src/gallium/drivers/r600/r600_pipe.c

   1 /*
   2  * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23 #include "r600_pipe.h"
  24 #include "r600_public.h"
  25 #include "r600_isa.h"
  26 #include "evergreen_compute.h"
  27 #include "r600d.h"
  28
  29 #include "sb/sb_public.h"
  30
  31 #include <errno.h>
  32 #include "pipe/p_shader_tokens.h"
  33 #include "util/u_debug.h"
  34 #include "util/u_memory.h"
  35 #include "util/u_simple_shaders.h"
  36 #include "util/u_upload_mgr.h"
  37 #include "util/u_math.h"
  38 #include "vl/vl_decoder.h"
  39 #include "vl/vl_video_buffer.h"
  40 #include "radeon/radeon_video.h"
  41 #include "radeon/radeon_uvd.h"
  42 #include "os/os_time.h"
  43
  44 static const struct debug_named_value r600_debug_options[] = {
  45         /* features */
  46 #if defined(R600_USE_LLVM)
  47         { "llvm", DBG_LLVM, "Enable the LLVM shader compiler" },
  48 #endif
  49         { "nocpdma", DBG_NO_CP_DMA, "Disable CP DMA" },
  50
  51         /* shader backend */
  52         { "nosb", DBG_NO_SB, "Disable sb backend for graphics shaders" },
  53         { "sbcl", DBG_SB_CS, "Enable sb backend for compute shaders" },
  54         { "sbdry", DBG_SB_DRY_RUN, "Don't use optimized bytecode (just print the dumps)" },
  55         { "sbstat", DBG_SB_STAT, "Print optimization statistics for shaders" },
  56         { "sbdump", DBG_SB_DUMP, "Print IR dumps after some optimization passes" },
  57         { "sbnofallback", DBG_SB_NO_FALLBACK, "Abort on errors instead of fallback" },
  58         { "sbdisasm", DBG_SB_DISASM, "Use sb disassembler for shader dumps" },
  59         { "sbsafemath", DBG_SB_SAFEMATH, "Disable unsafe math optimizations" },
  60
  61         DEBUG_NAMED_VALUE_END /* must be last */
  62 };
  63
  64 /*
  65  * pipe_context
  66  */
  67
  68 static void r600_destroy_context(struct pipe_context *context)
  69 {
  70         struct r600_context *rctx = (struct r600_context *)context;
  71
  72         r600_isa_destroy(rctx->isa);
  73
  74         r600_sb_context_destroy(rctx->sb_context);
  75
  76         pipe_resource_reference((struct pipe_resource**)&rctx->dummy_cmask, NULL);
  77         pipe_resource_reference((struct pipe_resource**)&rctx->dummy_fmask, NULL);
  78
  79         if (rctx->dummy_pixel_shader) {
  80                 rctx->b.b.delete_fs_state(&rctx->b.b, rctx->dummy_pixel_shader);
  81         }
  82         if (rctx->custom_dsa_flush) {
  83                 rctx->b.b.delete_depth_stencil_alpha_state(&rctx->b.b, rctx->custom_dsa_flush);
  84         }
  85         if (rctx->custom_blend_resolve) {
  86                 rctx->b.b.delete_blend_state(&rctx->b.b, rctx->custom_blend_resolve);
  87         }
  88         if (rctx->custom_blend_decompress) {
  89                 rctx->b.b.delete_blend_state(&rctx->b.b, rctx->custom_blend_decompress);
  90         }
  91         if (rctx->custom_blend_fastclear) {
  92                 rctx->b.b.delete_blend_state(&rctx->b.b, rctx->custom_blend_fastclear);
  93         }
  94         util_unreference_framebuffer_state(&rctx->framebuffer.state);
  95
  96         if (rctx->blitter) {
  97                 util_blitter_destroy(rctx->blitter);
  98         }
  99         if (rctx->allocator_fetch_shader) {
 100                 u_suballocator_destroy(rctx->allocator_fetch_shader);
 101         }
 102
 103         r600_release_command_buffer(&rctx->start_cs_cmd);
 104
 105         FREE(rctx->start_compute_cs_cmd.buf);
 106
 107         r600_common_context_cleanup(&rctx->b);
 108         FREE(rctx);
 109 }
 110
 111 static struct pipe_context *r600_create_context(struct pipe_screen *screen,
 112                                                 void *priv, unsigned flags)
 113 {
 114         struct r600_context *rctx = CALLOC_STRUCT(r600_context);
 115         struct r600_screen* rscreen = (struct r600_screen *)screen;
 116         struct radeon_winsys *ws = rscreen->b.ws;
 117
 118         if (rctx == NULL)
 119                 return NULL;
 120
 121         rctx->b.b.screen = screen;
 122         rctx->b.b.priv = priv;
 123         rctx->b.b.destroy = r600_destroy_context;
 124         rctx->b.set_atom_dirty = (void *)r600_set_atom_dirty;
 125
 126         if (!r600_common_context_init(&rctx->b, &rscreen->b))
 127                 goto fail;
 128
 129         rctx->screen = rscreen;
 130         rctx->keep_tiling_flags = rscreen->b.info.drm_minor >= 12;
 131
 132         r600_init_blit_functions(rctx);
 133
 134         if (rscreen->b.info.has_uvd) {
 135                 rctx->b.b.create_video_codec = r600_uvd_create_decoder;
 136                 rctx->b.b.create_video_buffer = r600_video_buffer_create;
 137         } else {
 138                 rctx->b.b.create_video_codec = vl_create_decoder;
 139                 rctx->b.b.create_video_buffer = vl_video_buffer_create;
 140         }
 141
 142         r600_init_common_state_functions(rctx);
 143
 144         switch (rctx->b.chip_class) {
 145         case R600:
 146         case R700:
 147                 r600_init_state_functions(rctx);
 148                 r600_init_atom_start_cs(rctx);
 149                 rctx->custom_dsa_flush = r600_create_db_flush_dsa(rctx);
 150                 rctx->custom_blend_resolve = rctx->b.chip_class == R700 ? r700_create_resolve_blend(rctx)
 151                                                                       : r600_create_resolve_blend(rctx);
 152                 rctx->custom_blend_decompress = r600_create_decompress_blend(rctx);
 153                 rctx->has_vertex_cache = !(rctx->b.family == CHIP_RV610 ||
 154                                            rctx->b.family == CHIP_RV620 ||
 155                                            rctx->b.family == CHIP_RS780 ||
 156                                            rctx->b.family == CHIP_RS880 ||
 157                                            rctx->b.family == CHIP_RV710);
 158                 break;
 159         case EVERGREEN:
 160         case CAYMAN:
 161                 evergreen_init_state_functions(rctx);
 162                 evergreen_init_atom_start_cs(rctx);
 163                 evergreen_init_atom_start_compute_cs(rctx);
 164                 rctx->custom_dsa_flush = evergreen_create_db_flush_dsa(rctx);
 165                 rctx->custom_blend_resolve = evergreen_create_resolve_blend(rctx);
 166                 rctx->custom_blend_decompress = evergreen_create_decompress_blend(rctx);
 167                 rctx->custom_blend_fastclear = evergreen_create_fastclear_blend(rctx);
 168                 rctx->has_vertex_cache = !(rctx->b.family == CHIP_CEDAR ||
 169                                            rctx->b.family == CHIP_PALM ||
 170                                            rctx->b.family == CHIP_SUMO ||
 171                                            rctx->b.family == CHIP_SUMO2 ||
 172                                            rctx->b.family == CHIP_CAICOS ||
 173                                            rctx->b.family == CHIP_CAYMAN ||
 174                                            rctx->b.family == CHIP_ARUBA);
 175                 break;
 176         default:
 177                 R600_ERR("Unsupported chip class %d.\n", rctx->b.chip_class);
 178                 goto fail;
 179         }
 180
 181         rctx->b.gfx.cs = ws->cs_create(rctx->b.ctx, RING_GFX,
 182                                        r600_context_gfx_flush, rctx,
 183                                        rscreen->b.trace_bo ?
 184                                                rscreen->b.trace_bo->cs_buf : NULL);
 185         rctx->b.gfx.flush = r600_context_gfx_flush;
 186
 187         rctx->allocator_fetch_shader = u_suballocator_create(&rctx->b.b, 64 * 1024, 256,
 188                                                              0, PIPE_USAGE_DEFAULT, FALSE);
 189         if (!rctx->allocator_fetch_shader)
 190                 goto fail;
 191
 192         rctx->isa = calloc(1, sizeof(struct r600_isa));
 193         if (!rctx->isa || r600_isa_init(rctx, rctx->isa))
 194                 goto fail;
 195
 196         if (rscreen->b.debug_flags & DBG_FORCE_DMA)
 197                 rctx->b.b.resource_copy_region = rctx->b.dma_copy;
 198
 199         rctx->blitter = util_blitter_create(&rctx->b.b);
 200         if (rctx->blitter == NULL)
 201                 goto fail;
 202         util_blitter_set_texture_multisample(rctx->blitter, rscreen->has_msaa);
 203         rctx->blitter->draw_rectangle = r600_draw_rectangle;
 204
 205         r600_begin_new_cs(rctx);
 206         r600_query_init_backend_mask(&rctx->b); /* this emits commands and must be last */
 207
 208         rctx->dummy_pixel_shader =
 209                 util_make_fragment_cloneinput_shader(&rctx->b.b, 0,
 210                                                      TGSI_SEMANTIC_GENERIC,
 211                                                      TGSI_INTERPOLATE_CONSTANT);
 212         rctx->b.b.bind_fs_state(&rctx->b.b, rctx->dummy_pixel_shader);
 213
 214         return &rctx->b.b;
 215
 216 fail:
 217         r600_destroy_context(&rctx->b.b);
 218         return NULL;
 219 }
 220
 221 /*
 222  * pipe_screen
 223  */
 224
 225 static int r600_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
 226 {
 227         struct r600_screen *rscreen = (struct r600_screen *)pscreen;
 228         enum radeon_family family = rscreen->b.family;
 229
 230         switch (param) {
 231         /* Supported features (boolean caps). */
 232         case PIPE_CAP_NPOT_TEXTURES:
 233         case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
 234         case PIPE_CAP_TWO_SIDED_STENCIL:
 235         case PIPE_CAP_ANISOTROPIC_FILTER:
 236         case PIPE_CAP_POINT_SPRITE:
 237         case PIPE_CAP_OCCLUSION_QUERY:
 238         case PIPE_CAP_TEXTURE_SHADOW_MAP:
 239         case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
 240         case PIPE_CAP_BLEND_EQUATION_SEPARATE:
 241         case PIPE_CAP_TEXTURE_SWIZZLE:
 242         case PIPE_CAP_DEPTH_CLIP_DISABLE:
 243         case PIPE_CAP_SHADER_STENCIL_EXPORT:
 244         case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
 245         case PIPE_CAP_MIXED_COLORBUFFER_FORMATS:
 246         case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
 247         case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
 248         case PIPE_CAP_SM3:
 249         case PIPE_CAP_SEAMLESS_CUBE_MAP:
 250         case PIPE_CAP_PRIMITIVE_RESTART:
 251         case PIPE_CAP_CONDITIONAL_RENDER:
 252         case PIPE_CAP_TEXTURE_BARRIER:
 253         case PIPE_CAP_VERTEX_COLOR_UNCLAMPED:
 254         case PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION:
 255         case PIPE_CAP_TGSI_INSTANCEID:
 256         case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
 257         case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
 258         case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
 259         case PIPE_CAP_USER_INDEX_BUFFERS:
 260         case PIPE_CAP_USER_CONSTANT_BUFFERS:
 261         case PIPE_CAP_START_INSTANCE:
 262         case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
 263         case PIPE_CAP_TEXTURE_BUFFER_OBJECTS:
 264         case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
 265         case PIPE_CAP_QUERY_PIPELINE_STATISTICS:
 266         case PIPE_CAP_TEXTURE_MULTISAMPLE:
 267         case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT:
 268         case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION:
 269         case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT:
 270         case PIPE_CAP_SAMPLE_SHADING:
 271         case PIPE_CAP_CLIP_HALFZ:
 272         case PIPE_CAP_POLYGON_OFFSET_CLAMP:
 273         case PIPE_CAP_CONDITIONAL_RENDER_INVERTED:
 274         case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
 275         case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR:
 276         case PIPE_CAP_TGSI_TXQS:
 277                 return 1;
 278
 279         case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
 280                 return rscreen->b.info.drm_major == 2 && rscreen->b.info.drm_minor >= 43;
 281
 282         case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
 283                 return !R600_BIG_ENDIAN && rscreen->b.info.has_userptr;
 284
 285         case PIPE_CAP_COMPUTE:
 286                 return rscreen->b.chip_class > R700;
 287
 288         case PIPE_CAP_TGSI_TEXCOORD:
 289                 return 0;
 290
 291         case PIPE_CAP_FAKE_SW_MSAA:
 292                 return 0;
 293
 294         case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
 295                 return MIN2(rscreen->b.info.vram_size, 0xFFFFFFFF);
 296
 297         case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT:
 298                 return R600_MAP_BUFFER_ALIGNMENT;
 299
 300         case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
 301                 return 256;
 302
 303         case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
 304                 return 1;
 305
 306         case PIPE_CAP_GLSL_FEATURE_LEVEL:
 307                 if (family >= CHIP_CEDAR)
 308                    return 410;
 309                 /* pre-evergreen geom shaders need newer kernel */
 310                 if (rscreen->b.info.drm_minor >= 37)
 311                    return 330;
 312                 return 140;
 313
 314         /* Supported except the original R600. */
 315         case PIPE_CAP_INDEP_BLEND_ENABLE:
 316         case PIPE_CAP_INDEP_BLEND_FUNC:
 317                 /* R600 doesn't support per-MRT blends */
 318                 return family == CHIP_R600 ? 0 : 1;
 319
 320         /* Supported on Evergreen. */
 321         case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
 322         case PIPE_CAP_CUBE_MAP_ARRAY:
 323         case PIPE_CAP_TEXTURE_GATHER_SM5:
 324         case PIPE_CAP_TEXTURE_QUERY_LOD:
 325         case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE:
 326         case PIPE_CAP_SAMPLER_VIEW_TARGET:
 327                 return family >= CHIP_CEDAR ? 1 : 0;
 328         case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
 329                 return family >= CHIP_CEDAR ? 4 : 0;
 330         case PIPE_CAP_DRAW_INDIRECT:
 331                 /* kernel command checker support is also required */
 332                 return family >= CHIP_CEDAR && rscreen->b.info.drm_minor >= 41;
 333
 334         /* Unsupported features. */
 335         case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
 336         case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
 337         case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
 338         case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
 339         case PIPE_CAP_VERTEX_COLOR_CLAMPED:
 340         case PIPE_CAP_USER_VERTEX_BUFFERS:
 341         case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
 342         case PIPE_CAP_VERTEXID_NOBASE:
 343         case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
 344         case PIPE_CAP_DEPTH_BOUNDS_TEST:
 345         case PIPE_CAP_FORCE_PERSAMPLE_INTERP:
 346         case PIPE_CAP_SHAREABLE_SHADERS:
 347         case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS:
 348         case PIPE_CAP_CLEAR_TEXTURE:
 349                 return 0;
 350
 351         /* Stream output. */
 352         case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
 353                 return rscreen->b.has_streamout ? 4 : 0;
 354         case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
 355                 return rscreen->b.has_streamout ? 1 : 0;
 356         case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
 357         case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
 358                 return 32*4;
 359
 360         /* Geometry shader output. */
 361         case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES:
 362                 return 1024;
 363         case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
 364                 return 16384;
 365         case PIPE_CAP_MAX_VERTEX_STREAMS:
 366                 return family >= CHIP_CEDAR ? 4 : 1;
 367
 368         case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE:
 369                 return 2047;
 370
 371         /* Texturing. */
 372         case PIPE_CAP_MAX_TEXTURE_2D_LEVELS:
 373         case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
 374                 if (family >= CHIP_CEDAR)
 375                         return 15;
 376                 else
 377                         return 14;
 378         case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
 379                 /* textures support 8192, but layered rendering supports 2048 */
 380                 return 12;
 381         case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
 382                 /* textures support 8192, but layered rendering supports 2048 */
 383                 return rscreen->b.info.drm_minor >= 9 ? 2048 : 0;
 384
 385         /* Render targets. */
 386         case PIPE_CAP_MAX_RENDER_TARGETS:
 387                 /* XXX some r6xx are buggy and can only do 4 */
 388                 return 8;
 389
 390         case PIPE_CAP_MAX_VIEWPORTS:
 391                 return R600_MAX_VIEWPORTS;
 392
 393         /* Timer queries, present when the clock frequency is non zero. */
 394         case PIPE_CAP_QUERY_TIME_ELAPSED:
 395                 return rscreen->b.info.r600_clock_crystal_freq != 0;
 396         case PIPE_CAP_QUERY_TIMESTAMP:
 397                 return rscreen->b.info.drm_minor >= 20 &&
 398                        rscreen->b.info.r600_clock_crystal_freq != 0;
 399
 400         case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
 401         case PIPE_CAP_MIN_TEXEL_OFFSET:
 402                 return -8;
 403
 404         case PIPE_CAP_MAX_TEXTURE_GATHER_OFFSET:
 405         case PIPE_CAP_MAX_TEXEL_OFFSET:
 406                 return 7;
 407
 408         case PIPE_CAP_TEXTURE_BORDER_COLOR_QUIRK:
 409                 return PIPE_QUIRK_TEXTURE_BORDER_COLOR_SWIZZLE_R600;
 410         case PIPE_CAP_ENDIANNESS:
 411                 return PIPE_ENDIAN_LITTLE;
 412
 413         case PIPE_CAP_VENDOR_ID:
 414                 return 0x1002;
 415         case PIPE_CAP_DEVICE_ID:
 416                 return rscreen->b.info.pci_id;
 417         case PIPE_CAP_ACCELERATED:
 418                 return 1;
 419         case PIPE_CAP_VIDEO_MEMORY:
 420                 return rscreen->b.info.vram_size >> 20;
 421         case PIPE_CAP_UMA:
 422                 return 0;
 423         case PIPE_CAP_MULTISAMPLE_Z_RESOLVE:
 424                 return rscreen->b.chip_class >= R700;
 425         }
 426         return 0;
 427 }
 428
 429 static int r600_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enum pipe_shader_cap param)
 430 {
 431         struct r600_screen *rscreen = (struct r600_screen *)pscreen;
 432
 433         switch(shader)
 434         {
 435         case PIPE_SHADER_FRAGMENT:
 436         case PIPE_SHADER_VERTEX:
 437         case PIPE_SHADER_COMPUTE:
 438                 break;
 439         case PIPE_SHADER_GEOMETRY:
 440                 if (rscreen->b.family >= CHIP_CEDAR)
 441                         break;
 442                 /* pre-evergreen geom shaders need newer kernel */
 443                 if (rscreen->b.info.drm_minor >= 37)
 444                         break;
 445                 return 0;
 446         default:
 447                 /* XXX: support tessellation on Evergreen */
 448                 return 0;
 449         }
 450
 451         switch (param) {
 452         case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
 453         case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
 454         case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
 455         case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
 456                 return 16384;
 457         case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
 458                 return 32;
 459         case PIPE_SHADER_CAP_MAX_INPUTS:
 460                 return shader == PIPE_SHADER_VERTEX ? 16 : 32;
 461         case PIPE_SHADER_CAP_MAX_OUTPUTS:
 462                 return shader == PIPE_SHADER_FRAGMENT ? 8 : 32;
 463         case PIPE_SHADER_CAP_MAX_TEMPS:
 464                 return 256; /* Max native temporaries. */
 465         case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
 466                 if (shader == PIPE_SHADER_COMPUTE) {
 467                         uint64_t max_const_buffer_size;
 468                         pscreen->get_compute_param(pscreen,
 469                                 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
 470                                 &max_const_buffer_size);
 471                         return max_const_buffer_size;
 472
 473                 } else {
 474                         return R600_MAX_CONST_BUFFER_SIZE;
 475                 }
 476         case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
 477                 return R600_MAX_USER_CONST_BUFFERS;
 478         case PIPE_SHADER_CAP_MAX_PREDS:
 479                 return 0; /* nothing uses this */
 480         case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
 481                 return 1;
 482         case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
 483                 return 1;
 484         case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
 485         case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
 486         case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
 487         case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
 488                 return 1;
 489         case PIPE_SHADER_CAP_SUBROUTINES:
 490                 return 0;
 491         case PIPE_SHADER_CAP_INTEGERS:
 492         case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
 493                 return 1;
 494         case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
 495         case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
 496                 return 16;
 497         case PIPE_SHADER_CAP_PREFERRED_IR:
 498                 if (shader == PIPE_SHADER_COMPUTE) {
 499 #if HAVE_LLVM < 0x0306
 500                         return PIPE_SHADER_IR_LLVM;
 501 #else
 502                         return PIPE_SHADER_IR_NATIVE;
 503 #endif
 504                 } else {
 505                         return PIPE_SHADER_IR_TGSI;
 506                 }
 507         case PIPE_SHADER_CAP_DOUBLES:
 508                 if (rscreen->b.family == CHIP_CYPRESS ||
 509                         rscreen->b.family == CHIP_CAYMAN || rscreen->b.family == CHIP_ARUBA)
 510                         return 1;
 511                 return 0;
 512         case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
 513         case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
 514         case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
 515                 return 0;
 516         case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
 517                 /* due to a bug in the shader compiler, some loops hang
 518                  * if they are not unrolled, see:
 519                  *    https://bugs.freedesktop.org/show_bug.cgi?id=86720
 520                  */
 521                 return 255;
 522         }
 523         return 0;
 524 }
 525
 526 static void r600_destroy_screen(struct pipe_screen* pscreen)
 527 {
 528         struct r600_screen *rscreen = (struct r600_screen *)pscreen;
 529
 530         if (rscreen == NULL)
 531                 return;
 532
 533         if (!rscreen->b.ws->unref(rscreen->b.ws))
 534                 return;
 535
 536         if (rscreen->global_pool) {
 537                 compute_memory_pool_delete(rscreen->global_pool);
 538         }
 539
 540         r600_destroy_common_screen(&rscreen->b);
 541 }
 542
 543 static struct pipe_resource *r600_resource_create(struct pipe_screen *screen,
 544                                                   const struct pipe_resource *templ)
 545 {
 546         if (templ->target == PIPE_BUFFER &&
 547             (templ->bind & PIPE_BIND_GLOBAL))
 548                 return r600_compute_global_buffer_create(screen, templ);
 549
 550         return r600_resource_create_common(screen, templ);
 551 }
 552
 553 struct pipe_screen *r600_screen_create(struct radeon_winsys *ws)
 554 {
 555         struct r600_screen *rscreen = CALLOC_STRUCT(r600_screen);
 556
 557         if (rscreen == NULL) {
 558                 return NULL;
 559         }
 560
 561         /* Set functions first. */
 562         rscreen->b.b.context_create = r600_create_context;
 563         rscreen->b.b.destroy = r600_destroy_screen;
 564         rscreen->b.b.get_param = r600_get_param;
 565         rscreen->b.b.get_shader_param = r600_get_shader_param;
 566         rscreen->b.b.resource_create = r600_resource_create;
 567
 568         if (!r600_common_screen_init(&rscreen->b, ws)) {
 569                 FREE(rscreen);
 570                 return NULL;
 571         }
 572
 573         if (rscreen->b.info.chip_class >= EVERGREEN) {
 574                 rscreen->b.b.is_format_supported = evergreen_is_format_supported;
 575         } else {
 576                 rscreen->b.b.is_format_supported = r600_is_format_supported;
 577         }
 578
 579         rscreen->b.debug_flags |= debug_get_flags_option("R600_DEBUG", r600_debug_options, 0);
 580         if (debug_get_bool_option("R600_DEBUG_COMPUTE", FALSE))
 581                 rscreen->b.debug_flags |= DBG_COMPUTE;
 582         if (debug_get_bool_option("R600_DUMP_SHADERS", FALSE))
 583                 rscreen->b.debug_flags |= DBG_FS | DBG_VS | DBG_GS | DBG_PS | DBG_CS;
 584         if (!debug_get_bool_option("R600_HYPERZ", TRUE))
 585                 rscreen->b.debug_flags |= DBG_NO_HYPERZ;
 586         if (debug_get_bool_option("R600_LLVM", FALSE))
 587                 rscreen->b.debug_flags |= DBG_LLVM;
 588
 589         if (rscreen->b.family == CHIP_UNKNOWN) {
 590                 fprintf(stderr, "r600: Unknown chipset 0x%04X\n", rscreen->b.info.pci_id);
 591                 FREE(rscreen);
 592                 return NULL;
 593         }
 594
 595         /* Figure out streamout kernel support. */
 596         switch (rscreen->b.chip_class) {
 597         case R600:
 598                 if (rscreen->b.family < CHIP_RS780) {
 599                         rscreen->b.has_streamout = rscreen->b.info.drm_minor >= 14;
 600                 } else {
 601                         rscreen->b.has_streamout = rscreen->b.info.drm_minor >= 23;
 602                 }
 603                 break;
 604         case R700:
 605                 rscreen->b.has_streamout = rscreen->b.info.drm_minor >= 17;
 606                 break;
 607         case EVERGREEN:
 608         case CAYMAN:
 609                 rscreen->b.has_streamout = rscreen->b.info.drm_minor >= 14;
 610                 break;
 611         default:
 612                 rscreen->b.has_streamout = FALSE;
 613                 break;
 614         }
 615
 616         /* MSAA support. */
 617         switch (rscreen->b.chip_class) {
 618         case R600:
 619         case R700:
 620                 rscreen->has_msaa = rscreen->b.info.drm_minor >= 22;
 621                 rscreen->has_compressed_msaa_texturing = false;
 622                 break;
 623         case EVERGREEN:
 624                 rscreen->has_msaa = rscreen->b.info.drm_minor >= 19;
 625                 rscreen->has_compressed_msaa_texturing = rscreen->b.info.drm_minor >= 24;
 626                 break;
 627         case CAYMAN:
 628                 rscreen->has_msaa = rscreen->b.info.drm_minor >= 19;
 629                 rscreen->has_compressed_msaa_texturing = true;
 630                 break;
 631         default:
 632                 rscreen->has_msaa = FALSE;
 633                 rscreen->has_compressed_msaa_texturing = false;
 634         }
 635
 636         rscreen->b.has_cp_dma = rscreen->b.info.drm_minor >= 27 &&
 637                               !(rscreen->b.debug_flags & DBG_NO_CP_DMA);
 638
 639         rscreen->global_pool = compute_memory_pool_new(rscreen);
 640
 641         /* Create the auxiliary context. This must be done last. */
 642         rscreen->b.aux_context = rscreen->b.b.context_create(&rscreen->b.b, NULL, 0);
 643
 644 #if 0 /* This is for testing whether aux_context and buffer clearing work correctly. */
 645         struct pipe_resource templ = {};
 646
 647         templ.width0 = 4;
 648         templ.height0 = 2048;
 649         templ.depth0 = 1;
 650         templ.array_size = 1;
 651         templ.target = PIPE_TEXTURE_2D;
 652         templ.format = PIPE_FORMAT_R8G8B8A8_UNORM;
 653         templ.usage = PIPE_USAGE_DEFAULT;
 654
 655         struct r600_resource *res = r600_resource(rscreen->screen.resource_create(&rscreen->screen, &templ));
 656         unsigned char *map = ws->buffer_map(res->cs_buf, NULL, PIPE_TRANSFER_WRITE);
 657
 658         memset(map, 0, 256);
 659
 660         r600_screen_clear_buffer(rscreen, &res->b.b, 4, 4, 0xCC);
 661         r600_screen_clear_buffer(rscreen, &res->b.b, 8, 4, 0xDD);
 662         r600_screen_clear_buffer(rscreen, &res->b.b, 12, 4, 0xEE);
 663         r600_screen_clear_buffer(rscreen, &res->b.b, 20, 4, 0xFF);
 664         r600_screen_clear_buffer(rscreen, &res->b.b, 32, 20, 0x87);
 665
 666         ws->buffer_wait(res->buf, RADEON_USAGE_WRITE);
 667
 668         int i;
 669         for (i = 0; i < 256; i++) {
 670                 printf("%02X", map[i]);
 671                 if (i % 16 == 15)
 672                         printf("\n");
 673         }
 674 #endif
 675
 676         return &rscreen->b.b;
 677 }