1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include "config/av1_rtcd.h"
13
14 #include "av1/common/enums.h"
15 #include "av1/common/av1_txfm.h"
16 #include "av1/common/x86/av1_txfm_sse2.h"
17 #include "av1/common/x86/highbd_txfm_utility_sse4.h"
18 #include "av1/encoder/av1_fwd_txfm1d_cfg.h"
19 #include "av1/encoder/x86/av1_txfm1d_sse4.h"
20 #include "av1/encoder/x86/av1_fwd_txfm_sse2.h"
21
int16_array_with_stride_to_int32_array_without_stride(const int16_t * input,int stride,int32_t * output,int txfm1d_size)22 static INLINE void int16_array_with_stride_to_int32_array_without_stride(
23 const int16_t *input, int stride, int32_t *output, int txfm1d_size) {
24 int r, c;
25 for (r = 0; r < txfm1d_size; r++) {
26 for (c = 0; c < txfm1d_size; c++) {
27 output[r * txfm1d_size + c] = (int32_t)input[r * stride + c];
28 }
29 }
30 }
31
32 typedef void (*TxfmFuncSSE2)(__m128i *input, __m128i *output,
33 const int8_t cos_bit, const int8_t *stage_range);
34
fdct32_new_sse4_1(__m128i * input,__m128i * output,const int8_t cos_bit,const int8_t * stage_range)35 static void fdct32_new_sse4_1(__m128i *input, __m128i *output,
36 const int8_t cos_bit, const int8_t *stage_range) {
37 const int txfm_size = 32;
38 const int num_per_128 = 4;
39 int col_num = txfm_size / num_per_128;
40 int col;
41 (void)stage_range;
42 for (col = 0; col < col_num; col++) {
43 av1_fdct32_new_sse4_1((input + col), (output + col), cos_bit, col_num);
44 }
45 }
46
fdct64_new_sse4_1(__m128i * input,__m128i * output,const int8_t cos_bit,const int8_t * stage_range)47 static void fdct64_new_sse4_1(__m128i *input, __m128i *output,
48 const int8_t cos_bit, const int8_t *stage_range) {
49 const int txfm_size = 64;
50 const int num_per_128 = 4;
51 int col_num = txfm_size / num_per_128;
52 (void)stage_range;
53 for (int col = 0; col < col_num; col++) {
54 av1_fdct64_new_sse4_1((input + col), (output + col), cos_bit, col_num,
55 col_num);
56 }
57 }
idtx32x32_sse4_1(__m128i * input,__m128i * output,const int8_t cos_bit,const int8_t * stage_range)58 static void idtx32x32_sse4_1(__m128i *input, __m128i *output,
59 const int8_t cos_bit, const int8_t *stage_range) {
60 (void)stage_range;
61
62 for (int i = 0; i < 8; i++) {
63 av1_idtx32_new_sse4_1(&input[i * 32], &output[i * 32], cos_bit, 1);
64 }
65 }
66
fwd_txfm_type_to_func(TXFM_TYPE txfm_type)67 static INLINE TxfmFuncSSE2 fwd_txfm_type_to_func(TXFM_TYPE txfm_type) {
68 switch (txfm_type) {
69 case TXFM_TYPE_DCT32: return fdct32_new_sse4_1; break;
70 case TXFM_TYPE_DCT64: return fdct64_new_sse4_1; break;
71 case TXFM_TYPE_IDENTITY32: return idtx32x32_sse4_1; break;
72 default: assert(0);
73 }
74 return NULL;
75 }
76
fwd_txfm2d_sse4_1(const int16_t * input,int32_t * output,const int stride,const TXFM_2D_FLIP_CFG * cfg,int32_t * txfm_buf)77 static INLINE void fwd_txfm2d_sse4_1(const int16_t *input, int32_t *output,
78 const int stride,
79 const TXFM_2D_FLIP_CFG *cfg,
80 int32_t *txfm_buf) {
81 // TODO(sarahparker) This does not currently support rectangular transforms
82 // and will break without splitting txfm_size out into row and col size.
83 // Rectangular transforms use c code only, so it should be ok for now.
84 // It will be corrected when there are sse implementations for rectangular
85 // transforms.
86 assert(cfg->tx_size < TX_SIZES);
87 const int txfm_size = tx_size_wide[cfg->tx_size];
88 const int8_t *shift = cfg->shift;
89 const int8_t *stage_range_col = cfg->stage_range_col;
90 const int8_t *stage_range_row = cfg->stage_range_row;
91 const int8_t cos_bit_col = cfg->cos_bit_col;
92 const int8_t cos_bit_row = cfg->cos_bit_row;
93 const TxfmFuncSSE2 txfm_func_col = fwd_txfm_type_to_func(cfg->txfm_type_col);
94 const TxfmFuncSSE2 txfm_func_row = fwd_txfm_type_to_func(cfg->txfm_type_row);
95
96 __m128i *buf_128 = (__m128i *)txfm_buf;
97 __m128i *out_128 = (__m128i *)output;
98 int num_per_128 = 4;
99 int txfm2d_size_128 = txfm_size * txfm_size / num_per_128;
100
101 int16_array_with_stride_to_int32_array_without_stride(input, stride, txfm_buf,
102 txfm_size);
103 av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[0]);
104 txfm_func_col(out_128, buf_128, cos_bit_col, stage_range_col);
105 av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[1]);
106 transpose_32(txfm_size, out_128, buf_128);
107 txfm_func_row(buf_128, out_128, cos_bit_row, stage_range_row);
108 av1_round_shift_array_32_sse4_1(out_128, buf_128, txfm2d_size_128, -shift[2]);
109 transpose_32(txfm_size, buf_128, out_128);
110 }
111
fwd_txfm2d_64x64_sse4_1(const int16_t * input,int32_t * output,const int stride,const TXFM_2D_FLIP_CFG * cfg,int32_t * txfm_buf)112 static INLINE void fwd_txfm2d_64x64_sse4_1(const int16_t *input,
113 int32_t *output, const int stride,
114 const TXFM_2D_FLIP_CFG *cfg,
115 int32_t *txfm_buf) {
116 assert(cfg->tx_size < TX_SIZES);
117 const int txfm_size = tx_size_wide[cfg->tx_size];
118 const int8_t *shift = cfg->shift;
119 const int8_t *stage_range_col = cfg->stage_range_col;
120 const int8_t cos_bit_col = cfg->cos_bit_col;
121 const int8_t cos_bit_row = cfg->cos_bit_row;
122 const TxfmFuncSSE2 txfm_func_col = fwd_txfm_type_to_func(cfg->txfm_type_col);
123 __m128i *buf_128 = (__m128i *)txfm_buf;
124 __m128i *out_128 = (__m128i *)output;
125
126 const int num_per_128 = 4;
127 int txfm2d_size_128 = txfm_size * txfm_size / num_per_128;
128 int col_num = txfm_size / num_per_128;
129
130 int16_array_with_stride_to_int32_array_without_stride(input, stride, output,
131 txfm_size);
132 /*col wise transform*/
133 txfm_func_col(out_128, buf_128, cos_bit_col, stage_range_col);
134 av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[1]);
135 transpose_32(txfm_size, out_128, buf_128);
136
137 /*row wise transform*/
138 for (int col = 0; col < (col_num >> 1); col++) {
139 av1_fdct64_new_sse4_1((buf_128 + col), (out_128 + col), cos_bit_row,
140 col_num, (col_num >> 1));
141 }
142
143 txfm2d_size_128 = (col_num >> 1) * (txfm_size >> 1);
144 av1_round_shift_array_32_sse4_1(out_128, buf_128, txfm2d_size_128, -shift[2]);
145 transpose_8nx8n(buf_128, out_128, 32, 32);
146 }
147
av1_fwd_txfm2d_32x32_sse4_1(const int16_t * input,int32_t * output,int stride,TX_TYPE tx_type,int bd)148 void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output,
149 int stride, TX_TYPE tx_type, int bd) {
150 DECLARE_ALIGNED(16, int32_t, txfm_buf[1024]);
151 TXFM_2D_FLIP_CFG cfg;
152 av1_get_fwd_txfm_cfg(tx_type, TX_32X32, &cfg);
153 (void)bd;
154 fwd_txfm2d_sse4_1(input, output, stride, &cfg, txfm_buf);
155 }
156
av1_fwd_txfm2d_64x64_sse4_1(const int16_t * input,int32_t * output,int stride,TX_TYPE tx_type,int bd)157 void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output,
158 int stride, TX_TYPE tx_type, int bd) {
159 DECLARE_ALIGNED(16, int32_t, txfm_buf[4096]);
160 TXFM_2D_FLIP_CFG cfg;
161 av1_get_fwd_txfm_cfg(tx_type, TX_64X64, &cfg);
162 (void)bd;
163 fwd_txfm2d_64x64_sse4_1(input, output, stride, &cfg, txfm_buf);
164 }
165
transpose_32_4x4x2(int stride,const __m128i * inputA,const __m128i * inputB,__m128i * output)166 static INLINE void transpose_32_4x4x2(int stride, const __m128i *inputA,
167 const __m128i *inputB, __m128i *output) {
168 __m128i temp0 = _mm_unpacklo_epi32(inputA[0], inputA[2]);
169 __m128i temp1 = _mm_unpackhi_epi32(inputA[0], inputA[2]);
170 __m128i temp2 = _mm_unpacklo_epi32(inputA[1], inputA[3]);
171 __m128i temp3 = _mm_unpackhi_epi32(inputA[1], inputA[3]);
172
173 output[0 * stride] = _mm_unpacklo_epi32(temp0, temp2);
174 output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2);
175 output[2 * stride] = _mm_unpacklo_epi32(temp1, temp3);
176 output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3);
177
178 temp0 = _mm_unpacklo_epi32(inputB[0], inputB[2]);
179 temp1 = _mm_unpackhi_epi32(inputB[0], inputB[2]);
180 temp2 = _mm_unpacklo_epi32(inputB[1], inputB[3]);
181 temp3 = _mm_unpackhi_epi32(inputB[1], inputB[3]);
182
183 output[4 * stride] = _mm_unpacklo_epi32(temp0, temp2);
184 output[5 * stride] = _mm_unpackhi_epi32(temp0, temp2);
185 output[6 * stride] = _mm_unpacklo_epi32(temp1, temp3);
186 output[7 * stride] = _mm_unpackhi_epi32(temp1, temp3);
187 }
188
lowbd_fwd_txfm2d_64x64_sse4_1(const int16_t * input,int32_t * output,int stride,TX_TYPE tx_type,int bd)189 static void lowbd_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output,
190 int stride, TX_TYPE tx_type, int bd) {
191 (void)bd;
192 (void)tx_type;
193 assert(tx_type == DCT_DCT);
194 const TX_SIZE tx_size = TX_64X64;
195 __m128i buf0[64], buf1[512];
196 const int8_t *shift = fwd_txfm_shift_ls[tx_size];
197 const int txw_idx = get_txw_idx(tx_size);
198 const int txh_idx = get_txh_idx(tx_size);
199 const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx];
200 const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx];
201 const int width = tx_size_wide[tx_size];
202 const int height = tx_size_high[tx_size];
203 const transform_1d_sse2 col_txfm = fdct8x64_new_sse2;
204 const int width_div8 = (width >> 3);
205 const int height_div8 = (height >> 3);
206
207 for (int i = 0; i < width_div8; i++) {
208 load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height);
209 round_shift_16bit(buf0, height, shift[0]);
210 col_txfm(buf0, buf0, cos_bit_col);
211 round_shift_16bit(buf0, height, shift[1]);
212 for (int j = 0; j < AOMMIN(4, height_div8); ++j) {
213 transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i);
214 }
215 }
216 for (int i = 0; i < AOMMIN(4, height_div8); i++) {
217 __m128i bufA[64];
218 __m128i bufB[64];
219 __m128i *buf = buf1 + width * i;
220 for (int j = 0; j < width; ++j) {
221 bufA[j] = _mm_cvtepi16_epi32(buf[j]);
222 bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j]));
223 }
224 av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row, 1, 1);
225 av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row, 1, 1);
226 av1_round_shift_array_32_sse4_1(bufA, bufA, 32, -shift[2]);
227 av1_round_shift_array_32_sse4_1(bufB, bufB, 32, -shift[2]);
228
229 int32_t *output8 = output + 8 * 32 * i;
230 for (int j = 0; j < width_div8; ++j) {
231 __m128i *out = (__m128i *)(output8 + 4 * j);
232 transpose_32_4x4x2(8, bufA + 4 * j, bufB + 4 * j, out);
233 }
234 }
235 }
236
lowbd_fwd_txfm2d_64x32_sse4_1(const int16_t * input,int32_t * output,int stride,TX_TYPE tx_type,int bd)237 static void lowbd_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output,
238 int stride, TX_TYPE tx_type, int bd) {
239 (void)bd;
240 const TX_SIZE tx_size = TX_64X32;
241 __m128i buf0[64], buf1[256];
242 const int8_t *shift = fwd_txfm_shift_ls[tx_size];
243 const int txw_idx = get_txw_idx(tx_size);
244 const int txh_idx = get_txh_idx(tx_size);
245 const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx];
246 const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx];
247 const int width = tx_size_wide[tx_size];
248 const int height = tx_size_high[tx_size];
249 const transform_1d_sse2 col_txfm = col_txfm8x32_arr[tx_type];
250 const int width_div8 = (width >> 3);
251 const int height_div8 = (height >> 3);
252
253 for (int i = 0; i < width_div8; i++) {
254 load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height);
255 round_shift_16bit(buf0, height, shift[0]);
256 col_txfm(buf0, buf0, cos_bit_col);
257 round_shift_16bit(buf0, height, shift[1]);
258 for (int j = 0; j < AOMMIN(4, height_div8); ++j) {
259 transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i);
260 }
261 }
262 assert(tx_type == DCT_DCT);
263 for (int i = 0; i < AOMMIN(4, height_div8); i++) {
264 __m128i bufA[64];
265 __m128i bufB[64];
266 __m128i *buf = buf1 + width * i;
267 for (int j = 0; j < width; ++j) {
268 bufA[j] = _mm_cvtepi16_epi32(buf[j]);
269 bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j]));
270 }
271 av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row, 1, 1);
272 av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row, 1, 1);
273 av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2], NewSqrt2);
274 av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2], NewSqrt2);
275
276 int32_t *output8 = output + 8 * 32 * i;
277 for (int j = 0; j < width_div8; ++j) {
278 __m128i *out = (__m128i *)(output8 + 4 * j);
279 transpose_32_4x4x2(8, bufA + 4 * j, bufB + 4 * j, out);
280 }
281 }
282 }
283
lowbd_fwd_txfm2d_32x64_sse4_1(const int16_t * input,int32_t * output,int stride,TX_TYPE tx_type,int bd)284 static void lowbd_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output,
285 int stride, TX_TYPE tx_type, int bd) {
286 (void)bd;
287 (void)tx_type;
288 assert(tx_type == DCT_DCT);
289 const TX_SIZE tx_size = TX_32X64;
290 __m128i buf0[64], buf1[256];
291 const int8_t *shift = fwd_txfm_shift_ls[tx_size];
292 const int txw_idx = get_txw_idx(tx_size);
293 const int txh_idx = get_txh_idx(tx_size);
294 const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx];
295 const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx];
296 const int width = tx_size_wide[tx_size];
297 const int height = tx_size_high[tx_size];
298 const transform_1d_sse2 col_txfm = fdct8x64_new_sse2;
299 const int width_div8 = (width >> 3);
300 const int height_div8 = (height >> 3);
301
302 for (int i = 0; i < width_div8; i++) {
303 load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height);
304 round_shift_16bit(buf0, height, shift[0]);
305 col_txfm(buf0, buf0, cos_bit_col);
306 round_shift_16bit(buf0, height, shift[1]);
307 for (int j = 0; j < AOMMIN(4, height_div8); ++j) {
308 transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i);
309 }
310 }
311
312 for (int i = 0; i < AOMMIN(4, height_div8); i++) {
313 __m128i bufA[32];
314 __m128i bufB[32];
315 __m128i *buf = buf1 + width * i;
316 for (int j = 0; j < width; ++j) {
317 bufA[j] = _mm_cvtepi16_epi32(buf[j]);
318 bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j]));
319 }
320 av1_fdct32_new_sse4_1(bufA, bufA, cos_bit_row, 1);
321 av1_fdct32_new_sse4_1(bufB, bufB, cos_bit_row, 1);
322 av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2], NewSqrt2);
323 av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2], NewSqrt2);
324
325 int32_t *output8 = output + 8 * 32 * i;
326 for (int j = 0; j < (32 / 4); ++j) {
327 __m128i *out = (__m128i *)(output8 + 4 * j);
328 transpose_32_4x4x2(8, bufA + 4 * j, bufB + 4 * j, out);
329 }
330 }
331 }
332
333 static FwdTxfm2dFunc fwd_txfm2d_func_ls[TX_SIZES_ALL] = {
334 av1_lowbd_fwd_txfm2d_4x4_sse2, // 4x4 transform
335 av1_lowbd_fwd_txfm2d_8x8_sse2, // 8x8 transform
336 av1_lowbd_fwd_txfm2d_16x16_sse2, // 16x16 transform
337 av1_lowbd_fwd_txfm2d_32x32_sse2, // 32x32 transform
338 lowbd_fwd_txfm2d_64x64_sse4_1, // 64x64 transform
339 av1_lowbd_fwd_txfm2d_4x8_sse2, // 4x8 transform
340 av1_lowbd_fwd_txfm2d_8x4_sse2, // 8x4 transform
341 av1_lowbd_fwd_txfm2d_8x16_sse2, // 8x16 transform
342 av1_lowbd_fwd_txfm2d_16x8_sse2, // 16x8 transform
343 av1_lowbd_fwd_txfm2d_16x32_sse2, // 16x32 transform
344 av1_lowbd_fwd_txfm2d_32x16_sse2, // 32x16 transform
345 lowbd_fwd_txfm2d_32x64_sse4_1, // 32x64 transform
346 lowbd_fwd_txfm2d_64x32_sse4_1, // 64x32 transform
347 av1_lowbd_fwd_txfm2d_4x16_sse2, // 4x16 transform
348 av1_lowbd_fwd_txfm2d_16x4_sse2, // 16x4 transform
349 av1_lowbd_fwd_txfm2d_8x32_sse2, // 8x32 transform
350 av1_lowbd_fwd_txfm2d_32x8_sse2, // 32x8 transform
351 av1_lowbd_fwd_txfm2d_16x64_sse2, // 16x64 transform
352 av1_lowbd_fwd_txfm2d_64x16_sse2, // 64x16 transform
353 };
354
av1_lowbd_fwd_txfm_sse4_1(const int16_t * src_diff,tran_low_t * coeff,int diff_stride,TxfmParam * txfm_param)355 void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff,
356 int diff_stride, TxfmParam *txfm_param) {
357 FwdTxfm2dFunc fwd_txfm2d_func = fwd_txfm2d_func_ls[txfm_param->tx_size];
358 if ((fwd_txfm2d_func == NULL) ||
359 (txfm_param->lossless && txfm_param->tx_size == TX_4X4)) {
360 av1_lowbd_fwd_txfm_c(src_diff, coeff, diff_stride, txfm_param);
361 } else {
362 fwd_txfm2d_func(src_diff, coeff, diff_stride, txfm_param->tx_type,
363 txfm_param->bd);
364 }
365 }
366