悔しいので公式どおりな離散コサイン変換 MMX 版

晒しておきます。
予め言っておきますが遅いです。
しかも汚くて読みづらいです。

static int icos_xu[8][8] = {
	{ 181,  251,  236,  212,  181,  142,  97,  49 },
	{ 181,  212,  97, -50, -182, -252, -237, -143 },
	{ 181,  142, -98, -252, -182,  49,  236,  212 },
	{ 181,  49, -237, -143,  181,  212, -98, -252 },
	{ 181, -50, -237,  142,  181, -213, -98,  251 },
	{ 181, -143, -98,  251, -182, -50,  236, -213 },
	{ 181, -213,  97,  49, -182,  251, -237,  142 },
	{ 181, -252,  236, -213,  181, -143,  97, -50 }
};

void DecodeDCT( int* _dct, u_int* _data )
{
	int* _icos_x = icos_xu[0];
	int* _icos_y = icos_xu[0];

	for(int y=0; y<8; y++) {
		for(int x=0; x<8; x++) {
/*
			int _tmp[8] = { 0 };
			for(int v=0; v<8; v++) {
				int* __dct = _dct[v];
				int ksu = _icos_y[v];
				for(int u=0; u<8; u++) {
					_tmp[u] += ksu * __dct[u];
				}
			}
			int sum = 0;
			for(int u=0; u<8; u++) sum += _tmp[u] * icos_xu[x][u];
*/
			_asm {
				;  eax = 汎用
				;  ebx = int* _icos_y
				;  ecx = ループカウンタ
				;  edx = 汎用
				;  esi = int* __dct = _dct[v]
				;  edi = int* icos_xu[x]
				;  mm0〜3 = _tmp[0〜7]
				;  mm4 = ksu = _icos_y[v]
				;  mm5〜7 = 汎用
				mov esi, _dct
				mov ebx, _icos_y
				mov ecx, 8
				pxor mm0, mm0
				pxor mm1, mm1
				pxor mm2, mm2
				pxor mm3, mm3
				
			MMX_DCT_LOOP:
				
				movd mm4, [ebx]
				movq mm5, mm4
				punpckldq mm4, mm5
				
				;   < Schedule Map >
				; mm5  RED MUL --> --> ADD         RED MUL --> --> ADD 
				; mm6          RED MUL --> -->             ADD         
				; mm7                      RED MUL --> -->     ADD     
				movq mm5, [esi]
				pmaddwd mm5, mm4
				movq mm6, [esi+8]
				pmaddwd mm6, mm4
				paddd mm0, mm5
				movq mm7, [esi+16]
				pmaddwd mm7, mm4
				movq mm5, [esi+24]
				pmaddwd mm5, mm4
				paddd mm1, mm6
				paddd mm2, mm7
				paddd mm3, mm5
				
				add esi, 32
				add ebx, 4
			loop MMX_DCT_LOOP
				
				mov edi, _icos_x
				
				;   < Schedule Map >
				;  RED SFT MUL --> -->                 RED SFT MUL --> --> ADD 
				;              RED SFT MUL --> -->                 ADD         
				;                          RED SFT MUL --> -->         ADD     
				movq mm5, [edi]
				psrad mm0, 10
				pmaddwd mm0, mm5
				
				movq mm6, [edi+8]
				psrad mm1, 10
				pmaddwd mm1, mm6
				
				movq mm7, [edi+16]
				psrad mm2, 10
				pmaddwd mm2, mm7
				
				movq mm5, [edi+24]
				psrad mm3, 10
				pmaddwd mm3, mm5
				
				paddd mm0, mm1
				paddd mm2, mm3
				paddd mm0, mm2
				
				psrlq mm0, 8
				movq mm1, mm0
				psrlq mm1, 32
				paddd mm0, mm1
				
				mov esi, _data
				movd [esi], mm0
				
				add esi, 4
				add edi, 32
				mov _data, esi
				mov _icos_x, edi
				
				emms
			}
		}
		_icos_x -= 64; _icos_y += 8;
	}
}