• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3  *
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.
9  */
10 
11 #include <emmintrin.h>  // SSE2
12 
13 #include "./vpx_dsp_rtcd.h"
14 #include "vpx_dsp/x86/inv_txfm_sse2.h"
15 #include "vpx_dsp/x86/transpose_sse2.h"
16 #include "vpx_dsp/x86/txfm_common_sse2.h"
17 
transpose_16bit_4(__m128i * res)18 static INLINE void transpose_16bit_4(__m128i *res) {
19   const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
20   const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
21 
22   res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
23   res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
24 }
25 
vpx_idct4x4_16_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)26 void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest,
27                              int stride) {
28   const __m128i eight = _mm_set1_epi16(8);
29   __m128i in[2];
30 
31   // Rows
32   in[0] = load_input_data8(input);
33   in[1] = load_input_data8(input + 8);
34   idct4_sse2(in);
35 
36   // Columns
37   idct4_sse2(in);
38 
39   // Final round and shift
40   in[0] = _mm_add_epi16(in[0], eight);
41   in[1] = _mm_add_epi16(in[1], eight);
42   in[0] = _mm_srai_epi16(in[0], 4);
43   in[1] = _mm_srai_epi16(in[1], 4);
44 
45   recon_and_store4x4_sse2(in, dest, stride);
46 }
47 
vpx_idct4x4_1_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)48 void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest,
49                             int stride) {
50   const __m128i zero = _mm_setzero_si128();
51   int a;
52   __m128i dc_value, d[2];
53 
54   a = (int)dct_const_round_shift((int16_t)input[0] * cospi_16_64);
55   a = (int)dct_const_round_shift(a * cospi_16_64);
56   a = ROUND_POWER_OF_TWO(a, 4);
57 
58   dc_value = _mm_set1_epi16(a);
59 
60   // Reconstruction and Store
61   d[0] = _mm_cvtsi32_si128(*(const int *)(dest));
62   d[1] = _mm_cvtsi32_si128(*(const int *)(dest + stride * 3));
63   d[0] = _mm_unpacklo_epi32(d[0],
64                             _mm_cvtsi32_si128(*(const int *)(dest + stride)));
65   d[1] = _mm_unpacklo_epi32(
66       _mm_cvtsi32_si128(*(const int *)(dest + stride * 2)), d[1]);
67   d[0] = _mm_unpacklo_epi8(d[0], zero);
68   d[1] = _mm_unpacklo_epi8(d[1], zero);
69   d[0] = _mm_add_epi16(d[0], dc_value);
70   d[1] = _mm_add_epi16(d[1], dc_value);
71   d[0] = _mm_packus_epi16(d[0], d[1]);
72 
73   *(int *)dest = _mm_cvtsi128_si32(d[0]);
74   d[0] = _mm_srli_si128(d[0], 4);
75   *(int *)(dest + stride) = _mm_cvtsi128_si32(d[0]);
76   d[0] = _mm_srli_si128(d[0], 4);
77   *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d[0]);
78   d[0] = _mm_srli_si128(d[0], 4);
79   *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d[0]);
80 }
81 
idct4_sse2(__m128i * const in)82 void idct4_sse2(__m128i *const in) {
83   const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
84   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
85   const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
86   const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
87   __m128i u[2];
88 
89   transpose_16bit_4(in);
90   // stage 1
91   u[0] = _mm_unpacklo_epi16(in[0], in[1]);
92   u[1] = _mm_unpackhi_epi16(in[0], in[1]);
93   u[0] = idct_calc_wraplow_sse2(k__cospi_p16_p16, k__cospi_p16_m16, u[0]);
94   u[1] = idct_calc_wraplow_sse2(k__cospi_p08_p24, k__cospi_p24_m08, u[1]);
95 
96   // stage 2
97   in[0] = _mm_add_epi16(u[0], u[1]);
98   in[1] = _mm_sub_epi16(u[0], u[1]);
99   in[1] = _mm_shuffle_epi32(in[1], 0x4E);
100 }
101 
iadst4_sse2(__m128i * const in)102 void iadst4_sse2(__m128i *const in) {
103   const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
104   const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
105   const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
106   const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
107   const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
108   const __m128i kZero = _mm_set1_epi16(0);
109   const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
110   __m128i u[8], v[8], in7;
111 
112   transpose_16bit_4(in);
113   in7 = _mm_srli_si128(in[1], 8);
114   in7 = _mm_add_epi16(in7, in[0]);
115   in7 = _mm_sub_epi16(in7, in[1]);
116 
117   u[0] = _mm_unpacklo_epi16(in[0], in[1]);
118   u[1] = _mm_unpackhi_epi16(in[0], in[1]);
119   u[2] = _mm_unpacklo_epi16(in7, kZero);
120   u[3] = _mm_unpackhi_epi16(in[0], kZero);
121 
122   v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04);  // s0 + s3
123   v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02);  // s2 + s5
124   v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x2
125   v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01);  // s1 - s4
126   v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04);  // s2 - s6
127   v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s2
128 
129   u[0] = _mm_add_epi32(v[0], v[1]);
130   u[1] = _mm_add_epi32(v[3], v[4]);
131   u[2] = v[2];
132   u[3] = _mm_add_epi32(u[0], u[1]);
133   u[4] = _mm_slli_epi32(v[5], 2);
134   u[5] = _mm_add_epi32(u[3], v[5]);
135   u[6] = _mm_sub_epi32(u[5], u[4]);
136 
137   v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
138   v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
139   v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
140   v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
141 
142   u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
143   u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
144   u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
145   u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
146 
147   in[0] = _mm_packs_epi32(u[0], u[1]);
148   in[1] = _mm_packs_epi32(u[2], u[3]);
149 }
150 
load_buffer_8x8(const tran_low_t * const input,__m128i * const in)151 static INLINE void load_buffer_8x8(const tran_low_t *const input,
152                                    __m128i *const in) {
153   in[0] = load_input_data8(input + 0 * 8);
154   in[1] = load_input_data8(input + 1 * 8);
155   in[2] = load_input_data8(input + 2 * 8);
156   in[3] = load_input_data8(input + 3 * 8);
157   in[4] = load_input_data8(input + 4 * 8);
158   in[5] = load_input_data8(input + 5 * 8);
159   in[6] = load_input_data8(input + 6 * 8);
160   in[7] = load_input_data8(input + 7 * 8);
161 }
162 
vpx_idct8x8_64_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)163 void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest,
164                              int stride) {
165   __m128i in[8];
166   int i;
167 
168   // Load input data.
169   load_buffer_8x8(input, in);
170 
171   // 2-D
172   for (i = 0; i < 2; i++) {
173     idct8_sse2(in);
174   }
175 
176   write_buffer_8x8(in, dest, stride);
177 }
178 
vpx_idct8x8_12_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)179 void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest,
180                              int stride) {
181   __m128i io[8];
182 
183   io[0] = load_input_data4(input + 0 * 8);
184   io[1] = load_input_data4(input + 1 * 8);
185   io[2] = load_input_data4(input + 2 * 8);
186   io[3] = load_input_data4(input + 3 * 8);
187 
188   idct8x8_12_add_kernel_sse2(io);
189   write_buffer_8x8(io, dest, stride);
190 }
191 
recon_and_store_8_dual(uint8_t * const dest,const __m128i in_x,const int stride)192 static INLINE void recon_and_store_8_dual(uint8_t *const dest,
193                                           const __m128i in_x,
194                                           const int stride) {
195   const __m128i zero = _mm_setzero_si128();
196   __m128i d0, d1;
197 
198   d0 = _mm_loadl_epi64((__m128i *)(dest + 0 * stride));
199   d1 = _mm_loadl_epi64((__m128i *)(dest + 1 * stride));
200   d0 = _mm_unpacklo_epi8(d0, zero);
201   d1 = _mm_unpacklo_epi8(d1, zero);
202   d0 = _mm_add_epi16(in_x, d0);
203   d1 = _mm_add_epi16(in_x, d1);
204   d0 = _mm_packus_epi16(d0, d1);
205   _mm_storel_epi64((__m128i *)(dest + 0 * stride), d0);
206   _mm_storeh_pi((__m64 *)(dest + 1 * stride), _mm_castsi128_ps(d0));
207 }
208 
vpx_idct8x8_1_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)209 void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest,
210                             int stride) {
211   __m128i dc_value;
212   tran_high_t a1;
213   tran_low_t out =
214       WRAPLOW(dct_const_round_shift((int16_t)input[0] * cospi_16_64));
215 
216   out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
217   a1 = ROUND_POWER_OF_TWO(out, 5);
218   dc_value = _mm_set1_epi16((int16_t)a1);
219 
220   recon_and_store_8_dual(dest, dc_value, stride);
221   dest += 2 * stride;
222   recon_and_store_8_dual(dest, dc_value, stride);
223   dest += 2 * stride;
224   recon_and_store_8_dual(dest, dc_value, stride);
225   dest += 2 * stride;
226   recon_and_store_8_dual(dest, dc_value, stride);
227 }
228 
idct8_sse2(__m128i * const in)229 void idct8_sse2(__m128i *const in) {
230   // 8x8 Transpose is copied from vpx_fdct8x8_sse2()
231   transpose_16bit_8x8(in, in);
232 
233   // 4-stage 1D idct8x8
234   idct8(in, in);
235 }
236 
iadst8_sse2(__m128i * const in)237 void iadst8_sse2(__m128i *const in) {
238   const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
239   const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
240   const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
241   const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
242   const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
243   const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
244   const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
245   const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
246   const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
247   const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
248   const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
249   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
250   const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
251   const __m128i k__const_0 = _mm_set1_epi16(0);
252   const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
253 
254   __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
255   __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
256   __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
257   __m128i s0, s1, s2, s3, s4, s5, s6, s7;
258   __m128i in0, in1, in2, in3, in4, in5, in6, in7;
259 
260   // transpose
261   transpose_16bit_8x8(in, in);
262 
263   // properly aligned for butterfly input
264   in0 = in[7];
265   in1 = in[0];
266   in2 = in[5];
267   in3 = in[2];
268   in4 = in[3];
269   in5 = in[4];
270   in6 = in[1];
271   in7 = in[6];
272 
273   // column transformation
274   // stage 1
275   // interleave and multiply/add into 32-bit integer
276   s0 = _mm_unpacklo_epi16(in0, in1);
277   s1 = _mm_unpackhi_epi16(in0, in1);
278   s2 = _mm_unpacklo_epi16(in2, in3);
279   s3 = _mm_unpackhi_epi16(in2, in3);
280   s4 = _mm_unpacklo_epi16(in4, in5);
281   s5 = _mm_unpackhi_epi16(in4, in5);
282   s6 = _mm_unpacklo_epi16(in6, in7);
283   s7 = _mm_unpackhi_epi16(in6, in7);
284 
285   u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
286   u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
287   u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
288   u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
289   u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
290   u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
291   u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
292   u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
293   u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
294   u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
295   u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
296   u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
297   u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
298   u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
299   u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
300   u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
301 
302   // addition
303   w0 = _mm_add_epi32(u0, u8);
304   w1 = _mm_add_epi32(u1, u9);
305   w2 = _mm_add_epi32(u2, u10);
306   w3 = _mm_add_epi32(u3, u11);
307   w4 = _mm_add_epi32(u4, u12);
308   w5 = _mm_add_epi32(u5, u13);
309   w6 = _mm_add_epi32(u6, u14);
310   w7 = _mm_add_epi32(u7, u15);
311   w8 = _mm_sub_epi32(u0, u8);
312   w9 = _mm_sub_epi32(u1, u9);
313   w10 = _mm_sub_epi32(u2, u10);
314   w11 = _mm_sub_epi32(u3, u11);
315   w12 = _mm_sub_epi32(u4, u12);
316   w13 = _mm_sub_epi32(u5, u13);
317   w14 = _mm_sub_epi32(u6, u14);
318   w15 = _mm_sub_epi32(u7, u15);
319 
320   // shift and rounding
321   v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
322   v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
323   v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
324   v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
325   v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
326   v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
327   v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
328   v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
329   v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
330   v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
331   v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
332   v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
333   v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
334   v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
335   v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
336   v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
337 
338   u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
339   u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
340   u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
341   u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
342   u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
343   u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
344   u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
345   u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
346   u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
347   u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
348   u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
349   u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
350   u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
351   u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
352   u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
353   u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
354 
355   // back to 16-bit and pack 8 integers into __m128i
356   in[0] = _mm_packs_epi32(u0, u1);
357   in[1] = _mm_packs_epi32(u2, u3);
358   in[2] = _mm_packs_epi32(u4, u5);
359   in[3] = _mm_packs_epi32(u6, u7);
360   in[4] = _mm_packs_epi32(u8, u9);
361   in[5] = _mm_packs_epi32(u10, u11);
362   in[6] = _mm_packs_epi32(u12, u13);
363   in[7] = _mm_packs_epi32(u14, u15);
364 
365   // stage 2
366   s0 = _mm_add_epi16(in[0], in[2]);
367   s1 = _mm_add_epi16(in[1], in[3]);
368   s2 = _mm_sub_epi16(in[0], in[2]);
369   s3 = _mm_sub_epi16(in[1], in[3]);
370   u0 = _mm_unpacklo_epi16(in[4], in[5]);
371   u1 = _mm_unpackhi_epi16(in[4], in[5]);
372   u2 = _mm_unpacklo_epi16(in[6], in[7]);
373   u3 = _mm_unpackhi_epi16(in[6], in[7]);
374 
375   v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
376   v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
377   v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
378   v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
379   v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
380   v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
381   v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
382   v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
383 
384   w0 = _mm_add_epi32(v0, v4);
385   w1 = _mm_add_epi32(v1, v5);
386   w2 = _mm_add_epi32(v2, v6);
387   w3 = _mm_add_epi32(v3, v7);
388   w4 = _mm_sub_epi32(v0, v4);
389   w5 = _mm_sub_epi32(v1, v5);
390   w6 = _mm_sub_epi32(v2, v6);
391   w7 = _mm_sub_epi32(v3, v7);
392 
393   v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
394   v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
395   v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
396   v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
397   v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
398   v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
399   v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
400   v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
401 
402   u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
403   u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
404   u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
405   u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
406   u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
407   u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
408   u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
409   u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
410 
411   // back to 16-bit intergers
412   s4 = _mm_packs_epi32(u0, u1);
413   s5 = _mm_packs_epi32(u2, u3);
414   s6 = _mm_packs_epi32(u4, u5);
415   s7 = _mm_packs_epi32(u6, u7);
416 
417   // stage 3
418   u0 = _mm_unpacklo_epi16(s2, s3);
419   u1 = _mm_unpackhi_epi16(s2, s3);
420   u2 = _mm_unpacklo_epi16(s6, s7);
421   u3 = _mm_unpackhi_epi16(s6, s7);
422 
423   s2 = idct_calc_wraplow_sse2(u0, u1, k__cospi_p16_p16);
424   s3 = idct_calc_wraplow_sse2(u0, u1, k__cospi_p16_m16);
425   s6 = idct_calc_wraplow_sse2(u2, u3, k__cospi_p16_p16);
426   s7 = idct_calc_wraplow_sse2(u2, u3, k__cospi_p16_m16);
427 
428   in[0] = s0;
429   in[1] = _mm_sub_epi16(k__const_0, s4);
430   in[2] = s6;
431   in[3] = _mm_sub_epi16(k__const_0, s2);
432   in[4] = s3;
433   in[5] = _mm_sub_epi16(k__const_0, s7);
434   in[6] = s5;
435   in[7] = _mm_sub_epi16(k__const_0, s1);
436 }
437 
idct16_load8x8(const tran_low_t * const input,__m128i * const in)438 static INLINE void idct16_load8x8(const tran_low_t *const input,
439                                   __m128i *const in) {
440   in[0] = load_input_data8(input + 0 * 16);
441   in[1] = load_input_data8(input + 1 * 16);
442   in[2] = load_input_data8(input + 2 * 16);
443   in[3] = load_input_data8(input + 3 * 16);
444   in[4] = load_input_data8(input + 4 * 16);
445   in[5] = load_input_data8(input + 5 * 16);
446   in[6] = load_input_data8(input + 6 * 16);
447   in[7] = load_input_data8(input + 7 * 16);
448 }
449 
vpx_idct16x16_256_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)450 void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
451                                 int stride) {
452   __m128i l[16], r[16], out[16], *in;
453   int i;
454 
455   in = l;
456   for (i = 0; i < 2; i++) {
457     idct16_load8x8(input, in);
458     transpose_16bit_8x8(in, in);
459     idct16_load8x8(input + 8, in + 8);
460     transpose_16bit_8x8(in + 8, in + 8);
461     idct16_8col(in, in);
462     in = r;
463     input += 128;
464   }
465 
466   for (i = 0; i < 16; i += 8) {
467     int j;
468     transpose_16bit_8x8(l + i, out);
469     transpose_16bit_8x8(r + i, out + 8);
470     idct16_8col(out, out);
471 
472     for (j = 0; j < 16; ++j) {
473       write_buffer_8x1(dest + j * stride, out[j]);
474     }
475 
476     dest += 8;
477   }
478 }
479 
vpx_idct16x16_38_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)480 void vpx_idct16x16_38_add_sse2(const tran_low_t *input, uint8_t *dest,
481                                int stride) {
482   __m128i in[16], temp[16], out[16];
483   int i;
484 
485   idct16_load8x8(input, in);
486   transpose_16bit_8x8(in, in);
487 
488   for (i = 8; i < 16; i++) {
489     in[i] = _mm_setzero_si128();
490   }
491   idct16_8col(in, temp);
492 
493   for (i = 0; i < 16; i += 8) {
494     int j;
495     transpose_16bit_8x8(temp + i, in);
496     idct16_8col(in, out);
497 
498     for (j = 0; j < 16; ++j) {
499       write_buffer_8x1(dest + j * stride, out[j]);
500     }
501 
502     dest += 8;
503   }
504 }
505 
vpx_idct16x16_10_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)506 void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest,
507                                int stride) {
508   __m128i in[16], l[16];
509   int i;
510 
511   // First 1-D inverse DCT
512   // Load input data.
513   in[0] = load_input_data4(input + 0 * 16);
514   in[1] = load_input_data4(input + 1 * 16);
515   in[2] = load_input_data4(input + 2 * 16);
516   in[3] = load_input_data4(input + 3 * 16);
517 
518   idct16x16_10_pass1(in, l);
519 
520   // Second 1-D inverse transform, performed per 8x16 block
521   for (i = 0; i < 16; i += 8) {
522     int j;
523     idct16x16_10_pass2(l + i, in);
524 
525     for (j = 0; j < 16; ++j) {
526       write_buffer_8x1(dest + j * stride, in[j]);
527     }
528 
529     dest += 8;
530   }
531 }
532 
recon_and_store_16(uint8_t * const dest,const __m128i in_x)533 static INLINE void recon_and_store_16(uint8_t *const dest, const __m128i in_x) {
534   const __m128i zero = _mm_setzero_si128();
535   __m128i d0, d1;
536 
537   d0 = _mm_load_si128((__m128i *)(dest));
538   d1 = _mm_unpackhi_epi8(d0, zero);
539   d0 = _mm_unpacklo_epi8(d0, zero);
540   d0 = _mm_add_epi16(in_x, d0);
541   d1 = _mm_add_epi16(in_x, d1);
542   d0 = _mm_packus_epi16(d0, d1);
543   _mm_store_si128((__m128i *)(dest), d0);
544 }
545 
vpx_idct16x16_1_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)546 void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest,
547                               int stride) {
548   __m128i dc_value;
549   int i;
550   tran_high_t a1;
551   tran_low_t out =
552       WRAPLOW(dct_const_round_shift((int16_t)input[0] * cospi_16_64));
553 
554   out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
555   a1 = ROUND_POWER_OF_TWO(out, 6);
556   dc_value = _mm_set1_epi16((int16_t)a1);
557 
558   for (i = 0; i < 16; ++i) {
559     recon_and_store_16(dest, dc_value);
560     dest += stride;
561   }
562 }
563 
iadst16_8col(__m128i * const in)564 static void iadst16_8col(__m128i *const in) {
565   // perform 16x16 1-D ADST for 8 columns
566   __m128i s[16], x[16], u[32], v[32];
567   const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
568   const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
569   const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
570   const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
571   const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
572   const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
573   const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
574   const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
575   const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
576   const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
577   const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
578   const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
579   const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
580   const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
581   const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
582   const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
583   const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
584   const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
585   const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
586   const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
587   const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
588   const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
589   const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
590   const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
591   const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
592   const __m128i k__cospi_m16_m16 = _mm_set1_epi16(-cospi_16_64);
593   const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
594   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
595   const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
596   const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
597   const __m128i kZero = _mm_set1_epi16(0);
598 
599   u[0] = _mm_unpacklo_epi16(in[15], in[0]);
600   u[1] = _mm_unpackhi_epi16(in[15], in[0]);
601   u[2] = _mm_unpacklo_epi16(in[13], in[2]);
602   u[3] = _mm_unpackhi_epi16(in[13], in[2]);
603   u[4] = _mm_unpacklo_epi16(in[11], in[4]);
604   u[5] = _mm_unpackhi_epi16(in[11], in[4]);
605   u[6] = _mm_unpacklo_epi16(in[9], in[6]);
606   u[7] = _mm_unpackhi_epi16(in[9], in[6]);
607   u[8] = _mm_unpacklo_epi16(in[7], in[8]);
608   u[9] = _mm_unpackhi_epi16(in[7], in[8]);
609   u[10] = _mm_unpacklo_epi16(in[5], in[10]);
610   u[11] = _mm_unpackhi_epi16(in[5], in[10]);
611   u[12] = _mm_unpacklo_epi16(in[3], in[12]);
612   u[13] = _mm_unpackhi_epi16(in[3], in[12]);
613   u[14] = _mm_unpacklo_epi16(in[1], in[14]);
614   u[15] = _mm_unpackhi_epi16(in[1], in[14]);
615 
616   v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
617   v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
618   v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
619   v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
620   v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
621   v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
622   v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
623   v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
624   v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
625   v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
626   v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
627   v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
628   v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
629   v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
630   v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
631   v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
632   v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
633   v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
634   v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
635   v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
636   v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
637   v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
638   v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
639   v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
640   v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
641   v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
642   v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
643   v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
644   v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
645   v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
646   v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
647   v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
648 
649   u[0] = _mm_add_epi32(v[0], v[16]);
650   u[1] = _mm_add_epi32(v[1], v[17]);
651   u[2] = _mm_add_epi32(v[2], v[18]);
652   u[3] = _mm_add_epi32(v[3], v[19]);
653   u[4] = _mm_add_epi32(v[4], v[20]);
654   u[5] = _mm_add_epi32(v[5], v[21]);
655   u[6] = _mm_add_epi32(v[6], v[22]);
656   u[7] = _mm_add_epi32(v[7], v[23]);
657   u[8] = _mm_add_epi32(v[8], v[24]);
658   u[9] = _mm_add_epi32(v[9], v[25]);
659   u[10] = _mm_add_epi32(v[10], v[26]);
660   u[11] = _mm_add_epi32(v[11], v[27]);
661   u[12] = _mm_add_epi32(v[12], v[28]);
662   u[13] = _mm_add_epi32(v[13], v[29]);
663   u[14] = _mm_add_epi32(v[14], v[30]);
664   u[15] = _mm_add_epi32(v[15], v[31]);
665   u[16] = _mm_sub_epi32(v[0], v[16]);
666   u[17] = _mm_sub_epi32(v[1], v[17]);
667   u[18] = _mm_sub_epi32(v[2], v[18]);
668   u[19] = _mm_sub_epi32(v[3], v[19]);
669   u[20] = _mm_sub_epi32(v[4], v[20]);
670   u[21] = _mm_sub_epi32(v[5], v[21]);
671   u[22] = _mm_sub_epi32(v[6], v[22]);
672   u[23] = _mm_sub_epi32(v[7], v[23]);
673   u[24] = _mm_sub_epi32(v[8], v[24]);
674   u[25] = _mm_sub_epi32(v[9], v[25]);
675   u[26] = _mm_sub_epi32(v[10], v[26]);
676   u[27] = _mm_sub_epi32(v[11], v[27]);
677   u[28] = _mm_sub_epi32(v[12], v[28]);
678   u[29] = _mm_sub_epi32(v[13], v[29]);
679   u[30] = _mm_sub_epi32(v[14], v[30]);
680   u[31] = _mm_sub_epi32(v[15], v[31]);
681 
682   v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
683   v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
684   v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
685   v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
686   v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
687   v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
688   v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
689   v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
690   v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
691   v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
692   v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
693   v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
694   v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
695   v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
696   v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
697   v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
698   v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
699   v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
700   v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
701   v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
702   v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
703   v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
704   v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
705   v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
706   v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
707   v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
708   v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
709   v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
710   v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
711   v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
712   v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
713   v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
714 
715   u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
716   u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
717   u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
718   u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
719   u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
720   u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
721   u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
722   u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
723   u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
724   u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
725   u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
726   u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
727   u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
728   u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
729   u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
730   u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
731   u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
732   u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
733   u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
734   u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
735   u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
736   u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
737   u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
738   u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
739   u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
740   u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
741   u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
742   u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
743   u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
744   u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
745   u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
746   u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
747 
748   s[0] = _mm_packs_epi32(u[0], u[1]);
749   s[1] = _mm_packs_epi32(u[2], u[3]);
750   s[2] = _mm_packs_epi32(u[4], u[5]);
751   s[3] = _mm_packs_epi32(u[6], u[7]);
752   s[4] = _mm_packs_epi32(u[8], u[9]);
753   s[5] = _mm_packs_epi32(u[10], u[11]);
754   s[6] = _mm_packs_epi32(u[12], u[13]);
755   s[7] = _mm_packs_epi32(u[14], u[15]);
756   s[8] = _mm_packs_epi32(u[16], u[17]);
757   s[9] = _mm_packs_epi32(u[18], u[19]);
758   s[10] = _mm_packs_epi32(u[20], u[21]);
759   s[11] = _mm_packs_epi32(u[22], u[23]);
760   s[12] = _mm_packs_epi32(u[24], u[25]);
761   s[13] = _mm_packs_epi32(u[26], u[27]);
762   s[14] = _mm_packs_epi32(u[28], u[29]);
763   s[15] = _mm_packs_epi32(u[30], u[31]);
764 
765   // stage 2
766   u[0] = _mm_unpacklo_epi16(s[8], s[9]);
767   u[1] = _mm_unpackhi_epi16(s[8], s[9]);
768   u[2] = _mm_unpacklo_epi16(s[10], s[11]);
769   u[3] = _mm_unpackhi_epi16(s[10], s[11]);
770   u[4] = _mm_unpacklo_epi16(s[12], s[13]);
771   u[5] = _mm_unpackhi_epi16(s[12], s[13]);
772   u[6] = _mm_unpacklo_epi16(s[14], s[15]);
773   u[7] = _mm_unpackhi_epi16(s[14], s[15]);
774 
775   v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
776   v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
777   v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
778   v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
779   v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
780   v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
781   v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
782   v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
783   v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
784   v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
785   v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
786   v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
787   v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
788   v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
789   v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
790   v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
791 
792   u[0] = _mm_add_epi32(v[0], v[8]);
793   u[1] = _mm_add_epi32(v[1], v[9]);
794   u[2] = _mm_add_epi32(v[2], v[10]);
795   u[3] = _mm_add_epi32(v[3], v[11]);
796   u[4] = _mm_add_epi32(v[4], v[12]);
797   u[5] = _mm_add_epi32(v[5], v[13]);
798   u[6] = _mm_add_epi32(v[6], v[14]);
799   u[7] = _mm_add_epi32(v[7], v[15]);
800   u[8] = _mm_sub_epi32(v[0], v[8]);
801   u[9] = _mm_sub_epi32(v[1], v[9]);
802   u[10] = _mm_sub_epi32(v[2], v[10]);
803   u[11] = _mm_sub_epi32(v[3], v[11]);
804   u[12] = _mm_sub_epi32(v[4], v[12]);
805   u[13] = _mm_sub_epi32(v[5], v[13]);
806   u[14] = _mm_sub_epi32(v[6], v[14]);
807   u[15] = _mm_sub_epi32(v[7], v[15]);
808 
809   v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
810   v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
811   v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
812   v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
813   v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
814   v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
815   v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
816   v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
817   v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
818   v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
819   v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
820   v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
821   v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
822   v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
823   v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
824   v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
825 
826   u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
827   u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
828   u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
829   u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
830   u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
831   u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
832   u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
833   u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
834   u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
835   u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
836   u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
837   u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
838   u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
839   u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
840   u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
841   u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
842 
843   x[0] = _mm_add_epi16(s[0], s[4]);
844   x[1] = _mm_add_epi16(s[1], s[5]);
845   x[2] = _mm_add_epi16(s[2], s[6]);
846   x[3] = _mm_add_epi16(s[3], s[7]);
847   x[4] = _mm_sub_epi16(s[0], s[4]);
848   x[5] = _mm_sub_epi16(s[1], s[5]);
849   x[6] = _mm_sub_epi16(s[2], s[6]);
850   x[7] = _mm_sub_epi16(s[3], s[7]);
851   x[8] = _mm_packs_epi32(u[0], u[1]);
852   x[9] = _mm_packs_epi32(u[2], u[3]);
853   x[10] = _mm_packs_epi32(u[4], u[5]);
854   x[11] = _mm_packs_epi32(u[6], u[7]);
855   x[12] = _mm_packs_epi32(u[8], u[9]);
856   x[13] = _mm_packs_epi32(u[10], u[11]);
857   x[14] = _mm_packs_epi32(u[12], u[13]);
858   x[15] = _mm_packs_epi32(u[14], u[15]);
859 
860   // stage 3
861   u[0] = _mm_unpacklo_epi16(x[4], x[5]);
862   u[1] = _mm_unpackhi_epi16(x[4], x[5]);
863   u[2] = _mm_unpacklo_epi16(x[6], x[7]);
864   u[3] = _mm_unpackhi_epi16(x[6], x[7]);
865   u[4] = _mm_unpacklo_epi16(x[12], x[13]);
866   u[5] = _mm_unpackhi_epi16(x[12], x[13]);
867   u[6] = _mm_unpacklo_epi16(x[14], x[15]);
868   u[7] = _mm_unpackhi_epi16(x[14], x[15]);
869 
870   v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
871   v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
872   v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
873   v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
874   v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
875   v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
876   v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
877   v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
878   v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
879   v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
880   v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
881   v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
882   v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
883   v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
884   v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
885   v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
886 
887   u[0] = _mm_add_epi32(v[0], v[4]);
888   u[1] = _mm_add_epi32(v[1], v[5]);
889   u[2] = _mm_add_epi32(v[2], v[6]);
890   u[3] = _mm_add_epi32(v[3], v[7]);
891   u[4] = _mm_sub_epi32(v[0], v[4]);
892   u[5] = _mm_sub_epi32(v[1], v[5]);
893   u[6] = _mm_sub_epi32(v[2], v[6]);
894   u[7] = _mm_sub_epi32(v[3], v[7]);
895   u[8] = _mm_add_epi32(v[8], v[12]);
896   u[9] = _mm_add_epi32(v[9], v[13]);
897   u[10] = _mm_add_epi32(v[10], v[14]);
898   u[11] = _mm_add_epi32(v[11], v[15]);
899   u[12] = _mm_sub_epi32(v[8], v[12]);
900   u[13] = _mm_sub_epi32(v[9], v[13]);
901   u[14] = _mm_sub_epi32(v[10], v[14]);
902   u[15] = _mm_sub_epi32(v[11], v[15]);
903 
904   u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
905   u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
906   u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
907   u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
908   u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
909   u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
910   u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
911   u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
912   u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
913   u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
914   u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
915   u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
916   u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
917   u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
918   u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
919   u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
920 
921   v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
922   v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
923   v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
924   v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
925   v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
926   v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
927   v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
928   v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
929   v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
930   v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
931   v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
932   v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
933   v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
934   v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
935   v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
936   v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
937 
938   s[0] = _mm_add_epi16(x[0], x[2]);
939   s[1] = _mm_add_epi16(x[1], x[3]);
940   s[2] = _mm_sub_epi16(x[0], x[2]);
941   s[3] = _mm_sub_epi16(x[1], x[3]);
942   s[4] = _mm_packs_epi32(v[0], v[1]);
943   s[5] = _mm_packs_epi32(v[2], v[3]);
944   s[6] = _mm_packs_epi32(v[4], v[5]);
945   s[7] = _mm_packs_epi32(v[6], v[7]);
946   s[8] = _mm_add_epi16(x[8], x[10]);
947   s[9] = _mm_add_epi16(x[9], x[11]);
948   s[10] = _mm_sub_epi16(x[8], x[10]);
949   s[11] = _mm_sub_epi16(x[9], x[11]);
950   s[12] = _mm_packs_epi32(v[8], v[9]);
951   s[13] = _mm_packs_epi32(v[10], v[11]);
952   s[14] = _mm_packs_epi32(v[12], v[13]);
953   s[15] = _mm_packs_epi32(v[14], v[15]);
954 
955   // stage 4
956   u[0] = _mm_unpacklo_epi16(s[2], s[3]);
957   u[1] = _mm_unpackhi_epi16(s[2], s[3]);
958   u[2] = _mm_unpacklo_epi16(s[6], s[7]);
959   u[3] = _mm_unpackhi_epi16(s[6], s[7]);
960   u[4] = _mm_unpacklo_epi16(s[10], s[11]);
961   u[5] = _mm_unpackhi_epi16(s[10], s[11]);
962   u[6] = _mm_unpacklo_epi16(s[14], s[15]);
963   u[7] = _mm_unpackhi_epi16(s[14], s[15]);
964 
965   in[7] = idct_calc_wraplow_sse2(u[0], u[1], k__cospi_m16_m16);
966   in[8] = idct_calc_wraplow_sse2(u[0], u[1], k__cospi_p16_m16);
967   in[4] = idct_calc_wraplow_sse2(u[2], u[3], k__cospi_p16_p16);
968   in[11] = idct_calc_wraplow_sse2(u[2], u[3], k__cospi_m16_p16);
969   in[6] = idct_calc_wraplow_sse2(u[4], u[5], k__cospi_p16_p16);
970   in[9] = idct_calc_wraplow_sse2(u[4], u[5], k__cospi_m16_p16);
971   in[5] = idct_calc_wraplow_sse2(u[6], u[7], k__cospi_m16_m16);
972   in[10] = idct_calc_wraplow_sse2(u[6], u[7], k__cospi_p16_m16);
973 
974   in[0] = s[0];
975   in[1] = _mm_sub_epi16(kZero, s[8]);
976   in[2] = s[12];
977   in[3] = _mm_sub_epi16(kZero, s[4]);
978   in[12] = s[5];
979   in[13] = _mm_sub_epi16(kZero, s[13]);
980   in[14] = s[9];
981   in[15] = _mm_sub_epi16(kZero, s[1]);
982 }
983 
idct16_sse2(__m128i * const in0,__m128i * const in1)984 void idct16_sse2(__m128i *const in0, __m128i *const in1) {
985   transpose_16bit_16x16(in0, in1);
986   idct16_8col(in0, in0);
987   idct16_8col(in1, in1);
988 }
989 
iadst16_sse2(__m128i * const in0,__m128i * const in1)990 void iadst16_sse2(__m128i *const in0, __m128i *const in1) {
991   transpose_16bit_16x16(in0, in1);
992   iadst16_8col(in0);
993   iadst16_8col(in1);
994 }
995 
996 // Group the coefficient calculation into smaller functions to prevent stack
997 // spillover in 32x32 idct optimizations:
998 // quarter_1: 0-7
999 // quarter_2: 8-15
1000 // quarter_3_4: 16-23, 24-31
1001 
1002 // For each 8x32 block __m128i in[32],
1003 // Input with index, 0, 4
1004 // output pixels: 0-7 in __m128i out[32]
idct32_34_8x32_quarter_1(const __m128i * const in,__m128i * const out)1005 static INLINE void idct32_34_8x32_quarter_1(const __m128i *const in /*in[32]*/,
1006                                             __m128i *const out /*out[8]*/) {
1007   const __m128i zero = _mm_setzero_si128();
1008   __m128i step1[8], step2[8];
1009 
1010   // stage 3
1011   butterfly(in[4], zero, cospi_28_64, cospi_4_64, &step1[4], &step1[7]);
1012 
1013   // stage 4
1014   step2[0] = butterfly_cospi16(in[0]);
1015   step2[4] = step1[4];
1016   step2[5] = step1[4];
1017   step2[6] = step1[7];
1018   step2[7] = step1[7];
1019 
1020   // stage 5
1021   step1[0] = step2[0];
1022   step1[1] = step2[0];
1023   step1[2] = step2[0];
1024   step1[3] = step2[0];
1025   step1[4] = step2[4];
1026   butterfly(step2[6], step2[5], cospi_16_64, cospi_16_64, &step1[5], &step1[6]);
1027   step1[7] = step2[7];
1028 
1029   // stage 6
1030   out[0] = _mm_add_epi16(step1[0], step1[7]);
1031   out[1] = _mm_add_epi16(step1[1], step1[6]);
1032   out[2] = _mm_add_epi16(step1[2], step1[5]);
1033   out[3] = _mm_add_epi16(step1[3], step1[4]);
1034   out[4] = _mm_sub_epi16(step1[3], step1[4]);
1035   out[5] = _mm_sub_epi16(step1[2], step1[5]);
1036   out[6] = _mm_sub_epi16(step1[1], step1[6]);
1037   out[7] = _mm_sub_epi16(step1[0], step1[7]);
1038 }
1039 
1040 // For each 8x32 block __m128i in[32],
1041 // Input with index, 2, 6
1042 // output pixels: 8-15 in __m128i out[32]
idct32_34_8x32_quarter_2(const __m128i * const in,__m128i * const out)1043 static INLINE void idct32_34_8x32_quarter_2(const __m128i *const in /*in[32]*/,
1044                                             __m128i *const out /*out[16]*/) {
1045   const __m128i zero = _mm_setzero_si128();
1046   __m128i step1[16], step2[16];
1047 
1048   // stage 2
1049   butterfly(in[2], zero, cospi_30_64, cospi_2_64, &step2[8], &step2[15]);
1050   butterfly(zero, in[6], cospi_6_64, cospi_26_64, &step2[11], &step2[12]);
1051 
1052   // stage 3
1053   step1[8] = step2[8];
1054   step1[9] = step2[8];
1055   step1[14] = step2[15];
1056   step1[15] = step2[15];
1057   step1[10] = step2[11];
1058   step1[11] = step2[11];
1059   step1[12] = step2[12];
1060   step1[13] = step2[12];
1061 
1062   idct32_8x32_quarter_2_stage_4_to_6(step1, out);
1063 }
1064 
idct32_34_8x32_quarter_1_2(const __m128i * const in,__m128i * const out)1065 static INLINE void idct32_34_8x32_quarter_1_2(
1066     const __m128i *const in /*in[32]*/, __m128i *const out /*out[32]*/) {
1067   __m128i temp[16];
1068   idct32_34_8x32_quarter_1(in, temp);
1069   idct32_34_8x32_quarter_2(in, temp);
1070   // stage 7
1071   add_sub_butterfly(temp, out, 16);
1072 }
1073 
1074 // For each 8x32 block __m128i in[32],
1075 // Input with odd index, 1, 3, 5, 7
1076 // output pixels: 16-23, 24-31 in __m128i out[32]
idct32_34_8x32_quarter_3_4(const __m128i * const in,__m128i * const out)1077 static INLINE void idct32_34_8x32_quarter_3_4(
1078     const __m128i *const in /*in[32]*/, __m128i *const out /*out[32]*/) {
1079   const __m128i zero = _mm_setzero_si128();
1080   __m128i step1[32];
1081 
1082   // stage 1
1083   butterfly(in[1], zero, cospi_31_64, cospi_1_64, &step1[16], &step1[31]);
1084   butterfly(zero, in[7], cospi_7_64, cospi_25_64, &step1[19], &step1[28]);
1085   butterfly(in[5], zero, cospi_27_64, cospi_5_64, &step1[20], &step1[27]);
1086   butterfly(zero, in[3], cospi_3_64, cospi_29_64, &step1[23], &step1[24]);
1087 
1088   // stage 3
1089   butterfly(step1[31], step1[16], cospi_28_64, cospi_4_64, &step1[17],
1090             &step1[30]);
1091   butterfly(step1[28], step1[19], -cospi_4_64, cospi_28_64, &step1[18],
1092             &step1[29]);
1093   butterfly(step1[27], step1[20], cospi_12_64, cospi_20_64, &step1[21],
1094             &step1[26]);
1095   butterfly(step1[24], step1[23], -cospi_20_64, cospi_12_64, &step1[22],
1096             &step1[25]);
1097 
1098   idct32_8x32_quarter_3_4_stage_4_to_7(step1, out);
1099 }
1100 
idct32_34_8x32_sse2(const __m128i * const in,__m128i * const out)1101 void idct32_34_8x32_sse2(const __m128i *const in /*in[32]*/,
1102                          __m128i *const out /*out[32]*/) {
1103   __m128i temp[32];
1104 
1105   idct32_34_8x32_quarter_1_2(in, temp);
1106   idct32_34_8x32_quarter_3_4(in, temp);
1107   // final stage
1108   add_sub_butterfly(temp, out, 32);
1109 }
1110 
1111 // Only upper-left 8x8 has non-zero coeff
vpx_idct32x32_34_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)1112 void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest,
1113                                int stride) {
1114   __m128i io[32], col[32];
1115   int i;
1116 
1117   // Load input data. Only need to load the top left 8x8 block.
1118   load_transpose_16bit_8x8(input, 32, io);
1119   idct32_34_8x32_sse2(io, col);
1120 
1121   for (i = 0; i < 32; i += 8) {
1122     int j;
1123     transpose_16bit_8x8(col + i, io);
1124     idct32_34_8x32_sse2(io, io);
1125 
1126     for (j = 0; j < 32; ++j) {
1127       write_buffer_8x1(dest + j * stride, io[j]);
1128     }
1129 
1130     dest += 8;
1131   }
1132 }
1133 
1134 // For each 8x32 block __m128i in[32],
1135 // Input with index, 0, 4, 8, 12, 16, 20, 24, 28
1136 // output pixels: 0-7 in __m128i out[32]
idct32_1024_8x32_quarter_1(const __m128i * const in,__m128i * const out)1137 static INLINE void idct32_1024_8x32_quarter_1(
1138     const __m128i *const in /*in[32]*/, __m128i *const out /*out[8]*/) {
1139   __m128i step1[8], step2[8];
1140 
1141   // stage 3
1142   butterfly(in[4], in[28], cospi_28_64, cospi_4_64, &step1[4], &step1[7]);
1143   butterfly(in[20], in[12], cospi_12_64, cospi_20_64, &step1[5], &step1[6]);
1144 
1145   // stage 4
1146   butterfly(in[0], in[16], cospi_16_64, cospi_16_64, &step2[1], &step2[0]);
1147   butterfly(in[8], in[24], cospi_24_64, cospi_8_64, &step2[2], &step2[3]);
1148   step2[4] = _mm_add_epi16(step1[4], step1[5]);
1149   step2[5] = _mm_sub_epi16(step1[4], step1[5]);
1150   step2[6] = _mm_sub_epi16(step1[7], step1[6]);
1151   step2[7] = _mm_add_epi16(step1[7], step1[6]);
1152 
1153   // stage 5
1154   step1[0] = _mm_add_epi16(step2[0], step2[3]);
1155   step1[1] = _mm_add_epi16(step2[1], step2[2]);
1156   step1[2] = _mm_sub_epi16(step2[1], step2[2]);
1157   step1[3] = _mm_sub_epi16(step2[0], step2[3]);
1158   step1[4] = step2[4];
1159   butterfly(step2[6], step2[5], cospi_16_64, cospi_16_64, &step1[5], &step1[6]);
1160   step1[7] = step2[7];
1161 
1162   // stage 6
1163   out[0] = _mm_add_epi16(step1[0], step1[7]);
1164   out[1] = _mm_add_epi16(step1[1], step1[6]);
1165   out[2] = _mm_add_epi16(step1[2], step1[5]);
1166   out[3] = _mm_add_epi16(step1[3], step1[4]);
1167   out[4] = _mm_sub_epi16(step1[3], step1[4]);
1168   out[5] = _mm_sub_epi16(step1[2], step1[5]);
1169   out[6] = _mm_sub_epi16(step1[1], step1[6]);
1170   out[7] = _mm_sub_epi16(step1[0], step1[7]);
1171 }
1172 
1173 // For each 8x32 block __m128i in[32],
1174 // Input with index, 2, 6, 10, 14, 18, 22, 26, 30
1175 // output pixels: 8-15 in __m128i out[32]
idct32_1024_8x32_quarter_2(const __m128i * const in,__m128i * const out)1176 static INLINE void idct32_1024_8x32_quarter_2(
1177     const __m128i *const in /*in[32]*/, __m128i *const out /*out[16]*/) {
1178   __m128i step1[16], step2[16];
1179 
1180   // stage 2
1181   butterfly(in[2], in[30], cospi_30_64, cospi_2_64, &step2[8], &step2[15]);
1182   butterfly(in[18], in[14], cospi_14_64, cospi_18_64, &step2[9], &step2[14]);
1183   butterfly(in[10], in[22], cospi_22_64, cospi_10_64, &step2[10], &step2[13]);
1184   butterfly(in[26], in[6], cospi_6_64, cospi_26_64, &step2[11], &step2[12]);
1185 
1186   // stage 3
1187   step1[8] = _mm_add_epi16(step2[8], step2[9]);
1188   step1[9] = _mm_sub_epi16(step2[8], step2[9]);
1189   step1[10] = _mm_sub_epi16(step2[11], step2[10]);
1190   step1[11] = _mm_add_epi16(step2[11], step2[10]);
1191   step1[12] = _mm_add_epi16(step2[12], step2[13]);
1192   step1[13] = _mm_sub_epi16(step2[12], step2[13]);
1193   step1[14] = _mm_sub_epi16(step2[15], step2[14]);
1194   step1[15] = _mm_add_epi16(step2[15], step2[14]);
1195 
1196   idct32_8x32_quarter_2_stage_4_to_6(step1, out);
1197 }
1198 
idct32_1024_8x32_quarter_1_2(const __m128i * const in,__m128i * const out)1199 static INLINE void idct32_1024_8x32_quarter_1_2(
1200     const __m128i *const in /*in[32]*/, __m128i *const out /*out[32]*/) {
1201   __m128i temp[16];
1202   idct32_1024_8x32_quarter_1(in, temp);
1203   idct32_1024_8x32_quarter_2(in, temp);
1204   // stage 7
1205   add_sub_butterfly(temp, out, 16);
1206 }
1207 
1208 // For each 8x32 block __m128i in[32],
1209 // Input with odd index,
1210 // 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
1211 // output pixels: 16-23, 24-31 in __m128i out[32]
idct32_1024_8x32_quarter_3_4(const __m128i * const in,__m128i * const out)1212 static INLINE void idct32_1024_8x32_quarter_3_4(
1213     const __m128i *const in /*in[32]*/, __m128i *const out /*out[32]*/) {
1214   __m128i step1[32], step2[32];
1215 
1216   // stage 1
1217   butterfly(in[1], in[31], cospi_31_64, cospi_1_64, &step1[16], &step1[31]);
1218   butterfly(in[17], in[15], cospi_15_64, cospi_17_64, &step1[17], &step1[30]);
1219   butterfly(in[9], in[23], cospi_23_64, cospi_9_64, &step1[18], &step1[29]);
1220   butterfly(in[25], in[7], cospi_7_64, cospi_25_64, &step1[19], &step1[28]);
1221 
1222   butterfly(in[5], in[27], cospi_27_64, cospi_5_64, &step1[20], &step1[27]);
1223   butterfly(in[21], in[11], cospi_11_64, cospi_21_64, &step1[21], &step1[26]);
1224 
1225   butterfly(in[13], in[19], cospi_19_64, cospi_13_64, &step1[22], &step1[25]);
1226   butterfly(in[29], in[3], cospi_3_64, cospi_29_64, &step1[23], &step1[24]);
1227 
1228   // stage 2
1229   step2[16] = _mm_add_epi16(step1[16], step1[17]);
1230   step2[17] = _mm_sub_epi16(step1[16], step1[17]);
1231   step2[18] = _mm_sub_epi16(step1[19], step1[18]);
1232   step2[19] = _mm_add_epi16(step1[19], step1[18]);
1233   step2[20] = _mm_add_epi16(step1[20], step1[21]);
1234   step2[21] = _mm_sub_epi16(step1[20], step1[21]);
1235   step2[22] = _mm_sub_epi16(step1[23], step1[22]);
1236   step2[23] = _mm_add_epi16(step1[23], step1[22]);
1237 
1238   step2[24] = _mm_add_epi16(step1[24], step1[25]);
1239   step2[25] = _mm_sub_epi16(step1[24], step1[25]);
1240   step2[26] = _mm_sub_epi16(step1[27], step1[26]);
1241   step2[27] = _mm_add_epi16(step1[27], step1[26]);
1242   step2[28] = _mm_add_epi16(step1[28], step1[29]);
1243   step2[29] = _mm_sub_epi16(step1[28], step1[29]);
1244   step2[30] = _mm_sub_epi16(step1[31], step1[30]);
1245   step2[31] = _mm_add_epi16(step1[31], step1[30]);
1246 
1247   // stage 3
1248   step1[16] = step2[16];
1249   step1[31] = step2[31];
1250   butterfly(step2[30], step2[17], cospi_28_64, cospi_4_64, &step1[17],
1251             &step1[30]);
1252   butterfly(step2[29], step2[18], -cospi_4_64, cospi_28_64, &step1[18],
1253             &step1[29]);
1254   step1[19] = step2[19];
1255   step1[20] = step2[20];
1256   butterfly(step2[26], step2[21], cospi_12_64, cospi_20_64, &step1[21],
1257             &step1[26]);
1258   butterfly(step2[25], step2[22], -cospi_20_64, cospi_12_64, &step1[22],
1259             &step1[25]);
1260   step1[23] = step2[23];
1261   step1[24] = step2[24];
1262   step1[27] = step2[27];
1263   step1[28] = step2[28];
1264 
1265   idct32_8x32_quarter_3_4_stage_4_to_7(step1, out);
1266 }
1267 
idct32_1024_8x32(const __m128i * const in,__m128i * const out)1268 void idct32_1024_8x32(const __m128i *const in /*in[32]*/,
1269                       __m128i *const out /*out[32]*/) {
1270   __m128i temp[32];
1271 
1272   idct32_1024_8x32_quarter_1_2(in, temp);
1273   idct32_1024_8x32_quarter_3_4(in, temp);
1274   // final stage
1275   add_sub_butterfly(temp, out, 32);
1276 }
1277 
vpx_idct32x32_1024_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)1278 void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest,
1279                                  int stride) {
1280   __m128i col[4][32], io[32];
1281   int i;
1282 
1283   // rows
1284   for (i = 0; i < 4; i++) {
1285     load_transpose_16bit_8x8(&input[0], 32, &io[0]);
1286     load_transpose_16bit_8x8(&input[8], 32, &io[8]);
1287     load_transpose_16bit_8x8(&input[16], 32, &io[16]);
1288     load_transpose_16bit_8x8(&input[24], 32, &io[24]);
1289     idct32_1024_8x32(io, col[i]);
1290     input += 32 << 3;
1291   }
1292 
1293   // columns
1294   for (i = 0; i < 32; i += 8) {
1295     // Transpose 32x8 block to 8x32 block
1296     transpose_16bit_8x8(col[0] + i, io);
1297     transpose_16bit_8x8(col[1] + i, io + 8);
1298     transpose_16bit_8x8(col[2] + i, io + 16);
1299     transpose_16bit_8x8(col[3] + i, io + 24);
1300 
1301     idct32_1024_8x32(io, io);
1302     store_buffer_8x32(io, dest, stride);
1303     dest += 8;
1304   }
1305 }
1306 
vpx_idct32x32_135_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)1307 void vpx_idct32x32_135_add_sse2(const tran_low_t *input, uint8_t *dest,
1308                                 int stride) {
1309   __m128i col[2][32], in[32], out[32];
1310   int i;
1311 
1312   for (i = 16; i < 32; i++) {
1313     in[i] = _mm_setzero_si128();
1314   }
1315 
1316   // rows
1317   for (i = 0; i < 2; i++) {
1318     load_transpose_16bit_8x8(&input[0], 32, &in[0]);
1319     load_transpose_16bit_8x8(&input[8], 32, &in[8]);
1320     idct32_1024_8x32(in, col[i]);
1321     input += 32 << 3;
1322   }
1323 
1324   // columns
1325   for (i = 0; i < 32; i += 8) {
1326     transpose_16bit_8x8(col[0] + i, in);
1327     transpose_16bit_8x8(col[1] + i, in + 8);
1328     idct32_1024_8x32(in, out);
1329     store_buffer_8x32(out, dest, stride);
1330     dest += 8;
1331   }
1332 }
1333 
vpx_idct32x32_1_add_sse2(const tran_low_t * input,uint8_t * dest,int stride)1334 void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest,
1335                               int stride) {
1336   __m128i dc_value;
1337   int j;
1338   tran_high_t a1;
1339   tran_low_t out =
1340       WRAPLOW(dct_const_round_shift((int16_t)input[0] * cospi_16_64));
1341 
1342   out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
1343   a1 = ROUND_POWER_OF_TWO(out, 6);
1344   dc_value = _mm_set1_epi16((int16_t)a1);
1345 
1346   for (j = 0; j < 32; ++j) {
1347     recon_and_store_16(dest + j * stride + 0, dc_value);
1348     recon_and_store_16(dest + j * stride + 16, dc_value);
1349   }
1350 }
1351