Assembly x64 Intro - Dct.asm of OpenH264 Encode
2015-12-10 15:06
543 查看
/****************************************************************************
* DCT functions
****************************************************************************/
void WelsDctT4_c (int16_t* pDct, uint8_t* pPixel1, int32_t iStride1, uint8_t* pPixel2, int32_t iStride2) {
int16_t i, pData[16], s[4];
for (i = 0 ; i < 16 ; i += 4) {
const int32_t kiI1 = 1 + i;
const int32_t kiI2 = 2 + i;
const int32_t kiI3 = 3 + i;
pData[i ] = pPixel1[0] - pPixel2[0];
pData[kiI1] = pPixel1[1] - pPixel2[1];
pData[kiI2] = pPixel1[2] - pPixel2[2];
pData[kiI3] = pPixel1[3] - pPixel2[3];
pPixel1 += iStride1;
pPixel2 += iStride2;
/*horizontal transform */ /* 行变形， 计算Y = X * T(Cf4) = T(Cf4 * T(X)) */
s[0] = pData[i] + pData[kiI3];
s[3] = pData[i] - pData[kiI3];
s[1] = pData[kiI1] + pData[kiI2];
s[2] = pData[kiI1] - pData[kiI2];
pDct[i ] = s[0] + s[1];
pDct[kiI2] = s[0] - s[1];
pDct[kiI1] = (s[3] << 1) + s[2];
pDct[kiI3] = s[3] - (s[2] << 1);
}
/* vertical transform */ /* 列变形， 计算 Cf4 * Y */
for (i = 0 ; i < 4 ; i ++) {
const int32_t kiI4 = 4 + i;
const int32_t kiI8 = 8 + i;
const int32_t kiI12 = 12 + i;
s[0] = pDct[i ] + pDct[kiI12];
s[3] = pDct[i ] - pDct[kiI12];
s[1] = pDct[kiI4] + pDct[kiI8 ];
s[2] = pDct[kiI4] - pDct[kiI8 ];
pDct[i ] = s[0] + s[1];
pDct[kiI8 ] = s[0] - s[1];
pDct[kiI4 ] = (s[3] << 1) + s[2];
pDct[kiI12] = s[3] - (s[2] << 1);
}
}
;*!
;* \copy
;* Copyright (c) 2009-2013, Cisco Systems
;* All rights reserved.
;*
;* Redistribution and use in source and binary forms, with or without
;* modification, are permitted 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.
;*
;* 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.
;*
;*
;* dct.asm
;*
;* Abstract
;* WelsDctFourT4_sse2
;*
;* History
;* 8/4/2009 Created
;*
;*
;*************************************************************************/
%include "asm_inc.asm"
SECTION .rodata align=16
;***********************************************************************
; Constant
;***********************************************************************
align 16
SSE2_DeQuant8 dw 10, 13, 10, 13, 13, 16, 13, 16,
dw 10, 13, 10, 13, 13, 16, 13, 16,
dw 11, 14, 11, 14, 14, 18, 14, 18,
dw 11, 14, 11, 14, 14, 18, 14, 18,
dw 13, 16, 13, 16, 16, 20, 16, 20,
dw 13, 16, 13, 16, 16, 20, 16, 20,
dw 14, 18, 14, 18, 18, 23, 18, 23,
dw 14, 18, 14, 18, 18, 23, 18, 23,
dw 16, 20, 16, 20, 20, 25, 20, 25,
dw 16, 20, 16, 20, 20, 25, 20, 25,
dw 18, 23, 18, 23, 23, 29, 23, 29,
dw 18, 23, 18, 23, 23, 29, 23, 29
;***********************************************************************
; MMX functions
;***********************************************************************
%macro MMX_LoadDiff4P 5
movd %1, [%3]
movd %2, [%4]
punpcklbw %1, %5
punpcklbw %2, %5
psubw %1, %2
%endmacro
%macro MMX_LoadDiff4x4P 10 ;d0, d1, d2, d3, pix1address, pix1stride, pix2address, pix2stride, tmp(mm), 0(mm)
MMX_LoadDiff4P %1, %9, %5, %7, %10
MMX_LoadDiff4P %2, %9, %5+%6, %7+%8, %10
lea %5, [%5+2*%6]
lea %7, [%7+2*%8]
MMX_LoadDiff4P %3, %9, %5, %7, %10
MMX_LoadDiff4P %4, %9, %5+%6, %7+%8, %10
%endmacro
%macro MMX_SumSubMul2 3
movq %3, %1
psllw %1, $01
paddw %1, %2
psllw %2, $01
psubw %3, %2
%endmacro
%macro MMX_SumSubDiv2 3
movq %3, %2
psraw %3, $01
paddw %3, %1
psraw %1, $01
psubw %1, %2
%endmacro
%macro MMX_SumSub 3
movq %3, %2
psubw %2, %1
paddw %1, %3
%endmacro
%macro MMX_DCT 6
MMX_SumSub %4, %1, %6
MMX_SumSub %3, %2, %6
MMX_SumSub %3, %4, %6
MMX_SumSubMul2 %1, %2, %5
%endmacro
%macro MMX_IDCT 6
MMX_SumSub %4, %5, %6
MMX_SumSubDiv2 %3, %2, %1
MMX_SumSub %1, %4, %6
MMX_SumSub %3, %5, %6
%endmacro
%macro MMX_StoreDiff4P 6
movd %2, %6
punpcklbw %2, %4
paddw %1, %3
psraw %1, $06
paddsw %1, %2
packuswb %1, %2
movd %5, %1
%endmacro
SECTION .text
;***********************************************************************
; void WelsDctT4_mmx( int16_t *pDct[4], uint8_t *pix1, int32_t i_pix1, uint8_t *pix2, int32_t i_pix2 )
;***********************************************************************
WELS_EXTERN WelsDctT4_mmx
%assign push_num 0
LOAD_5_PARA
SIGN_EXTENSION r2, r2d
SIGN_EXTENSION r4, r4d
WELS_Zero mm7
； 下面的汇编代码将使 mm1 保存的是 4x4 残差矩阵的第一行 ； mm2 保存的是 4x4 残差矩阵的第二行
; mm3 保存的是 4x4 残差矩阵的第三行； mm4 保存的是 4x4 残差矩阵的第四行
MMX_LoadDiff4x4P mm1, mm2, mm3, mm4, r1, r2, r3, r4, mm0, mm7
； 先执行列变换， 即 Y = Cf4 * X, mm1 denote first component of column vector, mm2 denote second component
; mm3 third component, mm4 4th component.
MMX_DCT mm1, mm2, mm3 ,mm4, mm5, mm6
； after transforming, mm3, mm1, mm4, mm5 denote first, second, third, 4th component of result vector
； mm3 = a3a2a1a0, mm1 = b3b2b1b0, mm4 = c3c2c1c0, mm5 = d3d2d1d0
; expand =>
; 1.) movq mm2, mm3 ; mm2 = mm3 = a3a2a1a0
; punpckhwd mm2, mm1 ; mm2 = b3a3b2a2
; punpcklwd mm3, mm1 ; mm3 = b1a1b0a0
;
; 2.) movq mm1, mm4 ; mm1 = mm4 = c3c2c1c0
; punpckhwd mm1, mm5 ; mm1 = d3c3d2c2
; punpcklwd mm4, mm5 ; mm4 = d1c1d0c0
;
; 3.) movq mm5, mm3 ; mm5 = mm3 = b1a1b0a0
; punpckhdq mm5, mm4 ; mm5 = d1c1b1a1
; punpckldq mm3, mm4 ; mm3 = d0c0b0a0
;
; 4.) movq mm4, mm2 ; mm4 = mm2 = b3a3b2a2
; punpckhdq mm4, mm1 ; mm4 = d3c3b3a3
; punpckldq mm2, mm1 ; mm2 = d2c2b2a2
MMX_Trans4x4W mm3, mm1, mm4, mm5, mm2
; 因此上面的转置后， mm3 = d0c0b0a0, mm5 = d1c1b1a1, mm2 = d2c2b2a2, mm4 = d3c3b3a3
; 这儿执行行变形 《=》 Z = Cf4 * T（Y）
MMX_DCT mm3, mm5, mm2 ,mm4, mm1, mm6
; 再次转置 T(Z) = Y * T(Cf4)
MMX_Trans4x4W mm2, mm3, mm4, mm1, mm5
movq [r0+ 0], mm2
movq [r0+ 8], mm1
movq [r0+16], mm5
movq [r0+24], mm4
WELSEMMS
LOAD_5_PARA_POP
ret
/****************************************************************************
* IDCT functions, final output = prediction(CS) + IDCT(scaled_coeff)
****************************************************************************/
void WelsIDctT4Rec_c (uint8_t* pRec, int32_t iStride, uint8_t* pPred, int32_t iPredStride, int16_t* pDct) {
int32_t i;
int16_t iTemp[16];
int32_t iDstStridex2 = iStride << 1;
int32_t iDstStridex3 = iStride + iDstStridex2;
int32_t iPredStridex2 = iPredStride << 1;
int32_t iPredStridex3 = iPredStride + iPredStridex2;
for (i = 0; i < 4; i ++) { // horizon; Row Transform, Y = Z * Ci4
int32_t iIdx = i << 2;
; kiHorSumU = a + c
const int32_t kiHorSumU = pDct[iIdx] + pDct[iIdx + 2]; // add 0-2
; kiHorDelU = a - c
const int32_t kiHorDelU = pDct[iIdx] - pDct[iIdx + 2]; // sub 0-2
; kiHorSumD = b + d / 2
const int32_t kiHorSumD = pDct[iIdx + 1] + (pDct[iIdx + 3] >> 1);
; kiHorDelD = b / 2 - d
const int32_t kiHorDelD = (pDct[iIdx + 1] >> 1) - pDct[iIdx + 3];
; iTemp[ildx] = a + b + c + d / 2
iTemp[iIdx ] = kiHorSumU + kiHorSumD;
; iTemp[ildx + 1] = a + b / 2 - c - d
iTemp[iIdx + 1] = kiHorDelU + kiHorDelD;
; iTemp[ildx + 2] = a - b / 2 - c + d
iTemp[iIdx + 2] = kiHorDelU - kiHorDelD;
; iTemp[ildx + 2] = a - b + c - d / 2
iTemp[iIdx + 3] = kiHorSumU - kiHorSumD;
}
for (i = 0; i < 4; i ++) { // vertical; Column Transform, X = T(Ci4) * Y
; kiVerSumL = a + c
const int32_t kiVerSumL = iTemp[i] + iTemp[8 + i];
; kiVerDelL = a - c
const int32_t kiVerDelL = iTemp[i] - iTemp[8 + i];
; kiVerDelR = b / 2 - d
const int32_t kiVerDelR = (iTemp[4 + i] >> 1) - iTemp[12 + i];
; kiVerSumR = b + d / 2
const int32_t kiVerSumR = iTemp[4 + i] + (iTemp[12 + i] >> 1);
; kiVerSumL + kiVerSumR = a + b + c + d / 2
pRec[i ] = WelsClip1 (pPred[i ] + ((kiVerSumL + kiVerSumR + 32) >> 6));
; kiVerDelL + kiVerDelR = a + b / 2 - c - d
pRec[iStride + i ] = WelsClip1 (pPred[iPredStride + i ] + ((kiVerDelL + kiVerDelR + 32) >> 6));
; kiVerDelL - kiVerDelR = a - b / 2 - c + d
pRec[iDstStridex2 + i] = WelsClip1 (pPred[iPredStridex2 + i] + ((kiVerDelL - kiVerDelR + 32) >> 6));
; kiVerSumL - kiVerSumR = a - b + c - d / 2
pRec[iDstStridex3 + i] = WelsClip1 (pPred[iPredStridex3 + i] + ((kiVerSumL - kiVerSumR + 32) >> 6));
}
}
;***********************************************************************
; void WelsIDctT4Rec_mmx(uint8_t *rec, int32_t stride, uint8_t *pred, int32_t pred_stride, int16_t *rs)
;***********************************************************************
WELS_EXTERN WelsIDctT4Rec_mmx
%assign push_num 0
LOAD_5_PARA
SIGN_EXTENSION r1, r1d
SIGN_EXTENSION r3, r3d
movq mm0, [r4+ 0]
movq mm1, [r4+ 8]
movq mm2, [r4+16]
movq mm3, [r4+24]
; mm0 为 4x4 矩阵的第一行， mm1 为 第二行；以此类推
; 转置后为 mm0 = a, mm3 = b, mm4 = c, mm2 = d
; 即输出结果为 a, 空， d, b, c
MMX_Trans4x4W mm0, mm1, mm2, mm3, mm4
; MMX_IDC mm1, d, b, c, a, mm6; 行变换，Y = Z * Ci4
; 结果为mm1 = a + b + c + d / 2, mm3 = a + b / 2 - c - d
; mm0 = a - b / 2 - c + d, mm4 = a - b + c - d / 2
MMX_IDCT mm1, mm2, mm3, mm4, mm0, mm6
; mm1 = a, mm3 = b, mm0 = c, mm4 = d => a, 空， d, b, c
; 即 mm1 = a, mm3, mm0 = d, mm4 = b, mm2 = c
MMX_Trans4x4W mm1, mm3, mm0, mm4, mm2
; MMX_IDCT mm3, mm0 = d, mm4 = b, mm2 = c, mm1 = a, mm6
; X = T(Ci4) * W; 列变换， W 为列向量
; 结果为 mm3, mm4, mm1, mm2
MMX_IDCT mm3, mm0, mm4, mm2, mm1, mm6
; mm7 = 0
WELS_Zero mm7
; mm6 = 32 32 32 32
WELS_DW32 mm6
; calc pRec[i ] = WelsClip1 (pPred[i ] + ((kiVerSumL + kiVerSumR + 32) >> 6)); one whole row
; %macro MMX_StoreDiff4P 6
; movd %2, %6
; punpcklbw %2, %4
; paddw %1, %3
; psraw %1, $06
; paddsw %1, %2
; packuswb %1, %2
; movd %5, %1
; %endmacro
; MMX_StoreDiff4P mm3, mm0, mm6, mm7, [r0], [r2] =>
; movd mm0, [r2]
; punpcklbw mm0, mm7 ; mm0, mm7 = 0 交叉组合 => mm0
; paddw mm3, mm6 ; mm6 = 32 32 32 32; mm3 += 32;
; psraw mm3, 6 ; mm3 = mm3 >> 6
; paddsw mm3, mm0 ; mm3 += mm0; mm0 即 [r2]
; packuswb mm3, mm0 ; 无符号饱和压缩 => mm3 (mm3为什么和mm0作无符号饱和压缩 ？？？)
; 哦，想明白了， mm0 在上面的饱和压缩运算中， 相当于一个哑元而已
; movd [r0], mm3
MMX_StoreDiff4P mm3, mm0, mm6, mm7, [r0], [r2]
; calc pRec[iStride + i ] = WelsClip1 (pPred[iPredStride + i ] + ((kiVerDelL + kiVerDelR + 32) >> 6)); one whole row
MMX_StoreDiff4P mm4, mm0, mm6, mm7, [r0+r1], [r2+r3]
lea r0, [r0+2*r1]
lea r2, [r2+2*r3]
; calc pRec[iDstStridex2 + i] = WelsClip1 (pPred[iPredStridex2 + i] + ((kiVerDelL - kiVerDelR + 32) >> 6)); one whole row
MMX_StoreDiff4P mm1, mm0, mm6, mm7, [r0], [r2]
; calc pRec[iDstStridex3 + i] = WelsClip1 (pPred[iPredStridex3 + i] + ((kiVerSumL - kiVerSumR + 32) >> 6)); one whole row
MMX_StoreDiff4P mm2, mm0, mm6, mm7, [r0+r1], [r2+r3]
WELSEMMS
LOAD_5_PARA_POP
ret
;for TRANSPOSE
%macro SSE2_XSawp 4
movdqa %4, %2 ; 移动16字节对齐的双四字
punpckl%1 %2, %3 ; 如 punpcklqdq，交叉组合低位四字
punpckh%1 %4, %3 ; 如 punpckhqdq, 交叉组合高位四字
%endmacro
;***********************************************************************
; SSE2 functions
;***********************************************************************
%macro SSE2_Store4x8p 6
SSE2_XSawp qdq, %2, %3, %6
SSE2_XSawp qdq, %4, %5, %3
MOVDQ [%1+0x00], %2
MOVDQ [%1+0x10], %4
MOVDQ [%1+0x20], %6
MOVDQ [%1+0x30], %3
%endmacro
%macro SSE2_Load4x8p 6
MOVDQ %2, [%1+0x00]
MOVDQ %4, [%1+0x10]
MOVDQ %6, [%1+0x20]
MOVDQ %3, [%1+0x30]
SSE2_XSawp qdq, %4, %3, %5
SSE2_XSawp qdq, %2, %6, %3
%endmacro
%macro SSE2_SumSubMul2 3
movdqa %3, %1
paddw %1, %1
paddw %1, %2
psubw %3, %2
psubw %3, %2
%endmacro
%macro SSE2_SumSubDiv2 4
movdqa %4, %1
movdqa %3, %2
psraw %2, $01
psraw %4, $01
paddw %1, %2
psubw %4, %3
%endmacro
%macro SSE2_StoreDiff8p 6
paddw %1, %3
psraw %1, $06
movq %2, %6
punpcklbw %2, %4
paddsw %2, %1
packuswb %2, %2
movq %5, %2
%endmacro
%macro SSE2_StoreDiff8p 5
movq %2, %5
punpcklbw %2, %3
paddsw %2, %1
packuswb %2, %2
movq %4, %2
%endmacro
%macro SSE2_Load8DC 6
movdqa %1, %6 ; %1 = dc0 dc1
paddw %1, %5
psraw %1, $06 ; (dc + 32) >> 6
movdqa %2, %1
psrldq %2, 4
punpcklwd %2, %2
punpckldq %2, %2 ; %2 = dc2 dc2 dc2 dc2 dc3 dc3 dc3 dc3
movdqa %3, %1
psrldq %3, 8
punpcklwd %3, %3
punpckldq %3, %3 ; %3 = dc4 dc4 dc4 dc4 dc5 dc5 dc5 dc5
movdqa %4, %1
psrldq %4, 12
punpcklwd %4, %4
punpckldq %4, %4 ; %4 = dc6 dc6 dc6 dc6 dc7 dc7 dc7 dc7
punpcklwd %1, %1
punpckldq %1, %1 ; %1 = dc0 dc0 dc0 dc0 dc1 dc1 dc1 dc1
%endmacro
%macro SSE2_DCT 6
SSE2_SumSub %6, %3, %5
SSE2_SumSub %1, %2, %5
SSE2_SumSub %3, %2, %5
SSE2_SumSubMul2 %6, %1, %4
%endmacro
%macro SSE2_IDCT 7
SSE2_SumSub %7, %2, %6
SSE2_SumSubDiv2 %1, %3, %5, %4
SSE2_SumSub %2, %1, %5
SSE2_SumSub %7, %4, %5
%endmacro
;***********************************************************************
; void WelsDctFourT4_sse2(int16_t *pDct, uint8_t *pix1, int32_t i_pix1, uint8_t *pix2, int32_t i_pix2 )
;***********************************************************************
WELS_EXTERN WelsDctFourT4_sse2
%assign push_num 0
LOAD_5_PARA
PUSH_XMM 8
SIGN_EXTENSION r2, r2d
SIGN_EXTENSION r4, r4d
pxor xmm7, xmm7
;Load 4x8
; ;xmm0, xmm6, xmm7, [eax], [ecx]
; ;xmm7 = 0, eax = pix1, ecx = pix2, xmm0 save the result
; %macro SSE2_LoadDiff8P 5
; movq %1, %4
; punpcklbw %1, %3
; movq %2, %5
; punpcklbw %2, %3
; psubw %1, %2
; %endmacro
; SSE2_LoadDiff8P xmm0, xmm6, xmm7, [r1], [r3] =>
; movq xmm0, [r1]
; punpcklbw xmm0, xmm7 ;交叉组合xmm0与xmm7低位四字中的字节,将结果放入xmm0
; movq xmm6, [r3]
; punpcklbw xmm6, xmm7 ;交叉组合xmm6与xmm7低位四字中的字节,将结果放入xmm6
; psubw xmm0, xmm6 ; xmm0 -= mm6;
; i.e xmm0 = [r1] - [r3]; 8个 16bit 相减
SSE2_LoadDiff8P xmm0, xmm6, xmm7, [r1], [r3]
; 同样地， xmml = [r1 + r2] - [r3 + r4]
SSE2_LoadDiff8P xmm1, xmm6, xmm7, [r1+r2], [r3+r4]
lea r1, [r1 + 2 * r2]
lea r3, [r3 + 2 * r4]
; 同样地， xmm2 = [r1 + 2r2] - [r3 + 2r4]
SSE2_LoadDiff8P xmm2, xmm6, xmm7, [r1], [r3]
; 同样地， xmm3 = [r1 + 3r2] - [r3 + 3r4]
SSE2_LoadDiff8P xmm3, xmm6, xmm7, [r1+r2], [r3+r4]
; %macro SSE2_DCT 6
; SSE2_SumSub %6, %3, %5
; SSE2_SumSub %1, %2, %5
; SSE2_SumSub %3, %2, %5
; SSE2_SumSubMul2 %6, %1, %4
; %endmacro
; m2 = m1 + m2, m1 = m1 - m2
; %macro SSE2_SumSub 3
; movdqa %3, %2
; paddw %2, %1
; psubw %1, %3
; %endmacro
; %macro SSE2_SumSubMul2 3
; movdqa %3, %1
; paddw %1, %1
; paddw %1, %2
; psubw %3, %2
; psubw %3, %2
; %endmacro
; Y = Cf4 * X, 列变换； result = xmm2, xmm0, xmm3, xmm4
; SSE2_DCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm0 =>
; SSE2_SumSub xmm0, xmm3, xmm5 =>
;;;; movdqa xmm5, xmm3 => xmm5 = xmm3 = d
;;;; paddw xmm3, xmm0 => xmm3 = xmm3 + xmm0 = a + d
;;;; psubw xmm0, xmm5 => xmm0 = xmm0 - xmm5 = a - d
; SSE2_SumSub xmm1, xmm2, xmm5
;;;; movdqa xmm5, xmm2 => xmm5 = xmm2 = c
;;;; paddw xmm2, xmm1 => xmm2 = xmm2 + xmm1 = b + c
;;;; psubw xmm1, xmm5 => xmm1 = xmm1 - xmm5 = b - c
; SSE2_SumSub xmm3, xmm2, xmm5
;;;; movdqa xmm5, xmm2 => xmm5 = xmm2 = b + c
;;;; paddw xmm2, xmm3 => xmm2 = xmm2 + xmm3 = a + b + c + d
;;;; psubw xmm3, xmm5 => xmm3 = xmm3 - xmm5 = a - b - c + d
; SSE2_SumSubMul2 xmm0, xmm1, xmm4
;;;; movdqa xmm4, xmm0 => xmm4 = xmm0 = a - d
;;;; paddw xmm0, xmm0 => xmm0 = 2(a - d)
;;;; paddw xmm0, xmm1 => xmm0 = xmm0 + xmm1 = 2a + b -c - 2d
;;;; psubw xmm4, xmm1 => xmm4 = xmm4 - xmm1 = a - b + c - d
;;;; psubw xmm4, xmm1 => xmm4 = xmm4 - xmm1 = a - 2b + 2c - d
; result => xmm2 = 1st, xmm0 = 2nd, xmm3 = 3rd, xmm4 = 4th
SSE2_DCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm0
SSE2_TransTwo4x4W xmm2, xmm0, xmm3, xmm4, xmm1
SSE2_DCT xmm0, xmm4, xmm1, xmm3, xmm5, xmm2
SSE2_TransTwo4x4W xmm4, xmm2, xmm1, xmm3, xmm0
SSE2_Store4x8p r0, xmm4, xmm2, xmm3, xmm0, xmm5
lea r1, [r1 + 2 * r2]
lea r3, [r3 + 2 * r4]
;Load 4x8
SSE2_LoadDiff8P xmm0, xmm6, xmm7, [r1 ], [r3 ]
SSE2_LoadDiff8P xmm1, xmm6, xmm7, [r1+r2 ], [r3+r4]
lea r1, [r1 + 2 * r2]
lea r3, [r3 + 2 * r4]
SSE2_LoadDiff8P xmm2, xmm6, xmm7, [r1], [r3]
SSE2_LoadDiff8P xmm3, xmm6, xmm7, [r1+r2], [r3+r4]
SSE2_DCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm0
SSE2_TransTwo4x4W xmm2, xmm0, xmm3, xmm4, xmm1
SSE2_DCT xmm0, xmm4, xmm1, xmm3, xmm5, xmm2
SSE2_TransTwo4x4W xmm4, xmm2, xmm1, xmm3, xmm0
lea r0, [r0+64]
SSE2_Store4x8p r0, xmm4, xmm2, xmm3, xmm0, xmm5
POP_XMM
LOAD_5_PARA_POP
ret
;***********************************************************************
; void WelsIDctFourT4Rec_sse2(uint8_t *rec, int32_t stride, uint8_t *pred, int32_t pred_stride, int16_t *rs);
;***********************************************************************
WELS_EXTERN WelsIDctFourT4Rec_sse2
%assign push_num 0
LOAD_5_PARA
PUSH_XMM 8
SIGN_EXTENSION r1, r1d
SIGN_EXTENSION r3, r3d
;Load 4x8
SSE2_Load4x8p r4, xmm0, xmm1, xmm4, xmm2, xmm5
SSE2_TransTwo4x4W xmm0, xmm1, xmm4, xmm2, xmm3
SSE2_IDCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm0
SSE2_TransTwo4x4W xmm1, xmm4, xmm0, xmm2, xmm3
SSE2_IDCT xmm4, xmm2, xmm3, xmm0, xmm5, xmm6, xmm1
WELS_Zero xmm7
WELS_DW32 xmm6
SSE2_StoreDiff8p xmm4, xmm5, xmm6, xmm7, [r0 ], [r2]
SSE2_StoreDiff8p xmm0, xmm5, xmm6, xmm7, [r0 + r1 ], [r2 + r3]
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff8p xmm1, xmm5, xmm6, xmm7, [r0], [r2]
SSE2_StoreDiff8p xmm2, xmm5, xmm6, xmm7, [r0 + r1 ], [r2 + r3]
add r4, 64
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_Load4x8p r4, xmm0, xmm1, xmm4, xmm2, xmm5
SSE2_TransTwo4x4W xmm0, xmm1, xmm4, xmm2, xmm3
SSE2_IDCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm0
SSE2_TransTwo4x4W xmm1, xmm4, xmm0, xmm2, xmm3
SSE2_IDCT xmm4, xmm2, xmm3, xmm0, xmm5, xmm6, xmm1
WELS_Zero xmm7
WELS_DW32 xmm6
SSE2_StoreDiff8p xmm4, xmm5, xmm6, xmm7, [r0 ], [r2]
SSE2_StoreDiff8p xmm0, xmm5, xmm6, xmm7, [r0 + r1 ], [r2 + r3]
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff8p xmm1, xmm5, xmm6, xmm7, [r0], [r2]
SSE2_StoreDiff8p xmm2, xmm5, xmm6, xmm7, [r0 + r1], [r2 + r3]
POP_XMM
LOAD_5_PARA_POP
; pop esi
; pop ebx
ret
%macro SSE2_StoreDiff4x8p 8
SSE2_StoreDiff8p %1, %3, %4, [%5], [%6]
SSE2_StoreDiff8p %1, %3, %4, [%5 + %7], [%6 + %8]
SSE2_StoreDiff8p %2, %3, %4, [%5 + 8], [%6 + 8]
SSE2_StoreDiff8p %2, %3, %4, [%5 + %7 + 8], [%6 + %8 + 8]
%endmacro
;***********************************************************************
; void WelsIDctRecI16x16Dc_sse2(uint8_t *rec, int32_t stride, uint8_t *pred, int32_t pred_stride, int16_t *dct_dc)
;***********************************************************************
WELS_EXTERN WelsIDctRecI16x16Dc_sse2
%assign push_num 0
LOAD_5_PARA
PUSH_XMM 8
SIGN_EXTENSION r1, r1d
SIGN_EXTENSION r3, r3d
pxor xmm7, xmm7
WELS_DW32 xmm6
SSE2_Load8DC xmm0, xmm1, xmm2, xmm3, xmm6, [r4]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
SSE2_Load8DC xmm0, xmm1, xmm2, xmm3, xmm6, [r4 + 16]
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
POP_XMM
LOAD_5_PARA_POP
ret
%macro SSE2_SumSubD 3
movdqa %3, %2
paddd %2, %1
psubd %1, %3
%endmacro
%macro SSE2_SumSubDiv2D 4
paddd %1, %2
paddd %1, %3
psrad %1, 1
movdqa %4, %1
psubd %4, %2
%endmacro
%macro SSE2_Load4Col 5
movsx r2, WORD[%5]
movd %1, r2d
movsx r2, WORD[%5 + 0x20]
movd %2, r2d
punpckldq %1, %2
movsx r2, WORD[%5 + 0x80]
movd %3, r2d
movsx r2, WORD[%5 + 0xa0]
movd %4, r2d
punpckldq %3, %4
punpcklqdq %1, %3
%endmacro
;***********************************************************************
;void WelsHadamardT4Dc_sse2( int16_t *luma_dc, int16_t *pDct)
;***********************************************************************
WELS_EXTERN WelsHadamardT4Dc_sse2
%assign push_num 0
LOAD_2_PARA
PUSH_XMM 8
SSE2_Load4Col xmm1, xmm5, xmm6, xmm0, r1
SSE2_Load4Col xmm2, xmm5, xmm6, xmm0, r1 + 0x40
SSE2_Load4Col xmm3, xmm5, xmm6, xmm0, r1 + 0x100
SSE2_Load4Col xmm4, xmm5, xmm6, xmm0, r1 + 0x140
SSE2_SumSubD xmm1, xmm2, xmm7
SSE2_SumSubD xmm3, xmm4, xmm7
SSE2_SumSubD xmm2, xmm4, xmm7
SSE2_SumSubD xmm1, xmm3, xmm7
SSE2_Trans4x4D xmm4, xmm2, xmm1, xmm3, xmm5 ; pOut: xmm4,xmm3,xmm5,xmm1
SSE2_SumSubD xmm4, xmm3, xmm7
SSE2_SumSubD xmm5, xmm1, xmm7
WELS_DD1 xmm6
SSE2_SumSubDiv2D xmm3, xmm1, xmm6, xmm0 ; pOut: xmm3 = (xmm3+xmm1+1)/2, xmm0 = (xmm3-xmm1+1)/2
SSE2_SumSubDiv2D xmm4, xmm5, xmm6, xmm1 ; pOut: xmm4 = (xmm4+xmm5+1)/2, xmm1 = (xmm4-xmm5+1)/2
SSE2_Trans4x4D xmm3, xmm0, xmm1, xmm4, xmm2 ; pOut: xmm3,xmm4,xmm2,xmm1
packssdw xmm3, xmm4
packssdw xmm2, xmm1
movdqa [r0+ 0], xmm3
movdqa [r0+16], xmm2
POP_XMM
ret
* DCT functions
****************************************************************************/
void WelsDctT4_c (int16_t* pDct, uint8_t* pPixel1, int32_t iStride1, uint8_t* pPixel2, int32_t iStride2) {
int16_t i, pData[16], s[4];
for (i = 0 ; i < 16 ; i += 4) {
const int32_t kiI1 = 1 + i;
const int32_t kiI2 = 2 + i;
const int32_t kiI3 = 3 + i;
pData[i ] = pPixel1[0] - pPixel2[0];
pData[kiI1] = pPixel1[1] - pPixel2[1];
pData[kiI2] = pPixel1[2] - pPixel2[2];
pData[kiI3] = pPixel1[3] - pPixel2[3];
pPixel1 += iStride1;
pPixel2 += iStride2;
/*horizontal transform */ /* 行变形， 计算Y = X * T(Cf4) = T(Cf4 * T(X)) */
s[0] = pData[i] + pData[kiI3];
s[3] = pData[i] - pData[kiI3];
s[1] = pData[kiI1] + pData[kiI2];
s[2] = pData[kiI1] - pData[kiI2];
pDct[i ] = s[0] + s[1];
pDct[kiI2] = s[0] - s[1];
pDct[kiI1] = (s[3] << 1) + s[2];
pDct[kiI3] = s[3] - (s[2] << 1);
}
/* vertical transform */ /* 列变形， 计算 Cf4 * Y */
for (i = 0 ; i < 4 ; i ++) {
const int32_t kiI4 = 4 + i;
const int32_t kiI8 = 8 + i;
const int32_t kiI12 = 12 + i;
s[0] = pDct[i ] + pDct[kiI12];
s[3] = pDct[i ] - pDct[kiI12];
s[1] = pDct[kiI4] + pDct[kiI8 ];
s[2] = pDct[kiI4] - pDct[kiI8 ];
pDct[i ] = s[0] + s[1];
pDct[kiI8 ] = s[0] - s[1];
pDct[kiI4 ] = (s[3] << 1) + s[2];
pDct[kiI12] = s[3] - (s[2] << 1);
}
}
;*!
;* \copy
;* Copyright (c) 2009-2013, Cisco Systems
;* All rights reserved.
;*
;* Redistribution and use in source and binary forms, with or without
;* modification, are permitted 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.
;*
;* 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.
;*
;*
;* dct.asm
;*
;* Abstract
;* WelsDctFourT4_sse2
;*
;* History
;* 8/4/2009 Created
;*
;*
;*************************************************************************/
%include "asm_inc.asm"
SECTION .rodata align=16
;***********************************************************************
; Constant
;***********************************************************************
align 16
SSE2_DeQuant8 dw 10, 13, 10, 13, 13, 16, 13, 16,
dw 10, 13, 10, 13, 13, 16, 13, 16,
dw 11, 14, 11, 14, 14, 18, 14, 18,
dw 11, 14, 11, 14, 14, 18, 14, 18,
dw 13, 16, 13, 16, 16, 20, 16, 20,
dw 13, 16, 13, 16, 16, 20, 16, 20,
dw 14, 18, 14, 18, 18, 23, 18, 23,
dw 14, 18, 14, 18, 18, 23, 18, 23,
dw 16, 20, 16, 20, 20, 25, 20, 25,
dw 16, 20, 16, 20, 20, 25, 20, 25,
dw 18, 23, 18, 23, 23, 29, 23, 29,
dw 18, 23, 18, 23, 23, 29, 23, 29
;***********************************************************************
; MMX functions
;***********************************************************************
%macro MMX_LoadDiff4P 5
movd %1, [%3]
movd %2, [%4]
punpcklbw %1, %5
punpcklbw %2, %5
psubw %1, %2
%endmacro
%macro MMX_LoadDiff4x4P 10 ;d0, d1, d2, d3, pix1address, pix1stride, pix2address, pix2stride, tmp(mm), 0(mm)
MMX_LoadDiff4P %1, %9, %5, %7, %10
MMX_LoadDiff4P %2, %9, %5+%6, %7+%8, %10
lea %5, [%5+2*%6]
lea %7, [%7+2*%8]
MMX_LoadDiff4P %3, %9, %5, %7, %10
MMX_LoadDiff4P %4, %9, %5+%6, %7+%8, %10
%endmacro
%macro MMX_SumSubMul2 3
movq %3, %1
psllw %1, $01
paddw %1, %2
psllw %2, $01
psubw %3, %2
%endmacro
%macro MMX_SumSubDiv2 3
movq %3, %2
psraw %3, $01
paddw %3, %1
psraw %1, $01
psubw %1, %2
%endmacro
%macro MMX_SumSub 3
movq %3, %2
psubw %2, %1
paddw %1, %3
%endmacro
%macro MMX_DCT 6
MMX_SumSub %4, %1, %6
MMX_SumSub %3, %2, %6
MMX_SumSub %3, %4, %6
MMX_SumSubMul2 %1, %2, %5
%endmacro
%macro MMX_IDCT 6
MMX_SumSub %4, %5, %6
MMX_SumSubDiv2 %3, %2, %1
MMX_SumSub %1, %4, %6
MMX_SumSub %3, %5, %6
%endmacro
%macro MMX_StoreDiff4P 6
movd %2, %6
punpcklbw %2, %4
paddw %1, %3
psraw %1, $06
paddsw %1, %2
packuswb %1, %2
movd %5, %1
%endmacro
SECTION .text
;***********************************************************************
; void WelsDctT4_mmx( int16_t *pDct[4], uint8_t *pix1, int32_t i_pix1, uint8_t *pix2, int32_t i_pix2 )
;***********************************************************************
WELS_EXTERN WelsDctT4_mmx
%assign push_num 0
LOAD_5_PARA
SIGN_EXTENSION r2, r2d
SIGN_EXTENSION r4, r4d
WELS_Zero mm7
； 下面的汇编代码将使 mm1 保存的是 4x4 残差矩阵的第一行 ； mm2 保存的是 4x4 残差矩阵的第二行
; mm3 保存的是 4x4 残差矩阵的第三行； mm4 保存的是 4x4 残差矩阵的第四行
MMX_LoadDiff4x4P mm1, mm2, mm3, mm4, r1, r2, r3, r4, mm0, mm7
； 先执行列变换， 即 Y = Cf4 * X, mm1 denote first component of column vector, mm2 denote second component
; mm3 third component, mm4 4th component.
MMX_DCT mm1, mm2, mm3 ,mm4, mm5, mm6
； after transforming, mm3, mm1, mm4, mm5 denote first, second, third, 4th component of result vector
； mm3 = a3a2a1a0, mm1 = b3b2b1b0, mm4 = c3c2c1c0, mm5 = d3d2d1d0
; expand =>
; 1.) movq mm2, mm3 ; mm2 = mm3 = a3a2a1a0
; punpckhwd mm2, mm1 ; mm2 = b3a3b2a2
; punpcklwd mm3, mm1 ; mm3 = b1a1b0a0
;
; 2.) movq mm1, mm4 ; mm1 = mm4 = c3c2c1c0
; punpckhwd mm1, mm5 ; mm1 = d3c3d2c2
; punpcklwd mm4, mm5 ; mm4 = d1c1d0c0
;
; 3.) movq mm5, mm3 ; mm5 = mm3 = b1a1b0a0
; punpckhdq mm5, mm4 ; mm5 = d1c1b1a1
; punpckldq mm3, mm4 ; mm3 = d0c0b0a0
;
; 4.) movq mm4, mm2 ; mm4 = mm2 = b3a3b2a2
; punpckhdq mm4, mm1 ; mm4 = d3c3b3a3
; punpckldq mm2, mm1 ; mm2 = d2c2b2a2
MMX_Trans4x4W mm3, mm1, mm4, mm5, mm2
; 因此上面的转置后， mm3 = d0c0b0a0, mm5 = d1c1b1a1, mm2 = d2c2b2a2, mm4 = d3c3b3a3
; 这儿执行行变形 《=》 Z = Cf4 * T（Y）
MMX_DCT mm3, mm5, mm2 ,mm4, mm1, mm6
; 再次转置 T(Z) = Y * T(Cf4)
MMX_Trans4x4W mm2, mm3, mm4, mm1, mm5
movq [r0+ 0], mm2
movq [r0+ 8], mm1
movq [r0+16], mm5
movq [r0+24], mm4
WELSEMMS
LOAD_5_PARA_POP
ret
/****************************************************************************
* IDCT functions, final output = prediction(CS) + IDCT(scaled_coeff)
****************************************************************************/
void WelsIDctT4Rec_c (uint8_t* pRec, int32_t iStride, uint8_t* pPred, int32_t iPredStride, int16_t* pDct) {
int32_t i;
int16_t iTemp[16];
int32_t iDstStridex2 = iStride << 1;
int32_t iDstStridex3 = iStride + iDstStridex2;
int32_t iPredStridex2 = iPredStride << 1;
int32_t iPredStridex3 = iPredStride + iPredStridex2;
for (i = 0; i < 4; i ++) { // horizon; Row Transform, Y = Z * Ci4
int32_t iIdx = i << 2;
; kiHorSumU = a + c
const int32_t kiHorSumU = pDct[iIdx] + pDct[iIdx + 2]; // add 0-2
; kiHorDelU = a - c
const int32_t kiHorDelU = pDct[iIdx] - pDct[iIdx + 2]; // sub 0-2
; kiHorSumD = b + d / 2
const int32_t kiHorSumD = pDct[iIdx + 1] + (pDct[iIdx + 3] >> 1);
; kiHorDelD = b / 2 - d
const int32_t kiHorDelD = (pDct[iIdx + 1] >> 1) - pDct[iIdx + 3];
; iTemp[ildx] = a + b + c + d / 2
iTemp[iIdx ] = kiHorSumU + kiHorSumD;
; iTemp[ildx + 1] = a + b / 2 - c - d
iTemp[iIdx + 1] = kiHorDelU + kiHorDelD;
; iTemp[ildx + 2] = a - b / 2 - c + d
iTemp[iIdx + 2] = kiHorDelU - kiHorDelD;
; iTemp[ildx + 2] = a - b + c - d / 2
iTemp[iIdx + 3] = kiHorSumU - kiHorSumD;
}
for (i = 0; i < 4; i ++) { // vertical; Column Transform, X = T(Ci4) * Y
; kiVerSumL = a + c
const int32_t kiVerSumL = iTemp[i] + iTemp[8 + i];
; kiVerDelL = a - c
const int32_t kiVerDelL = iTemp[i] - iTemp[8 + i];
; kiVerDelR = b / 2 - d
const int32_t kiVerDelR = (iTemp[4 + i] >> 1) - iTemp[12 + i];
; kiVerSumR = b + d / 2
const int32_t kiVerSumR = iTemp[4 + i] + (iTemp[12 + i] >> 1);
; kiVerSumL + kiVerSumR = a + b + c + d / 2
pRec[i ] = WelsClip1 (pPred[i ] + ((kiVerSumL + kiVerSumR + 32) >> 6));
; kiVerDelL + kiVerDelR = a + b / 2 - c - d
pRec[iStride + i ] = WelsClip1 (pPred[iPredStride + i ] + ((kiVerDelL + kiVerDelR + 32) >> 6));
; kiVerDelL - kiVerDelR = a - b / 2 - c + d
pRec[iDstStridex2 + i] = WelsClip1 (pPred[iPredStridex2 + i] + ((kiVerDelL - kiVerDelR + 32) >> 6));
; kiVerSumL - kiVerSumR = a - b + c - d / 2
pRec[iDstStridex3 + i] = WelsClip1 (pPred[iPredStridex3 + i] + ((kiVerSumL - kiVerSumR + 32) >> 6));
}
}
;***********************************************************************
; void WelsIDctT4Rec_mmx(uint8_t *rec, int32_t stride, uint8_t *pred, int32_t pred_stride, int16_t *rs)
;***********************************************************************
WELS_EXTERN WelsIDctT4Rec_mmx
%assign push_num 0
LOAD_5_PARA
SIGN_EXTENSION r1, r1d
SIGN_EXTENSION r3, r3d
movq mm0, [r4+ 0]
movq mm1, [r4+ 8]
movq mm2, [r4+16]
movq mm3, [r4+24]
; mm0 为 4x4 矩阵的第一行， mm1 为 第二行；以此类推
; 转置后为 mm0 = a, mm3 = b, mm4 = c, mm2 = d
; 即输出结果为 a, 空， d, b, c
MMX_Trans4x4W mm0, mm1, mm2, mm3, mm4
; MMX_IDC mm1, d, b, c, a, mm6; 行变换，Y = Z * Ci4
; 结果为mm1 = a + b + c + d / 2, mm3 = a + b / 2 - c - d
; mm0 = a - b / 2 - c + d, mm4 = a - b + c - d / 2
MMX_IDCT mm1, mm2, mm3, mm4, mm0, mm6
; mm1 = a, mm3 = b, mm0 = c, mm4 = d => a, 空， d, b, c
; 即 mm1 = a, mm3, mm0 = d, mm4 = b, mm2 = c
MMX_Trans4x4W mm1, mm3, mm0, mm4, mm2
; MMX_IDCT mm3, mm0 = d, mm4 = b, mm2 = c, mm1 = a, mm6
; X = T(Ci4) * W; 列变换， W 为列向量
; 结果为 mm3, mm4, mm1, mm2
MMX_IDCT mm3, mm0, mm4, mm2, mm1, mm6
; mm7 = 0
WELS_Zero mm7
; mm6 = 32 32 32 32
WELS_DW32 mm6
; calc pRec[i ] = WelsClip1 (pPred[i ] + ((kiVerSumL + kiVerSumR + 32) >> 6)); one whole row
; %macro MMX_StoreDiff4P 6
; movd %2, %6
; punpcklbw %2, %4
; paddw %1, %3
; psraw %1, $06
; paddsw %1, %2
; packuswb %1, %2
; movd %5, %1
; %endmacro
; MMX_StoreDiff4P mm3, mm0, mm6, mm7, [r0], [r2] =>
; movd mm0, [r2]
; punpcklbw mm0, mm7 ; mm0, mm7 = 0 交叉组合 => mm0
; paddw mm3, mm6 ; mm6 = 32 32 32 32; mm3 += 32;
; psraw mm3, 6 ; mm3 = mm3 >> 6
; paddsw mm3, mm0 ; mm3 += mm0; mm0 即 [r2]
; packuswb mm3, mm0 ; 无符号饱和压缩 => mm3 (mm3为什么和mm0作无符号饱和压缩 ？？？)
; 哦，想明白了， mm0 在上面的饱和压缩运算中， 相当于一个哑元而已
; movd [r0], mm3
MMX_StoreDiff4P mm3, mm0, mm6, mm7, [r0], [r2]
; calc pRec[iStride + i ] = WelsClip1 (pPred[iPredStride + i ] + ((kiVerDelL + kiVerDelR + 32) >> 6)); one whole row
MMX_StoreDiff4P mm4, mm0, mm6, mm7, [r0+r1], [r2+r3]
lea r0, [r0+2*r1]
lea r2, [r2+2*r3]
; calc pRec[iDstStridex2 + i] = WelsClip1 (pPred[iPredStridex2 + i] + ((kiVerDelL - kiVerDelR + 32) >> 6)); one whole row
MMX_StoreDiff4P mm1, mm0, mm6, mm7, [r0], [r2]
; calc pRec[iDstStridex3 + i] = WelsClip1 (pPred[iPredStridex3 + i] + ((kiVerSumL - kiVerSumR + 32) >> 6)); one whole row
MMX_StoreDiff4P mm2, mm0, mm6, mm7, [r0+r1], [r2+r3]
WELSEMMS
LOAD_5_PARA_POP
ret
;for TRANSPOSE
%macro SSE2_XSawp 4
movdqa %4, %2 ; 移动16字节对齐的双四字
punpckl%1 %2, %3 ; 如 punpcklqdq，交叉组合低位四字
punpckh%1 %4, %3 ; 如 punpckhqdq, 交叉组合高位四字
%endmacro
;***********************************************************************
; SSE2 functions
;***********************************************************************
%macro SSE2_Store4x8p 6
SSE2_XSawp qdq, %2, %3, %6
SSE2_XSawp qdq, %4, %5, %3
MOVDQ [%1+0x00], %2
MOVDQ [%1+0x10], %4
MOVDQ [%1+0x20], %6
MOVDQ [%1+0x30], %3
%endmacro
%macro SSE2_Load4x8p 6
MOVDQ %2, [%1+0x00]
MOVDQ %4, [%1+0x10]
MOVDQ %6, [%1+0x20]
MOVDQ %3, [%1+0x30]
SSE2_XSawp qdq, %4, %3, %5
SSE2_XSawp qdq, %2, %6, %3
%endmacro
%macro SSE2_SumSubMul2 3
movdqa %3, %1
paddw %1, %1
paddw %1, %2
psubw %3, %2
psubw %3, %2
%endmacro
%macro SSE2_SumSubDiv2 4
movdqa %4, %1
movdqa %3, %2
psraw %2, $01
psraw %4, $01
paddw %1, %2
psubw %4, %3
%endmacro
%macro SSE2_StoreDiff8p 6
paddw %1, %3
psraw %1, $06
movq %2, %6
punpcklbw %2, %4
paddsw %2, %1
packuswb %2, %2
movq %5, %2
%endmacro
%macro SSE2_StoreDiff8p 5
movq %2, %5
punpcklbw %2, %3
paddsw %2, %1
packuswb %2, %2
movq %4, %2
%endmacro
%macro SSE2_Load8DC 6
movdqa %1, %6 ; %1 = dc0 dc1
paddw %1, %5
psraw %1, $06 ; (dc + 32) >> 6
movdqa %2, %1
psrldq %2, 4
punpcklwd %2, %2
punpckldq %2, %2 ; %2 = dc2 dc2 dc2 dc2 dc3 dc3 dc3 dc3
movdqa %3, %1
psrldq %3, 8
punpcklwd %3, %3
punpckldq %3, %3 ; %3 = dc4 dc4 dc4 dc4 dc5 dc5 dc5 dc5
movdqa %4, %1
psrldq %4, 12
punpcklwd %4, %4
punpckldq %4, %4 ; %4 = dc6 dc6 dc6 dc6 dc7 dc7 dc7 dc7
punpcklwd %1, %1
punpckldq %1, %1 ; %1 = dc0 dc0 dc0 dc0 dc1 dc1 dc1 dc1
%endmacro
%macro SSE2_DCT 6
SSE2_SumSub %6, %3, %5
SSE2_SumSub %1, %2, %5
SSE2_SumSub %3, %2, %5
SSE2_SumSubMul2 %6, %1, %4
%endmacro
%macro SSE2_IDCT 7
SSE2_SumSub %7, %2, %6
SSE2_SumSubDiv2 %1, %3, %5, %4
SSE2_SumSub %2, %1, %5
SSE2_SumSub %7, %4, %5
%endmacro
;***********************************************************************
; void WelsDctFourT4_sse2(int16_t *pDct, uint8_t *pix1, int32_t i_pix1, uint8_t *pix2, int32_t i_pix2 )
;***********************************************************************
WELS_EXTERN WelsDctFourT4_sse2
%assign push_num 0
LOAD_5_PARA
PUSH_XMM 8
SIGN_EXTENSION r2, r2d
SIGN_EXTENSION r4, r4d
pxor xmm7, xmm7
;Load 4x8
; ;xmm0, xmm6, xmm7, [eax], [ecx]
; ;xmm7 = 0, eax = pix1, ecx = pix2, xmm0 save the result
; %macro SSE2_LoadDiff8P 5
; movq %1, %4
; punpcklbw %1, %3
; movq %2, %5
; punpcklbw %2, %3
; psubw %1, %2
; %endmacro
; SSE2_LoadDiff8P xmm0, xmm6, xmm7, [r1], [r3] =>
; movq xmm0, [r1]
; punpcklbw xmm0, xmm7 ;交叉组合xmm0与xmm7低位四字中的字节,将结果放入xmm0
; movq xmm6, [r3]
; punpcklbw xmm6, xmm7 ;交叉组合xmm6与xmm7低位四字中的字节,将结果放入xmm6
; psubw xmm0, xmm6 ; xmm0 -= mm6;
; i.e xmm0 = [r1] - [r3]; 8个 16bit 相减
SSE2_LoadDiff8P xmm0, xmm6, xmm7, [r1], [r3]
; 同样地， xmml = [r1 + r2] - [r3 + r4]
SSE2_LoadDiff8P xmm1, xmm6, xmm7, [r1+r2], [r3+r4]
lea r1, [r1 + 2 * r2]
lea r3, [r3 + 2 * r4]
; 同样地， xmm2 = [r1 + 2r2] - [r3 + 2r4]
SSE2_LoadDiff8P xmm2, xmm6, xmm7, [r1], [r3]
; 同样地， xmm3 = [r1 + 3r2] - [r3 + 3r4]
SSE2_LoadDiff8P xmm3, xmm6, xmm7, [r1+r2], [r3+r4]
; %macro SSE2_DCT 6
; SSE2_SumSub %6, %3, %5
; SSE2_SumSub %1, %2, %5
; SSE2_SumSub %3, %2, %5
; SSE2_SumSubMul2 %6, %1, %4
; %endmacro
; m2 = m1 + m2, m1 = m1 - m2
; %macro SSE2_SumSub 3
; movdqa %3, %2
; paddw %2, %1
; psubw %1, %3
; %endmacro
; %macro SSE2_SumSubMul2 3
; movdqa %3, %1
; paddw %1, %1
; paddw %1, %2
; psubw %3, %2
; psubw %3, %2
; %endmacro
; Y = Cf4 * X, 列变换； result = xmm2, xmm0, xmm3, xmm4
; SSE2_DCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm0 =>
; SSE2_SumSub xmm0, xmm3, xmm5 =>
;;;; movdqa xmm5, xmm3 => xmm5 = xmm3 = d
;;;; paddw xmm3, xmm0 => xmm3 = xmm3 + xmm0 = a + d
;;;; psubw xmm0, xmm5 => xmm0 = xmm0 - xmm5 = a - d
; SSE2_SumSub xmm1, xmm2, xmm5
;;;; movdqa xmm5, xmm2 => xmm5 = xmm2 = c
;;;; paddw xmm2, xmm1 => xmm2 = xmm2 + xmm1 = b + c
;;;; psubw xmm1, xmm5 => xmm1 = xmm1 - xmm5 = b - c
; SSE2_SumSub xmm3, xmm2, xmm5
;;;; movdqa xmm5, xmm2 => xmm5 = xmm2 = b + c
;;;; paddw xmm2, xmm3 => xmm2 = xmm2 + xmm3 = a + b + c + d
;;;; psubw xmm3, xmm5 => xmm3 = xmm3 - xmm5 = a - b - c + d
; SSE2_SumSubMul2 xmm0, xmm1, xmm4
;;;; movdqa xmm4, xmm0 => xmm4 = xmm0 = a - d
;;;; paddw xmm0, xmm0 => xmm0 = 2(a - d)
;;;; paddw xmm0, xmm1 => xmm0 = xmm0 + xmm1 = 2a + b -c - 2d
;;;; psubw xmm4, xmm1 => xmm4 = xmm4 - xmm1 = a - b + c - d
;;;; psubw xmm4, xmm1 => xmm4 = xmm4 - xmm1 = a - 2b + 2c - d
; result => xmm2 = 1st, xmm0 = 2nd, xmm3 = 3rd, xmm4 = 4th
SSE2_DCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm0
SSE2_TransTwo4x4W xmm2, xmm0, xmm3, xmm4, xmm1
SSE2_DCT xmm0, xmm4, xmm1, xmm3, xmm5, xmm2
SSE2_TransTwo4x4W xmm4, xmm2, xmm1, xmm3, xmm0
SSE2_Store4x8p r0, xmm4, xmm2, xmm3, xmm0, xmm5
lea r1, [r1 + 2 * r2]
lea r3, [r3 + 2 * r4]
;Load 4x8
SSE2_LoadDiff8P xmm0, xmm6, xmm7, [r1 ], [r3 ]
SSE2_LoadDiff8P xmm1, xmm6, xmm7, [r1+r2 ], [r3+r4]
lea r1, [r1 + 2 * r2]
lea r3, [r3 + 2 * r4]
SSE2_LoadDiff8P xmm2, xmm6, xmm7, [r1], [r3]
SSE2_LoadDiff8P xmm3, xmm6, xmm7, [r1+r2], [r3+r4]
SSE2_DCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm0
SSE2_TransTwo4x4W xmm2, xmm0, xmm3, xmm4, xmm1
SSE2_DCT xmm0, xmm4, xmm1, xmm3, xmm5, xmm2
SSE2_TransTwo4x4W xmm4, xmm2, xmm1, xmm3, xmm0
lea r0, [r0+64]
SSE2_Store4x8p r0, xmm4, xmm2, xmm3, xmm0, xmm5
POP_XMM
LOAD_5_PARA_POP
ret
;***********************************************************************
; void WelsIDctFourT4Rec_sse2(uint8_t *rec, int32_t stride, uint8_t *pred, int32_t pred_stride, int16_t *rs);
;***********************************************************************
WELS_EXTERN WelsIDctFourT4Rec_sse2
%assign push_num 0
LOAD_5_PARA
PUSH_XMM 8
SIGN_EXTENSION r1, r1d
SIGN_EXTENSION r3, r3d
;Load 4x8
SSE2_Load4x8p r4, xmm0, xmm1, xmm4, xmm2, xmm5
SSE2_TransTwo4x4W xmm0, xmm1, xmm4, xmm2, xmm3
SSE2_IDCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm0
SSE2_TransTwo4x4W xmm1, xmm4, xmm0, xmm2, xmm3
SSE2_IDCT xmm4, xmm2, xmm3, xmm0, xmm5, xmm6, xmm1
WELS_Zero xmm7
WELS_DW32 xmm6
SSE2_StoreDiff8p xmm4, xmm5, xmm6, xmm7, [r0 ], [r2]
SSE2_StoreDiff8p xmm0, xmm5, xmm6, xmm7, [r0 + r1 ], [r2 + r3]
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff8p xmm1, xmm5, xmm6, xmm7, [r0], [r2]
SSE2_StoreDiff8p xmm2, xmm5, xmm6, xmm7, [r0 + r1 ], [r2 + r3]
add r4, 64
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_Load4x8p r4, xmm0, xmm1, xmm4, xmm2, xmm5
SSE2_TransTwo4x4W xmm0, xmm1, xmm4, xmm2, xmm3
SSE2_IDCT xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm0
SSE2_TransTwo4x4W xmm1, xmm4, xmm0, xmm2, xmm3
SSE2_IDCT xmm4, xmm2, xmm3, xmm0, xmm5, xmm6, xmm1
WELS_Zero xmm7
WELS_DW32 xmm6
SSE2_StoreDiff8p xmm4, xmm5, xmm6, xmm7, [r0 ], [r2]
SSE2_StoreDiff8p xmm0, xmm5, xmm6, xmm7, [r0 + r1 ], [r2 + r3]
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff8p xmm1, xmm5, xmm6, xmm7, [r0], [r2]
SSE2_StoreDiff8p xmm2, xmm5, xmm6, xmm7, [r0 + r1], [r2 + r3]
POP_XMM
LOAD_5_PARA_POP
; pop esi
; pop ebx
ret
%macro SSE2_StoreDiff4x8p 8
SSE2_StoreDiff8p %1, %3, %4, [%5], [%6]
SSE2_StoreDiff8p %1, %3, %4, [%5 + %7], [%6 + %8]
SSE2_StoreDiff8p %2, %3, %4, [%5 + 8], [%6 + 8]
SSE2_StoreDiff8p %2, %3, %4, [%5 + %7 + 8], [%6 + %8 + 8]
%endmacro
;***********************************************************************
; void WelsIDctRecI16x16Dc_sse2(uint8_t *rec, int32_t stride, uint8_t *pred, int32_t pred_stride, int16_t *dct_dc)
;***********************************************************************
WELS_EXTERN WelsIDctRecI16x16Dc_sse2
%assign push_num 0
LOAD_5_PARA
PUSH_XMM 8
SIGN_EXTENSION r1, r1d
SIGN_EXTENSION r3, r3d
pxor xmm7, xmm7
WELS_DW32 xmm6
SSE2_Load8DC xmm0, xmm1, xmm2, xmm3, xmm6, [r4]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
SSE2_Load8DC xmm0, xmm1, xmm2, xmm3, xmm6, [r4 + 16]
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm0, xmm1, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
lea r0, [r0 + 2 * r1]
lea r2, [r2 + 2 * r3]
SSE2_StoreDiff4x8p xmm2, xmm3, xmm5, xmm7, r0, r2, r1, r3
POP_XMM
LOAD_5_PARA_POP
ret
%macro SSE2_SumSubD 3
movdqa %3, %2
paddd %2, %1
psubd %1, %3
%endmacro
%macro SSE2_SumSubDiv2D 4
paddd %1, %2
paddd %1, %3
psrad %1, 1
movdqa %4, %1
psubd %4, %2
%endmacro
%macro SSE2_Load4Col 5
movsx r2, WORD[%5]
movd %1, r2d
movsx r2, WORD[%5 + 0x20]
movd %2, r2d
punpckldq %1, %2
movsx r2, WORD[%5 + 0x80]
movd %3, r2d
movsx r2, WORD[%5 + 0xa0]
movd %4, r2d
punpckldq %3, %4
punpcklqdq %1, %3
%endmacro
;***********************************************************************
;void WelsHadamardT4Dc_sse2( int16_t *luma_dc, int16_t *pDct)
;***********************************************************************
WELS_EXTERN WelsHadamardT4Dc_sse2
%assign push_num 0
LOAD_2_PARA
PUSH_XMM 8
SSE2_Load4Col xmm1, xmm5, xmm6, xmm0, r1
SSE2_Load4Col xmm2, xmm5, xmm6, xmm0, r1 + 0x40
SSE2_Load4Col xmm3, xmm5, xmm6, xmm0, r1 + 0x100
SSE2_Load4Col xmm4, xmm5, xmm6, xmm0, r1 + 0x140
SSE2_SumSubD xmm1, xmm2, xmm7
SSE2_SumSubD xmm3, xmm4, xmm7
SSE2_SumSubD xmm2, xmm4, xmm7
SSE2_SumSubD xmm1, xmm3, xmm7
SSE2_Trans4x4D xmm4, xmm2, xmm1, xmm3, xmm5 ; pOut: xmm4,xmm3,xmm5,xmm1
SSE2_SumSubD xmm4, xmm3, xmm7
SSE2_SumSubD xmm5, xmm1, xmm7
WELS_DD1 xmm6
SSE2_SumSubDiv2D xmm3, xmm1, xmm6, xmm0 ; pOut: xmm3 = (xmm3+xmm1+1)/2, xmm0 = (xmm3-xmm1+1)/2
SSE2_SumSubDiv2D xmm4, xmm5, xmm6, xmm1 ; pOut: xmm4 = (xmm4+xmm5+1)/2, xmm1 = (xmm4-xmm5+1)/2
SSE2_Trans4x4D xmm3, xmm0, xmm1, xmm4, xmm2 ; pOut: xmm3,xmm4,xmm2,xmm1
packssdw xmm3, xmm4
packssdw xmm2, xmm1
movdqa [r0+ 0], xmm3
movdqa [r0+16], xmm2
POP_XMM
ret
内容来自用户分享和网络整理，不保证内容的准确性，如有侵权内容，可联系管理员处理 
相关文章推荐
- 使用shell和awk批量处理二进制数据
- 监控
- nginx配置location总结及rewrite规则写法
- Linux awk sed
- centos yum 错误处理
- Linux内核哈希表分析与应用
- shell记录
- eclipse 启动tomcat一直处在starting状态
- shell 常用命令之三 grep
- 深入浅出linux内核源代码之双向链表list_head(下)
- 深入浅出linux内核源代码之双向链表list_head(上)
- 查看Linux版本系统信息方法汇总
- imeOptions使用注意
- linux screen 命令行终端切换
- Corosync+Pacemaker实现web集群高可用
- 《Linux内核Makefile分析》之 if_changed_rule/cc_o_c/any-prereq/arg-check
- 需要交互的shell编程——EOF(转载)
- meizu手机的虚拟键盘被popwindow覆盖的解决办法
- Linux 编程学习笔记----ANSI C 文件I/O管理
- Linux 日常管理