This allows catching whether the functions write outside of the designated rectangle, and if run with "checkasm -v", it also prints out on which side of the rectangle the overwrite was.
Unified · Split
| @@ -44,9 +44,14 @@ static const int offsets[] = {0, 255, -1 }; · 43 unmodified lines | |||
| 44 | uint32_t r = rnd() & mask; \ | 44 | uint32_t r = rnd() & mask; \ |
| 45 | AV_WN32A(buf0 + k, r); \ | 45 | AV_WN32A(buf0 + k, r); \ |
| 46 | AV_WN32A(buf1 + k, r); \ | 46 | AV_WN32A(buf1 + k, r); \ |
| 47 | if (k >= BUF_SIZE) \ | 47 | } \ |
| 48 | continue; \ | 48 | } while (0) |
| 49 | r = rnd(); \ | 49 | |
| 50 | #define randomize_buffers_dst() \ | ||
| 51 | do { \ | ||
| 52 | int k; \ | ||
| 53 | for (k = 0; k < BUF_SIZE; k += 4) { \ | ||
| 54 | uint32_t r = rnd(); \ | ||
| 50 | AV_WN32A(dst0 + k, r); \ | 55 | AV_WN32A(dst0 + k, r); \ |
| 51 | AV_WN32A(dst1 + k, r); \ | 56 | AV_WN32A(dst1 + k, r); \ |
| 52 | } \ | 57 | } \ |
| @@ -100,6 +105,7 @@ static void checkasm_check_hevc_qpel(void) · 47 unmodified lines | |||
| 100 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { | 105 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { |
| 101 | int16_t *dstw0 = (int16_t *) dst0, *dstw1 = (int16_t *) dst1; | 106 | int16_t *dstw0 = (int16_t *) dst0, *dstw1 = (int16_t *) dst1; |
| 102 | randomize_buffers(); | 107 | randomize_buffers(); |
| 108 | randomize_buffers_dst(); | ||
| 103 | call_ref(dstw0, src0, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); | 109 | call_ref(dstw0, src0, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); |
| 104 | call_new(dstw1, src1, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); | 110 | call_new(dstw1, src1, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); |
| 105 | checkasm_check(int16_t, dstw0, MAX_PB_SIZE * sizeof(int16_t), | 111 | checkasm_check(int16_t, dstw0, MAX_PB_SIZE * sizeof(int16_t), |
| @@ -118,8 +124,8 @@ static void checkasm_check_hevc_qpel_uni(void) · 12 unmodified lines | |||
| 118 | { | 124 | { |
| 119 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); | 125 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); |
| 120 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); | 126 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); |
| 121 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 127 | PIXEL_RECT(dst0, 64, 64); |
| 122 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 128 | PIXEL_RECT(dst1, 64, 64); |
| 123 | 129 | ||
| 124 | HEVCDSPContext h; | 130 | HEVCDSPContext h; |
| 125 | int size, bit_depth, i, j; | 131 | int size, bit_depth, i, j; |
| @@ -143,16 +149,18 @@ static void checkasm_check_hevc_qpel_uni(void) · 17 unmodified lines | |||
| 143 | if (check_func(h.put_hevc_qpel_uni[size][j][i], | 149 | if (check_func(h.put_hevc_qpel_uni[size][j][i], |
| 144 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { | 150 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { |
| 145 | randomize_buffers(); | 151 | randomize_buffers(); |
| 146 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 152 | CLEAR_PIXEL_RECT(dst0); |
| 153 | CLEAR_PIXEL_RECT(dst1); | ||
| 154 | call_ref(dst0, dst0_stride, | ||
| 147 | src0, sizes[size] * SIZEOF_PIXEL, | 155 | src0, sizes[size] * SIZEOF_PIXEL, |
| 148 | sizes[size], i, j, sizes[size]); | 156 | sizes[size], i, j, sizes[size]); |
| 149 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 157 | call_new(dst1, dst1_stride, |
| 150 | src1, sizes[size] * SIZEOF_PIXEL, | 158 | src1, sizes[size] * SIZEOF_PIXEL, |
| 151 | sizes[size], i, j, sizes[size]); | 159 | sizes[size], i, j, sizes[size]); |
| 152 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 160 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 153 | dst1, sizes[size] * SIZEOF_PIXEL, | 161 | dst1, dst1_stride, |
| 154 | size[sizes], size[sizes], "dst"); | 162 | size[sizes], size[sizes], "dst"); |
| 155 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 163 | bench_new(dst1, dst1_stride, |
| 156 | src1, sizes[size] * SIZEOF_PIXEL, | 164 | src1, sizes[size] * SIZEOF_PIXEL, |
| 157 | sizes[size], i, j, sizes[size]); | 165 | sizes[size], i, j, sizes[size]); |
| 158 | } | 166 | } |
| @@ -167,8 +175,8 @@ static void checkasm_check_hevc_qpel_uni_w(void) · 8 unmodified lines | |||
| 167 | { | 175 | { |
| 168 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); | 176 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); |
| 169 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); | 177 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); |
| 170 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 178 | PIXEL_RECT(dst0, 64, 64); |
| 171 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 179 | PIXEL_RECT(dst1, 64, 64); |
| 172 | 180 | ||
| 173 | HEVCDSPContext h; | 181 | HEVCDSPContext h; |
| 174 | int size, bit_depth, i, j; | 182 | int size, bit_depth, i, j; |
| @@ -196,16 +204,18 @@ static void checkasm_check_hevc_qpel_uni_w(void) · 21 unmodified lines | |||
| 196 | for (wx = weights; *wx >= 0; wx++) { | 204 | for (wx = weights; *wx >= 0; wx++) { |
| 197 | for (ox = offsets; *ox >= 0; ox++) { | 205 | for (ox = offsets; *ox >= 0; ox++) { |
| 198 | randomize_buffers(); | 206 | randomize_buffers(); |
| 199 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 207 | CLEAR_PIXEL_RECT(dst0); |
| 208 | CLEAR_PIXEL_RECT(dst1); | ||
| 209 | call_ref(dst0, dst0_stride, | ||
| 200 | src0, sizes[size] * SIZEOF_PIXEL, | 210 | src0, sizes[size] * SIZEOF_PIXEL, |
| 201 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); | 211 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); |
| 202 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 212 | call_new(dst1, dst1_stride, |
| 203 | src1, sizes[size] * SIZEOF_PIXEL, | 213 | src1, sizes[size] * SIZEOF_PIXEL, |
| 204 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); | 214 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); |
| 205 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 215 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 206 | dst1, sizes[size] * SIZEOF_PIXEL, | 216 | dst1, dst1_stride, |
| 207 | size[sizes], size[sizes], "dst"); | 217 | size[sizes], size[sizes], "dst"); |
| 208 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 218 | bench_new(dst1, dst1_stride, |
| 209 | src1, sizes[size] * SIZEOF_PIXEL, | 219 | src1, sizes[size] * SIZEOF_PIXEL, |
| 210 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); | 220 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); |
| 211 | } | 221 | } |
| @@ -223,8 +233,8 @@ static void checkasm_check_hevc_qpel_bi(void) · 11 unmodified lines | |||
| 223 | { | 233 | { |
| 224 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); | 234 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); |
| 225 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); | 235 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); |
| 226 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 236 | PIXEL_RECT(dst0, 64, 64); |
| 227 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 237 | PIXEL_RECT(dst1, 64, 64); |
| 228 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); | 238 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); |
| 229 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); | 239 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); |
| 230 | 240 | ||
| @@ -251,16 +261,18 @@ static void checkasm_check_hevc_qpel_bi(void) · 20 unmodified lines | |||
| 251 | if (check_func(h.put_hevc_qpel_bi[size][j][i], | 261 | if (check_func(h.put_hevc_qpel_bi[size][j][i], |
| 252 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { | 262 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { |
| 253 | randomize_buffers_ref(); | 263 | randomize_buffers_ref(); |
| 254 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 264 | CLEAR_PIXEL_RECT(dst0); |
| 265 | CLEAR_PIXEL_RECT(dst1); | ||
| 266 | call_ref(dst0, dst0_stride, | ||
| 255 | src0, sizes[size] * SIZEOF_PIXEL, | 267 | src0, sizes[size] * SIZEOF_PIXEL, |
| 256 | ref0, sizes[size], i, j, sizes[size]); | 268 | ref0, sizes[size], i, j, sizes[size]); |
| 257 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 269 | call_new(dst1, dst1_stride, |
| 258 | src1, sizes[size] * SIZEOF_PIXEL, | 270 | src1, sizes[size] * SIZEOF_PIXEL, |
| 259 | ref1, sizes[size], i, j, sizes[size]); | 271 | ref1, sizes[size], i, j, sizes[size]); |
| 260 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 272 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 261 | dst1, sizes[size] * SIZEOF_PIXEL, | 273 | dst1, dst1_stride, |
| 262 | size[sizes], size[sizes], "dst"); | 274 | size[sizes], size[sizes], "dst"); |
| 263 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 275 | bench_new(dst1, dst1_stride, |
| 264 | src1, sizes[size] * SIZEOF_PIXEL, | 276 | src1, sizes[size] * SIZEOF_PIXEL, |
| 265 | ref1, sizes[size], i, j, sizes[size]); | 277 | ref1, sizes[size], i, j, sizes[size]); |
| 266 | } | 278 | } |
| @@ -275,8 +287,8 @@ static void checkasm_check_hevc_qpel_bi_w(void) · 8 unmodified lines | |||
| 275 | { | 287 | { |
| 276 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); | 288 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE + SRC_EXTRA]); |
| 277 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); | 289 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE + SRC_EXTRA]); |
| 278 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 290 | PIXEL_RECT(dst0, 64, 64); |
| 279 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 291 | PIXEL_RECT(dst1, 64, 64); |
| 280 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); | 292 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); |
| 281 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); | 293 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); |
| 282 | 294 | ||
| @@ -308,16 +320,18 @@ static void checkasm_check_hevc_qpel_bi_w(void) · 25 unmodified lines | |||
| 308 | for (wx = weights; *wx >= 0; wx++) { | 320 | for (wx = weights; *wx >= 0; wx++) { |
| 309 | for (ox = offsets; *ox >= 0; ox++) { | 321 | for (ox = offsets; *ox >= 0; ox++) { |
| 310 | randomize_buffers_ref(); | 322 | randomize_buffers_ref(); |
| 311 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 323 | CLEAR_PIXEL_RECT(dst0); |
| 324 | CLEAR_PIXEL_RECT(dst1); | ||
| 325 | call_ref(dst0, dst0_stride, | ||
| 312 | src0, sizes[size] * SIZEOF_PIXEL, | 326 | src0, sizes[size] * SIZEOF_PIXEL, |
| 313 | ref0, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); | 327 | ref0, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); |
| 314 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 328 | call_new(dst1, dst1_stride, |
| 315 | src1, sizes[size] * SIZEOF_PIXEL, | 329 | src1, sizes[size] * SIZEOF_PIXEL, |
| 316 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); | 330 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); |
| 317 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 331 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 318 | dst1, sizes[size] * SIZEOF_PIXEL, | 332 | dst1, dst1_stride, |
| 319 | size[sizes], size[sizes], "dst"); | 333 | size[sizes], size[sizes], "dst"); |
| 320 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 334 | bench_new(dst1, dst1_stride, |
| 321 | src1, sizes[size] * SIZEOF_PIXEL, | 335 | src1, sizes[size] * SIZEOF_PIXEL, |
| 322 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); | 336 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); |
| 323 | } | 337 | } |
| @@ -364,6 +378,7 @@ static void checkasm_check_hevc_epel(void) · 40 unmodified lines | |||
| 364 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { | 378 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { |
| 365 | int16_t *dstw0 = (int16_t *) dst0, *dstw1 = (int16_t *) dst1; | 379 | int16_t *dstw0 = (int16_t *) dst0, *dstw1 = (int16_t *) dst1; |
| 366 | randomize_buffers(); | 380 | randomize_buffers(); |
| 381 | randomize_buffers_dst(); | ||
| 367 | call_ref(dstw0, src0, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); | 382 | call_ref(dstw0, src0, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); |
| 368 | call_new(dstw1, src1, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); | 383 | call_new(dstw1, src1, sizes[size] * SIZEOF_PIXEL, sizes[size], i, j, sizes[size]); |
| 369 | checkasm_check(int16_t, dstw0, MAX_PB_SIZE * sizeof(int16_t), | 384 | checkasm_check(int16_t, dstw0, MAX_PB_SIZE * sizeof(int16_t), |
| @@ -382,8 +397,8 @@ static void checkasm_check_hevc_epel_uni(void) · 12 unmodified lines | |||
| 382 | { | 397 | { |
| 383 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); | 398 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); |
| 384 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); | 399 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); |
| 385 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 400 | PIXEL_RECT(dst0, 64, 64); |
| 386 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 401 | PIXEL_RECT(dst1, 64, 64); |
| 387 | 402 | ||
| 388 | HEVCDSPContext h; | 403 | HEVCDSPContext h; |
| 389 | int size, bit_depth, i, j; | 404 | int size, bit_depth, i, j; |
| @@ -407,16 +422,18 @@ static void checkasm_check_hevc_epel_uni(void) · 17 unmodified lines | |||
| 407 | if (check_func(h.put_hevc_epel_uni[size][j][i], | 422 | if (check_func(h.put_hevc_epel_uni[size][j][i], |
| 408 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { | 423 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { |
| 409 | randomize_buffers(); | 424 | randomize_buffers(); |
| 410 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 425 | CLEAR_PIXEL_RECT(dst0); |
| 426 | CLEAR_PIXEL_RECT(dst1); | ||
| 427 | call_ref(dst0, dst0_stride, | ||
| 411 | src0, sizes[size] * SIZEOF_PIXEL, | 428 | src0, sizes[size] * SIZEOF_PIXEL, |
| 412 | sizes[size], i, j, sizes[size]); | 429 | sizes[size], i, j, sizes[size]); |
| 413 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 430 | call_new(dst1, dst1_stride, |
| 414 | src1, sizes[size] * SIZEOF_PIXEL, | 431 | src1, sizes[size] * SIZEOF_PIXEL, |
| 415 | sizes[size], i, j, sizes[size]); | 432 | sizes[size], i, j, sizes[size]); |
| 416 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 433 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 417 | dst1, sizes[size] * SIZEOF_PIXEL, | 434 | dst1, dst1_stride, |
| 418 | size[sizes], size[sizes], "dst"); | 435 | size[sizes], size[sizes], "dst"); |
| 419 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 436 | bench_new(dst1, dst1_stride, |
| 420 | src1, sizes[size] * SIZEOF_PIXEL, | 437 | src1, sizes[size] * SIZEOF_PIXEL, |
| 421 | sizes[size], i, j, sizes[size]); | 438 | sizes[size], i, j, sizes[size]); |
| 422 | } | 439 | } |
| @@ -431,8 +448,8 @@ static void checkasm_check_hevc_epel_uni_w(void) · 8 unmodified lines | |||
| 431 | { | 448 | { |
| 432 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); | 449 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); |
| 433 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); | 450 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); |
| 434 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 451 | PIXEL_RECT(dst0, 64, 64); |
| 435 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 452 | PIXEL_RECT(dst1, 64, 64); |
| 436 | 453 | ||
| 437 | HEVCDSPContext h; | 454 | HEVCDSPContext h; |
| 438 | int size, bit_depth, i, j; | 455 | int size, bit_depth, i, j; |
| @@ -460,16 +477,18 @@ static void checkasm_check_hevc_epel_uni_w(void) · 21 unmodified lines | |||
| 460 | for (wx = weights; *wx >= 0; wx++) { | 477 | for (wx = weights; *wx >= 0; wx++) { |
| 461 | for (ox = offsets; *ox >= 0; ox++) { | 478 | for (ox = offsets; *ox >= 0; ox++) { |
| 462 | randomize_buffers(); | 479 | randomize_buffers(); |
| 463 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 480 | CLEAR_PIXEL_RECT(dst0); |
| 481 | CLEAR_PIXEL_RECT(dst1); | ||
| 482 | call_ref(dst0, dst0_stride, | ||
| 464 | src0, sizes[size] * SIZEOF_PIXEL, | 483 | src0, sizes[size] * SIZEOF_PIXEL, |
| 465 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); | 484 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); |
| 466 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 485 | call_new(dst1, dst1_stride, |
| 467 | src1, sizes[size] * SIZEOF_PIXEL, | 486 | src1, sizes[size] * SIZEOF_PIXEL, |
| 468 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); | 487 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); |
| 469 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 488 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 470 | dst1, sizes[size] * SIZEOF_PIXEL, | 489 | dst1, dst1_stride, |
| 471 | size[sizes], size[sizes], "dst"); | 490 | size[sizes], size[sizes], "dst"); |
| 472 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 491 | bench_new(dst1, dst1_stride, |
| 473 | src1, sizes[size] * SIZEOF_PIXEL, | 492 | src1, sizes[size] * SIZEOF_PIXEL, |
| 474 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); | 493 | sizes[size], *denom, *wx, *ox, i, j, sizes[size]); |
| 475 | } | 494 | } |
| @@ -487,8 +506,8 @@ static void checkasm_check_hevc_epel_bi(void) · 11 unmodified lines | |||
| 487 | { | 506 | { |
| 488 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); | 507 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); |
| 489 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); | 508 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); |
| 490 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 509 | PIXEL_RECT(dst0, 64, 64); |
| 491 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 510 | PIXEL_RECT(dst1, 64, 64); |
| 492 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); | 511 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); |
| 493 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); | 512 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); |
| 494 | 513 | ||
| @@ -515,16 +534,18 @@ static void checkasm_check_hevc_epel_bi(void) · 20 unmodified lines | |||
| 515 | if (check_func(h.put_hevc_epel_bi[size][j][i], | 534 | if (check_func(h.put_hevc_epel_bi[size][j][i], |
| 516 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { | 535 | "put_hevc_%s%d_%d", type, sizes[size], bit_depth)) { |
| 517 | randomize_buffers_ref(); | 536 | randomize_buffers_ref(); |
| 518 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 537 | CLEAR_PIXEL_RECT(dst0); |
| 538 | CLEAR_PIXEL_RECT(dst1); | ||
| 539 | call_ref(dst0, dst0_stride, | ||
| 519 | src0, sizes[size] * SIZEOF_PIXEL, | 540 | src0, sizes[size] * SIZEOF_PIXEL, |
| 520 | ref0, sizes[size], i, j, sizes[size]); | 541 | ref0, sizes[size], i, j, sizes[size]); |
| 521 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 542 | call_new(dst1, dst1_stride, |
| 522 | src1, sizes[size] * SIZEOF_PIXEL, | 543 | src1, sizes[size] * SIZEOF_PIXEL, |
| 523 | ref1, sizes[size], i, j, sizes[size]); | 544 | ref1, sizes[size], i, j, sizes[size]); |
| 524 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 545 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 525 | dst1, sizes[size] * SIZEOF_PIXEL, | 546 | dst1, dst1_stride, |
| 526 | size[sizes], size[sizes], "dst"); | 547 | size[sizes], size[sizes], "dst"); |
| 527 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 548 | bench_new(dst1, dst1_stride, |
| 528 | src1, sizes[size] * SIZEOF_PIXEL, | 549 | src1, sizes[size] * SIZEOF_PIXEL, |
| 529 | ref1, sizes[size], i, j, sizes[size]); | 550 | ref1, sizes[size], i, j, sizes[size]); |
| 530 | } | 551 | } |
| @@ -539,8 +560,8 @@ static void checkasm_check_hevc_epel_bi_w(void) · 8 unmodified lines | |||
| 539 | { | 560 | { |
| 540 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); | 561 | LOCAL_ALIGNED_32(uint8_t, buf0, [BUF_SIZE]); |
| 541 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); | 562 | LOCAL_ALIGNED_32(uint8_t, buf1, [BUF_SIZE]); |
| 542 | LOCAL_ALIGNED_32(uint8_t, dst0, [BUF_SIZE]); | 563 | PIXEL_RECT(dst0, 64, 64); |
| 543 | LOCAL_ALIGNED_32(uint8_t, dst1, [BUF_SIZE]); | 564 | PIXEL_RECT(dst1, 64, 64); |
| 544 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); | 565 | LOCAL_ALIGNED_32(int16_t, ref0, [BUF_SIZE]); |
| 545 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); | 566 | LOCAL_ALIGNED_32(int16_t, ref1, [BUF_SIZE]); |
| 546 | 567 | ||
| @@ -572,16 +593,18 @@ static void checkasm_check_hevc_epel_bi_w(void) · 25 unmodified lines | |||
| 572 | for (wx = weights; *wx >= 0; wx++) { | 593 | for (wx = weights; *wx >= 0; wx++) { |
| 573 | for (ox = offsets; *ox >= 0; ox++) { | 594 | for (ox = offsets; *ox >= 0; ox++) { |
| 574 | randomize_buffers_ref(); | 595 | randomize_buffers_ref(); |
| 575 | call_ref(dst0, sizes[size] * SIZEOF_PIXEL, | 596 | CLEAR_PIXEL_RECT(dst0); |
| 597 | CLEAR_PIXEL_RECT(dst1); | ||
| 598 | call_ref(dst0, dst0_stride, | ||
| 576 | src0, sizes[size] * SIZEOF_PIXEL, | 599 | src0, sizes[size] * SIZEOF_PIXEL, |
| 577 | ref0, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); | 600 | ref0, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); |
| 578 | call_new(dst1, sizes[size] * SIZEOF_PIXEL, | 601 | call_new(dst1, dst1_stride, |
| 579 | src1, sizes[size] * SIZEOF_PIXEL, | 602 | src1, sizes[size] * SIZEOF_PIXEL, |
| 580 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); | 603 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); |
| 581 | checkasm_check_pixel(dst0, sizes[size] * SIZEOF_PIXEL, | 604 | checkasm_check_pixel_padded(dst0, dst0_stride, |
| 582 | dst1, sizes[size] * SIZEOF_PIXEL, | 605 | dst1, dst1_stride, |
| 583 | size[sizes], size[sizes], "dst"); | 606 | size[sizes], size[sizes], "dst"); |
| 584 | bench_new(dst1, sizes[size] * SIZEOF_PIXEL, | 607 | bench_new(dst1, dst1_stride, |
| 585 | src1, sizes[size] * SIZEOF_PIXEL, | 608 | src1, sizes[size] * SIZEOF_PIXEL, |
| 586 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); | 609 | ref1, sizes[size], *denom, *wx, *wx, *ox, *ox, i, j, sizes[size]); |
| 587 | } | 610 | } |