From ddd446f274d33d8624da3fd842c23e9ede0e6b11 Mon Sep 17 00:00:00 2001 From: volzhs Date: Fri, 8 Jul 2016 19:29:58 +0900 Subject: Update webp driver to 0.5.1 (upstream) --- drivers/webp/dsp/enc_sse41.c | 102 ++++++++++++++----------------------------- 1 file changed, 33 insertions(+), 69 deletions(-) (limited to 'drivers/webp/dsp/enc_sse41.c') diff --git a/drivers/webp/dsp/enc_sse41.c b/drivers/webp/dsp/enc_sse41.c index 27f4189833..a1783901a6 100644 --- a/drivers/webp/dsp/enc_sse41.c +++ b/drivers/webp/dsp/enc_sse41.c @@ -17,6 +17,7 @@ #include #include // for abs() +#include "./common_sse2.h" #include "../enc/vp8enci.h" //------------------------------------------------------------------------------ @@ -67,55 +68,45 @@ static void CollectHistogram(const uint8_t* ref, const uint8_t* pred, // reconstructed samples. // Hadamard transform -// Returns the difference between the weighted sum of the absolute value of -// transformed coefficients. +// Returns the weighted sum of the absolute value of transformed coefficients. +// w[] contains a row-major 4 by 4 symmetric matrix. static int TTransform(const uint8_t* inA, const uint8_t* inB, const uint16_t* const w) { + int32_t sum[4]; __m128i tmp_0, tmp_1, tmp_2, tmp_3; - // Load, combine and transpose inputs. + // Load and combine inputs. { - const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]); - const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]); - const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]); + const __m128i inA_0 = _mm_loadu_si128((const __m128i*)&inA[BPS * 0]); + const __m128i inA_1 = _mm_loadu_si128((const __m128i*)&inA[BPS * 1]); + const __m128i inA_2 = _mm_loadu_si128((const __m128i*)&inA[BPS * 2]); + // In SSE4.1, with gcc 4.8 at least (maybe other versions), + // _mm_loadu_si128 is faster than _mm_loadl_epi64. But for the last lump + // of inA and inB, _mm_loadl_epi64 is still used not to have an out of + // bound read. const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]); - const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]); - const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]); - const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]); + const __m128i inB_0 = _mm_loadu_si128((const __m128i*)&inB[BPS * 0]); + const __m128i inB_1 = _mm_loadu_si128((const __m128i*)&inB[BPS * 1]); + const __m128i inB_2 = _mm_loadu_si128((const __m128i*)&inB[BPS * 2]); const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]); // Combine inA and inB (we'll do two transforms in parallel). - const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0); - const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1); - const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2); - const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3); - // a00 b00 a01 b01 a02 b03 a03 b03 0 0 0 0 0 0 0 0 - // a10 b10 a11 b11 a12 b12 a13 b13 0 0 0 0 0 0 0 0 - // a20 b20 a21 b21 a22 b22 a23 b23 0 0 0 0 0 0 0 0 - // a30 b30 a31 b31 a32 b32 a33 b33 0 0 0 0 0 0 0 0 - - // Transpose the two 4x4, discarding the filling zeroes. - const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2); - const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3); - // a00 a20 b00 b20 a01 a21 b01 b21 a02 a22 b02 b22 a03 a23 b03 b23 - // a10 a30 b10 b30 a11 a31 b11 b31 a12 a32 b12 b32 a13 a33 b13 b33 - const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1); - const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1); - // a00 a10 a20 a30 b00 b10 b20 b30 a01 a11 a21 a31 b01 b11 b21 b31 - // a02 a12 a22 a32 b02 b12 b22 b32 a03 a13 a23 a33 b03 b13 b23 b33 - - // Convert to 16b. - tmp_0 = _mm_cvtepu8_epi16(transpose1_0); - tmp_1 = _mm_cvtepu8_epi16(_mm_srli_si128(transpose1_0, 8)); - tmp_2 = _mm_cvtepu8_epi16(transpose1_1); - tmp_3 = _mm_cvtepu8_epi16(_mm_srli_si128(transpose1_1, 8)); - // a00 a10 a20 a30 b00 b10 b20 b30 - // a01 a11 a21 a31 b01 b11 b21 b31 - // a02 a12 a22 a32 b02 b12 b22 b32 - // a03 a13 a23 a33 b03 b13 b23 b33 + const __m128i inAB_0 = _mm_unpacklo_epi32(inA_0, inB_0); + const __m128i inAB_1 = _mm_unpacklo_epi32(inA_1, inB_1); + const __m128i inAB_2 = _mm_unpacklo_epi32(inA_2, inB_2); + const __m128i inAB_3 = _mm_unpacklo_epi32(inA_3, inB_3); + tmp_0 = _mm_cvtepu8_epi16(inAB_0); + tmp_1 = _mm_cvtepu8_epi16(inAB_1); + tmp_2 = _mm_cvtepu8_epi16(inAB_2); + tmp_3 = _mm_cvtepu8_epi16(inAB_3); + // a00 a01 a02 a03 b00 b01 b02 b03 + // a10 a11 a12 a13 b10 b11 b12 b13 + // a20 a21 a22 a23 b20 b21 b22 b23 + // a30 a31 a32 a33 b30 b31 b32 b33 } - // Horizontal pass and subsequent transpose. + // Vertical pass first to avoid a transpose (vertical and horizontal passes + // are commutative because w/kWeightY is symmetric) and subsequent transpose. { // Calculate a and b (two 4x4 at once). const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); @@ -132,33 +123,10 @@ static int TTransform(const uint8_t* inA, const uint8_t* inB, // a30 a31 a32 a33 b30 b31 b32 b33 // Transpose the two 4x4. - const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1); - const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3); - const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1); - const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3); - // a00 a10 a01 a11 a02 a12 a03 a13 - // a20 a30 a21 a31 a22 a32 a23 a33 - // b00 b10 b01 b11 b02 b12 b03 b13 - // b20 b30 b21 b31 b22 b32 b23 b33 - const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1); - const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3); - const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1); - const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3); - // a00 a10 a20 a30 a01 a11 a21 a31 - // b00 b10 b20 b30 b01 b11 b21 b31 - // a02 a12 a22 a32 a03 a13 a23 a33 - // b02 b12 a22 b32 b03 b13 b23 b33 - tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); - tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); - tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); - tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); - // a00 a10 a20 a30 b00 b10 b20 b30 - // a01 a11 a21 a31 b01 b11 b21 b31 - // a02 a12 a22 a32 b02 b12 b22 b32 - // a03 a13 a23 a33 b03 b13 b23 b33 + VP8Transpose_2_4x4_16b(&b0, &b1, &b2, &b3, &tmp_0, &tmp_1, &tmp_2, &tmp_3); } - // Vertical pass and difference of weighted sums. + // Horizontal pass and difference of weighted sums. { // Load all inputs. const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]); @@ -195,11 +163,9 @@ static int TTransform(const uint8_t* inA, const uint8_t* inB, // difference of weighted sums A_b2 = _mm_sub_epi32(A_b0, B_b0); - // cascading summation of the differences - B_b0 = _mm_hadd_epi32(A_b2, A_b2); - B_b2 = _mm_hadd_epi32(B_b0, B_b0); - return _mm_cvtsi128_si32(B_b2); + _mm_storeu_si128((__m128i*)&sum[0], A_b2); } + return sum[0] + sum[1] + sum[2] + sum[3]; } static int Disto4x4(const uint8_t* const a, const uint8_t* const b, @@ -240,8 +206,6 @@ static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16], __m128i packed_out; // Load all inputs. - // TODO(cduvivier): Make variable declarations and allocations aligned so that - // we can use _mm_load_si128 instead of _mm_loadu_si128. __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]); -- cgit v1.2.3