cpy2Dto1D_shr 515 source/common/arm/asm-primitives.cpp p.cu[BLOCK_4x4].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_4x4_neon); cpy2Dto1D_shr 516 source/common/arm/asm-primitives.cpp p.cu[BLOCK_8x8].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_8x8_neon); cpy2Dto1D_shr 517 source/common/arm/asm-primitives.cpp p.cu[BLOCK_16x16].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_16x16_neon); cpy2Dto1D_shr 518 source/common/arm/asm-primitives.cpp p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32x32_neon); cpy2Dto1D_shr 1003 source/common/pixel.cpp p.cu[BLOCK_ ## W ## x ## H].cpy2Dto1D_shr = cpy2Dto1D_shr<W>; \ cpy2Dto1D_shr 271 source/common/primitives.h cpy2Dto1D_shr_t cpy2Dto1D_shr; cpy2Dto1D_shr 422 source/common/quant.cpp primitives.cu[sizeIdx].cpy2Dto1D_shr(m_resiDctCoeff, residual, resiStride, -transformShift); cpy2Dto1D_shr 958 source/common/x86/asm-primitives.cpp ALL_LUMA_TU_S(cpy2Dto1D_shr, cpy2Dto1D_shr_, sse2); cpy2Dto1D_shr 1585 source/common/x86/asm-primitives.cpp p.cu[BLOCK_8x8].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_8_avx2); cpy2Dto1D_shr 1586 source/common/x86/asm-primitives.cpp p.cu[BLOCK_16x16].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_16_avx2); cpy2Dto1D_shr 1587 source/common/x86/asm-primitives.cpp p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32_avx2); cpy2Dto1D_shr 2301 source/common/x86/asm-primitives.cpp ALL_LUMA_TU_S(cpy2Dto1D_shr, cpy2Dto1D_shr_, sse2); cpy2Dto1D_shr 2916 source/common/x86/asm-primitives.cpp p.cu[BLOCK_8x8].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_8_avx2); cpy2Dto1D_shr 2917 source/common/x86/asm-primitives.cpp p.cu[BLOCK_16x16].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_16_avx2); cpy2Dto1D_shr 2918 source/common/x86/asm-primitives.cpp p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32_avx2); cpy2Dto1D_shr 32 source/common/x86/blockcopy8.h FUNCDEF_TU_S(void, cpy2Dto1D_shr, sse2, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift); cpy2Dto1D_shr 33 source/common/x86/blockcopy8.h FUNCDEF_TU_S(void, cpy2Dto1D_shr, sse4, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift); cpy2Dto1D_shr 34 source/common/x86/blockcopy8.h FUNCDEF_TU_S(void, cpy2Dto1D_shr, avx2, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift); cpy2Dto1D_shr 2412 source/test/pixelharness.cpp if (opt.cu[i].cpy2Dto1D_shr) cpy2Dto1D_shr 2414 source/test/pixelharness.cpp if (!check_cpy2Dto1D_shr_t(ref.cu[i].cpy2Dto1D_shr, opt.cu[i].cpy2Dto1D_shr)) cpy2Dto1D_shr 3047 source/test/pixelharness.cpp if ((i < BLOCK_64x64) && opt.cu[i].cpy2Dto1D_shr) cpy2Dto1D_shr 3050 source/test/pixelharness.cpp REPORT_SPEEDUP(opt.cu[i].cpy2Dto1D_shr, ref.cu[i].cpy2Dto1D_shr, sbuf1, sbuf2, STRIDE, 3);