* Copyright (C) 2014 Broadcom
* Copyright (C) 2018 Alyssa Rosenzweig
* Copyright (C) 2019 Collabora, Ltd.
+ * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
#include "util/u_debug.h"
#include "util/u_memory.h"
-#include "util/u_format.h"
-#include "util/u_format_s3tc.h"
+#include "util/format/u_format.h"
+#include "util/format/u_format_s3tc.h"
#include "util/u_video.h"
#include "util/u_screen.h"
#include "util/os_time.h"
+#include "util/u_process.h"
#include "pipe/p_defines.h"
#include "pipe/p_screen.h"
#include "draw/draw_context.h"
-#include <xf86drm.h>
#include <fcntl.h>
#include "drm-uapi/drm_fourcc.h"
+#include "drm-uapi/panfrost_drm.h"
+#include "pan_bo.h"
#include "pan_screen.h"
#include "pan_resource.h"
#include "pan_public.h"
#include "pan_context.h"
#include "midgard/midgard_compile.h"
+#include "panfrost-quirks.h"
static const struct debug_named_value debug_options[] = {
{"msgs", PAN_DBG_MSGS, "Print debug messages"},
{"trace", PAN_DBG_TRACE, "Trace the command stream"},
{"deqp", PAN_DBG_DEQP, "Hacks for dEQP"},
{"afbc", PAN_DBG_AFBC, "Enable non-conformant AFBC impl"},
+ {"sync", PAN_DBG_SYNC, "Wait for each job's completion and check for any GPU fault"},
+ {"precompile", PAN_DBG_PRECOMPILE, "Precompile shaders for shader-db"},
DEBUG_NAMED_VALUE_END
};
static const char *
panfrost_get_name(struct pipe_screen *screen)
{
- return "panfrost";
+ return panfrost_model_name(pan_screen(screen)->gpu_id);
}
static const char *
panfrost_get_vendor(struct pipe_screen *screen)
{
- return "panfrost";
+ return "Panfrost";
}
static const char *
case PIPE_CAP_MAX_RENDER_TARGETS:
return is_deqp ? 4 : 1;
+ /* Throttling frames breaks pipelining */
+ case PIPE_CAP_THROTTLE:
+ return 0;
case PIPE_CAP_OCCLUSION_QUERY:
return 1;
case PIPE_CAP_QUERY_SO_OVERFLOW:
return 0;
- case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
case PIPE_CAP_TEXTURE_SWIZZLE:
return 1;
+ case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
+ case PIPE_CAP_TEXTURE_MIRROR_CLAMP_TO_EDGE:
+ return 1;
+
case PIPE_CAP_TGSI_INSTANCEID:
case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
- return is_deqp ? 1 : 0;
+ return 1;
case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
return is_deqp ? 4 : 0;
case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
return is_deqp ? 64 : 0;
+ case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
+ return 1;
case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
- return is_deqp ? 256 : 0; /* for GL3 */
+ return 256;
case PIPE_CAP_GLSL_FEATURE_LEVEL:
case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY:
return is_deqp ? 300 : 120;
case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
- return is_deqp ? 16 : 0;
+ return 16;
case PIPE_CAP_CUBE_MAP_ARRAY:
return is_deqp;
case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
return 0;
+ /* I really don't want to set this CAP but let's not swim against the
+ * tide.. */
+ case PIPE_CAP_TGSI_TEXCOORD:
+ return 1;
+
case PIPE_CAP_SEAMLESS_CUBE_MAP:
case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
return 1;
case PIPE_CAP_MAX_VARYINGS:
return 16;
+ case PIPE_CAP_ALPHA_TEST:
+ case PIPE_CAP_FLATSHADE:
+ case PIPE_CAP_TWO_SIDED_COLOR:
+ case PIPE_CAP_CLIP_PLANES:
+ return 0;
+
default:
return u_pipe_screen_get_param_defaults(screen, param);
}
/* this is probably not totally correct.. but it's a start: */
switch (param) {
- case PIPE_SHADER_CAP_SCALAR_ISA:
- return 0;
-
case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
- return 0;
case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
return 16;
case PIPE_SHADER_CAP_MAX_OUTPUTS:
- return shader == PIPE_SHADER_FRAGMENT ? 4 : 8;
+ return shader == PIPE_SHADER_FRAGMENT ? 4 : 16;
case PIPE_SHADER_CAP_MAX_TEMPS:
return 256; /* GL_MAX_PROGRAM_TEMPORARIES_ARB */
return PIPE_SHADER_IR_NIR;
case PIPE_SHADER_CAP_SUPPORTED_IRS:
- return 0;
+ return (1 << PIPE_SHADER_IR_NIR) | (1 << PIPE_SHADER_IR_NIR_SERIALIZED);
case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
return 32;
case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
- return is_deqp;
+ return is_deqp ? 4 : 0;
case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTERS:
case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTER_BUFFERS:
return 0;
if (!format_desc)
return false;
- if (sample_count > 1)
+ /* MSAA 4x supported, but no more. Technically some revisions of the
+ * hardware can go up to 16x but we don't support higher modes yet. */
+
+ if (sample_count > 1 && !(pan_debug & PAN_DBG_DEQP))
+ return false;
+
+ if (sample_count > 4)
+ return false;
+
+ if (MAX2(sample_count, 1) != MAX2(storage_sample_count, 1))
return false;
/* Format wishlist */
if (scanout && renderable && !util_format_is_rgba8_variant(format_desc))
return false;
- if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN &&
- format_desc->layout != UTIL_FORMAT_LAYOUT_OTHER) {
- /* Compressed formats not yet hooked up. */
- return false;
+ switch (format_desc->layout) {
+ case UTIL_FORMAT_LAYOUT_PLAIN:
+ case UTIL_FORMAT_LAYOUT_OTHER:
+ break;
+ case UTIL_FORMAT_LAYOUT_ETC:
+ case UTIL_FORMAT_LAYOUT_ASTC:
+ return true;
+ default:
+ return false;
}
/* Internally, formats that are depth/stencil renderable are limited.
return true;
}
+static int
+panfrost_get_compute_param(struct pipe_screen *pscreen, enum pipe_shader_ir ir_type,
+ enum pipe_compute_cap param, void *ret)
+{
+ const char * const ir = "panfrost";
+
+ if (!(pan_debug & PAN_DBG_DEQP))
+ return 0;
+
+#define RET(x) do { \
+ if (ret) \
+ memcpy(ret, x, sizeof(x)); \
+ return sizeof(x); \
+} while (0)
+
+ switch (param) {
+ case PIPE_COMPUTE_CAP_ADDRESS_BITS:
+ RET((uint32_t []){ 64 });
+
+ case PIPE_COMPUTE_CAP_IR_TARGET:
+ if (ret)
+ sprintf(ret, "%s", ir);
+ return strlen(ir) * sizeof(char);
+
+ case PIPE_COMPUTE_CAP_GRID_DIMENSION:
+ RET((uint64_t []) { 3 });
+
+ case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
+ RET(((uint64_t []) { 65535, 65535, 65535 }));
+
+ case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
+ RET(((uint64_t []) { 1024, 1024, 64 }));
+
+ case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
+ RET((uint64_t []) { 1024 });
+
+ case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
+ RET((uint64_t []) { 1024*1024*512 /* Maybe get memory */ });
+
+ case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
+ RET((uint64_t []) { 32768 });
+
+ case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
+ case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
+ RET((uint64_t []) { 4096 });
+
+ case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
+ RET((uint64_t []) { 1024*1024*512 /* Maybe get memory */ });
+
+ case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
+ RET((uint32_t []) { 800 /* MHz -- TODO */ });
+
+ case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
+ RET((uint32_t []) { 9999 }); // TODO
+
+ case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
+ RET((uint32_t []) { 1 }); // TODO
+
+ case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
+ RET((uint32_t []) { 32 }); // TODO
+
+ case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
+ RET((uint64_t []) { 1024 }); // TODO
+ }
+
+ return 0;
+}
static void
panfrost_destroy_screen(struct pipe_screen *pscreen)
{
struct panfrost_screen *screen = pan_screen(pscreen);
panfrost_bo_cache_evict_all(screen);
+ pthread_mutex_destroy(&screen->bo_cache.lock);
+ pthread_mutex_destroy(&screen->active_bos_lock);
+ drmFreeVersion(screen->kernel_version);
ralloc_free(screen);
}
struct pipe_fence_handle **ptr,
struct pipe_fence_handle *fence)
{
- panfrost_drm_fence_reference(pscreen, ptr, fence);
+ struct panfrost_fence **p = (struct panfrost_fence **)ptr;
+ struct panfrost_fence *f = (struct panfrost_fence *)fence;
+ struct panfrost_fence *old = *p;
+
+ if (pipe_reference(&(*p)->reference, &f->reference)) {
+ util_dynarray_foreach(&old->syncfds, int, fd)
+ close(*fd);
+ util_dynarray_fini(&old->syncfds);
+ free(old);
+ }
+ *p = f;
}
static bool
struct pipe_fence_handle *fence,
uint64_t timeout)
{
- return panfrost_drm_fence_finish(pscreen, ctx, fence, timeout);
+ struct panfrost_screen *screen = pan_screen(pscreen);
+ struct panfrost_fence *f = (struct panfrost_fence *)fence;
+ struct util_dynarray syncobjs;
+ int ret;
+
+ /* All fences were already signaled */
+ if (!util_dynarray_num_elements(&f->syncfds, int))
+ return true;
+
+ util_dynarray_init(&syncobjs, NULL);
+ util_dynarray_foreach(&f->syncfds, int, fd) {
+ uint32_t syncobj;
+
+ ret = drmSyncobjCreate(screen->fd, 0, &syncobj);
+ assert(!ret);
+
+ ret = drmSyncobjImportSyncFile(screen->fd, syncobj, *fd);
+ assert(!ret);
+ util_dynarray_append(&syncobjs, uint32_t, syncobj);
+ }
+
+ uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
+ if (abs_timeout == OS_TIMEOUT_INFINITE)
+ abs_timeout = INT64_MAX;
+
+ ret = drmSyncobjWait(screen->fd, util_dynarray_begin(&syncobjs),
+ util_dynarray_num_elements(&syncobjs, uint32_t),
+ abs_timeout, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL,
+ NULL);
+
+ util_dynarray_foreach(&syncobjs, uint32_t, syncobj)
+ drmSyncobjDestroy(screen->fd, *syncobj);
+
+ return ret >= 0;
+}
+
+struct panfrost_fence *
+panfrost_fence_create(struct panfrost_context *ctx,
+ struct util_dynarray *fences)
+{
+ struct panfrost_screen *screen = pan_screen(ctx->base.screen);
+ struct panfrost_fence *f = calloc(1, sizeof(*f));
+ if (!f)
+ return NULL;
+
+ util_dynarray_init(&f->syncfds, NULL);
+
+ /* Export fences from all pending batches. */
+ util_dynarray_foreach(fences, struct panfrost_batch_fence *, fence) {
+ int fd = -1;
+
+ /* The fence is already signaled, no need to export it. */
+ if ((*fence)->signaled)
+ continue;
+
+ drmSyncobjExportSyncFile(screen->fd, (*fence)->syncobj, &fd);
+ if (fd == -1)
+ fprintf(stderr, "export failed: %m\n");
+
+ assert(fd != -1);
+ util_dynarray_append(&f->syncfds, int, fd);
+ }
+
+ pipe_reference_init(&f->reference, 1);
+
+ return f;
}
static const void *
return &midgard_nir_options;
}
+static uint32_t
+panfrost_active_bos_hash(const void *key)
+{
+ const struct panfrost_bo *bo = key;
+
+ return _mesa_hash_data(&bo->gem_handle, sizeof(bo->gem_handle));
+}
+
+static bool
+panfrost_active_bos_cmp(const void *keya, const void *keyb)
+{
+ const struct panfrost_bo *a = keya, *b = keyb;
+
+ return a->gem_handle == b->gem_handle;
+}
+
struct pipe_screen *
panfrost_create_screen(int fd, struct renderonly *ro)
{
- struct panfrost_screen *screen = rzalloc(NULL, struct panfrost_screen);
-
pan_debug = debug_get_option_pan_debug();
+ /* Blacklist apps known to be buggy under Panfrost */
+ const char *proc = util_get_process_name();
+ const char *blacklist[] = {
+ "chromium",
+ "chrome",
+ };
+
+ for (unsigned i = 0; i < ARRAY_SIZE(blacklist); ++i) {
+ if ((strcmp(blacklist[i], proc) == 0))
+ return NULL;
+ }
+
+ /* Create the screen */
+ struct panfrost_screen *screen = rzalloc(NULL, struct panfrost_screen);
+
if (!screen)
return NULL;
screen->fd = fd;
- screen->gpu_id = panfrost_drm_query_gpu_version(screen);
- screen->require_sfbd = screen->gpu_id < 0x0750; /* T760 is the first to support MFBD */
+ screen->gpu_id = panfrost_query_gpu_version(screen->fd);
+ screen->core_count = panfrost_query_core_count(screen->fd);
+ screen->thread_tls_alloc = panfrost_query_thread_tls_alloc(screen->fd);
+ screen->quirks = panfrost_get_quirks(screen->gpu_id);
+ screen->kernel_version = drmGetVersion(fd);
/* Check if we're loading against a supported GPU model. */
switch (screen->gpu_id) {
+ case 0x720: /* T720 */
case 0x750: /* T760 */
case 0x820: /* T820 */
case 0x860: /* T860 */
break;
default:
/* Fail to load against untested models */
- debug_printf("panfrost: Unsupported model %X",
- screen->gpu_id);
+ debug_printf("panfrost: Unsupported model %X", screen->gpu_id);
return NULL;
}
- util_dynarray_init(&screen->transient_bo, screen);
+ pthread_mutex_init(&screen->active_bos_lock, NULL);
+ screen->active_bos = _mesa_set_create(screen, panfrost_active_bos_hash,
+ panfrost_active_bos_cmp);
- for (unsigned i = 0; i < ARRAY_SIZE(screen->bo_cache); ++i)
- list_inithead(&screen->bo_cache[i]);
+ pthread_mutex_init(&screen->bo_cache.lock, NULL);
+ list_inithead(&screen->bo_cache.lru);
+ for (unsigned i = 0; i < ARRAY_SIZE(screen->bo_cache.buckets); ++i)
+ list_inithead(&screen->bo_cache.buckets[i]);
if (pan_debug & PAN_DBG_TRACE)
pandecode_initialize();
screen->base.get_device_vendor = panfrost_get_device_vendor;
screen->base.get_param = panfrost_get_param;
screen->base.get_shader_param = panfrost_get_shader_param;
+ screen->base.get_compute_param = panfrost_get_compute_param;
screen->base.get_paramf = panfrost_get_paramf;
screen->base.get_timestamp = panfrost_get_timestamp;
screen->base.is_format_supported = panfrost_is_format_supported;
screen->base.get_compiler_options = panfrost_screen_get_compiler_options;
screen->base.fence_reference = panfrost_fence_reference;
screen->base.fence_finish = panfrost_fence_finish;
-
- screen->last_fragment_flushed = true;
- screen->last_job = NULL;
+ screen->base.set_damage_region = panfrost_resource_set_damage_region;
panfrost_resource_screen_init(screen);
projects / mesa.git / blobdiff