Lines Matching full:xmm3
104 movq xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
110 por xmm3,xmm4
112 por xmm1,xmm3
132 movaps xmm3,xmm0
137 shufps xmm3,xmm3,0xFF ; xmm3=(03 03 03 03)
145 movaps XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
146 movaps XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
156 movq xmm3, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
166 punpcklwd xmm3,xmm3 ; xmm3=(60 60 61 61 62 62 63 63)
168 psrad xmm3,(DWORD_BIT-WORD_BIT) ; xmm3=in6=(60 61 62 63)
170 cvtdq2ps xmm3,xmm3 ; xmm3=in6=(60 61 62 63)
175 mulps xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
180 subps xmm1,xmm3
182 addps xmm5,xmm3 ; xmm5=tmp13
200 movq xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
205 punpcklwd xmm3,xmm3 ; xmm3=(30 30 31 31 32 32 33 33)
207 psrad xmm3,(DWORD_BIT-WORD_BIT) ; xmm3=in3=(30 31 32 33)
209 cvtdq2ps xmm3,xmm3 ; xmm3=in3=(30 31 32 33)
219 mulps xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
226 addps xmm5,xmm3 ; xmm5=z13
228 subps xmm0,xmm3 ; xmm0=z10
236 movaps xmm3,xmm0
239 mulps xmm3,[rel PD_M2_613] ; xmm3=(z10 * -2.613125930)
241 addps xmm3,xmm0 ; xmm3=tmp12
246 subps xmm3,xmm1 ; xmm3=tmp6
250 addps xmm7,xmm3 ; xmm7=data1=(10 11 12 13)
252 subps xmm0,xmm3 ; xmm0=data6=(60 61 62 63)
253 subps xmm2,xmm3 ; xmm2=tmp5
258 movaps xmm3,xmm0 ; transpose coefficients(phase 1)
260 unpckhps xmm3,xmm5 ; xmm3=(62 72 63 73)
266 movaps XMMWORD [wk(1)], xmm3 ; wk(1)=(62 72 63 73)
270 movaps xmm3,xmm5
274 subps xmm3,xmm4 ; xmm3=data3=(30 31 32 33)
277 unpcklps xmm7,xmm3 ; xmm7=(20 30 21 31)
278 unpckhps xmm2,xmm3 ; xmm2=(22 32 23 33)
283 movaps xmm3,xmm6 ; transpose coefficients(phase 2)
285 unpckhps2 xmm3,xmm7 ; xmm3=(01 11 21 31)
294 movaps XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
301 movaps xmm3,xmm4 ; transpose coefficients(phase 2)
303 unpckhps2 xmm3,xmm2 ; xmm3=(43 53 63 73)
308 movaps XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
338 movaps xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_FAST_FLOAT)]
343 subps xmm1,xmm3
345 addps xmm5,xmm3 ; xmm5=tmp13
363 movaps xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_FAST_FLOAT)]
370 addps xmm5,xmm3 ; xmm5=z13
372 subps xmm0,xmm3 ; xmm0=z10
380 movaps xmm3,xmm0
383 mulps xmm3,[rel PD_M2_613] ; xmm3=(z10 * -2.613125930)
385 addps xmm3,xmm0 ; xmm3=tmp12
390 subps xmm3,xmm1 ; xmm3=tmp6
394 xmm3 ; xmm7=data1=(01 11 21 31)
396 subps xmm0,xmm3 ; xmm0=data6=(06 16 26 36)
397 subps xmm2,xmm3 ; xmm2=tmp5
400 pcmpeqd xmm3,xmm3
401 psrld xmm3,WORD_BIT ; xmm3={0xFFFF 0x0000 0xFFFF 0x0000 ..}
408 pand xmm6,xmm3 ; xmm6=(00 -- 10 -- 20 -- 30 --)
410 pand xmm0,xmm3 ; xmm0=(06 -- 16 -- 26 -- 36 --)
416 movaps xmm3, XMMWORD [wk(1)] ; xmm3=tmp3
420 movaps xmm5,xmm3
422 addps xmm3,xmm4 ; xmm3=data4=(04 14 24 34)
430 addps xmm3,xmm2 ; xmm3=roundint(data4/8)=(04 ** 14 ** 24 ** 34 **)
435 pand xmm3,xmm4 ; xmm3=(04 -- 14 -- 24 -- 34 --)
439 por xmm3,xmm7 ; xmm3=(04 05 14 15 24 25 34 35)
444 packsswb xmm6,xmm3 ; xmm6=(00 01 10 11 20 21 30 31 04 05 14 15 24 25 34 35)
458 pshufd xmm3,xmm7,0x4E ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
467 movq XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm3
