/* ----------------------------------------------------------------------------- The copyright in this software is being made available under the Clear BSD License, included below. No patent rights, trademark rights and/or other Intellectual Property Rights other than the copyrights concerning the Software are granted under this license. The Clear BSD License Copyright (c) 2018-2023, Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. & The VVdeC Authors. All rights reserved. Redistribution and use in source and binary forms, with or without modification, are permitted (subject to the limitations in the disclaimer below) provided that the following conditions are met: * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. * Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. NO EXPRESS OR IMPLIED LICENSES TO ANY PARTY'S PATENT RIGHTS ARE GRANTED BY THIS LICENSE. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. ------------------------------------------------------------------------------------------- */ /** \file InterPredX86.h \brief SIMD for InterPrediction */ //! \ingroup CommonLib //! \{ #include "CommonLib/CommonDef.h" #include "CommonDefX86.h" #include "CommonLib/InterPrediction.h" namespace vvdec { #if ENABLE_SIMD_OPT_INTER #ifdef TARGET_SIMD_X86 template inline void PaddBIO_SIMD( const Pel* refPel, Pel* dstPel, unsigned width, const int shift ) { int w; __m128i off = _mm_set1_epi16( ( Pel ) IF_INTERNAL_OFFS ); if( width > 4 ) { for( w = 0; w < width; w += 8 ) { __m128i ref = _mm_lddqu_si128( ( __m128i const * )&refPel[w] ); ref = _mm_slli_epi16( ref, shift ); ref = _mm_sub_epi16( ref, off ); _mm_storeu_si128( ( __m128i * )&dstPel[w], ref ); } //2 * BIO_EXTEND_SIZE __m128i ref = _mm_lddqu_si128( ( __m128i const * )&refPel[w] ); ref = _mm_slli_epi16( ref, shift ); ref = _mm_sub_epi16( ref, off ); _mm_storeu_si32( ( __m128i * )&dstPel[w], ref ); } else { __m128i ref = _mm_lddqu_si128( ( __m128i const * )&refPel[0] ); ref = _mm_slli_epi16( ref, shift ); ref = _mm_sub_epi16( ref, off ); _mm_storel_epi64( ( __m128i * )&dstPel[0], ref ); ref = _mm_srli_si128( ref, 8 ); _mm_storeu_si32( ( __m128i * )&dstPel[4], ref ); } } static inline int rightShiftMSB( int numer, int denom ) { unsigned int shiftIdx = bit_scan_reverse( denom ); return ( numer >> shiftIdx ); } template inline void addBIOAvg4_2x_SSE(const int16_t* src0, const int16_t* src1, int16_t* dst, ptrdiff_t dstStride, const int16_t* gradX0, const int16_t* gradX1, const int16_t* gradY0, const int16_t* gradY1, ptrdiff_t widthG, int tmpx0, int tmpy0, int tmpx1, int tmpy1, int shift, int offset, const ClpRng& clpRng) { const ptrdiff_t src0Stride = widthG; const ptrdiff_t src1Stride = widthG; const ptrdiff_t gradStride = widthG; __m128i mm_tmpx0 = _mm_set1_epi32( ( tmpx0 & 0xffff ) | ( tmpy0 *(1<< 16 )) ); __m128i mm_tmpx1 = _mm_set1_epi32( ( tmpx1 & 0xffff ) | ( tmpy1 *(1<< 16 )) ); __m128i mm_offset = _mm_set1_epi32( offset ); __m128i vibdimin = _mm_set1_epi16( clpRng.min() ); __m128i vibdimax = _mm_set1_epi16( clpRng.max() ); __m128i mm_a; __m128i mm_b; __m128i mm_sum; __m128i mm_gx0, mm_gx1, mm_gy0, mm_gy1, mm_s0, mm_s1, mm_tmp; for( int y = 0; y < 4; y++, dst += dstStride, src0 += src0Stride, src1 += src1Stride, gradX0 += gradStride, gradX1 += gradStride, gradY0 += gradStride, gradY1 += gradStride ) { mm_gx0 = _mm_loadu_si128 ( ( const __m128i* ) gradX0 ); mm_gx1 = _mm_loadu_si128 ( ( const __m128i* ) gradX1 ); mm_gy0 = _mm_loadu_si128 ( ( const __m128i* ) gradY0 ); mm_gy1 = _mm_loadu_si128 ( ( const __m128i* ) gradY1 ); mm_s0 = _mm_loadu_si128 ( ( const __m128i* ) src0 ); mm_s1 = _mm_loadu_si128 ( ( const __m128i* ) src1 ); mm_a = _mm_unpacklo_epi16 ( mm_gx0, mm_gy0 ); mm_b = _mm_unpacklo_epi16 ( mm_gx1, mm_gy1 ); mm_a = _mm_sub_epi16 ( mm_a, mm_b ); mm_sum = _mm_madd_epi16 ( mm_a, mm_tmpx0 ); mm_a = _mm_cvtepi16_epi32 ( mm_s0 ); mm_b = _mm_cvtepi16_epi32 ( mm_s1 ); mm_tmp = _mm_add_epi32 ( _mm_add_epi32( mm_sum, mm_a ), _mm_add_epi32( mm_b, mm_offset ) ); mm_a = _mm_unpackhi_epi16 ( mm_gx0, mm_gy0 ); mm_b = _mm_unpackhi_epi16 ( mm_gx1, mm_gy1 ); mm_a = _mm_sub_epi16 ( mm_a, mm_b ); mm_sum = _mm_madd_epi16 ( mm_a, mm_tmpx1 ); mm_a = _mm_cvtepi16_epi32 ( _mm_unpackhi_epi64( mm_s0, mm_s0 ) ); mm_b = _mm_cvtepi16_epi32 ( _mm_unpackhi_epi64( mm_s1, mm_s0 ) ); mm_sum = _mm_add_epi32 ( _mm_add_epi32( mm_sum, mm_a ), _mm_add_epi32( mm_b, mm_offset ) ); mm_sum = _mm_packs_epi32 ( _mm_srai_epi32( mm_tmp, shift ), _mm_srai_epi32( mm_sum, shift ) ); mm_sum = _mm_min_epi16 ( vibdimax, _mm_max_epi16( vibdimin, mm_sum ) ); _mm_storeu_si128 ( (__m128i *) dst, mm_sum ); } } #if USE_AVX2 template<> inline void addBIOAvg4_2x_SSE(const int16_t* src0, const int16_t* src1, int16_t* dst, ptrdiff_t dstStride, const int16_t* gradX0, const int16_t* gradX1, const int16_t* gradY0, const int16_t* gradY1, ptrdiff_t widthG, int tmpx0, int tmpx1, int tmpy0, int tmpy1, int shift, int offset, const ClpRng& clpRng) { const ptrdiff_t src0Stride = widthG; const ptrdiff_t src1Stride = widthG; const ptrdiff_t gradStride = widthG; __m256i mm_tmpx = _mm256_inserti128_si256( _mm256_castsi128_si256( _mm_set1_epi32( ( tmpx0 & 0xffff ) | ( tmpy0 *(1<< 16 )) ) ), _mm_set1_epi32( ( tmpx1 & 0xffff ) | ( tmpy1 *(1<< 16 )) ), 1 ); __m256i mm_offset = _mm256_set1_epi32( offset ); __m256i vibdimin = _mm256_set1_epi32( clpRng.min() ); __m256i vibdimax = _mm256_set1_epi32( clpRng.max() ); __m256i mm_a; __m256i mm_b; __m256i mm_sum; __m128i xsrc0, xsrc1; for( int y = 0; y < 4; y++, dst += dstStride, src0 += src0Stride, src1 += src1Stride, gradX0 += gradStride, gradX1 += gradStride, gradY0 += gradStride, gradY1 += gradStride ) { xsrc0 = _mm_loadu_si128 ( ( const __m128i * ) gradX0 ); xsrc1 = _mm_loadu_si128 ( ( const __m128i * ) gradY0 ); mm_a = _mm256_castsi128_si256( _mm_unpacklo_epi16( xsrc0, xsrc1 ) ); mm_a = _mm256_inserti128_si256( mm_a, _mm_unpackhi_epi16( xsrc0, xsrc1 ), 1 ); xsrc0 = _mm_loadu_si128 ( ( const __m128i * ) gradX1 ); xsrc1 = _mm_loadu_si128 ( ( const __m128i * ) gradY1 ); mm_b = _mm256_castsi128_si256( _mm_unpacklo_epi16( xsrc0, xsrc1 ) ); mm_b = _mm256_inserti128_si256( mm_b, _mm_unpackhi_epi16( xsrc0, xsrc1 ), 1 ); mm_a = _mm256_sub_epi16 ( mm_a, mm_b ); mm_sum = _mm256_madd_epi16 ( mm_a, mm_tmpx ); mm_a = _mm256_cvtepi16_epi32 ( _mm_loadu_si128( (const __m128i *) ( src0 ) ) ); mm_b = _mm256_cvtepi16_epi32 ( _mm_loadu_si128( (const __m128i *) ( src1 ) ) ); mm_sum = _mm256_add_epi32 ( _mm256_add_epi32( mm_sum, mm_a ), _mm256_add_epi32( mm_b, mm_offset ) ); mm_sum = _mm256_srai_epi32 ( mm_sum, shift ); mm_sum = _mm256_min_epi32 ( vibdimax, _mm256_max_epi32( vibdimin, mm_sum ) ); _mm_storeu_si128 ( (__m128i *) dst, _mm256_cvtepi32_epi16x( mm_sum ) ); } } #endif template< X86_VEXT vext > static inline void calcBIOSums_SSE(const Pel* srcY0Tmp, const Pel* srcY1Tmp, const Pel* gradX0, const Pel* gradX1, const Pel* gradY0, const Pel* gradY1, const int widthG, const int bitDepth, int limit, int &tmpx, int &tmpy) { static constexpr int shift4 = 4; static constexpr int shift5 = 1; const int srcStride = widthG; __m128i sumAbsGXTmp = _mm_setzero_si128(); __m128i sumDIXTmp = _mm_setzero_si128(); __m128i sumAbsGYTmp = _mm_setzero_si128(); __m128i sumDIYTmp = _mm_setzero_si128(); __m128i sumSignGyGxTmp = _mm_setzero_si128(); for (int y = 0; y < 6; y++) { __m128i shiftSrcY0Tmp = _mm_srai_epi16(_mm_loadu_si128((__m128i*)(srcY0Tmp)), shift4); __m128i shiftSrcY1Tmp = _mm_srai_epi16(_mm_loadu_si128((__m128i*)(srcY1Tmp)), shift4); __m128i loadGradX0 = _mm_loadu_si128((__m128i*)(gradX0)); __m128i loadGradX1 = _mm_loadu_si128((__m128i*)(gradX1)); __m128i loadGradY0 = _mm_loadu_si128((__m128i*)(gradY0)); __m128i loadGradY1 = _mm_loadu_si128((__m128i*)(gradY1)); __m128i subTemp1 = _mm_sub_epi16(shiftSrcY1Tmp, shiftSrcY0Tmp); __m128i packTempX = _mm_srai_epi16(_mm_add_epi16(loadGradX0, loadGradX1), shift5); __m128i packTempY = _mm_srai_epi16(_mm_add_epi16(loadGradY0, loadGradY1), shift5); __m128i gX = _mm_abs_epi16(packTempX); __m128i gY = _mm_abs_epi16(packTempY); __m128i dIX = _mm_sign_epi16(subTemp1, packTempX ); __m128i dIY = _mm_sign_epi16(subTemp1, packTempY ); __m128i signGY_GX = _mm_sign_epi16(packTempX, packTempY ); sumAbsGXTmp = _mm_add_epi16(sumAbsGXTmp, gX); sumDIXTmp = _mm_add_epi16(sumDIXTmp, dIX); sumAbsGYTmp = _mm_add_epi16(sumAbsGYTmp, gY); sumDIYTmp = _mm_add_epi16(sumDIYTmp, dIY); sumSignGyGxTmp = _mm_add_epi16(sumSignGyGxTmp, signGY_GX); srcY0Tmp += srcStride; srcY1Tmp += srcStride; gradX0 += widthG; gradX1 += widthG; gradY0 += widthG; gradY1 += widthG; } sumAbsGXTmp = _mm_madd_epi16(sumAbsGXTmp, _mm_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0)); sumDIXTmp = _mm_madd_epi16(sumDIXTmp, _mm_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0)); sumAbsGYTmp = _mm_madd_epi16(sumAbsGYTmp, _mm_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0)); sumDIYTmp = _mm_madd_epi16(sumDIYTmp, _mm_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0)); sumSignGyGxTmp = _mm_madd_epi16(sumSignGyGxTmp, _mm_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0)); __m128i a12 = _mm_unpacklo_epi32(sumAbsGXTmp, sumAbsGYTmp); __m128i a3 = _mm_unpackhi_epi32(sumAbsGXTmp, sumAbsGYTmp); __m128i b12 = _mm_unpacklo_epi32(sumDIXTmp, sumDIYTmp); __m128i b3 = _mm_unpackhi_epi32(sumDIXTmp, sumDIYTmp); __m128i c1 = _mm_unpacklo_epi64(a12, b12); __m128i c2 = _mm_unpackhi_epi64(a12, b12); __m128i c3 = _mm_unpacklo_epi64(a3, b3); c1 = _mm_add_epi32(c1, c2); c1 = _mm_add_epi32(c1, c3); int sumAbsGX = _mm_cvtsi128_si32(c1); int sumAbsGY = _mm_cvtsi128_si32(_mm_shuffle_epi32(c1, 0x55)); int sumDIX = _mm_cvtsi128_si32(_mm_shuffle_epi32(c1, 0xaa)); int sumDIY = _mm_cvtsi128_si32(_mm_shuffle_epi32(c1, 0xff)); sumSignGyGxTmp = _mm_add_epi32(sumSignGyGxTmp, _mm_shuffle_epi32(sumSignGyGxTmp, 0x4e)); // 01001110 sumSignGyGxTmp = _mm_add_epi32(sumSignGyGxTmp, _mm_shuffle_epi32(sumSignGyGxTmp, 0xb1)); // 10110001 int sumSignGY_GX = _mm_cvtsi128_si32(sumSignGyGxTmp); tmpx = sumAbsGX == 0 ? 0 : rightShiftMSB( sumDIX *4, sumAbsGX ); tmpx = Clip3( -limit, limit, tmpx ); int mainsGxGy = sumSignGY_GX >> 12; int secsGxGy = sumSignGY_GX & ( ( 1 << 12 ) - 1 ); int tmpData = tmpx * mainsGxGy; tmpData = ( ( tmpData *(1<< 12 )) + tmpx * secsGxGy ) >> 1; tmpy = sumAbsGY == 0 ? 0 : rightShiftMSB( ( ( sumDIY *(1<< 2 )) - tmpData ), sumAbsGY ); tmpy = Clip3( -limit, limit, tmpy ); } #if USE_AVX2 static inline void calcBIOSums2x_AVX2(const Pel* srcY0Tmp, const Pel* srcY1Tmp, const Pel* gradX0, const Pel* gradX1, const Pel* gradY0, const Pel* gradY1, const int widthG, const int bitDepth, int limit, int &tmpx0, int &tmpx1, int &tmpy0, int &tmpy1 ) { static constexpr int shift4 = 4; static constexpr int shift5 = 1; const int srcStride = widthG; __m256i sumAbsGXTmp = _mm256_setzero_si256(); __m256i sumDIXTmp = _mm256_setzero_si256(); __m256i sumAbsGYTmp = _mm256_setzero_si256(); __m256i sumDIYTmp = _mm256_setzero_si256(); __m256i sumSignGyGxTmp = _mm256_setzero_si256(); #define _mm256_load2_si128_offset4(addr) _mm256_inserti128_si256( _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*) &addr[0])), _mm_loadu_si128((const __m128i*) &addr[4]), 1 ) for (int y = 0; y < 6; y++) { __m256i shiftSrcY0Tmp = _mm256_srai_epi16(_mm256_load2_si128_offset4(srcY0Tmp), shift4); __m256i shiftSrcY1Tmp = _mm256_srai_epi16(_mm256_load2_si128_offset4(srcY1Tmp), shift4); __m256i loadGradX0 = _mm256_load2_si128_offset4(gradX0); __m256i loadGradX1 = _mm256_load2_si128_offset4(gradX1); __m256i loadGradY0 = _mm256_load2_si128_offset4(gradY0); __m256i loadGradY1 = _mm256_load2_si128_offset4(gradY1); __m256i subTemp1 = _mm256_sub_epi16(shiftSrcY1Tmp, shiftSrcY0Tmp); __m256i packTempX = _mm256_srai_epi16(_mm256_add_epi16(loadGradX0, loadGradX1), shift5); __m256i packTempY = _mm256_srai_epi16(_mm256_add_epi16(loadGradY0, loadGradY1), shift5); __m256i gX = _mm256_abs_epi16(packTempX); __m256i gY = _mm256_abs_epi16(packTempY); __m256i dIX = _mm256_sign_epi16(subTemp1, packTempX ); __m256i dIY = _mm256_sign_epi16(subTemp1, packTempY ); __m256i signGY_GX = _mm256_sign_epi16(packTempX, packTempY ); sumAbsGXTmp = _mm256_add_epi16(sumAbsGXTmp, gX); sumDIXTmp = _mm256_add_epi16(sumDIXTmp, dIX); sumAbsGYTmp = _mm256_add_epi16(sumAbsGYTmp, gY); sumDIYTmp = _mm256_add_epi16(sumDIYTmp, dIY); sumSignGyGxTmp = _mm256_add_epi16(sumSignGyGxTmp, signGY_GX); srcY0Tmp += srcStride; srcY1Tmp += srcStride; gradX0 += widthG; gradX1 += widthG; gradY0 += widthG; gradY1 += widthG; } #undef _mm256_load2_si128_offset4 sumAbsGXTmp = _mm256_madd_epi16(sumAbsGXTmp, _mm256_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0)); sumDIXTmp = _mm256_madd_epi16(sumDIXTmp, _mm256_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0)); sumAbsGYTmp = _mm256_madd_epi16(sumAbsGYTmp, _mm256_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0)); sumDIYTmp = _mm256_madd_epi16(sumDIYTmp, _mm256_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0)); sumSignGyGxTmp = _mm256_madd_epi16(sumSignGyGxTmp, _mm256_setr_epi16(1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0)); __m256i a12 = _mm256_unpacklo_epi32(sumAbsGXTmp, sumAbsGYTmp); __m256i a3 = _mm256_unpackhi_epi32(sumAbsGXTmp, sumAbsGYTmp); __m256i b12 = _mm256_unpacklo_epi32(sumDIXTmp, sumDIYTmp); __m256i b3 = _mm256_unpackhi_epi32(sumDIXTmp, sumDIYTmp); __m256i c1 = _mm256_unpacklo_epi64(a12, b12); __m256i c2 = _mm256_unpackhi_epi64(a12, b12); __m256i c3 = _mm256_unpacklo_epi64(a3, b3); c1 = _mm256_add_epi32(c1, c2); c1 = _mm256_add_epi32(c1, c3); int tmpData[8]; _mm256_storeu_si256( ( __m256i* ) &tmpData[0], c1 ); #define sumAbsGX0 tmpData[0] #define sumAbsGX1 tmpData[4] #define sumAbsGY0 tmpData[1] #define sumAbsGY1 tmpData[5] #define sumDIX0 tmpData[2] #define sumDIX1 tmpData[6] #define sumDIY0 tmpData[3] #define sumDIY1 tmpData[7] sumSignGyGxTmp = _mm256_add_epi32(sumSignGyGxTmp, _mm256_shuffle_epi32(sumSignGyGxTmp, 0x4e)); // 01001110 sumSignGyGxTmp = _mm256_add_epi32(sumSignGyGxTmp, _mm256_shuffle_epi32(sumSignGyGxTmp, 0xb1)); // 10110001 int sumSignGY_GX0 = _mm256_extract_epi32( sumSignGyGxTmp, 0 ); int sumSignGY_GX1 = _mm256_extract_epi32( sumSignGyGxTmp, 4 ); #if 0 tmpx0 = sumAbsGX0 == 0 ? 0 : Clip3( -limit, limit, rightShiftMSB( sumDIX0 << 2, sumAbsGX0 ) ); tmpx1 = sumAbsGX1 == 0 ? 0 : Clip3( -limit, limit, rightShiftMSB( sumDIX1 << 2, sumAbsGX1 ) ); __m128i vtmpx = _mm_setr_epi32 ( tmpx0, tmpx1, 0, 0 ); __m128i vsumSignGY_GX = _mm_setr_epi32 ( sumSignGY_GX0, sumSignGY_GX1, 0, 0 ); __m128i vmainsGxGy = _mm_srai_epi32 ( vsumSignGY_GX, 12 ); __m128i vsecsGxGy = _mm_and_si128 ( vsumSignGY_GX, _mm_set1_epi32( ( 1 << 12 ) - 1 ) ); __m128i vtmpData = _mm_mullo_epi32( vtmpx, vmainsGxGy ); vtmpData = _mm_slli_epi32 ( vtmpData, 12 ); vtmpData = _mm_add_epi32 ( vtmpData, _mm_mullo_epi32( vtmpx, vsecsGxGy ) ); vtmpData = _mm_srai_epi32 ( vtmpData, 1 ); __m128i vtmpyIn = _mm_slli_epi32 ( _mm_setr_epi32( sumDIY0, sumDIY1, 0, 0 ), 2 ); vtmpyIn = _mm_sub_epi32 ( vtmpyIn, vtmpData ); tmpy0 = sumAbsGY0 == 0 ? 0 : Clip3( -limit, limit, rightShiftMSB( _mm_extract_epi32( vtmpyIn, 0 ), sumAbsGY0 ) ); tmpy1 = sumAbsGY1 == 0 ? 0 : Clip3( -limit, limit, rightShiftMSB( _mm_extract_epi32( vtmpyIn, 1 ), sumAbsGY1 ) ); #else tmpx0 = sumAbsGX0 == 0 ? 0 : rightShiftMSB( sumDIX0 *(1<< 2), sumAbsGX0 ); tmpx0 = Clip3( -limit, limit, tmpx0 ); int mainsGxGy0 = sumSignGY_GX0 >> 12; int secsGxGy0 = sumSignGY_GX0 & ( ( 1 << 12 ) - 1 ); int tmpData0 = tmpx0 * mainsGxGy0; tmpData0 = ( ( tmpData0 *(1<< 12 )) + tmpx0 * secsGxGy0 ) >> 1; tmpy0 = sumAbsGY0 == 0 ? 0 : rightShiftMSB( ( ( sumDIY0 *(1<< 2 )) - tmpData0 ), sumAbsGY0 ); tmpy0 = Clip3( -limit, limit, tmpy0 ); tmpx1 = sumAbsGX1 == 0 ? 0 : rightShiftMSB( sumDIX1 *(1<< 2), sumAbsGX1 ); tmpx1 = Clip3( -limit, limit, tmpx1 ); int mainsGxGy1 = sumSignGY_GX1 >> 12; int secsGxGy1 = sumSignGY_GX1 & ( ( 1 << 12 ) - 1 ); int tmpData1 = tmpx1 * mainsGxGy1; tmpData1 = ( ( tmpData1 *(1<< 12 )) + tmpx1 * secsGxGy1 ) >> 1; tmpy1 = sumAbsGY1 == 0 ? 0 : rightShiftMSB( ( ( sumDIY1 *(1<< 2) ) - tmpData1 ), sumAbsGY1 ); tmpy1 = Clip3( -limit, limit, tmpy1 ); #endif #undef sumAbsGX0 #undef sumAbsGX1 #undef sumAbsGY0 #undef sumAbsGY1 #undef sumDIX0 #undef sumDIX1 #undef sumDIY0 #undef sumDIY1 } #endif template< X86_VEXT vext> void BiOptFlowCoreSIMD( const Pel* srcY0, const Pel* srcY1, const Pel* gradX0, const Pel* gradX1, const Pel* gradY0, const Pel* gradY1, const int width, const int height, Pel* dstY, const ptrdiff_t dstStride, const int shiftNum, const int offset, const int limit, const ClpRng& clpRng, const int bitDepth ) { const int widthG = width + BIO_ALIGN_SIZE; const int stridePredMC = widthG; int offsetPos = widthG * BIO_EXTEND_SIZE + BIO_EXTEND_SIZE; const int xUnit = ( width >> 2 ); const int yUnit = ( height >> 2 ); const Pel* srcY0Temp; const Pel* srcY1Temp; Pel *dstY0; int OffPos; int OffPad = 0; int tmpx0, tmpy0, tmpx1, tmpy1; for( int yu = 0; yu < yUnit; yu++, srcY0 += ( stridePredMC << 2 ), srcY1 += ( stridePredMC << 2 ), dstY += ( dstStride << 2 ), offsetPos += ( widthG << 2 ) ) { srcY0Temp = srcY0; srcY1Temp = srcY1; dstY0 = dstY; OffPos = offsetPos; OffPad = ( ( yu * widthG ) << 2 ); for( int xu = 0; xu < xUnit; xu += 2, srcY0Temp += 8, srcY1Temp += 8, dstY0 += 8, OffPos += 8, OffPad += 8 ) { #if USE_AVX2 calcBIOSums2x_AVX2( srcY0Temp, srcY1Temp, gradX0 + OffPad, gradX1 + OffPad, gradY0 + OffPad, gradY1 + OffPad, stridePredMC, bitDepth, limit, tmpx0, tmpx1, tmpy0, tmpy1 ); addBIOAvg4_2x_SSE( srcY0Temp + stridePredMC + 1, srcY1Temp + stridePredMC + 1, dstY0, dstStride, gradX0 + OffPos, gradX1 + OffPos, gradY0 + OffPos, gradY1 + OffPos, widthG, tmpx0, tmpx1, tmpy0, tmpy1, shiftNum, offset, clpRng ); #else calcBIOSums_SSE( srcY0Temp + 0, srcY1Temp + 0, gradX0 + OffPad + 0, gradX1 + OffPad + 0, gradY0 + OffPad + 0, gradY1 + OffPad + 0, stridePredMC, bitDepth, limit, tmpx0, tmpy0 ); calcBIOSums_SSE( srcY0Temp + 4, srcY1Temp + 4, gradX0 + OffPad + 4, gradX1 + OffPad + 4, gradY0 + OffPad + 4, gradY1 + OffPad + 4, stridePredMC, bitDepth, limit, tmpx1, tmpy1 ); addBIOAvg4_2x_SSE( srcY0Temp + stridePredMC + 1, srcY1Temp + stridePredMC + 1, dstY0, dstStride, gradX0 + OffPos, gradX1 + OffPos, gradY0 + OffPos, gradY1 + OffPos, widthG, tmpx0, tmpy0, tmpx1, tmpy1, shiftNum, offset, clpRng ); #endif } // xu } // yu #if USE_AVX2 _mm256_zeroupper(); #endif } template< X86_VEXT vext, bool bi > void applyPROF_SSE(Pel* dstPel, ptrdiff_t dstStride, const Pel* srcPel, const Pel* gradX, const Pel* gradY, const int* dMvX, const int* dMvY, int shiftNum, Pel offset, const ClpRng& clpRng) { static constexpr ptrdiff_t srcStride = 6; static constexpr ptrdiff_t gradStride = 4; static constexpr ptrdiff_t dMvStride = 4; const int dILimit = 1 << std::max(clpRng.bd + 1, 13); #if USE_AVX2 __m256i mm_dmvx, mm_dmvy, mm_gradx, mm_grady, mm_dI, mm_dI0, mm_src; __m256i mm_offset = _mm256_set1_epi16( offset ); __m256i vibdimin = _mm256_set1_epi16( clpRng.min() ); __m256i vibdimax = _mm256_set1_epi16( clpRng.max() ); __m256i mm_dimin = _mm256_set1_epi32( -dILimit ); __m256i mm_dimax = _mm256_set1_epi32( dILimit - 1 ); const int *vX0 = dMvX, *vY0 = dMvY; const Pel *gX0 = gradX, *gY0 = gradY; // first two rows mm_dmvx = _mm256_loadu_si256( ( const __m256i * ) vX0 ); mm_dmvy = _mm256_loadu_si256( ( const __m256i * ) vY0 ); mm_dmvx = _mm256_packs_epi32( mm_dmvx, _mm256_setzero_si256() ); mm_dmvy = _mm256_packs_epi32( mm_dmvy, _mm256_setzero_si256() ); mm_gradx = _mm256_inserti128_si256( _mm256_castsi128_si256( _mm_loadl_epi64( ( __m128i* )gX0 ) ), _mm_loadl_epi64( ( __m128i* )( gX0 + gradStride ) ), 1 ); mm_grady = _mm256_inserti128_si256( _mm256_castsi128_si256( _mm_loadl_epi64( ( __m128i* )gY0 ) ), _mm_loadl_epi64( ( __m128i* )( gY0 + gradStride ) ), 1 ); mm_dI0 = _mm256_madd_epi16( _mm256_unpacklo_epi16( mm_dmvx, mm_dmvy ), _mm256_unpacklo_epi16( mm_gradx, mm_grady ) ); mm_dI0 = _mm256_min_epi32( mm_dimax, _mm256_max_epi32( mm_dimin, mm_dI0 ) ); // next two rows vX0 += ( dMvStride << 1 ); vY0 += ( dMvStride << 1 ); gX0 += ( gradStride << 1 ); gY0 += ( gradStride << 1 ); mm_dmvx = _mm256_loadu_si256( ( const __m256i * ) vX0 ); mm_dmvy = _mm256_loadu_si256( ( const __m256i * ) vY0 ); mm_dmvx = _mm256_packs_epi32( mm_dmvx, _mm256_setzero_si256() ); mm_dmvy = _mm256_packs_epi32( mm_dmvy, _mm256_setzero_si256() ); mm_gradx = _mm256_inserti128_si256( _mm256_castsi128_si256( _mm_loadl_epi64( ( __m128i* )gX0 ) ), _mm_loadl_epi64( ( __m128i* )( gX0 + gradStride ) ), 1 ); mm_grady = _mm256_inserti128_si256( _mm256_castsi128_si256( _mm_loadl_epi64( ( __m128i* )gY0 ) ), _mm_loadl_epi64( ( __m128i* )( gY0 + gradStride ) ), 1 ); mm_dI = _mm256_madd_epi16( _mm256_unpacklo_epi16( mm_dmvx, mm_dmvy ), _mm256_unpacklo_epi16( mm_gradx, mm_grady ) ); mm_dI = _mm256_min_epi32( mm_dimax, _mm256_max_epi32( mm_dimin, mm_dI ) ); // combine four rows mm_dI = _mm256_packs_epi32( mm_dI0, mm_dI ); const Pel* src0 = srcPel + srcStride; mm_src = _mm256_inserti128_si256( _mm256_castsi128_si256(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)srcPel), _mm_loadl_epi64((const __m128i *)(srcPel + (srcStride << 1))))), _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)src0), _mm_loadl_epi64((const __m128i *)(src0 + (srcStride << 1)))), 1 ); mm_dI = _mm256_add_epi16(mm_dI, mm_src); if (!bi) { mm_dI = _mm256_srai_epi16(_mm256_adds_epi16(mm_dI, mm_offset), shiftNum); mm_dI = _mm256_min_epi16(vibdimax, _mm256_max_epi16(vibdimin, mm_dI)); } // store final results __m128i dITmp = _mm256_extracti128_si256(mm_dI, 1); Pel* dst0 = dstPel; _mm_storel_epi64((__m128i *)dst0, _mm256_castsi256_si128(mm_dI)); dst0 += dstStride; _mm_storel_epi64((__m128i *)dst0, dITmp); dst0 += dstStride; _mm_storel_epi64((__m128i *)dst0, _mm_unpackhi_epi64(_mm256_castsi256_si128(mm_dI), _mm256_castsi256_si128(mm_dI))); dst0 += dstStride; _mm_storel_epi64((__m128i *)dst0, _mm_unpackhi_epi64(dITmp, dITmp)); #else __m128i mm_dmvx, mm_dmvy, mm_gradx, mm_grady, mm_dI, mm_dI0; __m128i mm_offset = _mm_set1_epi16( offset ); __m128i vibdimin = _mm_set1_epi16( clpRng.min() ); __m128i vibdimax = _mm_set1_epi16( clpRng.max() ); __m128i mm_dimin = _mm_set1_epi32( -dILimit ); __m128i mm_dimax = _mm_set1_epi32( dILimit - 1 ); static constexpr int height = 4; for( int h = 0; h < height; h += 2 ) { const int* vX = dMvX; const int* vY = dMvY; const Pel* gX = gradX; const Pel* gY = gradY; const Pel* src = srcPel; Pel* dst = dstPel; // first row mm_dmvx = _mm_packs_epi32( _mm_loadu_si128( ( const __m128i * ) vX ), _mm_setzero_si128() ); mm_dmvy = _mm_packs_epi32( _mm_loadu_si128( ( const __m128i * ) vY ), _mm_setzero_si128() ); mm_gradx = _mm_loadl_epi64( ( __m128i* ) gX ); mm_grady = _mm_loadl_epi64( ( __m128i* ) gY ); mm_dI0 = _mm_madd_epi16 ( _mm_unpacklo_epi16( mm_dmvx, mm_dmvy ), _mm_unpacklo_epi16( mm_gradx, mm_grady ) ); mm_dI0 = _mm_min_epi32 ( mm_dimax, _mm_max_epi32( mm_dimin, mm_dI0 ) ); // second row mm_dmvx = _mm_packs_epi32( _mm_loadu_si128( ( const __m128i * ) ( vX + dMvStride ) ), _mm_setzero_si128() ); mm_dmvy = _mm_packs_epi32( _mm_loadu_si128( ( const __m128i * ) ( vY + dMvStride ) ), _mm_setzero_si128() ); mm_gradx = _mm_loadl_epi64( ( __m128i* ) ( gX + gradStride ) ); mm_grady = _mm_loadl_epi64( ( __m128i* ) ( gY + gradStride ) ); mm_dI = _mm_madd_epi16 ( _mm_unpacklo_epi16( mm_dmvx, mm_dmvy ), _mm_unpacklo_epi16( mm_gradx, mm_grady ) ); mm_dI = _mm_min_epi32 ( mm_dimax, _mm_max_epi32( mm_dimin, mm_dI ) ); // combine both rows mm_dI = _mm_packs_epi32( mm_dI0, mm_dI ); mm_dI = _mm_add_epi16 ( _mm_unpacklo_epi64( _mm_loadl_epi64( ( const __m128i * )src ), _mm_loadl_epi64( ( const __m128i * )( src + srcStride ) ) ), mm_dI ); if (!bi) { mm_dI = _mm_srai_epi16(_mm_adds_epi16(mm_dI, mm_offset), shiftNum); mm_dI = _mm_min_epi16(vibdimax, _mm_max_epi16(vibdimin, mm_dI)); } _mm_storel_epi64( ( __m128i * ) dst, mm_dI ); _mm_storel_epi64( ( __m128i * )( dst + dstStride ), _mm_unpackhi_epi64( mm_dI, mm_dI ) ); dMvX += (dMvStride << 1); dMvY += (dMvStride << 1); gradX += (gradStride << 1); gradY += (gradStride << 1); srcPel += (srcStride << 1); dstPel += (dstStride << 1); } #endif } template< X86_VEXT vext > void roundIntVector_SIMD(int* v, int size, unsigned int nShift, const int dmvLimit) { CHECKD(size % 16 != 0, "Size must be multiple of 16!"); #ifdef USE_AVX512 if (vext >= AVX512 && size >= 16) { __m512i dMvMin = _mm256_set1_epi32(-dmvLimit); __m512i dMvMax = _mm256_set1_epi32(dmvLimit); __m512i nOffset = _mm512_set1_epi32((1 << (nShift - 1))); __m512i vones = _mm512_set1_epi32(1); __m512i vzero = _mm512_setzero_si512(); for (int i = 0; i < size; i += 16, v += 16) { __m512i src = _mm512_loadu_si512(v); __mmask16 mask = _mm512_cmpge_epi32_mask(src, vzero); src = __mm512_add_epi32(src, nOffset); __mm512i dst = _mm512_srai_epi32(_mm512_mask_sub_epi32(src, mask, src, vones), nShift); dst = _mm512_min_epi32(dMvMax, _mm512_max_epi32(dMvMin, dst)); _mm512_storeu_si512(v, dst); } } else #endif #ifdef USE_AVX2 if (vext >= AVX2 && size >= 8) { __m256i dMvMin = _mm256_set1_epi32(-dmvLimit); __m256i dMvMax = _mm256_set1_epi32(dmvLimit); __m256i nOffset = _mm256_set1_epi32(1 << (nShift - 1)); __m256i vzero = _mm256_setzero_si256(); for (int i = 0; i < size; i += 8, v += 8) { __m256i src = _mm256_lddqu_si256((__m256i*)v); __m256i of = _mm256_cmpgt_epi32(src, vzero); __m256i dst = _mm256_srai_epi32(_mm256_add_epi32(_mm256_add_epi32(src, nOffset), of), nShift); dst = _mm256_min_epi32(dMvMax, _mm256_max_epi32(dMvMin, dst)); _mm256_storeu_si256((__m256i*)v, dst); } } else #endif { __m128i dMvMin = _mm_set1_epi32(-dmvLimit); __m128i dMvMax = _mm_set1_epi32(dmvLimit); __m128i nOffset = _mm_set1_epi32((1 << (nShift - 1))); __m128i vzero = _mm_setzero_si128(); for (int i = 0; i < size; i += 4, v += 4) { __m128i src = _mm_loadu_si128((__m128i*)v); __m128i of = _mm_cmpgt_epi32(src, vzero); __m128i dst = _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(src, nOffset), of), nShift); dst = _mm_min_epi32(dMvMax, _mm_max_epi32(dMvMin, dst)); _mm_storeu_si128((__m128i*)v, dst); } } } template< X86_VEXT vext, bool PAD = true> void gradFilter_SSE( int16_t* src, ptrdiff_t _srcStride, int width, int height, ptrdiff_t _gradStride, int16_t* gradX, int16_t* gradY, const int bitDepth) { const int widthInside = PAD ? width - 2 * BIO_EXTEND_SIZE : 4; const int heightInside = PAD ? height - 2 * BIO_EXTEND_SIZE : 4; const ptrdiff_t gradStride = PAD ? _gradStride : 4; const ptrdiff_t srcStride = PAD ? _srcStride : 6; int16_t* srcTmp = PAD ? src + srcStride + 1 : src; int16_t* gradXTmp = PAD ? gradX + gradStride + 1 : gradX; int16_t* gradYTmp = PAD ? gradY + gradStride + 1 : gradY; const int shift1 = std::max( 6, bitDepth - 6 ); CHECKD( gradStride != _gradStride, "Wrong PROF stride!" ); CHECKD( srcStride != _srcStride, "Wrong PROF stride!" ); #if USE_AVX2 if( PAD && ( widthInside & 15 ) == 0 && vext >= AVX2 ) { for( int y = 0; y < heightInside; y++ ) { for( int x = 0; x < widthInside; x += 16 ) { __m256i mmPixTop = _mm256_srai_epi16( _mm256_loadu_si256( ( __m256i * ) ( srcTmp + x - srcStride ) ), shift1 ); __m256i mmPixBottom = _mm256_srai_epi16( _mm256_loadu_si256( ( __m256i * ) ( srcTmp + x + srcStride ) ), shift1 ); __m256i mmPixLeft = _mm256_srai_epi16( _mm256_loadu_si256( ( __m256i * ) ( srcTmp + x - 1 ) ), shift1 ); __m256i mmPixRight = _mm256_srai_epi16( _mm256_loadu_si256( ( __m256i * ) ( srcTmp + x + 1 ) ), shift1 ); __m256i mmGradVer = _mm256_sub_epi16( mmPixBottom, mmPixTop ); __m256i mmGradHor = _mm256_sub_epi16( mmPixRight, mmPixLeft ); _mm256_storeu_si256( ( __m256i * )( gradYTmp + x ), mmGradVer ); _mm256_storeu_si256( ( __m256i * )( gradXTmp + x ), mmGradHor ); } gradXTmp[widthInside] = gradXTmp[widthInside - 1]; gradYTmp[widthInside] = gradYTmp[widthInside - 1]; srcTmp [widthInside] = srcTmp [widthInside - 1]; gradXTmp[-1] = gradXTmp[0]; gradYTmp[-1] = gradYTmp[0]; srcTmp [-1] = srcTmp [0]; gradXTmp += gradStride; gradYTmp += gradStride; srcTmp += srcStride; } { gradXTmp = gradX + gradStride; gradYTmp = gradY + gradStride; srcTmp = src + srcStride; Pel * src0 = gradXTmp; Pel * src1 = gradXTmp + (height - 2 * BIO_EXTEND_SIZE - 1) * gradStride; Pel * src2 = gradYTmp; Pel * src3 = gradYTmp + (height - 2 * BIO_EXTEND_SIZE - 1) * gradStride; Pel * src4 = srcTmp; Pel * src5 = srcTmp + (height - 2 * BIO_EXTEND_SIZE - 1) * srcStride; Pel * dst0 = gradXTmp - gradStride; Pel * dst1 = gradXTmp + (height - 2 * BIO_EXTEND_SIZE) * gradStride; Pel * dst2 = gradYTmp - gradStride; Pel * dst3 = gradYTmp + (height - 2 * BIO_EXTEND_SIZE) * gradStride; Pel * dst4 = srcTmp - srcStride; Pel * dst5 = srcTmp + (height - 2 * BIO_EXTEND_SIZE) * srcStride; int x; for (x = 0; x < widthInside; x+=16) { __m256i s0 = _mm256_loadu_si256(( __m256i* )(src0 + x)); __m256i s1 = _mm256_loadu_si256(( __m256i* )(src1 + x)); __m256i s2 = _mm256_loadu_si256(( __m256i* )(src2 + x)); __m256i s3 = _mm256_loadu_si256(( __m256i* )(src3 + x)); __m256i s4 = _mm256_loadu_si256(( __m256i* )(src4 + x)); __m256i s5 = _mm256_loadu_si256(( __m256i* )(src5 + x)); _mm256_storeu_si256(( __m256i* )(dst0 + x), s0); _mm256_storeu_si256(( __m256i* )(dst1 + x), s1); _mm256_storeu_si256(( __m256i* )(dst2 + x), s2); _mm256_storeu_si256(( __m256i* )(dst3 + x), s3); _mm256_storeu_si256(( __m256i* )(dst4 + x), s4); _mm256_storeu_si256(( __m256i* )(dst5 + x), s5); } ((int32_t * )(dst0 + x))[0] = ((int32_t * )(src0 + x))[0]; ((int32_t * )(dst1 + x))[0] = ((int32_t * )(src1 + x))[0]; ((int32_t * )(dst2 + x))[0] = ((int32_t * )(src2 + x))[0]; ((int32_t * )(dst3 + x))[0] = ((int32_t * )(src3 + x))[0]; ((int32_t * )(dst4 + x))[0] = ((int32_t * )(src4 + x))[0]; ((int32_t * )(dst5 + x))[0] = ((int32_t * )(src5 + x))[0]; } } else #endif if( PAD && ( widthInside & 7 ) == 0 ) { for( int y = 0; y < heightInside; y++ ) { for( int x = 0; x < widthInside; x += 8 ) { __m128i mmPixTop = _mm_srai_epi16( _mm_loadu_si128( ( __m128i* )( srcTmp + x - srcStride ) ), shift1 ); __m128i mmPixBottom = _mm_srai_epi16( _mm_loadu_si128( ( __m128i* )( srcTmp + x + srcStride ) ), shift1 ); __m128i mmPixLeft = _mm_srai_epi16( _mm_loadu_si128( ( __m128i* )( srcTmp + x - 1 ) ), shift1 ); __m128i mmPixRight = _mm_srai_epi16( _mm_loadu_si128( ( __m128i* )( srcTmp + x + 1 ) ), shift1 ); __m128i mmGradVer = _mm_sub_epi16( mmPixBottom, mmPixTop ); __m128i mmGradHor = _mm_sub_epi16( mmPixRight, mmPixLeft ); _mm_storeu_si128((__m128i *) (gradYTmp + x), mmGradVer); _mm_storeu_si128((__m128i *) (gradXTmp + x), mmGradHor); } gradXTmp[widthInside] = gradXTmp[widthInside - 1]; gradYTmp[widthInside] = gradYTmp[widthInside - 1]; srcTmp [widthInside] = srcTmp [widthInside - 1]; gradXTmp[-1] = gradXTmp[0]; gradYTmp[-1] = gradYTmp[0]; srcTmp [-1] = srcTmp [0]; gradXTmp += gradStride; gradYTmp += gradStride; srcTmp += srcStride; } { gradXTmp = gradX + gradStride; gradYTmp = gradY + gradStride; srcTmp = src + srcStride; Pel * src0 = gradXTmp; Pel * src1 = gradXTmp + (height - 2 * BIO_EXTEND_SIZE - 1) * gradStride; Pel * src2 = gradYTmp; Pel * src3 = gradYTmp + (height - 2 * BIO_EXTEND_SIZE - 1) * gradStride; Pel * src4 = srcTmp; Pel * src5 = srcTmp + (height - 2 * BIO_EXTEND_SIZE - 1) * srcStride; Pel * dst0 = gradXTmp - gradStride; Pel * dst1 = gradXTmp + (height - 2 * BIO_EXTEND_SIZE) * gradStride; Pel * dst2 = gradYTmp - gradStride; Pel * dst3 = gradYTmp + (height - 2 * BIO_EXTEND_SIZE) * gradStride; Pel * dst4 = srcTmp - srcStride; Pel * dst5 = srcTmp + (height - 2 * BIO_EXTEND_SIZE) * srcStride; int x; for (x = 0; x < widthInside; x+=8) { __m128i s0 = _mm_loadu_si128(( __m128i* )(src0 + x)); __m128i s1 = _mm_loadu_si128(( __m128i* )(src1 + x)); __m128i s2 = _mm_loadu_si128(( __m128i* )(src2 + x)); __m128i s3 = _mm_loadu_si128(( __m128i* )(src3 + x)); __m128i s4 = _mm_loadu_si128(( __m128i* )(src4 + x)); __m128i s5 = _mm_loadu_si128(( __m128i* )(src5 + x)); _mm_storeu_si128(( __m128i* )(dst0 + x), s0); _mm_storeu_si128(( __m128i* )(dst1 + x), s1); _mm_storeu_si128(( __m128i* )(dst2 + x), s2); _mm_storeu_si128(( __m128i* )(dst3 + x), s3); _mm_storeu_si128(( __m128i* )(dst4 + x), s4); _mm_storeu_si128(( __m128i* )(dst5 + x), s5); } ((int32_t * )(dst0 + x))[0] = ((int32_t * )(src0 + x))[0]; ((int32_t * )(dst1 + x))[0] = ((int32_t * )(src1 + x))[0]; ((int32_t * )(dst2 + x))[0] = ((int32_t * )(src2 + x))[0]; ((int32_t * )(dst3 + x))[0] = ((int32_t * )(src3 + x))[0]; ((int32_t * )(dst4 + x))[0] = ((int32_t * )(src4 + x))[0]; ((int32_t * )(dst5 + x))[0] = ((int32_t * )(src5 + x))[0]; } } else { CHECK_RECOVERABLE( widthInside != 4, "Width needs to be '4'!" ); for( int y = 0; y < ( PAD ? heightInside : 4 ); y++ ) { __m128i mmPixTop = _mm_srai_epi16( _mm_loadl_epi64( ( __m128i* )( srcTmp - srcStride ) ), shift1 ); __m128i mmPixBottom = _mm_srai_epi16( _mm_loadl_epi64( ( __m128i* )( srcTmp + srcStride ) ), shift1 ); __m128i mmPixLeft = _mm_srai_epi16( _mm_loadl_epi64( ( __m128i* )( srcTmp - 1 ) ), shift1 ); __m128i mmPixRight = _mm_srai_epi16( _mm_loadl_epi64( ( __m128i* )( srcTmp + 1 ) ), shift1 ); __m128i mmGradVer = _mm_sub_epi16( mmPixBottom, mmPixTop ); __m128i mmGradHor = _mm_sub_epi16( mmPixRight, mmPixLeft ); _mm_storel_epi64( ( __m128i * )( gradYTmp ), mmGradVer ); _mm_storel_epi64( ( __m128i * )( gradXTmp ), mmGradHor ); if( PAD ) { gradXTmp[widthInside] = gradXTmp[widthInside - 1]; gradYTmp[widthInside] = gradYTmp[widthInside - 1]; srcTmp [widthInside] = srcTmp [widthInside - 1]; gradXTmp[-1] = gradXTmp[0]; gradYTmp[-1] = gradYTmp[0]; srcTmp [-1] = srcTmp [0]; } gradXTmp += gradStride; gradYTmp += gradStride; srcTmp += srcStride; if( PAD ) { gradXTmp = gradX + gradStride; gradYTmp = gradY + gradStride; srcTmp = src + srcStride; ::memcpy( gradXTmp + heightInside * gradStride, gradXTmp + ( heightInside - 1 ) * gradStride, sizeof( int16_t ) * ( width ) ); ::memcpy( gradYTmp + heightInside * gradStride, gradYTmp + ( heightInside - 1 ) * gradStride, sizeof( int16_t ) * ( width ) ); ::memcpy( srcTmp + heightInside * srcStride , srcTmp + ( heightInside - 1 ) * srcStride , sizeof( int16_t ) * ( width ) ); ::memcpy( gradXTmp - gradStride, gradXTmp, sizeof( int16_t ) * ( width ) ); ::memcpy( gradYTmp - gradStride, gradYTmp, sizeof( int16_t ) * ( width ) ); ::memcpy( srcTmp - srcStride , srcTmp , sizeof( int16_t ) * ( width ) ); } } } #if USE_AVX2 _mm256_zeroupper(); #endif } template void prefetchPadC_SSE( const Pel* src, const ptrdiff_t srcStride, Pel* dst, const ptrdiff_t dstStride, int width, int height ) { _mm_prefetch( ( const char* ) src, _MM_HINT_T0 ); _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); if( width == 7 ) { const __m128i sl = _mm_setr_epi8( 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 12, 13 ); __m128i l = _mm_shuffle_epi8( _mm_loadu_si128( ( const __m128i* ) src ), sl ); Pel pel0 = *src; dst[-1-dstStride] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride ), l ); dst[-1] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride ), l ); for( height--, dst += dstStride, src += srcStride; height > 0; height--, src += srcStride, dst += dstStride ) { _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); l = _mm_shuffle_epi8( _mm_loadu_si128( ( const __m128i* ) src ), sl ); pel0 = *src; dst[-1] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst ), l ); } dst[-1] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst ), l ); } else { const __m128i sl = _mm_setr_epi8( 0, 1, 2, 3, 4, 5, 4, 5, -1, -1, -1, -1, -1, -1, -1, -1 ); __m128i l0 = _mm_loadu_si128( ( const __m128i* ) &src[0] ); __m128i l1 = _mm_shuffle_epi8( _mm_loadl_epi64( ( const __m128i* ) &src[8] ), sl ); Pel pel0 = *src; dst[-1 - dstStride] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride - 0 ), l0 ); _mm_storel_epi64( ( __m128i* ) ( dst - 1 * dstStride + 8 ), l1 ); dst[-1] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride - 0 ), l0 ); _mm_storel_epi64( ( __m128i* ) ( dst - 0 * dstStride + 8 ), l1 ); for( height--, dst += dstStride, src += srcStride; height > 0; height--, src += srcStride, dst += dstStride ) { _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); l0 = _mm_loadu_si128( ( const __m128i* ) &src[0] ); l1 = _mm_shuffle_epi8( _mm_loadl_epi64( ( const __m128i* ) &src[8] ), sl ); pel0 = *src; dst[-1] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst - 0 ), l0 ); _mm_storel_epi64( ( __m128i* ) ( dst + 8 ), l1 ); } dst[-1] = pel0; _mm_storeu_si128( ( __m128i* ) ( dst - 0 ), l0 ); _mm_storel_epi64( ( __m128i* ) ( dst + 8 ), l1 ); } } template void prefetchPadL_SSE( const Pel* src, const ptrdiff_t srcStride, Pel* dst, const ptrdiff_t dstStride, int width, int height ) { _mm_prefetch( ( const char* ) src, _MM_HINT_T0 ); _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); if( width == 15 ) { const __m128i sb = _mm_setr_epi8( 0, 1, 0, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 ); __m128i l0 = _mm_shuffle_epi8( _mm_loadu_si128( ( const __m128i* ) &src[ 0] ), sb ); __m128i l1 = _mm_loadu_si128( ( const __m128i* ) &src[ 6] ); __m128i l3 = _mm_set1_epi16 ( src[14] ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 * dstStride - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 * dstStride + 6 ), l1 ); _mm_storel_epi64( ( __m128i* ) ( dst - 2 * dstStride + 14 ), l3 ); _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride + 6 ), l1 ); _mm_storel_epi64( ( __m128i* ) ( dst - 1 * dstStride + 14 ), l3 ); _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride + 6 ), l1 ); _mm_storel_epi64( ( __m128i* ) ( dst - 0 * dstStride + 14 ), l3 ); for( height--, dst += dstStride, src += srcStride; height > 0; height--, src += srcStride, dst += dstStride ) { _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); l0 = _mm_shuffle_epi8( _mm_loadu_si128( ( const __m128i* ) &src[ 0] ), sb ); l1 = _mm_loadu_si128( ( const __m128i* ) &src[ 6] ); l3 = _mm_set1_epi16 ( src[14] ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst + 6 ), l1 ); _mm_storel_epi64( ( __m128i* ) ( dst + 14 ), l3 ); } _mm_storeu_si128( ( __m128i* ) ( dst - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst + 6 ), l1 ); _mm_storel_epi64( ( __m128i* ) ( dst + 14 ), l3 ); dst += dstStride; _mm_storeu_si128( ( __m128i* ) ( dst - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst + 6 ), l1 ); _mm_storel_epi64( ( __m128i* ) ( dst + 14 ), l3 ); } else { const __m128i sb = _mm_setr_epi8( 0, 1, 0, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 ); __m128i l0 = _mm_shuffle_epi8( _mm_loadu_si128( ( const __m128i* ) &src[ 0] ), sb ); __m128i l1 = _mm_loadu_si128( ( const __m128i* ) &src[ 6] ); __m128i l2 = _mm_loadu_si128( ( const __m128i* ) &src[14] ); __m128i l3 = _mm_set1_epi16 ( src[22] ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 * dstStride - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 * dstStride + 6 ), l1 ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 * dstStride + 14 ), l2 ); _mm_storel_epi64( ( __m128i* ) ( dst - 2 * dstStride + 22 ), l3 ); _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride + 6 ), l1 ); _mm_storeu_si128( ( __m128i* ) ( dst - 1 * dstStride + 14 ), l2 ); _mm_storel_epi64( ( __m128i* ) ( dst - 1 * dstStride + 22 ), l3 ); _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride + 6 ), l1 ); _mm_storeu_si128( ( __m128i* ) ( dst - 0 * dstStride + 14 ), l2 ); _mm_storel_epi64( ( __m128i* ) ( dst - 0 * dstStride + 22 ), l3 ); for( height--, dst += dstStride, src += srcStride; height > 0; height--, src += srcStride, dst += dstStride ) { _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); l0 = _mm_shuffle_epi8( _mm_loadu_si128( ( const __m128i* ) &src[ 0] ), sb ); l1 = _mm_loadu_si128( ( const __m128i* ) &src[ 6] ); l2 = _mm_loadu_si128( ( const __m128i* ) &src[14] ); l3 = _mm_set1_epi16 ( src[22] ); _mm_storeu_si128( ( __m128i* ) ( dst - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst + 6 ), l1 ); _mm_storeu_si128( ( __m128i* ) ( dst + 14 ), l2 ); _mm_storel_epi64( ( __m128i* ) ( dst + 22 ), l3 ); } _mm_storeu_si128( ( __m128i* ) ( dst - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst + 6 ), l1 ); _mm_storeu_si128( ( __m128i* ) ( dst + 14 ), l2 ); _mm_storel_epi64( ( __m128i* ) ( dst + 22 ), l3 ); dst += dstStride; _mm_storeu_si128( ( __m128i* ) ( dst - 2 ), l0 ); _mm_storeu_si128( ( __m128i* ) ( dst + 6 ), l1 ); _mm_storeu_si128( ( __m128i* ) ( dst + 14 ), l2 ); _mm_storel_epi64( ( __m128i* ) ( dst + 22 ), l3 ); } } #if USE_AVX2 template<> void prefetchPadL_SSE( const Pel* src, const ptrdiff_t srcStride, Pel* dst, const ptrdiff_t dstStride, int width, int height ) { _mm_prefetch( ( const char* ) src, _MM_HINT_T0 ); _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); if( width == 15 ) { const __m256i sl = _mm256_setr_epi8( 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 10, 11, 10, 11 ); __m128i l0 = _mm_set1_epi16( src[0] ); __m256i l1 = _mm256_shuffle_epi8( _mm256_loadu_si256( ( const __m256i* ) &src[1] ), sl ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 * dstStride - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst - 2 * dstStride + 1 ), l1 ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 1 * dstStride - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst - 1 * dstStride + 1 ), l1 ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 0 * dstStride - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst - 0 * dstStride + 1 ), l1 ); for( height--, dst += dstStride, src += srcStride; height > 0; height--, src += srcStride, dst += dstStride ) { _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); l0 = _mm_set1_epi16( src[0] ); l1 = _mm256_shuffle_epi8( _mm256_loadu_si256( ( const __m256i* ) & src[1] ), sl ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst + 1 ), l1 ); } _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst + 1 ), l1 ); dst += dstStride; _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst + 1 ), l1 ); } else { const __m128i sl = _mm_setr_epi8( 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 10, 11, 10, 11 ); __m128i l0 = _mm_set1_epi16( src[ 0] ); __m256i l1 = _mm256_loadu_si256( ( const __m256i* ) &src[ 1] ); __m128i l2 = _mm_shuffle_epi8 ( _mm_loadu_si128 ( ( const __m128i* ) &src[17] ), sl ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 * dstStride - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst - 2 * dstStride + 1 ), l1 ); _mm_storeu_si128 ( ( __m128i* ) ( dst - 2 * dstStride + 17 ), l2 ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 1 * dstStride - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst - 1 * dstStride + 1 ), l1 ); _mm_storeu_si128 ( ( __m128i* ) ( dst - 1 * dstStride + 17 ), l2 ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 0 * dstStride - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst - 0 * dstStride + 1 ), l1 ); _mm_storeu_si128 ( ( __m128i* ) ( dst - 0 * dstStride + 17 ), l2 ); for( height--, dst += dstStride, src += srcStride; height > 0; height--, src += srcStride, dst += dstStride ) { _mm_prefetch( ( const char* ) &src[srcStride], _MM_HINT_T0 ); l0 = _mm_set1_epi16( src[ 0] ); l1 = _mm256_loadu_si256( ( const __m256i* ) &src[ 1] ); l2 = _mm_shuffle_epi8 ( _mm_loadu_si128 ( ( const __m128i* ) &src[17] ), sl ); _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst + 1 ), l1 ); _mm_storeu_si128 ( ( __m128i* ) ( dst + 17 ), l2 ); } _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst + 1 ), l1 ); _mm_storeu_si128 ( ( __m128i* ) ( dst + 17 ), l2 ); dst += dstStride; _mm_storel_epi64 ( ( __m128i* ) ( dst - 2 ), l0 ); _mm256_storeu_si256( ( __m256i* ) ( dst + 1 ), l1 ); _mm_storeu_si128 ( ( __m128i* ) ( dst + 17 ), l2 ); } _mm256_zeroupper(); } #endif template void InterPrediction::_initInterPredictionX86() { BiOptFlow = BiOptFlowCoreSIMD ; PaddBIO = PaddBIO_SIMD ; #if !defined( REAL_TARGET_WASM ) // profilings show those functions are slower with WASM SIMD emulation than C++->WASM BioGradFilter = gradFilter_SSE ; profGradFilter = gradFilter_SSE ; #endif applyPROF[0] = applyPROF_SSE ; applyPROF[1] = applyPROF_SSE ; roundIntVector = roundIntVector_SIMD; prefetchPad[0] = prefetchPadL_SSE ; prefetchPad[2] = prefetchPadC_SSE ; } template void InterPrediction::_initInterPredictionX86(); #endif // TARGET_SIMD_X86 #endif }