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 #include "util/u_surface.h"
  37
  38 #include "nouveau/nouveau_winsys.h"
  39 #include "nouveau/nouveau_screen.h"
  40 #include "nvfx_context.h"
  41 #include "nvfx_screen.h"
  42 #include "nvfx_resource.h"
  43 #include "nv04_2d.h"
  44
  45 #include <nouveau/nouveau_bo.h>
  46
  47 static INLINE void
  48 nvfx_region_set_format(struct nv04_region* rgn, enum pipe_format format)
  49 {
  50         unsigned bits = util_format_get_blocksizebits(format);
  51         unsigned shift = 0;
  52         rgn->one_bits = 0;
  53
  54         switch(bits)
  55         {
  56         case 8:
  57                 rgn->bpps = 0;
  58                 break;
  59         case 16:
  60                 rgn->bpps = 1;
  61                 if(format == PIPE_FORMAT_B5G5R5X1_UNORM)
  62                         rgn->one_bits = 1;
  63                 break;
  64         case 32:
  65                 rgn->bpps = 2;
  66                 if(format == PIPE_FORMAT_R8G8B8X8_UNORM || format == PIPE_FORMAT_B8G8R8X8_UNORM)
  67                         rgn->one_bits = 8;
  68                 break;
  69         case 64:
  70                 rgn->bpps = 2;
  71                 shift = 1;
  72                 break;
  73         case 128:
  74                 rgn->bpps = 2;
  75                 shift = 2;
  76                 break;
  77         }
  78
  79         if(shift) {
  80                 rgn->x = util_format_get_nblocksx(format, rgn->x) << shift;
  81                 rgn->y = util_format_get_nblocksy(format, rgn->y);
  82                 rgn->w <<= shift;
  83         }
  84 }
  85
  86 static INLINE void
  87 nvfx_region_init_for_surface(struct nv04_region* rgn, struct nvfx_surface* surf, unsigned x, unsigned y, boolean for_write)
  88 {
  89         rgn->x = x;
  90         rgn->y = y;
  91         rgn->z = 0;
  92
  93         if(surf->temp)
  94         {
  95                 rgn->bo = surf->temp->base.bo;
  96                 rgn->offset = 0;
  97                 rgn->pitch = surf->temp->linear_pitch;
  98
  99                 if(for_write)
 100                         util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(&surf->base.base), &surf->base);
 101         } else {
 102                 rgn->bo = ((struct nvfx_resource*)surf->base.base.texture)->bo;
 103                 rgn->offset = surf->offset;
 104
 105                 if(surf->base.base.texture->flags & NOUVEAU_RESOURCE_FLAG_LINEAR)
 106                         rgn->pitch = surf->pitch;
 107                 else
 108                 {
 109                         rgn->pitch = 0;
 110                         rgn->z = surf->base.base.u.tex.first_layer;
 111                         rgn->w = surf->base.base.width;
 112                         rgn->h = surf->base.base.height;
 113                         rgn->d = u_minify(surf->base.base.texture->depth0, surf->base.base.u.tex.level);
 114                 }
 115         }
 116
 117         nvfx_region_set_format(rgn, surf->base.base.format);
 118         if(!rgn->pitch)
 119                 nv04_region_try_to_linearize(rgn);
 120 }
 121
 122 static INLINE void
 123 nvfx_region_init_for_subresource(struct nv04_region* rgn, struct pipe_resource* pt, unsigned level, unsigned x, unsigned y, unsigned z, boolean for_write)
 124 {
 125         if(pt->target != PIPE_BUFFER)
 126         {
 127                 struct nvfx_surface* ns = (struct nvfx_surface*)util_surfaces_peek(&((struct nvfx_miptree*)pt)->surfaces, pt, level, z);
 128                 if(ns && util_dirty_surface_is_dirty(&ns->base))
 129                 {
 130                         nvfx_region_init_for_surface(rgn, ns, x, y, for_write);
 131                         return;
 132                 }
 133         }
 134
 135         rgn->bo = ((struct nvfx_resource*)pt)->bo;
 136         rgn->offset = nvfx_subresource_offset(pt, z, level, z);
 137         rgn->x = x;
 138         rgn->y = y;
 139
 140         if(pt->flags & NOUVEAU_RESOURCE_FLAG_LINEAR)
 141         {
 142                 rgn->pitch = nvfx_subresource_pitch(pt, level);
 143                 rgn->z = 0;
 144         }
 145         else
 146         {
 147                 rgn->pitch = 0;
 148                 rgn->z = z;
 149                 rgn->w = u_minify(pt->width0, level);
 150                 rgn->h = u_minify(pt->height0, level);
 151                 rgn->d = u_minify(pt->depth0, level);
 152         }
 153
 154         nvfx_region_set_format(rgn, pt->format);
 155         if(!rgn->pitch)
 156                 nv04_region_try_to_linearize(rgn);
 157 }
 158
 159 // don't save index buffer because blitter doesn't setit
 160 static struct blitter_context*
 161 nvfx_get_blitter(struct pipe_context* pipe, int copy)
 162 {
 163         struct nvfx_context* nvfx = nvfx_context(pipe);
 164         struct blitter_context** pblitter;
 165         struct blitter_context* blitter;
 166
 167         assert(nvfx->blitters_in_use < Elements(nvfx->blitter));
 168
 169         if(nvfx->query && !nvfx->blitters_in_use)
 170         {
 171                 struct nouveau_channel* chan = nvfx->screen->base.channel;
 172                 struct nouveau_grobj *eng3d = nvfx->screen->eng3d;
 173                 BEGIN_RING(chan, eng3d, NV30_3D_QUERY_ENABLE, 1);
 174                 OUT_RING(chan, 0);
 175         }
 176
 177         pblitter = &nvfx->blitter[nvfx->blitters_in_use++];
 178         if(!*pblitter)
 179                 *pblitter = util_blitter_create(pipe);
 180         blitter = *pblitter;
 181
 182         util_blitter_save_blend(blitter, nvfx->blend);
 183         util_blitter_save_depth_stencil_alpha(blitter, nvfx->zsa);
 184         util_blitter_save_stencil_ref(blitter, &nvfx->stencil_ref);
 185         util_blitter_save_rasterizer(blitter, nvfx->rasterizer);
 186         util_blitter_save_fragment_shader(blitter, nvfx->fragprog);
 187         util_blitter_save_vertex_shader(blitter, nvfx->vertprog);
 188         util_blitter_save_viewport(blitter, &nvfx->viewport);
 189         util_blitter_save_framebuffer(blitter, &nvfx->framebuffer);
 190         util_blitter_save_clip(blitter, &nvfx->clip);
 191         util_blitter_save_vertex_elements(blitter, nvfx->vtxelt);
 192         util_blitter_save_vertex_buffers(blitter, nvfx->vtxbuf_nr, nvfx->vtxbuf);
 193
 194         if(copy)
 195         {
 196                 util_blitter_save_fragment_sampler_states(blitter, nvfx->nr_samplers, (void**)nvfx->tex_sampler);
 197                 util_blitter_save_fragment_sampler_views(blitter, nvfx->nr_textures, nvfx->fragment_sampler_views);
 198         }
 199
 200         return blitter;
 201 }
 202
 203 static inline void
 204 nvfx_put_blitter(struct pipe_context* pipe, struct blitter_context* blitter)
 205 {
 206         struct nvfx_context* nvfx = nvfx_context(pipe);
 207         --nvfx->blitters_in_use;
 208         assert(nvfx->blitters_in_use >= 0);
 209
 210         if(nvfx->query && !nvfx->blitters_in_use)
 211         {
 212                 struct nouveau_channel* chan = nvfx->screen->base.channel;
 213                 struct nouveau_grobj *eng3d = nvfx->screen->eng3d;
 214                 BEGIN_RING(chan, eng3d, NV30_3D_QUERY_ENABLE, 1);
 215                 OUT_RING(chan, 1);
 216         }
 217 }
 218
 219 static unsigned
 220 nvfx_region_clone(struct nv04_2d_context* ctx, struct nv04_region* rgn, unsigned w, unsigned h, boolean for_read)
 221 {
 222         unsigned begin = nv04_region_begin(rgn, w, h);
 223         unsigned end = nv04_region_end(rgn, w, h);
 224         unsigned size = end - begin;
 225         struct nouveau_bo* bo = 0;
 226         nouveau_bo_new(rgn->bo->device, NOUVEAU_BO_MAP | NOUVEAU_BO_GART, 256, size, &bo);
 227
 228         if(for_read || (size > ((w * h) << rgn->bpps)))
 229                 nv04_memcpy(ctx, bo, 0, rgn->bo, rgn->offset + begin, size);
 230
 231         rgn->bo = bo;
 232         rgn->offset = -begin;
 233         return begin;
 234 }
 235
 236 static void
 237 nvfx_resource_copy_region(struct pipe_context *pipe,
 238                           struct pipe_resource *dstr, unsigned dst_level,
 239                           unsigned dstx, unsigned dsty, unsigned dstz,
 240                           struct pipe_resource *srcr, unsigned src_level,
 241                           const struct pipe_box *src_box)
 242 {
 243         static int copy_threshold = -1;
 244         struct nv04_2d_context *ctx = nvfx_screen(pipe->screen)->eng2d;
 245         struct nv04_region dst, src;
 246         int dst_to_gpu;
 247         int src_on_gpu;
 248         boolean small;
 249         int ret;
 250         unsigned w = src_box->width;
 251         unsigned h = src_box->height;
 252
 253         if(!w || !h)
 254                 return;
 255
 256         /* Fallback for buffers. */
 257         if (dstr->target == PIPE_BUFFER && srcr->target == PIPE_BUFFER) {
 258                 util_resource_copy_region(pipe, dstr, dst_level, dstx, dsty, dstz,
 259                                           srcr, src_level, src_box);
 260                 return;
 261         }
 262
 263         if(copy_threshold < 0)
 264                 copy_threshold = debug_get_num_option("NOUVEAU_COPY_THRESHOLD", 4);
 265
 266         dst_to_gpu = dstr->usage != PIPE_USAGE_DYNAMIC && dstr->usage != PIPE_USAGE_STAGING;
 267         src_on_gpu = nvfx_resource_on_gpu(srcr);
 268
 269         nvfx_region_init_for_subresource(&dst, dstr, dst_level, dstx, dsty, dstz, TRUE);
 270         nvfx_region_init_for_subresource(&src, srcr, src_level, src_box->x, src_box->y, src_box->z, FALSE);
 271         w = util_format_get_stride(dstr->format, w) >> dst.bpps;
 272         h = util_format_get_nblocksy(dstr->format, h);
 273
 274         small = (w * h <= copy_threshold);
 275         if((!dst_to_gpu || !src_on_gpu) && small)
 276                 ret = -1; /* use the CPU */
 277         else
 278                 ret = nv04_region_copy_2d(ctx, &dst, &src, w, h, dst_to_gpu, src_on_gpu);
 279         if(!ret)
 280         {}
 281         else if(ret > 0
 282                         && dstr->bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL)
 283                         && srcr->bind & PIPE_BIND_SAMPLER_VIEW)
 284         {
 285                 /* this currently works because we hack the bind flags on resource creation to be
 286                  * the maximum set that the resource type actually supports
 287                  *
 288                  * TODO: perhaps support reinterpreting the formats
 289                  */
 290                 struct blitter_context* blitter = nvfx_get_blitter(pipe, 1);
 291                 util_blitter_copy_texture(blitter, dstr, dst_level, dstx, dsty, dstz, srcr, src_level, src_box, TRUE);
 292                 nvfx_put_blitter(pipe, blitter);
 293         }
 294         else
 295         {
 296                 struct nv04_region dstt = dst;
 297                 struct nv04_region srct = src;
 298                 unsigned dstbegin = 0;
 299
 300                 if(!small)
 301                 {
 302                         if(src_on_gpu)
 303                                 nvfx_region_clone(ctx, &srct, w, h, TRUE);
 304
 305                         if(dst_to_gpu)
 306                                 dstbegin = nvfx_region_clone(ctx, &dstt, w, h, FALSE);
 307                 }
 308
 309                 nv04_region_copy_cpu(&dstt, &srct, w, h);
 310
 311                 if(srct.bo != src.bo)
 312                         nouveau_screen_bo_release(pipe->screen, srct.bo);
 313
 314                 if(dstt.bo != dst.bo)
 315                 {
 316                         nv04_memcpy(ctx, dst.bo, dst.offset + dstbegin, dstt.bo, 0, dstt.bo->size);
 317                         nouveau_screen_bo_release(pipe->screen, dstt.bo);
 318                 }
 319         }
 320 }
 321
 322 static int
 323 nvfx_surface_fill(struct pipe_context* pipe, struct pipe_surface *dsts,
 324                   unsigned dx, unsigned dy, unsigned w, unsigned h, unsigned value)
 325 {
 326         struct nv04_2d_context *ctx = nvfx_screen(pipe->screen)->eng2d;
 327         struct nv04_region dst;
 328         int ret;
 329         /* Always try to use the GPU right now, if possible
 330          * If the user wanted the surface data on the CPU, he would have cleared with memset (hopefully) */
 331
 332         // we don't care about interior pixel order since we set all them to the same value
 333         nvfx_region_init_for_surface(&dst, (struct nvfx_surface*)dsts, dx, dy, TRUE);
 334
 335         w = util_format_get_stride(dsts->format, w) >> dst.bpps;
 336         h = util_format_get_nblocksy(dsts->format, h);
 337
 338         ret = nv04_region_fill_2d(ctx, &dst, w, h, value);
 339         if(ret > 0 && dsts->texture->bind & PIPE_BIND_RENDER_TARGET)
 340                 return 1;
 341         else if(ret)
 342         {
 343                 struct nv04_region dstt = dst;
 344                 unsigned dstbegin = 0;
 345
 346                 if(nvfx_resource_on_gpu(dsts->texture))
 347                         dstbegin = nvfx_region_clone(ctx, &dstt, w, h, FALSE);
 348
 349                 nv04_region_fill_cpu(&dstt, w, h, value);
 350
 351                 if(dstt.bo != dst.bo)
 352                 {
 353                         nv04_memcpy(ctx, dst.bo, dst.offset + dstbegin, dstt.bo, 0, dstt.bo->size);
 354                         nouveau_screen_bo_release(pipe->screen, dstt.bo);
 355                 }
 356         }
 357
 358         return 0;
 359 }
 360
 361
 362 void
 363 nvfx_screen_surface_takedown(struct pipe_screen *pscreen)
 364 {
 365         nv04_2d_context_takedown(nvfx_screen(pscreen)->eng2d);
 366         nvfx_screen(pscreen)->eng2d = 0;
 367 }
 368
 369 int
 370 nvfx_screen_surface_init(struct pipe_screen *pscreen)
 371 {
 372         struct nv04_2d_context* ctx = nv04_2d_context_init(nouveau_screen(pscreen)->channel);
 373         if(!ctx)
 374                 return -1;
 375         nvfx_screen(pscreen)->eng2d = ctx;
 376         return 0;
 377 }
 378
 379 static void
 380 nvfx_surface_copy_temp(struct pipe_context* pipe, struct pipe_surface* surf, int to_temp)
 381 {
 382         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
 383         struct pipe_box box;
 384         struct nvfx_context* nvfx = nvfx_context(pipe);
 385         struct nvfx_miptree* temp;
 386         unsigned use_vertex_buffers;
 387         boolean use_index_buffer;
 388         unsigned base_vertex;
 389
 390         /* temporarily detach the temp, so it isn't used in place of the actual resource */
 391         temp = ns->temp;
 392         ns->temp = 0;
 393
 394         // TODO: we really should do this validation before setting these variable in draw calls
 395         use_vertex_buffers = nvfx->use_vertex_buffers;
 396         use_index_buffer = nvfx->use_index_buffer;
 397         base_vertex = nvfx->base_vertex;
 398
 399         box.x = box.y = 0;
 400         assert(surf->u.tex.first_layer == surf->u.tex.last_layer);
 401         box.width = surf->width;
 402         box.height = surf->height;
 403         box.depth = 1;
 404
 405         if(to_temp) {
 406                 box.z = surf->u.tex.first_layer;
 407                 nvfx_resource_copy_region(pipe, &temp->base.base, 0, 0, 0, 0, surf->texture, surf->u.tex.level, &box);
 408         }
 409         else {
 410                 box.z = 0;
 411                 nvfx_resource_copy_region(pipe, surf->texture, surf->u.tex.level, 0, 0, surf->u.tex.first_layer, &temp->base.base, 0, &box);
 412         }
 413
 414         /* If this triggers, it probably means we attempted to use the blitter
 415          * but failed due to non-renderability of the target.
 416          * Obviously, this would lead to infinite recursion if supported. */
 417         assert(!ns->temp);
 418
 419         ns->temp = temp;
 420
 421         nvfx->use_vertex_buffers = use_vertex_buffers;
 422         nvfx->use_index_buffer = use_index_buffer;
 423         nvfx->base_vertex = base_vertex;
 424
 425         nvfx->dirty |= NVFX_NEW_ARRAYS;
 426         nvfx->draw_dirty |= NVFX_NEW_ARRAYS;
 427 }
 428
 429 void
 430 nvfx_surface_create_temp(struct pipe_context* pipe, struct pipe_surface* surf)
 431 {
 432         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
 433         struct pipe_resource template;
 434         memset(&template, 0, sizeof(struct pipe_resource));
 435         template.target = PIPE_TEXTURE_2D;
 436         template.format = surf->format;
 437         template.width0 = surf->width;
 438         template.height0 = surf->height;
 439         template.depth0 = 1;
 440         template.nr_samples = surf->texture->nr_samples;
 441         template.flags = NOUVEAU_RESOURCE_FLAG_LINEAR;
 442
 443         assert(!ns->temp && !util_dirty_surface_is_dirty(&ns->base));
 444
 445         ns->temp = (struct nvfx_miptree*)nvfx_miptree_create(pipe->screen, &template);
 446         nvfx_surface_copy_temp(pipe, surf, 1);
 447 }
 448
 449 void
 450 nvfx_surface_flush(struct pipe_context* pipe, struct pipe_surface* surf)
 451 {
 452         struct nvfx_context* nvfx = (struct nvfx_context*)pipe;
 453         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
 454         boolean bound = FALSE;
 455
 456         nvfx_surface_copy_temp(pipe, surf, 0);
 457
 458         util_dirty_surface_set_clean(nvfx_surface_get_dirty_surfaces(surf), &ns->base);
 459
 460         if(nvfx->framebuffer.zsbuf == surf)
 461                 bound = TRUE;
 462         else
 463         {
 464                 for(unsigned i = 0; i < nvfx->framebuffer.nr_cbufs; ++i)
 465                 {
 466                         if(nvfx->framebuffer.cbufs[i] == surf)
 467                         {
 468                                 bound = TRUE;
 469                                 break;
 470                         }
 471                 }
 472         }
 473
 474         if(!bound)
 475                 pipe_resource_reference((struct pipe_resource**)&ns->temp, 0);
 476 }
 477
 478 static void
 479 nvfx_clear_render_target(struct pipe_context *pipe,
 480                          struct pipe_surface *dst,
 481                          const float *rgba,
 482                          unsigned dstx, unsigned dsty,
 483                          unsigned width, unsigned height)
 484 {
 485         union util_color uc;
 486         util_pack_color(rgba, dst->format, &uc);
 487
 488         if(util_format_get_blocksizebits(dst->format) > 32
 489                 || nvfx_surface_fill(pipe, dst, dstx, dsty, width, height, uc.ui))
 490         {
 491                 // TODO: probably should use hardware clear here instead if possible
 492                 struct blitter_context* blitter = nvfx_get_blitter(pipe, 0);
 493                 util_blitter_clear_render_target(blitter, dst, rgba, dstx, dsty, width, height);
 494                 nvfx_put_blitter(pipe, blitter);
 495         }
 496 }
 497
 498 static void
 499 nvfx_clear_depth_stencil(struct pipe_context *pipe,
 500                          struct pipe_surface *dst,
 501                          unsigned clear_flags,
 502                          double depth,
 503                          unsigned stencil,
 504                          unsigned dstx, unsigned dsty,
 505                          unsigned width, unsigned height)
 506 {
 507         if(util_format_get_blocksizebits(dst->format) > 32
 508                 || nvfx_surface_fill(pipe, dst, dstx, dsty, width, height, util_pack_z_stencil(dst->format, depth, stencil)))
 509         {
 510                 // TODO: probably should use hardware clear here instead if possible
 511                 struct blitter_context* blitter = nvfx_get_blitter(pipe, 0);
 512                 util_blitter_clear_depth_stencil(blitter, dst, clear_flags, depth, stencil, dstx, dsty, width, height);
 513                 nvfx_put_blitter(pipe, blitter);
 514         }
 515 }
 516
 517
 518 void
 519 nvfx_init_surface_functions(struct nvfx_context *nvfx)
 520 {
 521         nvfx->pipe.resource_copy_region = nvfx_resource_copy_region;
 522         nvfx->pipe.clear_render_target = nvfx_clear_render_target;
 523         nvfx->pipe.clear_depth_stencil = nvfx_clear_depth_stencil;
 524 }