OLD | NEW |
---|---|
1 /* | |
2 * Copyright 2016 The Android Open Source Project | |
mtklein
2015/06/26 14:05:50
Let's put 2013 (file created) or 2015 (now) here.
yang.zhang
2015/06/29 07:25:57
Done.
| |
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 | |
1 | 8 |
2 #include "SkBlitMask.h" | 9 #include "SkBlitMask.h" |
3 #include "SkColor_opts_neon.h" | 10 #include "SkColor_opts_neon.h" |
4 | 11 |
5 static void D32_A8_Black_neon(void* SK_RESTRICT dst, size_t dstRB, | 12 static void D32_A8_Black_neon(void* SK_RESTRICT dst, size_t dstRB, |
6 const void* SK_RESTRICT maskPtr, size_t maskRB, | 13 const void* SK_RESTRICT maskPtr, size_t maskRB, |
7 SkColor, int width, int height) { | 14 SkColor, int width, int height) { |
8 SkPMColor* SK_RESTRICT device = (SkPMColor*)dst; | 15 SkPMColor* SK_RESTRICT device = (SkPMColor*)dst; |
9 const uint8_t* SK_RESTRICT mask = (const uint8_t*)maskPtr; | 16 const uint8_t* SK_RESTRICT mask = (const uint8_t*)maskPtr; |
10 | 17 |
(...skipping 234 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... | |
245 | 252 |
246 dst += 8; | 253 dst += 8; |
247 src += 8; | 254 src += 8; |
248 width -= 8; | 255 width -= 8; |
249 } | 256 } |
250 | 257 |
251 for (int i = 0; i < width; i++) { | 258 for (int i = 0; i < width; i++) { |
252 dst[i] = SkBlendLCD16(colA, colR, colG, colB, dst[i], src[i]); | 259 dst[i] = SkBlendLCD16(colA, colR, colG, colB, dst[i], src[i]); |
253 } | 260 } |
254 } | 261 } |
262 | |
263 void SkRGB16BlitterBlitV_neon(uint16_t* device, | |
264 int height, | |
265 size_t deviceRB, | |
266 unsigned scale, | |
267 uint32_t src32) { | |
268 uint32x4_t vsrc32, vscale5; | |
mtklein
2015/06/26 14:05:50
Does writing it like this recover any of the slowd
yang.zhang
2015/06/29 07:25:57
Yeah. The setup code may have an effect on the cas
| |
269 uint16x8_t vmaskq_g16, vmaskq_ng16; | |
270 uint16x4_t vmask_g16, vmask_ng16; | |
271 uint16x8_t vdev; | |
272 uint16x8x2_t vdst32; | |
mtklein
2015/06/26 14:05:50
I'd prefer if if you could move the declarations o
yang.zhang
2015/06/29 07:25:57
Done.
| |
273 uint32x4_t vdst32_lo, vdst32_hi; | |
274 uint16x4_t vtmp_lo, vtmp_hi; | |
275 uint16x4_t vdst16_lo, vdst16_hi; | |
276 uint16_t* dst = device; | |
277 | |
278 // prepare constants | |
279 vdev = vdupq_n_u16(0); | |
280 vmaskq_g16 = vdupq_n_u16(SK_G16_MASK_IN_PLACE); | |
mtklein
2015/06/26 14:05:50
Why do we make four masks here when we can use van
yang.zhang
2015/06/29 07:25:57
Done.
| |
281 vmaskq_ng16 = vdupq_n_u16(~SK_G16_MASK_IN_PLACE); | |
282 vmask_g16 = vdup_n_u16(SK_G16_MASK_IN_PLACE); | |
283 vmask_ng16 = vdup_n_u16(~SK_G16_MASK_IN_PLACE); | |
284 vsrc32 = vdupq_n_u32(src32); | |
285 vscale5 = vdupq_n_u32((uint32_t)scale); | |
286 | |
287 while (height >= 8){ | |
288 vdev = vld1q_lane_u16(device, vdev, 0); | |
mtklein
2015/06/26 14:05:50
This code (and the stores) might read more clearly
yang.zhang
2015/06/29 07:25:57
Done.
| |
289 device = (uint16_t*)((char*)device + deviceRB); | |
290 vdev = vld1q_lane_u16(device, vdev, 1); | |
291 device = (uint16_t*)((char*)device + deviceRB); | |
292 vdev = vld1q_lane_u16(device, vdev, 2); | |
293 device = (uint16_t*)((char*)device + deviceRB); | |
294 vdev = vld1q_lane_u16(device, vdev, 3); | |
295 device = (uint16_t*)((char*)device + deviceRB); | |
296 vdev = vld1q_lane_u16(device, vdev, 4); | |
297 device = (uint16_t*)((char*)device + deviceRB); | |
298 vdev = vld1q_lane_u16(device, vdev, 5); | |
299 device = (uint16_t*)((char*)device + deviceRB); | |
300 vdev = vld1q_lane_u16(device, vdev, 6); | |
301 device = (uint16_t*)((char*)device + deviceRB); | |
302 vdev = vld1q_lane_u16(device, vdev, 7); | |
303 device = (uint16_t*)((char*)device + deviceRB); | |
304 | |
305 // Expand_rgb_16 | |
306 vdst32 = vzipq_u16((vdev & vmaskq_ng16), (vdev & vmaskq_g16)); | |
307 vdst32_lo = vmulq_u32(vreinterpretq_u32_u16(vdst32.val[0]), vscale5); | |
308 vdst32_hi = vmulq_u32(vreinterpretq_u32_u16(vdst32.val[1]), vscale5); | |
309 | |
310 // Compact_rgb_16 | |
311 vdst32_lo = vaddq_u32(vdst32_lo, vsrc32); | |
312 vdst32_hi = vaddq_u32(vdst32_hi, vsrc32); | |
313 vdst32_lo = vshrq_n_u32(vdst32_lo, 5); | |
314 vdst32_hi = vshrq_n_u32(vdst32_hi, 5); | |
315 | |
316 vtmp_lo = vmovn_u32(vdst32_lo) & vmask_ng16; | |
317 vtmp_hi = vshrn_n_u32(vdst32_lo, 16) & vmask_g16; | |
318 vdst16_lo = vorr_u16(vtmp_lo, vtmp_hi); | |
319 vtmp_lo = vmovn_u32(vdst32_hi) & vmask_ng16; | |
320 vtmp_hi = vshrn_n_u32(vdst32_hi, 16) & vmask_g16; | |
321 vdst16_hi = vorr_u16(vtmp_lo, vtmp_hi); | |
322 | |
323 vst1_lane_u16(dst, vdst16_lo, 0); | |
324 dst = (uint16_t*)((char*)dst + deviceRB); | |
325 vst1_lane_u16(dst, vdst16_lo, 1); | |
326 dst = (uint16_t*)((char*)dst + deviceRB); | |
327 vst1_lane_u16(dst, vdst16_lo, 2); | |
328 dst = (uint16_t*)((char*)dst + deviceRB); | |
329 vst1_lane_u16(dst, vdst16_lo, 3); | |
330 dst = (uint16_t*)((char*)dst + deviceRB); | |
331 vst1_lane_u16(dst, vdst16_hi, 0); | |
332 dst = (uint16_t*)((char*)dst + deviceRB); | |
333 vst1_lane_u16(dst, vdst16_hi, 1); | |
334 dst = (uint16_t*)((char*)dst + deviceRB); | |
335 vst1_lane_u16(dst, vdst16_hi, 2); | |
336 dst = (uint16_t*)((char*)dst + deviceRB); | |
337 vst1_lane_u16(dst, vdst16_hi, 3); | |
338 dst = (uint16_t*)((char*)dst + deviceRB); | |
339 height -= 8; | |
340 } | |
341 while (height != 0){ | |
342 uint32_t dst32 = SkExpand_rgb_16(*device) * scale; | |
343 *device = SkCompact_rgb_16((src32 + dst32) >> 5); | |
344 device = (uint16_t*)((char*)device + deviceRB); | |
345 height--; | |
346 } | |
347 } | |
348 | |
349 void SkRGB16BlitterBlitH_neon(uint16_t* device, | |
mtklein
2015/06/26 14:05:50
Let's leave this out until it's used?
| |
350 int height, | |
351 unsigned scale, | |
352 uint32_t src32) { | |
353 | |
354 uint32x4_t vsrc32, vscale5; | |
355 uint16x8_t vmaskq_g16, vmaskq_ng16; | |
356 uint16x4_t vmask_g16, vmask_ng16; | |
357 uint16x8_t vdev; | |
358 uint16x8x2_t vdst32; | |
359 uint32x4_t vdst32_lo, vdst32_hi; | |
360 uint16x4_t vtmp_lo, vtmp_hi; | |
361 uint16x4_t vdst16_lo, vdst16_hi; | |
362 | |
363 // prepare constants | |
364 vdev = vdupq_n_u16(0); | |
365 vmaskq_g16 = vdupq_n_u16(SK_G16_MASK_IN_PLACE); | |
366 vmaskq_ng16 = vdupq_n_u16(~SK_G16_MASK_IN_PLACE); | |
367 vmask_g16 = vdup_n_u16(SK_G16_MASK_IN_PLACE); | |
368 vmask_ng16 = vdup_n_u16(~SK_G16_MASK_IN_PLACE); | |
369 vsrc32 = vdupq_n_u32(src32); | |
370 vscale5 = vdupq_n_u32((uint32_t)scale); | |
371 | |
372 while (height >= 8){ | |
373 vdev = vld1q_u16(device); | |
374 | |
375 // Expand_rgb_16 | |
376 vdst32 = vzipq_u16((vdev & vmaskq_ng16), (vdev & vmaskq_g16)); | |
377 vdst32_lo = vmulq_u32(vreinterpretq_u32_u16(vdst32.val[0]), vscale5); | |
378 vdst32_hi = vmulq_u32(vreinterpretq_u32_u16(vdst32.val[1]), vscale5); | |
379 | |
380 // Compact_rgb_16 | |
381 vdst32_lo = vaddq_u32(vdst32_lo, vsrc32); | |
382 vdst32_hi = vaddq_u32(vdst32_hi, vsrc32); | |
383 vdst32_lo = vshrq_n_u32(vdst32_lo, 5); | |
384 vdst32_hi = vshrq_n_u32(vdst32_hi, 5); | |
385 | |
386 vtmp_lo = vmovn_u32(vdst32_lo) & vmask_ng16; | |
387 vtmp_hi = vshrn_n_u32(vdst32_lo, 16) & vmask_g16; | |
388 vdst16_lo = vorr_u16(vtmp_lo, vtmp_hi); | |
389 vtmp_lo = vmovn_u32(vdst32_hi) & vmask_ng16; | |
390 vtmp_hi = vshrn_n_u32(vdst32_hi, 16) & vmask_g16; | |
391 vdst16_hi = vorr_u16(vtmp_lo, vtmp_hi); | |
392 | |
393 vst1q_u16(device, vcombine_u16(vdst16_lo, vdst16_hi)); | |
394 device += 8; | |
395 height -= 8; | |
396 } | |
397 while (height != 0){ | |
398 uint32_t dst32 = SkExpand_rgb_16(*device) * scale; | |
399 *device++ = SkCompact_rgb_16((src32 + dst32) >> 5); | |
400 height--; | |
401 } | |
402 } | |
OLD | NEW |