1 /*
2 * Copyright (c) 2013 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 <math.h>
12 #include <stdlib.h>
13 #include <string.h>
14 #include <limits>
15 #include <tuple>
16
17 #include "gtest/gtest.h"
18
19 #include "./vp9_rtcd.h"
20 #include "./vpx_dsp_rtcd.h"
21 #include "test/acm_random.h"
22 #include "test/clear_system_state.h"
23 #include "test/register_state_check.h"
24 #include "test/util.h"
25 #include "vp9/common/vp9_blockd.h"
26 #include "vp9/common/vp9_scan.h"
27 #include "vpx/vpx_integer.h"
28 #include "vpx_config.h"
29 #include "vpx_ports/vpx_timer.h"
30
31 using libvpx_test::ACMRandom;
32
33 namespace {
34
35 typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
36 typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
37 typedef void (*InvTxfmWithBdFunc)(const tran_low_t *in, uint8_t *out,
38 int stride, int bd);
39
40 template <InvTxfmFunc fn>
wrapper(const tran_low_t * in,uint8_t * out,int stride,int bd)41 void wrapper(const tran_low_t *in, uint8_t *out, int stride, int bd) {
42 (void)bd;
43 fn(in, out, stride);
44 }
45
46 #if CONFIG_VP9_HIGHBITDEPTH
47 typedef void (*InvTxfmHighbdFunc)(const tran_low_t *in, uint16_t *out,
48 int stride, int bd);
49 template <InvTxfmHighbdFunc fn>
highbd_wrapper(const tran_low_t * in,uint8_t * out,int stride,int bd)50 void highbd_wrapper(const tran_low_t *in, uint8_t *out, int stride, int bd) {
51 fn(in, CAST_TO_SHORTPTR(out), stride, bd);
52 }
53 #endif
54
55 typedef std::tuple<FwdTxfmFunc, InvTxfmWithBdFunc, InvTxfmWithBdFunc, TX_SIZE,
56 int, int, int>
57 PartialInvTxfmParam;
58 const int kMaxNumCoeffs = 1024;
59 const int kCountTestBlock = 1000;
60
61 class PartialIDctTest : public ::testing::TestWithParam<PartialInvTxfmParam> {
62 public:
63 ~PartialIDctTest() override = default;
SetUp()64 void SetUp() override {
65 rnd_.Reset(ACMRandom::DeterministicSeed());
66 fwd_txfm_ = GET_PARAM(0);
67 full_inv_txfm_ = GET_PARAM(1);
68 partial_inv_txfm_ = GET_PARAM(2);
69 tx_size_ = GET_PARAM(3);
70 last_nonzero_ = GET_PARAM(4);
71 bit_depth_ = GET_PARAM(5);
72 pixel_size_ = GET_PARAM(6);
73 mask_ = (1 << bit_depth_) - 1;
74
75 switch (tx_size_) {
76 case TX_4X4: size_ = 4; break;
77 case TX_8X8: size_ = 8; break;
78 case TX_16X16: size_ = 16; break;
79 case TX_32X32: size_ = 32; break;
80 default: FAIL() << "Wrong Size!";
81 }
82
83 // Randomize stride_ to a value less than or equal to 1024
84 stride_ = rnd_(1024) + 1;
85 if (stride_ < size_) {
86 stride_ = size_;
87 }
88 // Align stride_ to 16 if it's bigger than 16.
89 if (stride_ > 16) {
90 stride_ &= ~15;
91 }
92
93 input_block_size_ = size_ * size_;
94 output_block_size_ = size_ * stride_;
95
96 input_block_ = reinterpret_cast<tran_low_t *>(
97 vpx_memalign(16, sizeof(*input_block_) * input_block_size_));
98 output_block_ = reinterpret_cast<uint8_t *>(
99 vpx_memalign(16, pixel_size_ * output_block_size_));
100 output_block_ref_ = reinterpret_cast<uint8_t *>(
101 vpx_memalign(16, pixel_size_ * output_block_size_));
102 }
103
TearDown()104 void TearDown() override {
105 vpx_free(input_block_);
106 input_block_ = nullptr;
107 vpx_free(output_block_);
108 output_block_ = nullptr;
109 vpx_free(output_block_ref_);
110 output_block_ref_ = nullptr;
111 libvpx_test::ClearSystemState();
112 }
113
InitMem()114 void InitMem() {
115 memset(input_block_, 0, sizeof(*input_block_) * input_block_size_);
116 if (pixel_size_ == 1) {
117 for (int j = 0; j < output_block_size_; ++j) {
118 output_block_[j] = output_block_ref_[j] = rnd_.Rand16() & mask_;
119 }
120 } else {
121 ASSERT_EQ(2, pixel_size_);
122 uint16_t *const output = reinterpret_cast<uint16_t *>(output_block_);
123 uint16_t *const output_ref =
124 reinterpret_cast<uint16_t *>(output_block_ref_);
125 for (int j = 0; j < output_block_size_; ++j) {
126 output[j] = output_ref[j] = rnd_.Rand16() & mask_;
127 }
128 }
129 }
130
InitInput()131 void InitInput() {
132 const int64_t max_coeff = (32766 << (bit_depth_ - 8)) / 4;
133 int64_t max_energy_leftover = max_coeff * max_coeff;
134 for (int j = 0; j < last_nonzero_; ++j) {
135 tran_low_t coeff = static_cast<tran_low_t>(
136 sqrt(1.0 * max_energy_leftover) * (rnd_.Rand16() - 32768) / 65536);
137 max_energy_leftover -= static_cast<int64_t>(coeff) * coeff;
138 if (max_energy_leftover < 0) {
139 max_energy_leftover = 0;
140 coeff = 0;
141 }
142 input_block_[vp9_default_scan_orders[tx_size_].scan[j]] = coeff;
143 }
144 }
145
PrintDiff()146 void PrintDiff() {
147 if (memcmp(output_block_ref_, output_block_,
148 pixel_size_ * output_block_size_)) {
149 uint16_t ref, opt;
150 for (int y = 0; y < size_; y++) {
151 for (int x = 0; x < size_; x++) {
152 if (pixel_size_ == 1) {
153 ref = output_block_ref_[y * stride_ + x];
154 opt = output_block_[y * stride_ + x];
155 } else {
156 ref = reinterpret_cast<uint16_t *>(
157 output_block_ref_)[y * stride_ + x];
158 opt = reinterpret_cast<uint16_t *>(output_block_)[y * stride_ + x];
159 }
160 if (ref != opt) {
161 printf("dest[%d][%d] diff:%6d (ref),%6d (opt)\n", y, x, ref, opt);
162 }
163 }
164 }
165
166 printf("\ninput_block_:\n");
167 for (int y = 0; y < size_; y++) {
168 for (int x = 0; x < size_; x++) {
169 printf("%6d,", input_block_[y * size_ + x]);
170 }
171 printf("\n");
172 }
173 }
174 }
175
176 protected:
177 int last_nonzero_;
178 TX_SIZE tx_size_;
179 tran_low_t *input_block_;
180 uint8_t *output_block_;
181 uint8_t *output_block_ref_;
182 int size_;
183 int stride_;
184 int pixel_size_;
185 int input_block_size_;
186 int output_block_size_;
187 int bit_depth_;
188 int mask_;
189 FwdTxfmFunc fwd_txfm_;
190 InvTxfmWithBdFunc full_inv_txfm_;
191 InvTxfmWithBdFunc partial_inv_txfm_;
192 ACMRandom rnd_;
193 };
194
TEST_P(PartialIDctTest,RunQuantCheck)195 TEST_P(PartialIDctTest, RunQuantCheck) {
196 const int count_test_block = (size_ != 4) ? kCountTestBlock : 65536;
197 DECLARE_ALIGNED(16, int16_t, input_extreme_block[kMaxNumCoeffs]);
198 DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kMaxNumCoeffs]);
199
200 InitMem();
201
202 for (int i = 0; i < count_test_block; ++i) {
203 // Initialize a test block with input range [-mask_, mask_].
204 if (size_ != 4) {
205 if (i == 0) {
206 for (int k = 0; k < input_block_size_; ++k) {
207 input_extreme_block[k] = mask_;
208 }
209 } else if (i == 1) {
210 for (int k = 0; k < input_block_size_; ++k) {
211 input_extreme_block[k] = -mask_;
212 }
213 } else {
214 for (int k = 0; k < input_block_size_; ++k) {
215 input_extreme_block[k] = rnd_.Rand8() % 2 ? mask_ : -mask_;
216 }
217 }
218 } else {
219 // Try all possible combinations.
220 for (int k = 0; k < input_block_size_; ++k) {
221 input_extreme_block[k] = (i & (1 << k)) ? mask_ : -mask_;
222 }
223 }
224
225 fwd_txfm_(input_extreme_block, output_ref_block, size_);
226
227 // quantization with minimum allowed step sizes
228 input_block_[0] = (output_ref_block[0] / 4) * 4;
229 for (int k = 1; k < last_nonzero_; ++k) {
230 const int pos = vp9_default_scan_orders[tx_size_].scan[k];
231 input_block_[pos] = (output_ref_block[pos] / 4) * 4;
232 }
233
234 ASM_REGISTER_STATE_CHECK(
235 full_inv_txfm_(input_block_, output_block_ref_, stride_, bit_depth_));
236 ASM_REGISTER_STATE_CHECK(
237 partial_inv_txfm_(input_block_, output_block_, stride_, bit_depth_));
238 ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
239 pixel_size_ * output_block_size_))
240 << "Error: partial inverse transform produces different results";
241 }
242 }
243
TEST_P(PartialIDctTest,ResultsMatch)244 TEST_P(PartialIDctTest, ResultsMatch) {
245 for (int i = 0; i < kCountTestBlock; ++i) {
246 InitMem();
247 InitInput();
248
249 ASM_REGISTER_STATE_CHECK(
250 full_inv_txfm_(input_block_, output_block_ref_, stride_, bit_depth_));
251 ASM_REGISTER_STATE_CHECK(
252 partial_inv_txfm_(input_block_, output_block_, stride_, bit_depth_));
253 ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
254 pixel_size_ * output_block_size_))
255 << "Error: partial inverse transform produces different results";
256 }
257 }
258
TEST_P(PartialIDctTest,AddOutputBlock)259 TEST_P(PartialIDctTest, AddOutputBlock) {
260 for (int i = 0; i < kCountTestBlock; ++i) {
261 InitMem();
262 for (int j = 0; j < last_nonzero_; ++j) {
263 input_block_[vp9_default_scan_orders[tx_size_].scan[j]] = 10;
264 }
265
266 ASM_REGISTER_STATE_CHECK(
267 full_inv_txfm_(input_block_, output_block_ref_, stride_, bit_depth_));
268 ASM_REGISTER_STATE_CHECK(
269 partial_inv_txfm_(input_block_, output_block_, stride_, bit_depth_));
270 ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
271 pixel_size_ * output_block_size_))
272 << "Error: Transform results are not correctly added to output.";
273 }
274 }
275
TEST_P(PartialIDctTest,SingleExtremeCoeff)276 TEST_P(PartialIDctTest, SingleExtremeCoeff) {
277 const int16_t max_coeff = std::numeric_limits<int16_t>::max();
278 const int16_t min_coeff = std::numeric_limits<int16_t>::min();
279 for (int i = 0; i < last_nonzero_; ++i) {
280 memset(input_block_, 0, sizeof(*input_block_) * input_block_size_);
281 // Run once for min and once for max.
282 for (int j = 0; j < 2; ++j) {
283 const int coeff = j ? min_coeff : max_coeff;
284
285 memset(output_block_, 0, pixel_size_ * output_block_size_);
286 memset(output_block_ref_, 0, pixel_size_ * output_block_size_);
287 input_block_[vp9_default_scan_orders[tx_size_].scan[i]] = coeff;
288
289 ASM_REGISTER_STATE_CHECK(
290 full_inv_txfm_(input_block_, output_block_ref_, stride_, bit_depth_));
291 ASM_REGISTER_STATE_CHECK(
292 partial_inv_txfm_(input_block_, output_block_, stride_, bit_depth_));
293 ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
294 pixel_size_ * output_block_size_))
295 << "Error: Fails with single coeff of " << coeff << " at " << i
296 << ".";
297 }
298 }
299 }
300
TEST_P(PartialIDctTest,DISABLED_Speed)301 TEST_P(PartialIDctTest, DISABLED_Speed) {
302 // Keep runtime stable with transform size.
303 const int kCountSpeedTestBlock = 500000000 / input_block_size_;
304 InitMem();
305 InitInput();
306
307 for (int i = 0; i < kCountSpeedTestBlock; ++i) {
308 ASM_REGISTER_STATE_CHECK(
309 full_inv_txfm_(input_block_, output_block_ref_, stride_, bit_depth_));
310 }
311 vpx_usec_timer timer;
312 vpx_usec_timer_start(&timer);
313 for (int i = 0; i < kCountSpeedTestBlock; ++i) {
314 partial_inv_txfm_(input_block_, output_block_, stride_, bit_depth_);
315 }
316 libvpx_test::ClearSystemState();
317 vpx_usec_timer_mark(&timer);
318 const int elapsed_time =
319 static_cast<int>(vpx_usec_timer_elapsed(&timer) / 1000);
320 printf("idct%dx%d_%d (%s %d) time: %5d ms\n", size_, size_, last_nonzero_,
321 (pixel_size_ == 1) ? "bitdepth" : "high bitdepth", bit_depth_,
322 elapsed_time);
323 ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
324 pixel_size_ * output_block_size_))
325 << "Error: partial inverse transform produces different results";
326 }
327
328 using std::make_tuple;
329
330 const PartialInvTxfmParam c_partial_idct_tests[] = {
331 #if CONFIG_VP9_HIGHBITDEPTH
332 make_tuple(
333 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
334 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>, TX_32X32, 1024, 8, 2),
335 make_tuple(
336 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
337 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>, TX_32X32, 1024, 10, 2),
338 make_tuple(
339 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
340 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>, TX_32X32, 1024, 12, 2),
341 make_tuple(
342 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
343 &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>, TX_32X32, 135, 8, 2),
344 make_tuple(
345 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
346 &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>, TX_32X32, 135, 10, 2),
347 make_tuple(
348 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
349 &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>, TX_32X32, 135, 12, 2),
350 make_tuple(
351 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
352 &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>, TX_32X32, 34, 8, 2),
353 make_tuple(
354 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
355 &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>, TX_32X32, 34, 10, 2),
356 make_tuple(
357 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
358 &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>, TX_32X32, 34, 12, 2),
359 make_tuple(&vpx_highbd_fdct32x32_c,
360 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
361 &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>, TX_32X32, 1, 8, 2),
362 make_tuple(&vpx_highbd_fdct32x32_c,
363 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
364 &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>, TX_32X32, 1, 10, 2),
365 make_tuple(&vpx_highbd_fdct32x32_c,
366 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
367 &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>, TX_32X32, 1, 12, 2),
368 make_tuple(
369 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
370 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>, TX_16X16, 256, 8, 2),
371 make_tuple(
372 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
373 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>, TX_16X16, 256, 10, 2),
374 make_tuple(
375 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
376 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>, TX_16X16, 256, 12, 2),
377 make_tuple(
378 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
379 &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>, TX_16X16, 38, 8, 2),
380 make_tuple(
381 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
382 &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>, TX_16X16, 38, 10, 2),
383 make_tuple(
384 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
385 &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>, TX_16X16, 38, 12, 2),
386 make_tuple(
387 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
388 &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>, TX_16X16, 10, 8, 2),
389 make_tuple(
390 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
391 &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>, TX_16X16, 10, 10, 2),
392 make_tuple(
393 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
394 &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>, TX_16X16, 10, 12, 2),
395 make_tuple(&vpx_highbd_fdct16x16_c,
396 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
397 &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>, TX_16X16, 1, 8, 2),
398 make_tuple(&vpx_highbd_fdct16x16_c,
399 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
400 &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>, TX_16X16, 1, 10, 2),
401 make_tuple(&vpx_highbd_fdct16x16_c,
402 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
403 &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>, TX_16X16, 1, 12, 2),
404 make_tuple(&vpx_highbd_fdct8x8_c,
405 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
406 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>, TX_8X8, 64, 8, 2),
407 make_tuple(&vpx_highbd_fdct8x8_c,
408 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
409 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>, TX_8X8, 64, 10, 2),
410 make_tuple(&vpx_highbd_fdct8x8_c,
411 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
412 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>, TX_8X8, 64, 12, 2),
413 make_tuple(&vpx_highbd_fdct8x8_c,
414 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
415 &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>, TX_8X8, 12, 8, 2),
416 make_tuple(&vpx_highbd_fdct8x8_c,
417 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
418 &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>, TX_8X8, 12, 10, 2),
419 make_tuple(&vpx_highbd_fdct8x8_c,
420 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
421 &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>, TX_8X8, 12, 12, 2),
422 make_tuple(&vpx_highbd_fdct8x8_c,
423 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
424 &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>, TX_8X8, 1, 8, 2),
425 make_tuple(&vpx_highbd_fdct8x8_c,
426 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
427 &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>, TX_8X8, 1, 10, 2),
428 make_tuple(&vpx_highbd_fdct8x8_c,
429 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
430 &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>, TX_8X8, 1, 12, 2),
431 make_tuple(&vpx_highbd_fdct4x4_c,
432 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
433 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>, TX_4X4, 16, 8, 2),
434 make_tuple(&vpx_highbd_fdct4x4_c,
435 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
436 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>, TX_4X4, 16, 10, 2),
437 make_tuple(&vpx_highbd_fdct4x4_c,
438 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
439 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>, TX_4X4, 16, 12, 2),
440 make_tuple(&vpx_highbd_fdct4x4_c,
441 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
442 &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>, TX_4X4, 1, 8, 2),
443 make_tuple(&vpx_highbd_fdct4x4_c,
444 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
445 &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>, TX_4X4, 1, 10, 2),
446 make_tuple(&vpx_highbd_fdct4x4_c,
447 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
448 &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>, TX_4X4, 1, 12, 2),
449 #endif // CONFIG_VP9_HIGHBITDEPTH
450 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
451 &wrapper<vpx_idct32x32_1024_add_c>, TX_32X32, 1024, 8, 1),
452 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
453 &wrapper<vpx_idct32x32_135_add_c>, TX_32X32, 135, 8, 1),
454 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
455 &wrapper<vpx_idct32x32_34_add_c>, TX_32X32, 34, 8, 1),
456 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
457 &wrapper<vpx_idct32x32_1_add_c>, TX_32X32, 1, 8, 1),
458 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
459 &wrapper<vpx_idct16x16_256_add_c>, TX_16X16, 256, 8, 1),
460 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
461 &wrapper<vpx_idct16x16_38_add_c>, TX_16X16, 38, 8, 1),
462 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
463 &wrapper<vpx_idct16x16_10_add_c>, TX_16X16, 10, 8, 1),
464 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
465 &wrapper<vpx_idct16x16_1_add_c>, TX_16X16, 1, 8, 1),
466 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
467 &wrapper<vpx_idct8x8_64_add_c>, TX_8X8, 64, 8, 1),
468 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
469 &wrapper<vpx_idct8x8_12_add_c>, TX_8X8, 12, 8, 1),
470 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
471 &wrapper<vpx_idct8x8_1_add_c>, TX_8X8, 1, 8, 1),
472 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
473 &wrapper<vpx_idct4x4_16_add_c>, TX_4X4, 16, 8, 1),
474 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
475 &wrapper<vpx_idct4x4_1_add_c>, TX_4X4, 1, 8, 1)
476 };
477
478 INSTANTIATE_TEST_SUITE_P(C, PartialIDctTest,
479 ::testing::ValuesIn(c_partial_idct_tests));
480
481 #if !CONFIG_EMULATE_HARDWARE
482
483 #if HAVE_NEON
484 const PartialInvTxfmParam neon_partial_idct_tests[] = {
485 #if CONFIG_VP9_HIGHBITDEPTH
486 make_tuple(&vpx_highbd_fdct32x32_c,
487 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
488 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_neon>, TX_32X32,
489 1024, 8, 2),
490 make_tuple(&vpx_highbd_fdct32x32_c,
491 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
492 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_neon>, TX_32X32,
493 1024, 10, 2),
494 make_tuple(&vpx_highbd_fdct32x32_c,
495 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
496 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_neon>, TX_32X32,
497 1024, 12, 2),
498 make_tuple(
499 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
500 &highbd_wrapper<vpx_highbd_idct32x32_135_add_neon>, TX_32X32, 135, 8, 2),
501 make_tuple(
502 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
503 &highbd_wrapper<vpx_highbd_idct32x32_135_add_neon>, TX_32X32, 135, 10, 2),
504 make_tuple(
505 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
506 &highbd_wrapper<vpx_highbd_idct32x32_135_add_neon>, TX_32X32, 135, 12, 2),
507 make_tuple(
508 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
509 &highbd_wrapper<vpx_highbd_idct32x32_34_add_neon>, TX_32X32, 34, 8, 2),
510 make_tuple(
511 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
512 &highbd_wrapper<vpx_highbd_idct32x32_34_add_neon>, TX_32X32, 34, 10, 2),
513 make_tuple(
514 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
515 &highbd_wrapper<vpx_highbd_idct32x32_34_add_neon>, TX_32X32, 34, 12, 2),
516 make_tuple(
517 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>,
518 &highbd_wrapper<vpx_highbd_idct32x32_1_add_neon>, TX_32X32, 1, 8, 2),
519 make_tuple(
520 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>,
521 &highbd_wrapper<vpx_highbd_idct32x32_1_add_neon>, TX_32X32, 1, 10, 2),
522 make_tuple(
523 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>,
524 &highbd_wrapper<vpx_highbd_idct32x32_1_add_neon>, TX_32X32, 1, 12, 2),
525 make_tuple(
526 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
527 &highbd_wrapper<vpx_highbd_idct16x16_256_add_neon>, TX_16X16, 256, 8, 2),
528 make_tuple(
529 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
530 &highbd_wrapper<vpx_highbd_idct16x16_256_add_neon>, TX_16X16, 256, 10, 2),
531 make_tuple(
532 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
533 &highbd_wrapper<vpx_highbd_idct16x16_256_add_neon>, TX_16X16, 256, 12, 2),
534 make_tuple(
535 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
536 &highbd_wrapper<vpx_highbd_idct16x16_38_add_neon>, TX_16X16, 38, 8, 2),
537 make_tuple(
538 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
539 &highbd_wrapper<vpx_highbd_idct16x16_38_add_neon>, TX_16X16, 38, 10, 2),
540 make_tuple(
541 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
542 &highbd_wrapper<vpx_highbd_idct16x16_38_add_neon>, TX_16X16, 38, 12, 2),
543 make_tuple(
544 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
545 &highbd_wrapper<vpx_highbd_idct16x16_10_add_neon>, TX_16X16, 10, 8, 2),
546 make_tuple(
547 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
548 &highbd_wrapper<vpx_highbd_idct16x16_10_add_neon>, TX_16X16, 10, 10, 2),
549 make_tuple(
550 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
551 &highbd_wrapper<vpx_highbd_idct16x16_10_add_neon>, TX_16X16, 10, 12, 2),
552 make_tuple(
553 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>,
554 &highbd_wrapper<vpx_highbd_idct16x16_1_add_neon>, TX_16X16, 1, 8, 2),
555 make_tuple(
556 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>,
557 &highbd_wrapper<vpx_highbd_idct16x16_1_add_neon>, TX_16X16, 1, 10, 2),
558 make_tuple(
559 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>,
560 &highbd_wrapper<vpx_highbd_idct16x16_1_add_neon>, TX_16X16, 1, 12, 2),
561 make_tuple(&vpx_highbd_fdct8x8_c,
562 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
563 &highbd_wrapper<vpx_highbd_idct8x8_64_add_neon>, TX_8X8, 64, 8, 2),
564 make_tuple(
565 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
566 &highbd_wrapper<vpx_highbd_idct8x8_64_add_neon>, TX_8X8, 64, 10, 2),
567 make_tuple(
568 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
569 &highbd_wrapper<vpx_highbd_idct8x8_64_add_neon>, TX_8X8, 64, 12, 2),
570 make_tuple(&vpx_highbd_fdct8x8_c,
571 &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
572 &highbd_wrapper<vpx_highbd_idct8x8_12_add_neon>, TX_8X8, 12, 8, 2),
573 make_tuple(
574 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
575 &highbd_wrapper<vpx_highbd_idct8x8_12_add_neon>, TX_8X8, 12, 10, 2),
576 make_tuple(
577 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
578 &highbd_wrapper<vpx_highbd_idct8x8_12_add_neon>, TX_8X8, 12, 12, 2),
579 make_tuple(&vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>,
580 &highbd_wrapper<vpx_highbd_idct8x8_1_add_neon>, TX_8X8, 1, 8, 2),
581 make_tuple(&vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>,
582 &highbd_wrapper<vpx_highbd_idct8x8_1_add_neon>, TX_8X8, 1, 10, 2),
583 make_tuple(&vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>,
584 &highbd_wrapper<vpx_highbd_idct8x8_1_add_neon>, TX_8X8, 1, 12, 2),
585 make_tuple(&vpx_highbd_fdct4x4_c,
586 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
587 &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 16, 8, 2),
588 make_tuple(
589 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
590 &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 16, 10, 2),
591 make_tuple(
592 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
593 &highbd_wrapper<vpx_highbd_idct4x4_16_add_neon>, TX_4X4, 16, 12, 2),
594 make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
595 &highbd_wrapper<vpx_highbd_idct4x4_1_add_neon>, TX_4X4, 1, 8, 2),
596 make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
597 &highbd_wrapper<vpx_highbd_idct4x4_1_add_neon>, TX_4X4, 1, 10, 2),
598 make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
599 &highbd_wrapper<vpx_highbd_idct4x4_1_add_neon>, TX_4X4, 1, 12, 2),
600 #endif // CONFIG_VP9_HIGHBITDEPTH
601 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
602 &wrapper<vpx_idct32x32_1024_add_neon>, TX_32X32, 1024, 8, 1),
603 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_135_add_c>,
604 &wrapper<vpx_idct32x32_135_add_neon>, TX_32X32, 135, 8, 1),
605 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_34_add_c>,
606 &wrapper<vpx_idct32x32_34_add_neon>, TX_32X32, 34, 8, 1),
607 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1_add_c>,
608 &wrapper<vpx_idct32x32_1_add_neon>, TX_32X32, 1, 8, 1),
609 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
610 &wrapper<vpx_idct16x16_256_add_neon>, TX_16X16, 256, 8, 1),
611 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_38_add_c>,
612 &wrapper<vpx_idct16x16_38_add_neon>, TX_16X16, 38, 8, 1),
613 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_10_add_c>,
614 &wrapper<vpx_idct16x16_10_add_neon>, TX_16X16, 10, 8, 1),
615 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_1_add_c>,
616 &wrapper<vpx_idct16x16_1_add_neon>, TX_16X16, 1, 8, 1),
617 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
618 &wrapper<vpx_idct8x8_64_add_neon>, TX_8X8, 64, 8, 1),
619 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_12_add_c>,
620 &wrapper<vpx_idct8x8_12_add_neon>, TX_8X8, 12, 8, 1),
621 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_1_add_c>,
622 &wrapper<vpx_idct8x8_1_add_neon>, TX_8X8, 1, 8, 1),
623 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
624 &wrapper<vpx_idct4x4_16_add_neon>, TX_4X4, 16, 8, 1),
625 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_1_add_c>,
626 &wrapper<vpx_idct4x4_1_add_neon>, TX_4X4, 1, 8, 1)
627 };
628
629 INSTANTIATE_TEST_SUITE_P(NEON, PartialIDctTest,
630 ::testing::ValuesIn(neon_partial_idct_tests));
631 #endif // HAVE_NEON
632
633 #if HAVE_SSE2
634 // 32x32_135_ is implemented using the 1024 version.
635 const PartialInvTxfmParam sse2_partial_idct_tests[] = {
636 #if CONFIG_VP9_HIGHBITDEPTH
637 make_tuple(&vpx_highbd_fdct32x32_c,
638 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
639 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_sse2>, TX_32X32,
640 1024, 8, 2),
641 make_tuple(&vpx_highbd_fdct32x32_c,
642 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
643 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_sse2>, TX_32X32,
644 1024, 10, 2),
645 make_tuple(&vpx_highbd_fdct32x32_c,
646 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
647 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_sse2>, TX_32X32,
648 1024, 12, 2),
649 make_tuple(
650 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
651 &highbd_wrapper<vpx_highbd_idct32x32_135_add_sse2>, TX_32X32, 135, 8, 2),
652 make_tuple(
653 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
654 &highbd_wrapper<vpx_highbd_idct32x32_135_add_sse2>, TX_32X32, 135, 10, 2),
655 make_tuple(
656 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
657 &highbd_wrapper<vpx_highbd_idct32x32_135_add_sse2>, TX_32X32, 135, 12, 2),
658 make_tuple(
659 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
660 &highbd_wrapper<vpx_highbd_idct32x32_34_add_sse2>, TX_32X32, 34, 8, 2),
661 make_tuple(
662 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
663 &highbd_wrapper<vpx_highbd_idct32x32_34_add_sse2>, TX_32X32, 34, 10, 2),
664 make_tuple(
665 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
666 &highbd_wrapper<vpx_highbd_idct32x32_34_add_sse2>, TX_32X32, 34, 12, 2),
667 make_tuple(
668 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>,
669 &highbd_wrapper<vpx_highbd_idct32x32_1_add_sse2>, TX_32X32, 1, 8, 2),
670 make_tuple(
671 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>,
672 &highbd_wrapper<vpx_highbd_idct32x32_1_add_sse2>, TX_32X32, 1, 10, 2),
673 make_tuple(
674 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1_add_c>,
675 &highbd_wrapper<vpx_highbd_idct32x32_1_add_sse2>, TX_32X32, 1, 12, 2),
676 make_tuple(
677 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
678 &highbd_wrapper<vpx_highbd_idct16x16_256_add_sse2>, TX_16X16, 256, 8, 2),
679 make_tuple(
680 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
681 &highbd_wrapper<vpx_highbd_idct16x16_256_add_sse2>, TX_16X16, 256, 10, 2),
682 make_tuple(
683 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
684 &highbd_wrapper<vpx_highbd_idct16x16_256_add_sse2>, TX_16X16, 256, 12, 2),
685 make_tuple(
686 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
687 &highbd_wrapper<vpx_highbd_idct16x16_38_add_sse2>, TX_16X16, 38, 8, 2),
688 make_tuple(
689 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
690 &highbd_wrapper<vpx_highbd_idct16x16_38_add_sse2>, TX_16X16, 38, 10, 2),
691 make_tuple(
692 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
693 &highbd_wrapper<vpx_highbd_idct16x16_38_add_sse2>, TX_16X16, 38, 12, 2),
694 make_tuple(
695 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
696 &highbd_wrapper<vpx_highbd_idct16x16_10_add_sse2>, TX_16X16, 10, 8, 2),
697 make_tuple(
698 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
699 &highbd_wrapper<vpx_highbd_idct16x16_10_add_sse2>, TX_16X16, 10, 10, 2),
700 make_tuple(
701 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
702 &highbd_wrapper<vpx_highbd_idct16x16_10_add_sse2>, TX_16X16, 10, 12, 2),
703 make_tuple(
704 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>,
705 &highbd_wrapper<vpx_highbd_idct16x16_1_add_sse2>, TX_16X16, 1, 8, 2),
706 make_tuple(
707 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>,
708 &highbd_wrapper<vpx_highbd_idct16x16_1_add_sse2>, TX_16X16, 1, 10, 2),
709 make_tuple(
710 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_1_add_c>,
711 &highbd_wrapper<vpx_highbd_idct16x16_1_add_sse2>, TX_16X16, 1, 12, 2),
712 make_tuple(&vpx_highbd_fdct8x8_c,
713 &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
714 &highbd_wrapper<vpx_highbd_idct8x8_64_add_sse2>, TX_8X8, 64, 8, 2),
715 make_tuple(
716 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
717 &highbd_wrapper<vpx_highbd_idct8x8_64_add_sse2>, TX_8X8, 64, 10, 2),
718 make_tuple(
719 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
720 &highbd_wrapper<vpx_highbd_idct8x8_64_add_sse2>, TX_8X8, 64, 12, 2),
721 make_tuple(&vpx_highbd_fdct8x8_c,
722 &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
723 &highbd_wrapper<vpx_highbd_idct8x8_12_add_sse2>, TX_8X8, 12, 8, 2),
724 make_tuple(
725 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
726 &highbd_wrapper<vpx_highbd_idct8x8_12_add_sse2>, TX_8X8, 12, 10, 2),
727 make_tuple(
728 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
729 &highbd_wrapper<vpx_highbd_idct8x8_12_add_sse2>, TX_8X8, 12, 12, 2),
730 make_tuple(&vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>,
731 &highbd_wrapper<vpx_highbd_idct8x8_1_add_sse2>, TX_8X8, 1, 8, 2),
732 make_tuple(&vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>,
733 &highbd_wrapper<vpx_highbd_idct8x8_1_add_sse2>, TX_8X8, 1, 10, 2),
734 make_tuple(&vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_1_add_c>,
735 &highbd_wrapper<vpx_highbd_idct8x8_1_add_sse2>, TX_8X8, 1, 12, 2),
736 make_tuple(&vpx_highbd_fdct4x4_c,
737 &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
738 &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 8, 2),
739 make_tuple(
740 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
741 &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 10, 2),
742 make_tuple(
743 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
744 &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 12, 2),
745 make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
746 &highbd_wrapper<vpx_highbd_idct4x4_1_add_sse2>, TX_4X4, 1, 8, 2),
747 make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
748 &highbd_wrapper<vpx_highbd_idct4x4_1_add_sse2>, TX_4X4, 1, 10, 2),
749 make_tuple(&vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_1_add_c>,
750 &highbd_wrapper<vpx_highbd_idct4x4_1_add_sse2>, TX_4X4, 1, 12, 2),
751 #endif // CONFIG_VP9_HIGHBITDEPTH
752 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
753 &wrapper<vpx_idct32x32_1024_add_sse2>, TX_32X32, 1024, 8, 1),
754 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_135_add_c>,
755 &wrapper<vpx_idct32x32_135_add_sse2>, TX_32X32, 135, 8, 1),
756 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_34_add_c>,
757 &wrapper<vpx_idct32x32_34_add_sse2>, TX_32X32, 34, 8, 1),
758 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1_add_c>,
759 &wrapper<vpx_idct32x32_1_add_sse2>, TX_32X32, 1, 8, 1),
760 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
761 &wrapper<vpx_idct16x16_256_add_sse2>, TX_16X16, 256, 8, 1),
762 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_38_add_c>,
763 &wrapper<vpx_idct16x16_38_add_sse2>, TX_16X16, 38, 8, 1),
764 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_10_add_c>,
765 &wrapper<vpx_idct16x16_10_add_sse2>, TX_16X16, 10, 8, 1),
766 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_1_add_c>,
767 &wrapper<vpx_idct16x16_1_add_sse2>, TX_16X16, 1, 8, 1),
768 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
769 &wrapper<vpx_idct8x8_64_add_sse2>, TX_8X8, 64, 8, 1),
770 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_12_add_c>,
771 &wrapper<vpx_idct8x8_12_add_sse2>, TX_8X8, 12, 8, 1),
772 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_1_add_c>,
773 &wrapper<vpx_idct8x8_1_add_sse2>, TX_8X8, 1, 8, 1),
774 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
775 &wrapper<vpx_idct4x4_16_add_sse2>, TX_4X4, 16, 8, 1),
776 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_1_add_c>,
777 &wrapper<vpx_idct4x4_1_add_sse2>, TX_4X4, 1, 8, 1)
778 };
779
780 INSTANTIATE_TEST_SUITE_P(SSE2, PartialIDctTest,
781 ::testing::ValuesIn(sse2_partial_idct_tests));
782
783 #endif // HAVE_SSE2
784
785 #if HAVE_SSSE3
786 const PartialInvTxfmParam ssse3_partial_idct_tests[] = {
787 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_135_add_c>,
788 &wrapper<vpx_idct32x32_135_add_ssse3>, TX_32X32, 135, 8, 1),
789 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_34_add_c>,
790 &wrapper<vpx_idct32x32_34_add_ssse3>, TX_32X32, 34, 8, 1),
791 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_12_add_c>,
792 &wrapper<vpx_idct8x8_12_add_ssse3>, TX_8X8, 12, 8, 1)
793 };
794
795 INSTANTIATE_TEST_SUITE_P(SSSE3, PartialIDctTest,
796 ::testing::ValuesIn(ssse3_partial_idct_tests));
797 #endif // HAVE_SSSE3
798
799 #if HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
800 const PartialInvTxfmParam sse4_1_partial_idct_tests[] = {
801 make_tuple(&vpx_highbd_fdct32x32_c,
802 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
803 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_sse4_1>, TX_32X32,
804 1024, 8, 2),
805 make_tuple(&vpx_highbd_fdct32x32_c,
806 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
807 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_sse4_1>, TX_32X32,
808 1024, 10, 2),
809 make_tuple(&vpx_highbd_fdct32x32_c,
810 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
811 &highbd_wrapper<vpx_highbd_idct32x32_1024_add_sse4_1>, TX_32X32,
812 1024, 12, 2),
813 make_tuple(&vpx_highbd_fdct32x32_c,
814 &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
815 &highbd_wrapper<vpx_highbd_idct32x32_135_add_sse4_1>, TX_32X32,
816 135, 8, 2),
817 make_tuple(&vpx_highbd_fdct32x32_c,
818 &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
819 &highbd_wrapper<vpx_highbd_idct32x32_135_add_sse4_1>, TX_32X32,
820 135, 10, 2),
821 make_tuple(&vpx_highbd_fdct32x32_c,
822 &highbd_wrapper<vpx_highbd_idct32x32_135_add_c>,
823 &highbd_wrapper<vpx_highbd_idct32x32_135_add_sse4_1>, TX_32X32,
824 135, 12, 2),
825 make_tuple(
826 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
827 &highbd_wrapper<vpx_highbd_idct32x32_34_add_sse4_1>, TX_32X32, 34, 8, 2),
828 make_tuple(
829 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
830 &highbd_wrapper<vpx_highbd_idct32x32_34_add_sse4_1>, TX_32X32, 34, 10, 2),
831 make_tuple(
832 &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_34_add_c>,
833 &highbd_wrapper<vpx_highbd_idct32x32_34_add_sse4_1>, TX_32X32, 34, 12, 2),
834 make_tuple(&vpx_highbd_fdct16x16_c,
835 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
836 &highbd_wrapper<vpx_highbd_idct16x16_256_add_sse4_1>, TX_16X16,
837 256, 8, 2),
838 make_tuple(&vpx_highbd_fdct16x16_c,
839 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
840 &highbd_wrapper<vpx_highbd_idct16x16_256_add_sse4_1>, TX_16X16,
841 256, 10, 2),
842 make_tuple(&vpx_highbd_fdct16x16_c,
843 &highbd_wrapper<vpx_highbd_idct16x16_256_add_c>,
844 &highbd_wrapper<vpx_highbd_idct16x16_256_add_sse4_1>, TX_16X16,
845 256, 12, 2),
846 make_tuple(
847 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
848 &highbd_wrapper<vpx_highbd_idct16x16_38_add_sse4_1>, TX_16X16, 38, 8, 2),
849 make_tuple(
850 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
851 &highbd_wrapper<vpx_highbd_idct16x16_38_add_sse4_1>, TX_16X16, 38, 10, 2),
852 make_tuple(
853 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_38_add_c>,
854 &highbd_wrapper<vpx_highbd_idct16x16_38_add_sse4_1>, TX_16X16, 38, 12, 2),
855 make_tuple(
856 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
857 &highbd_wrapper<vpx_highbd_idct16x16_10_add_sse4_1>, TX_16X16, 10, 8, 2),
858 make_tuple(
859 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
860 &highbd_wrapper<vpx_highbd_idct16x16_10_add_sse4_1>, TX_16X16, 10, 10, 2),
861 make_tuple(
862 &vpx_highbd_fdct16x16_c, &highbd_wrapper<vpx_highbd_idct16x16_10_add_c>,
863 &highbd_wrapper<vpx_highbd_idct16x16_10_add_sse4_1>, TX_16X16, 10, 12, 2),
864 make_tuple(
865 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
866 &highbd_wrapper<vpx_highbd_idct8x8_64_add_sse4_1>, TX_8X8, 64, 8, 2),
867 make_tuple(
868 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
869 &highbd_wrapper<vpx_highbd_idct8x8_64_add_sse4_1>, TX_8X8, 64, 10, 2),
870 make_tuple(
871 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_64_add_c>,
872 &highbd_wrapper<vpx_highbd_idct8x8_64_add_sse4_1>, TX_8X8, 64, 12, 2),
873 make_tuple(
874 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
875 &highbd_wrapper<vpx_highbd_idct8x8_12_add_sse4_1>, TX_8X8, 12, 8, 2),
876 make_tuple(
877 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
878 &highbd_wrapper<vpx_highbd_idct8x8_12_add_sse4_1>, TX_8X8, 12, 10, 2),
879 make_tuple(
880 &vpx_highbd_fdct8x8_c, &highbd_wrapper<vpx_highbd_idct8x8_12_add_c>,
881 &highbd_wrapper<vpx_highbd_idct8x8_12_add_sse4_1>, TX_8X8, 12, 12, 2),
882 make_tuple(
883 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
884 &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse4_1>, TX_4X4, 16, 8, 2),
885 make_tuple(
886 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
887 &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse4_1>, TX_4X4, 16, 10, 2),
888 make_tuple(
889 &vpx_highbd_fdct4x4_c, &highbd_wrapper<vpx_highbd_idct4x4_16_add_c>,
890 &highbd_wrapper<vpx_highbd_idct4x4_16_add_sse4_1>, TX_4X4, 16, 12, 2)
891 };
892
893 INSTANTIATE_TEST_SUITE_P(SSE4_1, PartialIDctTest,
894 ::testing::ValuesIn(sse4_1_partial_idct_tests));
895 #endif // HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
896
897 #if HAVE_DSPR2 && !CONFIG_VP9_HIGHBITDEPTH
898 const PartialInvTxfmParam dspr2_partial_idct_tests[] = {
899 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
900 &wrapper<vpx_idct32x32_1024_add_dspr2>, TX_32X32, 1024, 8, 1),
901 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_34_add_c>,
902 &wrapper<vpx_idct32x32_34_add_dspr2>, TX_32X32, 34, 8, 1),
903 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1_add_c>,
904 &wrapper<vpx_idct32x32_1_add_dspr2>, TX_32X32, 1, 8, 1),
905 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
906 &wrapper<vpx_idct16x16_256_add_dspr2>, TX_16X16, 256, 8, 1),
907 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_10_add_c>,
908 &wrapper<vpx_idct16x16_10_add_dspr2>, TX_16X16, 10, 8, 1),
909 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_1_add_c>,
910 &wrapper<vpx_idct16x16_1_add_dspr2>, TX_16X16, 1, 8, 1),
911 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
912 &wrapper<vpx_idct8x8_64_add_dspr2>, TX_8X8, 64, 8, 1),
913 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_12_add_c>,
914 &wrapper<vpx_idct8x8_12_add_dspr2>, TX_8X8, 12, 8, 1),
915 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_1_add_c>,
916 &wrapper<vpx_idct8x8_1_add_dspr2>, TX_8X8, 1, 8, 1),
917 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
918 &wrapper<vpx_idct4x4_16_add_dspr2>, TX_4X4, 16, 8, 1),
919 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_1_add_c>,
920 &wrapper<vpx_idct4x4_1_add_dspr2>, TX_4X4, 1, 8, 1)
921 };
922
923 INSTANTIATE_TEST_SUITE_P(DSPR2, PartialIDctTest,
924 ::testing::ValuesIn(dspr2_partial_idct_tests));
925 #endif // HAVE_DSPR2 && !CONFIG_VP9_HIGHBITDEPTH
926
927 #if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH
928 // 32x32_135_ is implemented using the 1024 version.
929 const PartialInvTxfmParam msa_partial_idct_tests[] = {
930 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
931 &wrapper<vpx_idct32x32_1024_add_msa>, TX_32X32, 1024, 8, 1),
932 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_34_add_c>,
933 &wrapper<vpx_idct32x32_34_add_msa>, TX_32X32, 34, 8, 1),
934 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1_add_c>,
935 &wrapper<vpx_idct32x32_1_add_msa>, TX_32X32, 1, 8, 1),
936 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
937 &wrapper<vpx_idct16x16_256_add_msa>, TX_16X16, 256, 8, 1),
938 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_10_add_c>,
939 &wrapper<vpx_idct16x16_10_add_msa>, TX_16X16, 10, 8, 1),
940 make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_1_add_c>,
941 &wrapper<vpx_idct16x16_1_add_msa>, TX_16X16, 1, 8, 1),
942 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_64_add_c>,
943 &wrapper<vpx_idct8x8_64_add_msa>, TX_8X8, 64, 8, 1),
944 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_12_add_c>,
945 &wrapper<vpx_idct8x8_12_add_msa>, TX_8X8, 12, 8, 1),
946 make_tuple(&vpx_fdct8x8_c, &wrapper<vpx_idct8x8_1_add_c>,
947 &wrapper<vpx_idct8x8_1_add_msa>, TX_8X8, 1, 8, 1),
948 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_16_add_c>,
949 &wrapper<vpx_idct4x4_16_add_msa>, TX_4X4, 16, 8, 1),
950 make_tuple(&vpx_fdct4x4_c, &wrapper<vpx_idct4x4_1_add_c>,
951 &wrapper<vpx_idct4x4_1_add_msa>, TX_4X4, 1, 8, 1)
952 };
953
954 INSTANTIATE_TEST_SUITE_P(MSA, PartialIDctTest,
955 ::testing::ValuesIn(msa_partial_idct_tests));
956 #endif // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH
957
958 #if HAVE_LSX && !CONFIG_VP9_HIGHBITDEPTH
959 const PartialInvTxfmParam lsx_partial_idct_tests[] = {
960 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1024_add_c>,
961 &wrapper<vpx_idct32x32_1024_add_lsx>, TX_32X32, 1024, 8, 1),
962 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_34_add_c>,
963 &wrapper<vpx_idct32x32_34_add_lsx>, TX_32X32, 34, 8, 1),
964 make_tuple(&vpx_fdct32x32_c, &wrapper<vpx_idct32x32_1_add_c>,
965 &wrapper<vpx_idct32x32_1_add_lsx>, TX_32X32, 1, 8, 1),
966 };
967
968 INSTANTIATE_TEST_SUITE_P(LSX, PartialIDctTest,
969 ::testing::ValuesIn(lsx_partial_idct_tests));
970 #endif // HAVE_LSX && !CONFIG_VP9_HIGHBITDEPTH
971
972 #endif // !CONFIG_EMULATE_HARDWARE
973
974 } // namespace
975