xref: /aosp_15_r20/external/webp/src/dsp/yuv_sse41.c (revision b2055c353e87c8814eb2b6b1b11112a1562253bd)
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 ([email protected])
13 
14 #include "src/dsp/yuv.h"
15 
16 #if defined(WEBP_USE_SSE41)
17 
18 #include <stdlib.h>
19 #include <smmintrin.h>
20 
21 #include "src/dsp/common_sse41.h"
22 #include "src/utils/utils.h"
23 
24 //-----------------------------------------------------------------------------
25 // Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
26 
27 // These constants are 14b fixed-point version of ITU-R BT.601 constants.
28 // R = (19077 * y             + 26149 * v - 14234) >> 6
29 // G = (19077 * y -  6419 * u - 13320 * v +  8708) >> 6
30 // 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)31 static void ConvertYUV444ToRGB_SSE41(const __m128i* const Y0,
32                                      const __m128i* const U0,
33                                      const __m128i* const V0,
34                                      __m128i* const R,
35                                      __m128i* const G,
36                                      __m128i* const B) {
37   const __m128i k19077 = _mm_set1_epi16(19077);
38   const __m128i k26149 = _mm_set1_epi16(26149);
39   const __m128i k14234 = _mm_set1_epi16(14234);
40   // 33050 doesn't fit in a signed short: only use this with unsigned arithmetic
41   const __m128i k33050 = _mm_set1_epi16((short)33050);
42   const __m128i k17685 = _mm_set1_epi16(17685);
43   const __m128i k6419  = _mm_set1_epi16(6419);
44   const __m128i k13320 = _mm_set1_epi16(13320);
45   const __m128i k8708  = _mm_set1_epi16(8708);
46 
47   const __m128i Y1 = _mm_mulhi_epu16(*Y0, k19077);
48 
49   const __m128i R0 = _mm_mulhi_epu16(*V0, k26149);
50   const __m128i R1 = _mm_sub_epi16(Y1, k14234);
51   const __m128i R2 = _mm_add_epi16(R1, R0);
52 
53   const __m128i G0 = _mm_mulhi_epu16(*U0, k6419);
54   const __m128i G1 = _mm_mulhi_epu16(*V0, k13320);
55   const __m128i G2 = _mm_add_epi16(Y1, k8708);
56   const __m128i G3 = _mm_add_epi16(G0, G1);
57   const __m128i G4 = _mm_sub_epi16(G2, G3);
58 
59   // be careful with the saturated *unsigned* arithmetic here!
60   const __m128i B0 = _mm_mulhi_epu16(*U0, k33050);
61   const __m128i B1 = _mm_adds_epu16(B0, Y1);
62   const __m128i B2 = _mm_subs_epu16(B1, k17685);
63 
64   // use logical shift for B2, which can be larger than 32767
65   *R = _mm_srai_epi16(R2, 6);   // range: [-14234, 30815]
66   *G = _mm_srai_epi16(G4, 6);   // range: [-10953, 27710]
67   *B = _mm_srli_epi16(B2, 6);   // range: [0, 34238]
68 }
69 
70 // Load the bytes into the *upper* part of 16b words. That's "<< 8", basically.
Load_HI_16_SSE41(const uint8_t * src)71 static WEBP_INLINE __m128i Load_HI_16_SSE41(const uint8_t* src) {
72   const __m128i zero = _mm_setzero_si128();
73   return _mm_unpacklo_epi8(zero, _mm_loadl_epi64((const __m128i*)src));
74 }
75 
76 // Load and replicate the U/V samples
Load_UV_HI_8_SSE41(const uint8_t * src)77 static WEBP_INLINE __m128i Load_UV_HI_8_SSE41(const uint8_t* src) {
78   const __m128i zero = _mm_setzero_si128();
79   const __m128i tmp0 = _mm_cvtsi32_si128(WebPMemToInt32(src));
80   const __m128i tmp1 = _mm_unpacklo_epi8(zero, tmp0);
81   return _mm_unpacklo_epi16(tmp1, tmp1);   // replicate samples
82 }
83 
84 // 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)85 static void YUV444ToRGB_SSE41(const uint8_t* const y,
86                               const uint8_t* const u,
87                               const uint8_t* const v,
88                               __m128i* const R, __m128i* const G,
89                               __m128i* const B) {
90   const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_HI_16_SSE41(u),
91                 V0 = Load_HI_16_SSE41(v);
92   ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B);
93 }
94 
95 // 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)96 static void YUV420ToRGB_SSE41(const uint8_t* const y,
97                               const uint8_t* const u,
98                               const uint8_t* const v,
99                               __m128i* const R, __m128i* const G,
100                               __m128i* const B) {
101   const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_UV_HI_8_SSE41(u),
102                 V0 = Load_UV_HI_8_SSE41(v);
103   ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B);
104 }
105 
106 // Pack the planar buffers
107 // rrrr... rrrr... gggg... gggg... bbbb... bbbb....
108 // 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)109 static WEBP_INLINE void PlanarTo24b_SSE41(
110     __m128i* const in0, __m128i* const in1, __m128i* const in2,
111     __m128i* const in3, __m128i* const in4, __m128i* const in5,
112     uint8_t* const rgb) {
113   // The input is 6 registers of sixteen 8b but for the sake of explanation,
114   // let's take 6 registers of four 8b values.
115   // To pack, we will keep taking one every two 8b integer and move it
116   // around as follows:
117   // Input:
118   //   r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7
119   // Split the 6 registers in two sets of 3 registers: the first set as the even
120   // 8b bytes, the second the odd ones:
121   //   r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7
122   // Repeat the same permutations twice more:
123   //   r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
124   //   r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
125   VP8PlanarTo24b_SSE41(in0, in1, in2, in3, in4, in5);
126 
127   _mm_storeu_si128((__m128i*)(rgb +  0), *in0);
128   _mm_storeu_si128((__m128i*)(rgb + 16), *in1);
129   _mm_storeu_si128((__m128i*)(rgb + 32), *in2);
130   _mm_storeu_si128((__m128i*)(rgb + 48), *in3);
131   _mm_storeu_si128((__m128i*)(rgb + 64), *in4);
132   _mm_storeu_si128((__m128i*)(rgb + 80), *in5);
133 }
134 
VP8YuvToRgb32_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst)135 void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
136                          uint8_t* dst) {
137   __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
138   __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
139 
140   YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0);
141   YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1);
142   YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2);
143   YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3);
144 
145   // Cast to 8b and store as RRRRGGGGBBBB.
146   rgb0 = _mm_packus_epi16(R0, R1);
147   rgb1 = _mm_packus_epi16(R2, R3);
148   rgb2 = _mm_packus_epi16(G0, G1);
149   rgb3 = _mm_packus_epi16(G2, G3);
150   rgb4 = _mm_packus_epi16(B0, B1);
151   rgb5 = _mm_packus_epi16(B2, B3);
152 
153   // Pack as RGBRGBRGBRGB.
154   PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
155 }
156 
VP8YuvToBgr32_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst)157 void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v,
158                          uint8_t* dst) {
159   __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
160   __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
161 
162   YUV444ToRGB_SSE41(y +  0, u +  0, v +  0, &R0, &G0, &B0);
163   YUV444ToRGB_SSE41(y +  8, u +  8, v +  8, &R1, &G1, &B1);
164   YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2);
165   YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3);
166 
167   // Cast to 8b and store as BBBBGGGGRRRR.
168   bgr0 = _mm_packus_epi16(B0, B1);
169   bgr1 = _mm_packus_epi16(B2, B3);
170   bgr2 = _mm_packus_epi16(G0, G1);
171   bgr3 = _mm_packus_epi16(G2, G3);
172   bgr4 = _mm_packus_epi16(R0, R1);
173   bgr5= _mm_packus_epi16(R2, R3);
174 
175   // Pack as BGRBGRBGRBGR.
176   PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst);
177 }
178 
179 //-----------------------------------------------------------------------------
180 // Arbitrary-length row conversion functions
181 
YuvToRgbRow_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst,int len)182 static void YuvToRgbRow_SSE41(const uint8_t* y,
183                               const uint8_t* u, const uint8_t* v,
184                               uint8_t* dst, int len) {
185   int n;
186   for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
187     __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
188     __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
189 
190     YUV420ToRGB_SSE41(y +  0, u +  0, v +  0, &R0, &G0, &B0);
191     YUV420ToRGB_SSE41(y +  8, u +  4, v +  4, &R1, &G1, &B1);
192     YUV420ToRGB_SSE41(y + 16, u +  8, v +  8, &R2, &G2, &B2);
193     YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3);
194 
195     // Cast to 8b and store as RRRRGGGGBBBB.
196     rgb0 = _mm_packus_epi16(R0, R1);
197     rgb1 = _mm_packus_epi16(R2, R3);
198     rgb2 = _mm_packus_epi16(G0, G1);
199     rgb3 = _mm_packus_epi16(G2, G3);
200     rgb4 = _mm_packus_epi16(B0, B1);
201     rgb5 = _mm_packus_epi16(B2, B3);
202 
203     // Pack as RGBRGBRGBRGB.
204     PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
205 
206     y += 32;
207     u += 16;
208     v += 16;
209   }
210   for (; n < len; ++n) {   // Finish off
211     VP8YuvToRgb(y[0], u[0], v[0], dst);
212     dst += 3;
213     y += 1;
214     u += (n & 1);
215     v += (n & 1);
216   }
217 }
218 
YuvToBgrRow_SSE41(const uint8_t * y,const uint8_t * u,const uint8_t * v,uint8_t * dst,int len)219 static void YuvToBgrRow_SSE41(const uint8_t* y,
220                               const uint8_t* u, const uint8_t* v,
221                               uint8_t* dst, int len) {
222   int n;
223   for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
224     __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
225     __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
226 
227     YUV420ToRGB_SSE41(y +  0, u +  0, v +  0, &R0, &G0, &B0);
228     YUV420ToRGB_SSE41(y +  8, u +  4, v +  4, &R1, &G1, &B1);
229     YUV420ToRGB_SSE41(y + 16, u +  8, v +  8, &R2, &G2, &B2);
230     YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3);
231 
232     // Cast to 8b and store as BBBBGGGGRRRR.
233     bgr0 = _mm_packus_epi16(B0, B1);
234     bgr1 = _mm_packus_epi16(B2, B3);
235     bgr2 = _mm_packus_epi16(G0, G1);
236     bgr3 = _mm_packus_epi16(G2, G3);
237     bgr4 = _mm_packus_epi16(R0, R1);
238     bgr5 = _mm_packus_epi16(R2, R3);
239 
240     // Pack as BGRBGRBGRBGR.
241     PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst);
242 
243     y += 32;
244     u += 16;
245     v += 16;
246   }
247   for (; n < len; ++n) {   // Finish off
248     VP8YuvToBgr(y[0], u[0], v[0], dst);
249     dst += 3;
250     y += 1;
251     u += (n & 1);
252     v += (n & 1);
253   }
254 }
255 
256 //------------------------------------------------------------------------------
257 // Entry point
258 
259 extern void WebPInitSamplersSSE41(void);
260 
WebPInitSamplersSSE41(void)261 WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE41(void) {
262   WebPSamplers[MODE_RGB]  = YuvToRgbRow_SSE41;
263   WebPSamplers[MODE_BGR]  = YuvToBgrRow_SSE41;
264 }
265 
266 //------------------------------------------------------------------------------
267 // RGB24/32 -> YUV converters
268 
269 // Load eight 16b-words from *src.
270 #define LOAD_16(src) _mm_loadu_si128((const __m128i*)(src))
271 // Store either 16b-words into *dst
272 #define STORE_16(V, dst) _mm_storeu_si128((__m128i*)(dst), (V))
273 
274 #define WEBP_SSE41_SHUFF(OUT)  do {                  \
275   const __m128i tmp0 = _mm_shuffle_epi8(A0, shuff0); \
276   const __m128i tmp1 = _mm_shuffle_epi8(A1, shuff1); \
277   const __m128i tmp2 = _mm_shuffle_epi8(A2, shuff2); \
278   const __m128i tmp3 = _mm_shuffle_epi8(A3, shuff0); \
279   const __m128i tmp4 = _mm_shuffle_epi8(A4, shuff1); \
280   const __m128i tmp5 = _mm_shuffle_epi8(A5, shuff2); \
281                                                      \
282   /* OR everything to get one channel */             \
283   const __m128i tmp6 = _mm_or_si128(tmp0, tmp1);     \
284   const __m128i tmp7 = _mm_or_si128(tmp3, tmp4);     \
285   out[OUT + 0] = _mm_or_si128(tmp6, tmp2);           \
286   out[OUT + 1] = _mm_or_si128(tmp7, tmp5);           \
287 } while (0);
288 
289 // Unpack the 8b input rgbrgbrgbrgb ... as contiguous registers:
290 // rrrr... rrrr... gggg... gggg... bbbb... bbbb....
291 // Similar to PlanarTo24bHelper(), but in reverse order.
RGB24PackedToPlanar_SSE41(const uint8_t * const rgb,__m128i * const out)292 static WEBP_INLINE void RGB24PackedToPlanar_SSE41(
293     const uint8_t* const rgb, __m128i* const out /*out[6]*/) {
294   const __m128i A0 = _mm_loadu_si128((const __m128i*)(rgb +  0));
295   const __m128i A1 = _mm_loadu_si128((const __m128i*)(rgb + 16));
296   const __m128i A2 = _mm_loadu_si128((const __m128i*)(rgb + 32));
297   const __m128i A3 = _mm_loadu_si128((const __m128i*)(rgb + 48));
298   const __m128i A4 = _mm_loadu_si128((const __m128i*)(rgb + 64));
299   const __m128i A5 = _mm_loadu_si128((const __m128i*)(rgb + 80));
300 
301   // Compute RR.
302   {
303     const __m128i shuff0 = _mm_set_epi8(
304         -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0);
305     const __m128i shuff1 = _mm_set_epi8(
306         -1, -1, -1, -1, -1, 14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1);
307     const __m128i shuff2 = _mm_set_epi8(
308         13, 10, 7, 4, 1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
309     WEBP_SSE41_SHUFF(0)
310   }
311   // Compute GG.
312   {
313     const __m128i shuff0 = _mm_set_epi8(
314         -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1);
315     const __m128i shuff1 = _mm_set_epi8(
316         -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1);
317     const __m128i shuff2 = _mm_set_epi8(
318         14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
319     WEBP_SSE41_SHUFF(2)
320   }
321   // Compute BB.
322   {
323     const __m128i shuff0 = _mm_set_epi8(
324         -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 14, 11, 8, 5, 2);
325     const __m128i shuff1 = _mm_set_epi8(
326         -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1, -1, -1, -1, -1, -1);
327     const __m128i shuff2 = _mm_set_epi8(
328         15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
329     WEBP_SSE41_SHUFF(4)
330   }
331 }
332 
333 #undef WEBP_SSE41_SHUFF
334 
335 // Convert 8 packed ARGB to r[], g[], b[]
RGB32PackedToPlanar_SSE41(const uint32_t * const argb,__m128i * const rgb)336 static WEBP_INLINE void RGB32PackedToPlanar_SSE41(
337     const uint32_t* const argb, __m128i* const rgb /*in[6]*/) {
338   const __m128i zero = _mm_setzero_si128();
339   __m128i a0 = LOAD_16(argb + 0);
340   __m128i a1 = LOAD_16(argb + 4);
341   __m128i a2 = LOAD_16(argb + 8);
342   __m128i a3 = LOAD_16(argb + 12);
343   VP8L32bToPlanar_SSE41(&a0, &a1, &a2, &a3);
344   rgb[0] = _mm_unpacklo_epi8(a1, zero);
345   rgb[1] = _mm_unpackhi_epi8(a1, zero);
346   rgb[2] = _mm_unpacklo_epi8(a2, zero);
347   rgb[3] = _mm_unpackhi_epi8(a2, zero);
348   rgb[4] = _mm_unpacklo_epi8(a3, zero);
349   rgb[5] = _mm_unpackhi_epi8(a3, zero);
350 }
351 
352 // This macro computes (RG * MULT_RG + GB * MULT_GB + ROUNDER) >> DESCALE_FIX
353 // It's a macro and not a function because we need to use immediate values with
354 // srai_epi32, e.g.
355 #define TRANSFORM(RG_LO, RG_HI, GB_LO, GB_HI, MULT_RG, MULT_GB, \
356                   ROUNDER, DESCALE_FIX, OUT) do {               \
357   const __m128i V0_lo = _mm_madd_epi16(RG_LO, MULT_RG);         \
358   const __m128i V0_hi = _mm_madd_epi16(RG_HI, MULT_RG);         \
359   const __m128i V1_lo = _mm_madd_epi16(GB_LO, MULT_GB);         \
360   const __m128i V1_hi = _mm_madd_epi16(GB_HI, MULT_GB);         \
361   const __m128i V2_lo = _mm_add_epi32(V0_lo, V1_lo);            \
362   const __m128i V2_hi = _mm_add_epi32(V0_hi, V1_hi);            \
363   const __m128i V3_lo = _mm_add_epi32(V2_lo, ROUNDER);          \
364   const __m128i V3_hi = _mm_add_epi32(V2_hi, ROUNDER);          \
365   const __m128i V5_lo = _mm_srai_epi32(V3_lo, DESCALE_FIX);     \
366   const __m128i V5_hi = _mm_srai_epi32(V3_hi, DESCALE_FIX);     \
367   (OUT) = _mm_packs_epi32(V5_lo, V5_hi);                        \
368 } while (0)
369 
370 #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)371 static WEBP_INLINE void ConvertRGBToY_SSE41(const __m128i* const R,
372                                             const __m128i* const G,
373                                             const __m128i* const B,
374                                             __m128i* const Y) {
375   const __m128i kRG_y = MK_CST_16(16839, 33059 - 16384);
376   const __m128i kGB_y = MK_CST_16(16384, 6420);
377   const __m128i kHALF_Y = _mm_set1_epi32((16 << YUV_FIX) + YUV_HALF);
378 
379   const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G);
380   const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G);
381   const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B);
382   const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B);
383   TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_y, kGB_y, kHALF_Y, YUV_FIX, *Y);
384 }
385 
ConvertRGBToUV_SSE41(const __m128i * const R,const __m128i * const G,const __m128i * const B,__m128i * const U,__m128i * const V)386 static WEBP_INLINE void ConvertRGBToUV_SSE41(const __m128i* const R,
387                                              const __m128i* const G,
388                                              const __m128i* const B,
389                                              __m128i* const U,
390                                              __m128i* const V) {
391   const __m128i kRG_u = MK_CST_16(-9719, -19081);
392   const __m128i kGB_u = MK_CST_16(0, 28800);
393   const __m128i kRG_v = MK_CST_16(28800, 0);
394   const __m128i kGB_v = MK_CST_16(-24116, -4684);
395   const __m128i kHALF_UV = _mm_set1_epi32(((128 << YUV_FIX) + YUV_HALF) << 2);
396 
397   const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G);
398   const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G);
399   const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B);
400   const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B);
401   TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_u, kGB_u,
402             kHALF_UV, YUV_FIX + 2, *U);
403   TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_v, kGB_v,
404             kHALF_UV, YUV_FIX + 2, *V);
405 }
406 
407 #undef MK_CST_16
408 #undef TRANSFORM
409 
ConvertRGB24ToY_SSE41(const uint8_t * rgb,uint8_t * y,int width)410 static void ConvertRGB24ToY_SSE41(const uint8_t* rgb, uint8_t* y, int width) {
411   const int max_width = width & ~31;
412   int i;
413   for (i = 0; i < max_width; rgb += 3 * 16 * 2) {
414     __m128i rgb_plane[6];
415     int j;
416 
417     RGB24PackedToPlanar_SSE41(rgb, rgb_plane);
418 
419     for (j = 0; j < 2; ++j, i += 16) {
420       const __m128i zero = _mm_setzero_si128();
421       __m128i r, g, b, Y0, Y1;
422 
423       // Convert to 16-bit Y.
424       r = _mm_unpacklo_epi8(rgb_plane[0 + j], zero);
425       g = _mm_unpacklo_epi8(rgb_plane[2 + j], zero);
426       b = _mm_unpacklo_epi8(rgb_plane[4 + j], zero);
427       ConvertRGBToY_SSE41(&r, &g, &b, &Y0);
428 
429       // Convert to 16-bit Y.
430       r = _mm_unpackhi_epi8(rgb_plane[0 + j], zero);
431       g = _mm_unpackhi_epi8(rgb_plane[2 + j], zero);
432       b = _mm_unpackhi_epi8(rgb_plane[4 + j], zero);
433       ConvertRGBToY_SSE41(&r, &g, &b, &Y1);
434 
435       // Cast to 8-bit and store.
436       STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
437     }
438   }
439   for (; i < width; ++i, rgb += 3) {   // left-over
440     y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
441   }
442 }
443 
ConvertBGR24ToY_SSE41(const uint8_t * bgr,uint8_t * y,int width)444 static void ConvertBGR24ToY_SSE41(const uint8_t* bgr, uint8_t* y, int width) {
445   const int max_width = width & ~31;
446   int i;
447   for (i = 0; i < max_width; bgr += 3 * 16 * 2) {
448     __m128i bgr_plane[6];
449     int j;
450 
451     RGB24PackedToPlanar_SSE41(bgr, bgr_plane);
452 
453     for (j = 0; j < 2; ++j, i += 16) {
454       const __m128i zero = _mm_setzero_si128();
455       __m128i r, g, b, Y0, Y1;
456 
457       // Convert to 16-bit Y.
458       b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero);
459       g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero);
460       r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero);
461       ConvertRGBToY_SSE41(&r, &g, &b, &Y0);
462 
463       // Convert to 16-bit Y.
464       b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero);
465       g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero);
466       r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero);
467       ConvertRGBToY_SSE41(&r, &g, &b, &Y1);
468 
469       // Cast to 8-bit and store.
470       STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
471     }
472   }
473   for (; i < width; ++i, bgr += 3) {  // left-over
474     y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
475   }
476 }
477 
ConvertARGBToY_SSE41(const uint32_t * argb,uint8_t * y,int width)478 static void ConvertARGBToY_SSE41(const uint32_t* argb, uint8_t* y, int width) {
479   const int max_width = width & ~15;
480   int i;
481   for (i = 0; i < max_width; i += 16) {
482     __m128i Y0, Y1, rgb[6];
483     RGB32PackedToPlanar_SSE41(&argb[i], rgb);
484     ConvertRGBToY_SSE41(&rgb[0], &rgb[2], &rgb[4], &Y0);
485     ConvertRGBToY_SSE41(&rgb[1], &rgb[3], &rgb[5], &Y1);
486     STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
487   }
488   for (; i < width; ++i) {   // left-over
489     const uint32_t p = argb[i];
490     y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >>  0) & 0xff,
491                      YUV_HALF);
492   }
493 }
494 
495 // Horizontal add (doubled) of two 16b values, result is 16b.
496 // 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)497 static void HorizontalAddPack_SSE41(const __m128i* const A,
498                                     const __m128i* const B,
499                                     __m128i* const out) {
500   const __m128i k2 = _mm_set1_epi16(2);
501   const __m128i C = _mm_madd_epi16(*A, k2);
502   const __m128i D = _mm_madd_epi16(*B, k2);
503   *out = _mm_packs_epi32(C, D);
504 }
505 
ConvertARGBToUV_SSE41(const uint32_t * argb,uint8_t * u,uint8_t * v,int src_width,int do_store)506 static void ConvertARGBToUV_SSE41(const uint32_t* argb,
507                                   uint8_t* u, uint8_t* v,
508                                   int src_width, int do_store) {
509   const int max_width = src_width & ~31;
510   int i;
511   for (i = 0; i < max_width; i += 32, u += 16, v += 16) {
512     __m128i rgb[6], U0, V0, U1, V1;
513     RGB32PackedToPlanar_SSE41(&argb[i], rgb);
514     HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]);
515     HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]);
516     HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]);
517     ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U0, &V0);
518 
519     RGB32PackedToPlanar_SSE41(&argb[i + 16], rgb);
520     HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]);
521     HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]);
522     HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]);
523     ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U1, &V1);
524 
525     U0 = _mm_packus_epi16(U0, U1);
526     V0 = _mm_packus_epi16(V0, V1);
527     if (!do_store) {
528       const __m128i prev_u = LOAD_16(u);
529       const __m128i prev_v = LOAD_16(v);
530       U0 = _mm_avg_epu8(U0, prev_u);
531       V0 = _mm_avg_epu8(V0, prev_v);
532     }
533     STORE_16(U0, u);
534     STORE_16(V0, v);
535   }
536   if (i < src_width) {  // left-over
537     WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store);
538   }
539 }
540 
541 // 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)542 static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE41(
543     const uint16_t* const rgbx,
544     __m128i* const r, __m128i* const g, __m128i* const b) {
545   const __m128i in0 = LOAD_16(rgbx +  0);  // r0 | g0 | b0 |x| r1 | g1 | b1 |x
546   const __m128i in1 = LOAD_16(rgbx +  8);  // r2 | g2 | b2 |x| r3 | g3 | b3 |x
547   const __m128i in2 = LOAD_16(rgbx + 16);  // r4 | ...
548   const __m128i in3 = LOAD_16(rgbx + 24);  // r6 | ...
549   // aarrggbb as 16-bit.
550   const __m128i shuff0 =
551       _mm_set_epi8(-1, -1, -1, -1, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
552   const __m128i shuff1 =
553       _mm_set_epi8(13, 12, 5, 4, -1, -1, -1, -1, 11, 10, 3, 2, 9, 8, 1, 0);
554   const __m128i A0 = _mm_shuffle_epi8(in0, shuff0);
555   const __m128i A1 = _mm_shuffle_epi8(in1, shuff1);
556   const __m128i A2 = _mm_shuffle_epi8(in2, shuff0);
557   const __m128i A3 = _mm_shuffle_epi8(in3, shuff1);
558   // R0R1G0G1
559   // B0B1****
560   // R2R3G2G3
561   // B2B3****
562   // (OR is used to free port 5 for the unpack)
563   const __m128i B0 = _mm_unpacklo_epi32(A0, A1);
564   const __m128i B1 = _mm_or_si128(A0, A1);
565   const __m128i B2 = _mm_unpacklo_epi32(A2, A3);
566   const __m128i B3 = _mm_or_si128(A2, A3);
567   // Gather the channels.
568   *r = _mm_unpacklo_epi64(B0, B2);
569   *g = _mm_unpackhi_epi64(B0, B2);
570   *b = _mm_unpackhi_epi64(B1, B3);
571 }
572 
ConvertRGBA32ToUV_SSE41(const uint16_t * rgb,uint8_t * u,uint8_t * v,int width)573 static void ConvertRGBA32ToUV_SSE41(const uint16_t* rgb,
574                                     uint8_t* u, uint8_t* v, int width) {
575   const int max_width = width & ~15;
576   const uint16_t* const last_rgb = rgb + 4 * max_width;
577   while (rgb < last_rgb) {
578     __m128i r, g, b, U0, V0, U1, V1;
579     RGBA32PackedToPlanar_16b_SSE41(rgb +  0, &r, &g, &b);
580     ConvertRGBToUV_SSE41(&r, &g, &b, &U0, &V0);
581     RGBA32PackedToPlanar_16b_SSE41(rgb + 32, &r, &g, &b);
582     ConvertRGBToUV_SSE41(&r, &g, &b, &U1, &V1);
583     STORE_16(_mm_packus_epi16(U0, U1), u);
584     STORE_16(_mm_packus_epi16(V0, V1), v);
585     u += 16;
586     v += 16;
587     rgb += 2 * 32;
588   }
589   if (max_width < width) {  // left-over
590     WebPConvertRGBA32ToUV_C(rgb, u, v, width - max_width);
591   }
592 }
593 
594 //------------------------------------------------------------------------------
595 
596 extern void WebPInitConvertARGBToYUVSSE41(void);
597 
WebPInitConvertARGBToYUVSSE41(void)598 WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE41(void) {
599   WebPConvertARGBToY = ConvertARGBToY_SSE41;
600   WebPConvertARGBToUV = ConvertARGBToUV_SSE41;
601 
602   WebPConvertRGB24ToY = ConvertRGB24ToY_SSE41;
603   WebPConvertBGR24ToY = ConvertBGR24ToY_SSE41;
604 
605   WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_SSE41;
606 }
607 
608 //------------------------------------------------------------------------------
609 
610 #else  // !WEBP_USE_SSE41
611 
612 WEBP_DSP_INIT_STUB(WebPInitSamplersSSE41)
613 WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE41)
614 
615 #endif  // WEBP_USE_SSE41
616