ref: c39a05ff61db40d39717df583b8aefd42cc3a2bf
parent: 2d0c11093e9854980a84d9415da2eb6a4cd42647
author: Scott LaVarnway <slavarnway@google.com>
date: Wed Aug 30 05:13:14 EDT 2017
vpxdsp: [x86] add highbd_dc_left_predictor functions C vs SSE2 speed gains: _4x4 : ~6.49x _8x8 : ~10.82x _16x16 : ~7.61x _32x32 : ~5.29x BUG=webm:1411 Change-Id: Ibc30c50cb7139049bf05298010803499e6ef949b
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -481,7 +481,8 @@
#if HAVE_SSE2
HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred4,
- vpx_highbd_dc_predictor_4x4_sse2, NULL,
+ vpx_highbd_dc_predictor_4x4_sse2,
+ vpx_highbd_dc_left_predictor_4x4_sse2,
vpx_highbd_dc_top_predictor_4x4_sse2, NULL,
vpx_highbd_v_predictor_4x4_sse2,
vpx_highbd_h_predictor_4x4_sse2, NULL, NULL, NULL, NULL,
@@ -488,7 +489,8 @@
NULL, NULL, vpx_highbd_tm_predictor_4x4_c)
HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred8,
- vpx_highbd_dc_predictor_8x8_sse2, NULL,
+ vpx_highbd_dc_predictor_8x8_sse2,
+ vpx_highbd_dc_left_predictor_8x8_sse2,
vpx_highbd_dc_top_predictor_8x8_sse2, NULL,
vpx_highbd_v_predictor_8x8_sse2,
vpx_highbd_h_predictor_8x8_sse2, NULL, NULL, NULL, NULL,
@@ -495,7 +497,8 @@
NULL, NULL, vpx_highbd_tm_predictor_8x8_sse2)
HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred16,
- vpx_highbd_dc_predictor_16x16_sse2, NULL,
+ vpx_highbd_dc_predictor_16x16_sse2,
+ vpx_highbd_dc_left_predictor_16x16_sse2,
vpx_highbd_dc_top_predictor_16x16_sse2, NULL,
vpx_highbd_v_predictor_16x16_sse2,
vpx_highbd_h_predictor_16x16_sse2, NULL, NULL, NULL,
@@ -502,7 +505,8 @@
NULL, NULL, NULL, vpx_highbd_tm_predictor_16x16_sse2)
HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred32,
- vpx_highbd_dc_predictor_32x32_sse2, NULL,
+ vpx_highbd_dc_predictor_32x32_sse2,
+ vpx_highbd_dc_left_predictor_32x32_sse2,
vpx_highbd_dc_top_predictor_32x32_sse2, NULL,
vpx_highbd_v_predictor_32x32_sse2,
vpx_highbd_h_predictor_32x32_sse2, NULL, NULL, NULL,
--- a/test/vp9_intrapred_test.cc
+++ b/test/vp9_intrapred_test.cc
@@ -471,6 +471,14 @@
INSTANTIATE_TEST_CASE_P(
SSE2_TO_C_8, VP9HighbdIntraPredTest,
::testing::Values(
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_4x4_sse2,
+ &vpx_highbd_dc_left_predictor_4x4_c, 4, 8),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_8x8_sse2,
+ &vpx_highbd_dc_left_predictor_8x8_c, 8, 8),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_16x16_sse2,
+ &vpx_highbd_dc_left_predictor_16x16_c, 16, 8),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_32x32_sse2,
+ &vpx_highbd_dc_left_predictor_32x32_c, 32, 8),
HighbdIntraPredParam(&vpx_highbd_dc_predictor_4x4_sse2,
&vpx_highbd_dc_predictor_4x4_c, 4, 8),
HighbdIntraPredParam(&vpx_highbd_dc_predictor_8x8_sse2,
@@ -515,6 +523,14 @@
INSTANTIATE_TEST_CASE_P(
SSE2_TO_C_10, VP9HighbdIntraPredTest,
::testing::Values(
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_4x4_sse2,
+ &vpx_highbd_dc_left_predictor_4x4_c, 4, 10),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_8x8_sse2,
+ &vpx_highbd_dc_left_predictor_8x8_c, 8, 10),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_16x16_sse2,
+ &vpx_highbd_dc_left_predictor_16x16_c, 16, 10),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_32x32_sse2,
+ &vpx_highbd_dc_left_predictor_32x32_c, 32, 10),
HighbdIntraPredParam(&vpx_highbd_dc_predictor_4x4_sse2,
&vpx_highbd_dc_predictor_4x4_c, 4, 10),
HighbdIntraPredParam(&vpx_highbd_dc_predictor_8x8_sse2,
@@ -559,6 +575,14 @@
INSTANTIATE_TEST_CASE_P(
SSE2_TO_C_12, VP9HighbdIntraPredTest,
::testing::Values(
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_4x4_sse2,
+ &vpx_highbd_dc_left_predictor_4x4_c, 4, 12),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_8x8_sse2,
+ &vpx_highbd_dc_left_predictor_8x8_c, 8, 12),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_16x16_sse2,
+ &vpx_highbd_dc_left_predictor_16x16_c, 16, 12),
+ HighbdIntraPredParam(&vpx_highbd_dc_left_predictor_32x32_sse2,
+ &vpx_highbd_dc_left_predictor_32x32_c, 32, 12),
HighbdIntraPredParam(&vpx_highbd_dc_predictor_4x4_sse2,
&vpx_highbd_dc_predictor_4x4_c, 4, 12),
HighbdIntraPredParam(&vpx_highbd_dc_predictor_8x8_sse2,
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -217,7 +217,7 @@
specialize qw/vpx_highbd_dc_top_predictor_4x4 neon sse2/;
add_proto qw/void vpx_highbd_dc_left_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
- specialize qw/vpx_highbd_dc_left_predictor_4x4 neon/;
+ specialize qw/vpx_highbd_dc_left_predictor_4x4 neon sse2/;
add_proto qw/void vpx_highbd_dc_128_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
specialize qw/vpx_highbd_dc_128_predictor_4x4 neon/;
@@ -252,7 +252,7 @@
specialize qw/vpx_highbd_dc_top_predictor_8x8 neon sse2/;
add_proto qw/void vpx_highbd_dc_left_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
- specialize qw/vpx_highbd_dc_left_predictor_8x8 neon/;
+ specialize qw/vpx_highbd_dc_left_predictor_8x8 neon sse2/;
add_proto qw/void vpx_highbd_dc_128_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
specialize qw/vpx_highbd_dc_128_predictor_8x8 neon/;
@@ -287,7 +287,7 @@
specialize qw/vpx_highbd_dc_top_predictor_16x16 neon sse2/;
add_proto qw/void vpx_highbd_dc_left_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
- specialize qw/vpx_highbd_dc_left_predictor_16x16 neon/;
+ specialize qw/vpx_highbd_dc_left_predictor_16x16 neon sse2/;
add_proto qw/void vpx_highbd_dc_128_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
specialize qw/vpx_highbd_dc_128_predictor_16x16 neon/;
@@ -322,7 +322,7 @@
specialize qw/vpx_highbd_dc_top_predictor_32x32 neon sse2/;
add_proto qw/void vpx_highbd_dc_left_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
- specialize qw/vpx_highbd_dc_left_predictor_32x32 neon/;
+ specialize qw/vpx_highbd_dc_left_predictor_32x32 neon sse2/;
add_proto qw/void vpx_highbd_dc_128_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
specialize qw/vpx_highbd_dc_128_predictor_32x32 neon/;
--- a/vpx_dsp/x86/highbd_intrapred_intrin_sse2.c
+++ b/vpx_dsp/x86/highbd_intrapred_intrin_sse2.c
@@ -177,6 +177,17 @@
}
}
+void vpx_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
+ const uint16_t *above,
+ const uint16_t *left, int bd) {+ const __m128i two = _mm_cvtsi32_si128(2);
+ const __m128i sum = dc_sum_4(left);
+ const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
+ (void)above;
+ (void)bd;
+ dc_store_4x4(dst, stride, &dc);
+}
+
void vpx_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
const uint16_t *above,
const uint16_t *left, int bd) {@@ -210,6 +221,17 @@
}
}
+void vpx_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
+ const uint16_t *above,
+ const uint16_t *left, int bd) {+ const __m128i four = _mm_cvtsi32_si128(4);
+ const __m128i sum = dc_sum_8(left);
+ const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
+ (void)above;
+ (void)bd;
+ dc_store_8x8(dst, stride, &dc);
+}
+
void vpx_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
const uint16_t *above,
const uint16_t *left, int bd) {@@ -241,6 +263,17 @@
}
}
+void vpx_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
+ const uint16_t *above,
+ const uint16_t *left, int bd) {+ const __m128i eight = _mm_cvtsi32_si128(8);
+ const __m128i sum = dc_sum_16(left);
+ const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+ (void)above;
+ (void)bd;
+ dc_store_16x16(dst, stride, &dc);
+}
+
void vpx_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
const uint16_t *above,
const uint16_t *left, int bd) {@@ -275,6 +308,17 @@
_mm_store_si128((__m128i *)(dst + 16), dc_dup);
_mm_store_si128((__m128i *)(dst + 24), dc_dup);
}
+}
+
+void vpx_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
+ const uint16_t *above,
+ const uint16_t *left, int bd) {+ const __m128i sixteen = _mm_cvtsi32_si128(16);
+ const __m128i sum = dc_sum_32(left);
+ const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
+ (void)above;
+ (void)bd;
+ dc_store_32x32(dst, stride, &dc);
}
void vpx_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,