OLD | NEW |
| (Empty) |
1 /* | |
2 * Copyright 2013 The Android Open Source Project | |
3 * | |
4 * Use of this source code is governed by a BSD-style license that can be | |
5 * found in the LICENSE file. | |
6 */ | |
7 | |
8 #if !defined(_MSC_VER) | |
9 | |
10 #define EXTRACT_ALPHA(var1, var2) \ | |
11 movdqa %var1, %var2; /* Clone source pixels to extract alpha
*/\ | |
12 psrlw $8, %var2; /* Discard red and blue, leaving alpha a
nd green */\ | |
13 pshufhw $0xF5, %var2, %var2; /* Repeat alpha for scaling (high) */\ | |
14 movdqa %xmm6, %xmm4; \ | |
15 pshuflw $0xF5, %var2, %var2; /* Repeat alpha for scaling (low) */\ | |
16 movdqa %xmm5, %xmm3; \ | |
17 psubw %var2, %xmm4 /* Finalize alpha calculations */ | |
18 | |
19 #define SCALE_PIXELS \ | |
20 psllw $8, %xmm5; /* Filter out red and blue components */
\ | |
21 pmulhuw %xmm4, %xmm5; /* Scale red and blue */\ | |
22 psrlw $8, %xmm3; /* Filter out alpha and green components
*/\ | |
23 pmullw %xmm4, %xmm3 /* Scale alpha and green */ | |
24 | |
25 | |
26 /* | |
27 * void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst, | |
28 * const SkPMColor* SK_RESTRICT src, | |
29 * int count, U8CPU alpha) | |
30 * | |
31 * This function is divided into six blocks: initialization, blit 4-15 pixels, | |
32 * blit 0-3 pixels, align destination for 16+ pixel blits, | |
33 * blit 16+ pixels with source unaligned, blit 16+ pixels with source aligned. | |
34 * There are some code reuse between the blocks. | |
35 * | |
36 * The primary optimization comes from checking the source pixels' alpha value. | |
37 * If the alpha is zero, the pixel can be skipped entirely. | |
38 * If the alpha is fully opaque, the pixel can be copied directly to the destina
tion. | |
39 * According to collected statistics, these two cases are the most common. | |
40 * The main loop(s) uses pre-loading and unrolling in an attempt to reduce the | |
41 * memory latency worse-case. | |
42 */ | |
43 | |
44 #ifdef __clang__ | |
45 .text | |
46 .global _S32A_Opaque_BlitRow32_SSE4_asm | |
47 #else | |
48 .section .text.sse4.2,"ax",@progbits | |
49 .type S32A_Opaque_BlitRow32_SSE4_asm, @function | |
50 .global S32A_Opaque_BlitRow32_SSE4_asm | |
51 #endif | |
52 | |
53 .p2align 4 | |
54 _S32A_Opaque_BlitRow32_SSE4_asm: | |
55 S32A_Opaque_BlitRow32_SSE4_asm: | |
56 .cfi_startproc | |
57 prefetcht0 (%rsi) | |
58 movl %edx, %ecx // Pixel count | |
59 movq %rdi, %rdx // Destination pointer | |
60 movq %rsi, %rax // Source pointer | |
61 | |
62 // Setup SSE constants | |
63 movdqa .LAlphaCheckMask(%rip), %xmm7 // 0xFF000000 mask to check alpha | |
64 movdqa .LInverseAlphaCalc(%rip), %xmm6// 16-bit 256 to calculate inv. a
lpha | |
65 movdqa .LResultMergeMask(%rip), %xmm0 // 0x00FF00FF mask (Must be in xm
m0 because of pblendvb) | |
66 | |
67 subl $4, %ecx // Check if we have only 0-3 pixels | |
68 js .LReallySmall | |
69 cmpl $11, %ecx // Do we have enough pixels to run the m
ain loop? | |
70 ja .LBigBlit | |
71 | |
72 // Handle small blits (4-15 pixels) | |
73 ////////////////////////////////////////////////////////////////////////////
//// | |
74 xorq %rdi, %rdi // Reset offset to zero | |
75 | |
76 .LSmallLoop: | |
77 lddqu (%rax, %rdi), %xmm1 // Load four source pixels | |
78 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
79 ja .LSmallAlphaNotOpaqueOrZero | |
80 jz .LSmallAlphaZero | |
81 movdqu %xmm1, (%rdx, %rdi) // Store four destination pixels | |
82 .LSmallAlphaZero: | |
83 addq $16, %rdi | |
84 subl $4, %ecx // Check if there are four additional pi
xels, at least | |
85 jns .LSmallLoop | |
86 jmp .LSmallRemaining | |
87 | |
88 // Handle mixed alphas (calculate and scale) | |
89 .p2align 4 | |
90 .LSmallAlphaNotOpaqueOrZero: | |
91 lddqu (%rdx, %rdi), %xmm5 // Load four destination pixels | |
92 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
93 SCALE_PIXELS // Scale pixels using alpha | |
94 | |
95 addq $16, %rdi | |
96 subl $4, %ecx // Check if we can store all four pixels | |
97 pblendvb %xmm0, %xmm5, %xmm3 | |
98 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
99 movdqu %xmm1, -16(%rdx, %rdi) // Store four destination pixels | |
100 jns .LSmallLoop | |
101 | |
102 // Handle the last 0-3 pixels (also used by the big unaligned loop) | |
103 .LSmallRemaining: | |
104 cmpl $-4, %ecx // Check if we are done | |
105 je .LSmallExit | |
106 sall $2, %ecx // Calculate offset for last pixels | |
107 movslq %ecx, %rcx | |
108 addq %rcx, %rdi | |
109 | |
110 lddqu (%rax, %rdi), %xmm1 // Load last four source pixels (overlap
ping) | |
111 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
112 jc .LSmallRemainingStoreAll// If all alphas are opaque, just store
(overlapping) | |
113 jz .LSmallExit // If all alphas are zero, skip the pixe
ls completely | |
114 | |
115 // Handle mixed alphas (calculate and scale) | |
116 lddqu (%rdx, %rdi), %xmm5 // Load last four destination pixels (ov
erlapping) | |
117 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
118 | |
119 psllw $8, %xmm3 // Filter out red and blue components | |
120 pmulhuw %xmm4, %xmm3 // Scale red and blue | |
121 movdqa %xmm5, %xmm2 | |
122 psrlw $8, %xmm2 // Filter out alpha and green components | |
123 pmullw %xmm4, %xmm2 // Scale alpha and green | |
124 | |
125 cmpl $-8, %ecx // Check how many pixels should be writt
en | |
126 pblendvb %xmm0, %xmm3, %xmm2 // Combine results | |
127 paddb %xmm2, %xmm1 // Add source and destination pixels tog
ether | |
128 jb .LSmallPixelsLeft1 | |
129 ja .LSmallPixelsLeft3 // To avoid double-blending the overlapp
ing pixels... | |
130 pblendw $0xF0, %xmm1, %xmm5 // Merge only the final two pixels to th
e destination | |
131 movdqu %xmm5, (%rdx, %rdi) // Store last two destination pixels | |
132 .LSmallExit: | |
133 ret | |
134 | |
135 .LSmallPixelsLeft1: | |
136 pblendw $0xC0, %xmm1, %xmm5 // Merge only the final pixel to the des
tination | |
137 movdqu %xmm5, (%rdx, %rdi) // Store last destination pixel | |
138 ret | |
139 | |
140 .LSmallPixelsLeft3: | |
141 pblendw $0xFC, %xmm1, %xmm5 // Merge only the final three pixels to
the destination | |
142 movdqu %xmm5, (%rdx, %rdi) // Store last three destination pixels | |
143 ret | |
144 | |
145 .LSmallRemainingStoreAll: | |
146 movdqu %xmm1, (%rdx, %rdi) // Store last destination pixels (overwr
ite) | |
147 ret | |
148 | |
149 // Handle really small blits (0-3 pixels) | |
150 ////////////////////////////////////////////////////////////////////////////
//// | |
151 .LReallySmall: | |
152 addl $4, %ecx | |
153 jle .LReallySmallExit | |
154 pcmpeqd %xmm1, %xmm1 | |
155 cmpl $2, %ecx // Check how many pixels should be read | |
156 pinsrd $0x0, (%rax), %xmm1 // Load one source pixel | |
157 pinsrd $0x0, (%rdx), %xmm5 // Load one destination pixel | |
158 jb .LReallySmallCalc | |
159 pinsrd $0x1, 4(%rax), %xmm1 // Load second source pixel | |
160 pinsrd $0x1, 4(%rdx), %xmm5 // Load second destination pixel | |
161 je .LReallySmallCalc | |
162 pinsrd $0x2, 8(%rax), %xmm1 // Load third source pixel | |
163 pinsrd $0x2, 8(%rdx), %xmm5 // Load third destination pixel | |
164 | |
165 .LReallySmallCalc: | |
166 ptest %xmm7, %xmm1 // Check if all alphas are opaque | |
167 jc .LReallySmallStore // If all alphas are opaque, just store | |
168 | |
169 // Handle mixed alphas (calculate and scale) | |
170 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
171 | |
172 pand %xmm0, %xmm5 // Filter out red and blue components | |
173 pmullw %xmm4, %xmm5 // Scale red and blue | |
174 psrlw $8, %xmm3 // Filter out alpha and green components | |
175 pmullw %xmm4, %xmm3 // Scale alpha and green | |
176 | |
177 psrlw $8, %xmm5 // Combine results | |
178 pblendvb %xmm0, %xmm5, %xmm3 | |
179 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
180 | |
181 .LReallySmallStore: | |
182 cmpl $2, %ecx // Check how many pixels should be writt
en | |
183 pextrd $0x0, %xmm1, (%rdx) // Store one destination pixel | |
184 jb .LReallySmallExit | |
185 pextrd $0x1, %xmm1, 4(%rdx) // Store second destination pixel | |
186 je .LReallySmallExit | |
187 pextrd $0x2, %xmm1, 8(%rdx) // Store third destination pixel | |
188 .LReallySmallExit: | |
189 ret | |
190 | |
191 // Handle bigger blit operations (16+ pixels) | |
192 ////////////////////////////////////////////////////////////////////////////
//// | |
193 .p2align 4 | |
194 .LBigBlit: | |
195 // Align destination? | |
196 testl $0xF, %edx | |
197 lddqu (%rax), %xmm1 // Pre-load four source pixels | |
198 jz .LAligned | |
199 | |
200 movq %rdx, %rdi // Calculate alignment of destination po
inter | |
201 negq %rdi | |
202 andl $0xF, %edi | |
203 | |
204 // Handle 1-3 pixels to align destination | |
205 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
206 jz .LAlignDone // If all alphas are zero, just skip | |
207 lddqu (%rdx), %xmm5 // Load four destination pixels | |
208 jc .LAlignStore // If all alphas are opaque, just store | |
209 | |
210 // Handle mixed alphas (calculate and scale) | |
211 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
212 | |
213 psllw $8, %xmm3 // Filter out red and blue components | |
214 pmulhuw %xmm4, %xmm3 // Scale red and blue | |
215 movdqa %xmm5, %xmm2 | |
216 psrlw $8, %xmm2 // Filter out alpha and green components | |
217 pmullw %xmm4, %xmm2 // Scale alpha and green | |
218 | |
219 pblendvb %xmm0, %xmm3, %xmm2 // Combine results | |
220 paddb %xmm2, %xmm1 // Add source and destination pixels tog
ether | |
221 | |
222 .LAlignStore: | |
223 cmpl $8, %edi // Check how many pixels should be writt
en | |
224 jb .LAlignPixelsLeft1 | |
225 ja .LAlignPixelsLeft3 | |
226 pblendw $0x0F, %xmm1, %xmm5 // Blend two pixels | |
227 jmp .LAlignStorePixels | |
228 | |
229 .LAlignPixelsLeft1: | |
230 pblendw $0x03, %xmm1, %xmm5 // Blend one pixel | |
231 jmp .LAlignStorePixels | |
232 | |
233 .LAlignPixelsLeft3: | |
234 pblendw $0x3F, %xmm1, %xmm5 // Blend three pixels | |
235 | |
236 .LAlignStorePixels: | |
237 movdqu %xmm5, (%rdx) // Store destination pixels | |
238 | |
239 .LAlignDone: | |
240 addq %rdi, %rax // Adjust pointers and pixel count | |
241 addq %rdi, %rdx | |
242 shrq $2, %rdi | |
243 lddqu (%rax), %xmm1 // Pre-load new source pixels (after ali
gnment) | |
244 subl %edi, %ecx | |
245 | |
246 .LAligned: // Destination is guaranteed to be 16 by
te aligned | |
247 xorq %rdi, %rdi // Reset offset to zero | |
248 subl $8, %ecx // Decrease counter (Reserve four pixels
for the cleanup) | |
249 testl $0xF, %eax // Check alignment of source pointer | |
250 jz .LAlignedLoop | |
251 | |
252 // Source not aligned to destination | |
253 ////////////////////////////////////////////////////////////////////////////
//// | |
254 .p2align 4 | |
255 .LUnalignedLoop: // Main loop for unaligned, handles eigh
t pixels per iteration | |
256 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
257 ja .LAlphaNotOpaqueOrZero00 | |
258 lddqu 16(%rax, %rdi), %xmm2 // Pre-load four source pixels | |
259 jz .LAlphaZero00 | |
260 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
261 | |
262 .LAlphaZero00: | |
263 ptest %xmm7, %xmm2 // Check if all alphas are zero or opaqu
e | |
264 ja .LAlphaNotOpaqueOrZero01 | |
265 lddqu 32(%rax, %rdi), %xmm1 // Pre-load four source pixels | |
266 jz .LAlphaZero01 | |
267 movdqa %xmm2, 16(%rdx, %rdi) // Store four destination pixels | |
268 | |
269 .LAlphaZero01: | |
270 addq $32, %rdi // Adjust offset and pixel count | |
271 subl $8, %ecx | |
272 jae .LUnalignedLoop | |
273 addl $8, %ecx // Adjust pixel count | |
274 jmp .LLoopCleanup0 | |
275 | |
276 .p2align 4 | |
277 .LAlphaNotOpaqueOrZero00: | |
278 movdqa (%rdx, %rdi), %xmm5 // Load four destination pixels | |
279 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
280 SCALE_PIXELS // Scale pixels using alpha | |
281 | |
282 lddqu 16(%rax, %rdi), %xmm2 // Pre-load four source pixels | |
283 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
284 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
285 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
286 | |
287 // Handle next four pixels | |
288 ptest %xmm7, %xmm2 // Check if all alphas are zero or opaqu
e | |
289 ja .LAlphaNotOpaqueOrZero01 | |
290 lddqu 32(%rax, %rdi), %xmm1 // Pre-load four source pixels | |
291 jz .LAlphaZero02 | |
292 movdqa %xmm2, 16(%rdx, %rdi) // Store four destination pixels | |
293 .LAlphaZero02: | |
294 addq $32, %rdi // Adjust offset and pixel count | |
295 subl $8, %ecx | |
296 jae .LUnalignedLoop | |
297 addl $8, %ecx // Adjust pixel count | |
298 jmp .LLoopCleanup0 | |
299 | |
300 .p2align 4 | |
301 .LAlphaNotOpaqueOrZero01: | |
302 movdqa 16(%rdx, %rdi), %xmm5 // Load four destination pixels | |
303 EXTRACT_ALPHA(xmm2, xmm1) // Extract and clone alpha value | |
304 SCALE_PIXELS // Scale pixels using alpha | |
305 | |
306 lddqu 32(%rax, %rdi), %xmm1 // Pre-load four source pixels | |
307 addq $32, %rdi | |
308 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
309 paddb %xmm3, %xmm2 // Add source and destination pixels tog
ether | |
310 subl $8, %ecx | |
311 movdqa %xmm2, -16(%rdx, %rdi) // Store four destination pixels | |
312 jae .LUnalignedLoop | |
313 addl $8, %ecx // Adjust pixel count | |
314 | |
315 // Cleanup - handle pending pixels from loop | |
316 .LLoopCleanup0: | |
317 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
318 ja .LAlphaNotOpaqueOrZero02 | |
319 jz .LAlphaZero03 | |
320 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
321 .LAlphaZero03: | |
322 addq $16, %rdi | |
323 subl $4, %ecx | |
324 js .LSmallRemaining // Reuse code from small loop | |
325 lddqu (%rax, %rdi), %xmm1 // Pre-load four source pixels | |
326 jmp .LLoopCleanup0 | |
327 | |
328 .LAlphaNotOpaqueOrZero02: | |
329 movdqa (%rdx, %rdi), %xmm5 // Load four destination pixels | |
330 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
331 SCALE_PIXELS // Scale pixels using alpha | |
332 | |
333 addq $16, %rdi | |
334 subl $4, %ecx | |
335 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
336 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
337 movdqa %xmm1, -16(%rdx, %rdi) // Store four destination pixels | |
338 js .LSmallRemaining // Reuse code from small loop | |
339 lddqu (%rax, %rdi), %xmm1 // Pre-load four source pixels | |
340 jmp .LLoopCleanup0 | |
341 | |
342 // Source aligned to destination | |
343 ////////////////////////////////////////////////////////////////////////////
//// | |
344 .p2align 4 | |
345 .LAlignedLoop: // Main loop for aligned, handles eight
pixels per iteration | |
346 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
347 ja .LAlphaNotOpaqueOrZero10 | |
348 movdqa 16(%rax, %rdi), %xmm2 // Pre-load four source pixels | |
349 jz .LAlphaZero10 | |
350 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
351 | |
352 .LAlphaZero10: | |
353 ptest %xmm7, %xmm2 // Check if all alphas are zero or opaqu
e | |
354 ja .LAlphaNotOpaqueOrZero11 | |
355 movdqa 32(%rax, %rdi), %xmm1 // Pre-load four source pixels | |
356 jz .LAlphaZero11 | |
357 movdqa %xmm2, 16(%rdx, %rdi) // Store four destination pixels | |
358 | |
359 .LAlphaZero11: | |
360 addq $32, %rdi // Adjust offset and pixel count | |
361 subl $8, %ecx | |
362 jae .LAlignedLoop | |
363 jmp .LLoopCleanup1 | |
364 | |
365 .p2align 4 | |
366 .LAlphaNotOpaqueOrZero10: | |
367 movdqa (%rdx, %rdi), %xmm5 // Load four destination pixels | |
368 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
369 SCALE_PIXELS // Scale pixels using alpha | |
370 | |
371 movdqa 16(%rax, %rdi), %xmm2 // Pre-load four source pixels | |
372 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
373 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
374 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
375 | |
376 // Handle next four pixels | |
377 ptest %xmm7, %xmm2 // Check if all alphas are zero or opaqu
e | |
378 ja .LAlphaNotOpaqueOrZero11 | |
379 movdqa 32(%rax, %rdi), %xmm1 // Pre-load four source pixels | |
380 jz .LAlphaZero12 | |
381 movdqa %xmm2, 16(%rdx, %rdi) // Store four destination pixels | |
382 .LAlphaZero12: | |
383 addq $32, %rdi // Adjust offset and pixel count | |
384 subl $8, %ecx | |
385 jae .LAlignedLoop | |
386 jmp .LLoopCleanup1 | |
387 | |
388 .p2align 4 | |
389 .LAlphaNotOpaqueOrZero11: | |
390 movdqa 16(%rdx, %rdi), %xmm5 // Load four destination pixels | |
391 EXTRACT_ALPHA(xmm2, xmm1) // Extract and clone alpha value | |
392 SCALE_PIXELS // Scale pixels using alpha | |
393 movdqa 32(%rax, %rdi), %xmm1 // Pre-load four source pixels | |
394 | |
395 addq $32, %rdi | |
396 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
397 paddb %xmm3, %xmm2 // Add source and destination pixels tog
ether | |
398 subl $8, %ecx | |
399 movdqa %xmm2, -16(%rdx, %rdi) // Store four destination pixels | |
400 jae .LAlignedLoop | |
401 | |
402 // Cleanup - handle four pending pixels from loop | |
403 .LLoopCleanup1: | |
404 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
405 ja .LAlphaNotOpaqueOrZero12 | |
406 jz .LAlphaZero13 | |
407 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
408 .LAlphaZero13: | |
409 addl $8, %ecx // Adjust offset and pixel count | |
410 jz .LExit | |
411 addq $16, %rdi | |
412 jmp .LRemainLoop1 | |
413 | |
414 .LAlphaNotOpaqueOrZero12: | |
415 movdqa (%rdx, %rdi), %xmm5 // Load four destination pixels | |
416 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
417 SCALE_PIXELS // Scale pixels using alpha | |
418 | |
419 addl $8, %ecx // Adjust offset and pixel count | |
420 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
421 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
422 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
423 jz .LExit | |
424 addq $16, %rdi | |
425 | |
426 // Handle last 1-7 pixels | |
427 .LRemainLoop1: | |
428 movdqa (%rax, %rdi), %xmm1 // Load four source pixels | |
429 ptest %xmm7, %xmm1 // Check if all alphas are zero or opaqu
e | |
430 ja .LRemainAlphaNotOpaqueOrZero1 | |
431 jz .LRemainAlphaZero1 | |
432 | |
433 // All alphas were opaque (copy) | |
434 subl $4, %ecx // Check if we have more than four pixel
s left | |
435 jle .LRemainStore | |
436 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
437 addq $16, %rdi | |
438 jmp .LRemainLoop1 | |
439 | |
440 // All alphas were zero (skip) | |
441 .p2align 4 | |
442 .LRemainAlphaZero1: | |
443 subl $4, %ecx // Check if we have more than four pixel
s left | |
444 jle .LExit | |
445 addq $16, %rdi | |
446 jmp .LRemainLoop1 | |
447 | |
448 // Handle mixed alphas (calculate and scale) | |
449 .p2align 4 | |
450 .LRemainAlphaNotOpaqueOrZero1: | |
451 movdqa (%rdx, %rdi), %xmm5 // Load four destination pixels | |
452 EXTRACT_ALPHA(xmm1, xmm2) // Extract and clone alpha value | |
453 SCALE_PIXELS // Scale pixels using alpha | |
454 | |
455 subl $4, %ecx | |
456 pblendvb %xmm0, %xmm5, %xmm3 // Combine results | |
457 paddb %xmm3, %xmm1 // Add source and destination pixels tog
ether | |
458 jle .LRemainStore | |
459 movdqa %xmm1, (%rdx, %rdi) // Store four destination pixels | |
460 addq $16, %rdi | |
461 jmp .LRemainLoop1 | |
462 | |
463 // Store the last 1-4 pixels | |
464 .p2align 4 | |
465 .LRemainStore: | |
466 jz .LRemainFull | |
467 movdqa (%rdx, %rdi), %xmm5 // Load four destination pixels | |
468 cmpl $-2, %ecx // Check how many pixels should be writt
en | |
469 jb .LRemainPixelsLeft11 | |
470 ja .LRemainPixelsLeft13 | |
471 pblendw $0x0F, %xmm1, %xmm5 | |
472 movdqa %xmm5, (%rdx, %rdi) // Store last 2 destination pixels | |
473 .LExit: | |
474 ret | |
475 | |
476 .LRemainPixelsLeft11: | |
477 pblendw $0x03, %xmm1, %xmm5 | |
478 movdqa %xmm5, (%rdx, %rdi) // Store last destination pixel | |
479 ret | |
480 | |
481 .LRemainPixelsLeft13: | |
482 pblendw $0x3F, %xmm1, %xmm5 | |
483 movdqa %xmm5, (%rdx, %rdi) // Store last 3 destination pixels | |
484 ret | |
485 | |
486 .LRemainFull: | |
487 movdqa %xmm1, (%rdx, %rdi) // Store last 4 destination pixels | |
488 ret | |
489 | |
490 .cfi_endproc | |
491 #ifndef __clang__ | |
492 .size S32A_Opaque_BlitRow32_SSE4_asm, .-S32A_Opaque_BlitRow32_SSE4_asm | |
493 #endif | |
494 | |
495 // Constants for SSE code | |
496 .pushsection .rodata | |
497 .p2align 4 | |
498 .LAlphaCheckMask: | |
499 .long 0xFF000000, 0xFF000000, 0xFF000000, 0xFF000000 | |
500 .LInverseAlphaCalc: | |
501 .word 256, 256, 256, 256, 256, 256, 256, 256 | |
502 .LResultMergeMask: | |
503 .long 0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF | |
504 .popsection | |
505 #endif | |
OLD | NEW |