1/* 2 * Copyright (C) 2017 Rob Clark <robclark@freedesktop.org> 3 * Copyright © 2018 Google, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice (including the next 13 * paragraph) shall be included in all copies or substantial portions of the 14 * Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 22 * SOFTWARE. 23 * 24 * Authors: 25 * Rob Clark <robclark@freedesktop.org> 26 */ 27 28#include "util/u_dump.h" 29 30#include "freedreno_blitter.h" 31#include "freedreno_fence.h" 32#include "freedreno_resource.h" 33 34#include "fd6_blitter.h" 35#include "fd6_format.h" 36#include "fd6_emit.h" 37 38/* Make sure none of the requested dimensions extend beyond the size of the 39 * resource. Not entirely sure why this happens, but sometimes it does, and 40 * w/ 2d blt doesn't have wrap modes like a sampler, so force those cases 41 * back to u_blitter 42 */ 43static bool 44ok_dims(const struct pipe_resource *r, const struct pipe_box *b, int lvl) 45{ 46 int last_layer = 47 r->target == PIPE_TEXTURE_3D ? u_minify(r->depth0, lvl) 48 : r->array_size; 49 50 return (b->x >= 0) && (b->x + b->width <= u_minify(r->width0, lvl)) && 51 (b->y >= 0) && (b->y + b->height <= u_minify(r->height0, lvl)) && 52 (b->z >= 0) && (b->z + b->depth <= last_layer); 53} 54 55static bool 56ok_format(enum pipe_format pfmt) 57{ 58 enum a6xx_color_fmt fmt = fd6_pipe2color(pfmt); 59 60 switch (pfmt) { 61 case PIPE_FORMAT_Z24_UNORM_S8_UINT: 62 case PIPE_FORMAT_Z24X8_UNORM: 63 case PIPE_FORMAT_Z16_UNORM: 64 case PIPE_FORMAT_Z32_UNORM: 65 case PIPE_FORMAT_Z32_FLOAT: 66 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT: 67 case PIPE_FORMAT_S8_UINT: 68 return true; 69 default: 70 break; 71 } 72 73 if (fmt == ~0) 74 return false; 75 76 if (fd6_ifmt(fmt) == 0) 77 return false; 78 79 return true; 80} 81 82#define DEBUG_BLIT_FALLBACK 0 83#define fail_if(cond) \ 84 do { \ 85 if (cond) { \ 86 if (DEBUG_BLIT_FALLBACK) { \ 87 fprintf(stderr, "falling back: %s for blit:\n", #cond); \ 88 util_dump_blit_info(stderr, info); \ 89 fprintf(stderr, "\nsrc: "); \ 90 util_dump_resource(stderr, info->src.resource); \ 91 fprintf(stderr, "\ndst: "); \ 92 util_dump_resource(stderr, info->dst.resource); \ 93 fprintf(stderr, "\n"); \ 94 } \ 95 return false; \ 96 } \ 97 } while (0) 98 99static bool 100can_do_blit(const struct pipe_blit_info *info) 101{ 102 /* I think we can do scaling, but not in z dimension since that would 103 * require blending.. 104 */ 105 fail_if(info->dst.box.depth != info->src.box.depth); 106 107 /* Fail if unsupported format: */ 108 fail_if(!ok_format(info->src.format)); 109 fail_if(!ok_format(info->dst.format)); 110 111 /* We can blit if both or neither formats are compressed formats... */ 112 fail_if(util_format_is_compressed(info->src.format) != 113 util_format_is_compressed(info->src.format)); 114 115 /* ... but only if they're the same compression format. */ 116 fail_if(util_format_is_compressed(info->src.format) && 117 info->src.format != info->dst.format); 118 119 fail_if(!ok_dims(info->src.resource, &info->src.box, info->src.level)); 120 121 fail_if(!ok_dims(info->dst.resource, &info->dst.box, info->dst.level)); 122 123 debug_assert(info->dst.box.width >= 0); 124 debug_assert(info->dst.box.height >= 0); 125 debug_assert(info->dst.box.depth >= 0); 126 127 /* We could probably blit between resources with equal sample count.. */ 128 fail_if(info->dst.resource->nr_samples > 1); 129 130 /* CP_BLIT supports resolving, but seems to pick one only of the samples 131 * (no blending). This doesn't work for RGBA resolves, so we fall back in 132 * that case. However, GL/GLES spec says: 133 * 134 * "If the source formats are integer types or stencil values, a single 135 * sample’s value is selected for each pixel. If the source formats are 136 * floating-point or normalized types, the sample values for each pixel 137 * are resolved in an implementationdependent manner. If the source 138 * formats are depth values, sample values are resolved in an 139 * implementation-dependent manner where the result will be between the 140 * minimum and maximum depth values in the pixel." 141 * 142 * so do those with CP_BLIT. 143 */ 144 fail_if((info->mask & PIPE_MASK_RGBA) && 145 info->src.resource->nr_samples > 1); 146 147 fail_if(info->window_rectangle_include); 148 149 fail_if(util_format_is_srgb(info->src.format)); 150 fail_if(util_format_is_srgb(info->dst.format)); 151 152 const struct util_format_description *src_desc = 153 util_format_description(info->src.format); 154 const struct util_format_description *dst_desc = 155 util_format_description(info->dst.format); 156 const int common_channels = MIN2(src_desc->nr_channels, dst_desc->nr_channels); 157 158 if (info->mask & PIPE_MASK_RGBA) { 159 for (int i = 0; i < common_channels; i++) { 160 fail_if(memcmp(&src_desc->channel[i], 161 &dst_desc->channel[i], 162 sizeof(src_desc->channel[0]))); 163 } 164 } 165 166 fail_if(info->alpha_blend); 167 168 return true; 169} 170 171static void 172emit_setup(struct fd_ringbuffer *ring) 173{ 174 OUT_PKT7(ring, CP_EVENT_WRITE, 1); 175 OUT_RING(ring, PC_CCU_INVALIDATE_COLOR); 176 177 OUT_PKT7(ring, CP_EVENT_WRITE, 1); 178 OUT_RING(ring, LRZ_FLUSH); 179 180 OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1); 181 OUT_RING(ring, 0x0); 182 183 OUT_WFI5(ring); 184 185 OUT_PKT4(ring, REG_A6XX_RB_CCU_CNTL, 1); 186 OUT_RING(ring, 0x10000000); 187} 188 189static uint32_t 190blit_control(enum a6xx_color_fmt fmt) 191{ 192 unsigned blit_cntl = 0xf00000; 193 blit_cntl |= A6XX_RB_2D_BLIT_CNTL_COLOR_FORMAT(fmt); 194 blit_cntl |= A6XX_RB_2D_BLIT_CNTL_IFMT(fd6_ifmt(fmt)); 195 return blit_cntl; 196} 197 198/* buffers need to be handled specially since x/width can exceed the bounds 199 * supported by hw.. if necessary decompose into (potentially) two 2D blits 200 */ 201static void 202emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info) 203{ 204 const struct pipe_box *sbox = &info->src.box; 205 const struct pipe_box *dbox = &info->dst.box; 206 struct fd_resource *src, *dst; 207 unsigned sshift, dshift; 208 209 if (DEBUG_BLIT_FALLBACK) { 210 fprintf(stderr, "buffer blit: "); 211 util_dump_blit_info(stderr, info); 212 fprintf(stderr, "\ndst resource: "); 213 util_dump_resource(stderr, info->dst.resource); 214 fprintf(stderr, "\nsrc resource: "); 215 util_dump_resource(stderr, info->src.resource); 216 fprintf(stderr, "\n"); 217 } 218 219 src = fd_resource(info->src.resource); 220 dst = fd_resource(info->dst.resource); 221 222 debug_assert(src->cpp == 1); 223 debug_assert(dst->cpp == 1); 224 debug_assert(info->src.resource->format == info->dst.resource->format); 225 debug_assert((sbox->y == 0) && (sbox->height == 1)); 226 debug_assert((dbox->y == 0) && (dbox->height == 1)); 227 debug_assert((sbox->z == 0) && (sbox->depth == 1)); 228 debug_assert((dbox->z == 0) && (dbox->depth == 1)); 229 debug_assert(sbox->width == dbox->width); 230 debug_assert(info->src.level == 0); 231 debug_assert(info->dst.level == 0); 232 233 /* 234 * Buffers can have dimensions bigger than max width, remap into 235 * multiple 1d blits to fit within max dimension 236 * 237 * Note that blob uses .ARRAY_PITCH=128 for blitting buffers, which 238 * seems to prevent overfetch related faults. Not quite sure what 239 * the deal is there. 240 * 241 * Low 6 bits of SRC/DST addresses need to be zero (ie. address 242 * aligned to 64) so we need to shift src/dst x1/x2 to make up the 243 * difference. On top of already splitting up the blit so width 244 * isn't > 16k. 245 * 246 * We perhaps could do a bit better, if src and dst are aligned but 247 * in the worst case this means we have to split the copy up into 248 * 16k (0x4000) minus 64 (0x40). 249 */ 250 251 sshift = sbox->x & 0x3f; 252 dshift = dbox->x & 0x3f; 253 254 OUT_PKT7(ring, CP_SET_MARKER, 1); 255 OUT_RING(ring, A2XX_CP_SET_MARKER_0_MODE(RM6_BLIT2DSCALE)); 256 257 uint32_t blit_cntl = blit_control(RB6_R8_UNORM) | 0x20000000; 258 OUT_PKT4(ring, REG_A6XX_RB_2D_BLIT_CNTL, 1); 259 OUT_RING(ring, blit_cntl); 260 261 OUT_PKT4(ring, REG_A6XX_GRAS_2D_BLIT_CNTL, 1); 262 OUT_RING(ring, blit_cntl); 263 264 for (unsigned off = 0; off < sbox->width; off += (0x4000 - 0x40)) { 265 unsigned soff, doff, w, p; 266 267 soff = (sbox->x + off) & ~0x3f; 268 doff = (dbox->x + off) & ~0x3f; 269 270 w = MIN2(sbox->width - off, (0x4000 - 0x40)); 271 p = align(w, 64); 272 273 debug_assert((soff + w) <= fd_bo_size(src->bo)); 274 debug_assert((doff + w) <= fd_bo_size(dst->bo)); 275 276 /* 277 * Emit source: 278 */ 279 OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 10); 280 OUT_RING(ring, A6XX_SP_PS_2D_SRC_INFO_COLOR_FORMAT(RB6_R8_UNORM) | 281 A6XX_SP_PS_2D_SRC_INFO_TILE_MODE(TILE6_LINEAR) | 282 A6XX_SP_PS_2D_SRC_INFO_COLOR_SWAP(WZYX) | 283 0x500000); 284 OUT_RING(ring, A6XX_SP_PS_2D_SRC_SIZE_WIDTH(sshift + w) | 285 A6XX_SP_PS_2D_SRC_SIZE_HEIGHT(1)); /* SP_PS_2D_SRC_SIZE */ 286 OUT_RELOC(ring, src->bo, soff, 0, 0); /* SP_PS_2D_SRC_LO/HI */ 287 OUT_RING(ring, A6XX_SP_PS_2D_SRC_PITCH_PITCH(p)); 288 289 OUT_RING(ring, 0x00000000); 290 OUT_RING(ring, 0x00000000); 291 OUT_RING(ring, 0x00000000); 292 OUT_RING(ring, 0x00000000); 293 OUT_RING(ring, 0x00000000); 294 295 /* 296 * Emit destination: 297 */ 298 OUT_PKT4(ring, REG_A6XX_RB_2D_DST_INFO, 9); 299 OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(RB6_R8_UNORM) | 300 A6XX_RB_2D_DST_INFO_TILE_MODE(TILE6_LINEAR) | 301 A6XX_RB_2D_DST_INFO_COLOR_SWAP(WZYX)); 302 OUT_RELOCW(ring, dst->bo, doff, 0, 0); /* RB_2D_DST_LO/HI */ 303 OUT_RING(ring, A6XX_RB_2D_DST_SIZE_PITCH(p)); 304 OUT_RING(ring, 0x00000000); 305 OUT_RING(ring, 0x00000000); 306 OUT_RING(ring, 0x00000000); 307 OUT_RING(ring, 0x00000000); 308 OUT_RING(ring, 0x00000000); 309 310 /* 311 * Blit command: 312 */ 313 OUT_PKT4(ring, REG_A6XX_GRAS_2D_SRC_TL_X, 4); 314 OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_X_X(sshift)); 315 OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_X_X(sshift + w - 1)); 316 OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_Y_Y(0)); 317 OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_Y_Y(0)); 318 319 OUT_PKT4(ring, REG_A6XX_GRAS_2D_DST_TL, 2); 320 OUT_RING(ring, A6XX_GRAS_2D_DST_TL_X(dshift) | A6XX_GRAS_2D_DST_TL_Y(0)); 321 OUT_RING(ring, A6XX_GRAS_2D_DST_BR_X(dshift + w - 1) | A6XX_GRAS_2D_DST_BR_Y(0)); 322 323 OUT_PKT7(ring, CP_EVENT_WRITE, 1); 324 OUT_RING(ring, 0x3f); 325 OUT_WFI5(ring); 326 327 OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8C01, 1); 328 OUT_RING(ring, 0); 329 330 OUT_PKT4(ring, REG_A6XX_SP_2D_SRC_FORMAT, 1); 331 OUT_RING(ring, 0xf180); 332 333 OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1); 334 OUT_RING(ring, 0x01000000); 335 336 OUT_PKT7(ring, CP_BLIT, 1); 337 OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_SCALE)); 338 339 OUT_WFI5(ring); 340 341 OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1); 342 OUT_RING(ring, 0); 343 } 344} 345 346static void 347emit_blit_texture(struct fd_ringbuffer *ring, const struct pipe_blit_info *info) 348{ 349 const struct pipe_box *sbox = &info->src.box; 350 const struct pipe_box *dbox = &info->dst.box; 351 struct fd_resource *src, *dst; 352 struct fd_resource_slice *sslice, *dslice; 353 enum a6xx_color_fmt sfmt, dfmt; 354 enum a6xx_tile_mode stile, dtile; 355 enum a3xx_color_swap sswap, dswap; 356 unsigned spitch, dpitch; 357 int sx1, sy1, sx2, sy2; 358 int dx1, dy1, dx2, dy2; 359 360 if (DEBUG_BLIT_FALLBACK) { 361 fprintf(stderr, "texture blit: "); 362 util_dump_blit_info(stderr, info); 363 fprintf(stderr, "\ndst resource: "); 364 util_dump_resource(stderr, info->dst.resource); 365 fprintf(stderr, "\nsrc resource: "); 366 util_dump_resource(stderr, info->src.resource); 367 fprintf(stderr, "\n"); 368 } 369 370 src = fd_resource(info->src.resource); 371 dst = fd_resource(info->dst.resource); 372 373 sslice = fd_resource_slice(src, info->src.level); 374 dslice = fd_resource_slice(dst, info->dst.level); 375 376 sfmt = fd6_pipe2color(info->src.format); 377 dfmt = fd6_pipe2color(info->dst.format); 378 379 int blocksize = util_format_get_blocksize(info->src.format); 380 int blockwidth = util_format_get_blockwidth(info->src.format); 381 int blockheight = util_format_get_blockheight(info->src.format); 382 int nelements; 383 384 stile = fd_resource_level_linear(info->src.resource, info->src.level) ? 385 TILE6_LINEAR : src->tile_mode; 386 dtile = fd_resource_level_linear(info->dst.resource, info->dst.level) ? 387 TILE6_LINEAR : dst->tile_mode; 388 389 sswap = stile ? WZYX : fd6_pipe2swap(info->src.format); 390 dswap = dtile ? WZYX : fd6_pipe2swap(info->dst.format); 391 392 if (util_format_is_compressed(info->src.format)) { 393 debug_assert(info->src.format == info->dst.format); 394 sfmt = dfmt = RB6_R8_UNORM; 395 nelements = blocksize; 396 } else { 397 debug_assert(!util_format_is_compressed(info->dst.format)); 398 nelements = 1; 399 } 400 401 spitch = DIV_ROUND_UP(sslice->pitch, blockwidth) * src->cpp; 402 dpitch = DIV_ROUND_UP(dslice->pitch, blockwidth) * dst->cpp; 403 404 sx1 = sbox->x / blockwidth * nelements; 405 sy1 = sbox->y / blockheight; 406 sx2 = DIV_ROUND_UP(sbox->x + sbox->width, blockwidth) * nelements - 1; 407 sy2 = DIV_ROUND_UP(sbox->y + sbox->height, blockheight) - 1; 408 409 dx1 = dbox->x / blockwidth * nelements; 410 dy1 = dbox->y / blockheight; 411 dx2 = DIV_ROUND_UP(dbox->x + dbox->width, blockwidth) * nelements - 1; 412 dy2 = DIV_ROUND_UP(dbox->y + dbox->height, blockheight) - 1; 413 414 uint32_t width = DIV_ROUND_UP(u_minify(src->base.width0, info->src.level), blockwidth) * nelements; 415 uint32_t height = DIV_ROUND_UP(u_minify(src->base.height0, info->src.level), blockheight); 416 417 OUT_PKT7(ring, CP_SET_MARKER, 1); 418 OUT_RING(ring, A2XX_CP_SET_MARKER_0_MODE(RM6_BLIT2DSCALE)); 419 420 uint32_t blit_cntl = blit_control(dfmt); 421 422 if (dtile != stile) 423 blit_cntl |= 0x20000000; 424 425 if (info->scissor_enable) { 426 OUT_PKT4(ring, REG_A6XX_GRAS_RESOLVE_CNTL_1, 2); 427 OUT_RING(ring, A6XX_GRAS_RESOLVE_CNTL_1_X(info->scissor.minx) | 428 A6XX_GRAS_RESOLVE_CNTL_1_Y(info->scissor.miny)); 429 OUT_RING(ring, A6XX_GRAS_RESOLVE_CNTL_1_X(info->scissor.maxx - 1) | 430 A6XX_GRAS_RESOLVE_CNTL_1_Y(info->scissor.maxy - 1)); 431 blit_cntl |= A6XX_RB_2D_BLIT_CNTL_SCISSOR; 432 } 433 434 OUT_PKT4(ring, REG_A6XX_RB_2D_BLIT_CNTL, 1); 435 OUT_RING(ring, blit_cntl); 436 437 OUT_PKT4(ring, REG_A6XX_GRAS_2D_BLIT_CNTL, 1); 438 OUT_RING(ring, blit_cntl); 439 440 for (unsigned i = 0; i < info->dst.box.depth; i++) { 441 unsigned soff = fd_resource_offset(src, info->src.level, sbox->z + i); 442 unsigned doff = fd_resource_offset(dst, info->dst.level, dbox->z + i); 443 unsigned subwcoff = fd_resource_ubwc_offset(src, info->src.level, sbox->z + i); 444 unsigned dubwcoff = fd_resource_ubwc_offset(dst, info->dst.level, dbox->z + i); 445 bool subwc_enabled = fd_resource_ubwc_enabled(src, info->src.level); 446 bool dubwc_enabled = fd_resource_ubwc_enabled(dst, info->dst.level); 447 448 /* 449 * Emit source: 450 */ 451 uint32_t filter = 0; 452 if (info->filter == PIPE_TEX_FILTER_LINEAR) 453 filter = A6XX_SP_PS_2D_SRC_INFO_FILTER; 454 455 enum a3xx_msaa_samples samples = fd_msaa_samples(src->base.nr_samples); 456 457 OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 10); 458 OUT_RING(ring, A6XX_SP_PS_2D_SRC_INFO_COLOR_FORMAT(sfmt) | 459 A6XX_SP_PS_2D_SRC_INFO_TILE_MODE(stile) | 460 A6XX_SP_PS_2D_SRC_INFO_COLOR_SWAP(sswap) | 461 A6XX_SP_PS_2D_SRC_INFO_SAMPLES(samples) | 462 COND(subwc_enabled, A6XX_SP_PS_2D_SRC_INFO_FLAGS) | 463 0x500000 | filter); 464 OUT_RING(ring, A6XX_SP_PS_2D_SRC_SIZE_WIDTH(width) | 465 A6XX_SP_PS_2D_SRC_SIZE_HEIGHT(height)); /* SP_PS_2D_SRC_SIZE */ 466 OUT_RELOC(ring, src->bo, soff, 0, 0); /* SP_PS_2D_SRC_LO/HI */ 467 OUT_RING(ring, A6XX_SP_PS_2D_SRC_PITCH_PITCH(spitch)); 468 469 OUT_RING(ring, 0x00000000); 470 OUT_RING(ring, 0x00000000); 471 OUT_RING(ring, 0x00000000); 472 OUT_RING(ring, 0x00000000); 473 OUT_RING(ring, 0x00000000); 474 475 if (subwc_enabled) { 476 OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_FLAGS_LO, 6); 477 OUT_RELOC(ring, src->bo, subwcoff, 0, 0); 478 OUT_RING(ring, A6XX_RB_MRT_FLAG_BUFFER_PITCH_PITCH(src->ubwc_pitch) | 479 A6XX_RB_MRT_FLAG_BUFFER_PITCH_ARRAY_PITCH(src->ubwc_size)); 480 OUT_RING(ring, 0x00000000); 481 OUT_RING(ring, 0x00000000); 482 OUT_RING(ring, 0x00000000); 483 } 484 485 /* 486 * Emit destination: 487 */ 488 OUT_PKT4(ring, REG_A6XX_RB_2D_DST_INFO, 9); 489 OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(dfmt) | 490 A6XX_RB_2D_DST_INFO_TILE_MODE(dtile) | 491 A6XX_RB_2D_DST_INFO_COLOR_SWAP(dswap) | 492 COND(dubwc_enabled, A6XX_RB_2D_DST_INFO_FLAGS)); 493 OUT_RELOCW(ring, dst->bo, doff, 0, 0); /* RB_2D_DST_LO/HI */ 494 OUT_RING(ring, A6XX_RB_2D_DST_SIZE_PITCH(dpitch)); 495 OUT_RING(ring, 0x00000000); 496 OUT_RING(ring, 0x00000000); 497 OUT_RING(ring, 0x00000000); 498 OUT_RING(ring, 0x00000000); 499 OUT_RING(ring, 0x00000000); 500 501 if (dubwc_enabled) { 502 OUT_PKT4(ring, REG_A6XX_RB_2D_DST_FLAGS_LO, 6); 503 OUT_RELOCW(ring, dst->bo, dubwcoff, 0, 0); 504 OUT_RING(ring, A6XX_RB_MRT_FLAG_BUFFER_PITCH_PITCH(dst->ubwc_pitch) | 505 A6XX_RB_MRT_FLAG_BUFFER_PITCH_ARRAY_PITCH(dst->ubwc_size)); 506 OUT_RING(ring, 0x00000000); 507 OUT_RING(ring, 0x00000000); 508 OUT_RING(ring, 0x00000000); 509 } 510 /* 511 * Blit command: 512 */ 513 OUT_PKT4(ring, REG_A6XX_GRAS_2D_SRC_TL_X, 4); 514 OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_X_X(sx1)); 515 OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_X_X(sx2)); 516 OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_Y_Y(sy1)); 517 OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_Y_Y(sy2)); 518 519 OUT_PKT4(ring, REG_A6XX_GRAS_2D_DST_TL, 2); 520 OUT_RING(ring, A6XX_GRAS_2D_DST_TL_X(dx1) | A6XX_GRAS_2D_DST_TL_Y(dy1)); 521 OUT_RING(ring, A6XX_GRAS_2D_DST_BR_X(dx2) | A6XX_GRAS_2D_DST_BR_Y(dy2)); 522 523 OUT_PKT7(ring, CP_EVENT_WRITE, 1); 524 OUT_RING(ring, 0x3f); 525 OUT_WFI5(ring); 526 527 OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8C01, 1); 528 OUT_RING(ring, 0); 529 530 OUT_PKT4(ring, REG_A6XX_SP_2D_SRC_FORMAT, 1); 531 OUT_RING(ring, A6XX_SP_2D_SRC_FORMAT_COLOR_FORMAT(sfmt) | 532 COND(util_format_is_pure_sint(info->src.format), 533 A6XX_SP_2D_SRC_FORMAT_SINT) | 534 COND(util_format_is_pure_uint(info->src.format), 535 A6XX_SP_2D_SRC_FORMAT_UINT) | 536 COND(util_format_is_snorm(info->src.format), 537 A6XX_SP_2D_SRC_FORMAT_SINT | 538 A6XX_SP_2D_SRC_FORMAT_NORM) | 539 COND(util_format_is_unorm(info->src.format), 540// TODO sometimes blob uses UINT+NORM but dEQP seems unhappy about that 541// A6XX_SP_2D_SRC_FORMAT_UINT | 542 A6XX_SP_2D_SRC_FORMAT_NORM) | 543 0xf000); 544 545 OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1); 546 OUT_RING(ring, 0x01000000); 547 548 OUT_PKT7(ring, CP_BLIT, 1); 549 OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_SCALE)); 550 551 OUT_WFI5(ring); 552 553 OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1); 554 OUT_RING(ring, 0); 555 } 556} 557 558static void 559rewrite_zs_blit(struct fd_ringbuffer *ring, const struct pipe_blit_info *info) 560{ 561 struct pipe_blit_info separate = *info; 562 563 if (DEBUG_BLIT_FALLBACK) { 564 fprintf(stderr, "---- rewrite_separate_zs_blit: "); 565 util_dump_blit_info(stderr, info); 566 fprintf(stderr, "\ndst resource: "); 567 util_dump_resource(stderr, info->dst.resource); 568 fprintf(stderr, "\nsrc resource: "); 569 util_dump_resource(stderr, info->src.resource); 570 fprintf(stderr, "\n\n"); 571 } 572 573 switch (info->src.format) { 574 case PIPE_FORMAT_S8_UINT: 575 debug_assert(info->mask == PIPE_MASK_S); 576 separate.mask = PIPE_MASK_R; 577 separate.src.format = PIPE_FORMAT_R8_UINT; 578 separate.dst.format = PIPE_FORMAT_R8_UINT; 579 emit_blit_texture(ring, &separate); 580 break; 581 582 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT: 583 if (info->mask & PIPE_MASK_Z) { 584 separate.mask = PIPE_MASK_R; 585 separate.src.format = PIPE_FORMAT_R32_FLOAT; 586 separate.dst.format = PIPE_FORMAT_R32_FLOAT; 587 emit_blit_texture(ring, &separate); 588 } 589 if (info->mask & PIPE_MASK_S) { 590 separate.mask = PIPE_MASK_R; 591 separate.src.format = PIPE_FORMAT_R8_UINT; 592 separate.dst.format = PIPE_FORMAT_R8_UINT; 593 separate.src.resource = &fd_resource(info->src.resource)->stencil->base; 594 separate.dst.resource = &fd_resource(info->dst.resource)->stencil->base; 595 emit_blit_texture(ring, &separate); 596 } 597 break; 598 599 case PIPE_FORMAT_Z16_UNORM: 600 separate.mask = PIPE_MASK_R; 601 separate.src.format = PIPE_FORMAT_R16_UNORM; 602 separate.dst.format = PIPE_FORMAT_R16_UNORM; 603 emit_blit_texture(ring, &separate); 604 break; 605 606 case PIPE_FORMAT_Z32_UNORM: 607 case PIPE_FORMAT_Z32_FLOAT: 608 debug_assert(info->mask == PIPE_MASK_Z); 609 separate.mask = PIPE_MASK_R; 610 separate.src.format = PIPE_FORMAT_R32_UINT; 611 separate.dst.format = PIPE_FORMAT_R32_UINT; 612 emit_blit_texture(ring, &separate); 613 break; 614 615 case PIPE_FORMAT_Z24_UNORM_S8_UINT: 616 debug_assert(info->mask == PIPE_MASK_ZS); 617 case PIPE_FORMAT_Z24X8_UNORM: 618 case PIPE_FORMAT_X8Z24_UNORM: 619 separate.mask = PIPE_MASK_R; 620 separate.src.format = PIPE_FORMAT_R32_UINT; 621 separate.dst.format = PIPE_FORMAT_R32_UINT; 622 emit_blit_texture(ring, &separate); 623 break; 624 625 default: 626 unreachable(""); 627 } 628} 629 630static void 631rewrite_combined_zs_blit(struct fd_context *ctx, const struct pipe_blit_info *info) 632{ 633 struct pipe_blit_info separate = *info; 634 635 if (DEBUG_BLIT_FALLBACK) { 636 fprintf(stderr, "---- rewrite_combined_zs_blit: "); 637 util_dump_blit_info(stderr, info); 638 fprintf(stderr, "\ndst resource: "); 639 util_dump_resource(stderr, info->dst.resource); 640 fprintf(stderr, "\nsrc resource: "); 641 util_dump_resource(stderr, info->src.resource); 642 fprintf(stderr, "\n"); 643 } 644 645 switch (info->mask) { 646 case PIPE_MASK_Z: 647 separate.mask = PIPE_MASK_R | PIPE_MASK_G | PIPE_MASK_B; 648 separate.src.format = PIPE_FORMAT_R8G8B8A8_UNORM; 649 separate.dst.format = PIPE_FORMAT_R8G8B8A8_UNORM; 650 651 fd_blitter_blit(ctx, &separate); 652 break; 653 654 case PIPE_MASK_S: 655 separate.mask = PIPE_MASK_A; 656 separate.src.format = PIPE_FORMAT_R8G8B8A8_UNORM; 657 separate.dst.format = PIPE_FORMAT_R8G8B8A8_UNORM; 658 659 fd_blitter_blit(ctx, &separate); 660 break; 661 662 default: 663 unreachable(""); 664 } 665} 666 667static bool 668fd6_blit(struct fd_context *ctx, const struct pipe_blit_info *info) 669{ 670 struct fd_batch *batch; 671 672 if (info->dst.format == PIPE_FORMAT_Z24_UNORM_S8_UINT && 673 info->mask != PIPE_MASK_ZS) { 674 rewrite_combined_zs_blit(ctx, info); 675 return true; 676 } 677 678 if (!can_do_blit(info)) 679 return false; 680 681 fd_fence_ref(ctx->base.screen, &ctx->last_fence, NULL); 682 683 batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true); 684 685 fd6_emit_restore(batch, batch->draw); 686 fd6_emit_lrz_flush(batch->draw); 687 688 mtx_lock(&ctx->screen->lock); 689 690 fd_batch_resource_used(batch, fd_resource(info->src.resource), false); 691 fd_batch_resource_used(batch, fd_resource(info->dst.resource), true); 692 693 mtx_unlock(&ctx->screen->lock); 694 695 emit_setup(batch->draw); 696 697 if ((info->src.resource->target == PIPE_BUFFER) && 698 (info->dst.resource->target == PIPE_BUFFER)) { 699 assert(fd_resource(info->src.resource)->tile_mode == TILE6_LINEAR); 700 assert(fd_resource(info->dst.resource)->tile_mode == TILE6_LINEAR); 701 emit_blit_buffer(batch->draw, info); 702 } else { 703 /* I don't *think* we need to handle blits between buffer <-> !buffer */ 704 debug_assert(info->src.resource->target != PIPE_BUFFER); 705 debug_assert(info->dst.resource->target != PIPE_BUFFER); 706 707 if (info->mask & (PIPE_MASK_ZS)) { 708 rewrite_zs_blit(batch->draw, info); 709 } else { 710 emit_blit_texture(batch->draw, info); 711 } 712 } 713 714 fd6_event_write(batch, batch->draw, 0x1d, true); 715 fd6_event_write(batch, batch->draw, FACENESS_FLUSH, true); 716 fd6_event_write(batch, batch->draw, CACHE_FLUSH_TS, true); 717 718 fd_resource(info->dst.resource)->valid = true; 719 batch->needs_flush = true; 720 721 fd_batch_flush(batch, false, false); 722 fd_batch_reference(&batch, NULL); 723 724 return true; 725} 726 727void 728fd6_blitter_init(struct pipe_context *pctx) 729{ 730 if (fd_mesa_debug & FD_DBG_NOBLIT) 731 return; 732 733 fd_context(pctx)->blit = fd6_blit; 734} 735 736unsigned 737fd6_tile_mode(const struct pipe_resource *tmpl) 738{ 739 /* basically just has to be a format we can blit, so uploads/downloads 740 * via linear staging buffer works: 741 */ 742 if (ok_format(tmpl->format)) 743 return TILE6_3; 744 745 return TILE6_LINEAR; 746} 747