2 * Copyright 2012 The LibYuv Project Authors. All rights reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
11 #include "libyuv/row.h"
13 #include "libyuv/basic_types.h"
20 // TODO(fbarchard): Consider 'any' functions handling any quantity of pixels.
21 // TODO(fbarchard): Consider 'any' functions handling odd alignment.
22 // YUV to RGB does multiple of 8 with SIMD and remainder with C.
23 #define YANY(NAMEANY, I420TORGB_SIMD, I420TORGB_C, UV_SHIFT, BPP, MASK) \
24 void NAMEANY(const uint8* y_buf, \
29 int n = width & ~MASK; \
30 I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n); \
31 I420TORGB_C(y_buf + n, \
32 u_buf + (n >> UV_SHIFT), \
33 v_buf + (n >> UV_SHIFT), \
34 rgb_buf + n * BPP, width & MASK); \
37 #ifdef HAS_I422TOARGBROW_SSSE3
38 YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C,
40 YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C,
42 YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C,
44 YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C,
46 YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C,
48 YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C,
50 // I422ToRGB565Row_SSSE3 is unaligned.
51 YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
53 YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
55 YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
57 // I422ToRGB24Row_SSSE3 is unaligned.
58 YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
59 YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
60 YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
61 YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
62 #endif // HAS_I422TOARGBROW_SSSE3
63 #ifdef HAS_I422TOARGBROW_AVX2
64 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
65 #endif // HAS_I422TOARGBROW_AVX2
66 #ifdef HAS_I422TOARGBROW_NEON
67 YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
68 YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
69 YANY(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, I411ToARGBRow_C, 2, 4, 7)
70 YANY(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, I422ToBGRARow_C, 1, 4, 7)
71 YANY(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, I422ToABGRRow_C, 1, 4, 7)
72 YANY(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, I422ToRGBARow_C, 1, 4, 7)
73 YANY(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, I422ToRGB24Row_C, 1, 3, 7)
74 YANY(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, I422ToRAWRow_C, 1, 3, 7)
75 YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
77 YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
79 YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
80 YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
81 YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
82 #endif // HAS_I422TOARGBROW_NEON
85 // Wrappers to handle odd width
86 #define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP) \
87 void NAMEANY(const uint8* y_buf, \
88 const uint8* uv_buf, \
92 NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n); \
93 NV12TORGB_C(y_buf + n, \
94 uv_buf + (n >> UV_SHIFT), \
95 rgb_buf + n * BPP, width & 7); \
98 #ifdef HAS_NV12TOARGBROW_SSSE3
99 NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C,
101 NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C,
103 #endif // HAS_NV12TOARGBROW_SSSE3
104 #ifdef HAS_NV12TOARGBROW_NEON
105 NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
106 NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
107 #endif // HAS_NV12TOARGBROW_NEON
108 #ifdef HAS_NV12TORGB565ROW_SSSE3
109 NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
111 NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
113 #endif // HAS_NV12TORGB565ROW_SSSE3
114 #ifdef HAS_NV12TORGB565ROW_NEON
115 NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
116 NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
117 #endif // HAS_NV12TORGB565ROW_NEON
120 #define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP) \
121 void NAMEANY(const uint8* src, \
124 int n = width & ~MASK; \
125 ARGBTORGB_SIMD(src, dst, n); \
126 ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK); \
129 #if defined(HAS_ARGBTORGB24ROW_SSSE3)
130 RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
132 RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
134 RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
136 RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
138 RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
140 RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C,
142 RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
144 RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C,
146 RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C,
148 // These require alignment on ARGB, so C is used for remainder.
149 RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
151 RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
153 RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
155 RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
157 RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
160 #if defined(HAS_ARGBTORGB24ROW_NEON)
161 RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
162 RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
163 RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
165 RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
167 RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
169 RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
171 RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
173 RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
175 RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
180 // ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
181 #define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP) \
182 void NAMEANY(const uint8* src, \
183 uint8* dst, uint32 selector, \
185 int n = width & ~MASK; \
186 ARGBTORGB_SIMD(src, dst, selector, n); \
187 ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK); \
190 #if defined(HAS_ARGBTOBAYERROW_SSSE3)
191 BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
194 #if defined(HAS_ARGBTOBAYERROW_NEON)
195 BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
200 // RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
201 #define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM) \
202 void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) { \
203 ARGBTOY_SIMD(src_argb, dst_y, width - NUM); \
204 ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP, \
205 dst_y + (width - NUM) * BPP, NUM); \
208 #ifdef HAS_ARGBTOYROW_AVX2
209 YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
210 YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
211 YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
212 YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
214 #ifdef HAS_ARGBTOYROW_SSSE3
215 YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16)
216 YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16)
217 YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16)
218 YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16)
219 YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16)
220 YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16)
222 #ifdef HAS_ARGBTOYJROW_SSSE3
223 YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16)
225 #ifdef HAS_ARGBTOYROW_NEON
226 YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
227 YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
228 YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
229 YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
230 YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
231 YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
232 YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
233 YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
234 YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
235 YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
236 YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
237 YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
238 YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
239 YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
240 YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
241 YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
242 YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
246 #define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK) \
247 void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) { \
248 int n = width & ~MASK; \
249 ARGBTOY_SIMD(src_argb, dst_y, n); \
250 ARGBTOY_C(src_argb + n * SBPP, \
251 dst_y + n * BPP, width & MASK); \
254 // Attenuate is destructive so last16 method can not be used due to overlap.
255 #ifdef HAS_ARGBATTENUATEROW_SSSE3
256 YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
259 #ifdef HAS_ARGBATTENUATEROW_SSE2
260 YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
263 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
264 YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
267 #ifdef HAS_ARGBATTENUATEROW_AVX2
268 YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
271 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
272 YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
275 #ifdef HAS_ARGBATTENUATEROW_NEON
276 YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
281 // RGB/YUV to UV does multiple of 16 with SIMD and remainder with C.
282 #define UVANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK) \
283 void NAMEANY(const uint8* src_argb, int src_stride_argb, \
284 uint8* dst_u, uint8* dst_v, int width) { \
285 int n = width & ~MASK; \
286 ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n); \
287 ANYTOUV_C(src_argb + n * BPP, src_stride_argb, \
293 #ifdef HAS_ARGBTOYROW_AVX2
294 UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
295 UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
296 UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
298 #ifdef HAS_ARGBTOUVROW_SSSE3
299 UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15)
300 UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C,
302 UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15)
303 UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15)
304 UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15)
305 UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15)
306 UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15)
308 #ifdef HAS_ARGBTOUVROW_NEON
309 UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
310 UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
311 UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
312 UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
313 UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
314 UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
315 UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
316 UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
317 UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
318 UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
319 UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15)
320 UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
324 #define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT) \
325 void NAMEANY(const uint8* src_uv, \
326 uint8* dst_u, uint8* dst_v, int width) { \
327 int n = width & ~MASK; \
328 ANYTOUV_SIMD(src_uv, dst_u, dst_v, n); \
329 ANYTOUV_C(src_uv + n * BPP, \
330 dst_u + (n >> SHIFT), \
331 dst_v + (n >> SHIFT), \
335 #ifdef HAS_ARGBTOUV444ROW_SSSE3
336 UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3,
337 ARGBToUV444Row_C, 4, 15, 0)
339 #ifdef HAS_YUY2TOUV422ROW_AVX2
340 UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
341 YUY2ToUV422Row_C, 2, 31, 1)
342 UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
343 UYVYToUV422Row_C, 2, 31, 1)
345 #ifdef HAS_ARGBTOUVROW_SSSE3
346 UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3,
347 ARGBToUV422Row_C, 4, 15, 1)
348 UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2,
349 YUY2ToUV422Row_C, 2, 15, 1)
350 UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2,
351 UYVYToUV422Row_C, 2, 15, 1)
353 #ifdef HAS_YUY2TOUV422ROW_NEON
354 UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
355 ARGBToUV444Row_C, 4, 7, 0)
356 UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
357 ARGBToUV422Row_C, 4, 15, 1)
358 UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
359 ARGBToUV411Row_C, 4, 31, 2)
360 UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
361 YUY2ToUV422Row_C, 2, 15, 1)
362 UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
363 UYVYToUV422Row_C, 2, 15, 1)
367 #define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK) \
368 void NAMEANY(const uint8* src_uv, \
369 uint8* dst_u, uint8* dst_v, int width) { \
370 int n = width & ~MASK; \
371 ANYTOUV_SIMD(src_uv, dst_u, dst_v, n); \
372 ANYTOUV_C(src_uv + n * 2, \
378 #ifdef HAS_SPLITUVROW_SSE2
379 SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15)
381 #ifdef HAS_SPLITUVROW_AVX2
382 SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
384 #ifdef HAS_SPLITUVROW_NEON
385 SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
387 #ifdef HAS_SPLITUVROW_MIPS_DSPR2
388 SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
393 #define MERGEUVROW_ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK) \
394 void NAMEANY(const uint8* src_u, const uint8* src_v, \
395 uint8* dst_uv, int width) { \
396 int n = width & ~MASK; \
397 ANYTOUV_SIMD(src_u, src_v, dst_uv, n); \
398 ANYTOUV_C(src_u + n, \
404 #ifdef HAS_MERGEUVROW_SSE2
405 MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15)
407 #ifdef HAS_MERGEUVROW_AVX2
408 MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
410 #ifdef HAS_MERGEUVROW_NEON
411 MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
413 #undef MERGEUVROW_ANY
415 #define MATHROW_ANY(NAMEANY, ARGBMATH_SIMD, ARGBMATH_C, MASK) \
416 void NAMEANY(const uint8* src_argb0, const uint8* src_argb1, \
417 uint8* dst_argb, int width) { \
418 int n = width & ~MASK; \
419 ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n); \
420 ARGBMATH_C(src_argb0 + n * 4, \
426 #ifdef HAS_ARGBMULTIPLYROW_SSE2
427 MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
430 #ifdef HAS_ARGBADDROW_SSE2
431 MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
433 #ifdef HAS_ARGBSUBTRACTROW_SSE2
434 MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
437 #ifdef HAS_ARGBMULTIPLYROW_AVX2
438 MATHROW_ANY(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, ARGBMultiplyRow_C,
441 #ifdef HAS_ARGBADDROW_AVX2
442 MATHROW_ANY(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, ARGBAddRow_C, 7)
444 #ifdef HAS_ARGBSUBTRACTROW_AVX2
445 MATHROW_ANY(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, ARGBSubtractRow_C,
448 #ifdef HAS_ARGBMULTIPLYROW_NEON
449 MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
452 #ifdef HAS_ARGBADDROW_NEON
453 MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
455 #ifdef HAS_ARGBSUBTRACTROW_NEON
456 MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
461 // Shuffle may want to work in place, so last16 method can not be used.
462 #define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK) \
463 void NAMEANY(const uint8* src_argb, uint8* dst_argb, \
464 const uint8* shuffler, int width) { \
465 int n = width & ~MASK; \
466 ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n); \
467 ARGBTOY_C(src_argb + n * SBPP, \
468 dst_argb + n * BPP, shuffler, width & MASK); \
471 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
472 YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3,
473 ARGBShuffleRow_C, 4, 4, 7)
475 #ifdef HAS_ARGBSHUFFLEROW_AVX2
476 YANY(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2,
477 ARGBShuffleRow_C, 4, 4, 15)
479 #ifdef HAS_ARGBSHUFFLEROW_NEON
480 YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
481 ARGBShuffleRow_C, 4, 4, 3)
485 // Interpolate may want to work in place, so last16 method can not be used.
486 #define NANY(NAMEANY, TERP_SIMD, TERP_C, SBPP, BPP, MASK) \
487 void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, \
488 ptrdiff_t src_stride_ptr, int width, \
489 int source_y_fraction) { \
490 int n = width & ~MASK; \
491 TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr, \
492 n, source_y_fraction); \
493 TERP_C(dst_ptr + n * BPP, \
494 src_ptr + n * SBPP, src_stride_ptr, \
495 width & MASK, source_y_fraction); \
498 #ifdef HAS_INTERPOLATEROW_SSSE3
499 NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3,
500 InterpolateRow_C, 1, 1, 15)
502 #ifdef HAS_INTERPOLATEROW_SSE2
503 NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2,
504 InterpolateRow_C, 1, 1, 15)
506 #ifdef HAS_INTERPOLATEROW_NEON
507 NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
508 InterpolateRow_C, 1, 1, 15)
510 #ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
511 NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
512 InterpolateRow_C, 1, 1, 3)
518 } // namespace libyuv