SSE2 version of vectorized 8-tap filtering.
About 20% overall encoder speedup (vs. about 30% for sse4 version).
Change-Id: Ibf608a6a1bc94b14ec47e8046d3206b275b5a8bd
diff --git a/vp8/common/rtcd_defs.sh b/vp8/common/rtcd_defs.sh
index 1cb5de3..66029f8 100644
--- a/vp8/common/rtcd_defs.sh
+++ b/vp8/common/rtcd_defs.sh
@@ -14,8 +14,8 @@
# compiles warning free but a dissassembly of generated code show bugs. To be
# on the safe side, only enabled when compiled with 'gcc'.
if [ "$CONFIG_GCC" = "yes" ]; then
- specialize vp8_filter_block2d_4x4_8 sse4_1
- specialize vp8_filter_block2d_8x4_8 sse4_1
- specialize vp8_filter_block2d_8x8_8 sse4_1
- specialize vp8_filter_block2d_16x16_8 sse4_1
+ specialize vp8_filter_block2d_4x4_8 sse4_1 sse2
+ specialize vp8_filter_block2d_8x4_8 sse4_1 sse2
+ specialize vp8_filter_block2d_8x8_8 sse4_1 sse2
+ specialize vp8_filter_block2d_16x16_8 sse4_1 sse2
fi
diff --git a/vp8/common/x86/filter_sse2.c b/vp8/common/x86/filter_sse2.c
new file mode 100644
index 0000000..fe57b4e
--- /dev/null
+++ b/vp8/common/x86/filter_sse2.c
@@ -0,0 +1,289 @@
+/*
+ * Copyright (c) 2012 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 <assert.h> // for alignment checks
+#include <emmintrin.h> // SSE2
+#include "vp8/common/filter.h"
+#include "vpx_ports/mem.h" // for DECLARE_ALIGNED
+#include "vpx_rtcd.h"
+
+// TODO(cd): After cleanup, commit faster versions for non 4x4 size. This is
+// just a quick partial snapshot so that other can already use some
+// speedup.
+// TODO(cd): Use vectorized 8 tap filtering code as speedup to pure C 6 tap
+// filtering.
+// TODO(cd): Add some comments, better variable naming.
+// TODO(cd): Maybe use _mm_maddubs_epi16 if smaller filter coeficients (no sum
+// of positive above 128), or have higher precision filter
+// coefficients.
+
+DECLARE_ALIGNED(16, static const unsigned int, rounding_c[4]) = {
+ VP8_FILTER_WEIGHT >> 1,
+ VP8_FILTER_WEIGHT >> 1,
+ VP8_FILTER_WEIGHT >> 1,
+ VP8_FILTER_WEIGHT >> 1,
+};
+
+// Creating a macro to do more than four pixels at once to hide instruction
+// latency is actually slower :-(
+#define DO_FOUR_PIXELS(result, src_ptr, offset) \
+ { \
+ /* Do shifted load to achieve require shuffles through unpacking */ \
+ const __m128i src0 = _mm_loadu_si128((const __m128i *)(src_ptr + offset + 0)); \
+ const __m128i src1 = _mm_loadu_si128((const __m128i *)(src_ptr + offset + 1)); \
+ const __m128i src2 = _mm_loadu_si128((const __m128i *)(src_ptr + offset + 2)); \
+ const __m128i src3 = _mm_loadu_si128((const __m128i *)(src_ptr + offset + 3)); \
+ const __m128i src01 = _mm_unpacklo_epi8(src0, src1); \
+ const __m128i src01_16 = _mm_unpacklo_epi8(src01, zero); \
+ const __m128i src23 = _mm_unpacklo_epi8(src2, src3); \
+ const __m128i src23_16 = _mm_unpacklo_epi8(src23, zero); \
+ /* Shit by 4 bytes through suffle to get additional shifted loads */ \
+ const __m128i src4 = _mm_shuffle_epi32(src0, _MM_SHUFFLE(3, 3, 2, 1)); \
+ const __m128i src5 = _mm_shuffle_epi32(src1, _MM_SHUFFLE(3, 3, 2, 1)); \
+ const __m128i src6 = _mm_shuffle_epi32(src2, _MM_SHUFFLE(3, 3, 2, 1)); \
+ const __m128i src7 = _mm_shuffle_epi32(src3, _MM_SHUFFLE(3, 3, 2, 1)); \
+ const __m128i src45 = _mm_unpacklo_epi8(src4, src5); \
+ const __m128i src45_16 = _mm_unpacklo_epi8(src45, zero); \
+ const __m128i src67 = _mm_unpacklo_epi8(src6, src7); \
+ const __m128i src67_16 = _mm_unpacklo_epi8(src67, zero); \
+ /* multiply accumulate them */ \
+ const __m128i mad01 = _mm_madd_epi16(src01_16, fil01); \
+ const __m128i mad23 = _mm_madd_epi16(src23_16, fil23); \
+ const __m128i mad45 = _mm_madd_epi16(src45_16, fil45); \
+ const __m128i mad67 = _mm_madd_epi16(src67_16, fil67); \
+ const __m128i mad0123 = _mm_add_epi32(mad01, mad23); \
+ const __m128i mad4567 = _mm_add_epi32(mad45, mad67); \
+ __m128i mad_all = _mm_add_epi32(mad0123, mad4567); \
+ mad_all = _mm_add_epi32(mad_all, rounding); \
+ result = _mm_srai_epi32(mad_all, VP8_FILTER_SHIFT); \
+ }
+
+void vp8_filter_block2d_4x4_8_sse2
+(
+ const unsigned char *src_ptr, const unsigned int src_stride,
+ const short *HFilter_aligned16, const short *VFilter_aligned16,
+ unsigned char *dst_ptr, unsigned int dst_stride
+) {
+ __m128i intermediateA, intermediateB, intermediateC;
+
+ const int kInterp_Extend = 4;
+
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i rounding = _mm_load_si128((const __m128i *)rounding_c);
+
+ // check alignment
+ assert(0 == ((long)HFilter_aligned16)%16);
+ assert(0 == ((long)VFilter_aligned16)%16);
+
+ {
+ __m128i transpose3_0;
+ __m128i transpose3_1;
+ __m128i transpose3_2;
+ __m128i transpose3_3;
+
+ // Horizontal pass (src -> intermediate).
+ {
+ const __m128i HFilter = _mm_load_si128((const __m128i *)HFilter_aligned16);
+ // get first two columns filter coefficients
+ __m128i fil01 = _mm_shuffle_epi32(HFilter, _MM_SHUFFLE(0, 0, 0, 0));
+ __m128i fil23 = _mm_shuffle_epi32(HFilter, _MM_SHUFFLE(1, 1, 1, 1));
+ __m128i fil45 = _mm_shuffle_epi32(HFilter, _MM_SHUFFLE(2, 2, 2, 2));
+ __m128i fil67 = _mm_shuffle_epi32(HFilter, _MM_SHUFFLE(3, 3, 3, 3));
+ src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+
+ {
+ __m128i mad_all0;
+ __m128i mad_all1;
+ __m128i mad_all2;
+ __m128i mad_all3;
+ DO_FOUR_PIXELS(mad_all0, src_ptr, 0*src_stride)
+ DO_FOUR_PIXELS(mad_all1, src_ptr, 1*src_stride)
+ DO_FOUR_PIXELS(mad_all2, src_ptr, 2*src_stride)
+ DO_FOUR_PIXELS(mad_all3, src_ptr, 3*src_stride)
+ mad_all0 = _mm_packs_epi32(mad_all0, mad_all1);
+ mad_all2 = _mm_packs_epi32(mad_all2, mad_all3);
+ intermediateA = _mm_packus_epi16(mad_all0, mad_all2);
+ // --
+ src_ptr += src_stride*4;
+ // --
+ DO_FOUR_PIXELS(mad_all0, src_ptr, 0*src_stride)
+ DO_FOUR_PIXELS(mad_all1, src_ptr, 1*src_stride)
+ DO_FOUR_PIXELS(mad_all2, src_ptr, 2*src_stride)
+ DO_FOUR_PIXELS(mad_all3, src_ptr, 3*src_stride)
+ mad_all0 = _mm_packs_epi32(mad_all0, mad_all1);
+ mad_all2 = _mm_packs_epi32(mad_all2, mad_all3);
+ intermediateB = _mm_packus_epi16(mad_all0, mad_all2);
+ // --
+ src_ptr += src_stride*4;
+ // --
+ DO_FOUR_PIXELS(mad_all0, src_ptr, 0*src_stride)
+ DO_FOUR_PIXELS(mad_all1, src_ptr, 1*src_stride)
+ DO_FOUR_PIXELS(mad_all2, src_ptr, 2*src_stride)
+ mad_all0 = _mm_packs_epi32(mad_all0, mad_all1);
+ mad_all2 = _mm_packs_epi32(mad_all2, mad_all2);
+ intermediateC = _mm_packus_epi16(mad_all0, mad_all2);
+ }
+ }
+
+ // Transpose result (intermediate -> transpose3_x)
+ {
+ // 00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33
+ // 40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73
+ // 80 81 82 83 90 91 92 93 A0 A1 A2 A3 xx xx xx xx
+ const __m128i transpose0_0 = _mm_unpacklo_epi8(intermediateA, intermediateB);
+ const __m128i transpose0_1 = _mm_unpackhi_epi8(intermediateA, intermediateB);
+ const __m128i transpose0_2 = _mm_unpacklo_epi8(intermediateC, intermediateC);
+ const __m128i transpose0_3 = _mm_unpackhi_epi8(intermediateC, intermediateC);
+ // 00 40 01 41 02 42 03 43 10 50 11 51 12 52 13 53
+ // 20 60 21 61 22 62 23 63 30 70 31 71 32 72 33 73
+ // 80 xx 81 xx 82 xx 83 xx 90 xx 91 xx 92 xx 93 xx
+ // A0 xx A1 xx A2 xx A3 xx xx xx xx xx xx xx xx xx
+ const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1);
+ const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1);
+ const __m128i transpose1_2 = _mm_unpacklo_epi8(transpose0_2, transpose0_3);
+ const __m128i transpose1_3 = _mm_unpackhi_epi8(transpose0_2, transpose0_3);
+ // 00 20 40 60 01 21 41 61 02 22 42 62 03 23 43 63
+ // 10 30 50 70 11 31 51 71 12 32 52 72 13 33 53 73
+ // 80 A0 xx xx 81 A1 xx xx 82 A2 xx xx 83 A3 xx xx
+ // 90 xx xx xx 91 xx xx xx 92 xx xx xx 93 xx xx xx
+ const __m128i transpose2_0 = _mm_unpacklo_epi8(transpose1_0, transpose1_1);
+ const __m128i transpose2_1 = _mm_unpackhi_epi8(transpose1_0, transpose1_1);
+ const __m128i transpose2_2 = _mm_unpacklo_epi8(transpose1_2, transpose1_3);
+ const __m128i transpose2_3 = _mm_unpackhi_epi8(transpose1_2, transpose1_3);
+ // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+ // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+ // 80 90 A0 xx xx xx xx xx 81 91 A1 xx xx xx xx xx
+ // 82 92 A2 xx xx xx xx xx 83 93 A3 xx xx xx xx xx
+ transpose3_0 = _mm_castps_si128(
+ _mm_shuffle_ps(_mm_castsi128_ps(transpose2_0),
+ _mm_castsi128_ps(transpose2_2),
+ _MM_SHUFFLE(1, 0, 1, 0)));
+ transpose3_1 = _mm_castps_si128(
+ _mm_shuffle_ps(_mm_castsi128_ps(transpose2_0),
+ _mm_castsi128_ps(transpose2_2),
+ _MM_SHUFFLE(3, 2, 3, 2)));
+ transpose3_2 = _mm_castps_si128(
+ _mm_shuffle_ps(_mm_castsi128_ps(transpose2_1),
+ _mm_castsi128_ps(transpose2_3),
+ _MM_SHUFFLE(1, 0, 1, 0)));
+ transpose3_3 = _mm_castps_si128(
+ _mm_shuffle_ps(_mm_castsi128_ps(transpose2_1),
+ _mm_castsi128_ps(transpose2_3),
+ _MM_SHUFFLE(3, 2, 3, 2)));
+ // 00 10 20 30 40 50 60 70 80 90 A0 xx xx xx xx xx
+ // 01 11 21 31 41 51 61 71 81 91 A1 xx xx xx xx xx
+ // 02 12 22 32 42 52 62 72 82 92 A2 xx xx xx xx xx
+ // 03 13 23 33 43 53 63 73 83 93 A3 xx xx xx xx xx
+ }
+
+ // Vertical pass (transpose3_x -> dst).
+ {
+ const __m128i VFilter = _mm_load_si128((const __m128i *)VFilter_aligned16);
+ // get first two columns filter coefficients
+ __m128i fil01 = _mm_shuffle_epi32(VFilter, _MM_SHUFFLE(0, 0, 0, 0));
+ __m128i fil23 = _mm_shuffle_epi32(VFilter, _MM_SHUFFLE(1, 1, 1, 1));
+ __m128i fil45 = _mm_shuffle_epi32(VFilter, _MM_SHUFFLE(2, 2, 2, 2));
+ __m128i fil67 = _mm_shuffle_epi32(VFilter, _MM_SHUFFLE(3, 3, 3, 3));
+ __m128i col0, col1, col2, col3;
+ DECLARE_ALIGNED(16, unsigned char, temp[32]);
+ {
+ _mm_store_si128((__m128i *)temp, transpose3_0);
+ DO_FOUR_PIXELS(col0, temp, 0);
+ }
+ {
+ _mm_store_si128((__m128i *)temp, transpose3_1);
+ DO_FOUR_PIXELS(col1, temp, 0);
+ }
+ {
+ _mm_store_si128((__m128i *)temp, transpose3_2);
+ DO_FOUR_PIXELS(col2, temp, 0);
+ }
+ {
+ _mm_store_si128((__m128i *)temp, transpose3_3);
+ DO_FOUR_PIXELS(col3, temp, 0);
+ }
+ // transpose
+ {
+ __m128i T0 = _mm_unpacklo_epi32(col0, col1);
+ __m128i T1 = _mm_unpacklo_epi32(col2, col3);
+ __m128i T2 = _mm_unpackhi_epi32(col0, col1);
+ __m128i T3 = _mm_unpackhi_epi32(col2, col3);
+ col0 = _mm_unpacklo_epi64(T0, T1);
+ col1 = _mm_unpackhi_epi64(T0, T1);
+ col2 = _mm_unpacklo_epi64(T2, T3);
+ col3 = _mm_unpackhi_epi64(T2, T3);
+ }
+ // saturate to 8 bit
+ {
+ col0 = _mm_packs_epi32(col0, col0);
+ col0 = _mm_packus_epi16(col0, col0);
+ col1 = _mm_packs_epi32(col1, col1);
+ col1 = _mm_packus_epi16(col1, col1);
+ col2 = _mm_packs_epi32 (col2, col2);
+ col2 = _mm_packus_epi16(col2, col2);
+ col3 = _mm_packs_epi32 (col3, col3);
+ col3 = _mm_packus_epi16(col3, col3);
+ }
+ // store
+ {
+ *((unsigned int *)&dst_ptr[dst_stride * 0]) = _mm_cvtsi128_si32(col0);
+ *((unsigned int *)&dst_ptr[dst_stride * 1]) = _mm_cvtsi128_si32(col1);
+ *((unsigned int *)&dst_ptr[dst_stride * 2]) = _mm_cvtsi128_si32(col2);
+ *((unsigned int *)&dst_ptr[dst_stride * 3]) = _mm_cvtsi128_si32(col3);
+ }
+ }
+ }
+}
+
+void vp8_filter_block2d_8x4_8_sse2
+(
+ const unsigned char *src_ptr, const unsigned int src_stride,
+ const short *HFilter_aligned16, const short *VFilter_aligned16,
+ unsigned char *dst_ptr, unsigned int dst_stride
+) {
+ int j;
+ for (j=0; j<8; j+=4) {
+ vp8_filter_block2d_4x4_8_sse2(src_ptr + j, src_stride,
+ HFilter_aligned16, VFilter_aligned16,
+ dst_ptr + j, dst_stride);
+ }
+}
+
+void vp8_filter_block2d_8x8_8_sse2
+(
+ const unsigned char *src_ptr, const unsigned int src_stride,
+ const short *HFilter_aligned16, const short *VFilter_aligned16,
+ unsigned char *dst_ptr, unsigned int dst_stride
+) {
+ int i, j;
+ for (i=0; i<8; i+=4) {
+ for (j=0; j<8; j+=4) {
+ vp8_filter_block2d_4x4_8_sse2(src_ptr + j + i*src_stride, src_stride,
+ HFilter_aligned16, VFilter_aligned16,
+ dst_ptr + j + i*dst_stride, dst_stride);
+ }
+ }
+}
+
+void vp8_filter_block2d_16x16_8_sse2
+(
+ const unsigned char *src_ptr, const unsigned int src_stride,
+ const short *HFilter_aligned16, const short *VFilter_aligned16,
+ unsigned char *dst_ptr, unsigned int dst_stride
+) {
+ int i, j;
+ for (i=0; i<16; i+=4) {
+ for (j=0; j<16; j+=4) {
+ vp8_filter_block2d_4x4_8_sse2(src_ptr + j + i*src_stride, src_stride,
+ HFilter_aligned16, VFilter_aligned16,
+ dst_ptr + j + i*dst_stride, dst_stride);
+ }
+ }
+}
diff --git a/vp8/common/x86/filter_sse4.c b/vp8/common/x86/filter_sse4.c
index a037622..c461db1 100644
--- a/vp8/common/x86/filter_sse4.c
+++ b/vp8/common/x86/filter_sse4.c
@@ -25,9 +25,6 @@
// TODO(cd): Maybe use _mm_maddubs_epi16 if smaller filter coeficients (no sum
// of positive above 128), or have higher precision filter
// coefficients.
-// TODO(cd): Remove use of _mm_extract_epi32 and _mm_extract_epi64, to not
-// require SSE4.1
-// TODO(cd): Remove use of _mm_shuffle_epi8 to not require SSSE3
DECLARE_ALIGNED(16, static const unsigned char, mask0123_c[16]) = {
0x00, 0x01,
diff --git a/vp8/vp8_common.mk b/vp8/vp8_common.mk
index 3595a85..28f1bd9 100644
--- a/vp8/vp8_common.mk
+++ b/vp8/vp8_common.mk
@@ -119,6 +119,11 @@
vp8/common/x86/filter_sse4.c.o: CFLAGS += -msse4
endif
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/filter_sse2.c
+ifeq ($(HAVE_SSE2),yes)
+vp8/common/x86/filter_sse2.c.o: CFLAGS += -msse2
+endif
+
VP8_COMMON_SRCS-$(ARCH_ARM) += common/arm/arm_systemdependent.c
VP8_COMMON_SRCS-$(ARCH_ARM) += common/arm/bilinearfilter_arm.c
VP8_COMMON_SRCS-$(ARCH_ARM) += common/arm/bilinearfilter_arm.h
