int16_t const *src_diff, int src_stride,
+                             tran_low_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 2; ++idx) {
+    int16_t const *src_ptr = src_diff + idx * 8 * src_stride;
+    hadamard_8x8x2_avx2(src_ptr, src_stride, coeff + (idx * 64 * 2));
+  }
+
+  for (idx = 0; idx < 64; idx += 16) {
+    const __m256i coeff0 = load_tran_low(coeff);
+    const __m256i coeff1 = load_tran_low(coeff + 64);
+    const __m256i coeff2 = load_tran_low(coeff + 128);
+    const __m256i coeff3 = load_tran_low(coeff + 192);
+    __m256i b0 = _mm256_add_epi16(coeff0, coeff1);
+    __m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
+    __m256i b2 = _mm256_add_epi16(coeff2, coeff3);
+    __m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm256_srai_epi16(b0, 1);
+    b1 = _mm256_srai_epi16(b1, 1);
+    b2 = _mm256_srai_epi16(b2, 1);
+    b3 = _mm256_srai_epi16(b3, 1);
+
+    store_tran_low_256(_mm256_add_epi16(b0, b2), coeff);
+    store_tran_low_256(_mm256_add_epi16(b1, b3), coeff + 64);
+    store_tran_low_256(_mm256_sub_epi16(b0, b2), coeff + 128);
+    store_tran_low_256(_mm256_sub_epi16(b1, b3), coeff + 192);
+
+    coeff += 16;
+  }
+}
diff --git a/vpx_dsp/x86/bitdepth_conversion_avx2.h b/vpx_dsp/x86/bitdepth_conversion_avx2.h
index b9116f049..7bbdac7ef 100644
--- a/vpx_dsp/x86/bitdepth_conversion_avx2.h
+++ b/vpx_dsp/x86/bitdepth_conversion_avx2.h
@@ -27,4 +27,33 @@ static INLINE __m256i load_tran_low(const tran_low_t *a) {
 #endif
 }
 
+// Store 8 16 bit values. If the destination is 32 bits then sign extend the
+// values by multiplying by 1.
+static INLINE void store_tran_low(__m128i a, tran_low_t *b) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i a_hi = _mm_mulhi_epi16(a, one);
+  const __m128i a_lo = _mm_mullo_epi16(a, one);
+  const __m128i a_1 = _mm_unpacklo_epi16(a_lo, a_hi);
+  const __m128i a_2 = _mm_unpackhi_epi16(a_lo, a_hi);
+  _mm_store_si128((__m128i *)(b), a_1);
+  _mm_store_si128((__m128i *)(b + 4), a_2);
+#else
+  _mm_store_si128((__m128i *)(b), a);
+#endif
+}
+
+static INLINE void store_tran_low_256(__m256i a, tran_low_t *b) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i a_hi = _mm256_mulhi_epi16(a, one);
+  const __m256i a_lo = _mm256_mullo_epi16(a, one);
+  const __m256i a_1 = _mm256_unpacklo_epi16(a_lo, a_hi);
+  const __m256i a_2 = _mm256_unpackhi_epi16(a_lo, a_hi);
+  _mm256_storeu_si256((__m256i *)(b), a_1);
+  _mm256_storeu_si256((__m256i *)(b + 8), a_2);
+#else
+  _mm256_storeu_si256((__m256i *)(b), a);
+#endif
+}
 #endif  // VPX_DSP_X86_BITDEPTH_CONVERSION_AVX2_H_