ref: b397b397502d940a22a0fda35f1d787336dbdaf2
parent: 2ecbc51c2b030f84c6832696b1d465bfded932b7
author: Sindre Aamås <saamas@cisco.com>
date: Wed Jul 20 06:46:21 EDT 2016
[Common] Add AVX2 motion compensation routines The resulting per-block speedup is measured to between ~1.96x and ~3.27x for horizontal filtering, between ~1.43x and ~2.09x for vertical filtering, and between ~1.47x and ~3.04x for 2-dimensional filtering on Haswell, depending on block size, as compared with the existing SSE2 routines. Outliers are 4-wide vertical and 2-dimensional filtering with a speedups of ~5.94x and ~8.41x respectively on Haswell (with clang-703.0.31) due to the C routines used previously.
--- a/codec/common/inc/mc.h
+++ b/codec/common/inc/mc.h
@@ -332,6 +332,28 @@
void McHorVer20Width9Or17U8ToS16_ssse3 (const uint8_t* pSrc, int32_t iSrcStride,
int16_t* pDst, int32_t iDstStride, int32_t iWidth, int32_t iHeight);
+//***************************************************************************//
+// AVX2 definition //
+//***************************************************************************//
+#ifdef HAVE_AVX2
+void McHorVer02_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight);
+void McHorVer02Width4S16ToU8_avx2 (const int16_t* pSrc, uint8_t* pDst, int32_t iDstStride, int32_t iHeight);
+void McHorVer02Width5S16ToU8_avx2 (const int16_t* pSrc, uint8_t* pDst, int32_t iDstStride, int32_t iHeight);
+void McHorVer02Width8S16ToU8_avx2 (const int16_t* pSrc, uint8_t* pDst, int32_t iDstStride, int32_t iHeight);
+void McHorVer02Width9S16ToU8_avx2 (const int16_t* pSrc, uint8_t* pDst, int32_t iDstStride, int32_t iHeight);
+void McHorVer02Width16Or17S16ToU8_avx2 (const int16_t* pSrc, int32_t iSrcStride,
+ uint8_t* pDst, int32_t iDstStride, int32_t iWidth, int32_t iHeight);
+void McHorVer20_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight);
+void McHorVer20Width5Or9Or17_avx2 (const uint8_t* pSrc, int32_t iSrcStride,
+ uint8_t* pDst, int32_t iDstStride, int32_t iWidth, int32_t iHeight);
+void McHorVer20Width4U8ToS16_avx2 (const uint8_t* pSrc, int32_t iSrcStride, int16_t* pDst, int32_t iHeight);
+void McHorVer20Width8U8ToS16_avx2 (const uint8_t* pSrc, int32_t iSrcStride, int16_t* pDst, int32_t iHeight);
+void McHorVer20Width16U8ToS16_avx2 (const uint8_t* pSrc, int32_t iSrcStride, int16_t* pDst, int32_t iHeight);
+void McHorVer20Width17U8ToS16_avx2 (const uint8_t* pSrc, int32_t iSrcStride, int16_t* pDst, int32_t iHeight);
+#endif //HAVE_AVX2
+
#endif //X86_ASM
#if defined(__cplusplus)
--- a/codec/common/src/mc.cpp
+++ b/codec/common/src/mc.cpp
@@ -905,6 +905,169 @@
McChromaWithFragMv_c (pSrc, iSrcStride, pDst, iDstStride, iMvX, iMvY, iWidth, iHeight);
}
+//***************************************************************************//
+// AVX2 implementation //
+//***************************************************************************//
+
+#ifdef HAVE_AVX2
+
+void McHorVer22_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (int16_t, pTmp, 16 + 5, 16, 32);
+ if (iWidth < 8) {+ McHorVer20Width4U8ToS16_avx2 (pSrc, iSrcStride, &pTmp[0][0], iHeight + 5);
+ McHorVer02Width4S16ToU8_avx2 (&pTmp[0][0], pDst, iDstStride, iHeight);
+ } else if (iWidth == 8) {+ McHorVer20Width8U8ToS16_avx2 (pSrc, iSrcStride, &pTmp[0][0], iHeight + 5);
+ McHorVer02Width8S16ToU8_avx2 (&pTmp[0][0], pDst, iDstStride, iHeight);
+ } else {+ McHorVer20Width16U8ToS16_avx2 (pSrc, iSrcStride, &pTmp[0][0], iHeight + 5);
+ McHorVer02Width16Or17S16ToU8_avx2 (&pTmp[0][0], sizeof *pTmp, pDst, iDstStride, iWidth, iHeight);
+ }
+}
+
+void McHorVer01_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pTmp, 16, 16, 16);
+ McHorVer02_avx2 (pSrc, iSrcStride, &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, pSrc, iSrcStride,
+ &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+}
+
+void McHorVer03_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pTmp, 16, 16, 16);
+ McHorVer02_avx2 (pSrc, iSrcStride, &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, pSrc + iSrcStride, iSrcStride,
+ &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+}
+
+void McHorVer10_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc, iSrcStride, &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, pSrc, iSrcStride,
+ &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+}
+
+void McHorVer11_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pHorTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pVerTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc, iSrcStride, &pHorTmp[0][0], sizeof *pHorTmp, iWidth, iHeight);
+ McHorVer02_avx2 (pSrc, iSrcStride, &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pHorTmp[0][0], sizeof *pHorTmp,
+ &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+}
+
+void McHorVer12_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pVerTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pCtrTmp, 16, 16, 16);
+ McHorVer02_avx2 (pSrc, iSrcStride, &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+ McHorVer22_avx2 (pSrc, iSrcStride, &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pVerTmp[0][0], sizeof *pVerTmp,
+ &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+}
+
+void McHorVer13_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pHorTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pVerTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc + iSrcStride, iSrcStride, &pHorTmp[0][0], sizeof *pHorTmp, iWidth, iHeight);
+ McHorVer02_avx2 (pSrc, iSrcStride, &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pHorTmp[0][0], sizeof *pHorTmp,
+ &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+}
+
+void McHorVer21_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pHorTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pCtrTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc, iSrcStride, &pHorTmp[0][0], sizeof *pHorTmp, iWidth, iHeight);
+ McHorVer22_avx2 (pSrc, iSrcStride, &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pHorTmp[0][0], sizeof *pHorTmp,
+ &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+}
+
+void McHorVer23_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pHorTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pCtrTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc + iSrcStride, iSrcStride, &pHorTmp[0][0], sizeof *pHorTmp, iWidth, iHeight);
+ McHorVer22_avx2 (pSrc, iSrcStride, &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pHorTmp[0][0], sizeof *pHorTmp,
+ &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+}
+
+void McHorVer30_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc, iSrcStride, &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, pSrc + 1, iSrcStride, &pTmp[0][0], sizeof *pTmp, iWidth, iHeight);
+}
+
+void McHorVer31_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pHorTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pVerTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc, iSrcStride, &pHorTmp[0][0], sizeof *pHorTmp, iWidth, iHeight);
+ McHorVer02_avx2 (pSrc + 1, iSrcStride, &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pHorTmp[0][0], sizeof *pHorTmp,
+ &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+}
+
+void McHorVer32_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pVerTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pCtrTmp, 16, 16, 16);
+ McHorVer02_avx2 (pSrc + 1, iSrcStride, &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+ McHorVer22_avx2 (pSrc, iSrcStride, &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pVerTmp[0][0], sizeof *pVerTmp,
+ &pCtrTmp[0][0], sizeof *pCtrTmp, iWidth, iHeight);
+}
+
+void McHorVer33_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ ENFORCE_STACK_ALIGN_2D (uint8_t, pHorTmp, 16, 16, 16);
+ ENFORCE_STACK_ALIGN_2D (uint8_t, pVerTmp, 16, 16, 16);
+ McHorVer20_avx2 (pSrc + iSrcStride, iSrcStride, &pHorTmp[0][0], sizeof *pHorTmp, iWidth, iHeight);
+ McHorVer02_avx2 (pSrc + 1, iSrcStride, &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+ PixelAvgWidth4Or8Or16_sse2 (pDst, iDstStride, &pHorTmp[0][0], sizeof *pHorTmp,
+ &pVerTmp[0][0], sizeof *pVerTmp, iWidth, iHeight);
+}
+
+void McHorVer22Width5Or9Or17_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int32_t iWidth, int32_t iHeight) {+ if (iWidth < 9) {+ ENFORCE_STACK_ALIGN_2D (int16_t, pTmp, 9 + 5, WELS_ALIGN(5, 16 / sizeof (int16_t)), 16)
+ McHorVer20Width8U8ToS16_avx2 (pSrc, iSrcStride, &pTmp[0][0], iHeight + 5);
+ McHorVer02Width5S16ToU8_avx2 (&pTmp[0][0], pDst, iDstStride, iHeight);
+ } else if (iWidth == 9) {+ ENFORCE_STACK_ALIGN_2D (int16_t, pTmp, 17 + 5, 16, 32)
+ McHorVer20Width16U8ToS16_avx2 (pSrc, iSrcStride, &pTmp[0][0], iHeight + 5);
+ McHorVer02Width9S16ToU8_avx2 (&pTmp[0][0], pDst, iDstStride, iHeight);
+ } else {+ ENFORCE_STACK_ALIGN_2D (int16_t, pTmp, 17 + 5, WELS_ALIGN(17, 32 / sizeof (int16_t)), 32)
+ McHorVer20Width17U8ToS16_avx2 (pSrc, iSrcStride, &pTmp[0][0], iHeight + 5);
+ McHorVer02Width16Or17S16ToU8_avx2 (&pTmp[0][0], sizeof *pTmp, pDst, iDstStride, iWidth, iHeight);
+ }
+}
+
+void McLuma_avx2 (const uint8_t* pSrc, int32_t iSrcStride, uint8_t* pDst, int32_t iDstStride,
+ int16_t iMvX, int16_t iMvY, int32_t iWidth, int32_t iHeight) {+ static const PWelsMcWidthHeightFunc pWelsMcFunc[4][4] = {+ {McCopy_sse3, McHorVer01_avx2, McHorVer02_avx2, McHorVer03_avx2},+ {McHorVer10_avx2, McHorVer11_avx2, McHorVer12_avx2, McHorVer13_avx2},+ {McHorVer20_avx2, McHorVer21_avx2, McHorVer22_avx2, McHorVer23_avx2},+ {McHorVer30_avx2, McHorVer31_avx2, McHorVer32_avx2, McHorVer33_avx2},+ };
+
+ pWelsMcFunc[iMvX & 0x03][iMvY & 0x03] (pSrc, iSrcStride, pDst, iDstStride, iWidth, iHeight);
+}
+
+#endif //HAVE_AVX2
+
void PixelAvg_sse2 (uint8_t* pDst, int32_t iDstStride, const uint8_t* pSrcA, int32_t iSrcAStride,
const uint8_t* pSrcB, int32_t iSrcBStride, int32_t iWidth, int32_t iHeight) { static const PWelsSampleWidthAveragingFunc kpfFuncs[2] = {@@ -1523,6 +1686,14 @@
pMcFuncs->pMcChromaFunc = McChroma_ssse3;
pMcFuncs->pMcLumaFunc = McLuma_ssse3;
}
+#ifdef HAVE_AVX2
+ if (uiCpuFlag & WELS_CPU_AVX2) {+ pMcFuncs->pfLumaHalfpelHor = McHorVer20Width5Or9Or17_avx2;
+ pMcFuncs->pfLumaHalfpelVer = McHorVer02_avx2;
+ pMcFuncs->pfLumaHalfpelCen = McHorVer22Width5Or9Or17_avx2;
+ pMcFuncs->pMcLumaFunc = McLuma_avx2;
+ }
+#endif
#endif //(X86_ASM)
#if defined(HAVE_NEON)
--- a/codec/common/x86/mc_luma.asm
+++ b/codec/common/x86/mc_luma.asm
@@ -44,12 +44,38 @@
;*******************************************************************************
; Local Data (Read Only)
;*******************************************************************************
-SECTION .rodata align=16
+SECTION .rodata align=32
;*******************************************************************************
; Various memory constants (trigonometric values or rounding values)
;*******************************************************************************
+%ifdef HAVE_AVX2
+ALIGN 32
+dwm32768_256:
+ times 16 dw -32768
+maddubsw_m2p10_m40m40_p10m2_p0p0_256:
+ times 4 db -2, 10, -40, -40, 10, -2, 0, 0
+dwm1024_256:
+ times 16 dw -1024
+dd32768_256:
+ times 8 dd 32768
+maddubsw_p1m5_256:
+ times 16 db 1, -5
+maddubsw_m5p1_256:
+ times 16 db -5, 1
+db20_256:
+ times 32 db 20
+maddubsw_m5p20_256:
+ times 16 db -5, 20
+maddubsw_p20m5_256:
+ times 16 db 20, -5
+h264_w0x10_256:
+ times 16 dw 16
+dw32_256:
+ times 16 dw 32
+%endif ; HAVE_AVX2
+
ALIGN 16
shufb_32435465768798A9:
db 3, 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7, 9, 8, 10, 9
@@ -2874,3 +2900,1450 @@
%undef i_width
%undef i_height
%undef i_srcstride3
+
+
+%ifdef HAVE_AVX2
+
+; pixels=%1 shufb_32435465768798A9=%2 shufb_011267784556ABBC=%3 maddubsw_p1m5_p1m5_m5p1_m5p1=%4 tmp=%5,%6
+%macro AVX2_FilterHorizontalbw_16px 6
+ vpshufb %5, %1, %3
+ vpshufb %1, %1, %2
+ vpshufd %6, %1, 10110001b
+ vpmaddubsw %1, %1, [db20_256]
+ vpmaddubsw %5, %5, %4
+ vpmaddubsw %6, %6, %4
+ vpaddw %1, %1, %5
+ vpaddw %1, %1, %6
+%endmacro
+
+; pixels=%1 shufb_32435465768798A9=%2 shufb_011267784556ABBC=%3 db20=%4 tmp=%5,%6
+%macro AVX2_FilterHorizontal_16px 6
+ AVX2_FilterHorizontalbw_16px %1, %2, %3, %4, %5, %6
+ vpaddw %1, %1, [h264_w0x10_256]
+ vpsraw %1, %1, 5
+%endmacro
+
+; px0=%1 px1=%2 shufb_32435465768798A9=%3 shufb_011267784556ABBC=%4 maddubsw_p1m5_p1m5_m5p1_m5p1=%5 tmp=%6,%7
+%macro AVX2_FilterHorizontalbw_4x4px 7
+ vpshufb %6, %1, %4
+ vpshufb %7, %2, %4
+ vpshufb %1, %1, %3
+ vpshufb %2, %2, %3
+ vpunpcklqdq %1, %1, %2
+ vpunpcklqdq %6, %6, %7
+ vpshufd %7, %1, 10110001b
+ vpmaddubsw %1, %1, [db20_256]
+ vpmaddubsw %6, %6, %5
+ vpmaddubsw %7, %7, %5
+ vpaddw %1, %1, %6
+ vpaddw %1, %1, %7
+%endmacro
+
+; px0=%1 px1=%2 shufb_32435465768798A9=%3 shufb_011267784556ABBC=%4 db20=%5 tmp=%6,%7
+%macro AVX2_FilterHorizontal_4x4px 7
+ AVX2_FilterHorizontalbw_4x4px %1, %2, %3, %4, %5, %6, %7
+ vpaddw %1, %1, [h264_w0x10_256]
+ vpsraw %1, %1, 5
+%endmacro
+
+; pixels=%1 -32768>>scale=%2 tmp=%3
+%macro AVX2_FilterHorizontalbw_4px 3
+ vpmaddubsw %1, %1, [maddubsw_m2p10_m40m40_p10m2_p0p0_256]
+ vpmaddwd %1, %1, %2
+ vpshufd %3, %1, 10110001b
+ vpaddd %1, %1, %3
+%endmacro
+
+; pixels=%1 tmp=%2
+%macro AVX2_FilterHorizontal_4px 2
+ AVX2_FilterHorizontalbw_4px %1, [dwm1024_256], %2
+ vpaddd %1, %1, [dd32768_256]
+%endmacro
+
+; px_ab=%1 px_cd=%2 px_ef=%3 maddubsw_ab=%4 maddubsw_cd=%5 maddubsw_ef=%6 tmp=%7
+%macro AVX2_FilterVertical_16px 7
+ vpmaddubsw %1, %1, %4
+ vpmaddubsw %7, %2, %5
+ vpaddw %1, %1, %7
+ vpmaddubsw %7, %3, %6
+ vpaddw %1, %1, %7
+ vpaddw %1, %1, [h264_w0x10_256]
+ vpsraw %1, %1, 5
+%endmacro
+
+; px_a=%1 px_f=%2 px_bc=%3 px_de=%4 maddubsw_bc=%5 maddubsw_de=%6 tmp=%7,%8
+%macro AVX2_FilterVertical2_16px 8
+ vpxor %7, %7, %7
+ vpunpcklbw %1, %1, %7
+ vpunpcklbw %8, %2, %7
+ vpaddw %1, %1, %8
+ vpmaddubsw %7, %3, %5
+ vpaddw %1, %1, %7
+ vpmaddubsw %7, %4, %6
+ vpaddw %1, %1, %7
+ vpaddw %1, %1, [h264_w0x10_256]
+ vpsraw %1, %1, 5
+%endmacro
+
+; px0=%1 px1=%2 px2=%3 px3=%4 px4=%5 px5=%6 tmp=%7
+%macro AVX2_FilterVerticalw_16px 7
+ vpaddw %1, %1, %6
+ vpaddw %7, %2, %5
+ vpsubw %1, %1, %7
+ vpsraw %1, %1, 2
+ vpsubw %1, %1, %7
+ vpaddw %7, %3, %4
+ vpaddw %1, %1, %7
+ vpsraw %1, %1, 2
+ vpaddw %7, %7, [dw32_256]
+ vpaddw %1, %1, %7
+ vpsraw %1, %1, 6
+%endmacro
+
+;***********************************************************************
+; void McHorVer02_avx2(const uint8_t *pSrc,
+; int32_t iSrcStride,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iWidth,
+; int32_t iHeight)
+;***********************************************************************
+
+WELS_EXTERN McHorVer02_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_dststride r3
+%define i_width r4
+%define i_height r5
+%define i_srcstride3 r6
+ %assign push_num 0
+%ifdef X86_32
+ push r6
+ %assign push_num 1
+%endif
+ LOAD_6_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ SIGN_EXTENSION r4, r4d
+ SIGN_EXTENSION r5, r5d
+ sub p_src, i_srcstride
+ sub p_src, i_srcstride
+ lea i_srcstride3, [3 * i_srcstride]
+ cmp i_width, 8
+ je .width8
+ jg .width16
+; .width4:
+ vmovd xmm0, [p_src]
+ vpbroadcastd xmm5, [p_src + i_srcstride]
+ vpunpcklbw xmm0, xmm0, xmm5
+ vpbroadcastd ymm1, [p_src + 2 * i_srcstride]
+ vpunpcklbw xmm5, xmm5, xmm1
+ vpblendd xmm0, xmm0, xmm5, 1100b
+ vpbroadcastd ymm5, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpcklbw ymm1, ymm1, ymm5
+ vpbroadcastd ymm2, [p_src]
+ vpunpcklbw ymm5, ymm5, ymm2
+ vpblendd ymm1, ymm1, ymm5, 11001100b
+ vpblendd ymm0, ymm0, ymm1, 11110000b
+ vpbroadcastd ymm5, [p_src + i_srcstride]
+ lea p_src, [p_src + 2 * i_srcstride]
+ vpunpcklbw ymm2, ymm2, ymm5
+ vpbroadcastd ymm3, [p_src]
+ vpunpcklbw ymm5, ymm5, ymm3
+ vpblendd ymm2, ymm2, ymm5, 11001100b
+ vpblendd ymm1, ymm1, ymm2, 11110000b
+ vpbroadcastd ymm5, [p_src + i_srcstride]
+ vpunpcklbw ymm3, ymm3, ymm5
+ vpbroadcastd ymm4, [p_src + 2 * i_srcstride]
+ vpunpcklbw ymm5, ymm5, ymm4
+ vpblendd ymm3, ymm3, ymm5, 11001100b
+ vpblendd ymm2, ymm2, ymm3, 11110000b
+ vbroadcasti128 ymm6, [db20_128]
+ AVX2_FilterVertical_16px ymm0, ymm1, ymm2, [maddubsw_p1m5_256], ymm6, [maddubsw_m5p1_256], ymm5
+ vpackuswb ymm0, ymm0, ymm0
+ vmovd [p_dst], xmm0
+ vpsrlq xmm5, xmm0, 32
+ vmovd [p_dst + i_dststride], xmm5
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vextracti128 xmm0, ymm0, 1
+ vmovd [p_dst], xmm0
+ vpsrlq xmm5, xmm0, 32
+ vmovd [p_dst + i_dststride], xmm5
+ cmp i_height, 5
+ jl .width4_done
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vpbroadcastd ymm5, [p_src + i_srcstride3]
+ vpunpcklbw ymm4, ymm4, ymm5
+ jg .width4_height_ge8
+ AVX2_FilterVertical_16px xmm2, xmm3, xmm4, [maddubsw_p1m5_256], xmm6, [maddubsw_m5p1_256], xmm5
+ vpackuswb xmm2, xmm2, xmm2
+ vmovd [p_dst], xmm2
+ jmp .width4_done
+.width4_height_ge8:
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpbroadcastd ymm1, [p_src]
+ vpunpcklbw ymm5, ymm5, ymm1
+ vpblendd ymm4, ymm4, ymm5, 11001100b
+ vpblendd ymm3, ymm3, ymm4, 11110000b
+ vpbroadcastd ymm5, [p_src + i_srcstride]
+ vpunpcklbw ymm1, ymm5
+ vpbroadcastd ymm0, [p_src + 2 * i_srcstride]
+ vpunpcklbw ymm5, ymm5, ymm0
+ vpblendd ymm1, ymm1, ymm5, 11001100b
+ vpblendd ymm4, ymm4, ymm1, 11110000b
+ AVX2_FilterVertical_16px ymm2, ymm3, ymm4, [maddubsw_p1m5_256], ymm6, [maddubsw_m5p1_256], ymm5
+ vpackuswb ymm2, ymm2, ymm2
+ vmovd [p_dst], xmm2
+ vpsrlq xmm5, xmm2, 32
+ vmovd [p_dst + i_dststride], xmm5
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vextracti128 xmm2, ymm2, 1
+ vmovd [p_dst], xmm2
+ vpsrlq xmm5, xmm2, 32
+ vmovd [p_dst + i_dststride], xmm5
+ cmp i_height, 9
+ jl .width4_done
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovd xmm5, [p_src + i_srcstride3]
+ vpunpcklbw xmm0, xmm0, xmm5
+ AVX2_FilterVertical_16px xmm4, xmm1, xmm0, [maddubsw_p1m5_256], xmm6, [maddubsw_m5p1_256], xmm5
+ vpackuswb xmm4, xmm4, xmm4
+ vmovd [p_dst], xmm4
+.width4_done:
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+%ifdef X86_32
+ pop r6
+%endif
+ ret
+
+.width8:
+ sub i_height, 1
+ vmovq xmm0, [p_src]
+ vmovq xmm4, [p_src + i_srcstride]
+ vpunpcklbw xmm0, xmm0, xmm4
+ vmovq xmm1, [p_src + 2 * i_srcstride]
+ vpunpcklbw xmm4, xmm4, xmm1
+ vinserti128 ymm0, ymm0, xmm4, 1
+ vmovq xmm4, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpcklbw xmm1, xmm1, xmm4
+ vmovq xmm6, [p_src]
+ vpunpcklbw xmm4, xmm4, xmm6
+ vinserti128 ymm1, ymm1, xmm4, 1
+.width8_yloop:
+ vmovq xmm4, [p_src + i_srcstride]
+ vpunpcklbw xmm2, xmm6, xmm4
+ vmovq xmm3, [p_src + 2 * i_srcstride]
+ vpunpcklbw xmm4, xmm4, xmm3
+ vinserti128 ymm2, ymm2, xmm4, 1
+ vbroadcasti128 ymm5, [db20_128]
+ AVX2_FilterVertical_16px ymm0, ymm1, ymm2, [maddubsw_p1m5_256], ymm5, [maddubsw_m5p1_256], ymm4
+ vmovq xmm4, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpcklbw xmm3, xmm3, xmm4
+ vmovq xmm6, [p_src]
+ vpunpcklbw xmm4, xmm4, xmm6
+ vinserti128 ymm3, ymm3, xmm4, 1
+ AVX2_FilterVertical_16px ymm1, ymm2, ymm3, [maddubsw_p1m5_256], ymm5, [maddubsw_m5p1_256], ymm4
+ vpackuswb ymm0, ymm0, ymm1
+ vmovlps [p_dst], xmm0
+ vextracti128 xmm1, ymm0, 1
+ vmovlps [p_dst + i_dststride], xmm1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovhps [p_dst], xmm0
+ vmovhps [p_dst + i_dststride], xmm1
+ cmp i_height, 4
+ jl .width8_done
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovq xmm4, [p_src + i_srcstride]
+ vpunpcklbw xmm0, xmm6, xmm4
+ jg .width8_height_ge8
+ AVX2_FilterVertical_16px xmm2, xmm3, xmm0, [maddubsw_p1m5_256], xmm5, [maddubsw_m5p1_256], xmm4
+ vpackuswb xmm2, xmm2, xmm2
+ vmovlps [p_dst], xmm2
+ jmp .width8_done
+.width8_height_ge8:
+ vmovq xmm1, [p_src + 2 * i_srcstride]
+ vpunpcklbw xmm4, xmm4, xmm1
+ vinserti128 ymm0, ymm0, xmm4, 1
+ AVX2_FilterVertical_16px ymm2, ymm3, ymm0, [maddubsw_p1m5_256], ymm5, [maddubsw_m5p1_256], ymm4
+ vmovq xmm4, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpcklbw xmm1, xmm1, xmm4
+ vmovq xmm6, [p_src]
+ vpunpcklbw xmm4, xmm4, xmm6
+ vinserti128 ymm1, ymm1, xmm4, 1
+ AVX2_FilterVertical_16px ymm3, ymm0, ymm1, [maddubsw_p1m5_256], ymm5, [maddubsw_m5p1_256], ymm4
+ vpackuswb ymm2, ymm2, ymm3
+ vmovlps [p_dst], xmm2
+ vextracti128 xmm3, ymm2, 1
+ vmovlps [p_dst + i_dststride], xmm3
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovhps [p_dst], xmm2
+ vmovhps [p_dst + i_dststride], xmm3
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 8
+ jg .width8_yloop
+ jl .width8_done
+ vmovq xmm4, [p_src + i_srcstride]
+ vpunpcklbw xmm2, xmm6, xmm4
+ AVX2_FilterVertical_16px xmm0, xmm1, xmm2, [maddubsw_p1m5_256], xmm5, [maddubsw_m5p1_256], xmm4
+ vpackuswb xmm0, xmm0, xmm0
+ vmovlps [p_dst], xmm0
+.width8_done:
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+%ifdef X86_32
+ pop r6
+%endif
+ ret
+
+.width16:
+ sub i_height, 1
+ test i_height, 1
+ jnz .width16_yloop_begin_even
+ vmovq xmm0, [p_src]
+ vpbroadcastq ymm1, [p_src + 8]
+ vpblendd ymm0, ymm0, ymm1, 11110000b
+ vmovq xmm1, [p_src + i_srcstride]
+ vpbroadcastq ymm2, [p_src + i_srcstride + 8]
+ vpblendd ymm1, ymm1, ymm2, 11110000b
+ vpunpcklbw ymm0, ymm0, ymm1
+ vmovq xmm2, [p_src + 2 * i_srcstride]
+ vpbroadcastq ymm3, [p_src + 2 * i_srcstride + 8]
+ vpblendd ymm2, ymm2, ymm3, 11110000b
+ vmovq xmm3, [p_src + i_srcstride3]
+ vpbroadcastq ymm4, [p_src + i_srcstride3 + 8]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpblendd ymm3, ymm3, ymm4, 11110000b
+ vpunpcklbw ymm2, ymm2, ymm3
+ vmovq xmm4, [p_src]
+ vpbroadcastq ymm5, [p_src + 8]
+ vpblendd ymm4, ymm4, ymm5, 11110000b
+ vmovq xmm5, [p_src + i_srcstride]
+ vpbroadcastq ymm6, [p_src + i_srcstride + 8]
+ lea p_src, [p_src + 2 * i_srcstride]
+ vpblendd ymm5, ymm5, ymm6, 11110000b
+ vpunpcklbw ymm4, ymm4, ymm5
+ AVX2_FilterVertical_16px ymm0, ymm2, ymm4, [maddubsw_p1m5_256], [db20_256], [maddubsw_m5p1_256], ymm7
+ vpackuswb ymm0, ymm0, ymm0
+ vpermq ymm0, ymm0, 1000b
+ vmovdqa [p_dst], xmm0
+ add p_dst, i_dststride
+ jmp .width16_yloop
+.width16_yloop_begin_even:
+ vmovq xmm1, [p_src]
+ vpbroadcastq ymm2, [p_src + 8]
+ vpblendd ymm1, ymm1, ymm2, 11110000b
+ vmovq xmm2, [p_src + i_srcstride]
+ vpbroadcastq ymm3, [p_src + i_srcstride + 8]
+ vpblendd ymm2, ymm2, ymm3, 11110000b
+ vmovq xmm3, [p_src + 2 * i_srcstride]
+ vpbroadcastq ymm4, [p_src + 2 * i_srcstride + 8]
+ add p_src, i_srcstride3
+ vpblendd ymm3, ymm3, ymm4, 11110000b
+ vpunpcklbw ymm2, ymm2, ymm3
+ vmovq xmm4, [p_src]
+ vpbroadcastq ymm5, [p_src + 8]
+ vpblendd ymm4, ymm4, ymm5, 11110000b
+ vmovq xmm5, [p_src + i_srcstride]
+ vpbroadcastq ymm6, [p_src + i_srcstride + 8]
+ lea p_src, [p_src + 2 * i_srcstride]
+ vpblendd ymm5, ymm5, ymm6, 11110000b
+ vpunpcklbw ymm4, ymm4, ymm5
+.width16_yloop:
+ vmovq xmm6, [p_src]
+ vpbroadcastq ymm7, [p_src + 8]
+ vpblendd ymm6, ymm6, ymm7, 11110000b
+ AVX2_FilterVertical2_16px ymm1, ymm6, ymm2, ymm4, [maddubsw_m5p20_256], [maddubsw_p20m5_256], ymm0, ymm7
+ vmovq xmm7, [p_src + i_srcstride]
+ vpbroadcastq ymm0, [p_src + i_srcstride + 8]
+ vpblendd ymm7, ymm7, ymm0, 11110000b
+ vpunpcklbw ymm6, ymm6, ymm7
+ AVX2_FilterVertical_16px ymm2, ymm4, ymm6, [maddubsw_p1m5_256], [db20_256], [maddubsw_m5p1_256], ymm0
+ vpackuswb ymm1, ymm1, ymm2
+ vpermq ymm1, ymm1, 11011000b
+ vmovdqa [p_dst], xmm1
+ vextracti128 [p_dst + i_dststride], ymm1, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovq xmm0, [p_src + 2 * i_srcstride]
+ vpbroadcastq ymm1, [p_src + 2 * i_srcstride + 8]
+ vpblendd ymm0, ymm0, ymm1, 11110000b
+ AVX2_FilterVertical2_16px ymm3, ymm0, ymm4, ymm6, [maddubsw_m5p20_256], [maddubsw_p20m5_256], ymm2, ymm1
+ vmovq xmm1, [p_src + i_srcstride3]
+ vpbroadcastq ymm2, [p_src + i_srcstride3 + 8]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpblendd ymm1, ymm1, ymm2, 11110000b
+ vpunpcklbw ymm0, ymm0, ymm1
+ AVX2_FilterVertical_16px ymm4, ymm6, ymm0, [maddubsw_p1m5_256], [db20_256], [maddubsw_m5p1_256], ymm2
+ vpackuswb ymm3, ymm3, ymm4
+ vpermq ymm3, ymm3, 11011000b
+ vmovdqa [p_dst], xmm3
+ vextracti128 [p_dst + i_dststride], ymm3, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovq xmm2, [p_src]
+ vpbroadcastq ymm3, [p_src + 8]
+ vpblendd ymm2, ymm2, ymm3, 11110000b
+ AVX2_FilterVertical2_16px ymm5, ymm2, ymm6, ymm0, [maddubsw_m5p20_256], [maddubsw_p20m5_256], ymm4, ymm3
+ vmovq xmm3, [p_src + i_srcstride]
+ vpbroadcastq ymm4, [p_src + i_srcstride + 8]
+ vpblendd ymm3, ymm3, ymm4, 11110000b
+ vpunpcklbw ymm2, ymm2, ymm3
+ AVX2_FilterVertical_16px ymm6, ymm0, ymm2, [maddubsw_p1m5_256], [db20_256], [maddubsw_m5p1_256], ymm4
+ vpackuswb ymm5, ymm5, ymm6
+ vpermq ymm5, ymm5, 11011000b
+ vmovdqa [p_dst], xmm5
+ vextracti128 [p_dst + i_dststride], ymm5, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovq xmm4, [p_src + 2 * i_srcstride]
+ vpbroadcastq ymm5, [p_src + 2 * i_srcstride + 8]
+ vpblendd ymm4, ymm4, ymm5, 11110000b
+ AVX2_FilterVertical2_16px ymm7, ymm4, ymm0, ymm2, [maddubsw_m5p20_256], [maddubsw_p20m5_256], ymm6, ymm5
+ vmovq xmm5, [p_src + i_srcstride3]
+ vpbroadcastq ymm6, [p_src + i_srcstride3 + 8]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpblendd ymm5, ymm5, ymm6, 11110000b
+ vpunpcklbw ymm4, ymm4, ymm5
+ AVX2_FilterVertical_16px ymm0, ymm2, ymm4, [maddubsw_p1m5_256], [db20_256], [maddubsw_m5p1_256], ymm6
+ vpackuswb ymm7, ymm7, ymm0
+ vpermq ymm7, ymm7, 11011000b
+ vmovdqa [p_dst], xmm7
+ vextracti128 [p_dst + i_dststride], ymm7, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 8
+ jg .width16_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+%ifdef X86_32
+ pop r6
+%endif
+ ret
+%undef p_src
+%undef i_srcstride
+%undef i_srcstride3
+%undef p_dst
+%undef i_dststride
+%undef i_width
+%undef i_height
+%undef i_ycnt
+
+
+;*******************************************************************************
+; void McHorVer20_avx2(const uint8_t *pSrc,
+; int iSrcStride,
+; uint8_t *pDst,
+; int iDstStride,
+; int iWidth,
+; int iHeight);
+;*******************************************************************************
+
+WELS_EXTERN McHorVer20_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_dststride r3
+%define i_width r4
+%define i_height r5
+ %assign push_num 0
+ LOAD_6_PARA
+ PUSH_XMM 7
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ SIGN_EXTENSION r4, r4d
+ SIGN_EXTENSION r5, r5d
+ vbroadcasti128 ymm4, [shufb_32435465768798A9]
+ vbroadcasti128 ymm5, [shufb_011267784556ABBC]
+ vbroadcasti128 ymm6, [maddubsw_p1m5_p1m5_m5p1_m5p1_128]
+ cmp i_width, 8
+ je .width8
+ jg .width16_yloop
+%xdefine i_srcstride3 i_width
+%undef i_width
+ lea i_srcstride3, [3 * i_srcstride]
+.width4_yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm1, [p_src + i_srcstride - 2]
+ vinserti128 ymm0, ymm0, [p_src + 2 * i_srcstride - 2], 1
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride3 - 2], 1
+ lea p_src, [p_src + 4 * i_srcstride]
+ AVX2_FilterHorizontal_4x4px ymm0, ymm1, ymm4, ymm5, ymm6, ymm2, ymm3
+ vpackuswb ymm0, ymm0, ymm0
+ vmovd [p_dst], xmm0
+ vpsrlq xmm1, xmm0, 32
+ vmovd [p_dst + i_dststride], xmm1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vextracti128 xmm0, ymm0, 1
+ vmovd [p_dst], xmm0
+ vpsrlq xmm1, xmm0, 32
+ vmovd [p_dst + i_dststride], xmm1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 4
+ jg .width4_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+ ret
+.width8:
+ lea i_srcstride3, [3 * i_srcstride]
+.width8_yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm1, [p_src + i_srcstride - 2]
+ vinserti128 ymm0, ymm0, [p_src + 2 * i_srcstride - 2], 1
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride3 - 2], 1
+ lea p_src, [p_src + 4 * i_srcstride]
+ AVX2_FilterHorizontal_16px ymm0, ymm4, ymm5, ymm6, ymm2, ymm3
+ AVX2_FilterHorizontal_16px ymm1, ymm4, ymm5, ymm6, ymm2, ymm3
+ vpackuswb ymm0, ymm0, ymm1
+ vmovlps [p_dst], xmm0
+ vmovhps [p_dst + i_dststride], xmm0
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vextracti128 xmm0, ymm0, 1
+ vmovlps [p_dst], xmm0
+ vmovhps [p_dst + i_dststride], xmm0
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 4
+ jg .width8_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+ ret
+%undef i_srcstride3
+.width16_yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm1, [p_src + 6]
+ vinserti128 ymm0, ymm0, [p_src + i_srcstride - 2], 1
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride + 6], 1
+ lea p_src, [p_src + 2 * i_srcstride]
+ AVX2_FilterHorizontal_16px ymm0, ymm4, ymm5, ymm6, ymm2, ymm3
+ AVX2_FilterHorizontal_16px ymm1, ymm4, ymm5, ymm6, ymm2, ymm3
+ vpackuswb ymm0, ymm0, ymm1
+ vmovdqa [p_dst], xmm0
+ vextracti128 [p_dst + i_dststride], ymm0, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 2
+ jg .width16_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_dststride
+%undef i_width
+%undef i_height
+
+
+;***********************************************************************
+; void McHorVer20Width5Or9Or17_avx2(const uint8_t *pSrc,
+; int32_t iSrcStride,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iWidth,
+; int32_t iHeight);
+;***********************************************************************
+
+WELS_EXTERN McHorVer20Width5Or9Or17_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_dststride r3
+%define i_width r4
+%define i_height r5
+ %assign push_num 0
+ LOAD_6_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ SIGN_EXTENSION r4, r4d
+ SIGN_EXTENSION r5, r5d
+ vbroadcasti128 ymm5, [shufb_32435465768798A9]
+ vbroadcasti128 ymm6, [shufb_011267784556ABBC]
+ vbroadcasti128 ymm7, [maddubsw_p1m5_p1m5_m5p1_m5p1_128]
+ cmp i_width, 9
+ je .width9
+ jg .width17
+.width5_yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vinserti128 ymm0, ymm0, [p_src + i_srcstride - 2], 1
+ lea p_src, [p_src + 2 * i_srcstride]
+ AVX2_FilterHorizontal_16px ymm0, ymm5, ymm6, ymm7, ymm1, ymm2
+ vpackuswb ymm0, ymm0, ymm0
+ vpsrlq xmm1, xmm0, 8
+ vmovd [p_dst + 1], xmm1
+ vmovd [p_dst], xmm0
+ add p_dst, i_dststride
+ vextracti128 xmm0, ymm0, 1
+ vpsrlq xmm1, xmm0, 8
+ vmovd [p_dst + 1], xmm1
+ vmovd [p_dst], xmm0
+ add p_dst, i_dststride
+ sub i_height, 2
+ jg .width5_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+ ret
+.width9:
+%xdefine i_srcstride3 i_width
+%undef i_width
+ lea i_srcstride3, [3 * i_srcstride]
+.width9_yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm4, [p_src + i_srcstride - 2]
+ vinserti128 ymm0, ymm0, [p_src + 2 * i_srcstride - 2], 1
+ vinserti128 ymm4, ymm4, [p_src + i_srcstride3 - 2], 1
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpckhqdq ymm3, ymm0, ymm4
+ AVX2_FilterHorizontal_4px ymm3, ymm2
+ AVX2_FilterHorizontal_16px ymm0, ymm5, ymm6, ymm7, ymm1, ymm2
+ vpackuswb ymm3, ymm3, ymm0
+ vmovd [p_dst + 5], xmm3
+ vmovhps [p_dst], xmm3
+ add p_dst, i_dststride
+ AVX2_FilterHorizontal_16px ymm4, ymm5, ymm6, ymm7, ymm1, ymm2
+ vpackuswb ymm4, ymm4, ymm4
+ vpsrlq xmm2, xmm3, 32
+ vmovd [p_dst + 5], xmm2
+ vmovlps [p_dst], xmm4
+ add p_dst, i_dststride
+ vextracti128 xmm3, ymm3, 1
+ vextracti128 xmm4, ymm4, 1
+ vmovd [p_dst + 5], xmm3
+ vmovhps [p_dst], xmm3
+ add p_dst, i_dststride
+ vpsrlq xmm2, xmm3, 32
+ vmovd [p_dst + 5], xmm2
+ vmovlps [p_dst], xmm4
+ add p_dst, i_dststride
+ sub i_height, 4
+ jg .width9_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+ ret
+.width17:
+ lea i_srcstride3, [3 * i_srcstride]
+.width17_yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm3, [p_src + 6]
+ vinserti128 ymm0, ymm0, [p_src + i_srcstride - 2], 1
+ vinserti128 ymm3, ymm3, [p_src + i_srcstride + 6], 1
+ vmovdqa ymm4, ymm3
+ AVX2_FilterHorizontal_16px ymm0, ymm5, ymm6, ymm7, ymm1, ymm2
+ AVX2_FilterHorizontal_16px ymm3, ymm5, ymm6, ymm7, ymm1, ymm2
+ vpackuswb ymm0, ymm0, ymm3
+ vmovdqu xmm1, [p_src + 2 * i_srcstride - 2]
+ vmovdqu xmm3, [p_src + 2 * i_srcstride + 6]
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride3 - 2], 1
+ vinserti128 ymm3, ymm3, [p_src + i_srcstride3 + 6], 1
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpckhqdq ymm4, ymm4, ymm3
+ AVX2_FilterHorizontal_4px ymm4, ymm2
+ vpackuswb ymm4, ymm4, ymm4
+ vmovd [p_dst + 13], xmm4
+ vmovdqa [p_dst], xmm0
+ add p_dst, i_dststride
+ vextracti128 xmm2, ymm4, 1
+ vmovd [p_dst + 13], xmm2
+ vextracti128 [p_dst], ymm0, 1
+ add p_dst, i_dststride
+ vpsrlq xmm4, xmm4, 32
+ vmovd [p_dst + 13], xmm4
+ AVX2_FilterHorizontal_16px ymm1, ymm5, ymm6, ymm7, ymm0, ymm4
+ AVX2_FilterHorizontal_16px ymm3, ymm5, ymm6, ymm7, ymm0, ymm4
+ vpackuswb ymm1, ymm1, ymm3
+ vmovdqa [p_dst], xmm1
+ add p_dst, i_dststride
+ vpsrlq xmm2, xmm2, 32
+ vmovd [p_dst + 13], xmm2
+ vextracti128 [p_dst], ymm1, 1
+ add p_dst, i_dststride
+ sub i_height, 4
+ jg .width17_yloop
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+ ret
+%undef i_srcstride3
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_dststride
+%undef i_width
+%undef i_height
+
+
+;*******************************************************************************
+; void McHorVer20Width4U8ToS16_avx2(const uint8_t *pSrc,
+; int iSrcStride,
+; int16_t *pDst,
+; int iHeight);
+;*******************************************************************************
+
+WELS_EXTERN McHorVer20Width4U8ToS16_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_height r3
+%define i_srcstride3 r4
+%define i_dststride 8
+ %assign push_num 0
+%ifdef X86_32
+ push r4
+ %assign push_num 1
+%endif
+ LOAD_4_PARA
+ PUSH_XMM 7
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ sub p_src, i_srcstride
+ sub p_src, i_srcstride
+ lea i_srcstride3, [3 * i_srcstride]
+ vbroadcasti128 ymm4, [shufb_32435465768798A9]
+ vbroadcasti128 ymm5, [shufb_011267784556ABBC]
+ vbroadcasti128 ymm6, [maddubsw_p1m5_p1m5_m5p1_m5p1_128]
+ sub i_height, 3
+.yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm1, [p_src + i_srcstride - 2]
+ vinserti128 ymm0, ymm0, [p_src + 2 * i_srcstride - 2], 1
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride3 - 2], 1
+ lea p_src, [p_src + 4 * i_srcstride]
+ AVX2_FilterHorizontalbw_4x4px ymm0, ymm1, ymm4, ymm5, ymm6, ymm2, ymm3
+ vmovdqa [p_dst], ymm0
+ add p_dst, 4 * i_dststride
+ sub i_height, 4
+ jg .yloop
+ ; Height % 4 remaining single.
+ vmovdqu xmm0, [p_src - 2]
+ AVX2_FilterHorizontalbw_16px xmm0, xmm4, xmm5, xmm6, xmm2, xmm3
+ vmovlps [p_dst], xmm0
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+%ifdef X86_32
+ pop r4
+%endif
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_height
+%undef i_srcstride3
+%undef i_dststride
+
+
+;***********************************************************************
+; void McHorVer02Width4S16ToU8_avx2(const int16_t *pSrc,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iHeight);
+;***********************************************************************
+
+WELS_EXTERN McHorVer02Width4S16ToU8_avx2
+%define p_src r0
+%define p_dst r1
+%define i_dststride r2
+%define i_height r3
+%define i_dststride3 r4
+%define i_srcstride 8
+ %assign push_num 0
+%ifdef X86_32
+ push r4
+ %assign push_num 1
+%endif
+ LOAD_4_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r2, r2d
+ SIGN_EXTENSION r3, r3d
+ lea i_dststride3, [3 * i_dststride]
+ vmovdqu ymm0, [p_src + 0 * i_srcstride]
+ vmovdqu ymm1, [p_src + 1 * i_srcstride]
+ vmovdqu ymm2, [p_src + 2 * i_srcstride]
+ vmovdqu ymm3, [p_src + 3 * i_srcstride]
+ vmovdqu ymm4, [p_src + 4 * i_srcstride]
+ vmovdqu ymm5, [p_src + 5 * i_srcstride]
+ vmovdqu ymm6, [p_src + 6 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm7
+ vpackuswb ymm0, ymm0, ymm0
+ vmovd [p_dst], xmm0
+ vpsrlq xmm7, xmm0, 32
+ vmovd [p_dst + i_dststride], xmm7
+ vextracti128 xmm0, ymm0, 1
+ vmovd [p_dst + 2 * i_dststride], xmm0
+ vpsrlq xmm7, xmm0, 32
+ vmovd [p_dst + i_dststride3], xmm7
+ cmp i_height, 4
+ jle .done
+ lea p_dst, [p_dst + 4 * i_dststride]
+ vmovdqu ymm7, [p_src + 7 * i_srcstride]
+ vmovdqu ymm0, [p_src + 8 * i_srcstride]
+ vmovdqu ymm1, [p_src + 9 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm4, ymm5, ymm6, ymm7, ymm0, ymm1, ymm3
+ vpackuswb ymm4, ymm4, ymm4
+ vmovd [p_dst], xmm4
+ vpsrlq xmm3, xmm4, 32
+ vmovd [p_dst + i_dststride], xmm3
+ vextracti128 xmm4, ymm4, 1
+ vmovd [p_dst + 2 * i_dststride], xmm4
+ vpsrlq xmm3, xmm4, 32
+ vmovd [p_dst + i_dststride3], xmm3
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+%ifdef X86_32
+ pop r4
+%endif
+ ret
+%undef p_src
+%undef p_dst
+%undef i_dststride
+%undef i_height
+%undef i_srcstride
+%undef i_dststride3
+
+
+;*******************************************************************************
+; void McHorVer20Width8U8ToS16_avx2(const uint8_t *pSrc,
+; int iSrcStride,
+; int16_t *pDst,
+; int iHeight);
+;*******************************************************************************
+
+WELS_EXTERN McHorVer20Width8U8ToS16_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_height r3
+%define i_dststride 16
+ %assign push_num 0
+ LOAD_4_PARA
+ PUSH_XMM 6
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ sub p_src, i_srcstride
+ sub p_src, i_srcstride
+ vbroadcasti128 ymm3, [shufb_32435465768798A9]
+ vbroadcasti128 ymm4, [shufb_011267784556ABBC]
+ vbroadcasti128 ymm5, [maddubsw_p1m5_p1m5_m5p1_m5p1_128]
+ sub i_height, 1
+.yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vinserti128 ymm0, ymm0, [p_src + i_srcstride - 2], 1
+ lea p_src, [p_src + 2 * i_srcstride]
+ AVX2_FilterHorizontalbw_16px ymm0, ymm3, ymm4, ymm5, ymm1, ymm2
+ vmovdqu [p_dst], ymm0
+ add p_dst, 2 * i_dststride
+ sub i_height, 2
+ jg .yloop
+ jl .done
+ vmovdqu xmm0, [p_src - 2]
+ AVX2_FilterHorizontalbw_16px xmm0, xmm3, xmm4, xmm5, xmm1, xmm2
+ vmovdqa [p_dst], xmm0
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_height
+%undef i_dststride
+
+
+;***********************************************************************
+; void McHorVer02Width5S16ToU8_avx2(const int16_t *pSrc,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iHeight);
+;***********************************************************************
+
+WELS_EXTERN McHorVer02Width5S16ToU8_avx2
+%define p_src r0
+%define p_dst r1
+%define i_dststride r2
+%define i_height r3
+%define i_srcstride 16
+ %assign push_num 0
+ LOAD_4_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r2, r2d
+ SIGN_EXTENSION r3, r3d
+ vmovdqu ymm0, [p_src + 0 * i_srcstride]
+ vmovdqu ymm2, [p_src + 2 * i_srcstride]
+ vmovdqu ymm4, [p_src + 4 * i_srcstride]
+ vmovdqu ymm6, [p_src + 6 * i_srcstride]
+ vperm2i128 ymm1, ymm0, ymm2, 00100001b
+ vperm2i128 ymm3, ymm2, ymm4, 00100001b
+ vperm2i128 ymm5, ymm4, ymm6, 00100001b
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm7
+ vpackuswb ymm0, ymm0, ymm0
+ vpsrlq xmm7, xmm0, 8
+ vmovd [p_dst + 1], xmm7
+ vmovd [p_dst], xmm0
+ add p_dst, i_dststride
+ vextracti128 xmm0, ymm0, 1
+ vpsrlq xmm7, xmm0, 8
+ vmovd [p_dst + 1], xmm7
+ vmovd [p_dst], xmm0
+ add p_dst, i_dststride
+ vmovdqu ymm7, [p_src + 7 * i_srcstride]
+ vmovdqu ymm0, [p_src + 8 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm1
+ vpackuswb ymm2, ymm2, ymm2
+ vpsrlq xmm1, xmm2, 8
+ vmovd [p_dst + 1], xmm1
+ vmovd [p_dst], xmm2
+ add p_dst, i_dststride
+ vextracti128 xmm2, ymm2, 1
+ vpsrlq xmm1, xmm2, 8
+ vmovd [p_dst + 1], xmm1
+ vmovd [p_dst], xmm2
+ add p_dst, i_dststride
+ vmovdqu ymm1, [p_src + 9 * i_srcstride]
+ vmovdqu ymm2, [p_src + 10 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm4, ymm5, ymm6, ymm7, ymm0, ymm1, ymm3
+ vpackuswb ymm4, ymm4, ymm4
+ vpsrlq xmm3, xmm4, 8
+ vmovd [p_dst + 1], xmm3
+ vmovd [p_dst], xmm4
+ cmp i_height, 5
+ jle .done
+ add p_dst, i_dststride
+ vextracti128 xmm4, ymm4, 1
+ vpsrlq xmm3, xmm4, 8
+ vmovd [p_dst + 1], xmm3
+ vmovd [p_dst], xmm4
+ add p_dst, i_dststride
+ vmovdqu ymm3, [p_src + 11 * i_srcstride]
+ vmovdqu xmm4, [p_src + 12 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm6, ymm7, ymm0, ymm1, ymm2, ymm3, ymm5
+ vpackuswb ymm6, ymm6, ymm6
+ vpsrlq xmm5, xmm6, 8
+ vmovd [p_dst + 1], xmm5
+ vmovd [p_dst], xmm6
+ add p_dst, i_dststride
+ vextracti128 xmm6, ymm6, 1
+ vpsrlq xmm5, xmm6, 8
+ vmovd [p_dst + 1], xmm5
+ vmovd [p_dst], xmm6
+ add p_dst, i_dststride
+ vmovdqu xmm5, [p_src + 13 * i_srcstride]
+ AVX2_FilterVerticalw_16px xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm7
+ vpackuswb xmm0, xmm0, xmm0
+ vpsrlq xmm7, xmm0, 8
+ vmovd [p_dst + 1], xmm7
+ vmovd [p_dst], xmm0
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+ ret
+%undef p_src
+%undef p_dst
+%undef i_dststride
+%undef i_height
+%undef i_srcstride
+
+
+;***********************************************************************
+; void McHorVer02Width8S16ToU8_avx2(const int16_t *pSrc,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iHeight);
+;***********************************************************************
+
+WELS_EXTERN McHorVer02Width8S16ToU8_avx2
+%define p_src r0
+%define p_dst r1
+%define i_dststride r2
+%define i_height r3
+%define i_dststride3 r4
+%define i_srcstride 16
+ %assign push_num 0
+%ifdef X86_32
+ push r4
+ %assign push_num 1
+%endif
+ LOAD_4_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r2, r2d
+ SIGN_EXTENSION r3, r3d
+ lea i_dststride3, [3 * i_dststride]
+ vmovdqa ymm0, [p_src + 0 * i_srcstride]
+ vmovdqa ymm2, [p_src + 2 * i_srcstride]
+ vmovdqa ymm4, [p_src + 4 * i_srcstride]
+ vperm2i128 ymm1, ymm0, ymm2, 00100001b
+ vperm2i128 ymm3, ymm2, ymm4, 00100001b
+.yloop:
+ vmovdqa ymm6, [p_src + 6 * i_srcstride]
+ vperm2i128 ymm5, ymm4, ymm6, 00100001b
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm7
+ vmovdqu ymm7, [p_src + 7 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm1
+ vpackuswb ymm1, ymm0, ymm2
+ vmovdqa ymm0, [p_src + 8 * i_srcstride]
+ vextracti128 xmm2, ymm1, 1
+ vmovlps [p_dst], xmm1
+ vmovlps [p_dst + i_dststride], xmm2
+ vmovhps [p_dst + 2 * i_dststride], xmm1
+ vmovhps [p_dst + i_dststride3], xmm2
+ cmp i_height, 4
+ jle .done
+ lea p_dst, [p_dst + 4 * i_dststride]
+ vmovdqu ymm1, [p_src + 9 * i_srcstride]
+ vmovdqa ymm2, [p_src + 10 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm4, ymm5, ymm6, ymm7, ymm0, ymm1, ymm3
+ vmovdqu ymm3, [p_src + 11 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm6, ymm7, ymm0, ymm1, ymm2, ymm3, ymm5
+ vpackuswb ymm5, ymm4, ymm6
+ vmovdqa ymm4, [p_src + 12 * i_srcstride]
+ add p_src, 8 * i_srcstride
+ vextracti128 xmm6, ymm5, 1
+ vmovlps [p_dst], xmm5
+ vmovlps [p_dst + i_dststride], xmm6
+ vmovhps [p_dst + 2 * i_dststride], xmm5
+ vmovhps [p_dst + i_dststride3], xmm6
+ lea p_dst, [p_dst + 4 * i_dststride]
+ sub i_height, 8
+ jg .yloop
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+%ifdef X86_32
+ pop r4
+%endif
+ ret
+%undef p_src
+%undef p_dst
+%undef i_dststride
+%undef i_height
+%undef i_dststride3
+%undef i_srcstride
+
+
+;*******************************************************************************
+; void McHorVer20Width16U8ToS16_avx2(const uint8_t *pSrc,
+; int32_t iSrcStride,
+; int16_t *pDst,
+; int32_t iHeight);
+;*******************************************************************************
+
+WELS_EXTERN McHorVer20Width16U8ToS16_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_height r3
+%define i_dststride 32
+ %assign push_num 0
+ LOAD_4_PARA
+ PUSH_XMM 7
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ sub p_src, i_srcstride
+ sub p_src, i_srcstride
+ vbroadcasti128 ymm4, [shufb_32435465768798A9]
+ vbroadcasti128 ymm5, [shufb_011267784556ABBC]
+ vbroadcasti128 ymm6, [maddubsw_p1m5_p1m5_m5p1_m5p1_128]
+ sub i_height, 1
+.yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vinserti128 ymm0, ymm0, [p_src + 6], 1
+ vmovdqu xmm1, [p_src + i_srcstride - 2]
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride + 6], 1
+ lea p_src, [p_src + 2 * i_srcstride]
+ AVX2_FilterHorizontalbw_16px ymm0, ymm4, ymm5, ymm6, ymm2, ymm3
+ vmovdqa [p_dst], ymm0
+ AVX2_FilterHorizontalbw_16px ymm1, ymm4, ymm5, ymm6, ymm2, ymm3
+ vmovdqa [p_dst + i_dststride], ymm1
+ add p_dst, 2 * i_dststride
+ sub i_height, 2
+ jg .yloop
+ jl .done
+ vmovdqu xmm0, [p_src - 2]
+ vinserti128 ymm0, ymm0, [p_src + 6], 1
+ AVX2_FilterHorizontalbw_16px ymm0, ymm4, ymm5, ymm6, ymm1, ymm2
+ vmovdqa [p_dst], ymm0
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_height
+%undef i_dststride
+
+
+;***********************************************************************
+; void McHorVer02Width9S16ToU8_avx2(const int16_t *pSrc,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iHeight);
+;***********************************************************************
+
+WELS_EXTERN McHorVer02Width9S16ToU8_avx2
+%define p_src r0
+%define p_dst r1
+%define i_dststride r2
+%define i_height r3
+%define i_srcstride 32
+ %assign push_num 0
+ LOAD_4_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r2, r2d
+ SIGN_EXTENSION r3, r3d
+ vmovdqa ymm0, [p_src + 0 * i_srcstride]
+ vmovdqa ymm1, [p_src + 1 * i_srcstride]
+ vmovdqa ymm2, [p_src + 2 * i_srcstride]
+ vmovdqa ymm3, [p_src + 3 * i_srcstride]
+ vmovdqa ymm4, [p_src + 4 * i_srcstride]
+ sub i_height, 1
+.height_loop:
+ vmovdqa ymm5, [p_src + 5 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6
+ vmovdqa ymm6, [p_src + 6 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
+ vmovdqa ymm7, [p_src + 7 * i_srcstride]
+ vpackuswb ymm0, ymm0, ymm1
+ vextracti128 xmm1, ymm0, 1
+ vpsllq xmm1, xmm1, 56
+ vmovlps [p_dst + 1], xmm1
+ vmovlps [p_dst], xmm0
+ add p_dst, i_dststride
+ vmovhps [p_dst + 1], xmm1
+ vmovhps [p_dst], xmm0
+ add p_dst, i_dststride
+ AVX2_FilterVerticalw_16px ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm0
+ vmovdqa ymm0, [p_src + 8 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm3, ymm4, ymm5, ymm6, ymm7, ymm0, ymm1
+ vpackuswb ymm2, ymm2, ymm3
+ vextracti128 xmm3, ymm2, 1
+ vpsllq xmm3, xmm3, 56
+ vmovlps [p_dst + 1], xmm3
+ vmovlps [p_dst], xmm2
+ add p_dst, i_dststride
+ vmovhps [p_dst + 1], xmm3
+ vmovhps [p_dst], xmm2
+ add p_dst, i_dststride
+ vmovdqa ymm1, [p_src + 9 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm4, ymm5, ymm6, ymm7, ymm0, ymm1, ymm2
+ vmovdqa ymm2, [p_src + 10 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm5, ymm6, ymm7, ymm0, ymm1, ymm2, ymm3
+ vmovdqa ymm3, [p_src + 11 * i_srcstride]
+ vpackuswb ymm4, ymm4, ymm5
+ vextracti128 xmm5, ymm4, 1
+ vpsllq xmm5, xmm5, 56
+ vmovlps [p_dst + 1], xmm5
+ vmovlps [p_dst], xmm4
+ cmp i_height, 4
+ jle .done
+ add p_dst, i_dststride
+ vmovhps [p_dst + 1], xmm5
+ vmovhps [p_dst], xmm4
+ add p_dst, i_dststride
+ AVX2_FilterVerticalw_16px ymm6, ymm7, ymm0, ymm1, ymm2, ymm3, ymm4
+ vmovdqa ymm4, [p_src + 12 * i_srcstride]
+ add p_src, 8 * i_srcstride
+ AVX2_FilterVerticalw_16px ymm7, ymm0, ymm1, ymm2, ymm3, ymm4, ymm5
+ vpackuswb ymm6, ymm6, ymm7
+ vextracti128 xmm7, ymm6, 1
+ vpsllq xmm7, xmm7, 56
+ vmovlps [p_dst + 1], xmm7
+ vmovlps [p_dst], xmm6
+ add p_dst, i_dststride
+ vmovhps [p_dst + 1], xmm7
+ vmovhps [p_dst], xmm6
+ add p_dst, i_dststride
+ sub i_height, 8
+ jg .height_loop
+ vmovdqa ymm5, [p_src + 5 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6
+ vpackuswb ymm0, ymm0, ymm0
+ vextracti128 xmm1, ymm0, 1
+ vpsllq xmm1, xmm1, 56
+ vmovlps [p_dst + 1], xmm1
+ vmovlps [p_dst], xmm0
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_dststride
+%undef i_height
+
+
+;*******************************************************************************
+; void McHorVer20Width17U8ToS16_avx2(const uint8_t *pSrc,
+; int32_t iSrcStride,
+; int16_t *pDst,
+; int32_t iHeight);
+;*******************************************************************************
+
+WELS_EXTERN McHorVer20Width17U8ToS16_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_height r3
+%define i_srcstride3 r4
+%define i_dststride 64
+ %assign push_num 0
+%ifdef X86_32
+ push r4
+ %assign push_num 1
+%endif
+ LOAD_4_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ sub p_src, i_srcstride
+ sub p_src, i_srcstride
+ lea i_srcstride3, [3 * i_srcstride]
+ vbroadcasti128 ymm5, [shufb_32435465768798A9]
+ vbroadcasti128 ymm6, [shufb_011267784556ABBC]
+ vbroadcasti128 ymm7, [maddubsw_p1m5_p1m5_m5p1_m5p1_128]
+ sub i_height, 3
+.yloop:
+ vmovdqu xmm0, [p_src - 2]
+ vmovdqu xmm3, [p_src + 6]
+ vinserti128 ymm0, ymm0, [p_src + i_srcstride - 2], 1
+ vinserti128 ymm3, ymm3, [p_src + i_srcstride + 6], 1
+ vmovdqa ymm4, ymm3
+ AVX2_FilterHorizontalbw_16px ymm0, ymm5, ymm6, ymm7, ymm1, ymm2
+ vmovdqa [p_dst], xmm0
+ vextracti128 [p_dst + i_dststride], ymm0, 1
+ AVX2_FilterHorizontalbw_16px ymm3, ymm5, ymm6, ymm7, ymm1, ymm2
+ vmovdqu xmm1, [p_src + 2 * i_srcstride - 2]
+ vmovdqu xmm0, [p_src + 2 * i_srcstride + 6]
+ vinserti128 ymm1, ymm1, [p_src + i_srcstride3 - 2], 1
+ vinserti128 ymm0, ymm0, [p_src + i_srcstride3 + 6], 1
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpckhqdq ymm4, ymm4, ymm0
+ AVX2_FilterHorizontalbw_4px ymm4, [dwm32768_256], ymm2
+ vmovlps [p_dst + 26], xmm4
+ vmovdqa [p_dst + 16], xmm3
+ vextracti128 xmm2, ymm4, 1
+ vmovlps [p_dst + i_dststride + 26], xmm2
+ vextracti128 [p_dst + i_dststride + 16], ymm3, 1
+ vmovhps [p_dst + 2 * i_dststride + 26], xmm4
+ AVX2_FilterHorizontalbw_16px ymm1, ymm5, ymm6, ymm7, ymm3, ymm4
+ vmovdqa [p_dst + 2 * i_dststride], xmm1
+ AVX2_FilterHorizontalbw_16px ymm0, ymm5, ymm6, ymm7, ymm3, ymm4
+ vmovdqa [p_dst + 2 * i_dststride + 16], xmm0
+ vextracti128 [p_dst + 3 * i_dststride], ymm1, 1
+ vmovhps [p_dst + 3 * i_dststride + 26], xmm2
+ vextracti128 [p_dst + 3 * i_dststride + 16], ymm0, 1
+ add p_dst, 4 * i_dststride
+ sub i_height, 4
+ jg .yloop
+ ; Handle remaining 2 lines after 4x unrolled loop.
+ vmovdqu xmm0, [p_src - 2]
+ vinserti128 ymm0, ymm0, [p_src + 6], 1
+ vmovdqu xmm3, [p_src + i_srcstride - 2]
+ vinserti128 ymm3, ymm3, [p_src + i_srcstride + 6], 1
+ vpunpckhqdq ymm4, ymm0, ymm3
+ AVX2_FilterHorizontalbw_4px ymm4, [dwm32768_256], ymm2
+ AVX2_FilterHorizontalbw_16px ymm0, ymm5, ymm6, ymm7, ymm1, ymm2
+ AVX2_FilterHorizontalbw_16px ymm3, ymm5, ymm6, ymm7, ymm1, ymm2
+ vextracti128 xmm4, ymm4, 1
+ vmovlps [p_dst + 26], xmm4
+ vmovdqa [p_dst], ymm0
+ vmovhps [p_dst + i_dststride + 26], xmm4
+ vmovdqa [p_dst + i_dststride], ymm3
+ vzeroupper
+ POP_XMM
+ LOAD_4_PARA_POP
+%ifdef X86_32
+ pop r4
+%endif
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_dststride
+%undef i_height
+%undef i_srcstride3
+
+
+;***********************************************************************
+; void McHorVer02Width16Or17S16ToU8_avx2(const int16_t *pSrc,
+; int32_t iSrcStride,
+; uint8_t *pDst,
+; int32_t iDstStride,
+; int32_t iWidth,
+; int32_t iHeight);
+;***********************************************************************
+
+WELS_EXTERN McHorVer02Width16Or17S16ToU8_avx2
+%define p_src r0
+%define i_srcstride r1
+%define p_dst r2
+%define i_dststride r3
+%define i_width r4
+%define i_height r5
+%define i_srcstride3 r6
+ %assign push_num 0
+%ifdef X86_32
+ push r6
+ %assign push_num 1
+%endif
+ LOAD_6_PARA
+ PUSH_XMM 8
+ SIGN_EXTENSION r1, r1d
+ SIGN_EXTENSION r3, r3d
+ SIGN_EXTENSION r4, r4d
+ SIGN_EXTENSION r5, r5d
+ sub i_height, 1
+ lea i_srcstride3, [3 * i_srcstride]
+ test i_width, 1
+ jz .align_begin
+ push i_height
+ push p_src
+ push p_dst
+ lea p_src, [p_src + 2 * i_width - 2]
+ add p_dst, i_width
+ vmovd xmm0, [p_src]
+ vpunpcklwd xmm0, xmm0, [p_src + i_srcstride]
+ vmovd xmm1, [p_src + 2 * i_srcstride]
+ add p_src, i_srcstride3
+ vpunpcklwd xmm1, xmm1, [p_src]
+ vpunpckldq xmm0, xmm0, xmm1
+ vmovd xmm1, [p_src + i_srcstride]
+ vpunpcklwd xmm1, xmm1, [p_src + 2 * i_srcstride]
+ vmovd xmm2, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpcklwd xmm2, xmm2, [p_src]
+ vpunpckldq xmm1, xmm1, xmm2
+ vpunpcklqdq xmm0, xmm0, xmm1
+.height_loop_unalign:
+ vmovd xmm1, [p_src + i_srcstride]
+ vpalignr xmm1, xmm1, xmm0, 2
+ vmovd xmm2, [p_src + 2 * i_srcstride]
+ vpalignr xmm2, xmm2, xmm1, 2
+ vmovd xmm3, [p_src + i_srcstride3]
+ vpalignr xmm3, xmm3, xmm2, 2
+ lea p_src, [p_src + 4 * i_srcstride]
+ vmovd xmm4, [p_src]
+ vpalignr xmm4, xmm4, xmm3, 2
+ vmovd xmm5, [p_src + i_srcstride]
+ vpalignr xmm5, xmm5, xmm4, 2
+ AVX2_FilterVerticalw_16px xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm7
+ vpackuswb xmm0, xmm0, xmm0
+ vpslld xmm6, xmm0, 24
+ vmovd [p_dst - 4], xmm6
+ vmovlps [p_dst + 4 * i_dststride - 8], xmm6
+ add p_dst, i_dststride
+ vpslld xmm6, xmm0, 16
+ vmovd [p_dst - 4], xmm6
+ vmovlps [p_dst + 4 * i_dststride - 8], xmm6
+ add p_dst, i_dststride
+ vpslld xmm6, xmm0, 8
+ vmovd [p_dst - 4], xmm6
+ vmovd [p_dst + i_dststride - 4], xmm0
+ lea p_dst, [p_dst + 4 * i_dststride]
+ vmovlps [p_dst - 8], xmm6
+ vmovlps [p_dst + i_dststride - 8], xmm0
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 8
+ jle .height_loop_unalign_exit
+ vmovd xmm1, [p_src + 2 * i_srcstride]
+ vpalignr xmm1, xmm1, xmm5, 2
+ vmovd xmm0, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpunpcklwd xmm0, xmm0, [p_src]
+ vpalignr xmm0, xmm0, xmm1, 4
+ jmp .height_loop_unalign
+.height_loop_unalign_exit:
+ vpbroadcastq xmm6, [p_src + 2 * i_srcstride - 6]
+ AVX2_FilterVerticalw_16px xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
+ vpackuswb xmm1, xmm1, xmm1
+ vmovlps [p_dst - 8], xmm1
+ pop p_dst
+ pop p_src
+ pop i_height
+.align_begin:
+ vmovdqa ymm0, [p_src]
+ vmovdqa ymm1, [p_src + i_srcstride]
+ vmovdqa ymm2, [p_src + 2 * i_srcstride]
+ vmovdqa ymm3, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vmovdqa ymm4, [p_src]
+.height_loop:
+ vmovdqa ymm5, [p_src + i_srcstride]
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6
+ vmovdqa ymm6, [p_src + 2 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
+ vmovdqa ymm7, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpackuswb ymm0, ymm0, ymm1
+ vpermq ymm0, ymm0, 11011000b
+ vmovdqa [p_dst], xmm0
+ vextracti128 [p_dst + i_dststride], ymm0, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ AVX2_FilterVerticalw_16px ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm0
+ vmovdqa ymm0, [p_src]
+ AVX2_FilterVerticalw_16px ymm3, ymm4, ymm5, ymm6, ymm7, ymm0, ymm1
+ vpackuswb ymm2, ymm2, ymm3
+ vpermq ymm2, ymm2, 11011000b
+ vmovdqa [p_dst], xmm2
+ vextracti128 [p_dst + i_dststride], ymm2, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ vmovdqa ymm1, [p_src + i_srcstride]
+ AVX2_FilterVerticalw_16px ymm4, ymm5, ymm6, ymm7, ymm0, ymm1, ymm2
+ vmovdqa ymm2, [p_src + 2 * i_srcstride]
+ AVX2_FilterVerticalw_16px ymm5, ymm6, ymm7, ymm0, ymm1, ymm2, ymm3
+ vmovdqa ymm3, [p_src + i_srcstride3]
+ lea p_src, [p_src + 4 * i_srcstride]
+ vpackuswb ymm4, ymm4, ymm5
+ vpermq ymm4, ymm4, 11011000b
+ vmovdqa [p_dst], xmm4
+ vextracti128 [p_dst + i_dststride], ymm4, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ AVX2_FilterVerticalw_16px ymm6, ymm7, ymm0, ymm1, ymm2, ymm3, ymm4
+ vmovdqa ymm4, [p_src]
+ AVX2_FilterVerticalw_16px ymm7, ymm0, ymm1, ymm2, ymm3, ymm4, ymm5
+ vpackuswb ymm6, ymm6, ymm7
+ vpermq ymm6, ymm6, 11011000b
+ vmovdqa [p_dst], xmm6
+ vextracti128 [p_dst + i_dststride], ymm6, 1
+ lea p_dst, [p_dst + 2 * i_dststride]
+ sub i_height, 8
+ jg .height_loop
+ jl .done
+ vmovdqa ymm5, [p_src + i_srcstride]
+ AVX2_FilterVerticalw_16px ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6
+ vpackuswb ymm0, ymm0, ymm0
+ vpermq ymm0, ymm0, 11011000b
+ vmovdqa [p_dst], xmm0
+.done:
+ vzeroupper
+ POP_XMM
+ LOAD_6_PARA_POP
+%ifdef X86_32
+ pop r6
+%endif
+ ret
+%undef p_src
+%undef i_srcstride
+%undef p_dst
+%undef i_dststride
+%undef i_width
+%undef i_height
+%undef i_srcstride3
+
+%endif ; HAVE_AVX2
--- a/test/encoder/EncUT_MotionCompensation.cpp
+++ b/test/encoder/EncUT_MotionCompensation.cpp
@@ -222,7 +222,10 @@
#ifdef X86_ASM
DEF_LUMA_MCTESTS(WELS_CPU_SSE2, sse2)
DEF_LUMA_MCTESTS(WELS_CPU_SSE2 | WELS_CPU_SSSE3, ssse3)
+#ifdef HAVE_AVX2
+DEF_LUMA_MCTESTS(WELS_CPU_SSE2 | WELS_CPU_SSSE3 | WELS_CPU_AVX2, avx2)
#endif
+#endif
#define DEF_CHROMA_MCTEST(iW,iH) \
TEST(McChroma,iW##x##iH) \
@@ -392,4 +395,7 @@
#ifdef X86_ASM
DEF_HALFPEL_MCTESTS(WELS_CPU_SSE2, sse2)
DEF_HALFPEL_MCTESTS(WELS_CPU_SSE2 | WELS_CPU_SSSE3, ssse3)
+#ifdef HAVE_AVX2
+DEF_HALFPEL_MCTESTS(WELS_CPU_SSE2 | WELS_CPU_SSSE3 | WELS_CPU_AVX2, avx2)
+#endif
#endif