src/gallium/drivers/panfrost/pan_screen.c

   1 /*
   2  * Copyright (C) 2008 VMware, Inc.
   3  * Copyright (C) 2014 Broadcom
   4  * Copyright (C) 2018 Alyssa Rosenzweig
   5  * Copyright (C) 2019 Collabora, Ltd.
   6  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  24  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  25  * SOFTWARE.
  26  *
  27  */
  28
  29 #include "util/u_debug.h"
  30 #include "util/u_memory.h"
  31 #include "util/format/u_format.h"
  32 #include "util/format/u_format_s3tc.h"
  33 #include "util/u_video.h"
  34 #include "util/u_screen.h"
  35 #include "util/os_time.h"
  36 #include "util/u_process.h"
  37 #include "pipe/p_defines.h"
  38 #include "pipe/p_screen.h"
  39 #include "draw/draw_context.h"
  40
  41 #include <fcntl.h>
  42
  43 #include "drm-uapi/drm_fourcc.h"
  44 #include "drm-uapi/panfrost_drm.h"
  45
  46 #include "pan_bo.h"
  47 #include "pan_screen.h"
  48 #include "pan_resource.h"
  49 #include "pan_public.h"
  50 #include "pan_util.h"
  51 #include "pandecode/decode.h"
  52
  53 #include "pan_context.h"
  54 #include "midgard/midgard_compile.h"
  55 #include "panfrost-quirks.h"
  56
  57 static const struct debug_named_value debug_options[] = {
  58         {"msgs",      PAN_DBG_MSGS,     "Print debug messages"},
  59         {"trace",     PAN_DBG_TRACE,    "Trace the command stream"},
  60         {"deqp",      PAN_DBG_DEQP,     "Hacks for dEQP"},
  61         {"afbc",      PAN_DBG_AFBC,     "Enable non-conformant AFBC impl"},
  62         {"sync",      PAN_DBG_SYNC,     "Wait for each job's completion and check for any GPU fault"},
  63         DEBUG_NAMED_VALUE_END
  64 };
  65
  66 DEBUG_GET_ONCE_FLAGS_OPTION(pan_debug, "PAN_MESA_DEBUG", debug_options, 0)
  67
  68 int pan_debug = 0;
  69
  70 static const char *
  71 panfrost_get_name(struct pipe_screen *screen)
  72 {
  73         return "panfrost";
  74 }
  75
  76 static const char *
  77 panfrost_get_vendor(struct pipe_screen *screen)
  78 {
  79         return "panfrost";
  80 }
  81
  82 static const char *
  83 panfrost_get_device_vendor(struct pipe_screen *screen)
  84 {
  85         return "Arm";
  86 }
  87
  88 static int
  89 panfrost_get_param(struct pipe_screen *screen, enum pipe_cap param)
  90 {
  91         /* We expose in-dev stuff for dEQP that we don't want apps to use yet */
  92         bool is_deqp = pan_debug & PAN_DBG_DEQP;
  93
  94         switch (param) {
  95         case PIPE_CAP_NPOT_TEXTURES:
  96         case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
  97         case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
  98         case PIPE_CAP_FRAGMENT_SHADER_TEXTURE_LOD:
  99         case PIPE_CAP_VERTEX_SHADER_SATURATE:
 100         case PIPE_CAP_POINT_SPRITE:
 101                 return 1;
 102
 103         case PIPE_CAP_MAX_RENDER_TARGETS:
 104                 return is_deqp ? 4 : 1;
 105
 106         /* Throttling frames breaks pipelining */
 107         case PIPE_CAP_THROTTLE:
 108                 return 0;
 109
 110         case PIPE_CAP_OCCLUSION_QUERY:
 111                 return 1;
 112         case PIPE_CAP_QUERY_TIME_ELAPSED:
 113         case PIPE_CAP_QUERY_PIPELINE_STATISTICS:
 114         case PIPE_CAP_QUERY_TIMESTAMP:
 115         case PIPE_CAP_QUERY_SO_OVERFLOW:
 116                 return 0;
 117
 118         case PIPE_CAP_TEXTURE_SWIZZLE:
 119                 return 1;
 120
 121         case PIPE_CAP_TGSI_INSTANCEID:
 122         case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
 123                 return is_deqp ? 1 : 0;
 124
 125         case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
 126                 return is_deqp ? 4 : 0;
 127         case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
 128         case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
 129                 return is_deqp ? 64 : 0;
 130         case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
 131                 return 1;
 132
 133         case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
 134                 return is_deqp ? 256 : 0; /* for GL3 */
 135
 136         case PIPE_CAP_GLSL_FEATURE_LEVEL:
 137         case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY:
 138                 return is_deqp ? 140 : 120;
 139         case PIPE_CAP_ESSL_FEATURE_LEVEL:
 140                 return is_deqp ? 300 : 120;
 141
 142         case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
 143                 return is_deqp ? 16 : 0;
 144
 145         case PIPE_CAP_CUBE_MAP_ARRAY:
 146                 return is_deqp;
 147
 148         /* For faking GLES 3.1 for dEQP-GLES31 */
 149         case PIPE_CAP_TEXTURE_MULTISAMPLE:
 150         case PIPE_CAP_MAX_COMBINED_HW_ATOMIC_COUNTERS:
 151         case PIPE_CAP_MAX_COMBINED_HW_ATOMIC_COUNTER_BUFFERS:
 152         case PIPE_CAP_IMAGE_LOAD_FORMATTED:
 153                 return is_deqp;
 154
 155         /* For faking compute shaders */
 156         case PIPE_CAP_COMPUTE:
 157                 return is_deqp;
 158
 159         /* TODO: Where does this req come from in practice? */
 160         case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
 161                 return 1;
 162
 163         case PIPE_CAP_MAX_TEXTURE_2D_SIZE:
 164                 return 4096;
 165         case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
 166         case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
 167                 return 13;
 168
 169         case PIPE_CAP_BLEND_EQUATION_SEPARATE:
 170         case PIPE_CAP_INDEP_BLEND_ENABLE:
 171         case PIPE_CAP_INDEP_BLEND_FUNC:
 172                 return 1;
 173
 174         case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
 175                 /* Hardware is natively upper left */
 176                 return 0;
 177
 178         case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
 179         case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
 180         case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
 181         case PIPE_CAP_GENERATE_MIPMAP:
 182                 return 1;
 183
 184         /* We would prefer varyings */
 185         case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
 186         case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
 187                 return 0;
 188
 189         /* I really don't want to set this CAP but let's not swim against the
 190          * tide.. */
 191         case PIPE_CAP_TGSI_TEXCOORD:
 192                 return 1;
 193
 194         case PIPE_CAP_SEAMLESS_CUBE_MAP:
 195         case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
 196                 return 1;
 197
 198         case PIPE_CAP_MAX_VERTEX_ELEMENT_SRC_OFFSET:
 199                 return 0xffff;
 200
 201         case PIPE_CAP_TEXTURE_BUFFER_OBJECTS:
 202                 return 1;
 203
 204         case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
 205                 return 65536;
 206
 207         case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
 208                 return 0;
 209
 210         case PIPE_CAP_ENDIANNESS:
 211                 return PIPE_ENDIAN_NATIVE;
 212
 213         case PIPE_CAP_SAMPLER_VIEW_TARGET:
 214                 return 1;
 215
 216         case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
 217                 return -8;
 218
 219         case PIPE_CAP_MAX_TEXTURE_GATHER_OFFSET:
 220                 return 7;
 221
 222         case PIPE_CAP_VENDOR_ID:
 223         case PIPE_CAP_DEVICE_ID:
 224                 return 0xFFFFFFFF;
 225
 226         case PIPE_CAP_ACCELERATED:
 227         case PIPE_CAP_UMA:
 228         case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
 229         case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR:
 230         case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS:
 231         case PIPE_CAP_TGSI_ARRAY_COMPONENTS:
 232                 return 1;
 233
 234         case PIPE_CAP_VIDEO_MEMORY: {
 235                 uint64_t system_memory;
 236
 237                 if (!os_get_total_physical_memory(&system_memory))
 238                         return 0;
 239
 240                 return (int)(system_memory >> 20);
 241         }
 242
 243         case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
 244                 return 4;
 245
 246         case PIPE_CAP_MAX_VARYINGS:
 247                 return 16;
 248
 249         case PIPE_CAP_ALPHA_TEST:
 250                 return 0;
 251
 252         default:
 253                 return u_pipe_screen_get_param_defaults(screen, param);
 254         }
 255 }
 256
 257 static int
 258 panfrost_get_shader_param(struct pipe_screen *screen,
 259                           enum pipe_shader_type shader,
 260                           enum pipe_shader_cap param)
 261 {
 262         bool is_deqp = pan_debug & PAN_DBG_DEQP;
 263
 264         if (shader != PIPE_SHADER_VERTEX &&
 265             shader != PIPE_SHADER_FRAGMENT &&
 266             !(shader == PIPE_SHADER_COMPUTE && is_deqp))
 267                 return 0;
 268
 269         /* this is probably not totally correct.. but it's a start: */
 270         switch (param) {
 271         case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
 272         case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
 273         case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
 274         case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
 275                 return 16384;
 276
 277         case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
 278                 return 1024;
 279
 280         case PIPE_SHADER_CAP_MAX_INPUTS:
 281                 return 16;
 282
 283         case PIPE_SHADER_CAP_MAX_OUTPUTS:
 284                 return shader == PIPE_SHADER_FRAGMENT ? 4 : 8;
 285
 286         case PIPE_SHADER_CAP_MAX_TEMPS:
 287                 return 256; /* GL_MAX_PROGRAM_TEMPORARIES_ARB */
 288
 289         case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
 290                 return 16 * 1024 * sizeof(float);
 291
 292         case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
 293                 return PAN_MAX_CONST_BUFFERS;
 294
 295         case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
 296                 return 0;
 297
 298         case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
 299                 return 1;
 300         case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
 301                 return 0;
 302
 303         case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
 304                 return 0;
 305
 306         case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
 307                 return 1;
 308
 309         case PIPE_SHADER_CAP_SUBROUTINES:
 310                 return 0;
 311
 312         case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
 313                 return 0;
 314
 315         case PIPE_SHADER_CAP_INTEGERS:
 316                 return 1;
 317
 318         case PIPE_SHADER_CAP_INT64_ATOMICS:
 319         case PIPE_SHADER_CAP_FP16:
 320         case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
 321         case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
 322         case PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED:
 323         case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
 324         case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
 325                 return 0;
 326
 327         case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
 328         case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
 329                 return 16; /* XXX: How many? */
 330
 331         case PIPE_SHADER_CAP_PREFERRED_IR:
 332                 return PIPE_SHADER_IR_NIR;
 333
 334         case PIPE_SHADER_CAP_SUPPORTED_IRS:
 335                 return (1 << PIPE_SHADER_IR_NIR) | (1 << PIPE_SHADER_IR_NIR_SERIALIZED);
 336
 337         case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
 338                 return 32;
 339
 340         case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
 341         case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
 342                 return is_deqp ? 4 : 0;
 343         case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTERS:
 344         case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTER_BUFFERS:
 345                 return 0;
 346
 347         case PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS:
 348         case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD:
 349                 return 0;
 350
 351         default:
 352                 fprintf(stderr, "unknown shader param %d\n", param);
 353                 return 0;
 354         }
 355
 356         return 0;
 357 }
 358
 359 static float
 360 panfrost_get_paramf(struct pipe_screen *screen, enum pipe_capf param)
 361 {
 362         switch (param) {
 363         case PIPE_CAPF_MAX_LINE_WIDTH:
 364
 365         /* fall-through */
 366         case PIPE_CAPF_MAX_LINE_WIDTH_AA:
 367                 return 255.0; /* arbitrary */
 368
 369         case PIPE_CAPF_MAX_POINT_WIDTH:
 370
 371         /* fall-through */
 372         case PIPE_CAPF_MAX_POINT_WIDTH_AA:
 373                 return 1024.0;
 374
 375         case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
 376                 return 16.0;
 377
 378         case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
 379                 return 16.0; /* arbitrary */
 380
 381         case PIPE_CAPF_MIN_CONSERVATIVE_RASTER_DILATE:
 382         case PIPE_CAPF_MAX_CONSERVATIVE_RASTER_DILATE:
 383         case PIPE_CAPF_CONSERVATIVE_RASTER_DILATE_GRANULARITY:
 384                 return 0.0f;
 385
 386         default:
 387                 debug_printf("Unexpected PIPE_CAPF %d query\n", param);
 388                 return 0.0;
 389         }
 390 }
 391
 392 /**
 393  * Query format support for creating a texture, drawing surface, etc.
 394  * \param format  the format to test
 395  * \param type  one of PIPE_TEXTURE, PIPE_SURFACE
 396  */
 397 static bool
 398 panfrost_is_format_supported( struct pipe_screen *screen,
 399                               enum pipe_format format,
 400                               enum pipe_texture_target target,
 401                               unsigned sample_count,
 402                               unsigned storage_sample_count,
 403                               unsigned bind)
 404 {
 405         const struct util_format_description *format_desc;
 406
 407         assert(target == PIPE_BUFFER ||
 408                target == PIPE_TEXTURE_1D ||
 409                target == PIPE_TEXTURE_1D_ARRAY ||
 410                target == PIPE_TEXTURE_2D ||
 411                target == PIPE_TEXTURE_2D_ARRAY ||
 412                target == PIPE_TEXTURE_RECT ||
 413                target == PIPE_TEXTURE_3D ||
 414                target == PIPE_TEXTURE_CUBE ||
 415                target == PIPE_TEXTURE_CUBE_ARRAY);
 416
 417         format_desc = util_format_description(format);
 418
 419         if (!format_desc)
 420                 return false;
 421
 422         if (sample_count > 1)
 423                 return false;
 424
 425         /* Format wishlist */
 426         if (format == PIPE_FORMAT_X8Z24_UNORM)
 427                 return false;
 428
 429         if (format == PIPE_FORMAT_A1B5G5R5_UNORM || format == PIPE_FORMAT_X1B5G5R5_UNORM)
 430                 return false;
 431
 432         /* TODO */
 433         if (format == PIPE_FORMAT_B5G5R5A1_UNORM)
 434                 return FALSE;
 435
 436         /* Don't confuse poorly written apps (workaround dEQP bug) that expect
 437          * more alpha than they ask for */
 438
 439         bool scanout = bind & (PIPE_BIND_SCANOUT | PIPE_BIND_SHARED | PIPE_BIND_DISPLAY_TARGET);
 440         bool renderable = bind & PIPE_BIND_RENDER_TARGET;
 441
 442         if (scanout && renderable && !util_format_is_rgba8_variant(format_desc))
 443                 return false;
 444
 445         if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN &&
 446             format_desc->layout != UTIL_FORMAT_LAYOUT_OTHER) {
 447                 /* Compressed formats not yet hooked up. */
 448                 return false;
 449         }
 450
 451         /* Internally, formats that are depth/stencil renderable are limited.
 452          *
 453          * In particular: Z16, Z24, Z24S8, S8 are all identical from the GPU
 454          * rendering perspective. That is, we render to Z24S8 (which we can
 455          * AFBC compress), ignore the different when texturing (who cares?),
 456          * and then in the off-chance there's a CPU read we blit back to
 457          * staging.
 458          *
 459          * ...alternatively, we can make the state tracker deal with that. */
 460
 461         if (bind & PIPE_BIND_DEPTH_STENCIL) {
 462                 switch (format) {
 463                         case PIPE_FORMAT_Z24_UNORM_S8_UINT:
 464                         case PIPE_FORMAT_Z24X8_UNORM:
 465                         case PIPE_FORMAT_Z32_UNORM:
 466                         case PIPE_FORMAT_Z32_FLOAT:
 467                         case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
 468                                 return true;
 469
 470                         default:
 471                                 return false;
 472                 }
 473         }
 474
 475         return true;
 476 }
 477
 478 static int
 479 panfrost_get_compute_param(struct pipe_screen *pscreen, enum pipe_shader_ir ir_type,
 480                 enum pipe_compute_cap param, void *ret)
 481 {
 482         const char * const ir = "panfrost";
 483
 484         if (!(pan_debug & PAN_DBG_DEQP))
 485                 return 0;
 486
 487 #define RET(x) do {                  \
 488    if (ret)                          \
 489       memcpy(ret, x, sizeof(x));     \
 490    return sizeof(x);                 \
 491 } while (0)
 492
 493         switch (param) {
 494         case PIPE_COMPUTE_CAP_ADDRESS_BITS:
 495                 RET((uint32_t []){ 64 });
 496
 497         case PIPE_COMPUTE_CAP_IR_TARGET:
 498                 if (ret)
 499                         sprintf(ret, "%s", ir);
 500                 return strlen(ir) * sizeof(char);
 501
 502         case PIPE_COMPUTE_CAP_GRID_DIMENSION:
 503                 RET((uint64_t []) { 3 });
 504
 505         case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
 506                 RET(((uint64_t []) { 65535, 65535, 65535 }));
 507
 508         case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
 509                 RET(((uint64_t []) { 1024, 1024, 64 }));
 510
 511         case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
 512                 RET((uint64_t []) { 1024 });
 513
 514         case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
 515                 RET((uint64_t []) { 1024*1024*512 /* Maybe get memory */ });
 516
 517         case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
 518                 RET((uint64_t []) { 32768 });
 519
 520         case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
 521         case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
 522                 RET((uint64_t []) { 4096 });
 523
 524         case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
 525                 RET((uint64_t []) { 1024*1024*512 /* Maybe get memory */ });
 526
 527         case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
 528                 RET((uint32_t []) { 800 /* MHz -- TODO */ });
 529
 530         case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
 531                 RET((uint32_t []) { 9999 });  // TODO
 532
 533         case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
 534                 RET((uint32_t []) { 1 }); // TODO
 535
 536         case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
 537                 RET((uint32_t []) { 32 });  // TODO
 538
 539         case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
 540                 RET((uint64_t []) { 1024 }); // TODO
 541         }
 542
 543         return 0;
 544 }
 545
 546 static void
 547 panfrost_destroy_screen(struct pipe_screen *pscreen)
 548 {
 549         struct panfrost_screen *screen = pan_screen(pscreen);
 550         panfrost_bo_cache_evict_all(screen);
 551         pthread_mutex_destroy(&screen->bo_cache.lock);
 552         pthread_mutex_destroy(&screen->active_bos_lock);
 553         drmFreeVersion(screen->kernel_version);
 554         ralloc_free(screen);
 555 }
 556
 557 static void
 558 panfrost_flush_frontbuffer(struct pipe_screen *_screen,
 559                            struct pipe_resource *resource,
 560                            unsigned level, unsigned layer,
 561                            void *context_private,
 562                            struct pipe_box *sub_box)
 563 {
 564         /* TODO: Display target integration */
 565 }
 566
 567 static uint64_t
 568 panfrost_get_timestamp(struct pipe_screen *_screen)
 569 {
 570         return os_time_get_nano();
 571 }
 572
 573 static void
 574 panfrost_fence_reference(struct pipe_screen *pscreen,
 575                          struct pipe_fence_handle **ptr,
 576                          struct pipe_fence_handle *fence)
 577 {
 578         struct panfrost_fence **p = (struct panfrost_fence **)ptr;
 579         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 580         struct panfrost_fence *old = *p;
 581
 582         if (pipe_reference(&(*p)->reference, &f->reference)) {
 583                 util_dynarray_foreach(&old->syncfds, int, fd)
 584                         close(*fd);
 585                 util_dynarray_fini(&old->syncfds);
 586                 free(old);
 587         }
 588         *p = f;
 589 }
 590
 591 static bool
 592 panfrost_fence_finish(struct pipe_screen *pscreen,
 593                       struct pipe_context *ctx,
 594                       struct pipe_fence_handle *fence,
 595                       uint64_t timeout)
 596 {
 597         struct panfrost_screen *screen = pan_screen(pscreen);
 598         struct panfrost_fence *f = (struct panfrost_fence *)fence;
 599         struct util_dynarray syncobjs;
 600         int ret;
 601
 602         /* All fences were already signaled */
 603         if (!util_dynarray_num_elements(&f->syncfds, int))
 604                 return true;
 605
 606         util_dynarray_init(&syncobjs, NULL);
 607         util_dynarray_foreach(&f->syncfds, int, fd) {
 608                 uint32_t syncobj;
 609
 610                 ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
 611                 assert(!ret);
 612
 613                 ret = drmSyncobjImportSyncFile(screen->fd, syncobj, *fd);
 614                 assert(!ret);
 615                 util_dynarray_append(&syncobjs, uint32_t, syncobj);
 616         }
 617
 618         uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 619         if (abs_timeout == OS_TIMEOUT_INFINITE)
 620                 abs_timeout = INT64_MAX;
 621
 622         ret = drmSyncobjWait(screen->fd, util_dynarray_begin(&syncobjs),
 623                              util_dynarray_num_elements(&syncobjs, uint32_t),
 624                              abs_timeout, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL,
 625                              NULL);
 626
 627         util_dynarray_foreach(&syncobjs, uint32_t, syncobj)
 628                 drmSyncobjDestroy(screen->fd, *syncobj);
 629
 630         return ret >= 0;
 631 }
 632
 633 struct panfrost_fence *
 634 panfrost_fence_create(struct panfrost_context *ctx,
 635                       struct util_dynarray *fences)
 636 {
 637         struct panfrost_screen *screen = pan_screen(ctx->base.screen);
 638         struct panfrost_fence *f = calloc(1, sizeof(*f));
 639         if (!f)
 640                 return NULL;
 641
 642         util_dynarray_init(&f->syncfds, NULL);
 643
 644         /* Export fences from all pending batches. */
 645         util_dynarray_foreach(fences, struct panfrost_batch_fence *, fence) {
 646                 int fd = -1;
 647
 648                 /* The fence is already signaled, no need to export it. */
 649                 if ((*fence)->signaled)
 650                         continue;
 651
 652                 drmSyncobjExportSyncFile(screen->fd, (*fence)->syncobj, &fd);
 653                 if (fd == -1)
 654                         fprintf(stderr, "export failed: %m\n");
 655
 656                 assert(fd != -1);
 657                 util_dynarray_append(&f->syncfds, int, fd);
 658         }
 659
 660         pipe_reference_init(&f->reference, 1);
 661
 662         return f;
 663 }
 664
 665 static const void *
 666 panfrost_screen_get_compiler_options(struct pipe_screen *pscreen,
 667                                      enum pipe_shader_ir ir,
 668                                      enum pipe_shader_type shader)
 669 {
 670         return &midgard_nir_options;
 671 }
 672
 673 static __u64
 674 panfrost_query_raw(
 675                 struct panfrost_screen *screen,
 676                 enum drm_panfrost_param param,
 677                 bool required)
 678 {
 679         struct drm_panfrost_get_param get_param = {0,};
 680         ASSERTED int ret;
 681
 682         get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
 683         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
 684
 685         assert(!(ret && required));
 686
 687         return get_param.value;
 688 }
 689
 690 static unsigned
 691 panfrost_query_gpu_version(struct panfrost_screen *screen)
 692 {
 693         return panfrost_query_raw(screen, DRM_PANFROST_PARAM_GPU_PROD_ID, true);
 694 }
 695
 696 static uint32_t
 697 panfrost_active_bos_hash(const void *key)
 698 {
 699         const struct panfrost_bo *bo = key;
 700
 701         return _mesa_hash_data(&bo->gem_handle, sizeof(bo->gem_handle));
 702 }
 703
 704 static bool
 705 panfrost_active_bos_cmp(const void *keya, const void *keyb)
 706 {
 707         const struct panfrost_bo *a = keya, *b = keyb;
 708
 709         return a->gem_handle == b->gem_handle;
 710 }
 711
 712 struct pipe_screen *
 713 panfrost_create_screen(int fd, struct renderonly *ro)
 714 {
 715         pan_debug = debug_get_option_pan_debug();
 716
 717         /* Blacklist apps known to be buggy under Panfrost */
 718         const char *proc = util_get_process_name();
 719         const char *blacklist[] = {
 720                 "chromium",
 721                 "chrome",
 722         };
 723
 724         for (unsigned i = 0; i < ARRAY_SIZE(blacklist); ++i) {
 725                 if ((strcmp(blacklist[i], proc) == 0))
 726                         return NULL;
 727         }
 728
 729         /* Create the screen */
 730         struct panfrost_screen *screen = rzalloc(NULL, struct panfrost_screen);
 731
 732         if (!screen)
 733                 return NULL;
 734
 735         if (ro) {
 736                 screen->ro = renderonly_dup(ro);
 737                 if (!screen->ro) {
 738                         fprintf(stderr, "Failed to dup renderonly object\n");
 739                         free(screen);
 740                         return NULL;
 741                 }
 742         }
 743
 744         screen->fd = fd;
 745
 746         screen->gpu_id = panfrost_query_gpu_version(screen);
 747         screen->quirks = panfrost_get_quirks(screen->gpu_id);
 748         screen->kernel_version = drmGetVersion(fd);
 749
 750         /* Check if we're loading against a supported GPU model. */
 751
 752         switch (screen->gpu_id) {
 753         case 0x720: /* T720 */
 754         case 0x750: /* T760 */
 755         case 0x820: /* T820 */
 756         case 0x860: /* T860 */
 757                 break;
 758         default:
 759                 /* Fail to load against untested models */
 760                 debug_printf("panfrost: Unsupported model %X", screen->gpu_id);
 761                 return NULL;
 762         }
 763
 764         pthread_mutex_init(&screen->active_bos_lock, NULL);
 765         screen->active_bos = _mesa_set_create(screen, panfrost_active_bos_hash,
 766                                               panfrost_active_bos_cmp);
 767
 768         pthread_mutex_init(&screen->bo_cache.lock, NULL);
 769         list_inithead(&screen->bo_cache.lru);
 770         for (unsigned i = 0; i < ARRAY_SIZE(screen->bo_cache.buckets); ++i)
 771                 list_inithead(&screen->bo_cache.buckets[i]);
 772
 773         if (pan_debug & PAN_DBG_TRACE)
 774                 pandecode_initialize();
 775
 776         screen->base.destroy = panfrost_destroy_screen;
 777
 778         screen->base.get_name = panfrost_get_name;
 779         screen->base.get_vendor = panfrost_get_vendor;
 780         screen->base.get_device_vendor = panfrost_get_device_vendor;
 781         screen->base.get_param = panfrost_get_param;
 782         screen->base.get_shader_param = panfrost_get_shader_param;
 783         screen->base.get_compute_param = panfrost_get_compute_param;
 784         screen->base.get_paramf = panfrost_get_paramf;
 785         screen->base.get_timestamp = panfrost_get_timestamp;
 786         screen->base.is_format_supported = panfrost_is_format_supported;
 787         screen->base.context_create = panfrost_create_context;
 788         screen->base.flush_frontbuffer = panfrost_flush_frontbuffer;
 789         screen->base.get_compiler_options = panfrost_screen_get_compiler_options;
 790         screen->base.fence_reference = panfrost_fence_reference;
 791         screen->base.fence_finish = panfrost_fence_finish;
 792         screen->base.set_damage_region = panfrost_resource_set_damage_region;
 793
 794         panfrost_resource_screen_init(screen);
 795
 796         return &screen->base;
 797 }