Assembly x64 Intro - SSE2 IDCT I16X16 DC
/*
* pfIDctI16x16Dc: do luma idct of an MB for I16x16 mode, when only dc value are non-zero
*/
void WelsIDctRecI16x16Dc_c (uint8_t* pRec, int32_t iStride, uint8_t* pPred, int32_t iPredStride, int16_t* pDctDc)
{
int32_t i, j;
for (i = 0; i < 16; i ++) {
for (j = 0; j < 16; j++) {
pRec[j] = WelsClip1 (pPred[j] + ((pDctDc[ (i & 0x0C) + (j >> 2)] + 32) >> 6));
}
pRec += iStride;
pPred += iPredStride;
}
}
;***********************************************************************
; 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