Lines 114-125
Link Here
|
114 |
"movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0 |
114 |
"movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0 |
115 |
"punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0 |
115 |
"punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0 |
116 |
"punpckldq %%mm1,%%mm1\n" // 00 v1 00 v0 00 v1 00 v0 |
116 |
"punpckldq %%mm1,%%mm1\n" // 00 v1 00 v0 00 v1 00 v0 |
117 |
"psubw %[_MMX_0080w],%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 |
117 |
"psubw _MMX_0080w,%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 |
118 |
|
118 |
|
119 |
// create Cr_g (result in mm0) |
119 |
// create Cr_g (result in mm0) |
120 |
"movq %%mm1,%%mm0\n" // r1 r1 r0 r0 r1 r1 r0 r0 |
120 |
"movq %%mm1,%%mm0\n" // r1 r1 r0 r0 r1 r1 r0 r0 |
121 |
"pmullw %[_MMX_VgrnRGB],%%mm0\n"// red*-46dec=0.7136*64 |
121 |
"pmullw _MMX_VgrnRGB,%%mm0\n"// red*-46dec=0.7136*64 |
122 |
"pmullw %[_MMX_VredRGB],%%mm1\n"// red*89dec=1.4013*64 |
122 |
"pmullw _MMX_VredRGB,%%mm1\n"// red*89dec=1.4013*64 |
123 |
"psraw $6, %%mm0\n" // red=red/64 |
123 |
"psraw $6, %%mm0\n" // red=red/64 |
124 |
"psraw $6, %%mm1\n" // red=red/64 |
124 |
"psraw $6, %%mm1\n" // red=red/64 |
125 |
|
125 |
|
Lines 128-135
Link Here
|
128 |
"movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0 |
128 |
"movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0 |
129 |
"punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0 |
129 |
"punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0 |
130 |
"movq %%mm2,%%mm4\n" // L3 L2 L1 L0 l3 l2 l1 l0 |
130 |
"movq %%mm2,%%mm4\n" // L3 L2 L1 L0 l3 l2 l1 l0 |
131 |
"pand %[_MMX_FF00w],%%mm2\n" // L3 0 L1 0 l3 0 l1 0 |
131 |
"pand _MMX_FF00w,%%mm2\n" // L3 0 L1 0 l3 0 l1 0 |
132 |
"pand %[_MMX_00FFw],%%mm4\n" // 0 L2 0 L0 0 l2 0 l0 |
132 |
"pand _MMX_00FFw,%%mm4\n" // 0 L2 0 L0 0 l2 0 l0 |
133 |
"psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1 |
133 |
"psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1 |
134 |
|
134 |
|
135 |
// create R (result in mm6) |
135 |
// create R (result in mm6) |
Lines 146-156
Link Here
|
146 |
"movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0 |
146 |
"movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0 |
147 |
"punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0 |
147 |
"punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0 |
148 |
"punpckldq %%mm1,%%mm1\n" // 00 u1 00 u0 00 u1 00 u0 |
148 |
"punpckldq %%mm1,%%mm1\n" // 00 u1 00 u0 00 u1 00 u0 |
149 |
"psubw %[_MMX_0080w],%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 |
149 |
"psubw _MMX_0080w,%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 |
150 |
// create Cb_g (result in mm5) |
150 |
// create Cb_g (result in mm5) |
151 |
"movq %%mm1,%%mm5\n" // u1 u1 u0 u0 u1 u1 u0 u0 |
151 |
"movq %%mm1,%%mm5\n" // u1 u1 u0 u0 u1 u1 u0 u0 |
152 |
"pmullw %[_MMX_UgrnRGB],%%mm5\n" // blue*-109dec=1.7129*64 |
152 |
"pmullw _MMX_UgrnRGB,%%mm5\n" // blue*-109dec=1.7129*64 |
153 |
"pmullw %[_MMX_UbluRGB],%%mm1\n" // blue*114dec=1.78125*64 |
153 |
"pmullw _MMX_UbluRGB,%%mm1\n" // blue*114dec=1.78125*64 |
154 |
"psraw $6, %%mm5\n" // blue=red/64 |
154 |
"psraw $6, %%mm5\n" // blue=red/64 |
155 |
"psraw $6, %%mm1\n" // blue=blue/64 |
155 |
"psraw $6, %%mm1\n" // blue=blue/64 |
156 |
|
156 |
|
Lines 232-245
Link Here
|
232 |
"popl %%ebx\n" |
232 |
"popl %%ebx\n" |
233 |
: |
233 |
: |
234 |
: "m" (cr), "r"(cb),"r"(lum), |
234 |
: "m" (cr), "r"(cb),"r"(lum), |
235 |
"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod), |
235 |
"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod) |
236 |
[_MMX_0080w] "m" (*MMX_0080w), |
236 |
: "%ebx" |
237 |
[_MMX_00FFw] "m" (*MMX_00FFw), |
|
|
238 |
[_MMX_FF00w] "m" (*MMX_FF00w), |
239 |
[_MMX_VgrnRGB] "m" (*MMX_VgrnRGB), |
240 |
[_MMX_VredRGB] "m" (*MMX_VredRGB), |
241 |
[_MMX_UgrnRGB] "m" (*MMX_UgrnRGB), |
242 |
[_MMX_UbluRGB] "m" (*MMX_UbluRGB) |
243 |
); |
237 |
); |
244 |
} |
238 |
} |
245 |
|
239 |
|
Lines 269-289
Link Here
|
269 |
"movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0 |
263 |
"movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0 |
270 |
"punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0 |
264 |
"punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0 |
271 |
"punpcklbw %%mm7, %%mm1\n" // 4 W cr 0 v3 0 v2 0 v1 0 v0 |
265 |
"punpcklbw %%mm7, %%mm1\n" // 4 W cr 0 v3 0 v2 0 v1 0 v0 |
272 |
"psubw %[_MMX_0080w], %%mm0\n" |
266 |
"psubw _MMX_0080w, %%mm0\n" |
273 |
"psubw %[_MMX_0080w], %%mm1\n" |
267 |
"psubw _MMX_0080w, %%mm1\n" |
274 |
"movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0 |
268 |
"movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0 |
275 |
"movq %%mm1, %%mm3\n" // Cr |
269 |
"movq %%mm1, %%mm3\n" // Cr |
276 |
"pmullw %[_MMX_Ugrn565], %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0 |
270 |
"pmullw _MMX_Ugrn565, %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0 |
277 |
"movq (%2), %%mm6\n" // L1 l7 L6 L5 L4 L3 L2 L1 L0 |
271 |
"movq (%2), %%mm6\n" // L1 l7 L6 L5 L4 L3 L2 L1 L0 |
278 |
"pmullw %[_MMX_Ublu5x5], %%mm0\n" // Cb2blue |
272 |
"pmullw _MMX_Ublu5x5, %%mm0\n" // Cb2blue |
279 |
"pand %[_MMX_00FFw], %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0 |
273 |
"pand _MMX_00FFw, %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0 |
280 |
"pmullw %[_MMX_Vgrn565], %%mm3\n" // Cr2green |
274 |
"pmullw _MMX_Vgrn565, %%mm3\n" // Cr2green |
281 |
"movq (%2), %%mm7\n" // L2 |
275 |
"movq (%2), %%mm7\n" // L2 |
282 |
"pmullw %[_MMX_Vred5x5], %%mm1\n" // Cr2red |
276 |
"pmullw _MMX_Vred5x5, %%mm1\n" // Cr2red |
283 |
"psrlw $8, %%mm7\n" // L2 00 L7 00 L5 00 L3 00 L1 |
277 |
"psrlw $8, %%mm7\n" // L2 00 L7 00 L5 00 L3 00 L1 |
284 |
"pmullw %[_MMX_Ycoeff], %%mm6\n" // lum1 |
278 |
"pmullw _MMX_Ycoeff, %%mm6\n" // lum1 |
285 |
"paddw %%mm3, %%mm2\n" // Cb2green + Cr2green == green |
279 |
"paddw %%mm3, %%mm2\n" // Cb2green + Cr2green == green |
286 |
"pmullw %[_MMX_Ycoeff], %%mm7\n" // lum2 |
280 |
"pmullw _MMX_Ycoeff, %%mm7\n" // lum2 |
287 |
|
281 |
|
288 |
"movq %%mm6, %%mm4\n" // lum1 |
282 |
"movq %%mm6, %%mm4\n" // lum1 |
289 |
"paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0 |
283 |
"paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0 |
Lines 301-311
Link Here
|
301 |
"punpcklbw %%mm4, %%mm4\n" |
295 |
"punpcklbw %%mm4, %%mm4\n" |
302 |
"punpcklbw %%mm5, %%mm5\n" |
296 |
"punpcklbw %%mm5, %%mm5\n" |
303 |
|
297 |
|
304 |
"pand %[_MMX_red565], %%mm4\n" |
298 |
"pand _MMX_red565, %%mm4\n" |
305 |
"psllw $3, %%mm5\n" // GREEN 1 |
299 |
"psllw $3, %%mm5\n" // GREEN 1 |
306 |
"punpcklbw %%mm6, %%mm6\n" |
300 |
"punpcklbw %%mm6, %%mm6\n" |
307 |
"pand %[_MMX_grn565], %%mm5\n" |
301 |
"pand _MMX_grn565, %%mm5\n" |
308 |
"pand %[_MMX_red565], %%mm6\n" |
302 |
"pand _MMX_red565, %%mm6\n" |
309 |
"por %%mm5, %%mm4\n" // |
303 |
"por %%mm5, %%mm4\n" // |
310 |
"psrlw $11, %%mm6\n" // BLUE 1 |
304 |
"psrlw $11, %%mm6\n" // BLUE 1 |
311 |
"movq %%mm3, %%mm5\n" // lum2 |
305 |
"movq %%mm3, %%mm5\n" // lum2 |
Lines 319-341
Link Here
|
319 |
"packuswb %%mm3, %%mm3\n" |
313 |
"packuswb %%mm3, %%mm3\n" |
320 |
"packuswb %%mm5, %%mm5\n" |
314 |
"packuswb %%mm5, %%mm5\n" |
321 |
"packuswb %%mm7, %%mm7\n" |
315 |
"packuswb %%mm7, %%mm7\n" |
322 |
"pand %[_MMX_00FFw], %%mm6\n" // L3 |
316 |
"pand _MMX_00FFw, %%mm6\n" // L3 |
323 |
"punpcklbw %%mm3, %%mm3\n" |
317 |
"punpcklbw %%mm3, %%mm3\n" |
324 |
"punpcklbw %%mm5, %%mm5\n" |
318 |
"punpcklbw %%mm5, %%mm5\n" |
325 |
"pmullw %[_MMX_Ycoeff], %%mm6\n" // lum3 |
319 |
"pmullw _MMX_Ycoeff, %%mm6\n" // lum3 |
326 |
"punpcklbw %%mm7, %%mm7\n" |
320 |
"punpcklbw %%mm7, %%mm7\n" |
327 |
"psllw $3, %%mm5\n" // GREEN 2 |
321 |
"psllw $3, %%mm5\n" // GREEN 2 |
328 |
"pand %[_MMX_red565], %%mm7\n" |
322 |
"pand _MMX_red565, %%mm7\n" |
329 |
"pand %[_MMX_red565], %%mm3\n" |
323 |
"pand _MMX_red565, %%mm3\n" |
330 |
"psrlw $11, %%mm7\n" // BLUE 2 |
324 |
"psrlw $11, %%mm7\n" // BLUE 2 |
331 |
"pand %[_MMX_grn565], %%mm5\n" |
325 |
"pand _MMX_grn565, %%mm5\n" |
332 |
"por %%mm7, %%mm3\n" |
326 |
"por %%mm7, %%mm3\n" |
333 |
"movq (%2,%4), %%mm7\n" // L4 load lum2 |
327 |
"movq (%2,%4), %%mm7\n" // L4 load lum2 |
334 |
"por %%mm5, %%mm3\n" // |
328 |
"por %%mm5, %%mm3\n" // |
335 |
"psrlw $8, %%mm7\n" // L4 |
329 |
"psrlw $8, %%mm7\n" // L4 |
336 |
"movq %%mm4, %%mm5\n" |
330 |
"movq %%mm4, %%mm5\n" |
337 |
"punpcklwd %%mm3, %%mm4\n" |
331 |
"punpcklwd %%mm3, %%mm4\n" |
338 |
"pmullw %[_MMX_Ycoeff], %%mm7\n" // lum4 |
332 |
"pmullw _MMX_Ycoeff, %%mm7\n" // lum4 |
339 |
"punpckhwd %%mm3, %%mm5\n" |
333 |
"punpckhwd %%mm3, %%mm5\n" |
340 |
|
334 |
|
341 |
"movq %%mm4, (%3)\n" // write row1 |
335 |
"movq %%mm4, (%3)\n" // write row1 |
Lines 362-372
Link Here
|
362 |
"punpcklbw %%mm5, %%mm5\n" |
356 |
"punpcklbw %%mm5, %%mm5\n" |
363 |
"punpcklbw %%mm6, %%mm6\n" |
357 |
"punpcklbw %%mm6, %%mm6\n" |
364 |
"psllw $3, %%mm5\n" // GREEN 3 |
358 |
"psllw $3, %%mm5\n" // GREEN 3 |
365 |
"pand %[_MMX_red565], %%mm4\n" |
359 |
"pand _MMX_red565, %%mm4\n" |
366 |
"psraw $6, %%mm3\n" // psr 6 |
360 |
"psraw $6, %%mm3\n" // psr 6 |
367 |
"psraw $6, %%mm0\n" |
361 |
"psraw $6, %%mm0\n" |
368 |
"pand %[_MMX_red565], %%mm6\n" // BLUE |
362 |
"pand _MMX_red565, %%mm6\n" // BLUE |
369 |
"pand %[_MMX_grn565], %%mm5\n" |
363 |
"pand _MMX_grn565, %%mm5\n" |
370 |
"psrlw $11, %%mm6\n" // BLUE 3 |
364 |
"psrlw $11, %%mm6\n" // BLUE 3 |
371 |
"por %%mm5, %%mm4\n" |
365 |
"por %%mm5, %%mm4\n" |
372 |
"psraw $6, %%mm7\n" |
366 |
"psraw $6, %%mm7\n" |
Lines 377-387
Link Here
|
377 |
"punpcklbw %%mm3, %%mm3\n" |
371 |
"punpcklbw %%mm3, %%mm3\n" |
378 |
"punpcklbw %%mm0, %%mm0\n" |
372 |
"punpcklbw %%mm0, %%mm0\n" |
379 |
"punpcklbw %%mm7, %%mm7\n" |
373 |
"punpcklbw %%mm7, %%mm7\n" |
380 |
"pand %[_MMX_red565], %%mm3\n" |
374 |
"pand _MMX_red565, %%mm3\n" |
381 |
"pand %[_MMX_red565], %%mm7\n" // BLUE |
375 |
"pand _MMX_red565, %%mm7\n" // BLUE |
382 |
"psllw $3, %%mm0\n" // GREEN 4 |
376 |
"psllw $3, %%mm0\n" // GREEN 4 |
383 |
"psrlw $11, %%mm7\n" |
377 |
"psrlw $11, %%mm7\n" |
384 |
"pand %[_MMX_grn565], %%mm0\n" |
378 |
"pand _MMX_grn565, %%mm0\n" |
385 |
"por %%mm7, %%mm3\n" |
379 |
"por %%mm7, %%mm3\n" |
386 |
"por %%mm0, %%mm3\n" |
380 |
"por %%mm0, %%mm3\n" |
387 |
|
381 |
|
Lines 413-428
Link Here
|
413 |
"popl %%ebx\n" |
407 |
"popl %%ebx\n" |
414 |
: |
408 |
: |
415 |
:"m" (cr), "r"(cb),"r"(lum), |
409 |
:"m" (cr), "r"(cb),"r"(lum), |
416 |
"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod), |
410 |
"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod) |
417 |
[_MMX_0080w] "m" (*MMX_0080w), |
411 |
: "%ebx" |
418 |
[_MMX_Ugrn565] "m" (*MMX_Ugrn565), |
|
|
419 |
[_MMX_Ublu5x5] "m" (*MMX_Ublu5x5), |
420 |
[_MMX_00FFw] "m" (*MMX_00FFw), |
421 |
[_MMX_Vgrn565] "m" (*MMX_Vgrn565), |
422 |
[_MMX_Vred5x5] "m" (*MMX_Vred5x5), |
423 |
[_MMX_Ycoeff] "m" (*MMX_Ycoeff), |
424 |
[_MMX_red565] "m" (*MMX_red565), |
425 |
[_MMX_grn565] "m" (*MMX_grn565) |
426 |
); |
412 |
); |
427 |
} |
413 |
} |
428 |
|
414 |
|