src/gallium/drivers/nvfx/nvfx_surface.c

   1
   2 /**************************************************************************
   3  *
   4  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 #include "pipe/p_context.h"
  30 #include "pipe/p_format.h"
  31 #include "util/u_format.h"
  32 #include "util/u_math.h"
  33 #include "util/u_memory.h"
  34 #include "util/u_pack_color.h"
  35 #include "util/u_blitter.h"
  36
  37 #include "nouveau/nouveau_winsys.h"
  38 #include "nouveau/nouveau_screen.h"
  39 #include "nvfx_context.h"
  40 #include "nvfx_screen.h"
  41 #include "nvfx_resource.h"
  42 #include "nv04_2d.h"
  43
  44 #include <nouveau/nouveau_bo.h>
  45
  46 static INLINE void
  47 nvfx_region_set_format(struct nv04_region* rgn, enum pipe_format format)
  48 {
  49         unsigned bits = util_format_get_blocksizebits(format);
  50         unsigned shift = 0;
  51         rgn->one_bits = 0;
  52
  53         switch(bits)
  54         {
  55         case 8:
  56                 rgn->bpps = 0;
  57                 break;
  58         case 16:
  59                 rgn->bpps = 1;
  60                 if(format == PIPE_FORMAT_B5G5R5X1_UNORM)
  61                         rgn->one_bits = 1;
  62                 break;
  63         case 32:
  64                 rgn->bpps = 2;
  65                 if(format == PIPE_FORMAT_R8G8B8X8_UNORM || format == PIPE_FORMAT_B8G8R8X8_UNORM)
  66                         rgn->one_bits = 8;
  67                 break;
  68         case 64:
  69                 rgn->bpps = 2;
  70                 shift = 1;
  71                 break;
  72         case 128:
  73                 rgn->bpps = 2;
  74                 shift = 2;
  75                 break;
  76         }
  77
  78         if(shift) {
  79                 rgn->x = util_format_get_nblocksx(format, rgn->x) << shift;
  80                 rgn->y = util_format_get_nblocksy(format, rgn->y);
  81                 rgn->w <<= shift;
  82         }
  83 }
  84
  85 static INLINE void
  86 nvfx_region_init_for_surface(struct nv04_region* rgn, struct nvfx_surface* surf, unsigned x, unsigned y, bool for_write)
  87 {
  88         rgn->x = x;
  89         rgn->y = y;
  90         rgn->z = 0;
  91
  92         if(surf->temp)
  93         {
  94                 rgn->bo = surf->temp->base.bo;
  95                 rgn->offset = 0;
  96                 rgn->pitch = surf->temp->linear_pitch;
  97
  98                 if(for_write)
  99                         util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(&surf->base.base), &surf->base);
 100         } else {
 101                 rgn->bo = ((struct nvfx_resource*)surf->base.base.texture)->bo;
 102                 rgn->offset = surf->offset;
 103
 104                 if(surf->base.base.texture->flags & NVFX_RESOURCE_FLAG_LINEAR)
 105                         rgn->pitch = surf->pitch;
 106                 else
 107                 {
 108                         rgn->pitch = 0;
 109                         rgn->z = surf->base.base.u.tex.first_layer;
 110                         rgn->w = surf->base.base.width;
 111                         rgn->h = surf->base.base.height;
 112                         rgn->d = u_minify(surf->base.base.texture->depth0, surf->base.base.u.tex.level);
 113                 }
 114         }
 115
 116         nvfx_region_set_format(rgn, surf->base.base.format);
 117         if(!rgn->pitch)
 118                 nv04_region_try_to_linearize(rgn);
 119 }
 120
 121 static INLINE void
 122 nvfx_region_init_for_subresource(struct nv04_region* rgn, struct pipe_resource* pt, unsigned level, unsigned x, unsigned y, unsigned z, bool for_write)
 123 {
 124         if(pt->target != PIPE_BUFFER)
 125         {
 126                 struct nvfx_surface* ns = (struct nvfx_surface*)util_surfaces_peek(&((struct nvfx_miptree*)pt)->surfaces, pt, level, z);
 127                 if(ns && util_dirty_surface_is_dirty(&ns->base))
 128                 {
 129                         nvfx_region_init_for_surface(rgn, ns, x, y, for_write);
 130                         return;
 131                 }
 132         }
 133
 134         rgn->bo = ((struct nvfx_resource*)pt)->bo;
 135         rgn->offset = nvfx_subresource_offset(pt, z, level, z);
 136         rgn->x = x;
 137         rgn->y = y;
 138
 139         if(pt->flags & NVFX_RESOURCE_FLAG_LINEAR)
 140         {
 141                 rgn->pitch = nvfx_subresource_pitch(pt, level);
 142                 rgn->z = 0;
 143         }
 144         else
 145         {
 146                 rgn->pitch = 0;
 147                 rgn->z = z;
 148                 rgn->w = u_minify(pt->width0, level);
 149                 rgn->h = u_minify(pt->height0, level);
 150                 rgn->d = u_minify(pt->depth0, level);
 151         }
 152
 153         nvfx_region_set_format(rgn, pt->format);
 154         if(!rgn->pitch)
 155                 nv04_region_try_to_linearize(rgn);
 156 }
 157
 158 // don't save index buffer because blitter doesn't setit
 159 static struct blitter_context*
 160 nvfx_get_blitter(struct pipe_context* pipe, int copy)
 161 {
 162         struct nvfx_context* nvfx = nvfx_context(pipe);
 163         struct blitter_context** pblitter;
 164         struct blitter_context* blitter;
 165
 166         assert(nvfx->blitters_in_use < Elements(nvfx->blitter));
 167
 168         if(nvfx->query && !nvfx->blitters_in_use)
 169         {
 170                 struct nouveau_channel* chan = nvfx->screen->base.channel;
 171                 struct nouveau_grobj *eng3d = nvfx->screen->eng3d;
 172                 BEGIN_RING(chan, eng3d, NV30_3D_QUERY_ENABLE, 1);
 173                 OUT_RING(chan, 0);
 174         }
 175
 176         pblitter = &nvfx->blitter[nvfx->blitters_in_use++];
 177         if(!*pblitter)
 178                 *pblitter = util_blitter_create(pipe);
 179         blitter = *pblitter;
 180
 181         util_blitter_save_blend(blitter, nvfx->blend);
 182         util_blitter_save_depth_stencil_alpha(blitter, nvfx->zsa);
 183         util_blitter_save_stencil_ref(blitter, &nvfx->stencil_ref);
 184         util_blitter_save_rasterizer(blitter, nvfx->rasterizer);
 185         util_blitter_save_fragment_shader(blitter, nvfx->fragprog);
 186         util_blitter_save_vertex_shader(blitter, nvfx->vertprog);
 187         util_blitter_save_viewport(blitter, &nvfx->viewport);
 188         util_blitter_save_framebuffer(blitter, &nvfx->framebuffer);
 189         util_blitter_save_clip(blitter, &nvfx->clip);
 190         util_blitter_save_vertex_elements(blitter, nvfx->vtxelt);
 191         util_blitter_save_vertex_buffers(blitter, nvfx->vtxbuf_nr, nvfx->vtxbuf);
 192
 193         if(copy)
 194         {
 195                 util_blitter_save_fragment_sampler_states(blitter, nvfx->nr_samplers, (void**)nvfx->tex_sampler);
 196                 util_blitter_save_fragment_sampler_views(blitter, nvfx->nr_textures, nvfx->fragment_sampler_views);
 197         }
 198
 199         return blitter;
 200 }
 201
 202 static inline void
 203 nvfx_put_blitter(struct pipe_context* pipe, struct blitter_context* blitter)
 204 {
 205         struct nvfx_context* nvfx = nvfx_context(pipe);
 206         --nvfx->blitters_in_use;
 207         assert(nvfx->blitters_in_use >= 0);
 208
 209         if(nvfx->query && !nvfx->blitters_in_use)
 210         {
 211                 struct nouveau_channel* chan = nvfx->screen->base.channel;
 212                 struct nouveau_grobj *eng3d = nvfx->screen->eng3d;
 213                 BEGIN_RING(chan, eng3d, NV30_3D_QUERY_ENABLE, 1);
 214                 OUT_RING(chan, 1);
 215         }
 216 }
 217
 218 static unsigned
 219 nvfx_region_clone(struct nv04_2d_context* ctx, struct nv04_region* rgn, unsigned w, unsigned h, boolean for_read)
 220 {
 221         unsigned begin = nv04_region_begin(rgn, w, h);
 222         unsigned end = nv04_region_end(rgn, w, h);
 223         unsigned size = end - begin;
 224         struct nouveau_bo* bo = 0;
 225         nouveau_bo_new(rgn->bo->device, NOUVEAU_BO_MAP | NOUVEAU_BO_GART, 256, size, &bo);
 226
 227         if(for_read || (size > ((w * h) << rgn->bpps)))
 228                 nv04_memcpy(ctx, bo, 0, rgn->bo, rgn->offset + begin, size);
 229
 230         rgn->bo = bo;
 231         rgn->offset = -begin;
 232         return begin;
 233 }
 234
 235 static void
 236 nvfx_resource_copy_region(struct pipe_context *pipe,
 237                           struct pipe_resource *dstr, unsigned dst_level,
 238                           unsigned dstx, unsigned dsty, unsigned dstz,
 239                           struct pipe_resource *srcr, unsigned src_level,
 240                           const struct pipe_box *src_box)
 241 {
 242         static int copy_threshold = -1;
 243         struct nv04_2d_context *ctx = nvfx_screen(pipe->screen)->eng2d;
 244         struct nv04_region dst, src;
 245         int dst_to_gpu;
 246         int src_on_gpu;
 247         boolean small;
 248         int ret;
 249         unsigned w = src_box->width;
 250         unsigned h = src_box->height;
 251
 252         if(!w || !h)
 253                 return;
 254
 255         if(copy_threshold < 0)
 256                 copy_threshold = debug_get_num_option("NOUVEAU_COPY_THRESHOLD", 4);
 257
 258         dst_to_gpu = dstr->usage != PIPE_USAGE_DYNAMIC && dstr->usage != PIPE_USAGE_STAGING;
 259         src_on_gpu = nvfx_resource_on_gpu(srcr);
 260
 261         nvfx_region_init_for_subresource(&dst, dstr, dst_level, dstx, dsty, dstz, TRUE);
 262         nvfx_region_init_for_subresource(&src, srcr, src_level, src_box->x, src_box->y, src_box->z, FALSE);
 263         w = util_format_get_stride(dstr->format, w) >> dst.bpps;
 264         h = util_format_get_nblocksy(dstr->format, h);
 265
 266         small = (w * h <= copy_threshold);
 267         if((!dst_to_gpu || !src_on_gpu) && small)
 268                 ret = -1; /* use the CPU */
 269         else
 270                 ret = nv04_region_copy_2d(ctx, &dst, &src, w, h, dst_to_gpu, src_on_gpu);
 271         if(!ret)
 272         {}
 273         else if(ret > 0
 274                         && dstr->bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL)
 275                         && srcr->bind & PIPE_BIND_SAMPLER_VIEW)
 276         {
 277                 /* this currently works because we hack the bind flags on resource creation to be
 278                  * the maximum set that the resource type actually supports
 279                  *
 280                  * TODO: perhaps support reinterpreting the formats
 281                  */
 282                 struct blitter_context* blitter = nvfx_get_blitter(pipe, 1);
 283                 util_blitter_copy_region(blitter, dstr, dst_level, dstx, dsty, dstz, srcr, src_level, src_box, TRUE);
 284                 nvfx_put_blitter(pipe, blitter);
 285         }
 286         else
 287         {
 288                 struct nv04_region dstt = dst;
 289                 struct nv04_region srct = src;
 290                 unsigned dstbegin = 0;
 291
 292                 if(!small)
 293                 {
 294                         if(src_on_gpu)
 295                                 nvfx_region_clone(ctx, &srct, w, h, TRUE);
 296
 297                         if(dst_to_gpu)
 298                                 dstbegin = nvfx_region_clone(ctx, &dstt, w, h, FALSE);
 299                 }
 300
 301                 nv04_region_copy_cpu(&dstt, &srct, w, h);
 302
 303                 if(srct.bo != src.bo)
 304                         nouveau_screen_bo_release(pipe->screen, srct.bo);
 305
 306                 if(dstt.bo != dst.bo)
 307                 {
 308                         nv04_memcpy(ctx, dst.bo, dst.offset + dstbegin, dstt.bo, 0, dstt.bo->size);
 309                         nouveau_screen_bo_release(pipe->screen, dstt.bo);
 310                 }
 311         }
 312 }
 313
 314 static int
 315 nvfx_surface_fill(struct pipe_context* pipe, struct pipe_surface *dsts,
 316                   unsigned dx, unsigned dy, unsigned w, unsigned h, unsigned value)
 317 {
 318         struct nv04_2d_context *ctx = nvfx_screen(pipe->screen)->eng2d;
 319         struct nv04_region dst;
 320         int ret;
 321         /* Always try to use the GPU right now, if possible
 322          * If the user wanted the surface data on the CPU, he would have cleared with memset (hopefully) */
 323
 324         // we don't care about interior pixel order since we set all them to the same value
 325         nvfx_region_init_for_surface(&dst, (struct nvfx_surface*)dsts, dx, dy, TRUE);
 326
 327         w = util_format_get_stride(dsts->format, w) >> dst.bpps;
 328         h = util_format_get_nblocksy(dsts->format, h);
 329
 330         ret = nv04_region_fill_2d(ctx, &dst, w, h, value);
 331         if(ret > 0 && dsts->texture->bind & PIPE_BIND_RENDER_TARGET)
 332                 return 1;
 333         else if(ret)
 334         {
 335                 struct nv04_region dstt = dst;
 336                 unsigned dstbegin = 0;
 337
 338                 if(nvfx_resource_on_gpu(dsts->texture))
 339                         dstbegin = nvfx_region_clone(ctx, &dstt, w, h, FALSE);
 340
 341                 nv04_region_fill_cpu(&dstt, w, h, value);
 342
 343                 if(dstt.bo != dst.bo)
 344                 {
 345                         nv04_memcpy(ctx, dst.bo, dst.offset + dstbegin, dstt.bo, 0, dstt.bo->size);
 346                         nouveau_screen_bo_release(pipe->screen, dstt.bo);
 347                 }
 348         }
 349
 350         return 0;
 351 }
 352
 353
 354 void
 355 nvfx_screen_surface_takedown(struct pipe_screen *pscreen)
 356 {
 357         nv04_2d_context_takedown(nvfx_screen(pscreen)->eng2d);
 358         nvfx_screen(pscreen)->eng2d = 0;
 359 }
 360
 361 int
 362 nvfx_screen_surface_init(struct pipe_screen *pscreen)
 363 {
 364         struct nv04_2d_context* ctx = nv04_2d_context_init(nouveau_screen(pscreen)->channel);
 365         if(!ctx)
 366                 return -1;
 367         nvfx_screen(pscreen)->eng2d = ctx;
 368         return 0;
 369 }
 370
 371 static void
 372 nvfx_surface_copy_temp(struct pipe_context* pipe, struct pipe_surface* surf, int to_temp)
 373 {
 374         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
 375         struct pipe_box box;
 376         struct nvfx_context* nvfx = nvfx_context(pipe);
 377         struct nvfx_miptree* temp;
 378         unsigned use_vertex_buffers;
 379         boolean use_index_buffer;
 380         unsigned base_vertex;
 381
 382         /* temporarily detach the temp, so it isn't used in place of the actual resource */
 383         temp = ns->temp;
 384         ns->temp = 0;
 385
 386         // TODO: we really should do this validation before setting these variable in draw calls
 387         use_vertex_buffers = nvfx->use_vertex_buffers;
 388         use_index_buffer = nvfx->use_index_buffer;
 389         base_vertex = nvfx->base_vertex;
 390
 391         box.x = box.y = 0;
 392         assert(surf->u.tex.first_layer == surf->u.tex.last_layer);
 393         box.width = surf->width;
 394         box.height = surf->height;
 395         box.depth = 1;
 396
 397         if(to_temp) {
 398                 box.z = surf->u.tex.first_layer;
 399                 nvfx_resource_copy_region(pipe, &temp->base.base, 0, 0, 0, 0, surf->texture, surf->u.tex.level, &box);
 400         }
 401         else {
 402                 box.z = 0;
 403                 nvfx_resource_copy_region(pipe, surf->texture, surf->u.tex.level, 0, 0, surf->u.tex.first_layer, &temp->base.base, 0, &box);
 404         }
 405
 406         /* If this triggers, it probably means we attempted to use the blitter
 407          * but failed due to non-renderability of the target.
 408          * Obviously, this would lead to infinite recursion if supported. */
 409         assert(!ns->temp);
 410
 411         ns->temp = temp;
 412
 413         nvfx->use_vertex_buffers = use_vertex_buffers;
 414         nvfx->use_index_buffer = use_index_buffer;
 415         nvfx->base_vertex = base_vertex;
 416
 417         nvfx->dirty |= NVFX_NEW_ARRAYS;
 418         nvfx->draw_dirty |= NVFX_NEW_ARRAYS;
 419 }
 420
 421 void
 422 nvfx_surface_create_temp(struct pipe_context* pipe, struct pipe_surface* surf)
 423 {
 424         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
 425         struct pipe_resource template;
 426         memset(&template, 0, sizeof(struct pipe_resource));
 427         template.target = PIPE_TEXTURE_2D;
 428         template.format = surf->format;
 429         template.width0 = surf->width;
 430         template.height0 = surf->height;
 431         template.depth0 = 1;
 432         template.nr_samples = surf->texture->nr_samples;
 433         template.flags = NVFX_RESOURCE_FLAG_LINEAR;
 434
 435         assert(!ns->temp && !util_dirty_surface_is_dirty(&ns->base));
 436
 437         ns->temp = (struct nvfx_miptree*)nvfx_miptree_create(pipe->screen, &template);
 438         nvfx_surface_copy_temp(pipe, surf, 1);
 439 }
 440
 441 void
 442 nvfx_surface_flush(struct pipe_context* pipe, struct pipe_surface* surf)
 443 {
 444         struct nvfx_context* nvfx = (struct nvfx_context*)pipe;
 445         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
 446         boolean bound = FALSE;
 447
 448         nvfx_surface_copy_temp(pipe, surf, 0);
 449
 450         util_dirty_surface_set_clean(nvfx_surface_get_dirty_surfaces(surf), &ns->base);
 451
 452         if(nvfx->framebuffer.zsbuf == surf)
 453                 bound = TRUE;
 454         else
 455         {
 456                 for(unsigned i = 0; i < nvfx->framebuffer.nr_cbufs; ++i)
 457                 {
 458                         if(nvfx->framebuffer.cbufs[i] == surf)
 459                         {
 460                                 bound = TRUE;
 461                                 break;
 462                         }
 463                 }
 464         }
 465
 466         if(!bound)
 467                 pipe_resource_reference((struct pipe_resource**)&ns->temp, 0);
 468 }
 469
 470 static void
 471 nvfx_clear_render_target(struct pipe_context *pipe,
 472                          struct pipe_surface *dst,
 473                          const float *rgba,
 474                          unsigned dstx, unsigned dsty,
 475                          unsigned width, unsigned height)
 476 {
 477         union util_color uc;
 478         util_pack_color(rgba, dst->format, &uc);
 479
 480         if(util_format_get_blocksizebits(dst->format) > 32
 481                 || nvfx_surface_fill(pipe, dst, dstx, dsty, width, height, uc.ui))
 482         {
 483                 // TODO: probably should use hardware clear here instead if possible
 484                 struct blitter_context* blitter = nvfx_get_blitter(pipe, 0);
 485                 util_blitter_clear_render_target(blitter, dst, rgba, dstx, dsty, width, height);
 486                 nvfx_put_blitter(pipe, blitter);
 487         }
 488 }
 489
 490 static void
 491 nvfx_clear_depth_stencil(struct pipe_context *pipe,
 492                          struct pipe_surface *dst,
 493                          unsigned clear_flags,
 494                          double depth,
 495                          unsigned stencil,
 496                          unsigned dstx, unsigned dsty,
 497                          unsigned width, unsigned height)
 498 {
 499         if(util_format_get_blocksizebits(dst->format) > 32
 500                 || nvfx_surface_fill(pipe, dst, dstx, dsty, width, height, util_pack_z_stencil(dst->format, depth, stencil)))
 501         {
 502                 // TODO: probably should use hardware clear here instead if possible
 503                 struct blitter_context* blitter = nvfx_get_blitter(pipe, 0);
 504                 util_blitter_clear_depth_stencil(blitter, dst, clear_flags, depth, stencil, dstx, dsty, width, height);
 505                 nvfx_put_blitter(pipe, blitter);
 506         }
 507 }
 508
 509
 510 void
 511 nvfx_init_surface_functions(struct nvfx_context *nvfx)
 512 {
 513         nvfx->pipe.resource_copy_region = nvfx_resource_copy_region;
 514         nvfx->pipe.clear_render_target = nvfx_clear_render_target;
 515         nvfx->pipe.clear_depth_stencil = nvfx_clear_depth_stencil;
 516 }