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(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C,
40 #endif // HAS_I422TOARGBROW_SSSE3
41 #ifdef HAS_I444TOARGBROW_SSSE3
42 YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C,
44 YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C,
46 YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C,
48 YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C,
50 YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C,
52 // I422ToRGB565Row_SSSE3 is unaligned.
53 YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
55 YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
57 YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
59 // I422ToRGB24Row_SSSE3 is unaligned.
60 YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
61 YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
62 YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
63 YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
64 #endif // HAS_I444TOARGBROW_SSSE3
65 #ifdef HAS_I422TOARGBROW_AVX2
66 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
67 #endif // HAS_I422TOARGBROW_AVX2
68 #ifdef HAS_I422TOARGBROW_NEON
69 YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
70 YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
71 YANY(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, I411ToARGBRow_C, 2, 4, 7)
72 YANY(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, I422ToBGRARow_C, 1, 4, 7)
73 YANY(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, I422ToABGRRow_C, 1, 4, 7)
74 YANY(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, I422ToRGBARow_C, 1, 4, 7)
75 YANY(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, I422ToRGB24Row_C, 1, 3, 7)
76 YANY(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, I422ToRAWRow_C, 1, 3, 7)
77 YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
79 YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
81 YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
82 YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
83 YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
84 #endif // HAS_I422TOARGBROW_NEON
87 // Wrappers to handle odd width
88 #define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP) \
89 void NAMEANY(const uint8* y_buf, \
90 const uint8* uv_buf, \
94 NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n); \
95 NV12TORGB_C(y_buf + n, \
96 uv_buf + (n >> UV_SHIFT), \
97 rgb_buf + n * BPP, width & 7); \
100 #ifdef HAS_NV12TOARGBROW_SSSE3
101 NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C,
103 NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C,
105 #endif // HAS_NV12TOARGBROW_SSSE3
106 #ifdef HAS_NV12TOARGBROW_NEON
107 NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
108 NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
109 #endif // HAS_NV12TOARGBROW_NEON
110 #ifdef HAS_NV12TORGB565ROW_SSSE3
111 NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
113 NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
115 #endif // HAS_NV12TORGB565ROW_SSSE3
116 #ifdef HAS_NV12TORGB565ROW_NEON
117 NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
118 NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
119 #endif // HAS_NV12TORGB565ROW_NEON
122 #define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP) \
123 void NAMEANY(const uint8* src, \
126 int n = width & ~MASK; \
127 ARGBTORGB_SIMD(src, dst, n); \
128 ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK); \
131 #if defined(HAS_ARGBTORGB24ROW_SSSE3)
132 RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
134 RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
136 RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
138 RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
140 RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
143 #if defined(HAS_I400TOARGBROW_SSE2)
144 RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C,
147 #if defined(HAS_YTOARGBROW_SSE2)
148 RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
150 RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C,
152 RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C,
154 // These require alignment on ARGB, so C is used for remainder.
155 RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
157 RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
159 RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
161 RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
163 RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
166 #if defined(HAS_ARGBTORGB24ROW_NEON)
167 RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
168 RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
169 RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
171 RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
173 RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
175 RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
177 RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
179 RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
181 RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
186 // ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
187 #define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP) \
188 void NAMEANY(const uint8* src, \
189 uint8* dst, uint32 selector, \
191 int n = width & ~MASK; \
192 ARGBTORGB_SIMD(src, dst, selector, n); \
193 ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK); \
196 #if defined(HAS_ARGBTOBAYERROW_SSSE3)
197 BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
200 #if defined(HAS_ARGBTOBAYERROW_NEON)
201 BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
204 #if defined(HAS_ARGBTOBAYERGGROW_SSE2)
205 BAYERANY(ARGBToBayerGGRow_Any_SSE2, ARGBToBayerGGRow_SSE2, ARGBToBayerGGRow_C,
208 #if defined(HAS_ARGBTOBAYERGGROW_NEON)
209 BAYERANY(ARGBToBayerGGRow_Any_NEON, ARGBToBayerGGRow_NEON, ARGBToBayerGGRow_C,
215 // RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
216 #define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM) \
217 void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) { \
218 ARGBTOY_SIMD(src_argb, dst_y, width - NUM); \
219 ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP, \
220 dst_y + (width - NUM) * BPP, NUM); \
223 #ifdef HAS_ARGBTOYROW_AVX2
224 YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
225 YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
226 YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
227 YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
229 #ifdef HAS_ARGBTOYROW_SSSE3
230 YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16)
232 #ifdef HAS_BGRATOYROW_SSSE3
233 YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16)
234 YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16)
235 YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16)
236 YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16)
237 YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16)
239 #ifdef HAS_ARGBTOYJROW_SSSE3
240 YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16)
242 #ifdef HAS_ARGBTOYROW_NEON
243 YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
244 YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
245 YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
246 YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
247 YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
248 YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
249 YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
250 YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
251 YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
252 YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
253 YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
254 YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
255 YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
256 YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
257 YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
258 YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
259 YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
263 #define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK) \
264 void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) { \
265 int n = width & ~MASK; \
266 ARGBTOY_SIMD(src_argb, dst_y, n); \
267 ARGBTOY_C(src_argb + n * SBPP, \
268 dst_y + n * BPP, width & MASK); \
271 // Attenuate is destructive so last16 method can not be used due to overlap.
272 #ifdef HAS_ARGBATTENUATEROW_SSSE3
273 YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
276 #ifdef HAS_ARGBATTENUATEROW_SSE2
277 YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
280 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
281 YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
284 #ifdef HAS_ARGBATTENUATEROW_AVX2
285 YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
288 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
289 YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
292 #ifdef HAS_ARGBATTENUATEROW_NEON
293 YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
298 // RGB/YUV to UV does multiple of 16 with SIMD and remainder with C.
299 #define UVANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK) \
300 void NAMEANY(const uint8* src_argb, int src_stride_argb, \
301 uint8* dst_u, uint8* dst_v, int width) { \
302 int n = width & ~MASK; \
303 ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n); \
304 ANYTOUV_C(src_argb + n * BPP, src_stride_argb, \
310 #ifdef HAS_ARGBTOUVROW_AVX2
311 UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
312 UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
313 UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
315 #ifdef HAS_ARGBTOUVROW_SSSE3
316 UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15)
317 UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C,
319 UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15)
320 UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15)
321 UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15)
322 UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15)
323 UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15)
325 #ifdef HAS_ARGBTOUVROW_NEON
326 UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
327 UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
328 UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
329 UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
330 UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
331 UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
332 UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
333 UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
334 UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
335 UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
336 UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15)
337 UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
341 #define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT) \
342 void NAMEANY(const uint8* src_uv, \
343 uint8* dst_u, uint8* dst_v, int width) { \
344 int n = width & ~MASK; \
345 ANYTOUV_SIMD(src_uv, dst_u, dst_v, n); \
346 ANYTOUV_C(src_uv + n * BPP, \
347 dst_u + (n >> SHIFT), \
348 dst_v + (n >> SHIFT), \
352 #ifdef HAS_ARGBTOUV444ROW_SSSE3
353 UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3,
354 ARGBToUV444Row_C, 4, 15, 0)
356 #ifdef HAS_YUY2TOUV422ROW_AVX2
357 UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
358 YUY2ToUV422Row_C, 2, 31, 1)
359 UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
360 UYVYToUV422Row_C, 2, 31, 1)
362 #ifdef HAS_ARGBTOUVROW_SSSE3
363 UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3,
364 ARGBToUV422Row_C, 4, 15, 1)
365 UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2,
366 YUY2ToUV422Row_C, 2, 15, 1)
367 UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2,
368 UYVYToUV422Row_C, 2, 15, 1)
370 #ifdef HAS_YUY2TOUV422ROW_NEON
371 UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
372 ARGBToUV444Row_C, 4, 7, 0)
373 UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
374 ARGBToUV422Row_C, 4, 15, 1)
375 UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
376 ARGBToUV411Row_C, 4, 31, 2)
377 UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
378 YUY2ToUV422Row_C, 2, 15, 1)
379 UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
380 UYVYToUV422Row_C, 2, 15, 1)
384 #define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK) \
385 void NAMEANY(const uint8* src_uv, \
386 uint8* dst_u, uint8* dst_v, int width) { \
387 int n = width & ~MASK; \
388 ANYTOUV_SIMD(src_uv, dst_u, dst_v, n); \
389 ANYTOUV_C(src_uv + n * 2, \
395 #ifdef HAS_SPLITUVROW_SSE2
396 SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15)
398 #ifdef HAS_SPLITUVROW_AVX2
399 SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
401 #ifdef HAS_SPLITUVROW_NEON
402 SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
404 #ifdef HAS_SPLITUVROW_MIPS_DSPR2
405 SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
410 #define MERGEUVROW_ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK) \
411 void NAMEANY(const uint8* src_u, const uint8* src_v, \
412 uint8* dst_uv, int width) { \
413 int n = width & ~MASK; \
414 ANYTOUV_SIMD(src_u, src_v, dst_uv, n); \
415 ANYTOUV_C(src_u + n, \
421 #ifdef HAS_MERGEUVROW_SSE2
422 MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15)
424 #ifdef HAS_MERGEUVROW_AVX2
425 MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
427 #ifdef HAS_MERGEUVROW_NEON
428 MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
430 #undef MERGEUVROW_ANY
432 #define MATHROW_ANY(NAMEANY, ARGBMATH_SIMD, ARGBMATH_C, MASK) \
433 void NAMEANY(const uint8* src_argb0, const uint8* src_argb1, \
434 uint8* dst_argb, int width) { \
435 int n = width & ~MASK; \
436 ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n); \
437 ARGBMATH_C(src_argb0 + n * 4, \
443 #ifdef HAS_ARGBMULTIPLYROW_SSE2
444 MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
447 #ifdef HAS_ARGBADDROW_SSE2
448 MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
450 #ifdef HAS_ARGBSUBTRACTROW_SSE2
451 MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
454 #ifdef HAS_ARGBMULTIPLYROW_AVX2
455 MATHROW_ANY(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, ARGBMultiplyRow_C,
458 #ifdef HAS_ARGBADDROW_AVX2
459 MATHROW_ANY(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, ARGBAddRow_C, 7)
461 #ifdef HAS_ARGBSUBTRACTROW_AVX2
462 MATHROW_ANY(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, ARGBSubtractRow_C,
465 #ifdef HAS_ARGBMULTIPLYROW_NEON
466 MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
469 #ifdef HAS_ARGBADDROW_NEON
470 MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
472 #ifdef HAS_ARGBSUBTRACTROW_NEON
473 MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
478 // Shuffle may want to work in place, so last16 method can not be used.
479 #define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK) \
480 void NAMEANY(const uint8* src_argb, uint8* dst_argb, \
481 const uint8* shuffler, int width) { \
482 int n = width & ~MASK; \
483 ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n); \
484 ARGBTOY_C(src_argb + n * SBPP, \
485 dst_argb + n * BPP, shuffler, width & MASK); \
488 #ifdef HAS_ARGBSHUFFLEROW_SSE2
489 YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2,
490 ARGBShuffleRow_C, 4, 4, 3)
492 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
493 YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3,
494 ARGBShuffleRow_C, 4, 4, 7)
496 #ifdef HAS_ARGBSHUFFLEROW_AVX2
497 YANY(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2,
498 ARGBShuffleRow_C, 4, 4, 15)
500 #ifdef HAS_ARGBSHUFFLEROW_NEON
501 YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
502 ARGBShuffleRow_C, 4, 4, 3)
506 // Interpolate may want to work in place, so last16 method can not be used.
507 #define NANY(NAMEANY, TERP_SIMD, TERP_C, SBPP, BPP, MASK) \
508 void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, \
509 ptrdiff_t src_stride_ptr, int width, \
510 int source_y_fraction) { \
511 int n = width & ~MASK; \
512 TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr, \
513 n, source_y_fraction); \
514 TERP_C(dst_ptr + n * BPP, \
515 src_ptr + n * SBPP, src_stride_ptr, \
516 width & MASK, source_y_fraction); \
519 #ifdef HAS_INTERPOLATEROW_AVX2
520 NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2,
521 InterpolateRow_C, 1, 1, 32)
523 #ifdef HAS_INTERPOLATEROW_SSSE3
524 NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3,
525 InterpolateRow_C, 1, 1, 15)
527 #ifdef HAS_INTERPOLATEROW_SSE2
528 NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2,
529 InterpolateRow_C, 1, 1, 15)
531 #ifdef HAS_INTERPOLATEROW_NEON
532 NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
533 InterpolateRow_C, 1, 1, 15)
535 #ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
536 NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
537 InterpolateRow_C, 1, 1, 3)
543 } // namespace libyuv