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