~4% improvement with a very rudimentary speed test
Unified · Split
| @@ -13,9 +13,10 @@ · 12 unmodified lines | |||
| 13 | 13 | ||
| 14 | #include "third_party/googletest/src/include/gtest/gtest.h" | 14 | #include "third_party/googletest/src/include/gtest/gtest.h" |
| 15 | 15 | ||
| 16 | #include "./vpx_config.h" | ||
| 17 | #include "./vp8_rtcd.h" | 16 | #include "./vp8_rtcd.h" |
| 17 | #include "./vpx_config.h" | ||
| 18 | #include "test/acm_random.h" | 18 | #include "test/acm_random.h" |
| 19 | #include "test/bench.h" | ||
| 19 | #include "test/clear_system_state.h" | 20 | #include "test/clear_system_state.h" |
| 20 | #include "test/register_state_check.h" | 21 | #include "test/register_state_check.h" |
| 21 | #include "test/util.h" | 22 | #include "test/util.h" |
| @@ -117,7 +118,8 @@ class QuantizeTestBase { · 95 unmodified lines | |||
| 117 | }; | 118 | }; |
| 118 | 119 | ||
| 119 | class QuantizeTest : public QuantizeTestBase, | 120 | class QuantizeTest : public QuantizeTestBase, |
| 120 | public ::testing::TestWithParam<VP8QuantizeParam> { | 121 | public ::testing::TestWithParam<VP8QuantizeParam>, |
| 122 | public AbstractBench { | ||
| 121 | protected: | 123 | protected: |
| 122 | virtual void SetUp() { | 124 | virtual void SetUp() { |
| 123 | SetupCompressor(); | 125 | SetupCompressor(); |
| @@ -125,6 +127,10 @@ class QuantizeTest : public QuantizeTestBase, · 1 unmodified line | |||
| 125 | c_quant_ = GET_PARAM(1); | 127 | c_quant_ = GET_PARAM(1); |
| 126 | } | 128 | } |
| 127 | 129 | ||
| 130 | virtual void Run() { | ||
| 131 | asm_quant_(&vp8_comp_->mb.block[0], ¯oblockd_dst_->block[0]); | ||
| 132 | } | ||
| 133 | |||
| 128 | void RunComparison() { | 134 | void RunComparison() { |
| 129 | for (int i = 0; i < kNumBlocks; ++i) { | 135 | for (int i = 0; i < kNumBlocks; ++i) { |
| 130 | ASM_REGISTER_STATE_CHECK( | 136 | ASM_REGISTER_STATE_CHECK( |
| @@ -167,6 +173,13 @@ TEST_P(QuantizeTest, TestMultipleQ) { · 36 unmodified lines | |||
| 167 | } | 173 | } |
| 168 | } | 174 | } |
| 169 | 175 | ||
| 176 | TEST_P(QuantizeTest, DISABLED_Speed) { | ||
| 177 | FillCoeffRandom(); | ||
| 178 | |||
| 179 | RunNTimes(10000000); | ||
| 180 | PrintMedian("vp8 quantize"); | ||
| 181 | } | ||
| 182 | |||
| 170 | #if HAVE_SSE2 | 183 | #if HAVE_SSE2 |
| 171 | INSTANTIATE_TEST_CASE_P( | 184 | INSTANTIATE_TEST_CASE_P( |
| 172 | SSE2, QuantizeTest, | 185 | SSE2, QuantizeTest, |
| @@ -11,8 +11,8 @@ · 10 unmodified lines | |||
| 11 | #include <smmintrin.h> /* SSE4.1 */ | 11 | #include <smmintrin.h> /* SSE4.1 */ |
| 12 | 12 | ||
| 13 | #include "./vp8_rtcd.h" | 13 | #include "./vp8_rtcd.h" |
| 14 | #include "vp8/encoder/block.h" | ||
| 15 | #include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */ | 14 | #include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */ |
| 15 | #include "vp8/encoder/block.h" | ||
| 16 | 16 | ||
| 17 | #define SELECT_EOB(i, z, x, y, q) \ | 17 | #define SELECT_EOB(i, z, x, y, q) \ |
| 18 | do { \ | 18 | do { \ |
| @@ -31,8 +31,7 @@ void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) { · 12 unmodified lines | |||
| 31 | char eob = 0; | 31 | char eob = 0; |
| 32 | short *zbin_boost_ptr = b->zrun_zbin_boost; | 32 | short *zbin_boost_ptr = b->zrun_zbin_boost; |
| 33 | 33 | ||
| 34 | __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1, dqcoeff0, | 34 | __m128i x0, x1, y0, y1, x_minus_zbin0, x_minus_zbin1, dqcoeff0, dqcoeff1; |
| 35 | dqcoeff1; | ||
| 36 | __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift)); | 35 | __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift)); |
| 37 | __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8)); | 36 | __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8)); |
| 38 | __m128i z0 = _mm_load_si128((__m128i *)(b->coeff)); | 37 | __m128i z0 = _mm_load_si128((__m128i *)(b->coeff)); |
| @@ -53,15 +52,9 @@ void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) { · 14 unmodified lines | |||
| 53 | zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0); | 52 | zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0); |
| 54 | zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra); | 53 | zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra); |
| 55 | 54 | ||
| 56 | /* Sign of z: z >> 15 */ | 55 | /* x = abs(z) */ |
| 57 | sz0 = _mm_srai_epi16(z0, 15); | 56 | x0 = _mm_abs_epi16(z0); |
| 58 | sz1 = _mm_srai_epi16(z1, 15); | 57 | x1 = _mm_abs_epi16(z1); |
| 59 | |||
| 60 | /* x = abs(z): (z ^ sz) - sz */ | ||
| 61 | x0 = _mm_xor_si128(z0, sz0); | ||
| 62 | x1 = _mm_xor_si128(z1, sz1); | ||
| 63 | x0 = _mm_sub_epi16(x0, sz0); | ||
| 64 | x1 = _mm_sub_epi16(x1, sz1); | ||
| 65 | 58 | ||
| 66 | /* zbin[] + zbin_extra */ | 59 | /* zbin[] + zbin_extra */ |
| 67 | zbin0 = _mm_add_epi16(zbin0, zbin_extra); | 60 | zbin0 = _mm_add_epi16(zbin0, zbin_extra); |
| @@ -89,11 +82,9 @@ void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) { · 21 unmodified lines | |||
| 89 | y0 = _mm_mulhi_epi16(y0, quant_shift0); | 82 | y0 = _mm_mulhi_epi16(y0, quant_shift0); |
| 90 | y1 = _mm_mulhi_epi16(y1, quant_shift1); | 83 | y1 = _mm_mulhi_epi16(y1, quant_shift1); |
| 91 | 84 | ||
| 92 | /* Return the sign: (y ^ sz) - sz */ | 85 | /* Restore the sign. */ |
| 93 | y0 = _mm_xor_si128(y0, sz0); | 86 | y0 = _mm_sign_epi16(y0, z0); |
| 94 | y1 = _mm_xor_si128(y1, sz1); | 87 | y1 = _mm_sign_epi16(y1, z1); |
| 95 | y0 = _mm_sub_epi16(y0, sz0); | ||
| 96 | y1 = _mm_sub_epi16(y1, sz1); | ||
| 97 | 88 | ||
| 98 | /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */ | 89 | /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */ |
| 99 | SELECT_EOB(1, 0, x_minus_zbin0, y0, qcoeff0); | 90 | SELECT_EOB(1, 0, x_minus_zbin0, y0, qcoeff0); |