src/gallium/drivers/freedreno/freedreno_resource.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "util/u_format.h"
  30 #include "util/u_inlines.h"
  31 #include "util/u_transfer.h"
  32 #include "util/u_string.h"
  33 #include "util/u_surface.h"
  34
  35 #include "freedreno_resource.h"
  36 #include "freedreno_screen.h"
  37 #include "freedreno_surface.h"
  38 #include "freedreno_context.h"
  39 #include "freedreno_query_hw.h"
  40 #include "freedreno_util.h"
  41
  42 #include <errno.h>
  43
  44 static void
  45 realloc_bo(struct fd_resource *rsc, uint32_t size)
  46 {
  47         struct fd_screen *screen = fd_screen(rsc->base.b.screen);
  48         uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
  49                         DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */
  50
  51         /* if we start using things other than write-combine,
  52          * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT
  53          */
  54
  55         if (rsc->bo)
  56                 fd_bo_del(rsc->bo);
  57
  58         rsc->bo = fd_bo_new(screen->dev, size, flags);
  59         rsc->timestamp = 0;
  60         rsc->dirty = rsc->reading = false;
  61         list_delinit(&rsc->list);
  62 }
  63
  64 static void fd_resource_transfer_flush_region(struct pipe_context *pctx,
  65                 struct pipe_transfer *ptrans,
  66                 const struct pipe_box *box)
  67 {
  68 }
  69
  70 static void
  71 fd_resource_transfer_unmap(struct pipe_context *pctx,
  72                 struct pipe_transfer *ptrans)
  73 {
  74         struct fd_context *ctx = fd_context(pctx);
  75         struct fd_resource *rsc = fd_resource(ptrans->resource);
  76         if (!(ptrans->usage & PIPE_TRANSFER_UNSYNCHRONIZED))
  77                 fd_bo_cpu_fini(rsc->bo);
  78         pipe_resource_reference(&ptrans->resource, NULL);
  79         util_slab_free(&ctx->transfer_pool, ptrans);
  80 }
  81
  82 static void *
  83 fd_resource_transfer_map(struct pipe_context *pctx,
  84                 struct pipe_resource *prsc,
  85                 unsigned level, unsigned usage,
  86                 const struct pipe_box *box,
  87                 struct pipe_transfer **pptrans)
  88 {
  89         struct fd_context *ctx = fd_context(pctx);
  90         struct fd_resource *rsc = fd_resource(prsc);
  91         struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
  92         struct pipe_transfer *ptrans;
  93         enum pipe_format format = prsc->format;
  94         uint32_t op = 0;
  95         uint32_t offset;
  96         char *buf;
  97         int ret = 0;
  98
  99         DBG("prsc=%p, level=%u, usage=%x", prsc, level, usage);
 100
 101         ptrans = util_slab_alloc(&ctx->transfer_pool);
 102         if (!ptrans)
 103                 return NULL;
 104
 105         /* util_slab_alloc() doesn't zero: */
 106         memset(ptrans, 0, sizeof(*ptrans));
 107
 108         pipe_resource_reference(&ptrans->resource, prsc);
 109         ptrans->level = level;
 110         ptrans->usage = usage;
 111         ptrans->box = *box;
 112         ptrans->stride = slice->pitch * rsc->cpp;
 113         ptrans->layer_stride = slice->size0;
 114
 115         if (usage & PIPE_TRANSFER_READ)
 116                 op |= DRM_FREEDRENO_PREP_READ;
 117
 118         if (usage & PIPE_TRANSFER_WRITE)
 119                 op |= DRM_FREEDRENO_PREP_WRITE;
 120
 121         if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 122                 realloc_bo(rsc, fd_bo_size(rsc->bo));
 123         } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
 124                 /* If the GPU is writing to the resource, or if it is reading from the
 125                  * resource and we're trying to write to it, flush the renders.
 126                  */
 127                 if (rsc->dirty ||
 128                         ((ptrans->usage & PIPE_TRANSFER_WRITE) && rsc->reading))
 129                         fd_context_render(pctx);
 130
 131                 /* The GPU keeps track of how the various bo's are being used, and
 132                  * will wait if necessary for the proper operation to have
 133                  * completed.
 134                  */
 135                 ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
 136                 if (ret)
 137                         goto fail;
 138         }
 139
 140         buf = fd_bo_map(rsc->bo);
 141         if (!buf) {
 142                 fd_resource_transfer_unmap(pctx, ptrans);
 143                 return NULL;
 144         }
 145
 146         *pptrans = ptrans;
 147
 148         if (rsc->layer_first) {
 149                 offset = slice->offset +
 150                         box->y / util_format_get_blockheight(format) * ptrans->stride +
 151                         box->x / util_format_get_blockwidth(format) * rsc->cpp +
 152                         box->z * rsc->layer_size;
 153         } else {
 154                 offset = slice->offset +
 155                         box->y / util_format_get_blockheight(format) * ptrans->stride +
 156                         box->x / util_format_get_blockwidth(format) * rsc->cpp +
 157                         box->z * slice->size0;
 158         }
 159
 160         return buf + offset;
 161
 162 fail:
 163         fd_resource_transfer_unmap(pctx, ptrans);
 164         return NULL;
 165 }
 166
 167 static void
 168 fd_resource_destroy(struct pipe_screen *pscreen,
 169                 struct pipe_resource *prsc)
 170 {
 171         struct fd_resource *rsc = fd_resource(prsc);
 172         if (rsc->bo)
 173                 fd_bo_del(rsc->bo);
 174         list_delinit(&rsc->list);
 175         FREE(rsc);
 176 }
 177
 178 static boolean
 179 fd_resource_get_handle(struct pipe_screen *pscreen,
 180                 struct pipe_resource *prsc,
 181                 struct winsys_handle *handle)
 182 {
 183         struct fd_resource *rsc = fd_resource(prsc);
 184
 185         return fd_screen_bo_get_handle(pscreen, rsc->bo,
 186                         rsc->slices[0].pitch * rsc->cpp, handle);
 187 }
 188
 189
 190 static const struct u_resource_vtbl fd_resource_vtbl = {
 191                 .resource_get_handle      = fd_resource_get_handle,
 192                 .resource_destroy         = fd_resource_destroy,
 193                 .transfer_map             = fd_resource_transfer_map,
 194                 .transfer_flush_region    = fd_resource_transfer_flush_region,
 195                 .transfer_unmap           = fd_resource_transfer_unmap,
 196                 .transfer_inline_write    = u_default_transfer_inline_write,
 197 };
 198
 199 static uint32_t
 200 setup_slices(struct fd_resource *rsc, uint32_t alignment)
 201 {
 202         struct pipe_resource *prsc = &rsc->base.b;
 203         uint32_t level, size = 0;
 204         uint32_t width = prsc->width0;
 205         uint32_t height = prsc->height0;
 206         uint32_t depth = prsc->depth0;
 207         /* in layer_first layout, the level (slice) contains just one
 208          * layer (since in fact the layer contains the slices)
 209          */
 210         uint32_t layers_in_level = rsc->layer_first ? 1 : prsc->array_size;
 211
 212         for (level = 0; level <= prsc->last_level; level++) {
 213                 struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
 214
 215                 slice->pitch = width = align(width, 32);
 216                 slice->offset = size;
 217                 /* 1d array and 2d array textures must all have the same layer size
 218                  * for each miplevel on a3xx. 3d textures can have different layer
 219                  * sizes for high levels, but the hw auto-sizer is buggy (or at least
 220                  * different than what this code does), so as soon as the layer size
 221                  * range gets into range, we stop reducing it.
 222                  */
 223                 if (prsc->target == PIPE_TEXTURE_3D && (
 224                                         level == 1 ||
 225                                         (level > 1 && rsc->slices[level - 1].size0 > 0xf000)))
 226                         slice->size0 = align(slice->pitch * height * rsc->cpp, alignment);
 227                 else if (level == 0 || rsc->layer_first || alignment == 1)
 228                         slice->size0 = align(slice->pitch * height * rsc->cpp, alignment);
 229                 else
 230                         slice->size0 = rsc->slices[level - 1].size0;
 231
 232                 size += slice->size0 * depth * layers_in_level;
 233
 234                 width = u_minify(width, 1);
 235                 height = u_minify(height, 1);
 236                 depth = u_minify(depth, 1);
 237         }
 238
 239         return size;
 240 }
 241
 242 static uint32_t
 243 slice_alignment(struct pipe_screen *pscreen, const struct pipe_resource *tmpl)
 244 {
 245         /* on a3xx, 2d array and 3d textures seem to want their
 246          * layers aligned to page boundaries:
 247          */
 248         switch (tmpl->target) {
 249         case PIPE_TEXTURE_3D:
 250         case PIPE_TEXTURE_1D_ARRAY:
 251         case PIPE_TEXTURE_2D_ARRAY:
 252                 return 4096;
 253         default:
 254                 return 1;
 255         }
 256 }
 257
 258 /**
 259  * Create a new texture object, using the given template info.
 260  */
 261 static struct pipe_resource *
 262 fd_resource_create(struct pipe_screen *pscreen,
 263                 const struct pipe_resource *tmpl)
 264 {
 265         struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
 266         struct pipe_resource *prsc = &rsc->base.b;
 267         uint32_t size;
 268
 269         DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
 270                         "nr_samples=%u, usage=%u, bind=%x, flags=%x",
 271                         tmpl->target, util_format_name(tmpl->format),
 272                         tmpl->width0, tmpl->height0, tmpl->depth0,
 273                         tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
 274                         tmpl->usage, tmpl->bind, tmpl->flags);
 275
 276         if (!rsc)
 277                 return NULL;
 278
 279         *prsc = *tmpl;
 280
 281         pipe_reference_init(&prsc->reference, 1);
 282         list_inithead(&rsc->list);
 283         prsc->screen = pscreen;
 284
 285         rsc->base.vtbl = &fd_resource_vtbl;
 286         rsc->cpp = util_format_get_blocksize(tmpl->format);
 287
 288         assert(rsc->cpp);
 289
 290         if (is_a4xx(fd_screen(pscreen))) {
 291                 switch (tmpl->target) {
 292                 case PIPE_TEXTURE_3D:
 293                         /* TODO 3D_ARRAY? */
 294                         rsc->layer_first = false;
 295                         break;
 296                 default:
 297                         rsc->layer_first = true;
 298                         break;
 299                 }
 300         }
 301
 302         size = setup_slices(rsc, slice_alignment(pscreen, tmpl));
 303
 304         if (rsc->layer_first) {
 305                 rsc->layer_size = align(size, 4096);
 306                 size = rsc->layer_size * prsc->array_size;
 307         }
 308
 309         realloc_bo(rsc, size);
 310         if (!rsc->bo)
 311                 goto fail;
 312
 313         return prsc;
 314 fail:
 315         fd_resource_destroy(pscreen, prsc);
 316         return NULL;
 317 }
 318
 319 /**
 320  * Create a texture from a winsys_handle. The handle is often created in
 321  * another process by first creating a pipe texture and then calling
 322  * resource_get_handle.
 323  */
 324 static struct pipe_resource *
 325 fd_resource_from_handle(struct pipe_screen *pscreen,
 326                 const struct pipe_resource *tmpl,
 327                 struct winsys_handle *handle)
 328 {
 329         struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
 330         struct fd_resource_slice *slice = &rsc->slices[0];
 331         struct pipe_resource *prsc = &rsc->base.b;
 332
 333         DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
 334                         "nr_samples=%u, usage=%u, bind=%x, flags=%x",
 335                         tmpl->target, util_format_name(tmpl->format),
 336                         tmpl->width0, tmpl->height0, tmpl->depth0,
 337                         tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
 338                         tmpl->usage, tmpl->bind, tmpl->flags);
 339
 340         if (!rsc)
 341                 return NULL;
 342
 343         *prsc = *tmpl;
 344
 345         pipe_reference_init(&prsc->reference, 1);
 346         list_inithead(&rsc->list);
 347         prsc->screen = pscreen;
 348
 349         rsc->bo = fd_screen_bo_from_handle(pscreen, handle, &slice->pitch);
 350         if (!rsc->bo)
 351                 goto fail;
 352
 353         rsc->base.vtbl = &fd_resource_vtbl;
 354         rsc->cpp = util_format_get_blocksize(tmpl->format);
 355         slice->pitch /= rsc->cpp;
 356
 357         assert(rsc->cpp);
 358
 359         return prsc;
 360
 361 fail:
 362         fd_resource_destroy(pscreen, prsc);
 363         return NULL;
 364 }
 365
 366 static void fd_blitter_pipe_begin(struct fd_context *ctx);
 367 static void fd_blitter_pipe_end(struct fd_context *ctx);
 368
 369 /**
 370  * _copy_region using pipe (3d engine)
 371  */
 372 static bool
 373 fd_blitter_pipe_copy_region(struct fd_context *ctx,
 374                 struct pipe_resource *dst,
 375                 unsigned dst_level,
 376                 unsigned dstx, unsigned dsty, unsigned dstz,
 377                 struct pipe_resource *src,
 378                 unsigned src_level,
 379                 const struct pipe_box *src_box)
 380 {
 381         /* not until we allow rendertargets to be buffers */
 382         if (dst->target == PIPE_BUFFER || src->target == PIPE_BUFFER)
 383                 return false;
 384
 385         if (!util_blitter_is_copy_supported(ctx->blitter, dst, src))
 386                 return false;
 387
 388         fd_blitter_pipe_begin(ctx);
 389         util_blitter_copy_texture(ctx->blitter,
 390                         dst, dst_level, dstx, dsty, dstz,
 391                         src, src_level, src_box);
 392         fd_blitter_pipe_end(ctx);
 393
 394         return true;
 395 }
 396
 397 /**
 398  * Copy a block of pixels from one resource to another.
 399  * The resource must be of the same format.
 400  * Resources with nr_samples > 1 are not allowed.
 401  */
 402 static void
 403 fd_resource_copy_region(struct pipe_context *pctx,
 404                 struct pipe_resource *dst,
 405                 unsigned dst_level,
 406                 unsigned dstx, unsigned dsty, unsigned dstz,
 407                 struct pipe_resource *src,
 408                 unsigned src_level,
 409                 const struct pipe_box *src_box)
 410 {
 411         struct fd_context *ctx = fd_context(pctx);
 412
 413         /* TODO if we have 2d core, or other DMA engine that could be used
 414          * for simple copies and reasonably easily synchronized with the 3d
 415          * core, this is where we'd plug it in..
 416          */
 417
 418         /* try blit on 3d pipe: */
 419         if (fd_blitter_pipe_copy_region(ctx,
 420                         dst, dst_level, dstx, dsty, dstz,
 421                         src, src_level, src_box))
 422                 return;
 423
 424         /* else fallback to pure sw: */
 425         util_resource_copy_region(pctx,
 426                         dst, dst_level, dstx, dsty, dstz,
 427                         src, src_level, src_box);
 428 }
 429
 430 /**
 431  * Optimal hardware path for blitting pixels.
 432  * Scaling, format conversion, up- and downsampling (resolve) are allowed.
 433  */
 434 static void
 435 fd_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
 436 {
 437         struct fd_context *ctx = fd_context(pctx);
 438         struct pipe_blit_info info = *blit_info;
 439
 440         if (info.src.resource->nr_samples > 1 &&
 441                         info.dst.resource->nr_samples <= 1 &&
 442                         !util_format_is_depth_or_stencil(info.src.resource->format) &&
 443                         !util_format_is_pure_integer(info.src.resource->format)) {
 444                 DBG("color resolve unimplemented");
 445                 return;
 446         }
 447
 448         if (util_try_blit_via_copy_region(pctx, &info)) {
 449                 return; /* done */
 450         }
 451
 452         if (info.mask & PIPE_MASK_S) {
 453                 DBG("cannot blit stencil, skipping");
 454                 info.mask &= ~PIPE_MASK_S;
 455         }
 456
 457         if (!util_blitter_is_blit_supported(ctx->blitter, &info)) {
 458                 DBG("blit unsupported %s -> %s",
 459                                 util_format_short_name(info.src.resource->format),
 460                                 util_format_short_name(info.dst.resource->format));
 461                 return;
 462         }
 463
 464         fd_blitter_pipe_begin(ctx);
 465         util_blitter_blit(ctx->blitter, &info);
 466         fd_blitter_pipe_end(ctx);
 467 }
 468
 469 static void
 470 fd_blitter_pipe_begin(struct fd_context *ctx)
 471 {
 472         util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb);
 473         util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx);
 474         util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp);
 475         util_blitter_save_rasterizer(ctx->blitter, ctx->rasterizer);
 476         util_blitter_save_viewport(ctx->blitter, &ctx->viewport);
 477         util_blitter_save_scissor(ctx->blitter, &ctx->scissor);
 478         util_blitter_save_fragment_shader(ctx->blitter, ctx->prog.fp);
 479         util_blitter_save_blend(ctx->blitter, ctx->blend);
 480         util_blitter_save_depth_stencil_alpha(ctx->blitter, ctx->zsa);
 481         util_blitter_save_stencil_ref(ctx->blitter, &ctx->stencil_ref);
 482         util_blitter_save_sample_mask(ctx->blitter, ctx->sample_mask);
 483         util_blitter_save_framebuffer(ctx->blitter, &ctx->framebuffer);
 484         util_blitter_save_fragment_sampler_states(ctx->blitter,
 485                         ctx->fragtex.num_samplers,
 486                         (void **)ctx->fragtex.samplers);
 487         util_blitter_save_fragment_sampler_views(ctx->blitter,
 488                         ctx->fragtex.num_textures, ctx->fragtex.textures);
 489
 490         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_BLIT);
 491 }
 492
 493 static void
 494 fd_blitter_pipe_end(struct fd_context *ctx)
 495 {
 496         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 497 }
 498
 499 static void
 500 fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 501 {
 502         struct fd_resource *rsc = fd_resource(prsc);
 503
 504         if (rsc->dirty)
 505                 fd_context_render(pctx);
 506 }
 507
 508 void
 509 fd_resource_screen_init(struct pipe_screen *pscreen)
 510 {
 511         pscreen->resource_create = fd_resource_create;
 512         pscreen->resource_from_handle = fd_resource_from_handle;
 513         pscreen->resource_get_handle = u_resource_get_handle_vtbl;
 514         pscreen->resource_destroy = u_resource_destroy_vtbl;
 515 }
 516
 517 void
 518 fd_resource_context_init(struct pipe_context *pctx)
 519 {
 520         pctx->transfer_map = u_transfer_map_vtbl;
 521         pctx->transfer_flush_region = u_transfer_flush_region_vtbl;
 522         pctx->transfer_unmap = u_transfer_unmap_vtbl;
 523         pctx->transfer_inline_write = u_transfer_inline_write_vtbl;
 524         pctx->create_surface = fd_create_surface;
 525         pctx->surface_destroy = fd_surface_destroy;
 526         pctx->resource_copy_region = fd_resource_copy_region;
 527         pctx->blit = fd_blit;
 528         pctx->flush_resource = fd_flush_resource;
 529 }