src/gallium/drivers/llvmpipe/lp_screen.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2008 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28
  29 #include "util/u_memory.h"
  30 #include "util/u_math.h"
  31 #include "util/u_cpu_detect.h"
  32 #include "util/format/u_format.h"
  33 #include "util/u_screen.h"
  34 #include "util/u_string.h"
  35 #include "util/format/u_format_s3tc.h"
  36 #include "pipe/p_defines.h"
  37 #include "pipe/p_screen.h"
  38 #include "draw/draw_context.h"
  39 #include "gallivm/lp_bld_type.h"
  40 #include "gallivm/lp_bld_nir.h"
  41 #include "util/disk_cache.h"
  42 #include "util/os_misc.h"
  43 #include "util/os_time.h"
  44 #include "lp_texture.h"
  45 #include "lp_fence.h"
  46 #include "lp_jit.h"
  47 #include "lp_screen.h"
  48 #include "lp_context.h"
  49 #include "lp_debug.h"
  50 #include "lp_public.h"
  51 #include "lp_limits.h"
  52 #include "lp_rast.h"
  53 #include "lp_cs_tpool.h"
  54
  55 #include "frontend/sw_winsys.h"
  56
  57 #include "nir.h"
  58
  59 #ifdef DEBUG
  60 int LP_DEBUG = 0;
  61
  62 static const struct debug_named_value lp_debug_flags[] = {
  63    { "pipe",   DEBUG_PIPE, NULL },
  64    { "tgsi",   DEBUG_TGSI, NULL },
  65    { "tex",    DEBUG_TEX, NULL },
  66    { "setup",  DEBUG_SETUP, NULL },
  67    { "rast",   DEBUG_RAST, NULL },
  68    { "query",  DEBUG_QUERY, NULL },
  69    { "screen", DEBUG_SCREEN, NULL },
  70    { "counters", DEBUG_COUNTERS, NULL },
  71    { "scene", DEBUG_SCENE, NULL },
  72    { "fence", DEBUG_FENCE, NULL },
  73    { "mem", DEBUG_MEM, NULL },
  74    { "fs", DEBUG_FS, NULL },
  75    { "cs", DEBUG_CS, NULL },
  76    { "tgsi_ir", DEBUG_TGSI_IR, NULL },
  77    { "cl", DEBUG_CL, NULL },
  78    { "cache_stats", DEBUG_CACHE_STATS, NULL },
  79    DEBUG_NAMED_VALUE_END
  80 };
  81 #endif
  82
  83 int LP_PERF = 0;
  84 static const struct debug_named_value lp_perf_flags[] = {
  85    { "texmem",         PERF_TEX_MEM, NULL },
  86    { "no_mipmap",      PERF_NO_MIPMAPS, NULL },
  87    { "no_linear",      PERF_NO_LINEAR, NULL },
  88    { "no_mip_linear",  PERF_NO_MIP_LINEAR, NULL },
  89    { "no_tex",         PERF_NO_TEX, NULL },
  90    { "no_blend",       PERF_NO_BLEND, NULL },
  91    { "no_depth",       PERF_NO_DEPTH, NULL },
  92    { "no_alphatest",   PERF_NO_ALPHATEST, NULL },
  93    DEBUG_NAMED_VALUE_END
  94 };
  95
  96
  97 static const char *
  98 llvmpipe_get_vendor(struct pipe_screen *screen)
  99 {
 100    return "Mesa/X.org";
 101 }
 102
 103
 104 static const char *
 105 llvmpipe_get_name(struct pipe_screen *screen)
 106 {
 107    static char buf[100];
 108    snprintf(buf, sizeof(buf), "llvmpipe (LLVM " MESA_LLVM_VERSION_STRING ", %u bits)",
 109             lp_native_vector_width );
 110    return buf;
 111 }
 112
 113
 114 static int
 115 llvmpipe_get_param(struct pipe_screen *screen, enum pipe_cap param)
 116 {
 117    switch (param) {
 118    case PIPE_CAP_NPOT_TEXTURES:
 119    case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
 120    case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
 121       return 1;
 122    case PIPE_CAP_FRAGMENT_SHADER_TEXTURE_LOD:
 123    case PIPE_CAP_FRAGMENT_SHADER_DERIVATIVES:
 124    case PIPE_CAP_VERTEX_SHADER_SATURATE:
 125       return 1;
 126    case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
 127       return 1;
 128    case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
 129       return PIPE_MAX_SO_BUFFERS;
 130    case PIPE_CAP_ANISOTROPIC_FILTER:
 131       return 0;
 132    case PIPE_CAP_POINT_SPRITE:
 133       return 1;
 134    case PIPE_CAP_MAX_RENDER_TARGETS:
 135       return PIPE_MAX_COLOR_BUFS;
 136    case PIPE_CAP_OCCLUSION_QUERY:
 137       return 1;
 138    case PIPE_CAP_QUERY_TIME_ELAPSED:
 139       return 0;
 140    case PIPE_CAP_QUERY_TIMESTAMP:
 141       return 1;
 142    case PIPE_CAP_QUERY_PIPELINE_STATISTICS:
 143       return 1;
 144    case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
 145    case PIPE_CAP_TEXTURE_MIRROR_CLAMP_TO_EDGE:
 146       return 1;
 147    case PIPE_CAP_TEXTURE_SWIZZLE:
 148       return 1;
 149    case PIPE_CAP_TEXTURE_BORDER_COLOR_QUIRK:
 150    case PIPE_CAP_MAX_TEXTURE_UPLOAD_MEMORY_BUDGET:
 151       return 0;
 152    case PIPE_CAP_MAX_TEXTURE_2D_SIZE:
 153       return 1 << (LP_MAX_TEXTURE_2D_LEVELS - 1);
 154    case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
 155       return LP_MAX_TEXTURE_3D_LEVELS;
 156    case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
 157       return LP_MAX_TEXTURE_CUBE_LEVELS;
 158    case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
 159       return LP_MAX_TEXTURE_ARRAY_LAYERS;
 160    case PIPE_CAP_BLEND_EQUATION_SEPARATE:
 161       return 1;
 162    case PIPE_CAP_INDEP_BLEND_ENABLE:
 163       return 1;
 164    case PIPE_CAP_INDEP_BLEND_FUNC:
 165       return 1;
 166    case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
 167    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
 168    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
 169       return 1;
 170    case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
 171    case PIPE_CAP_DEPTH_CLIP_DISABLE_SEPARATE:
 172       return 0;
 173    case PIPE_CAP_PRIMITIVE_RESTART:
 174    case PIPE_CAP_PRIMITIVE_RESTART_FIXED_INDEX:
 175       return 1;
 176    case PIPE_CAP_DEPTH_CLIP_DISABLE:
 177       return 1;
 178    case PIPE_CAP_SHADER_STENCIL_EXPORT:
 179       return 1;
 180    case PIPE_CAP_TGSI_INSTANCEID:
 181    case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
 182    case PIPE_CAP_START_INSTANCE:
 183       return 1;
 184    case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
 185       return 0;
 186    case PIPE_CAP_MIXED_COLORBUFFER_FORMATS:
 187       return 1;
 188    case PIPE_CAP_SEAMLESS_CUBE_MAP:
 189    case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
 190       return 1;
 191    /* this is a lie could support arbitrary large offsets */
 192    case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
 193    case PIPE_CAP_MIN_TEXEL_OFFSET:
 194       return -32;
 195    case PIPE_CAP_MAX_TEXTURE_GATHER_OFFSET:
 196    case PIPE_CAP_MAX_TEXEL_OFFSET:
 197       return 31;
 198    case PIPE_CAP_CONDITIONAL_RENDER:
 199    case PIPE_CAP_TEXTURE_BARRIER:
 200       return 1;
 201    case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
 202    case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
 203       return 16*4;
 204    case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES:
 205    case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
 206       return 1024;
 207    case PIPE_CAP_MAX_VERTEX_STREAMS: {
 208       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 209       return lscreen->use_tgsi ? 1 : 4;
 210    }
 211    case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE:
 212       return 2048;
 213    case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
 214    case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
 215       return 1;
 216    case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
 217       return 0;
 218    case PIPE_CAP_VERTEX_COLOR_UNCLAMPED:
 219    case PIPE_CAP_VERTEX_COLOR_CLAMPED:
 220       return 1;
 221    case PIPE_CAP_GLSL_FEATURE_LEVEL: {
 222       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 223       return lscreen->use_tgsi ? 330 : 420;
 224    }
 225    case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY:
 226       return 140;
 227    case PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION:
 228       return 0;
 229    case PIPE_CAP_COMPUTE:
 230       return GALLIVM_HAVE_CORO;
 231    case PIPE_CAP_USER_VERTEX_BUFFERS:
 232       return 1;
 233    case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
 234    case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
 235    case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
 236       return 0;
 237    case PIPE_CAP_TGSI_TEXCOORD:
 238    case PIPE_CAP_DRAW_INDIRECT:
 239       return 1;
 240
 241    case PIPE_CAP_CUBE_MAP_ARRAY:
 242       return 1;
 243    case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
 244       return 16;
 245    case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT:
 246       return 64;
 247    case PIPE_CAP_TEXTURE_BUFFER_OBJECTS:
 248       return 1;
 249    case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
 250       return 65536;
 251    case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
 252       return 16;
 253    case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
 254       return 0;
 255    case PIPE_CAP_MAX_VIEWPORTS:
 256       return PIPE_MAX_VIEWPORTS;
 257    case PIPE_CAP_ENDIANNESS:
 258       return PIPE_ENDIAN_NATIVE;
 259    case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT:
 260       return 1;
 261    case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT:
 262       return 1;
 263    case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
 264       return 4;
 265    case PIPE_CAP_TEXTURE_GATHER_SM5:
 266    case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
 267       return 0;
 268    case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION:
 269       return 1;
 270    case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE:
 271       return 1;
 272    case PIPE_CAP_TGSI_TEX_TXF_LZ:
 273    case PIPE_CAP_SAMPLER_VIEW_TARGET:
 274       return 1;
 275    case PIPE_CAP_FAKE_SW_MSAA: {
 276       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 277       return lscreen->use_tgsi ? 1 : 0;
 278    }
 279    case PIPE_CAP_TEXTURE_QUERY_LOD:
 280    case PIPE_CAP_CONDITIONAL_RENDER_INVERTED:
 281    case PIPE_CAP_TGSI_ARRAY_COMPONENTS:
 282    case PIPE_CAP_DOUBLES:
 283    case PIPE_CAP_INT64:
 284    case PIPE_CAP_INT64_DIVMOD:
 285    case PIPE_CAP_QUERY_SO_OVERFLOW:
 286    case PIPE_CAP_TGSI_DIV:
 287       return 1;
 288
 289    case PIPE_CAP_VENDOR_ID:
 290       return 0xFFFFFFFF;
 291    case PIPE_CAP_DEVICE_ID:
 292       return 0xFFFFFFFF;
 293    case PIPE_CAP_ACCELERATED:
 294       return 0;
 295    case PIPE_CAP_VIDEO_MEMORY: {
 296       /* XXX: Do we want to return the full amount fo system memory ? */
 297       uint64_t system_memory;
 298
 299       if (!os_get_total_physical_memory(&system_memory))
 300          return 0;
 301
 302       if (sizeof(void *) == 4)
 303          /* Cap to 2 GB on 32 bits system. We do this because llvmpipe does
 304           * eat application memory, which is quite limited on 32 bits. App
 305           * shouldn't expect too much available memory. */
 306          system_memory = MIN2(system_memory, 2048 << 20);
 307
 308       return (int)(system_memory >> 20);
 309    }
 310    case PIPE_CAP_UMA:
 311       return 0;
 312    case PIPE_CAP_CLIP_HALFZ:
 313       return 1;
 314    case PIPE_CAP_VERTEXID_NOBASE:
 315       return 0;
 316    case PIPE_CAP_POLYGON_OFFSET_CLAMP:
 317    case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
 318    case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR:
 319       return 1;
 320    case PIPE_CAP_CULL_DISTANCE:
 321       return 1;
 322    case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS:
 323       return 1;
 324    case PIPE_CAP_CLEAR_TEXTURE:
 325       return 1;
 326    case PIPE_CAP_MAX_VARYINGS:
 327       return 32;
 328    case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
 329       return 16;
 330    case PIPE_CAP_QUERY_BUFFER_OBJECT:
 331       return 1;
 332    case PIPE_CAP_DRAW_PARAMETERS:
 333       return 1;
 334    case PIPE_CAP_MULTI_DRAW_INDIRECT:
 335    case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
 336       return 1;
 337    case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
 338       return 32;
 339    case PIPE_CAP_RASTERIZER_SUBPIXEL_BITS:
 340       return 8;
 341    case PIPE_CAP_MULTISAMPLE_Z_RESOLVE:
 342    case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
 343    case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
 344    case PIPE_CAP_DEPTH_BOUNDS_TEST:
 345    case PIPE_CAP_FORCE_PERSAMPLE_INTERP:
 346    case PIPE_CAP_SHAREABLE_SHADERS:
 347    case PIPE_CAP_TGSI_PACK_HALF_FLOAT:
 348    case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
 349    case PIPE_CAP_INVALIDATE_BUFFER:
 350    case PIPE_CAP_GENERATE_MIPMAP:
 351    case PIPE_CAP_STRING_MARKER:
 352    case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
 353    case PIPE_CAP_SURFACE_REINTERPRET_BLOCKS:
 354    case PIPE_CAP_QUERY_MEMORY_INFO:
 355    case PIPE_CAP_PCI_GROUP:
 356    case PIPE_CAP_PCI_BUS:
 357    case PIPE_CAP_PCI_DEVICE:
 358    case PIPE_CAP_PCI_FUNCTION:
 359    case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR:
 360    case PIPE_CAP_PRIMITIVE_RESTART_FOR_PATCHES:
 361    case PIPE_CAP_MAX_WINDOW_RECTANGLES:
 362    case PIPE_CAP_POLYGON_OFFSET_UNITS_UNSCALED:
 363    case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS:
 364    case PIPE_CAP_TGSI_CAN_READ_OUTPUTS:
 365    case PIPE_CAP_NATIVE_FENCE_FD:
 366    case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY:
 367    case PIPE_CAP_FBFETCH:
 368    case PIPE_CAP_TGSI_MUL_ZERO_WINS:
 369    case PIPE_CAP_TGSI_CLOCK:
 370    case PIPE_CAP_POLYGON_MODE_FILL_RECTANGLE:
 371    case PIPE_CAP_SPARSE_BUFFER_PAGE_SIZE:
 372    case PIPE_CAP_TGSI_BALLOT:
 373    case PIPE_CAP_TGSI_TES_LAYER_VIEWPORT:
 374    case PIPE_CAP_CAN_BIND_CONST_BUFFER_AS_VERTEX:
 375    case PIPE_CAP_ALLOW_MAPPED_BUFFERS_DURING_EXECUTION:
 376    case PIPE_CAP_BINDLESS_TEXTURE:
 377    case PIPE_CAP_NIR_SAMPLERS_AS_DEREF:
 378    case PIPE_CAP_MEMOBJ:
 379    case PIPE_CAP_TGSI_ANY_REG_AS_ADDRESS:
 380    case PIPE_CAP_TILE_RASTER_ORDER:
 381    case PIPE_CAP_MAX_COMBINED_SHADER_OUTPUT_RESOURCES:
 382    case PIPE_CAP_FRAMEBUFFER_MSAA_CONSTRAINTS:
 383    case PIPE_CAP_SIGNED_VERTEX_BUFFER_OFFSET:
 384    case PIPE_CAP_CONTEXT_PRIORITY_MASK:
 385    case PIPE_CAP_FENCE_SIGNAL:
 386    case PIPE_CAP_CONSTBUF0_FLAGS:
 387    case PIPE_CAP_CONSERVATIVE_RASTER_POST_SNAP_TRIANGLES:
 388    case PIPE_CAP_CONSERVATIVE_RASTER_POST_SNAP_POINTS_LINES:
 389    case PIPE_CAP_CONSERVATIVE_RASTER_PRE_SNAP_TRIANGLES:
 390    case PIPE_CAP_CONSERVATIVE_RASTER_PRE_SNAP_POINTS_LINES:
 391    case PIPE_CAP_CONSERVATIVE_RASTER_POST_DEPTH_COVERAGE:
 392    case PIPE_CAP_MAX_CONSERVATIVE_RASTER_SUBPIXEL_PRECISION_BIAS:
 393    case PIPE_CAP_PROGRAMMABLE_SAMPLE_LOCATIONS:
 394       return 0;
 395    case PIPE_CAP_MAX_GS_INVOCATIONS:
 396       return 32;
 397    case PIPE_CAP_MAX_SHADER_BUFFER_SIZE:
 398       return LP_MAX_TGSI_SHADER_BUFFER_SIZE;
 399    case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT:
 400    case PIPE_CAP_TGSI_TG4_COMPONENT_IN_SWIZZLE:
 401    case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
 402       return 1;
 403    case PIPE_CAP_TGSI_TXQS:
 404    case PIPE_CAP_TGSI_VOTE:
 405    case PIPE_CAP_LOAD_CONSTBUF:
 406    case PIPE_CAP_TEXTURE_MULTISAMPLE:
 407    case PIPE_CAP_SAMPLE_SHADING:
 408    case PIPE_CAP_POST_DEPTH_COVERAGE:
 409    case PIPE_CAP_PACKED_UNIFORMS: {
 410       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 411       return !lscreen->use_tgsi;
 412    }
 413    default:
 414       return u_pipe_screen_get_param_defaults(screen, param);
 415    }
 416 }
 417
 418 static int
 419 llvmpipe_get_shader_param(struct pipe_screen *screen,
 420                           enum pipe_shader_type shader,
 421                           enum pipe_shader_cap param)
 422 {
 423    struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 424    switch(shader)
 425    {
 426    case PIPE_SHADER_COMPUTE:
 427       if ((LP_DEBUG & DEBUG_CL) && param == PIPE_SHADER_CAP_SUPPORTED_IRS)
 428          return (1 << PIPE_SHADER_IR_TGSI) | (1 << PIPE_SHADER_IR_NIR) | (1 << PIPE_SHADER_IR_NIR_SERIALIZED);
 429       /* fallthrough */
 430    case PIPE_SHADER_FRAGMENT:
 431       if (param == PIPE_SHADER_CAP_PREFERRED_IR) {
 432          if (lscreen->use_tgsi)
 433             return PIPE_SHADER_IR_TGSI;
 434          else
 435             return PIPE_SHADER_IR_NIR;
 436       }
 437       switch (param) {
 438       default:
 439          return gallivm_get_shader_param(param);
 440       }
 441       /* fallthrough */
 442    case PIPE_SHADER_TESS_CTRL:
 443    case PIPE_SHADER_TESS_EVAL:
 444       /* Tessellation shader needs llvm coroutines support */
 445       if (!GALLIVM_HAVE_CORO || lscreen->use_tgsi)
 446          return 0;
 447       /* fallthrough */
 448    case PIPE_SHADER_VERTEX:
 449    case PIPE_SHADER_GEOMETRY:
 450       if (param == PIPE_SHADER_CAP_PREFERRED_IR) {
 451          if (lscreen->use_tgsi)
 452             return PIPE_SHADER_IR_TGSI;
 453          else
 454             return PIPE_SHADER_IR_NIR;
 455       }
 456
 457       switch (param) {
 458       case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
 459          /* At this time, the draw module and llvmpipe driver only
 460           * support vertex shader texture lookups when LLVM is enabled in
 461           * the draw module.
 462           */
 463          if (debug_get_bool_option("DRAW_USE_LLVM", TRUE))
 464             return PIPE_MAX_SAMPLERS;
 465          else
 466             return 0;
 467       case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
 468          if (debug_get_bool_option("DRAW_USE_LLVM", TRUE))
 469             return PIPE_MAX_SHADER_SAMPLER_VIEWS;
 470          else
 471             return 0;
 472       default:
 473          return draw_get_shader_param(shader, param);
 474       }
 475    default:
 476       return 0;
 477    }
 478 }
 479
 480 static float
 481 llvmpipe_get_paramf(struct pipe_screen *screen, enum pipe_capf param)
 482 {
 483    switch (param) {
 484    case PIPE_CAPF_MAX_LINE_WIDTH:
 485       /* fall-through */
 486    case PIPE_CAPF_MAX_LINE_WIDTH_AA:
 487       return 255.0; /* arbitrary */
 488    case PIPE_CAPF_MAX_POINT_WIDTH:
 489       /* fall-through */
 490    case PIPE_CAPF_MAX_POINT_WIDTH_AA:
 491       return LP_MAX_POINT_WIDTH; /* arbitrary */
 492    case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
 493       return 16.0; /* not actually signficant at this time */
 494    case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
 495       return 16.0; /* arbitrary */
 496    case PIPE_CAPF_MIN_CONSERVATIVE_RASTER_DILATE:
 497       return 0.0;
 498    case PIPE_CAPF_MAX_CONSERVATIVE_RASTER_DILATE:
 499       return 0.0;
 500    case PIPE_CAPF_CONSERVATIVE_RASTER_DILATE_GRANULARITY:
 501       return 0.0;
 502    }
 503    /* should only get here on unhandled cases */
 504    debug_printf("Unexpected PIPE_CAP %d query\n", param);
 505    return 0.0;
 506 }
 507
 508 static int
 509 llvmpipe_get_compute_param(struct pipe_screen *_screen,
 510                            enum pipe_shader_ir ir_type,
 511                            enum pipe_compute_cap param,
 512                            void *ret)
 513 {
 514    switch (param) {
 515    case PIPE_COMPUTE_CAP_IR_TARGET:
 516       return 0;
 517    case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
 518       if (ret) {
 519          uint64_t *grid_size = ret;
 520          grid_size[0] = 65535;
 521          grid_size[1] = 65535;
 522          grid_size[2] = 65535;
 523       }
 524       return 3 * sizeof(uint64_t) ;
 525    case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
 526       if (ret) {
 527          uint64_t *block_size = ret;
 528          block_size[0] = 1024;
 529          block_size[1] = 1024;
 530          block_size[2] = 1024;
 531       }
 532       return 3 * sizeof(uint64_t);
 533    case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
 534       if (ret) {
 535          uint64_t *max_threads_per_block = ret;
 536          *max_threads_per_block = 1024;
 537       }
 538       return sizeof(uint64_t);
 539    case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
 540       if (ret) {
 541          uint64_t *max_local_size = ret;
 542          *max_local_size = 32768;
 543       }
 544       return sizeof(uint64_t);
 545    case PIPE_COMPUTE_CAP_GRID_DIMENSION:
 546       if (ret) {
 547          uint32_t *grid_dim = ret;
 548          *grid_dim = 3;
 549       }
 550       return sizeof(uint32_t);
 551    case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
 552       if (ret) {
 553          uint64_t *max_global_size = ret;
 554          *max_global_size = (1ULL << 31);
 555       }
 556       return sizeof(uint64_t);
 557    case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
 558       if (ret) {
 559          uint64_t *max_mem_alloc_size = ret;
 560          *max_mem_alloc_size = (1ULL << 31);
 561       }
 562       return sizeof(uint64_t);
 563    case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
 564       if (ret) {
 565          uint64_t *max_private = ret;
 566          *max_private = (1UL << 31);
 567       }
 568       return sizeof(uint64_t);
 569    case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
 570       if (ret) {
 571          uint64_t *max_input = ret;
 572          *max_input = 4096;
 573       }
 574       return sizeof(uint64_t);
 575    case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
 576       if (ret) {
 577          uint32_t *images = ret;
 578          *images = 0;
 579       }
 580       return sizeof(uint32_t);
 581    case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
 582       return 0;
 583    case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
 584       if (ret) {
 585          uint32_t *subgroup_size = ret;
 586          *subgroup_size = 32;
 587       }
 588       return sizeof(uint32_t);
 589    case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
 590       if (ret) {
 591          uint32_t *max_compute_units = ret;
 592          *max_compute_units = 8;
 593       }
 594       return sizeof(uint32_t);
 595    case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
 596       if (ret) {
 597          uint32_t *max_clock_freq = ret;
 598          *max_clock_freq = 300;
 599       }
 600       return sizeof(uint32_t);
 601    case PIPE_COMPUTE_CAP_ADDRESS_BITS:
 602       if (ret) {
 603          uint32_t *address_bits = ret;
 604          *address_bits = 64;
 605       }
 606       return sizeof(uint32_t);
 607    }
 608    return 0;
 609 }
 610
 611 static const struct nir_shader_compiler_options gallivm_nir_options = {
 612    .lower_scmp = true,
 613    .lower_flrp32 = true,
 614    .lower_flrp64 = true,
 615    .lower_fsat = true,
 616    .lower_bitfield_insert_to_shifts = true,
 617    .lower_bitfield_extract_to_shifts = true,
 618    .lower_sub = true,
 619    .lower_ffma = true,
 620    .lower_fmod = true,
 621    .lower_hadd = true,
 622    .lower_add_sat = true,
 623    .lower_pack_snorm_2x16 = true,
 624    .lower_pack_snorm_4x8 = true,
 625    .lower_pack_unorm_2x16 = true,
 626    .lower_pack_unorm_4x8 = true,
 627    .lower_unpack_snorm_2x16 = true,
 628    .lower_unpack_snorm_4x8 = true,
 629    .lower_unpack_unorm_2x16 = true,
 630    .lower_unpack_unorm_4x8 = true,
 631    .lower_extract_byte = true,
 632    .lower_extract_word = true,
 633    .lower_rotate = true,
 634    .lower_ifind_msb = true,
 635    .max_unroll_iterations = 32,
 636    .use_interpolated_input_intrinsics = true,
 637    .lower_to_scalar = true,
 638 };
 639
 640 static void
 641 llvmpipe_finalize_nir(struct pipe_screen *screen,
 642                       void *nirptr,
 643                       bool optimize)
 644 {
 645    struct nir_shader *nir = (struct nir_shader *)nirptr;
 646    lp_build_opt_nir(nir);
 647 }
 648
 649 static inline const void *
 650 llvmpipe_get_compiler_options(struct pipe_screen *screen,
 651                               enum pipe_shader_ir ir,
 652                               enum pipe_shader_type shader)
 653 {
 654    assert(ir == PIPE_SHADER_IR_NIR);
 655    return &gallivm_nir_options;
 656 }
 657
 658 /**
 659  * Query format support for creating a texture, drawing surface, etc.
 660  * \param format  the format to test
 661  * \param type  one of PIPE_TEXTURE, PIPE_SURFACE
 662  */
 663 static bool
 664 llvmpipe_is_format_supported( struct pipe_screen *_screen,
 665                               enum pipe_format format,
 666                               enum pipe_texture_target target,
 667                               unsigned sample_count,
 668                               unsigned storage_sample_count,
 669                               unsigned bind)
 670 {
 671    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 672    struct sw_winsys *winsys = screen->winsys;
 673    const struct util_format_description *format_desc;
 674
 675    format_desc = util_format_description(format);
 676    if (!format_desc)
 677       return false;
 678
 679    assert(target == PIPE_BUFFER ||
 680           target == PIPE_TEXTURE_1D ||
 681           target == PIPE_TEXTURE_1D_ARRAY ||
 682           target == PIPE_TEXTURE_2D ||
 683           target == PIPE_TEXTURE_2D_ARRAY ||
 684           target == PIPE_TEXTURE_RECT ||
 685           target == PIPE_TEXTURE_3D ||
 686           target == PIPE_TEXTURE_CUBE ||
 687           target == PIPE_TEXTURE_CUBE_ARRAY);
 688
 689    if (sample_count != 0 && sample_count != 1 && sample_count != 4)
 690       return false;
 691
 692    if (MAX2(1, sample_count) != MAX2(1, storage_sample_count))
 693       return false;
 694
 695    if (bind & PIPE_BIND_RENDER_TARGET) {
 696       if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
 697          /* this is a lie actually other formats COULD exist where we would fail */
 698          if (format_desc->nr_channels < 3)
 699             return false;
 700       }
 701       else if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB)
 702          return false;
 703
 704       if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN &&
 705           format != PIPE_FORMAT_R11G11B10_FLOAT)
 706          return false;
 707
 708       assert(format_desc->block.width == 1);
 709       assert(format_desc->block.height == 1);
 710
 711       if (format_desc->is_mixed)
 712          return false;
 713
 714       if (!format_desc->is_array && !format_desc->is_bitmask &&
 715           format != PIPE_FORMAT_R11G11B10_FLOAT)
 716          return false;
 717    }
 718
 719    if ((bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_SAMPLER_VIEW)) &&
 720        ((bind & PIPE_BIND_DISPLAY_TARGET) == 0)) {
 721       /* Disable all 3-channel formats, where channel size != 32 bits.
 722        * In some cases we run into crashes (in generate_unswizzled_blend()),
 723        * for 3-channel RGB16 variants, there was an apparent LLVM bug.
 724        * In any case, disabling the shallower 3-channel formats avoids a
 725        * number of issues with GL_ARB_copy_image support.
 726        */
 727       if (format_desc->is_array &&
 728           format_desc->nr_channels == 3 &&
 729           format_desc->block.bits != 96) {
 730          return false;
 731       }
 732    }
 733
 734    if (bind & PIPE_BIND_DISPLAY_TARGET) {
 735       if(!winsys->is_displaytarget_format_supported(winsys, bind, format))
 736          return false;
 737    }
 738
 739    if (bind & PIPE_BIND_DEPTH_STENCIL) {
 740       if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
 741          return false;
 742
 743       if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
 744          return false;
 745    }
 746
 747    if (format_desc->layout == UTIL_FORMAT_LAYOUT_ASTC ||
 748        format_desc->layout == UTIL_FORMAT_LAYOUT_ATC ||
 749        format_desc->layout == UTIL_FORMAT_LAYOUT_FXT1) {
 750       /* Software decoding is not hooked up. */
 751       return false;
 752    }
 753
 754    if (format_desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
 755        format != PIPE_FORMAT_ETC1_RGB8)
 756       return false;
 757
 758    /*
 759     * Everything can be supported by u_format
 760     * (those without fetch_rgba_float might be not but shouldn't hit that)
 761     */
 762
 763    return true;
 764 }
 765
 766
 767
 768
 769 static void
 770 llvmpipe_flush_frontbuffer(struct pipe_screen *_screen,
 771                            struct pipe_resource *resource,
 772                            unsigned level, unsigned layer,
 773                            void *context_private,
 774                            struct pipe_box *sub_box)
 775 {
 776    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 777    struct sw_winsys *winsys = screen->winsys;
 778    struct llvmpipe_resource *texture = llvmpipe_resource(resource);
 779
 780    assert(texture->dt);
 781    if (texture->dt)
 782       winsys->displaytarget_display(winsys, texture->dt, context_private, sub_box);
 783 }
 784
 785 static void
 786 llvmpipe_destroy_screen( struct pipe_screen *_screen )
 787 {
 788    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 789    struct sw_winsys *winsys = screen->winsys;
 790
 791    if (screen->cs_tpool)
 792       lp_cs_tpool_destroy(screen->cs_tpool);
 793
 794    if (screen->rast)
 795       lp_rast_destroy(screen->rast);
 796
 797    lp_jit_screen_cleanup(screen);
 798
 799    if (LP_DEBUG & DEBUG_CACHE_STATS)
 800       printf("disk shader cache:   hits = %u, misses = %u\n", screen->num_disk_shader_cache_hits,
 801              screen->num_disk_shader_cache_misses);
 802    disk_cache_destroy(screen->disk_shader_cache);
 803    if(winsys->destroy)
 804       winsys->destroy(winsys);
 805
 806    glsl_type_singleton_decref();
 807
 808    mtx_destroy(&screen->rast_mutex);
 809    mtx_destroy(&screen->cs_mutex);
 810    FREE(screen);
 811 }
 812
 813
 814
 815
 816 /**
 817  * Fence reference counting.
 818  */
 819 static void
 820 llvmpipe_fence_reference(struct pipe_screen *screen,
 821                          struct pipe_fence_handle **ptr,
 822                          struct pipe_fence_handle *fence)
 823 {
 824    struct lp_fence **old = (struct lp_fence **) ptr;
 825    struct lp_fence *f = (struct lp_fence *) fence;
 826
 827    lp_fence_reference(old, f);
 828 }
 829
 830
 831 /**
 832  * Wait for the fence to finish.
 833  */
 834 static bool
 835 llvmpipe_fence_finish(struct pipe_screen *screen,
 836                       struct pipe_context *ctx,
 837                       struct pipe_fence_handle *fence_handle,
 838                       uint64_t timeout)
 839 {
 840    struct lp_fence *f = (struct lp_fence *) fence_handle;
 841
 842    if (!timeout)
 843       return lp_fence_signalled(f);
 844
 845    if (!lp_fence_signalled(f)) {
 846       if (timeout != PIPE_TIMEOUT_INFINITE)
 847          return lp_fence_timedwait(f, timeout);
 848
 849       lp_fence_wait(f);
 850    }
 851    return true;
 852 }
 853
 854 static uint64_t
 855 llvmpipe_get_timestamp(struct pipe_screen *_screen)
 856 {
 857    return os_time_get_nano();
 858 }
 859
 860 static void lp_disk_cache_create(struct llvmpipe_screen *screen)
 861 {
 862    struct mesa_sha1 ctx;
 863    unsigned char sha1[20];
 864    char cache_id[20 * 2 + 1];
 865    _mesa_sha1_init(&ctx);
 866
 867    if (!disk_cache_get_function_identifier(lp_disk_cache_create, &ctx) ||
 868        !disk_cache_get_function_identifier(LLVMLinkInMCJIT, &ctx))
 869       return;
 870
 871    _mesa_sha1_final(&ctx, sha1);
 872    disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
 873
 874    screen->disk_shader_cache = disk_cache_create("llvmpipe", cache_id, 0);
 875 }
 876
 877 static struct disk_cache *lp_get_disk_shader_cache(struct pipe_screen *_screen)
 878 {
 879    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 880
 881    return screen->disk_shader_cache;
 882 }
 883
 884 void lp_disk_cache_find_shader(struct llvmpipe_screen *screen,
 885                                struct lp_cached_code *cache,
 886                                unsigned char ir_sha1_cache_key[20])
 887 {
 888    unsigned char sha1[CACHE_KEY_SIZE];
 889
 890    if (!screen->disk_shader_cache)
 891       return;
 892    disk_cache_compute_key(screen->disk_shader_cache, ir_sha1_cache_key, 20, sha1);
 893
 894    size_t binary_size;
 895    uint8_t *buffer = disk_cache_get(screen->disk_shader_cache, sha1, &binary_size);
 896    if (!buffer) {
 897       cache->data_size = 0;
 898       p_atomic_inc(&screen->num_disk_shader_cache_misses);
 899       return;
 900    }
 901    cache->data_size = binary_size;
 902    cache->data = buffer;
 903    p_atomic_inc(&screen->num_disk_shader_cache_hits);
 904 }
 905
 906 void lp_disk_cache_insert_shader(struct llvmpipe_screen *screen,
 907                                  struct lp_cached_code *cache,
 908                                  unsigned char ir_sha1_cache_key[20])
 909 {
 910    unsigned char sha1[CACHE_KEY_SIZE];
 911
 912    if (!screen->disk_shader_cache || !cache->data_size || cache->dont_cache)
 913       return;
 914    disk_cache_compute_key(screen->disk_shader_cache, ir_sha1_cache_key, 20, sha1);
 915    disk_cache_put(screen->disk_shader_cache, sha1, cache->data, cache->data_size, NULL);
 916 }
 917 /**
 918  * Create a new pipe_screen object
 919  * Note: we're not presently subclassing pipe_screen (no llvmpipe_screen).
 920  */
 921 struct pipe_screen *
 922 llvmpipe_create_screen(struct sw_winsys *winsys)
 923 {
 924    struct llvmpipe_screen *screen;
 925
 926    util_cpu_detect();
 927
 928    glsl_type_singleton_init_or_ref();
 929
 930 #ifdef DEBUG
 931    LP_DEBUG = debug_get_flags_option("LP_DEBUG", lp_debug_flags, 0 );
 932 #endif
 933
 934    LP_PERF = debug_get_flags_option("LP_PERF", lp_perf_flags, 0 );
 935
 936    screen = CALLOC_STRUCT(llvmpipe_screen);
 937    if (!screen)
 938       return NULL;
 939
 940    if (!lp_jit_screen_init(screen)) {
 941       FREE(screen);
 942       return NULL;
 943    }
 944
 945    screen->winsys = winsys;
 946
 947    screen->base.destroy = llvmpipe_destroy_screen;
 948
 949    screen->base.get_name = llvmpipe_get_name;
 950    screen->base.get_vendor = llvmpipe_get_vendor;
 951    screen->base.get_device_vendor = llvmpipe_get_vendor; // TODO should be the CPU vendor
 952    screen->base.get_param = llvmpipe_get_param;
 953    screen->base.get_shader_param = llvmpipe_get_shader_param;
 954    screen->base.get_compute_param = llvmpipe_get_compute_param;
 955    screen->base.get_paramf = llvmpipe_get_paramf;
 956    screen->base.get_compiler_options = llvmpipe_get_compiler_options;
 957    screen->base.is_format_supported = llvmpipe_is_format_supported;
 958
 959    screen->base.context_create = llvmpipe_create_context;
 960    screen->base.flush_frontbuffer = llvmpipe_flush_frontbuffer;
 961    screen->base.fence_reference = llvmpipe_fence_reference;
 962    screen->base.fence_finish = llvmpipe_fence_finish;
 963
 964    screen->base.get_timestamp = llvmpipe_get_timestamp;
 965
 966    screen->base.finalize_nir = llvmpipe_finalize_nir;
 967
 968    screen->base.get_disk_shader_cache = lp_get_disk_shader_cache;
 969    llvmpipe_init_screen_resource_funcs(&screen->base);
 970
 971    screen->use_tgsi = (LP_DEBUG & DEBUG_TGSI_IR);
 972    screen->num_threads = util_cpu_caps.nr_cpus > 1 ? util_cpu_caps.nr_cpus : 0;
 973 #ifdef EMBEDDED_DEVICE
 974    screen->num_threads = 0;
 975 #endif
 976    screen->num_threads = debug_get_num_option("LP_NUM_THREADS", screen->num_threads);
 977    screen->num_threads = MIN2(screen->num_threads, LP_MAX_THREADS);
 978
 979    screen->rast = lp_rast_create(screen->num_threads);
 980    if (!screen->rast) {
 981       lp_jit_screen_cleanup(screen);
 982       FREE(screen);
 983       return NULL;
 984    }
 985    (void) mtx_init(&screen->rast_mutex, mtx_plain);
 986
 987    screen->cs_tpool = lp_cs_tpool_create(screen->num_threads);
 988    if (!screen->cs_tpool) {
 989       lp_rast_destroy(screen->rast);
 990       lp_jit_screen_cleanup(screen);
 991       FREE(screen);
 992       return NULL;
 993    }
 994    (void) mtx_init(&screen->cs_mutex, mtx_plain);
 995
 996    lp_disk_cache_create(screen);
 997    return &screen->base;
 998 }