Imported Upstream version 2.0.1
[platform/upstream/libjpeg-turbo.git] / simd / loongson / jidctint-mmi.c
1 /*
2  * Loongson MMI optimizations for libjpeg-turbo
3  *
4  * Copyright (C) 2014-2015, 2018, D. R. Commander.  All Rights Reserved.
5  * Copyright (C) 2016-2017, Loongson Technology Corporation Limited, BeiJing.
6  *                          All Rights Reserved.
7  * Authors:  ZhuChen     <zhuchen@loongson.cn>
8  *           CaiWanwei   <caiwanwei@loongson.cn>
9  *           SunZhangzhi <sunzhangzhi-cq@loongson.cn>
10  *
11  * Based on the x86 SIMD extension for IJG JPEG library
12  * Copyright (C) 1999-2006, MIYASAKA Masaru.
13  *
14  * This software is provided 'as-is', without any express or implied
15  * warranty.  In no event will the authors be held liable for any damages
16  * arising from the use of this software.
17  *
18  * Permission is granted to anyone to use this software for any purpose,
19  * including commercial applications, and to alter it and redistribute it
20  * freely, subject to the following restrictions:
21  *
22  * 1. The origin of this software must not be misrepresented; you must not
23  *    claim that you wrote the original software. If you use this software
24  *    in a product, an acknowledgment in the product documentation would be
25  *    appreciated but is not required.
26  * 2. Altered source versions must be plainly marked as such, and must not be
27  *    misrepresented as being the original software.
28  * 3. This notice may not be removed or altered from any source distribution.
29  */
30
31 /* SLOW INTEGER INVERSE DCT */
32
33 #include "jsimd_mmi.h"
34
35
36 #define CONST_BITS  13
37 #define PASS1_BITS  2
38 #define DESCALE_P1  (CONST_BITS - PASS1_BITS)
39 #define DESCALE_P2  (CONST_BITS + PASS1_BITS + 3)
40 #define CENTERJSAMPLE  128
41
42 #define FIX_0_298  ((short)2446)  /* FIX(0.298631336) */
43 #define FIX_0_390  ((short)3196)  /* FIX(0.390180644) */
44 #define FIX_0_899  ((short)7373)  /* FIX(0.899976223) */
45 #define FIX_0_541  ((short)4433)  /* FIX(0.541196100) */
46 #define FIX_0_765  ((short)6270)  /* FIX(0.765366865) */
47 #define FIX_1_175  ((short)9633)  /* FIX(1.175875602) */
48 #define FIX_1_501  ((short)12299) /* FIX(1.501321110) */
49 #define FIX_1_847  ((short)15137) /* FIX(1.847759065) */
50 #define FIX_1_961  ((short)16069) /* FIX(1.961570560) */
51 #define FIX_2_053  ((short)16819) /* FIX(2.053119869) */
52 #define FIX_2_562  ((short)20995) /* FIX(2.562915447) */
53 #define FIX_3_072  ((short)25172) /* FIX(3.072711026) */
54
55 enum const_index {
56   index_PW_F130_F054,
57   index_PW_F054_MF130,
58   index_PW_MF078_F117,
59   index_PW_F117_F078,
60   index_PW_MF060_MF089,
61   index_PW_MF089_F060,
62   index_PW_MF050_MF256,
63   index_PW_MF256_F050,
64   index_PD_DESCALE_P1,
65   index_PD_DESCALE_P2,
66   index_PB_CENTERJSAMP
67 };
68
69 static uint64_t const_value[] = {
70   _uint64_set_pi16(FIX_0_541, (FIX_0_541 + FIX_0_765),
71                    FIX_0_541, (FIX_0_541 + FIX_0_765)),
72   _uint64_set_pi16((FIX_0_541 - FIX_1_847), FIX_0_541,
73                    (FIX_0_541 - FIX_1_847), FIX_0_541),
74   _uint64_set_pi16(FIX_1_175, (FIX_1_175 - FIX_1_961),
75                    FIX_1_175, (FIX_1_175 - FIX_1_961)),
76   _uint64_set_pi16((FIX_1_175 - FIX_0_390), FIX_1_175,
77                    (FIX_1_175 - FIX_0_390), FIX_1_175),
78   _uint64_set_pi16(-FIX_0_899, (FIX_0_298 - FIX_0_899),
79                    -FIX_0_899, (FIX_0_298 - FIX_0_899)),
80   _uint64_set_pi16((FIX_1_501 - FIX_0_899), -FIX_0_899,
81                    (FIX_1_501 - FIX_0_899), -FIX_0_899),
82   _uint64_set_pi16(-FIX_2_562, (FIX_2_053 - FIX_2_562),
83                    -FIX_2_562, (FIX_2_053 - FIX_2_562)),
84   _uint64_set_pi16((FIX_3_072 - FIX_2_562), -FIX_2_562,
85                    (FIX_3_072 - FIX_2_562), -FIX_2_562),
86   _uint64_set_pi32((1 << (DESCALE_P1 - 1)), (1 << (DESCALE_P1 - 1))),
87   _uint64_set_pi32((1 << (DESCALE_P2 - 1)), (1 << (DESCALE_P2 - 1))),
88   _uint64_set_pi8(CENTERJSAMPLE, CENTERJSAMPLE, CENTERJSAMPLE, CENTERJSAMPLE,
89                   CENTERJSAMPLE, CENTERJSAMPLE, CENTERJSAMPLE, CENTERJSAMPLE)
90 };
91
92 #define PW_F130_F054    get_const_value(index_PW_F130_F054)
93 #define PW_F054_MF130   get_const_value(index_PW_F054_MF130)
94 #define PW_MF078_F117   get_const_value(index_PW_MF078_F117)
95 #define PW_F117_F078    get_const_value(index_PW_F117_F078)
96 #define PW_MF060_MF089  get_const_value(index_PW_MF060_MF089)
97 #define PW_MF089_F060   get_const_value(index_PW_MF089_F060)
98 #define PW_MF050_MF256  get_const_value(index_PW_MF050_MF256)
99 #define PW_MF256_F050   get_const_value(index_PW_MF256_F050)
100 #define PD_DESCALE_P1   get_const_value(index_PD_DESCALE_P1)
101 #define PD_DESCALE_P2   get_const_value(index_PD_DESCALE_P2)
102 #define PB_CENTERJSAMP  get_const_value(index_PB_CENTERJSAMP)
103
104
105 #define test_m32_zero(mm32)  (!(*(uint32_t *)&mm32))
106 #define test_m64_zero(mm64)  (!(*(uint64_t *)&mm64))
107
108
109 #define DO_IDCT_COMMON(PASS) { \
110   __m64 tmp0_3l, tmp0_3h, tmp1_2l, tmp1_2h; \
111   __m64 tmp0l, tmp0h, tmp1l, tmp1h, tmp2l, tmp2h, tmp3l, tmp3h; \
112   __m64 z34l, z34h, z3l, z3h, z4l, z4h, z3, z4; \
113   __m64 out0l, out0h, out1l, out1h, out2l, out2h, out3l, out3h; \
114   __m64 out4l, out4h, out5l, out5h, out6l, out6h, out7l, out7h; \
115   \
116   z3 = _mm_add_pi16(tmp0, tmp2); \
117   z4 = _mm_add_pi16(tmp1, tmp3); \
118   \
119   /* (Original) \
120    * z5 = (z3 + z4) * 1.175875602; \
121    * z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644; \
122    * z3 += z5;  z4 += z5; \
123    * \
124    * (This implementation) \
125    * z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; \
126    * z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); \
127    */ \
128   \
129   z34l = _mm_unpacklo_pi16(z3, z4); \
130   z34h = _mm_unpackhi_pi16(z3, z4); \
131   z3l = _mm_madd_pi16(z34l, PW_MF078_F117); \
132   z3h = _mm_madd_pi16(z34h, PW_MF078_F117); \
133   z4l = _mm_madd_pi16(z34l, PW_F117_F078); \
134   z4h = _mm_madd_pi16(z34h, PW_F117_F078); \
135   \
136   /* (Original) \
137    * z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2; \
138    * tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869; \
139    * tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110; \
140    * z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447; \
141    * tmp0 += z1 + z3;  tmp1 += z2 + z4; \
142    * tmp2 += z2 + z3;  tmp3 += z1 + z4; \
143    * \
144    * (This implementation) \
145    * tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223; \
146    * tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447; \
147    * tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447); \
148    * tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223); \
149    * tmp0 += z3;  tmp1 += z4; \
150    * tmp2 += z3;  tmp3 += z4; \
151    */ \
152   \
153   tmp0_3l = _mm_unpacklo_pi16(tmp0, tmp3); \
154   tmp0_3h = _mm_unpackhi_pi16(tmp0, tmp3); \
155   \
156   tmp0l = _mm_madd_pi16(tmp0_3l, PW_MF060_MF089); \
157   tmp0h = _mm_madd_pi16(tmp0_3h, PW_MF060_MF089); \
158   tmp3l = _mm_madd_pi16(tmp0_3l, PW_MF089_F060); \
159   tmp3h = _mm_madd_pi16(tmp0_3h, PW_MF089_F060); \
160   \
161   tmp0l = _mm_add_pi32(tmp0l, z3l); \
162   tmp0h = _mm_add_pi32(tmp0h, z3h); \
163   tmp3l = _mm_add_pi32(tmp3l, z4l); \
164   tmp3h = _mm_add_pi32(tmp3h, z4h); \
165   \
166   tmp1_2l = _mm_unpacklo_pi16(tmp1, tmp2); \
167   tmp1_2h = _mm_unpackhi_pi16(tmp1, tmp2); \
168   \
169   tmp1l = _mm_madd_pi16(tmp1_2l, PW_MF050_MF256); \
170   tmp1h = _mm_madd_pi16(tmp1_2h, PW_MF050_MF256); \
171   tmp2l = _mm_madd_pi16(tmp1_2l, PW_MF256_F050); \
172   tmp2h = _mm_madd_pi16(tmp1_2h, PW_MF256_F050); \
173   \
174   tmp1l = _mm_add_pi32(tmp1l, z4l); \
175   tmp1h = _mm_add_pi32(tmp1h, z4h); \
176   tmp2l = _mm_add_pi32(tmp2l, z3l); \
177   tmp2h = _mm_add_pi32(tmp2h, z3h); \
178   \
179   /* Final output stage */ \
180   \
181   out0l = _mm_add_pi32(tmp10l, tmp3l); \
182   out0h = _mm_add_pi32(tmp10h, tmp3h); \
183   out7l = _mm_sub_pi32(tmp10l, tmp3l); \
184   out7h = _mm_sub_pi32(tmp10h, tmp3h); \
185   \
186   out0l = _mm_add_pi32(out0l, PD_DESCALE_P##PASS); \
187   out0h = _mm_add_pi32(out0h, PD_DESCALE_P##PASS); \
188   out0l = _mm_srai_pi32(out0l, DESCALE_P##PASS); \
189   out0h = _mm_srai_pi32(out0h, DESCALE_P##PASS); \
190   \
191   out7l = _mm_add_pi32(out7l, PD_DESCALE_P##PASS); \
192   out7h = _mm_add_pi32(out7h, PD_DESCALE_P##PASS); \
193   out7l = _mm_srai_pi32(out7l, DESCALE_P##PASS); \
194   out7h = _mm_srai_pi32(out7h, DESCALE_P##PASS); \
195   \
196   out0 = _mm_packs_pi32(out0l, out0h); \
197   out7 = _mm_packs_pi32(out7l, out7h); \
198   \
199   out1l = _mm_add_pi32(tmp11l, tmp2l); \
200   out1h = _mm_add_pi32(tmp11h, tmp2h); \
201   out6l = _mm_sub_pi32(tmp11l, tmp2l); \
202   out6h = _mm_sub_pi32(tmp11h, tmp2h); \
203   \
204   out1l = _mm_add_pi32(out1l, PD_DESCALE_P##PASS); \
205   out1h = _mm_add_pi32(out1h, PD_DESCALE_P##PASS); \
206   out1l = _mm_srai_pi32(out1l, DESCALE_P##PASS); \
207   out1h = _mm_srai_pi32(out1h, DESCALE_P##PASS); \
208   \
209   out6l = _mm_add_pi32(out6l, PD_DESCALE_P##PASS); \
210   out6h = _mm_add_pi32(out6h, PD_DESCALE_P##PASS); \
211   out6l = _mm_srai_pi32(out6l, DESCALE_P##PASS); \
212   out6h = _mm_srai_pi32(out6h, DESCALE_P##PASS); \
213   \
214   out1 = _mm_packs_pi32(out1l, out1h); \
215   out6 = _mm_packs_pi32(out6l, out6h); \
216   \
217   out2l = _mm_add_pi32(tmp12l, tmp1l); \
218   out2h = _mm_add_pi32(tmp12h, tmp1h); \
219   out5l = _mm_sub_pi32(tmp12l, tmp1l); \
220   out5h = _mm_sub_pi32(tmp12h, tmp1h); \
221   \
222   out2l = _mm_add_pi32(out2l, PD_DESCALE_P##PASS); \
223   out2h = _mm_add_pi32(out2h, PD_DESCALE_P##PASS); \
224   out2l = _mm_srai_pi32(out2l, DESCALE_P##PASS); \
225   out2h = _mm_srai_pi32(out2h, DESCALE_P##PASS); \
226   \
227   out5l = _mm_add_pi32(out5l, PD_DESCALE_P##PASS); \
228   out5h = _mm_add_pi32(out5h, PD_DESCALE_P##PASS); \
229   out5l = _mm_srai_pi32(out5l, DESCALE_P##PASS); \
230   out5h = _mm_srai_pi32(out5h, DESCALE_P##PASS); \
231   \
232   out2 = _mm_packs_pi32(out2l, out2h); \
233   out5 = _mm_packs_pi32(out5l, out5h); \
234   \
235   out3l = _mm_add_pi32(tmp13l, tmp0l); \
236   out3h = _mm_add_pi32(tmp13h, tmp0h); \
237   \
238   out4l = _mm_sub_pi32(tmp13l, tmp0l); \
239   out4h = _mm_sub_pi32(tmp13h, tmp0h); \
240   \
241   out3l = _mm_add_pi32(out3l, PD_DESCALE_P##PASS); \
242   out3h = _mm_add_pi32(out3h, PD_DESCALE_P##PASS); \
243   out3l = _mm_srai_pi32(out3l, DESCALE_P##PASS); \
244   out3h = _mm_srai_pi32(out3h, DESCALE_P##PASS); \
245   \
246   out4l = _mm_add_pi32(out4l, PD_DESCALE_P##PASS); \
247   out4h = _mm_add_pi32(out4h, PD_DESCALE_P##PASS); \
248   out4l = _mm_srai_pi32(out4l, DESCALE_P##PASS); \
249   out4h = _mm_srai_pi32(out4h, DESCALE_P##PASS); \
250   \
251   out3 = _mm_packs_pi32(out3l, out3h); \
252   out4 = _mm_packs_pi32(out4l, out4h); \
253 }
254
255 #define DO_IDCT_PASS1(iter) { \
256   __m64 col0l, col1l, col2l, col3l, col4l, col5l, col6l, col7l; \
257   __m64 quant0l, quant1l, quant2l, quant3l; \
258   __m64 quant4l, quant5l, quant6l, quant7l; \
259   __m64 z23, z2, z3, z23l, z23h; \
260   __m64 row01a, row01b, row01c, row01d, row23a, row23b, row23c, row23d; \
261   __m64 row0l, row0h, row1l, row1h, row2l, row2h, row3l, row3h; \
262   __m64 tmp0l, tmp0h, tmp1l, tmp1h, tmp2l, tmp2h, tmp3l, tmp3h; \
263   __m64 tmp10l, tmp10h, tmp11l, tmp11h, tmp12l, tmp12h, tmp13l, tmp13h; \
264   __m32 col0a, col1a, mm0; \
265   \
266   col0a = _mm_load_si32((__m32 *)&inptr[DCTSIZE * 1]); \
267   col1a = _mm_load_si32((__m32 *)&inptr[DCTSIZE * 2]); \
268   mm0 = _mm_or_si32(col0a, col1a); \
269   \
270   if (test_m32_zero(mm0)) { \
271     __m64 mm1, mm2; \
272     \
273     col0l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 0]); \
274     col1l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 1]); \
275     col2l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 2]); \
276     col3l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 3]); \
277     col4l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 4]); \
278     col5l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 5]); \
279     col6l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 6]); \
280     col7l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 7]); \
281     \
282     mm1 = _mm_or_si64(col1l, col3l); \
283     mm2 = _mm_or_si64(col2l, col4l); \
284     mm1 = _mm_or_si64(mm1, col5l); \
285     mm2 = _mm_or_si64(mm2, col6l); \
286     mm1 = _mm_or_si64(mm1, col7l); \
287     mm1 = _mm_or_si64(mm1, mm2); \
288     \
289     if (test_m64_zero(mm1)) { \
290       __m64 dcval, dcvall, dcvalh, row0, row1, row2, row3; \
291       \
292       /* AC terms all zero */ \
293       \
294       quant0l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 0]); \
295       \
296       dcval = _mm_mullo_pi16(col0l, quant0l); \
297       dcval = _mm_slli_pi16(dcval, PASS1_BITS);  /* dcval=(00 10 20 30) */ \
298       \
299       dcvall = _mm_unpacklo_pi16(dcval, dcval);  /* dcvall=(00 00 10 10) */ \
300       dcvalh = _mm_unpackhi_pi16(dcval, dcval);  /* dcvalh=(20 20 30 30) */ \
301       \
302       row0 = _mm_unpacklo_pi32(dcvall, dcvall);  /* row0=(00 00 00 00) */ \
303       row1 = _mm_unpackhi_pi32(dcvall, dcvall);  /* row1=(10 10 10 10) */ \
304       row2 = _mm_unpacklo_pi32(dcvalh, dcvalh);  /* row2=(20 20 20 20) */ \
305       row3 = _mm_unpackhi_pi32(dcvalh, dcvalh);  /* row3=(30 30 30 30) */ \
306       \
307       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 0], row0); \
308       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 0 + 4], row0); \
309       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 1], row1); \
310       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 1 + 4], row1); \
311       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 2], row2); \
312       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 2 + 4], row2); \
313       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 3], row3); \
314       _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 3 + 4], row3); \
315       \
316       goto nextcolumn##iter; \
317     } \
318   } \
319   \
320   /* Even part \
321    * \
322    * (Original) \
323    * z1 = (z2 + z3) * 0.541196100; \
324    * tmp2 = z1 + z3 * -1.847759065; \
325    * tmp3 = z1 + z2 * 0.765366865; \
326    * \
327    * (This implementation) \
328    * tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065); \
329    * tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100; \
330    */ \
331   \
332   col0l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 0]);  /* (00 10 20 30) */ \
333   col2l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 2]);  /* (02 12 22 32) */ \
334   col4l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 4]);  /* (04 14 24 34) */ \
335   col6l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 6]);  /* (06 16 26 36) */ \
336   \
337   quant0l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 0]); \
338   quant2l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 2]); \
339   quant4l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 4]); \
340   quant6l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 6]); \
341   \
342   z2 = _mm_mullo_pi16(col2l, quant2l); \
343   z3 = _mm_mullo_pi16(col6l, quant6l); \
344   \
345   z23l = _mm_unpacklo_pi16(z2, z3); \
346   z23h = _mm_unpackhi_pi16(z2, z3); \
347   tmp3l = _mm_madd_pi16(z23l, PW_F130_F054); \
348   tmp3h = _mm_madd_pi16(z23h, PW_F130_F054); \
349   tmp2l = _mm_madd_pi16(z23l, PW_F054_MF130); \
350   tmp2h = _mm_madd_pi16(z23h, PW_F054_MF130); \
351   \
352   z2 = _mm_mullo_pi16(col0l, quant0l); \
353   z3 = _mm_mullo_pi16(col4l, quant4l); \
354   \
355   z23 = _mm_add_pi16(z2, z3); \
356   tmp0l = _mm_loadlo_pi16_f(z23); \
357   tmp0h = _mm_loadhi_pi16_f(z23); \
358   tmp0l = _mm_srai_pi32(tmp0l, (16 - CONST_BITS)); \
359   tmp0h = _mm_srai_pi32(tmp0h, (16 - CONST_BITS)); \
360   \
361   tmp10l = _mm_add_pi32(tmp0l, tmp3l); \
362   tmp10h = _mm_add_pi32(tmp0h, tmp3h); \
363   tmp13l = _mm_sub_pi32(tmp0l, tmp3l); \
364   tmp13h = _mm_sub_pi32(tmp0h, tmp3h); \
365   \
366   z23 = _mm_sub_pi16(z2, z3); \
367   tmp1l = _mm_loadlo_pi16_f(z23); \
368   tmp1h = _mm_loadhi_pi16_f(z23); \
369   tmp1l = _mm_srai_pi32(tmp1l, (16 - CONST_BITS)); \
370   tmp1h = _mm_srai_pi32(tmp1h, (16 - CONST_BITS)); \
371   \
372   tmp11l = _mm_add_pi32(tmp1l, tmp2l); \
373   tmp11h = _mm_add_pi32(tmp1h, tmp2h); \
374   tmp12l = _mm_sub_pi32(tmp1l, tmp2l); \
375   tmp12h = _mm_sub_pi32(tmp1h, tmp2h); \
376   \
377   /* Odd part */ \
378   \
379   col1l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 1]);  /* (01 11 21 31) */ \
380   col3l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 3]);  /* (03 13 23 33) */ \
381   col5l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 5]);  /* (05 15 25 35) */ \
382   col7l = _mm_load_si64((__m64 *)&inptr[DCTSIZE * 7]);  /* (07 17 27 37) */ \
383   \
384   quant1l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 1]); \
385   quant3l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 3]); \
386   quant5l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 5]); \
387   quant7l = _mm_load_si64((__m64 *)&quantptr[DCTSIZE * 7]); \
388   \
389   tmp0 = _mm_mullo_pi16(col7l, quant7l); \
390   tmp1 = _mm_mullo_pi16(col5l, quant5l); \
391   tmp2 = _mm_mullo_pi16(col3l, quant3l); \
392   tmp3 = _mm_mullo_pi16(col1l, quant1l); \
393   \
394   DO_IDCT_COMMON(1) \
395   \
396   /* out0=(00 10 20 30), out1=(01 11 21 31) */ \
397   /* out2=(02 12 22 32), out3=(03 13 23 33) */ \
398   /* out4=(04 14 24 34), out5=(05 15 25 35) */ \
399   /* out6=(06 16 26 36), out7=(07 17 27 37) */ \
400   \
401   /* Transpose coefficients */ \
402   \
403   row01a = _mm_unpacklo_pi16(out0, out1);     /* row01a=(00 01 10 11) */ \
404   row23a = _mm_unpackhi_pi16(out0, out1);     /* row23a=(20 21 30 31) */ \
405   row01d = _mm_unpacklo_pi16(out6, out7);     /* row01d=(06 07 16 17) */ \
406   row23d = _mm_unpackhi_pi16(out6, out7);     /* row23d=(26 27 36 37) */ \
407   \
408   row01b = _mm_unpacklo_pi16(out2, out3);     /* row01b=(02 03 12 13) */ \
409   row23b = _mm_unpackhi_pi16(out2, out3);     /* row23b=(22 23 32 33) */ \
410   row01c = _mm_unpacklo_pi16(out4, out5);     /* row01c=(04 05 14 15) */ \
411   row23c = _mm_unpackhi_pi16(out4, out5);     /* row23c=(24 25 34 35) */ \
412   \
413   row0l = _mm_unpacklo_pi32(row01a, row01b);  /* row0l=(00 01 02 03) */ \
414   row1l = _mm_unpackhi_pi32(row01a, row01b);  /* row1l=(10 11 12 13) */ \
415   row2l = _mm_unpacklo_pi32(row23a, row23b);  /* row2l=(20 21 22 23) */ \
416   row3l = _mm_unpackhi_pi32(row23a, row23b);  /* row3l=(30 31 32 33) */ \
417   \
418   row0h = _mm_unpacklo_pi32(row01c, row01d);  /* row0h=(04 05 06 07) */ \
419   row1h = _mm_unpackhi_pi32(row01c, row01d);  /* row1h=(14 15 16 17) */ \
420   row2h = _mm_unpacklo_pi32(row23c, row23d);  /* row2h=(24 25 26 27) */ \
421   row3h = _mm_unpackhi_pi32(row23c, row23d);  /* row3h=(34 35 36 37) */ \
422   \
423   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 0], row0l); \
424   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 0 + 4], row0h); \
425   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 1], row1l); \
426   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 1 + 4], row1h); \
427   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 2], row2l); \
428   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 2 + 4], row2h); \
429   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 3], row3l); \
430   _mm_store_si64((__m64 *)&wsptr[DCTSIZE * 3 + 4], row3h); \
431 }
432
433 #define DO_IDCT_PASS2(ctr) { \
434   __m64 row0l, row1l, row2l, row3l, row4l, row5l, row6l, row7l; \
435   __m64 z23, z23l, z23h; \
436   __m64 col0123a, col0123b, col0123c, col0123d; \
437   __m64 col01l, col01h, col23l, col23h, row06, row17, row24, row35; \
438   __m64 col0, col1, col2, col3; \
439   __m64 tmp0l, tmp0h, tmp1l, tmp1h, tmp2l, tmp2h, tmp3l, tmp3h; \
440   __m64 tmp10l, tmp10h, tmp11l, tmp11h, tmp12l, tmp12h, tmp13l, tmp13h; \
441   \
442   row0l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 0]);  /* (00 01 02 03) */ \
443   row1l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 1]);  /* (10 11 12 13) */ \
444   row2l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 2]);  /* (20 21 22 23) */ \
445   row3l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 3]);  /* (30 31 32 33) */ \
446   row4l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 4]);  /* (40 41 42 43) */ \
447   row5l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 5]);  /* (50 51 52 53) */ \
448   row6l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 6]);  /* (60 61 62 63) */ \
449   row7l = _mm_load_si64((__m64 *)&wsptr[DCTSIZE * 7]);  /* (70 71 72 73) */ \
450   \
451   /* Even part \
452    * \
453    * (Original) \
454    * z1 = (z2 + z3) * 0.541196100; \
455    * tmp2 = z1 + z3 * -1.847759065; \
456    * tmp3 = z1 + z2 * 0.765366865; \
457    * \
458    * (This implementation) \
459    * tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065); \
460    * tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100; \
461    */ \
462   \
463   z23l = _mm_unpacklo_pi16(row2l, row6l); \
464   z23h = _mm_unpackhi_pi16(row2l, row6l); \
465   \
466   tmp3l = _mm_madd_pi16(z23l, PW_F130_F054); \
467   tmp3h = _mm_madd_pi16(z23h, PW_F130_F054); \
468   tmp2l = _mm_madd_pi16(z23l, PW_F054_MF130); \
469   tmp2h = _mm_madd_pi16(z23h, PW_F054_MF130); \
470   \
471   z23 = _mm_add_pi16(row0l, row4l); \
472   tmp0l = _mm_loadlo_pi16_f(z23); \
473   tmp0h = _mm_loadhi_pi16_f(z23); \
474   tmp0l = _mm_srai_pi32(tmp0l, (16 - CONST_BITS)); \
475   tmp0h = _mm_srai_pi32(tmp0h, (16 - CONST_BITS)); \
476   \
477   tmp10l = _mm_add_pi32(tmp0l, tmp3l); \
478   tmp10h = _mm_add_pi32(tmp0h, tmp3h); \
479   tmp13l = _mm_sub_pi32(tmp0l, tmp3l); \
480   tmp13h = _mm_sub_pi32(tmp0h, tmp3h); \
481   \
482   z23 = _mm_sub_pi16(row0l, row4l); \
483   tmp1l = _mm_loadlo_pi16_f(z23); \
484   tmp1h = _mm_loadhi_pi16_f(z23); \
485   tmp1l = _mm_srai_pi32(tmp1l, (16 - CONST_BITS)); \
486   tmp1h = _mm_srai_pi32(tmp1h, (16 - CONST_BITS)); \
487   \
488   tmp11l = _mm_add_pi32(tmp1l, tmp2l); \
489   tmp11h = _mm_add_pi32(tmp1h, tmp2h); \
490   tmp12l = _mm_sub_pi32(tmp1l, tmp2l); \
491   tmp12h = _mm_sub_pi32(tmp1h, tmp2h); \
492   \
493   /* Odd part */ \
494   \
495   tmp0 = row7l; \
496   tmp1 = row5l; \
497   tmp2 = row3l; \
498   tmp3 = row1l; \
499   \
500   DO_IDCT_COMMON(2) \
501   \
502   /* out0=(00 01 02 03), out1=(10 11 12 13) */ \
503   /* out2=(20 21 22 23), out3=(30 31 32 33) */ \
504   /* out4=(40 41 42 43), out5=(50 51 52 53) */ \
505   /* out6=(60 61 62 63), out7=(70 71 72 73) */ \
506   \
507   row06 = _mm_packs_pi16(out0, out6);  /* row06=(00 01 02 03 60 61 62 63) */ \
508   row17 = _mm_packs_pi16(out1, out7);  /* row17=(10 11 12 13 70 71 72 73) */ \
509   row24 = _mm_packs_pi16(out2, out4);  /* row24=(20 21 22 23 40 41 42 43) */ \
510   row35 = _mm_packs_pi16(out3, out5);  /* row35=(30 31 32 33 50 51 52 53) */ \
511   \
512   row06 = _mm_add_pi8(row06, PB_CENTERJSAMP); \
513   row17 = _mm_add_pi8(row17, PB_CENTERJSAMP); \
514   row24 = _mm_add_pi8(row24, PB_CENTERJSAMP); \
515   row35 = _mm_add_pi8(row35, PB_CENTERJSAMP); \
516   \
517   /* Transpose coefficients */ \
518   \
519   col0123a = _mm_unpacklo_pi8(row06, row17);  /* col0123a=(00 10 01 11 02 12 03 13) */ \
520   col0123d = _mm_unpackhi_pi8(row06, row17);  /* col0123d=(60 70 61 71 62 72 63 73) */ \
521   col0123b = _mm_unpacklo_pi8(row24, row35);  /* col0123b=(20 30 21 31 22 32 23 33) */ \
522   col0123c = _mm_unpackhi_pi8(row24, row35);  /* col0123c=(40 50 41 51 42 52 43 53) */ \
523   \
524   col01l = _mm_unpacklo_pi16(col0123a, col0123b);  /* col01l=(00 10 20 30 01 11 21 31) */ \
525   col23l = _mm_unpackhi_pi16(col0123a, col0123b);  /* col23l=(02 12 22 32 03 13 23 33) */ \
526   col01h = _mm_unpacklo_pi16(col0123c, col0123d);  /* col01h=(40 50 60 70 41 51 61 71) */ \
527   col23h = _mm_unpackhi_pi16(col0123c, col0123d);  /* col23h=(42 52 62 72 43 53 63 73) */ \
528   \
529   col0 = _mm_unpacklo_pi32(col01l, col01h);   /* col0=(00 10 20 30 40 50 60 70) */ \
530   col1 = _mm_unpackhi_pi32(col01l, col01h);   /* col1=(01 11 21 31 41 51 61 71) */ \
531   col2 = _mm_unpacklo_pi32(col23l, col23h);   /* col2=(02 12 22 32 42 52 62 72) */ \
532   col3 = _mm_unpackhi_pi32(col23l, col23h);   /* col3=(03 13 23 33 43 53 63 73) */ \
533   \
534   _mm_store_si64((__m64 *)(output_buf[ctr + 0] + output_col), col0); \
535   _mm_store_si64((__m64 *)(output_buf[ctr + 1] + output_col), col1); \
536   _mm_store_si64((__m64 *)(output_buf[ctr + 2] + output_col), col2); \
537   _mm_store_si64((__m64 *)(output_buf[ctr + 3] + output_col), col3); \
538 }
539
540 void jsimd_idct_islow_mmi(void *dct_table, JCOEFPTR coef_block,
541                           JSAMPARRAY output_buf, JDIMENSION output_col)
542 {
543   __m64 tmp0, tmp1, tmp2, tmp3;
544   __m64 out0, out1, out2, out3, out4, out5, out6, out7;
545   JCOEFPTR inptr;
546   ISLOW_MULT_TYPE *quantptr;
547   JCOEF *wsptr;
548   JCOEF workspace[DCTSIZE2];  /* buffers data between passes */
549
550   /* Pass 1: process columns. */
551
552   inptr = coef_block;
553   quantptr = (ISLOW_MULT_TYPE *)dct_table;
554   wsptr = workspace;
555
556   DO_IDCT_PASS1(1)
557 nextcolumn1:
558   inptr += 4;
559   quantptr += 4;
560   wsptr += DCTSIZE * 4;
561   DO_IDCT_PASS1(2)
562 nextcolumn2:
563
564   /* Pass 2: process rows. */
565
566   wsptr = workspace;
567
568   DO_IDCT_PASS2(0)
569   wsptr += 4;
570   DO_IDCT_PASS2(4)
571 }