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_POINT_SPRITE:
 131       return 1;
 132    case PIPE_CAP_MAX_RENDER_TARGETS:
 133       return PIPE_MAX_COLOR_BUFS;
 134    case PIPE_CAP_OCCLUSION_QUERY:
 135    case PIPE_CAP_QUERY_TIMESTAMP:
 136       return 1;
 137    case PIPE_CAP_QUERY_PIPELINE_STATISTICS:
 138       return 1;
 139    case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
 140    case PIPE_CAP_TEXTURE_MIRROR_CLAMP_TO_EDGE:
 141       return 1;
 142    case PIPE_CAP_TEXTURE_SWIZZLE:
 143    case PIPE_CAP_TEXTURE_SHADOW_LOD:
 144       return 1;
 145    case PIPE_CAP_MAX_TEXTURE_2D_SIZE:
 146       return 1 << (LP_MAX_TEXTURE_2D_LEVELS - 1);
 147    case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
 148       return LP_MAX_TEXTURE_3D_LEVELS;
 149    case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
 150       return LP_MAX_TEXTURE_CUBE_LEVELS;
 151    case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
 152       return LP_MAX_TEXTURE_ARRAY_LAYERS;
 153    case PIPE_CAP_BLEND_EQUATION_SEPARATE:
 154       return 1;
 155    case PIPE_CAP_INDEP_BLEND_ENABLE:
 156       return 1;
 157    case PIPE_CAP_INDEP_BLEND_FUNC:
 158       return 1;
 159    case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
 160    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
 161    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
 162       return 1;
 163    case PIPE_CAP_PRIMITIVE_RESTART:
 164    case PIPE_CAP_PRIMITIVE_RESTART_FIXED_INDEX:
 165       return 1;
 166    case PIPE_CAP_DEPTH_CLIP_DISABLE:
 167       return 1;
 168    case PIPE_CAP_SHADER_STENCIL_EXPORT:
 169       return 1;
 170    case PIPE_CAP_TGSI_INSTANCEID:
 171    case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
 172    case PIPE_CAP_START_INSTANCE:
 173    case PIPE_CAP_MIXED_COLORBUFFER_FORMATS:
 174       return 1;
 175    case PIPE_CAP_SEAMLESS_CUBE_MAP:
 176    case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
 177       return 1;
 178    /* this is a lie could support arbitrary large offsets */
 179    case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
 180    case PIPE_CAP_MIN_TEXEL_OFFSET:
 181       return -32;
 182    case PIPE_CAP_MAX_TEXTURE_GATHER_OFFSET:
 183    case PIPE_CAP_MAX_TEXEL_OFFSET:
 184       return 31;
 185    case PIPE_CAP_CONDITIONAL_RENDER:
 186    case PIPE_CAP_TEXTURE_BARRIER:
 187       return 1;
 188    case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
 189    case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
 190       return 16*4;
 191    case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES:
 192    case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
 193       return 1024;
 194    case PIPE_CAP_MAX_VERTEX_STREAMS: {
 195       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 196       return lscreen->use_tgsi ? 1 : 4;
 197    }
 198    case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE:
 199       return 2048;
 200    case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
 201    case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
 202    case PIPE_CAP_VERTEX_COLOR_UNCLAMPED:
 203    case PIPE_CAP_VERTEX_COLOR_CLAMPED:
 204       return 1;
 205    case PIPE_CAP_GLSL_FEATURE_LEVEL: {
 206       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 207       return lscreen->use_tgsi ? 330 : 450;
 208    }
 209    case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY:
 210       return 140;
 211    case PIPE_CAP_COMPUTE:
 212       return GALLIVM_HAVE_CORO;
 213    case PIPE_CAP_USER_VERTEX_BUFFERS:
 214       return 1;
 215    case PIPE_CAP_TGSI_TEXCOORD:
 216    case PIPE_CAP_DRAW_INDIRECT:
 217       return 1;
 218
 219    case PIPE_CAP_CUBE_MAP_ARRAY:
 220       return 1;
 221    case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
 222       return 16;
 223    case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT:
 224       return 64;
 225    case PIPE_CAP_TEXTURE_BUFFER_OBJECTS:
 226       return 1;
 227    case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
 228       return 65536;
 229    case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
 230       return 16;
 231    case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
 232       return 0;
 233    case PIPE_CAP_MAX_VIEWPORTS:
 234       return PIPE_MAX_VIEWPORTS;
 235    case PIPE_CAP_ENDIANNESS:
 236       return PIPE_ENDIAN_NATIVE;
 237    case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT:
 238       return 1;
 239    case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT:
 240       return 1;
 241    case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
 242       return 4;
 243    case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION:
 244       return 1;
 245    case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE:
 246       return 1;
 247    case PIPE_CAP_TGSI_TEX_TXF_LZ:
 248    case PIPE_CAP_SAMPLER_VIEW_TARGET:
 249       return 1;
 250    case PIPE_CAP_FAKE_SW_MSAA: {
 251       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 252       return lscreen->use_tgsi ? 1 : 0;
 253    }
 254    case PIPE_CAP_TEXTURE_QUERY_LOD:
 255    case PIPE_CAP_CONDITIONAL_RENDER_INVERTED:
 256    case PIPE_CAP_TGSI_ARRAY_COMPONENTS:
 257    case PIPE_CAP_DOUBLES:
 258    case PIPE_CAP_INT64:
 259    case PIPE_CAP_INT64_DIVMOD:
 260    case PIPE_CAP_QUERY_SO_OVERFLOW:
 261    case PIPE_CAP_TGSI_DIV:
 262       return 1;
 263
 264    case PIPE_CAP_VENDOR_ID:
 265       return 0xFFFFFFFF;
 266    case PIPE_CAP_DEVICE_ID:
 267       return 0xFFFFFFFF;
 268    case PIPE_CAP_ACCELERATED:
 269       return 0;
 270    case PIPE_CAP_VIDEO_MEMORY: {
 271       /* XXX: Do we want to return the full amount fo system memory ? */
 272       uint64_t system_memory;
 273
 274       if (!os_get_total_physical_memory(&system_memory))
 275          return 0;
 276
 277       if (sizeof(void *) == 4)
 278          /* Cap to 2 GB on 32 bits system. We do this because llvmpipe does
 279           * eat application memory, which is quite limited on 32 bits. App
 280           * shouldn't expect too much available memory. */
 281          system_memory = MIN2(system_memory, 2048 << 20);
 282
 283       return (int)(system_memory >> 20);
 284    }
 285    case PIPE_CAP_UMA:
 286       return 0;
 287    case PIPE_CAP_CLIP_HALFZ:
 288       return 1;
 289    case PIPE_CAP_POLYGON_OFFSET_CLAMP:
 290    case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
 291    case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR:
 292       return 1;
 293    case PIPE_CAP_CULL_DISTANCE:
 294       return 1;
 295    case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS:
 296       return 1;
 297    case PIPE_CAP_CLEAR_TEXTURE:
 298       return 1;
 299    case PIPE_CAP_MAX_VARYINGS:
 300       return 32;
 301    case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
 302       return 16;
 303    case PIPE_CAP_QUERY_BUFFER_OBJECT:
 304       return 1;
 305    case PIPE_CAP_DRAW_PARAMETERS:
 306       return 1;
 307    case PIPE_CAP_FBFETCH:
 308       return 8;
 309    case PIPE_CAP_FBFETCH_COHERENT:
 310       return 0;
 311    case PIPE_CAP_MULTI_DRAW_INDIRECT:
 312    case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
 313       return 1;
 314    case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
 315    case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR:
 316       return 1;
 317    case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
 318       return 32;
 319    case PIPE_CAP_RASTERIZER_SUBPIXEL_BITS:
 320       return 8;
 321    case PIPE_CAP_PCI_GROUP:
 322    case PIPE_CAP_PCI_BUS:
 323    case PIPE_CAP_PCI_DEVICE:
 324    case PIPE_CAP_PCI_FUNCTION:
 325    case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY:
 326    case PIPE_CAP_ALLOW_MAPPED_BUFFERS_DURING_EXECUTION:
 327       return 0;
 328    case PIPE_CAP_MAX_GS_INVOCATIONS:
 329       return 32;
 330    case PIPE_CAP_MAX_SHADER_BUFFER_SIZE:
 331       return LP_MAX_TGSI_SHADER_BUFFER_SIZE;
 332    case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT:
 333    case PIPE_CAP_TGSI_TG4_COMPONENT_IN_SWIZZLE:
 334    case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
 335       return 1;
 336    case PIPE_CAP_TGSI_TXQS:
 337    case PIPE_CAP_TGSI_VOTE:
 338    case PIPE_CAP_LOAD_CONSTBUF:
 339    case PIPE_CAP_TEXTURE_MULTISAMPLE:
 340    case PIPE_CAP_SAMPLE_SHADING:
 341    case PIPE_CAP_POST_DEPTH_COVERAGE:
 342    case PIPE_CAP_PACKED_UNIFORMS: {
 343       struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 344       return !lscreen->use_tgsi;
 345    }
 346    default:
 347       return u_pipe_screen_get_param_defaults(screen, param);
 348    }
 349 }
 350
 351 static int
 352 llvmpipe_get_shader_param(struct pipe_screen *screen,
 353                           enum pipe_shader_type shader,
 354                           enum pipe_shader_cap param)
 355 {
 356    struct llvmpipe_screen *lscreen = llvmpipe_screen(screen);
 357    switch(shader)
 358    {
 359    case PIPE_SHADER_COMPUTE:
 360       if ((LP_DEBUG & DEBUG_CL) && param == PIPE_SHADER_CAP_SUPPORTED_IRS)
 361          return (1 << PIPE_SHADER_IR_TGSI) | (1 << PIPE_SHADER_IR_NIR) | (1 << PIPE_SHADER_IR_NIR_SERIALIZED);
 362       /* fallthrough */
 363    case PIPE_SHADER_FRAGMENT:
 364       if (param == PIPE_SHADER_CAP_PREFERRED_IR) {
 365          if (lscreen->use_tgsi)
 366             return PIPE_SHADER_IR_TGSI;
 367          else
 368             return PIPE_SHADER_IR_NIR;
 369       }
 370       switch (param) {
 371       default:
 372          return gallivm_get_shader_param(param);
 373       }
 374       /* fallthrough */
 375    case PIPE_SHADER_TESS_CTRL:
 376    case PIPE_SHADER_TESS_EVAL:
 377       /* Tessellation shader needs llvm coroutines support */
 378       if (!GALLIVM_HAVE_CORO || lscreen->use_tgsi)
 379          return 0;
 380       /* fallthrough */
 381    case PIPE_SHADER_VERTEX:
 382    case PIPE_SHADER_GEOMETRY:
 383       if (param == PIPE_SHADER_CAP_PREFERRED_IR) {
 384          if (lscreen->use_tgsi)
 385             return PIPE_SHADER_IR_TGSI;
 386          else
 387             return PIPE_SHADER_IR_NIR;
 388       }
 389
 390       switch (param) {
 391       case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
 392          /* At this time, the draw module and llvmpipe driver only
 393           * support vertex shader texture lookups when LLVM is enabled in
 394           * the draw module.
 395           */
 396          if (debug_get_bool_option("DRAW_USE_LLVM", TRUE))
 397             return PIPE_MAX_SAMPLERS;
 398          else
 399             return 0;
 400       case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
 401          if (debug_get_bool_option("DRAW_USE_LLVM", TRUE))
 402             return PIPE_MAX_SHADER_SAMPLER_VIEWS;
 403          else
 404             return 0;
 405       default:
 406          return draw_get_shader_param(shader, param);
 407       }
 408    default:
 409       return 0;
 410    }
 411 }
 412
 413 static float
 414 llvmpipe_get_paramf(struct pipe_screen *screen, enum pipe_capf param)
 415 {
 416    switch (param) {
 417    case PIPE_CAPF_MAX_LINE_WIDTH:
 418       /* fall-through */
 419    case PIPE_CAPF_MAX_LINE_WIDTH_AA:
 420       return 255.0; /* arbitrary */
 421    case PIPE_CAPF_MAX_POINT_WIDTH:
 422       /* fall-through */
 423    case PIPE_CAPF_MAX_POINT_WIDTH_AA:
 424       return LP_MAX_POINT_WIDTH; /* arbitrary */
 425    case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
 426       return 16.0; /* not actually signficant at this time */
 427    case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
 428       return 16.0; /* arbitrary */
 429    case PIPE_CAPF_MIN_CONSERVATIVE_RASTER_DILATE:
 430       return 0.0;
 431    case PIPE_CAPF_MAX_CONSERVATIVE_RASTER_DILATE:
 432       return 0.0;
 433    case PIPE_CAPF_CONSERVATIVE_RASTER_DILATE_GRANULARITY:
 434       return 0.0;
 435    }
 436    /* should only get here on unhandled cases */
 437    debug_printf("Unexpected PIPE_CAP %d query\n", param);
 438    return 0.0;
 439 }
 440
 441 static int
 442 llvmpipe_get_compute_param(struct pipe_screen *_screen,
 443                            enum pipe_shader_ir ir_type,
 444                            enum pipe_compute_cap param,
 445                            void *ret)
 446 {
 447    switch (param) {
 448    case PIPE_COMPUTE_CAP_IR_TARGET:
 449       return 0;
 450    case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
 451       if (ret) {
 452          uint64_t *grid_size = ret;
 453          grid_size[0] = 65535;
 454          grid_size[1] = 65535;
 455          grid_size[2] = 65535;
 456       }
 457       return 3 * sizeof(uint64_t) ;
 458    case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
 459       if (ret) {
 460          uint64_t *block_size = ret;
 461          block_size[0] = 1024;
 462          block_size[1] = 1024;
 463          block_size[2] = 1024;
 464       }
 465       return 3 * sizeof(uint64_t);
 466    case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
 467       if (ret) {
 468          uint64_t *max_threads_per_block = ret;
 469          *max_threads_per_block = 1024;
 470       }
 471       return sizeof(uint64_t);
 472    case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
 473       if (ret) {
 474          uint64_t *max_local_size = ret;
 475          *max_local_size = 32768;
 476       }
 477       return sizeof(uint64_t);
 478    case PIPE_COMPUTE_CAP_GRID_DIMENSION:
 479       if (ret) {
 480          uint32_t *grid_dim = ret;
 481          *grid_dim = 3;
 482       }
 483       return sizeof(uint32_t);
 484    case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
 485       if (ret) {
 486          uint64_t *max_global_size = ret;
 487          *max_global_size = (1ULL << 31);
 488       }
 489       return sizeof(uint64_t);
 490    case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
 491       if (ret) {
 492          uint64_t *max_mem_alloc_size = ret;
 493          *max_mem_alloc_size = (1ULL << 31);
 494       }
 495       return sizeof(uint64_t);
 496    case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
 497       if (ret) {
 498          uint64_t *max_private = ret;
 499          *max_private = (1UL << 31);
 500       }
 501       return sizeof(uint64_t);
 502    case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
 503       if (ret) {
 504          uint64_t *max_input = ret;
 505          *max_input = 4096;
 506       }
 507       return sizeof(uint64_t);
 508    case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
 509       if (ret) {
 510          uint32_t *images = ret;
 511          *images = 0;
 512       }
 513       return sizeof(uint32_t);
 514    case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
 515       return 0;
 516    case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
 517       if (ret) {
 518          uint32_t *subgroup_size = ret;
 519          *subgroup_size = 32;
 520       }
 521       return sizeof(uint32_t);
 522    case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
 523       if (ret) {
 524          uint32_t *max_compute_units = ret;
 525          *max_compute_units = 8;
 526       }
 527       return sizeof(uint32_t);
 528    case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
 529       if (ret) {
 530          uint32_t *max_clock_freq = ret;
 531          *max_clock_freq = 300;
 532       }
 533       return sizeof(uint32_t);
 534    case PIPE_COMPUTE_CAP_ADDRESS_BITS:
 535       if (ret) {
 536          uint32_t *address_bits = ret;
 537          *address_bits = 64;
 538       }
 539       return sizeof(uint32_t);
 540    }
 541    return 0;
 542 }
 543
 544 static const struct nir_shader_compiler_options gallivm_nir_options = {
 545    .lower_scmp = true,
 546    .lower_flrp32 = true,
 547    .lower_flrp64 = true,
 548    .lower_fsat = true,
 549    .lower_bitfield_insert_to_shifts = true,
 550    .lower_bitfield_extract_to_shifts = true,
 551    .lower_sub = true,
 552    .lower_ffma = true,
 553    .lower_fmod = true,
 554    .lower_hadd = true,
 555    .lower_add_sat = true,
 556    .lower_ldexp = true,
 557    .lower_pack_snorm_2x16 = true,
 558    .lower_pack_snorm_4x8 = true,
 559    .lower_pack_unorm_2x16 = true,
 560    .lower_pack_unorm_4x8 = true,
 561    .lower_pack_half_2x16 = true,
 562    .lower_pack_split = true,
 563    .lower_unpack_snorm_2x16 = true,
 564    .lower_unpack_snorm_4x8 = true,
 565    .lower_unpack_unorm_2x16 = true,
 566    .lower_unpack_unorm_4x8 = true,
 567    .lower_unpack_half_2x16 = true,
 568    .lower_extract_byte = true,
 569    .lower_extract_word = true,
 570    .lower_rotate = true,
 571    .lower_uadd_carry = true,
 572    .lower_usub_borrow = true,
 573    .lower_mul_2x32_64 = true,
 574    .lower_ifind_msb = true,
 575    .max_unroll_iterations = 32,
 576    .use_interpolated_input_intrinsics = true,
 577    .lower_to_scalar = true,
 578    .lower_cs_local_index_from_id = true,
 579 };
 580
 581 static void
 582 llvmpipe_finalize_nir(struct pipe_screen *screen,
 583                       void *nirptr,
 584                       bool optimize)
 585 {
 586    struct nir_shader *nir = (struct nir_shader *)nirptr;
 587    lp_build_opt_nir(nir);
 588 }
 589
 590 static inline const void *
 591 llvmpipe_get_compiler_options(struct pipe_screen *screen,
 592                               enum pipe_shader_ir ir,
 593                               enum pipe_shader_type shader)
 594 {
 595    assert(ir == PIPE_SHADER_IR_NIR);
 596    return &gallivm_nir_options;
 597 }
 598
 599 /**
 600  * Query format support for creating a texture, drawing surface, etc.
 601  * \param format  the format to test
 602  * \param type  one of PIPE_TEXTURE, PIPE_SURFACE
 603  */
 604 static bool
 605 llvmpipe_is_format_supported( struct pipe_screen *_screen,
 606                               enum pipe_format format,
 607                               enum pipe_texture_target target,
 608                               unsigned sample_count,
 609                               unsigned storage_sample_count,
 610                               unsigned bind)
 611 {
 612    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 613    struct sw_winsys *winsys = screen->winsys;
 614    const struct util_format_description *format_desc;
 615
 616    format_desc = util_format_description(format);
 617    if (!format_desc)
 618       return false;
 619
 620    assert(target == PIPE_BUFFER ||
 621           target == PIPE_TEXTURE_1D ||
 622           target == PIPE_TEXTURE_1D_ARRAY ||
 623           target == PIPE_TEXTURE_2D ||
 624           target == PIPE_TEXTURE_2D_ARRAY ||
 625           target == PIPE_TEXTURE_RECT ||
 626           target == PIPE_TEXTURE_3D ||
 627           target == PIPE_TEXTURE_CUBE ||
 628           target == PIPE_TEXTURE_CUBE_ARRAY);
 629
 630    if (sample_count != 0 && sample_count != 1 && sample_count != 4)
 631       return false;
 632
 633    if (MAX2(1, sample_count) != MAX2(1, storage_sample_count))
 634       return false;
 635
 636    if (bind & PIPE_BIND_RENDER_TARGET) {
 637       if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
 638          /* this is a lie actually other formats COULD exist where we would fail */
 639          if (format_desc->nr_channels < 3)
 640             return false;
 641       }
 642       else if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB)
 643          return false;
 644
 645       if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN &&
 646           format != PIPE_FORMAT_R11G11B10_FLOAT)
 647          return false;
 648
 649       assert(format_desc->block.width == 1);
 650       assert(format_desc->block.height == 1);
 651
 652       if (format_desc->is_mixed)
 653          return false;
 654
 655       if (!format_desc->is_array && !format_desc->is_bitmask &&
 656           format != PIPE_FORMAT_R11G11B10_FLOAT)
 657          return false;
 658    }
 659
 660    if ((bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_SAMPLER_VIEW)) &&
 661        ((bind & PIPE_BIND_DISPLAY_TARGET) == 0)) {
 662       /* Disable all 3-channel formats, where channel size != 32 bits.
 663        * In some cases we run into crashes (in generate_unswizzled_blend()),
 664        * for 3-channel RGB16 variants, there was an apparent LLVM bug.
 665        * In any case, disabling the shallower 3-channel formats avoids a
 666        * number of issues with GL_ARB_copy_image support.
 667        */
 668       if (format_desc->is_array &&
 669           format_desc->nr_channels == 3 &&
 670           format_desc->block.bits != 96) {
 671          return false;
 672       }
 673    }
 674
 675    if (bind & PIPE_BIND_DISPLAY_TARGET) {
 676       if(!winsys->is_displaytarget_format_supported(winsys, bind, format))
 677          return false;
 678    }
 679
 680    if (bind & PIPE_BIND_DEPTH_STENCIL) {
 681       if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
 682          return false;
 683
 684       if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
 685          return false;
 686    }
 687
 688    if (format_desc->layout == UTIL_FORMAT_LAYOUT_ASTC ||
 689        format_desc->layout == UTIL_FORMAT_LAYOUT_ATC ||
 690        format_desc->layout == UTIL_FORMAT_LAYOUT_FXT1) {
 691       /* Software decoding is not hooked up. */
 692       return false;
 693    }
 694
 695    if (format_desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
 696        format != PIPE_FORMAT_ETC1_RGB8)
 697       return false;
 698
 699    /*
 700     * Everything can be supported by u_format
 701     * (those without fetch_rgba_float might be not but shouldn't hit that)
 702     */
 703
 704    return true;
 705 }
 706
 707
 708
 709
 710 static void
 711 llvmpipe_flush_frontbuffer(struct pipe_screen *_screen,
 712                            struct pipe_resource *resource,
 713                            unsigned level, unsigned layer,
 714                            void *context_private,
 715                            struct pipe_box *sub_box)
 716 {
 717    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 718    struct sw_winsys *winsys = screen->winsys;
 719    struct llvmpipe_resource *texture = llvmpipe_resource(resource);
 720
 721    assert(texture->dt);
 722    if (texture->dt)
 723       winsys->displaytarget_display(winsys, texture->dt, context_private, sub_box);
 724 }
 725
 726 static void
 727 llvmpipe_destroy_screen( struct pipe_screen *_screen )
 728 {
 729    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 730    struct sw_winsys *winsys = screen->winsys;
 731
 732    if (screen->cs_tpool)
 733       lp_cs_tpool_destroy(screen->cs_tpool);
 734
 735    if (screen->rast)
 736       lp_rast_destroy(screen->rast);
 737
 738    lp_jit_screen_cleanup(screen);
 739
 740    if (LP_DEBUG & DEBUG_CACHE_STATS)
 741       printf("disk shader cache:   hits = %u, misses = %u\n", screen->num_disk_shader_cache_hits,
 742              screen->num_disk_shader_cache_misses);
 743    disk_cache_destroy(screen->disk_shader_cache);
 744    if(winsys->destroy)
 745       winsys->destroy(winsys);
 746
 747    glsl_type_singleton_decref();
 748
 749    mtx_destroy(&screen->rast_mutex);
 750    mtx_destroy(&screen->cs_mutex);
 751    FREE(screen);
 752 }
 753
 754
 755
 756
 757 /**
 758  * Fence reference counting.
 759  */
 760 static void
 761 llvmpipe_fence_reference(struct pipe_screen *screen,
 762                          struct pipe_fence_handle **ptr,
 763                          struct pipe_fence_handle *fence)
 764 {
 765    struct lp_fence **old = (struct lp_fence **) ptr;
 766    struct lp_fence *f = (struct lp_fence *) fence;
 767
 768    lp_fence_reference(old, f);
 769 }
 770
 771
 772 /**
 773  * Wait for the fence to finish.
 774  */
 775 static bool
 776 llvmpipe_fence_finish(struct pipe_screen *screen,
 777                       struct pipe_context *ctx,
 778                       struct pipe_fence_handle *fence_handle,
 779                       uint64_t timeout)
 780 {
 781    struct lp_fence *f = (struct lp_fence *) fence_handle;
 782
 783    if (!timeout)
 784       return lp_fence_signalled(f);
 785
 786    if (!lp_fence_signalled(f)) {
 787       if (timeout != PIPE_TIMEOUT_INFINITE)
 788          return lp_fence_timedwait(f, timeout);
 789
 790       lp_fence_wait(f);
 791    }
 792    return true;
 793 }
 794
 795 static uint64_t
 796 llvmpipe_get_timestamp(struct pipe_screen *_screen)
 797 {
 798    return os_time_get_nano();
 799 }
 800
 801 static void lp_disk_cache_create(struct llvmpipe_screen *screen)
 802 {
 803    struct mesa_sha1 ctx;
 804    unsigned char sha1[20];
 805    char cache_id[20 * 2 + 1];
 806    _mesa_sha1_init(&ctx);
 807
 808    if (!disk_cache_get_function_identifier(lp_disk_cache_create, &ctx) ||
 809        !disk_cache_get_function_identifier(LLVMLinkInMCJIT, &ctx))
 810       return;
 811
 812    _mesa_sha1_final(&ctx, sha1);
 813    disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
 814
 815    screen->disk_shader_cache = disk_cache_create("llvmpipe", cache_id, 0);
 816 }
 817
 818 static struct disk_cache *lp_get_disk_shader_cache(struct pipe_screen *_screen)
 819 {
 820    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
 821
 822    return screen->disk_shader_cache;
 823 }
 824
 825 void lp_disk_cache_find_shader(struct llvmpipe_screen *screen,
 826                                struct lp_cached_code *cache,
 827                                unsigned char ir_sha1_cache_key[20])
 828 {
 829    unsigned char sha1[CACHE_KEY_SIZE];
 830
 831    if (!screen->disk_shader_cache)
 832       return;
 833    disk_cache_compute_key(screen->disk_shader_cache, ir_sha1_cache_key, 20, sha1);
 834
 835    size_t binary_size;
 836    uint8_t *buffer = disk_cache_get(screen->disk_shader_cache, sha1, &binary_size);
 837    if (!buffer) {
 838       cache->data_size = 0;
 839       p_atomic_inc(&screen->num_disk_shader_cache_misses);
 840       return;
 841    }
 842    cache->data_size = binary_size;
 843    cache->data = buffer;
 844    p_atomic_inc(&screen->num_disk_shader_cache_hits);
 845 }
 846
 847 void lp_disk_cache_insert_shader(struct llvmpipe_screen *screen,
 848                                  struct lp_cached_code *cache,
 849                                  unsigned char ir_sha1_cache_key[20])
 850 {
 851    unsigned char sha1[CACHE_KEY_SIZE];
 852
 853    if (!screen->disk_shader_cache || !cache->data_size || cache->dont_cache)
 854       return;
 855    disk_cache_compute_key(screen->disk_shader_cache, ir_sha1_cache_key, 20, sha1);
 856    disk_cache_put(screen->disk_shader_cache, sha1, cache->data, cache->data_size, NULL);
 857 }
 858 /**
 859  * Create a new pipe_screen object
 860  * Note: we're not presently subclassing pipe_screen (no llvmpipe_screen).
 861  */
 862 struct pipe_screen *
 863 llvmpipe_create_screen(struct sw_winsys *winsys)
 864 {
 865    struct llvmpipe_screen *screen;
 866
 867    util_cpu_detect();
 868
 869    glsl_type_singleton_init_or_ref();
 870
 871 #ifdef DEBUG
 872    LP_DEBUG = debug_get_flags_option("LP_DEBUG", lp_debug_flags, 0 );
 873 #endif
 874
 875    LP_PERF = debug_get_flags_option("LP_PERF", lp_perf_flags, 0 );
 876
 877    screen = CALLOC_STRUCT(llvmpipe_screen);
 878    if (!screen)
 879       return NULL;
 880
 881    if (!lp_jit_screen_init(screen)) {
 882       FREE(screen);
 883       return NULL;
 884    }
 885
 886    screen->winsys = winsys;
 887
 888    screen->base.destroy = llvmpipe_destroy_screen;
 889
 890    screen->base.get_name = llvmpipe_get_name;
 891    screen->base.get_vendor = llvmpipe_get_vendor;
 892    screen->base.get_device_vendor = llvmpipe_get_vendor; // TODO should be the CPU vendor
 893    screen->base.get_param = llvmpipe_get_param;
 894    screen->base.get_shader_param = llvmpipe_get_shader_param;
 895    screen->base.get_compute_param = llvmpipe_get_compute_param;
 896    screen->base.get_paramf = llvmpipe_get_paramf;
 897    screen->base.get_compiler_options = llvmpipe_get_compiler_options;
 898    screen->base.is_format_supported = llvmpipe_is_format_supported;
 899
 900    screen->base.context_create = llvmpipe_create_context;
 901    screen->base.flush_frontbuffer = llvmpipe_flush_frontbuffer;
 902    screen->base.fence_reference = llvmpipe_fence_reference;
 903    screen->base.fence_finish = llvmpipe_fence_finish;
 904
 905    screen->base.get_timestamp = llvmpipe_get_timestamp;
 906
 907    screen->base.finalize_nir = llvmpipe_finalize_nir;
 908
 909    screen->base.get_disk_shader_cache = lp_get_disk_shader_cache;
 910    llvmpipe_init_screen_resource_funcs(&screen->base);
 911
 912    screen->use_tgsi = (LP_DEBUG & DEBUG_TGSI_IR);
 913    screen->num_threads = util_cpu_caps.nr_cpus > 1 ? util_cpu_caps.nr_cpus : 0;
 914 #ifdef EMBEDDED_DEVICE
 915    screen->num_threads = 0;
 916 #endif
 917    screen->num_threads = debug_get_num_option("LP_NUM_THREADS", screen->num_threads);
 918    screen->num_threads = MIN2(screen->num_threads, LP_MAX_THREADS);
 919
 920    screen->rast = lp_rast_create(screen->num_threads);
 921    if (!screen->rast) {
 922       lp_jit_screen_cleanup(screen);
 923       FREE(screen);
 924       return NULL;
 925    }
 926    (void) mtx_init(&screen->rast_mutex, mtx_plain);
 927
 928    screen->cs_tpool = lp_cs_tpool_create(screen->num_threads);
 929    if (!screen->cs_tpool) {
 930       lp_rast_destroy(screen->rast);
 931       lp_jit_screen_cleanup(screen);
 932       FREE(screen);
 933       return NULL;
 934    }
 935    (void) mtx_init(&screen->cs_mutex, mtx_plain);
 936
 937    lp_disk_cache_create(screen);
 938    return &screen->base;
 939 }