1 // Copyright 2014 Google Inc. All Rights Reserved.
2 //
3 // Use of this source code is governed by a BSD-style license
4 // that can be found in the COPYING file in the root of the source
5 // tree. An additional intellectual property rights grant can be found
6 // in the file PATENTS. All contributing project authors may
7 // be found in the AUTHORS file in the root of the source tree.
8 // -----------------------------------------------------------------------------
9 //
10 // YUV->RGB conversion functions
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13
14 #include "src/dsp/yuv.h"
15
16 #if defined(WEBP_USE_SSE41)
17
18 #include "src/dsp/common_sse41.h"
19 #include <stdlib.h>
20 #include <smmintrin.h>
21
22 //-----------------------------------------------------------------------------
23 // Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
24
25 // These constants are 14b fixed-point version of ITU-R BT.601 constants.
26 // R = (19077 * y + 26149 * v - 14234) >> 6
27 // G = (19077 * y - 6419 * u - 13320 * v + 8708) >> 6
28 // B = (19077 * y + 33050 * u - 17685) >> 6
ConvertYUV444ToRGB_SSE41(const __m128i * const Y0,const __m128i * const U0,const __m128i * const V0,__m128i * const R,__m128i * const G,__m128i * const B)29 static void ConvertYUV444ToRGB_SSE41(const __m128i* const Y0,
30 const __m128i* const U0,
31 const __m128i* const V0,
32 __m128i* const R,
33 __m128i* const G,
34 __m128i* const B) {
35 const __m128i k19077 = _mm_set1_epi16(19077);
36 const __m128i k26149 = _mm_set1_epi16(26149);
37 const __m128i k14234 = _mm_set1_epi16(14234);
38 // 33050 doesn't fit in a signed short: only use this with unsigned arithmetic
39 const __m128i k33050 = _mm_set1_epi16((short)33050);
40 const __m128i k17685 = _mm_set1_epi16(17685);
41 const __m128i k6419 = _mm_set1_epi16(6419);
42 const __m128i k13320 = _mm_set1_epi16(13320);
43 const __m128i k8708 = _mm_set1_epi16(8708);
44
45 const __m128i Y1 = _mm_mulhi_epu16(*Y0, k19077);
46
47 const __m128i R0 = _mm_mulhi_epu16(*V0, k26149);
48 const __m128i R1 = _mm_sub_epi16(Y1, k14234);
49 const __m128i R2 = _mm_add_epi16(R1, R0);
50
51 const __m128i G0 = _mm_mulhi_epu16(*U0, k6419);
52 const __m128i G1 = _mm_mulhi_epu16(*V0, k13320);
53 const __m128i G2 = _mm_add_epi16(Y1, k8708);
54 const __m128i G3 = _mm_add_epi16(G0, G1);
55 const __m128i G4 = _mm_sub_epi16(G2, G3);
56
57 // be careful with the saturated *unsigned* arithmetic here!
58 const __m128i B0 = _mm_mulhi_epu16(*U0, k33050);
59 const __m128i B1 = _mm_adds_epu16(B0, Y1);
60 const __m128i B2 = _mm_subs_epu16(B1, k17685);
61
62 // use logical shift for B2, which can be larger than 32767
63 *R = _mm_srai_epi16(R2, 6); // range: [-14234, 30815]
64 *G = _mm_srai_epi16(G4, 6); // range: [-10953, 27710]
65 *B = _mm_srli_epi16(B2, 6); // range: [0, 34238]
66 }
67
68 // Load the bytes into the *upper* part of 16b words. That's "<< 8", basically.
Load_HI_16_SSE41(const uint8_t * src)69 static WEBP_INLINE __m128i Load_HI_16_SSE41(const uint8_t* src) {
70 const __m128i zero = _mm_setzero_si128();
71 return _mm_unpacklo_epi8(zero, _mm_loadl_epi64((const __m128i*)src));
72 }
73
74 // Load and replicate the U/V samples
Load_UV_HI_8_SSE41(const uint8_t * src)75 static WEBP_INLINE __m128i Load_UV_HI_8_SSE41(const uint8_t* src) {
76 const __m128i zero = _mm_setzero_si128();
77 const __m128i tmp0 = _mm_cvtsi32_si128(*(const uint32_t*)src);
78 const __m128i tmp1 = _mm_unpacklo_epi8(zero, tmp0);
79 return _mm_unpacklo_epi16(tmp1, tmp1); // replicate samples
80 }
81
82 // Convert 32 samples of YUV444 to R/G/B
YUV444ToRGB_SSE41(const uint8_t * const y,const uint8_t * const u,const uint8_t * const v,__m128i * const R,__m128i * const G,__m128i * const B)83 static void YUV444ToRGB_SSE41(const uint8_t* const y,
84 const uint8_t* const u,
85 const uint8_t* const v,
86 __m128i* const R, __m128i* const G,
87 __m128i* const B) {
88 const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_HI_16_SSE41(u),
89 V0 = Load_HI_16_SSE41(v);
90 ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B);
91 }
92
93 // Convert 32 samples of YUV420 to R/G/B
YUV420ToRGB_SSE41(const uint8_t * const y,const uint8_t * const u,const uint8_t * const v,__m128i * const R,__m128i * const G,__m128i * const B)94 static void YUV420ToRGB_SSE41(const uint8_t* const y,
95 const uint8_t* const u,
96 const uint8_t* const v,
97 __m128i* const R, __m128i* const G,
98 __m128i* const B) {
99 const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_UV_HI_8_SSE41(u),
100 V0 = Load_UV_HI_8_SSE41(v);
101 ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B);
102 }
103
104 // Pack the planar buffers
105 // rrrr... rrrr... gggg... gggg... bbbb... bbbb....
106 // triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
PlanarTo24b_SSE41(__m128i * const in0,__m128i * const in1,__m128i * const in2,__m128i * const in3,__m128i * const in4,__m128i * const in5,uint8_t * const rgb)107 static WEBP_INLINE void PlanarTo24b_SSE41(
108 __m128i* const in0, __m128i* const in1, __m128i* const in2,
109 __m128i* const in3, __m128i* const in4, __m128i* const in5,
110 uint8_t* const rgb) {
111 // The input is 6 registers of sixteen 8b but for the sake of explanation,
112 // let's take 6 registers of four 8b values.
113 // To pack, we will keep taking one every two 8b integer and move it
114 // around as follows:
115 // Input:
116 // r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7
117 // Split the 6 registers in two sets of 3 registers: the first set as the even
118 // 8b bytes, the second the odd ones:
119 // r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7
120 // Repeat the same permutations twice more:
121 // r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
122 // r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
123 VP8PlanarTo24b_SSE41(in0, in1, in2, in3, in4, in5);
124
125 _mm_storeu_si128((__m128i*)(rgb + 0), *in0);
126 _mm_storeu_si128((__m128i*)(rgb + 16), *in1);
127 _mm_storeu_si128((__m128i*)(rgb + 32), *in2);
128 _mm_storeu_si128((__m128i*)(rgb + 48), *in3);
129 _mm_storeu_si128((__m128i*)(rgb + 64), *in4);
130 _mm_storeu_si128((__m128i*)(rgb + 80), *in5);
131 }
132
VP8YuvToRgb32_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst)133 void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
134 uint8_t* dst) {
135 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
136 __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
137
138 YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0);
139 YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1);
140 YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2);
141 YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3);
142
143 // Cast to 8b and store as RRRRGGGGBBBB.
144 rgb0 = _mm_packus_epi16(R0, R1);
145 rgb1 = _mm_packus_epi16(R2, R3);
146 rgb2 = _mm_packus_epi16(G0, G1);
147 rgb3 = _mm_packus_epi16(G2, G3);
148 rgb4 = _mm_packus_epi16(B0, B1);
149 rgb5 = _mm_packus_epi16(B2, B3);
150
151 // Pack as RGBRGBRGBRGB.
152 PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
153 }
154
VP8YuvToBgr32_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst)155 void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
156 uint8_t* dst) {
157 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
158 __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
159
160 YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0);
161 YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1);
162 YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2);
163 YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3);
164
165 // Cast to 8b and store as BBBBGGGGRRRR.
166 bgr0 = _mm_packus_epi16(B0, B1);
167 bgr1 = _mm_packus_epi16(B2, B3);
168 bgr2 = _mm_packus_epi16(G0, G1);
169 bgr3 = _mm_packus_epi16(G2, G3);
170 bgr4 = _mm_packus_epi16(R0, R1);
171 bgr5= _mm_packus_epi16(R2, R3);
172
173 // Pack as BGRBGRBGRBGR.
174 PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst);
175 }
176
177 //-----------------------------------------------------------------------------
178 // Arbitrary-length row conversion functions
179
YuvToRgbRow_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst,int len)180 static void YuvToRgbRow_SSE41(const uint8_t* y,
181 const uint8_t* u, const uint8_t* v,
182 uint8_t* dst, int len) {
183 int n;
184 for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
185 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
186 __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
187
188 YUV420ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0);
189 YUV420ToRGB_SSE41(y + 8, u + 4, v + 4, &R1, &G1, &B1);
190 YUV420ToRGB_SSE41(y + 16, u + 8, v + 8, &R2, &G2, &B2);
191 YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3);
192
193 // Cast to 8b and store as RRRRGGGGBBBB.
194 rgb0 = _mm_packus_epi16(R0, R1);
195 rgb1 = _mm_packus_epi16(R2, R3);
196 rgb2 = _mm_packus_epi16(G0, G1);
197 rgb3 = _mm_packus_epi16(G2, G3);
198 rgb4 = _mm_packus_epi16(B0, B1);
199 rgb5 = _mm_packus_epi16(B2, B3);
200
201 // Pack as RGBRGBRGBRGB.
202 PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
203
204 y += 32;
205 u += 16;
206 v += 16;
207 }
208 for (; n < len; ++n) { // Finish off
209 VP8YuvToRgb(y[0], u[0], v[0], dst);
210 dst += 3;
211 y += 1;
212 u += (n & 1);
213 v += (n & 1);
214 }
215 }
216
YuvToBgrRow_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst,int len)217 static void YuvToBgrRow_SSE41(const uint8_t* y,
218 const uint8_t* u, const uint8_t* v,
219 uint8_t* dst, int len) {
220 int n;
221 for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
222 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
223 __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
224
225 YUV420ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0);
226 YUV420ToRGB_SSE41(y + 8, u + 4, v + 4, &R1, &G1, &B1);
227 YUV420ToRGB_SSE41(y + 16, u + 8, v + 8, &R2, &G2, &B2);
228 YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3);
229
230 // Cast to 8b and store as BBBBGGGGRRRR.
231 bgr0 = _mm_packus_epi16(B0, B1);
232 bgr1 = _mm_packus_epi16(B2, B3);
233 bgr2 = _mm_packus_epi16(G0, G1);
234 bgr3 = _mm_packus_epi16(G2, G3);
235 bgr4 = _mm_packus_epi16(R0, R1);
236 bgr5 = _mm_packus_epi16(R2, R3);
237
238 // Pack as BGRBGRBGRBGR.
239 PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst);
240
241 y += 32;
242 u += 16;
243 v += 16;
244 }
245 for (; n < len; ++n) { // Finish off
246 VP8YuvToBgr(y[0], u[0], v[0], dst);
247 dst += 3;
248 y += 1;
249 u += (n & 1);
250 v += (n & 1);
251 }
252 }
253
254 //------------------------------------------------------------------------------
255 // Entry point
256
257 extern void WebPInitSamplersSSE41(void);
258
WebPInitSamplersSSE41(void)259 WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE41(void) {
260 WebPSamplers[MODE_RGB] = YuvToRgbRow_SSE41;
261 WebPSamplers[MODE_BGR] = YuvToBgrRow_SSE41;
262 }
263
264 //------------------------------------------------------------------------------
265 // RGB24/32 -> YUV converters
266
267 // Load eight 16b-words from *src.
268 #define LOAD_16(src) _mm_loadu_si128((const __m128i*)(src))
269 // Store either 16b-words into *dst
270 #define STORE_16(V, dst) _mm_storeu_si128((__m128i*)(dst), (V))
271
272 #define WEBP_SSE41_SHUFF(OUT) do { \
273 const __m128i tmp0 = _mm_shuffle_epi8(A0, shuff0); \
274 const __m128i tmp1 = _mm_shuffle_epi8(A1, shuff1); \
275 const __m128i tmp2 = _mm_shuffle_epi8(A2, shuff2); \
276 const __m128i tmp3 = _mm_shuffle_epi8(A3, shuff0); \
277 const __m128i tmp4 = _mm_shuffle_epi8(A4, shuff1); \
278 const __m128i tmp5 = _mm_shuffle_epi8(A5, shuff2); \
279 \
280 /* OR everything to get one channel */ \
281 const __m128i tmp6 = _mm_or_si128(tmp0, tmp1); \
282 const __m128i tmp7 = _mm_or_si128(tmp3, tmp4); \
283 out[OUT + 0] = _mm_or_si128(tmp6, tmp2); \
284 out[OUT + 1] = _mm_or_si128(tmp7, tmp5); \
285 } while (0);
286
287 // Unpack the 8b input rgbrgbrgbrgb ... as contiguous registers:
288 // rrrr... rrrr... gggg... gggg... bbbb... bbbb....
289 // Similar to PlanarTo24bHelper(), but in reverse order.
RGB24PackedToPlanar_SSE41(const uint8_t * const rgb,__m128i * const out)290 static WEBP_INLINE void RGB24PackedToPlanar_SSE41(
291 const uint8_t* const rgb, __m128i* const out /*out[6]*/) {
292 const __m128i A0 = _mm_loadu_si128((const __m128i*)(rgb + 0));
293 const __m128i A1 = _mm_loadu_si128((const __m128i*)(rgb + 16));
294 const __m128i A2 = _mm_loadu_si128((const __m128i*)(rgb + 32));
295 const __m128i A3 = _mm_loadu_si128((const __m128i*)(rgb + 48));
296 const __m128i A4 = _mm_loadu_si128((const __m128i*)(rgb + 64));
297 const __m128i A5 = _mm_loadu_si128((const __m128i*)(rgb + 80));
298
299 // Compute RR.
300 {
301 const __m128i shuff0 = _mm_set_epi8(
302 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0);
303 const __m128i shuff1 = _mm_set_epi8(
304 -1, -1, -1, -1, -1, 14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1);
305 const __m128i shuff2 = _mm_set_epi8(
306 13, 10, 7, 4, 1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
307 WEBP_SSE41_SHUFF(0)
308 }
309 // Compute GG.
310 {
311 const __m128i shuff0 = _mm_set_epi8(
312 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1);
313 const __m128i shuff1 = _mm_set_epi8(
314 -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1);
315 const __m128i shuff2 = _mm_set_epi8(
316 14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
317 WEBP_SSE41_SHUFF(2)
318 }
319 // Compute BB.
320 {
321 const __m128i shuff0 = _mm_set_epi8(
322 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 14, 11, 8, 5, 2);
323 const __m128i shuff1 = _mm_set_epi8(
324 -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1, -1, -1, -1, -1, -1);
325 const __m128i shuff2 = _mm_set_epi8(
326 15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
327 WEBP_SSE41_SHUFF(4)
328 }
329 }
330
331 #undef WEBP_SSE41_SHUFF
332
333 // Convert 8 packed ARGB to r[], g[], b[]
RGB32PackedToPlanar_SSE41(const uint32_t * const argb,__m128i * const rgb)334 static WEBP_INLINE void RGB32PackedToPlanar_SSE41(
335 const uint32_t* const argb, __m128i* const rgb /*in[6]*/) {
336 const __m128i zero = _mm_setzero_si128();
337 __m128i a0 = LOAD_16(argb + 0);
338 __m128i a1 = LOAD_16(argb + 4);
339 __m128i a2 = LOAD_16(argb + 8);
340 __m128i a3 = LOAD_16(argb + 12);
341 VP8L32bToPlanar_SSE41(&a0, &a1, &a2, &a3);
342 rgb[0] = _mm_unpacklo_epi8(a1, zero);
343 rgb[1] = _mm_unpackhi_epi8(a1, zero);
344 rgb[2] = _mm_unpacklo_epi8(a2, zero);
345 rgb[3] = _mm_unpackhi_epi8(a2, zero);
346 rgb[4] = _mm_unpacklo_epi8(a3, zero);
347 rgb[5] = _mm_unpackhi_epi8(a3, zero);
348 }
349
350 // This macro computes (RG * MULT_RG + GB * MULT_GB + ROUNDER) >> DESCALE_FIX
351 // It's a macro and not a function because we need to use immediate values with
352 // srai_epi32, e.g.
353 #define TRANSFORM(RG_LO, RG_HI, GB_LO, GB_HI, MULT_RG, MULT_GB, \
354 ROUNDER, DESCALE_FIX, OUT) do { \
355 const __m128i V0_lo = _mm_madd_epi16(RG_LO, MULT_RG); \
356 const __m128i V0_hi = _mm_madd_epi16(RG_HI, MULT_RG); \
357 const __m128i V1_lo = _mm_madd_epi16(GB_LO, MULT_GB); \
358 const __m128i V1_hi = _mm_madd_epi16(GB_HI, MULT_GB); \
359 const __m128i V2_lo = _mm_add_epi32(V0_lo, V1_lo); \
360 const __m128i V2_hi = _mm_add_epi32(V0_hi, V1_hi); \
361 const __m128i V3_lo = _mm_add_epi32(V2_lo, ROUNDER); \
362 const __m128i V3_hi = _mm_add_epi32(V2_hi, ROUNDER); \
363 const __m128i V5_lo = _mm_srai_epi32(V3_lo, DESCALE_FIX); \
364 const __m128i V5_hi = _mm_srai_epi32(V3_hi, DESCALE_FIX); \
365 (OUT) = _mm_packs_epi32(V5_lo, V5_hi); \
366 } while (0)
367
368 #define MK_CST_16(A, B) _mm_set_epi16((B), (A), (B), (A), (B), (A), (B), (A))
ConvertRGBToY_SSE41(const __m128i * const R,const __m128i * const G,const __m128i * const B,__m128i * const Y)369 static WEBP_INLINE void ConvertRGBToY_SSE41(const __m128i* const R,
370 const __m128i* const G,
371 const __m128i* const B,
372 __m128i* const Y) {
373 const __m128i kRG_y = MK_CST_16(16839, 33059 - 16384);
374 const __m128i kGB_y = MK_CST_16(16384, 6420);
375 const __m128i kHALF_Y = _mm_set1_epi32((16 << YUV_FIX) + YUV_HALF);
376
377 const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G);
378 const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G);
379 const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B);
380 const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B);
381 TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_y, kGB_y, kHALF_Y, YUV_FIX, *Y);
382 }
383
ConvertRGBToUV_SSE41(const __m128i * const R,const __m128i * const G,const __m128i * const B,__m128i * const U,__m128i * const V)384 static WEBP_INLINE void ConvertRGBToUV_SSE41(const __m128i* const R,
385 const __m128i* const G,
386 const __m128i* const B,
387 __m128i* const U,
388 __m128i* const V) {
389 const __m128i kRG_u = MK_CST_16(-9719, -19081);
390 const __m128i kGB_u = MK_CST_16(0, 28800);
391 const __m128i kRG_v = MK_CST_16(28800, 0);
392 const __m128i kGB_v = MK_CST_16(-24116, -4684);
393 const __m128i kHALF_UV = _mm_set1_epi32(((128 << YUV_FIX) + YUV_HALF) << 2);
394
395 const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G);
396 const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G);
397 const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B);
398 const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B);
399 TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_u, kGB_u,
400 kHALF_UV, YUV_FIX + 2, *U);
401 TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_v, kGB_v,
402 kHALF_UV, YUV_FIX + 2, *V);
403 }
404
405 #undef MK_CST_16
406 #undef TRANSFORM
407
ConvertRGB24ToY_SSE41(const uint8_t * rgb,uint8_t * y,int width)408 static void ConvertRGB24ToY_SSE41(const uint8_t* rgb, uint8_t* y, int width) {
409 const int max_width = width & ~31;
410 int i;
411 for (i = 0; i < max_width; rgb += 3 * 16 * 2) {
412 __m128i rgb_plane[6];
413 int j;
414
415 RGB24PackedToPlanar_SSE41(rgb, rgb_plane);
416
417 for (j = 0; j < 2; ++j, i += 16) {
418 const __m128i zero = _mm_setzero_si128();
419 __m128i r, g, b, Y0, Y1;
420
421 // Convert to 16-bit Y.
422 r = _mm_unpacklo_epi8(rgb_plane[0 + j], zero);
423 g = _mm_unpacklo_epi8(rgb_plane[2 + j], zero);
424 b = _mm_unpacklo_epi8(rgb_plane[4 + j], zero);
425 ConvertRGBToY_SSE41(&r, &g, &b, &Y0);
426
427 // Convert to 16-bit Y.
428 r = _mm_unpackhi_epi8(rgb_plane[0 + j], zero);
429 g = _mm_unpackhi_epi8(rgb_plane[2 + j], zero);
430 b = _mm_unpackhi_epi8(rgb_plane[4 + j], zero);
431 ConvertRGBToY_SSE41(&r, &g, &b, &Y1);
432
433 // Cast to 8-bit and store.
434 STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
435 }
436 }
437 for (; i < width; ++i, rgb += 3) { // left-over
438 y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
439 }
440 }
441
ConvertBGR24ToY_SSE41(const uint8_t * bgr,uint8_t * y,int width)442 static void ConvertBGR24ToY_SSE41(const uint8_t* bgr, uint8_t* y, int width) {
443 const int max_width = width & ~31;
444 int i;
445 for (i = 0; i < max_width; bgr += 3 * 16 * 2) {
446 __m128i bgr_plane[6];
447 int j;
448
449 RGB24PackedToPlanar_SSE41(bgr, bgr_plane);
450
451 for (j = 0; j < 2; ++j, i += 16) {
452 const __m128i zero = _mm_setzero_si128();
453 __m128i r, g, b, Y0, Y1;
454
455 // Convert to 16-bit Y.
456 b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero);
457 g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero);
458 r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero);
459 ConvertRGBToY_SSE41(&r, &g, &b, &Y0);
460
461 // Convert to 16-bit Y.
462 b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero);
463 g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero);
464 r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero);
465 ConvertRGBToY_SSE41(&r, &g, &b, &Y1);
466
467 // Cast to 8-bit and store.
468 STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
469 }
470 }
471 for (; i < width; ++i, bgr += 3) { // left-over
472 y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
473 }
474 }
475
ConvertARGBToY_SSE41(const uint32_t * argb,uint8_t * y,int width)476 static void ConvertARGBToY_SSE41(const uint32_t* argb, uint8_t* y, int width) {
477 const int max_width = width & ~15;
478 int i;
479 for (i = 0; i < max_width; i += 16) {
480 __m128i Y0, Y1, rgb[6];
481 RGB32PackedToPlanar_SSE41(&argb[i], rgb);
482 ConvertRGBToY_SSE41(&rgb[0], &rgb[2], &rgb[4], &Y0);
483 ConvertRGBToY_SSE41(&rgb[1], &rgb[3], &rgb[5], &Y1);
484 STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
485 }
486 for (; i < width; ++i) { // left-over
487 const uint32_t p = argb[i];
488 y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff,
489 YUV_HALF);
490 }
491 }
492
493 // Horizontal add (doubled) of two 16b values, result is 16b.
494 // in: A | B | C | D | ... -> out: 2*(A+B) | 2*(C+D) | ...
HorizontalAddPack_SSE41(const __m128i * const A,const __m128i * const B,__m128i * const out)495 static void HorizontalAddPack_SSE41(const __m128i* const A,
496 const __m128i* const B,
497 __m128i* const out) {
498 const __m128i k2 = _mm_set1_epi16(2);
499 const __m128i C = _mm_madd_epi16(*A, k2);
500 const __m128i D = _mm_madd_epi16(*B, k2);
501 *out = _mm_packs_epi32(C, D);
502 }
503
ConvertARGBToUV_SSE41(const uint32_t * argb,uint8_t * u,uint8_t * v,int src_width,int do_store)504 static void ConvertARGBToUV_SSE41(const uint32_t* argb,
505 uint8_t* u, uint8_t* v,
506 int src_width, int do_store) {
507 const int max_width = src_width & ~31;
508 int i;
509 for (i = 0; i < max_width; i += 32, u += 16, v += 16) {
510 __m128i rgb[6], U0, V0, U1, V1;
511 RGB32PackedToPlanar_SSE41(&argb[i], rgb);
512 HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]);
513 HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]);
514 HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]);
515 ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U0, &V0);
516
517 RGB32PackedToPlanar_SSE41(&argb[i + 16], rgb);
518 HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]);
519 HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]);
520 HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]);
521 ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U1, &V1);
522
523 U0 = _mm_packus_epi16(U0, U1);
524 V0 = _mm_packus_epi16(V0, V1);
525 if (!do_store) {
526 const __m128i prev_u = LOAD_16(u);
527 const __m128i prev_v = LOAD_16(v);
528 U0 = _mm_avg_epu8(U0, prev_u);
529 V0 = _mm_avg_epu8(V0, prev_v);
530 }
531 STORE_16(U0, u);
532 STORE_16(V0, v);
533 }
534 if (i < src_width) { // left-over
535 WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store);
536 }
537 }
538
539 // Convert 16 packed ARGB 16b-values to r[], g[], b[]
RGBA32PackedToPlanar_16b_SSE41(const uint16_t * const rgbx,__m128i * const r,__m128i * const g,__m128i * const b)540 static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE41(
541 const uint16_t* const rgbx,
542 __m128i* const r, __m128i* const g, __m128i* const b) {
543 const __m128i in0 = LOAD_16(rgbx + 0); // r0 | g0 | b0 |x| r1 | g1 | b1 |x
544 const __m128i in1 = LOAD_16(rgbx + 8); // r2 | g2 | b2 |x| r3 | g3 | b3 |x
545 const __m128i in2 = LOAD_16(rgbx + 16); // r4 | ...
546 const __m128i in3 = LOAD_16(rgbx + 24); // r6 | ...
547 // aarrggbb as 16-bit.
548 const __m128i shuff0 =
549 _mm_set_epi8(-1, -1, -1, -1, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
550 const __m128i shuff1 =
551 _mm_set_epi8(13, 12, 5, 4, -1, -1, -1, -1, 11, 10, 3, 2, 9, 8, 1, 0);
552 const __m128i A0 = _mm_shuffle_epi8(in0, shuff0);
553 const __m128i A1 = _mm_shuffle_epi8(in1, shuff1);
554 const __m128i A2 = _mm_shuffle_epi8(in2, shuff0);
555 const __m128i A3 = _mm_shuffle_epi8(in3, shuff1);
556 // R0R1G0G1
557 // B0B1****
558 // R2R3G2G3
559 // B2B3****
560 // (OR is used to free port 5 for the unpack)
561 const __m128i B0 = _mm_unpacklo_epi32(A0, A1);
562 const __m128i B1 = _mm_or_si128(A0, A1);
563 const __m128i B2 = _mm_unpacklo_epi32(A2, A3);
564 const __m128i B3 = _mm_or_si128(A2, A3);
565 // Gather the channels.
566 *r = _mm_unpacklo_epi64(B0, B2);
567 *g = _mm_unpackhi_epi64(B0, B2);
568 *b = _mm_unpackhi_epi64(B1, B3);
569 }
570
ConvertRGBA32ToUV_SSE41(const uint16_t * rgb,uint8_t * u,uint8_t * v,int width)571 static void ConvertRGBA32ToUV_SSE41(const uint16_t* rgb,
572 uint8_t* u, uint8_t* v, int width) {
573 const int max_width = width & ~15;
574 const uint16_t* const last_rgb = rgb + 4 * max_width;
575 while (rgb < last_rgb) {
576 __m128i r, g, b, U0, V0, U1, V1;
577 RGBA32PackedToPlanar_16b_SSE41(rgb + 0, &r, &g, &b);
578 ConvertRGBToUV_SSE41(&r, &g, &b, &U0, &V0);
579 RGBA32PackedToPlanar_16b_SSE41(rgb + 32, &r, &g, &b);
580 ConvertRGBToUV_SSE41(&r, &g, &b, &U1, &V1);
581 STORE_16(_mm_packus_epi16(U0, U1), u);
582 STORE_16(_mm_packus_epi16(V0, V1), v);
583 u += 16;
584 v += 16;
585 rgb += 2 * 32;
586 }
587 if (max_width < width) { // left-over
588 WebPConvertRGBA32ToUV_C(rgb, u, v, width - max_width);
589 }
590 }
591
592 //------------------------------------------------------------------------------
593
594 extern void WebPInitConvertARGBToYUVSSE41(void);
595
WebPInitConvertARGBToYUVSSE41(void)596 WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE41(void) {
597 WebPConvertARGBToY = ConvertARGBToY_SSE41;
598 WebPConvertARGBToUV = ConvertARGBToUV_SSE41;
599
600 WebPConvertRGB24ToY = ConvertRGB24ToY_SSE41;
601 WebPConvertBGR24ToY = ConvertBGR24ToY_SSE41;
602
603 WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_SSE41;
604 }
605
606 //------------------------------------------------------------------------------
607
608 #else // !WEBP_USE_SSE41
609
610 WEBP_DSP_INIT_STUB(WebPInitSamplersSSE41)
611 WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE41)
612
613 #endif // WEBP_USE_SSE41
614