* IN THE SOFTWARE.
*/
+#include "dirent.h"
+#include <errno.h>
+#include <fcntl.h>
+#include <linux/audit.h>
+#include <linux/bpf.h>
+#include <linux/filter.h>
+#include <linux/seccomp.h>
+#include <linux/unistd.h>
#include <stdbool.h>
+#include <stddef.h>
+#include <stdio.h>
#include <string.h>
+#include <sys/prctl.h>
+#include <sys/wait.h>
#include <unistd.h>
#include <fcntl.h>
#include <llvm/Config/llvm-config.h>
+
#include "radv_debug.h"
#include "radv_private.h"
#include "radv_shader.h"
#include "radv_cs.h"
#include "util/disk_cache.h"
-#include "util/strtod.h"
#include "vk_util.h"
#include <xf86drm.h>
#include <amdgpu.h>
#include "util/build_id.h"
#include "util/debug.h"
#include "util/mesa-sha1.h"
+#include "util/timespec.h"
+#include "util/u_atomic.h"
#include "compiler/glsl_types.h"
#include "util/xmlpool.h"
+static struct radv_timeline_point *
+radv_timeline_find_point_at_least_locked(struct radv_device *device,
+ struct radv_timeline *timeline,
+ uint64_t p);
+
+static struct radv_timeline_point *
+radv_timeline_add_point_locked(struct radv_device *device,
+ struct radv_timeline *timeline,
+ uint64_t p);
+
+static void
+radv_timeline_trigger_waiters_locked(struct radv_timeline *timeline,
+ struct list_head *processing_list);
+
+static
+void radv_destroy_semaphore_part(struct radv_device *device,
+ struct radv_semaphore_part *part);
+
static int
radv_device_get_cache_uuid(enum radeon_family family, void *uuid)
{
ac_compute_device_uuid(info, uuid, VK_UUID_SIZE);
}
-static void
-radv_get_device_name(enum radeon_family family, char *name, size_t name_len, bool aco)
-{
- const char *chip_string;
-
- switch (family) {
- case CHIP_TAHITI: chip_string = "TAHITI"; break;
- case CHIP_PITCAIRN: chip_string = "PITCAIRN"; break;
- case CHIP_VERDE: chip_string = "CAPE VERDE"; break;
- case CHIP_OLAND: chip_string = "OLAND"; break;
- case CHIP_HAINAN: chip_string = "HAINAN"; break;
- case CHIP_BONAIRE: chip_string = "BONAIRE"; break;
- case CHIP_KAVERI: chip_string = "KAVERI"; break;
- case CHIP_KABINI: chip_string = "KABINI"; break;
- case CHIP_HAWAII: chip_string = "HAWAII"; break;
- case CHIP_TONGA: chip_string = "TONGA"; break;
- case CHIP_ICELAND: chip_string = "ICELAND"; break;
- case CHIP_CARRIZO: chip_string = "CARRIZO"; break;
- case CHIP_FIJI: chip_string = "FIJI"; break;
- case CHIP_POLARIS10: chip_string = "POLARIS10"; break;
- case CHIP_POLARIS11: chip_string = "POLARIS11"; break;
- case CHIP_POLARIS12: chip_string = "POLARIS12"; break;
- case CHIP_STONEY: chip_string = "STONEY"; break;
- case CHIP_VEGAM: chip_string = "VEGA M"; break;
- case CHIP_VEGA10: chip_string = "VEGA10"; break;
- case CHIP_VEGA12: chip_string = "VEGA12"; break;
- case CHIP_VEGA20: chip_string = "VEGA20"; break;
- case CHIP_RAVEN: chip_string = "RAVEN"; break;
- case CHIP_RAVEN2: chip_string = "RAVEN2"; break;
- case CHIP_NAVI10: chip_string = "NAVI10"; break;
- case CHIP_NAVI12: chip_string = "NAVI12"; break;
- case CHIP_NAVI14: chip_string = "NAVI14"; break;
- default: chip_string = "unknown"; break;
- }
-
- snprintf(name, name_len, "AMD RADV%s %s (LLVM " MESA_LLVM_VERSION_STRING ")", aco ? "/ACO" : "", chip_string);
-}
-
static uint64_t
radv_get_visible_vram_size(struct radv_physical_device *device)
{
return device->rad_info.vram_size - radv_get_visible_vram_size(device);
}
+static bool
+radv_is_mem_type_vram(enum radv_mem_type type)
+{
+ return type == RADV_MEM_TYPE_VRAM ||
+ type == RADV_MEM_TYPE_VRAM_UNCACHED;
+}
+
+static bool
+radv_is_mem_type_vram_visible(enum radv_mem_type type)
+{
+ return type == RADV_MEM_TYPE_VRAM_CPU_ACCESS ||
+ type == RADV_MEM_TYPE_VRAM_CPU_ACCESS_UNCACHED;
+}
+static bool
+radv_is_mem_type_gtt_wc(enum radv_mem_type type)
+{
+ return type == RADV_MEM_TYPE_GTT_WRITE_COMBINE ||
+ type == RADV_MEM_TYPE_GTT_WRITE_COMBINE_VRAM_UNCACHED;
+}
+
+static bool
+radv_is_mem_type_gtt_cached(enum radv_mem_type type)
+{
+ return type == RADV_MEM_TYPE_GTT_CACHED ||
+ type == RADV_MEM_TYPE_GTT_CACHED_VRAM_UNCACHED;
+}
+
+static bool
+radv_is_mem_type_uncached(enum radv_mem_type type)
+{
+ return type == RADV_MEM_TYPE_VRAM_UNCACHED ||
+ type == RADV_MEM_TYPE_VRAM_CPU_ACCESS_UNCACHED ||
+ type == RADV_MEM_TYPE_GTT_WRITE_COMBINE_VRAM_UNCACHED ||
+ type == RADV_MEM_TYPE_GTT_CACHED_VRAM_UNCACHED;
+}
+
static void
radv_physical_device_init_mem_types(struct radv_physical_device *device)
{
};
}
device->memory_properties.memoryTypeCount = type_count;
+
+ if (device->rad_info.has_l2_uncached) {
+ for (int i = 0; i < device->memory_properties.memoryTypeCount; i++) {
+ VkMemoryType mem_type = device->memory_properties.memoryTypes[i];
+
+ if ((mem_type.propertyFlags & (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)) ||
+ mem_type.propertyFlags == VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) {
+ enum radv_mem_type mem_type_id;
+
+ switch (device->mem_type_indices[i]) {
+ case RADV_MEM_TYPE_VRAM:
+ mem_type_id = RADV_MEM_TYPE_VRAM_UNCACHED;
+ break;
+ case RADV_MEM_TYPE_VRAM_CPU_ACCESS:
+ mem_type_id = RADV_MEM_TYPE_VRAM_CPU_ACCESS_UNCACHED;
+ break;
+ case RADV_MEM_TYPE_GTT_WRITE_COMBINE:
+ mem_type_id = RADV_MEM_TYPE_GTT_WRITE_COMBINE_VRAM_UNCACHED;
+ break;
+ case RADV_MEM_TYPE_GTT_CACHED:
+ mem_type_id = RADV_MEM_TYPE_GTT_CACHED_VRAM_UNCACHED;
+ break;
+ default:
+ unreachable("invalid memory type");
+ }
+
+ VkMemoryPropertyFlags property_flags = mem_type.propertyFlags |
+ VK_MEMORY_PROPERTY_DEVICE_COHERENT_BIT_AMD |
+ VK_MEMORY_PROPERTY_DEVICE_UNCACHED_BIT_AMD;
+
+ device->mem_type_indices[type_count] = mem_type_id;
+ device->memory_properties.memoryTypes[type_count++] = (VkMemoryType) {
+ .propertyFlags = property_flags,
+ .heapIndex = mem_type.heapIndex,
+ };
+ }
+ }
+ device->memory_properties.memoryTypeCount = type_count;
+ }
}
static void
if (!strcmp(family, ac_get_llvm_processor_name(i))) {
/* Override family and chip_class. */
device->rad_info.family = i;
+ device->rad_info.name = "OVERRIDDEN";
if (i >= CHIP_NAVI10)
device->rad_info.chip_class = GFX10;
else
device->rad_info.chip_class = GFX6;
+ /* Don't submit any IBs. */
+ device->instance->debug_flags |= RADV_DEBUG_NOOP;
return;
}
}
radv_handle_env_var_force_family(device);
device->use_aco = instance->perftest_flags & RADV_PERFTEST_ACO;
- if ((device->rad_info.chip_class < GFX8 ||
- device->rad_info.chip_class > GFX9) && device->use_aco) {
- fprintf(stderr, "WARNING: disabling ACO on unsupported GPUs.\n");
- device->use_aco = false;
- }
- radv_get_device_name(device->rad_info.family, device->name, sizeof(device->name), device->use_aco);
+ snprintf(device->name, sizeof(device->name),
+ "AMD RADV%s %s (LLVM " MESA_LLVM_VERSION_STRING ")", device->use_aco ? "/ACO" : "",
+ device->rad_info.name);
if (radv_device_get_cache_uuid(device->rad_info.family, device->cache_uuid)) {
device->ws->destroy(device->ws);
/* These flags affect shader compilation. */
uint64_t shader_env_flags =
(device->instance->perftest_flags & RADV_PERFTEST_SISCHED ? 0x1 : 0) |
- (device->instance->debug_flags & RADV_DEBUG_UNSAFE_MATH ? 0x2 : 0) |
- (device->use_aco ? 0x4 : 0);
+ (device->use_aco ? 0x2 : 0);
/* The gpu id is already embedded in the uuid so we just pass "radv"
* when creating the cache.
disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
device->disk_cache = disk_cache_create(device->name, buf, shader_env_flags);
- if (device->rad_info.chip_class < GFX8 ||
- device->rad_info.chip_class > GFX9)
+ if (device->rad_info.chip_class < GFX8)
fprintf(stderr, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
radv_get_driver_uuid(&device->driver_uuid);
device->dcc_msaa_allowed =
(device->instance->perftest_flags & RADV_PERFTEST_DCC_MSAA);
- device->use_shader_ballot = device->rad_info.chip_class >= GFX8 &&
- (device->use_aco || device->instance->perftest_flags & RADV_PERFTEST_SHADER_BALLOT);
+ device->use_shader_ballot = (device->use_aco && device->rad_info.chip_class >= GFX8) ||
+ (device->instance->perftest_flags & RADV_PERFTEST_SHADER_BALLOT);
+
+ device->use_ngg = device->rad_info.chip_class >= GFX10 &&
+ device->rad_info.family != CHIP_NAVI14 &&
+ !(device->instance->debug_flags & RADV_DEBUG_NO_NGG);
+ if (device->use_aco && device->use_ngg) {
+ fprintf(stderr, "WARNING: disabling NGG because ACO is used.\n");
+ device->use_ngg = false;
+ }
device->use_ngg_streamout = false;
{"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS},
{"nohiz", RADV_DEBUG_NO_HIZ},
{"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE},
- {"unsafemath", RADV_DEBUG_UNSAFE_MATH},
{"allbos", RADV_DEBUG_ALL_BOS},
{"noibs", RADV_DEBUG_NO_IBS},
{"spirv", RADV_DEBUG_DUMP_SPIRV},
{"nongg", RADV_DEBUG_NO_NGG},
{"noshaderballot", RADV_DEBUG_NO_SHADER_BALLOT},
{"allentrypoints", RADV_DEBUG_ALL_ENTRYPOINTS},
+ {"metashaders", RADV_DEBUG_DUMP_META_SHADERS},
+ {"nomemorycache", RADV_DEBUG_NO_MEMORY_CACHE},
+ {"noop", RADV_DEBUG_NOOP},
{NULL, 0}
};
if (LLVM_VERSION_MAJOR < 9)
instance->debug_flags |= RADV_DEBUG_NO_LOAD_STORE_OPT;
} else if (!strcmp(name, "Wolfenstein: Youngblood")) {
- if (!(instance->debug_flags & RADV_DEBUG_NO_SHADER_BALLOT)) {
+ if (!(instance->debug_flags & RADV_DEBUG_NO_SHADER_BALLOT) &&
+ !(instance->perftest_flags & RADV_PERFTEST_ACO)) {
/* Force enable VK_AMD_shader_ballot because it looks
* safe and it gives a nice boost (+20% on Vega 56 at
- * this time).
+ * this time). It also prevents corruption on LLVM.
*/
instance->perftest_flags |= RADV_PERFTEST_SHADER_BALLOT;
}
DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
DRI_CONF_SECTION_END
+
+ DRI_CONF_SECTION_DEBUG
+ DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
+ DRI_CONF_SECTION_END
DRI_CONF_END;
static void radv_init_dri_options(struct radv_instance *instance)
instance->apiVersion = client_version;
instance->physicalDeviceCount = -1;
+ /* Get secure compile thread count. NOTE: We cap this at 32 */
+#define MAX_SC_PROCS 32
+ char *num_sc_threads = getenv("RADV_SECURE_COMPILE_THREADS");
+ if (num_sc_threads)
+ instance->num_sc_threads = MIN2(strtoul(num_sc_threads, NULL, 10), MAX_SC_PROCS);
+
instance->debug_flags = parse_debug_string(getenv("RADV_DEBUG"),
radv_debug_options);
+ /* Disable memory cache when secure compile is set */
+ if (radv_device_use_secure_compile(instance))
+ instance->debug_flags |= RADV_DEBUG_NO_MEMORY_CACHE;
+
instance->perftest_flags = parse_debug_string(getenv("RADV_PERFTEST"),
radv_perftest_options);
VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
instance->engineVersion = engine_version;
- _mesa_locale_init();
glsl_type_singleton_init_or_ref();
VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
VG(VALGRIND_DESTROY_MEMPOOL(instance));
glsl_type_singleton_decref();
- _mesa_locale_fini();
driDestroyOptionCache(&instance->dri_options);
driDestroyOptionInfo(&instance->available_dri_options);
features->samplerYcbcrConversion = true;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
- VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
- (VkPhysicalDeviceDescriptorIndexingFeaturesEXT*)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES: {
+ VkPhysicalDeviceDescriptorIndexingFeatures *features =
+ (VkPhysicalDeviceDescriptorIndexingFeatures*)ext;
features->shaderInputAttachmentArrayDynamicIndexing = true;
features->shaderUniformTexelBufferArrayDynamicIndexing = true;
features->shaderStorageTexelBufferArrayDynamicIndexing = true;
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
(VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
- features->vertexAttributeInstanceRateDivisor = VK_TRUE;
- features->vertexAttributeInstanceRateZeroDivisor = VK_TRUE;
+ features->vertexAttributeInstanceRateDivisor = true;
+ features->vertexAttributeInstanceRateZeroDivisor = true;
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
features->geometryStreams = !pdevice->use_ngg_streamout;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
- VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
- (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES: {
+ VkPhysicalDeviceScalarBlockLayoutFeatures *features =
+ (VkPhysicalDeviceScalarBlockLayoutFeatures *)ext;
features->scalarBlockLayout = pdevice->rad_info.chip_class >= GFX7;
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT: {
VkPhysicalDeviceMemoryPriorityFeaturesEXT *features =
(VkPhysicalDeviceMemoryPriorityFeaturesEXT *)ext;
- features->memoryPriority = VK_TRUE;
+ features->memoryPriority = true;
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
features->bufferDeviceAddressMultiDevice = false;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES: {
+ VkPhysicalDeviceBufferDeviceAddressFeatures *features =
+ (VkPhysicalDeviceBufferDeviceAddressFeatures *)ext;
+ features->bufferDeviceAddress = true;
+ features->bufferDeviceAddressCaptureReplay = false;
+ features->bufferDeviceAddressMultiDevice = false;
+ break;
+ }
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
(VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
features->depthClipEnable = true;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
- VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
- (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES: {
+ VkPhysicalDeviceHostQueryResetFeatures *features =
+ (VkPhysicalDeviceHostQueryResetFeatures *)ext;
features->hostQueryReset = true;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
- VkPhysicalDevice8BitStorageFeaturesKHR *features =
- (VkPhysicalDevice8BitStorageFeaturesKHR*)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES: {
+ VkPhysicalDevice8BitStorageFeatures *features =
+ (VkPhysicalDevice8BitStorageFeatures *)ext;
bool enabled = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
features->storageBuffer8BitAccess = enabled;
features->uniformAndStorageBuffer8BitAccess = enabled;
features->storagePushConstant8 = enabled;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
- VkPhysicalDeviceFloat16Int8FeaturesKHR *features =
- (VkPhysicalDeviceFloat16Int8FeaturesKHR*)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_FLOAT16_INT8_FEATURES: {
+ VkPhysicalDeviceShaderFloat16Int8Features *features =
+ (VkPhysicalDeviceShaderFloat16Int8Features*)ext;
features->shaderFloat16 = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
features->shaderInt8 = !pdevice->use_aco;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
- VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features =
- (VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES: {
+ VkPhysicalDeviceShaderAtomicInt64Features *features =
+ (VkPhysicalDeviceShaderAtomicInt64Features *)ext;
features->shaderBufferInt64Atomics = LLVM_VERSION_MAJOR >= 9;
features->shaderSharedInt64Atomics = LLVM_VERSION_MAJOR >= 9;
break;
features->ycbcrImageArrays = true;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
- VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
- (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES: {
+ VkPhysicalDeviceUniformBufferStandardLayoutFeatures *features =
+ (VkPhysicalDeviceUniformBufferStandardLayoutFeatures *)ext;
features->uniformBufferStandardLayout = true;
break;
}
features->indexTypeUint8 = pdevice->rad_info.chip_class >= GFX8;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
- VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
- (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES: {
+ VkPhysicalDeviceImagelessFramebufferFeatures *features =
+ (VkPhysicalDeviceImagelessFramebufferFeatures *)ext;
features->imagelessFramebuffer = true;
break;
}
features->pipelineExecutableInfo = true;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
+ VkPhysicalDeviceShaderClockFeaturesKHR *features =
+ (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
+ features->shaderSubgroupClock = true;
+ features->shaderDeviceClock = false;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
+ VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
+ (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
+ features->texelBufferAlignment = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES: {
+ VkPhysicalDeviceTimelineSemaphoreFeatures *features =
+ (VkPhysicalDeviceTimelineSemaphoreFeatures *) ext;
+ features->timelineSemaphore = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
+ VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
+ (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
+ features->subgroupSizeControl = true;
+ features->computeFullSubgroups = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COHERENT_MEMORY_FEATURES_AMD: {
+ VkPhysicalDeviceCoherentMemoryFeaturesAMD *features =
+ (VkPhysicalDeviceCoherentMemoryFeaturesAMD *)ext;
+ features->deviceCoherentMemory = pdevice->rad_info.has_l2_uncached;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES: {
+ VkPhysicalDeviceShaderSubgroupExtendedTypesFeatures *features =
+ (VkPhysicalDeviceShaderSubgroupExtendedTypesFeatures *)ext;
+ features->shaderSubgroupExtendedTypes = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
+ VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
+ (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
+ features->separateDepthStencilLayouts = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES: {
+ VkPhysicalDeviceVulkan11Features *features =
+ (VkPhysicalDeviceVulkan11Features *)ext;
+ features->storageBuffer16BitAccess = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->uniformAndStorageBuffer16BitAccess = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->storagePushConstant16 = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->storageInputOutput16 = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco && LLVM_VERSION_MAJOR >= 9;
+ features->multiview = true;
+ features->multiviewGeometryShader = true;
+ features->multiviewTessellationShader = true;
+ features->variablePointersStorageBuffer = true;
+ features->variablePointers = true;
+ features->protectedMemory = false;
+ features->samplerYcbcrConversion = true;
+ features->shaderDrawParameters = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES: {
+ VkPhysicalDeviceVulkan12Features *features =
+ (VkPhysicalDeviceVulkan12Features *)ext;
+ features->samplerMirrorClampToEdge = true;
+ features->drawIndirectCount = true;
+ features->storageBuffer8BitAccess = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->uniformAndStorageBuffer8BitAccess = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->storagePushConstant8 = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->shaderBufferInt64Atomics = LLVM_VERSION_MAJOR >= 9;
+ features->shaderSharedInt64Atomics = LLVM_VERSION_MAJOR >= 9;
+ features->shaderFloat16 = pdevice->rad_info.chip_class >= GFX8 && !pdevice->use_aco;
+ features->shaderInt8 = !pdevice->use_aco;
+ features->descriptorIndexing = true;
+ features->shaderInputAttachmentArrayDynamicIndexing = true;
+ features->shaderUniformTexelBufferArrayDynamicIndexing = true;
+ features->shaderStorageTexelBufferArrayDynamicIndexing = true;
+ features->shaderUniformBufferArrayNonUniformIndexing = true;
+ features->shaderSampledImageArrayNonUniformIndexing = true;
+ features->shaderStorageBufferArrayNonUniformIndexing = true;
+ features->shaderStorageImageArrayNonUniformIndexing = true;
+ features->shaderInputAttachmentArrayNonUniformIndexing = true;
+ features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
+ features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
+ features->descriptorBindingUniformBufferUpdateAfterBind = true;
+ features->descriptorBindingSampledImageUpdateAfterBind = true;
+ features->descriptorBindingStorageImageUpdateAfterBind = true;
+ features->descriptorBindingStorageBufferUpdateAfterBind = true;
+ features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
+ features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
+ features->descriptorBindingUpdateUnusedWhilePending = true;
+ features->descriptorBindingPartiallyBound = true;
+ features->descriptorBindingVariableDescriptorCount = true;
+ features->runtimeDescriptorArray = true;
+ features->samplerFilterMinmax = pdevice->rad_info.chip_class >= GFX7;
+ features->scalarBlockLayout = pdevice->rad_info.chip_class >= GFX7;
+ features->imagelessFramebuffer = true;
+ features->uniformBufferStandardLayout = true;
+ features->shaderSubgroupExtendedTypes = true;
+ features->separateDepthStencilLayouts = true;
+ features->hostQueryReset = true;
+ features->timelineSemaphore = pdevice->rad_info.has_syncobj_wait_for_submit;
+ features->bufferDeviceAddress = true;
+ features->bufferDeviceAddressCaptureReplay = false;
+ features->bufferDeviceAddressMultiDevice = false;
+ features->vulkanMemoryModel = false;
+ features->vulkanMemoryModelDeviceScope = false;
+ features->vulkanMemoryModelAvailabilityVisibilityChains = false;
+ features->shaderOutputViewportIndex = true;
+ features->shaderOutputLayer = true;
+ features->subgroupBroadcastDynamicId = true;
+ break;
+ }
default:
break;
}
return radv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
}
-void radv_GetPhysicalDeviceProperties(
- VkPhysicalDevice physicalDevice,
- VkPhysicalDeviceProperties* pProperties)
+static size_t
+radv_max_descriptor_set_size()
{
- RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
- VkSampleCountFlags sample_counts = 0xf;
-
/* make sure that the entire descriptor set is addressable with a signed
* 32-bit int. So the sum of all limits scaled by descriptor size has to
* be at most 2 GiB. the combined image & samples object count as one of
* both. This limit is for the pipeline layout, not for the set layout, but
* there is no set limit, so we just set a pipeline limit. I don't think
* any app is going to hit this soon. */
- size_t max_descriptor_set_size = ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
+ return ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
+ - MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_INLINE_UNIFORM_BLOCK_COUNT) /
(32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
32 /* storage buffer, 32 due to potential space wasted on alignment */ +
32 /* sampler, largest when combined with image */ +
64 /* sampled image */ +
64 /* storage image */);
+}
+
+void radv_GetPhysicalDeviceProperties(
+ VkPhysicalDevice physicalDevice,
+ VkPhysicalDeviceProperties* pProperties)
+{
+ RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
+ VkSampleCountFlags sample_counts = 0xf;
+
+ size_t max_descriptor_set_size = radv_max_descriptor_set_size();
VkPhysicalDeviceLimits limits = {
.maxImageDimension1D = (1 << 14),
.maxFragmentCombinedOutputResources = 8,
.maxComputeSharedMemorySize = 32768,
.maxComputeWorkGroupCount = { 65535, 65535, 65535 },
- .maxComputeWorkGroupInvocations = 2048,
+ .maxComputeWorkGroupInvocations = 1024,
.maxComputeWorkGroupSize = {
- 2048,
- 2048,
- 2048
+ 1024,
+ 1024,
+ 1024
},
.subPixelPrecisionBits = 8,
.subTexelPrecisionBits = 8,
.viewportBoundsRange = { INT16_MIN, INT16_MAX },
.viewportSubPixelBits = 8,
.minMemoryMapAlignment = 4096, /* A page */
- .minTexelBufferOffsetAlignment = 1,
+ .minTexelBufferOffsetAlignment = 4,
.minUniformBufferOffsetAlignment = 4,
.minStorageBufferOffsetAlignment = 4,
.minTexelOffset = -32,
.framebufferNoAttachmentsSampleCounts = sample_counts,
.maxColorAttachments = MAX_RTS,
.sampledImageColorSampleCounts = sample_counts,
- .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
+ .sampledImageIntegerSampleCounts = sample_counts,
.sampledImageDepthSampleCounts = sample_counts,
.sampledImageStencilSampleCounts = sample_counts,
.storageImageSampleCounts = pdevice->rad_info.chip_class >= GFX8 ? sample_counts : VK_SAMPLE_COUNT_1_BIT,
.maxCombinedClipAndCullDistances = 8,
.discreteQueuePriorities = 2,
.pointSizeRange = { 0.0, 8192.0 },
- .lineWidthRange = { 0.0, 7.9921875 },
+ .lineWidthRange = { 0.0, 8192.0 },
.pointSizeGranularity = (1.0 / 8.0),
- .lineWidthGranularity = (1.0 / 128.0),
+ .lineWidthGranularity = (1.0 / 8.0),
.strictLines = false, /* FINISHME */
.standardSampleLocations = true,
.optimalBufferCopyOffsetAlignment = 128,
memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
}
+static void
+radv_get_physical_device_properties_1_1(struct radv_physical_device *pdevice,
+ VkPhysicalDeviceVulkan11Properties *p)
+{
+ assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
+
+ memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
+ memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
+ memset(p->deviceLUID, 0, VK_LUID_SIZE);
+ /* The LUID is for Windows. */
+ p->deviceLUIDValid = false;
+ p->deviceNodeMask = 0;
+
+ p->subgroupSize = RADV_SUBGROUP_SIZE;
+ p->subgroupSupportedStages = VK_SHADER_STAGE_ALL;
+ p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
+ VK_SUBGROUP_FEATURE_VOTE_BIT |
+ VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
+ VK_SUBGROUP_FEATURE_BALLOT_BIT |
+ VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
+ VK_SUBGROUP_FEATURE_QUAD_BIT;
+
+ if (pdevice->rad_info.chip_class == GFX8 ||
+ pdevice->rad_info.chip_class == GFX9) {
+ p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
+ VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT;
+ }
+ p->subgroupQuadOperationsInAllStages = true;
+
+ p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
+ p->maxMultiviewViewCount = MAX_VIEWS;
+ p->maxMultiviewInstanceIndex = INT_MAX;
+ p->protectedNoFault = false;
+ p->maxPerSetDescriptors = RADV_MAX_PER_SET_DESCRIPTORS;
+ p->maxMemoryAllocationSize = RADV_MAX_MEMORY_ALLOCATION_SIZE;
+}
+
+static void
+radv_get_physical_device_properties_1_2(struct radv_physical_device *pdevice,
+ VkPhysicalDeviceVulkan12Properties *p)
+{
+ assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
+
+ p->driverID = VK_DRIVER_ID_MESA_RADV;
+ snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE, "radv");
+ snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE,
+ "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
+ " (LLVM " MESA_LLVM_VERSION_STRING ")");
+ p->conformanceVersion = (VkConformanceVersion) {
+ .major = 1,
+ .minor = 2,
+ .subminor = 0,
+ .patch = 0,
+ };
+
+ /* On AMD hardware, denormals and rounding modes for fp16/fp64 are
+ * controlled by the same config register.
+ */
+ p->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR;
+ p->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR;
+
+ /* Do not allow both preserving and flushing denorms because different
+ * shaders in the same pipeline can have different settings and this
+ * won't work for merged shaders. To make it work, this requires LLVM
+ * support for changing the register. The same logic applies for the
+ * rounding modes because they are configured with the same config
+ * register. TODO: we can enable a lot of these for ACO when it
+ * supports all stages.
+ */
+ p->shaderDenormFlushToZeroFloat32 = true;
+ p->shaderDenormPreserveFloat32 = false;
+ p->shaderRoundingModeRTEFloat32 = true;
+ p->shaderRoundingModeRTZFloat32 = false;
+ p->shaderSignedZeroInfNanPreserveFloat32 = true;
+
+ p->shaderDenormFlushToZeroFloat16 = false;
+ p->shaderDenormPreserveFloat16 = pdevice->rad_info.chip_class >= GFX8;
+ p->shaderRoundingModeRTEFloat16 = pdevice->rad_info.chip_class >= GFX8;
+ p->shaderRoundingModeRTZFloat16 = false;
+ p->shaderSignedZeroInfNanPreserveFloat16 = pdevice->rad_info.chip_class >= GFX8;
+
+ p->shaderDenormFlushToZeroFloat64 = false;
+ p->shaderDenormPreserveFloat64 = pdevice->rad_info.chip_class >= GFX8;
+ p->shaderRoundingModeRTEFloat64 = pdevice->rad_info.chip_class >= GFX8;
+ p->shaderRoundingModeRTZFloat64 = false;
+ p->shaderSignedZeroInfNanPreserveFloat64 = pdevice->rad_info.chip_class >= GFX8;
+
+ p->maxUpdateAfterBindDescriptorsInAllPools = UINT32_MAX / 64;
+ p->shaderUniformBufferArrayNonUniformIndexingNative = false;
+ p->shaderSampledImageArrayNonUniformIndexingNative = false;
+ p->shaderStorageBufferArrayNonUniformIndexingNative = false;
+ p->shaderStorageImageArrayNonUniformIndexingNative = false;
+ p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
+ p->robustBufferAccessUpdateAfterBind = false;
+ p->quadDivergentImplicitLod = false;
+
+ size_t max_descriptor_set_size = ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS -
+ MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_INLINE_UNIFORM_BLOCK_COUNT) /
+ (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
+ 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
+ 32 /* sampler, largest when combined with image */ +
+ 64 /* sampled image */ +
+ 64 /* storage image */);
+ p->maxPerStageDescriptorUpdateAfterBindSamplers = max_descriptor_set_size;
+ p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = max_descriptor_set_size;
+ p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = max_descriptor_set_size;
+ p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_descriptor_set_size;
+ p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_descriptor_set_size;
+ p->maxPerStageDescriptorUpdateAfterBindInputAttachments = max_descriptor_set_size;
+ p->maxPerStageUpdateAfterBindResources = max_descriptor_set_size;
+ p->maxDescriptorSetUpdateAfterBindSamplers = max_descriptor_set_size;
+ p->maxDescriptorSetUpdateAfterBindUniformBuffers = max_descriptor_set_size;
+ p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS;
+ p->maxDescriptorSetUpdateAfterBindStorageBuffers = max_descriptor_set_size;
+ p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS;
+ p->maxDescriptorSetUpdateAfterBindSampledImages = max_descriptor_set_size;
+ p->maxDescriptorSetUpdateAfterBindStorageImages = max_descriptor_set_size;
+ p->maxDescriptorSetUpdateAfterBindInputAttachments = max_descriptor_set_size;
+
+ /* We support all of the depth resolve modes */
+ p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
+ VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
+ VK_RESOLVE_MODE_MIN_BIT_KHR |
+ VK_RESOLVE_MODE_MAX_BIT_KHR;
+
+ /* Average doesn't make sense for stencil so we don't support that */
+ p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
+ VK_RESOLVE_MODE_MIN_BIT_KHR |
+ VK_RESOLVE_MODE_MAX_BIT_KHR;
+
+ p->independentResolveNone = true;
+ p->independentResolve = true;
+
+ /* GFX6-8 only support single channel min/max filter. */
+ p->filterMinmaxImageComponentMapping = pdevice->rad_info.chip_class >= GFX9;
+ p->filterMinmaxSingleComponentFormats = true;
+
+ p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
+
+ p->framebufferIntegerColorSampleCounts = VK_SAMPLE_COUNT_1_BIT;
+}
+
void radv_GetPhysicalDeviceProperties2(
VkPhysicalDevice physicalDevice,
VkPhysicalDeviceProperties2 *pProperties)
RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
radv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
+ VkPhysicalDeviceVulkan11Properties core_1_1 = {
+ .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
+ };
+ radv_get_physical_device_properties_1_1(pdevice, &core_1_1);
+
+ VkPhysicalDeviceVulkan12Properties core_1_2 = {
+ .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
+ };
+ radv_get_physical_device_properties_1_2(pdevice, &core_1_2);
+
+#define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
+ memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
+ sizeof(core_##major##_##minor.core_property))
+
+#define CORE_PROPERTY(major, minor, property) \
+ CORE_RENAMED_PROPERTY(major, minor, property, property)
+
vk_foreach_struct(ext, pProperties->pNext) {
switch (ext->sType) {
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
VkPhysicalDeviceIDProperties *properties = (VkPhysicalDeviceIDProperties*)ext;
- memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
- memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
- properties->deviceLUIDValid = false;
+ CORE_PROPERTY(1, 1, deviceUUID);
+ CORE_PROPERTY(1, 1, driverUUID);
+ CORE_PROPERTY(1, 1, deviceLUID);
+ CORE_PROPERTY(1, 1, deviceLUIDValid);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
VkPhysicalDeviceMultiviewProperties *properties = (VkPhysicalDeviceMultiviewProperties*)ext;
- properties->maxMultiviewViewCount = MAX_VIEWS;
- properties->maxMultiviewInstanceIndex = INT_MAX;
+ CORE_PROPERTY(1, 1, maxMultiviewViewCount);
+ CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
VkPhysicalDevicePointClippingProperties *properties =
(VkPhysicalDevicePointClippingProperties*)ext;
- properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
+ CORE_PROPERTY(1, 1, pointClippingBehavior);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT: {
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
VkPhysicalDeviceSubgroupProperties *properties =
(VkPhysicalDeviceSubgroupProperties*)ext;
- properties->subgroupSize = 64;
- properties->supportedStages = VK_SHADER_STAGE_ALL;
- properties->supportedOperations =
- VK_SUBGROUP_FEATURE_BASIC_BIT |
- VK_SUBGROUP_FEATURE_BALLOT_BIT |
- VK_SUBGROUP_FEATURE_QUAD_BIT |
- VK_SUBGROUP_FEATURE_VOTE_BIT;
- if (pdevice->rad_info.chip_class >= GFX8) {
- properties->supportedOperations |=
- VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
- VK_SUBGROUP_FEATURE_CLUSTERED_BIT |
- VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
- VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT;
- }
- properties->quadOperationsInAllStages = true;
+ CORE_PROPERTY(1, 1, subgroupSize);
+ CORE_RENAMED_PROPERTY(1, 1, supportedStages,
+ subgroupSupportedStages);
+ CORE_RENAMED_PROPERTY(1, 1, supportedOperations,
+ subgroupSupportedOperations);
+ CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages,
+ subgroupQuadOperationsInAllStages);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
VkPhysicalDeviceMaintenance3Properties *properties =
(VkPhysicalDeviceMaintenance3Properties*)ext;
- /* Make sure everything is addressable by a signed 32-bit int, and
- * our largest descriptors are 96 bytes. */
- properties->maxPerSetDescriptors = (1ull << 31) / 96;
- /* Our buffer size fields allow only this much */
- properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
+ CORE_PROPERTY(1, 1, maxPerSetDescriptors);
+ CORE_PROPERTY(1, 1, maxMemoryAllocationSize);
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
- VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
- (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
- /* GFX6-8 only support single channel min/max filter. */
- properties->filterMinmaxImageComponentMapping = pdevice->rad_info.chip_class >= GFX9;
- properties->filterMinmaxSingleComponentFormats = true;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES: {
+ VkPhysicalDeviceSamplerFilterMinmaxProperties *properties =
+ (VkPhysicalDeviceSamplerFilterMinmaxProperties *)ext;
+ CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping);
+ CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD: {
properties->maxVertexAttribDivisor = UINT32_MAX;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
- VkPhysicalDeviceDescriptorIndexingPropertiesEXT *properties =
- (VkPhysicalDeviceDescriptorIndexingPropertiesEXT*)ext;
- properties->maxUpdateAfterBindDescriptorsInAllPools = UINT32_MAX / 64;
- properties->shaderUniformBufferArrayNonUniformIndexingNative = false;
- properties->shaderSampledImageArrayNonUniformIndexingNative = false;
- properties->shaderStorageBufferArrayNonUniformIndexingNative = false;
- properties->shaderStorageImageArrayNonUniformIndexingNative = false;
- properties->shaderInputAttachmentArrayNonUniformIndexingNative = false;
- properties->robustBufferAccessUpdateAfterBind = false;
- properties->quadDivergentImplicitLod = false;
-
- size_t max_descriptor_set_size = ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS -
- MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_INLINE_UNIFORM_BLOCK_COUNT) /
- (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
- 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
- 32 /* sampler, largest when combined with image */ +
- 64 /* sampled image */ +
- 64 /* storage image */);
- properties->maxPerStageDescriptorUpdateAfterBindSamplers = max_descriptor_set_size;
- properties->maxPerStageDescriptorUpdateAfterBindUniformBuffers = max_descriptor_set_size;
- properties->maxPerStageDescriptorUpdateAfterBindStorageBuffers = max_descriptor_set_size;
- properties->maxPerStageDescriptorUpdateAfterBindSampledImages = max_descriptor_set_size;
- properties->maxPerStageDescriptorUpdateAfterBindStorageImages = max_descriptor_set_size;
- properties->maxPerStageDescriptorUpdateAfterBindInputAttachments = max_descriptor_set_size;
- properties->maxPerStageUpdateAfterBindResources = max_descriptor_set_size;
- properties->maxDescriptorSetUpdateAfterBindSamplers = max_descriptor_set_size;
- properties->maxDescriptorSetUpdateAfterBindUniformBuffers = max_descriptor_set_size;
- properties->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS;
- properties->maxDescriptorSetUpdateAfterBindStorageBuffers = max_descriptor_set_size;
- properties->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS;
- properties->maxDescriptorSetUpdateAfterBindSampledImages = max_descriptor_set_size;
- properties->maxDescriptorSetUpdateAfterBindStorageImages = max_descriptor_set_size;
- properties->maxDescriptorSetUpdateAfterBindInputAttachments = max_descriptor_set_size;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES: {
+ VkPhysicalDeviceDescriptorIndexingProperties *properties =
+ (VkPhysicalDeviceDescriptorIndexingProperties*)ext;
+ CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools);
+ CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative);
+ CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative);
+ CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative);
+ CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative);
+ CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative);
+ CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind);
+ CORE_PROPERTY(1, 2, quadDivergentImplicitLod);
+ CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers);
+ CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers);
+ CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers);
+ CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages);
+ CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages);
+ CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments);
+ CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages);
+ CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
VkPhysicalDeviceProtectedMemoryProperties *properties =
(VkPhysicalDeviceProtectedMemoryProperties *)ext;
- properties->protectedNoFault = false;
+ CORE_PROPERTY(1, 1, protectedNoFault);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT: {
properties->primitiveOverestimationSize = 0;
properties->maxExtraPrimitiveOverestimationSize = 0;
properties->extraPrimitiveOverestimationSizeGranularity = 0;
- properties->primitiveUnderestimation = VK_FALSE;
- properties->conservativePointAndLineRasterization = VK_FALSE;
- properties->degenerateTrianglesRasterized = VK_FALSE;
- properties->degenerateLinesRasterized = VK_FALSE;
- properties->fullyCoveredFragmentShaderInputVariable = VK_FALSE;
- properties->conservativeRasterizationPostDepthCoverage = VK_FALSE;
+ properties->primitiveUnderestimation = false;
+ properties->conservativePointAndLineRasterization = false;
+ properties->degenerateTrianglesRasterized = false;
+ properties->degenerateLinesRasterized = false;
+ properties->fullyCoveredFragmentShaderInputVariable = false;
+ properties->conservativeRasterizationPostDepthCoverage = false;
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
properties->pciFunction = pdevice->bus_info.func;
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
- VkPhysicalDeviceDriverPropertiesKHR *driver_props =
- (VkPhysicalDeviceDriverPropertiesKHR *) ext;
-
- driver_props->driverID = VK_DRIVER_ID_MESA_RADV_KHR;
- snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR, "radv");
- snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
- "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
- " (LLVM " MESA_LLVM_VERSION_STRING ")");
-
- driver_props->conformanceVersion = (VkConformanceVersionKHR) {
- .major = 1,
- .minor = 1,
- .subminor = 2,
- .patch = 0,
- };
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES: {
+ VkPhysicalDeviceDriverProperties *properties =
+ (VkPhysicalDeviceDriverProperties *) ext;
+ CORE_PROPERTY(1, 2, driverID);
+ CORE_PROPERTY(1, 2, driverName);
+ CORE_PROPERTY(1, 2, driverInfo);
+ CORE_PROPERTY(1, 2, conformanceVersion);
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
properties->sampleLocationCoordinateRange[0] = 0.0f;
properties->sampleLocationCoordinateRange[1] = 0.9375f;
properties->sampleLocationSubPixelBits = 4;
- properties->variableSampleLocations = VK_FALSE;
+ properties->variableSampleLocations = false;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES: {
+ VkPhysicalDeviceDepthStencilResolveProperties *properties =
+ (VkPhysicalDeviceDepthStencilResolveProperties *)ext;
+ CORE_PROPERTY(1, 2, supportedDepthResolveModes);
+ CORE_PROPERTY(1, 2, supportedStencilResolveModes);
+ CORE_PROPERTY(1, 2, independentResolveNone);
+ CORE_PROPERTY(1, 2, independentResolve);
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
+ VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *properties =
+ (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
+ properties->storageTexelBufferOffsetAlignmentBytes = 4;
+ properties->storageTexelBufferOffsetSingleTexelAlignment = true;
+ properties->uniformTexelBufferOffsetAlignmentBytes = 4;
+ properties->uniformTexelBufferOffsetSingleTexelAlignment = true;
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES : {
+ VkPhysicalDeviceFloatControlsProperties *properties =
+ (VkPhysicalDeviceFloatControlsProperties *)ext;
+ CORE_PROPERTY(1, 2, denormBehaviorIndependence);
+ CORE_PROPERTY(1, 2, roundingModeIndependence);
+ CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16);
+ CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16);
+ CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16);
+ CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16);
+ CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16);
+ CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32);
+ CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32);
+ CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32);
+ CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32);
+ CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32);
+ CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64);
+ CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64);
+ CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64);
+ CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64);
+ CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64);
+ break;
+ }
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES: {
+ VkPhysicalDeviceTimelineSemaphoreProperties *properties =
+ (VkPhysicalDeviceTimelineSemaphoreProperties *) ext;
+ CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference);
break;
}
- case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
- VkPhysicalDeviceDepthStencilResolvePropertiesKHR *properties =
- (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
-
- /* We support all of the depth resolve modes */
- properties->supportedDepthResolveModes =
- VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
- VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
- VK_RESOLVE_MODE_MIN_BIT_KHR |
- VK_RESOLVE_MODE_MAX_BIT_KHR;
-
- /* Average doesn't make sense for stencil so we don't support that */
- properties->supportedStencilResolveModes =
- VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
- VK_RESOLVE_MODE_MIN_BIT_KHR |
- VK_RESOLVE_MODE_MAX_BIT_KHR;
-
- properties->independentResolveNone = VK_TRUE;
- properties->independentResolve = VK_TRUE;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
+ VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
+ (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
+ props->minSubgroupSize = 64;
+ props->maxSubgroupSize = 64;
+ props->maxComputeWorkgroupSubgroups = UINT32_MAX;
+ props->requiredSubgroupSizeStages = 0;
+
+ if (pdevice->rad_info.chip_class >= GFX10) {
+ /* Only GFX10+ supports wave32. */
+ props->minSubgroupSize = 32;
+ props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
+ }
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES:
+ radv_get_physical_device_properties_1_1(pdevice, (void *)ext);
+ break;
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES:
+ radv_get_physical_device_properties_1_2(pdevice, (void *)ext);
+ break;
default:
break;
}
{
int num_queue_families = 1;
int idx;
- if (pdevice->rad_info.num_compute_rings > 0 &&
+ if (pdevice->rad_info.num_rings[RING_COMPUTE] > 0 &&
!(pdevice->instance->debug_flags & RADV_DEBUG_NO_COMPUTE_QUEUE))
num_queue_families++;
idx++;
}
- if (pdevice->rad_info.num_compute_rings > 0 &&
+ if (pdevice->rad_info.num_rings[RING_COMPUTE] > 0 &&
!(pdevice->instance->debug_flags & RADV_DEBUG_NO_COMPUTE_QUEUE)) {
if (*pCount > idx) {
*pQueueFamilyProperties[idx] = (VkQueueFamilyProperties) {
.queueFlags = VK_QUEUE_COMPUTE_BIT |
VK_QUEUE_TRANSFER_BIT |
VK_QUEUE_SPARSE_BINDING_BIT,
- .queueCount = pdevice->rad_info.num_compute_rings,
+ .queueCount = pdevice->rad_info.num_rings[RING_COMPUTE],
.timestampValidBits = 64,
.minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
};
for (int i = 0; i < device->memory_properties.memoryTypeCount; i++) {
uint32_t heap_index = device->memory_properties.memoryTypes[i].heapIndex;
- switch (device->mem_type_indices[i]) {
- case RADV_MEM_TYPE_VRAM:
+ if (radv_is_mem_type_vram(device->mem_type_indices[i])) {
heap_usage = device->ws->query_value(device->ws,
RADEON_ALLOCATED_VRAM);
memoryBudget->heapBudget[heap_index] = heap_budget;
memoryBudget->heapUsage[heap_index] = heap_usage;
- break;
- case RADV_MEM_TYPE_VRAM_CPU_ACCESS:
+ } else if (radv_is_mem_type_vram_visible(device->mem_type_indices[i])) {
heap_usage = device->ws->query_value(device->ws,
RADEON_ALLOCATED_VRAM_VIS);
memoryBudget->heapBudget[heap_index] = heap_budget;
memoryBudget->heapUsage[heap_index] = heap_usage;
- break;
- case RADV_MEM_TYPE_GTT_WRITE_COMBINE:
+ } else if (radv_is_mem_type_gtt_wc(device->mem_type_indices[i])) {
heap_usage = device->ws->query_value(device->ws,
RADEON_ALLOCATED_GTT);
memoryBudget->heapBudget[heap_index] = heap_budget;
memoryBudget->heapUsage[heap_index] = heap_usage;
- break;
- default:
- break;
}
}
const struct radv_physical_device *physical_device = device->physical_device;
uint32_t memoryTypeBits = 0;
for (int i = 0; i < physical_device->memory_properties.memoryTypeCount; i++) {
- if (physical_device->mem_type_indices[i] == RADV_MEM_TYPE_GTT_CACHED) {
+ if (radv_is_mem_type_gtt_cached(physical_device->mem_type_indices[i])) {
memoryTypeBits = (1 << i);
break;
}
if (!queue->hw_ctx)
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+ list_inithead(&queue->pending_submissions);
+ pthread_mutex_init(&queue->pending_mutex, NULL);
+
return VK_SUCCESS;
}
static void
radv_queue_finish(struct radv_queue *queue)
{
+ pthread_mutex_destroy(&queue->pending_mutex);
+
if (queue->hw_ctx)
queue->device->ws->ctx_destroy(queue->hw_ctx);
return result;
}
-VkResult radv_CreateDevice(
- VkPhysicalDevice physicalDevice,
- const VkDeviceCreateInfo* pCreateInfo,
- const VkAllocationCallbacks* pAllocator,
- VkDevice* pDevice)
-{
- RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
- VkResult result;
- struct radv_device *device;
+static int install_seccomp_filter() {
- bool keep_shader_info = false;
+ struct sock_filter filter[] = {
+ /* Check arch is 64bit x86 */
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, arch))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, AUDIT_ARCH_X86_64, 0, 12),
- /* Check enabled features */
- if (pCreateInfo->pEnabledFeatures) {
- VkPhysicalDeviceFeatures supported_features;
- radv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
- VkBool32 *supported_feature = (VkBool32 *)&supported_features;
- VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
- unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
- for (uint32_t i = 0; i < num_features; i++) {
- if (enabled_feature[i] && !supported_feature[i])
- return vk_error(physical_device->instance, VK_ERROR_FEATURE_NOT_PRESENT);
- }
- }
+ /* Futex is required for mutex locks */
+ #if defined __NR__newselect
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR__newselect, 11, 0),
+ #elif defined __NR_select
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_select, 11, 0),
+ #else
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_pselect6, 11, 0),
+ #endif
- device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
- sizeof(*device), 8,
- VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
- if (!device)
- return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+ /* Allow system exit calls for the forked process */
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_exit_group, 9, 0),
- device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
- device->instance = physical_device->instance;
- device->physical_device = physical_device;
+ /* Allow system read calls */
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_read, 7, 0),
- device->ws = physical_device->ws;
- if (pAllocator)
- device->alloc = *pAllocator;
- else
- device->alloc = physical_device->instance->alloc;
+ /* Allow system write calls */
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_write, 5, 0),
- for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
- const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
- int index = radv_get_device_extension_index(ext_name);
- if (index < 0 || !physical_device->supported_extensions.extensions[index]) {
- vk_free(&device->alloc, device);
- return vk_error(physical_device->instance, VK_ERROR_EXTENSION_NOT_PRESENT);
- }
+ /* Allow system brk calls (we need this for malloc) */
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_brk, 3, 0),
- device->enabled_extensions.extensions[index] = true;
- }
+ /* Futex is required for mutex locks */
+ BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_futex, 1, 0),
- keep_shader_info = device->enabled_extensions.AMD_shader_info;
+ /* Return error if we hit a system call not on the whitelist */
+ BPF_STMT(BPF_RET + BPF_K, SECCOMP_RET_ERRNO | (EPERM & SECCOMP_RET_DATA)),
- /* With update after bind we can't attach bo's to the command buffer
- * from the descriptor set anymore, so we have to use a global BO list.
- */
- device->use_global_bo_list =
- (device->instance->perftest_flags & RADV_PERFTEST_BO_LIST) ||
- device->enabled_extensions.EXT_descriptor_indexing ||
- device->enabled_extensions.EXT_buffer_device_address;
+ /* Allow whitelisted system calls */
+ BPF_STMT(BPF_RET + BPF_K, SECCOMP_RET_ALLOW),
+ };
- device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
- pCreateInfo->pEnabledFeatures->robustBufferAccess;
+ struct sock_fprog prog = {
+ .len = (unsigned short)(sizeof(filter) / sizeof(filter[0])),
+ .filter = filter,
+ };
- mtx_init(&device->shader_slab_mutex, mtx_plain);
- list_inithead(&device->shader_slabs);
+ if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0))
+ return -1;
- radv_bo_list_init(&device->bo_list);
+ if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &prog))
+ return -1;
- for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
- const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
- uint32_t qfi = queue_create->queueFamilyIndex;
- const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority =
- vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
+ return 0;
+}
- assert(!global_priority || device->physical_device->rad_info.has_ctx_priority);
+/* Helper function with timeout support for reading from the pipe between
+ * processes used for secure compile.
+ */
+bool radv_sc_read(int fd, void *buf, size_t size, bool timeout)
+{
+ fd_set fds;
+ struct timeval tv;
- device->queues[qfi] = vk_alloc(&device->alloc,
- queue_create->queueCount * sizeof(struct radv_queue), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
- if (!device->queues[qfi]) {
- result = VK_ERROR_OUT_OF_HOST_MEMORY;
- goto fail;
- }
+ FD_ZERO(&fds);
+ FD_SET(fd, &fds);
- memset(device->queues[qfi], 0, queue_create->queueCount * sizeof(struct radv_queue));
+ while (true) {
+ /* We can't rely on the value of tv after calling select() so
+ * we must reset it on each iteration of the loop.
+ */
+ tv.tv_sec = 5;
+ tv.tv_usec = 0;
- device->queue_count[qfi] = queue_create->queueCount;
+ int rval = select(fd + 1, &fds, NULL, NULL, timeout ? &tv : NULL);
- for (unsigned q = 0; q < queue_create->queueCount; q++) {
- result = radv_queue_init(device, &device->queues[qfi][q],
- qfi, q, queue_create->flags,
- global_priority);
- if (result != VK_SUCCESS)
- goto fail;
+ if (rval == -1) {
+ /* select error */
+ return false;
+ } else if (rval) {
+ ssize_t bytes_read = read(fd, buf, size);
+ if (bytes_read < 0)
+ return false;
+
+ buf += bytes_read;
+ size -= bytes_read;
+ if (size == 0)
+ return true;
+ } else {
+ /* select timeout */
+ return false;
}
}
+}
- device->pbb_allowed = device->physical_device->rad_info.chip_class >= GFX9 &&
- !(device->instance->debug_flags & RADV_DEBUG_NOBINNING);
+static bool radv_close_all_fds(const int *keep_fds, int keep_fd_count)
+{
+ DIR *d;
+ struct dirent *dir;
+ d = opendir("/proc/self/fd");
+ if (!d)
+ return false;
+ int dir_fd = dirfd(d);
- /* Disable DFSM by default. As of 2019-09-15 Talos on Low is still 3% slower on Raven. */
- device->dfsm_allowed = device->pbb_allowed &&
- (device->instance->perftest_flags & RADV_PERFTEST_DFSM);
+ while ((dir = readdir(d)) != NULL) {
+ if (dir->d_name[0] == '.')
+ continue;
-#ifdef ANDROID
- device->always_use_syncobj = device->physical_device->rad_info.has_syncobj_wait_for_submit;
-#endif
+ int fd = atoi(dir->d_name);
+ if (fd == dir_fd)
+ continue;
- /* The maximum number of scratch waves. Scratch space isn't divided
- * evenly between CUs. The number is only a function of the number of CUs.
- * We can decrease the constant to decrease the scratch buffer size.
- *
- * sctx->scratch_waves must be >= the maximum possible size of
- * 1 threadgroup, so that the hw doesn't hang from being unable
+ bool keep = false;
+ for (int i = 0; !keep && i < keep_fd_count; ++i)
+ if (keep_fds[i] == fd)
+ keep = true;
+
+ if (keep)
+ continue;
+
+ close(fd);
+ }
+ closedir(d);
+ return true;
+}
+
+static bool secure_compile_open_fifo_fds(struct radv_secure_compile_state *sc,
+ int *fd_server, int *fd_client,
+ unsigned process, bool make_fifo)
+{
+ bool result = false;
+ char *fifo_server_path = NULL;
+ char *fifo_client_path = NULL;
+
+ if (asprintf(&fifo_server_path, "/tmp/radv_server_%s_%u", sc->uid, process) == -1)
+ goto open_fifo_exit;
+
+ if (asprintf(&fifo_client_path, "/tmp/radv_client_%s_%u", sc->uid, process) == -1)
+ goto open_fifo_exit;
+
+ if (make_fifo) {
+ int file1 = mkfifo(fifo_server_path, 0666);
+ if(file1 < 0)
+ goto open_fifo_exit;
+
+ int file2 = mkfifo(fifo_client_path, 0666);
+ if(file2 < 0)
+ goto open_fifo_exit;
+ }
+
+ *fd_server = open(fifo_server_path, O_RDWR);
+ if(*fd_server < 1)
+ goto open_fifo_exit;
+
+ *fd_client = open(fifo_client_path, O_RDWR);
+ if(*fd_client < 1) {
+ close(*fd_server);
+ goto open_fifo_exit;
+ }
+
+ result = true;
+
+open_fifo_exit:
+ free(fifo_server_path);
+ free(fifo_client_path);
+
+ return result;
+}
+
+static void run_secure_compile_device(struct radv_device *device, unsigned process,
+ int fd_idle_device_output)
+{
+ int fd_secure_input;
+ int fd_secure_output;
+ bool fifo_result = secure_compile_open_fifo_fds(device->sc_state,
+ &fd_secure_input,
+ &fd_secure_output,
+ process, false);
+
+ enum radv_secure_compile_type sc_type;
+
+ const int needed_fds[] = {
+ fd_secure_input,
+ fd_secure_output,
+ fd_idle_device_output,
+ };
+
+ if (!fifo_result || !radv_close_all_fds(needed_fds, ARRAY_SIZE(needed_fds)) ||
+ install_seccomp_filter() == -1) {
+ sc_type = RADV_SC_TYPE_INIT_FAILURE;
+ } else {
+ sc_type = RADV_SC_TYPE_INIT_SUCCESS;
+ device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input;
+ device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output;
+ }
+
+ write(fd_idle_device_output, &sc_type, sizeof(sc_type));
+
+ if (sc_type == RADV_SC_TYPE_INIT_FAILURE)
+ goto secure_compile_exit;
+
+ while (true) {
+ radv_sc_read(fd_secure_input, &sc_type, sizeof(sc_type), false);
+
+ if (sc_type == RADV_SC_TYPE_COMPILE_PIPELINE) {
+ struct radv_pipeline *pipeline;
+ bool sc_read = true;
+
+ pipeline = vk_zalloc2(&device->alloc, NULL, sizeof(*pipeline), 8,
+ VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+ pipeline->device = device;
+
+ /* Read pipeline layout */
+ struct radv_pipeline_layout layout;
+ sc_read = radv_sc_read(fd_secure_input, &layout, sizeof(struct radv_pipeline_layout), true);
+ sc_read &= radv_sc_read(fd_secure_input, &layout.num_sets, sizeof(uint32_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ for (uint32_t set = 0; set < layout.num_sets; set++) {
+ uint32_t layout_size;
+ sc_read &= radv_sc_read(fd_secure_input, &layout_size, sizeof(uint32_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ layout.set[set].layout = malloc(layout_size);
+ layout.set[set].layout->layout_size = layout_size;
+ sc_read &= radv_sc_read(fd_secure_input, layout.set[set].layout,
+ layout.set[set].layout->layout_size, true);
+ }
+
+ pipeline->layout = &layout;
+
+ /* Read pipeline key */
+ struct radv_pipeline_key key;
+ sc_read &= radv_sc_read(fd_secure_input, &key, sizeof(struct radv_pipeline_key), true);
+
+ /* Read pipeline create flags */
+ VkPipelineCreateFlags flags;
+ sc_read &= radv_sc_read(fd_secure_input, &flags, sizeof(VkPipelineCreateFlags), true);
+
+ /* Read stage and shader information */
+ uint32_t num_stages;
+ const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
+ sc_read &= radv_sc_read(fd_secure_input, &num_stages, sizeof(uint32_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ for (uint32_t i = 0; i < num_stages; i++) {
+
+ /* Read stage */
+ gl_shader_stage stage;
+ sc_read &= radv_sc_read(fd_secure_input, &stage, sizeof(gl_shader_stage), true);
+
+ VkPipelineShaderStageCreateInfo *pStage = calloc(1, sizeof(VkPipelineShaderStageCreateInfo));
+
+ /* Read entry point name */
+ size_t name_size;
+ sc_read &= radv_sc_read(fd_secure_input, &name_size, sizeof(size_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ char *ep_name = malloc(name_size);
+ sc_read &= radv_sc_read(fd_secure_input, ep_name, name_size, true);
+ pStage->pName = ep_name;
+
+ /* Read shader module */
+ size_t module_size;
+ sc_read &= radv_sc_read(fd_secure_input, &module_size, sizeof(size_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ struct radv_shader_module *module = malloc(module_size);
+ sc_read &= radv_sc_read(fd_secure_input, module, module_size, true);
+ pStage->module = radv_shader_module_to_handle(module);
+
+ /* Read specialization info */
+ bool has_spec_info;
+ sc_read &= radv_sc_read(fd_secure_input, &has_spec_info, sizeof(bool), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ if (has_spec_info) {
+ VkSpecializationInfo *specInfo = malloc(sizeof(VkSpecializationInfo));
+ pStage->pSpecializationInfo = specInfo;
+
+ sc_read &= radv_sc_read(fd_secure_input, &specInfo->dataSize, sizeof(size_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ void *si_data = malloc(specInfo->dataSize);
+ sc_read &= radv_sc_read(fd_secure_input, si_data, specInfo->dataSize, true);
+ specInfo->pData = si_data;
+
+ sc_read &= radv_sc_read(fd_secure_input, &specInfo->mapEntryCount, sizeof(uint32_t), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+
+ VkSpecializationMapEntry *mapEntries = malloc(sizeof(VkSpecializationMapEntry) * specInfo->mapEntryCount);
+ for (uint32_t j = 0; j < specInfo->mapEntryCount; j++) {
+ sc_read &= radv_sc_read(fd_secure_input, &mapEntries[j], sizeof(VkSpecializationMapEntry), true);
+ if (!sc_read)
+ goto secure_compile_exit;
+ }
+
+ specInfo->pMapEntries = mapEntries;
+ }
+
+ pStages[stage] = pStage;
+ }
+
+ /* Compile the shaders */
+ VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
+ radv_create_shaders(pipeline, device, NULL, &key, pStages, flags, NULL, stage_feedbacks);
+
+ /* free memory allocated above */
+ for (uint32_t set = 0; set < layout.num_sets; set++)
+ free(layout.set[set].layout);
+
+ for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
+ if (!pStages[i])
+ continue;
+
+ free((void *) pStages[i]->pName);
+ free(radv_shader_module_from_handle(pStages[i]->module));
+ if (pStages[i]->pSpecializationInfo) {
+ free((void *) pStages[i]->pSpecializationInfo->pData);
+ free((void *) pStages[i]->pSpecializationInfo->pMapEntries);
+ free((void *) pStages[i]->pSpecializationInfo);
+ }
+ free((void *) pStages[i]);
+ }
+
+ vk_free(&device->alloc, pipeline);
+
+ sc_type = RADV_SC_TYPE_COMPILE_PIPELINE_FINISHED;
+ write(fd_secure_output, &sc_type, sizeof(sc_type));
+
+ } else if (sc_type == RADV_SC_TYPE_DESTROY_DEVICE) {
+ goto secure_compile_exit;
+ }
+ }
+
+secure_compile_exit:
+ close(fd_secure_input);
+ close(fd_secure_output);
+ close(fd_idle_device_output);
+ _exit(0);
+}
+
+static enum radv_secure_compile_type fork_secure_compile_device(struct radv_device *device, unsigned process)
+{
+ int fd_secure_input[2];
+ int fd_secure_output[2];
+
+ /* create pipe descriptors (used to communicate between processes) */
+ if (pipe(fd_secure_input) == -1 || pipe(fd_secure_output) == -1)
+ return RADV_SC_TYPE_INIT_FAILURE;
+
+
+ int sc_pid;
+ if ((sc_pid = fork()) == 0) {
+ device->sc_state->secure_compile_thread_counter = process;
+ run_secure_compile_device(device, process, fd_secure_output[1]);
+ } else {
+ if (sc_pid == -1)
+ return RADV_SC_TYPE_INIT_FAILURE;
+
+ /* Read the init result returned from the secure process */
+ enum radv_secure_compile_type sc_type;
+ bool sc_read = radv_sc_read(fd_secure_output[0], &sc_type, sizeof(sc_type), true);
+
+ if (sc_type == RADV_SC_TYPE_INIT_FAILURE || !sc_read) {
+ close(fd_secure_input[0]);
+ close(fd_secure_input[1]);
+ close(fd_secure_output[1]);
+ close(fd_secure_output[0]);
+ int status;
+ waitpid(sc_pid, &status, 0);
+
+ return RADV_SC_TYPE_INIT_FAILURE;
+ } else {
+ assert(sc_type == RADV_SC_TYPE_INIT_SUCCESS);
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output, &sc_type, sizeof(sc_type));
+
+ close(fd_secure_input[0]);
+ close(fd_secure_input[1]);
+ close(fd_secure_output[1]);
+ close(fd_secure_output[0]);
+
+ int status;
+ waitpid(sc_pid, &status, 0);
+ }
+ }
+
+ return RADV_SC_TYPE_INIT_SUCCESS;
+}
+
+/* Run a bare bones fork of a device that was forked right after its creation.
+ * This device will have low overhead when it is forked again before each
+ * pipeline compilation. This device sits idle and its only job is to fork
+ * itself.
+ */
+static void run_secure_compile_idle_device(struct radv_device *device, unsigned process,
+ int fd_secure_input, int fd_secure_output)
+{
+ enum radv_secure_compile_type sc_type = RADV_SC_TYPE_INIT_SUCCESS;
+ device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input;
+ device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output;
+
+ write(fd_secure_output, &sc_type, sizeof(sc_type));
+
+ while (true) {
+ radv_sc_read(fd_secure_input, &sc_type, sizeof(sc_type), false);
+
+ if (sc_type == RADV_SC_TYPE_FORK_DEVICE) {
+ sc_type = fork_secure_compile_device(device, process);
+
+ if (sc_type == RADV_SC_TYPE_INIT_FAILURE)
+ goto secure_compile_exit;
+
+ } else if (sc_type == RADV_SC_TYPE_DESTROY_DEVICE) {
+ goto secure_compile_exit;
+ }
+ }
+
+secure_compile_exit:
+ close(fd_secure_input);
+ close(fd_secure_output);
+ _exit(0);
+}
+
+static void destroy_secure_compile_device(struct radv_device *device, unsigned process)
+{
+ int fd_secure_input = device->sc_state->secure_compile_processes[process].fd_secure_input;
+
+ enum radv_secure_compile_type sc_type = RADV_SC_TYPE_DESTROY_DEVICE;
+ write(fd_secure_input, &sc_type, sizeof(sc_type));
+
+ close(device->sc_state->secure_compile_processes[process].fd_secure_input);
+ close(device->sc_state->secure_compile_processes[process].fd_secure_output);
+
+ int status;
+ waitpid(device->sc_state->secure_compile_processes[process].sc_pid, &status, 0);
+}
+
+static VkResult fork_secure_compile_idle_device(struct radv_device *device)
+{
+ device->sc_state = vk_zalloc(&device->alloc,
+ sizeof(struct radv_secure_compile_state),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+
+ mtx_init(&device->sc_state->secure_compile_mutex, mtx_plain);
+
+ pid_t upid = getpid();
+ time_t seconds = time(NULL);
+
+ char *uid;
+ if (asprintf(&uid, "%ld_%ld", (long) upid, (long) seconds) == -1)
+ return VK_ERROR_INITIALIZATION_FAILED;
+
+ device->sc_state->uid = uid;
+
+ uint8_t sc_threads = device->instance->num_sc_threads;
+ int fd_secure_input[MAX_SC_PROCS][2];
+ int fd_secure_output[MAX_SC_PROCS][2];
+
+ /* create pipe descriptors (used to communicate between processes) */
+ for (unsigned i = 0; i < sc_threads; i++) {
+ if (pipe(fd_secure_input[i]) == -1 ||
+ pipe(fd_secure_output[i]) == -1) {
+ return VK_ERROR_INITIALIZATION_FAILED;
+ }
+ }
+
+ device->sc_state->secure_compile_processes = vk_zalloc(&device->alloc,
+ sizeof(struct radv_secure_compile_process) * sc_threads, 8,
+ VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+
+ for (unsigned process = 0; process < sc_threads; process++) {
+ if ((device->sc_state->secure_compile_processes[process].sc_pid = fork()) == 0) {
+ device->sc_state->secure_compile_thread_counter = process;
+ run_secure_compile_idle_device(device, process, fd_secure_input[process][0], fd_secure_output[process][1]);
+ } else {
+ if (device->sc_state->secure_compile_processes[process].sc_pid == -1)
+ return VK_ERROR_INITIALIZATION_FAILED;
+
+ /* Read the init result returned from the secure process */
+ enum radv_secure_compile_type sc_type;
+ bool sc_read = radv_sc_read(fd_secure_output[process][0], &sc_type, sizeof(sc_type), true);
+
+ bool fifo_result;
+ if (sc_read && sc_type == RADV_SC_TYPE_INIT_SUCCESS) {
+ fifo_result = secure_compile_open_fifo_fds(device->sc_state,
+ &device->sc_state->secure_compile_processes[process].fd_server,
+ &device->sc_state->secure_compile_processes[process].fd_client,
+ process, true);
+
+ device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input[process][1];
+ device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output[process][0];
+ }
+
+ if (sc_type == RADV_SC_TYPE_INIT_FAILURE || !sc_read || !fifo_result) {
+ close(fd_secure_input[process][0]);
+ close(fd_secure_input[process][1]);
+ close(fd_secure_output[process][1]);
+ close(fd_secure_output[process][0]);
+ int status;
+ waitpid(device->sc_state->secure_compile_processes[process].sc_pid, &status, 0);
+
+ /* Destroy any forks that were created sucessfully */
+ for (unsigned i = 0; i < process; i++) {
+ destroy_secure_compile_device(device, i);
+ }
+
+ return VK_ERROR_INITIALIZATION_FAILED;
+ }
+ }
+ }
+ return VK_SUCCESS;
+}
+
+static VkResult
+radv_create_pthread_cond(pthread_cond_t *cond)
+{
+ pthread_condattr_t condattr;
+ if (pthread_condattr_init(&condattr)) {
+ return VK_ERROR_INITIALIZATION_FAILED;
+ }
+
+ if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC)) {
+ pthread_condattr_destroy(&condattr);
+ return VK_ERROR_INITIALIZATION_FAILED;
+ }
+ if (pthread_cond_init(cond, &condattr)) {
+ pthread_condattr_destroy(&condattr);
+ return VK_ERROR_INITIALIZATION_FAILED;
+ }
+ pthread_condattr_destroy(&condattr);
+ return VK_SUCCESS;
+}
+
+VkResult radv_CreateDevice(
+ VkPhysicalDevice physicalDevice,
+ const VkDeviceCreateInfo* pCreateInfo,
+ const VkAllocationCallbacks* pAllocator,
+ VkDevice* pDevice)
+{
+ RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
+ VkResult result;
+ struct radv_device *device;
+
+ bool keep_shader_info = false;
+
+ /* Check enabled features */
+ if (pCreateInfo->pEnabledFeatures) {
+ VkPhysicalDeviceFeatures supported_features;
+ radv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
+ VkBool32 *supported_feature = (VkBool32 *)&supported_features;
+ VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
+ unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
+ for (uint32_t i = 0; i < num_features; i++) {
+ if (enabled_feature[i] && !supported_feature[i])
+ return vk_error(physical_device->instance, VK_ERROR_FEATURE_NOT_PRESENT);
+ }
+ }
+
+ device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
+ sizeof(*device), 8,
+ VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+ if (!device)
+ return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+
+ device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+ device->instance = physical_device->instance;
+ device->physical_device = physical_device;
+
+ device->ws = physical_device->ws;
+ if (pAllocator)
+ device->alloc = *pAllocator;
+ else
+ device->alloc = physical_device->instance->alloc;
+
+ for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
+ const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
+ int index = radv_get_device_extension_index(ext_name);
+ if (index < 0 || !physical_device->supported_extensions.extensions[index]) {
+ vk_free(&device->alloc, device);
+ return vk_error(physical_device->instance, VK_ERROR_EXTENSION_NOT_PRESENT);
+ }
+
+ device->enabled_extensions.extensions[index] = true;
+ }
+
+ keep_shader_info = device->enabled_extensions.AMD_shader_info;
+
+ /* With update after bind we can't attach bo's to the command buffer
+ * from the descriptor set anymore, so we have to use a global BO list.
+ */
+ device->use_global_bo_list =
+ (device->instance->perftest_flags & RADV_PERFTEST_BO_LIST) ||
+ device->enabled_extensions.EXT_descriptor_indexing ||
+ device->enabled_extensions.EXT_buffer_device_address ||
+ device->enabled_extensions.KHR_buffer_device_address;
+
+ device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
+ pCreateInfo->pEnabledFeatures->robustBufferAccess;
+
+ mtx_init(&device->shader_slab_mutex, mtx_plain);
+ list_inithead(&device->shader_slabs);
+
+ radv_bo_list_init(&device->bo_list);
+
+ for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
+ const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
+ uint32_t qfi = queue_create->queueFamilyIndex;
+ const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority =
+ vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
+
+ assert(!global_priority || device->physical_device->rad_info.has_ctx_priority);
+
+ device->queues[qfi] = vk_alloc(&device->alloc,
+ queue_create->queueCount * sizeof(struct radv_queue), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+ if (!device->queues[qfi]) {
+ result = VK_ERROR_OUT_OF_HOST_MEMORY;
+ goto fail;
+ }
+
+ memset(device->queues[qfi], 0, queue_create->queueCount * sizeof(struct radv_queue));
+
+ device->queue_count[qfi] = queue_create->queueCount;
+
+ for (unsigned q = 0; q < queue_create->queueCount; q++) {
+ result = radv_queue_init(device, &device->queues[qfi][q],
+ qfi, q, queue_create->flags,
+ global_priority);
+ if (result != VK_SUCCESS)
+ goto fail;
+ }
+ }
+
+ device->pbb_allowed = device->physical_device->rad_info.chip_class >= GFX9 &&
+ !(device->instance->debug_flags & RADV_DEBUG_NOBINNING);
+
+ /* Disable DFSM by default. As of 2019-09-15 Talos on Low is still 3% slower on Raven. */
+ device->dfsm_allowed = device->pbb_allowed &&
+ (device->instance->perftest_flags & RADV_PERFTEST_DFSM);
+
+ device->always_use_syncobj = device->physical_device->rad_info.has_syncobj_wait_for_submit;
+
+ /* The maximum number of scratch waves. Scratch space isn't divided
+ * evenly between CUs. The number is only a function of the number of CUs.
+ * We can decrease the constant to decrease the scratch buffer size.
+ *
+ * sctx->scratch_waves must be >= the maximum possible size of
+ * 1 threadgroup, so that the hw doesn't hang from being unable
* to start any.
*
* The recommended value is 4 per CU at most. Higher numbers don't
device->scratch_waves = MAX2(32 * physical_device->rad_info.num_good_compute_units,
max_threads_per_block / 64);
- device->dispatch_initiator = S_00B800_COMPUTE_SHADER_EN(1) |
- S_00B800_CS_W32_EN(device->physical_device->cs_wave_size == 32);
+ device->dispatch_initiator = S_00B800_COMPUTE_SHADER_EN(1);
if (device->physical_device->rad_info.chip_class >= GFX7) {
/* If the KMD allows it (there is a KMD hw register for it),
radv_dump_enabled_options(device, stderr);
}
- device->keep_shader_info = keep_shader_info;
+ /* Temporarily disable secure compile while we create meta shaders, etc */
+ uint8_t sc_threads = device->instance->num_sc_threads;
+ if (sc_threads)
+ device->instance->num_sc_threads = 0;
+ device->keep_shader_info = keep_shader_info;
result = radv_device_init_meta(device);
if (result != VK_SUCCESS)
goto fail;
device->empty_cs[family] = device->ws->cs_create(device->ws, family);
switch (family) {
case RADV_QUEUE_GENERAL:
- /* Since amdgpu version 3.6.0, CONTEXT_CONTROL is emitted by the kernel */
- if (device->physical_device->rad_info.drm_minor < 6) {
- radeon_emit(device->empty_cs[family], PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
- radeon_emit(device->empty_cs[family], CONTEXT_CONTROL_LOAD_ENABLE(1));
- radeon_emit(device->empty_cs[family], CONTEXT_CONTROL_SHADOW_ENABLE(1));
- }
+ radeon_emit(device->empty_cs[family], PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
+ radeon_emit(device->empty_cs[family], CONTEXT_CONTROL_LOAD_ENABLE(1));
+ radeon_emit(device->empty_cs[family], CONTEXT_CONTROL_SHADOW_ENABLE(1));
break;
case RADV_QUEUE_COMPUTE:
radeon_emit(device->empty_cs[family], PKT3(PKT3_NOP, 0, 0));
device->mem_cache = radv_pipeline_cache_from_handle(pc);
+ result = radv_create_pthread_cond(&device->timeline_cond);
+ if (result != VK_SUCCESS)
+ goto fail_mem_cache;
+
device->force_aniso =
MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
if (device->force_aniso >= 0) {
1 << util_logbase2(device->force_aniso));
}
+ /* Fork device for secure compile as required */
+ device->instance->num_sc_threads = sc_threads;
+ if (radv_device_use_secure_compile(device->instance)) {
+
+ result = fork_secure_compile_idle_device(device);
+ if (result != VK_SUCCESS)
+ goto fail_meta;
+ }
+
*pDevice = radv_device_to_handle(device);
return VK_SUCCESS;
+fail_mem_cache:
+ radv_DestroyPipelineCache(radv_device_to_handle(device), pc, NULL);
fail_meta:
radv_device_finish_meta(device);
fail:
radv_destroy_shader_slabs(device);
+ pthread_cond_destroy(&device->timeline_cond);
radv_bo_list_finish(&device->bo_list);
+ if (radv_device_use_secure_compile(device->instance)) {
+ for (unsigned i = 0; i < device->instance->num_sc_threads; i++ ) {
+ destroy_secure_compile_device(device, i);
+ }
+ }
+
+ if (device->sc_state) {
+ free(device->sc_state->uid);
+ vk_free(&device->alloc, device->sc_state->secure_compile_processes);
+ }
+ vk_free(&device->alloc, device->sc_state);
vk_free(&device->alloc, device);
}
if (queue->device->physical_device->rad_info.chip_class >= GFX10) {
desc[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
- S_008F0C_OOB_SELECT(2) |
+ S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
S_008F0C_RESOURCE_LEVEL(1);
} else {
desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
if (queue->device->physical_device->rad_info.chip_class >= GFX10) {
desc[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
- S_008F0C_OOB_SELECT(2) |
+ S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
S_008F0C_RESOURCE_LEVEL(1);
} else {
desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
if (queue->device->physical_device->rad_info.chip_class >= GFX10) {
desc[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
- S_008F0C_OOB_SELECT(2) |
+ S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
S_008F0C_RESOURCE_LEVEL(1);
} else {
desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
if (queue->device->physical_device->rad_info.chip_class >= GFX10) {
desc[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
- S_008F0C_OOB_SELECT(2) |
+ S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
S_008F0C_RESOURCE_LEVEL(1);
} else {
desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
if (queue->device->physical_device->rad_info.chip_class >= GFX10) {
desc[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
- S_008F0C_OOB_SELECT(3) |
+ S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
S_008F0C_RESOURCE_LEVEL(1);
} else {
desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
if (queue->device->physical_device->rad_info.chip_class >= GFX10) {
desc[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
- S_008F0C_OOB_SELECT(3) |
+ S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
S_008F0C_RESOURCE_LEVEL(1);
} else {
desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
}
}
+static void
+radv_emit_graphics_scratch(struct radv_queue *queue, struct radeon_cmdbuf *cs,
+ uint32_t size_per_wave, uint32_t waves,
+ struct radeon_winsys_bo *scratch_bo)
+{
+ if (queue->queue_family_index != RADV_QUEUE_GENERAL)
+ return;
+
+ if (!scratch_bo)
+ return;
+
+ radv_cs_add_buffer(queue->device->ws, cs, scratch_bo);
+
+ radeon_set_context_reg(cs, R_0286E8_SPI_TMPRING_SIZE,
+ S_0286E8_WAVES(waves) |
+ S_0286E8_WAVESIZE(round_up_u32(size_per_wave, 1024)));
+}
+
static void
radv_emit_compute_scratch(struct radv_queue *queue, struct radeon_cmdbuf *cs,
- struct radeon_winsys_bo *compute_scratch_bo)
+ uint32_t size_per_wave, uint32_t waves,
+ struct radeon_winsys_bo *compute_scratch_bo)
{
uint64_t scratch_va;
radeon_emit(cs, scratch_va);
radeon_emit(cs, S_008F04_BASE_ADDRESS_HI(scratch_va >> 32) |
S_008F04_SWIZZLE_ENABLE(1));
+
+ radeon_set_sh_reg(cs, R_00B860_COMPUTE_TMPRING_SIZE,
+ S_00B860_WAVES(waves) |
+ S_00B860_WAVESIZE(round_up_u32(size_per_wave, 1024)));
}
static void
static VkResult
radv_get_preamble_cs(struct radv_queue *queue,
- uint32_t scratch_size,
- uint32_t compute_scratch_size,
+ uint32_t scratch_size_per_wave,
+ uint32_t scratch_waves,
+ uint32_t compute_scratch_size_per_wave,
+ uint32_t compute_scratch_waves,
uint32_t esgs_ring_size,
uint32_t gsvs_ring_size,
bool needs_tess_rings,
bool needs_gds,
+ bool needs_gds_oa,
bool needs_sample_positions,
struct radeon_cmdbuf **initial_full_flush_preamble_cs,
struct radeon_cmdbuf **initial_preamble_cs,
struct radeon_winsys_bo *gds_bo = NULL;
struct radeon_winsys_bo *gds_oa_bo = NULL;
struct radeon_cmdbuf *dest_cs[3] = {0};
- bool add_tess_rings = false, add_gds = false, add_sample_positions = false;
+ bool add_tess_rings = false, add_gds = false, add_gds_oa = false, add_sample_positions = false;
unsigned tess_factor_ring_size = 0, tess_offchip_ring_size = 0;
unsigned max_offchip_buffers;
unsigned hs_offchip_param = 0;
if (needs_gds)
add_gds = true;
}
+ if (!queue->has_gds_oa) {
+ if (needs_gds_oa)
+ add_gds_oa = true;
+ }
if (!queue->has_sample_positions) {
if (needs_sample_positions)
add_sample_positions = true;
tess_offchip_ring_size = max_offchip_buffers *
queue->device->tess_offchip_block_dw_size * 4;
- if (scratch_size <= queue->scratch_size &&
- compute_scratch_size <= queue->compute_scratch_size &&
+ scratch_size_per_wave = MAX2(scratch_size_per_wave, queue->scratch_size_per_wave);
+ if (scratch_size_per_wave)
+ scratch_waves = MIN2(scratch_waves, UINT32_MAX / scratch_size_per_wave);
+ else
+ scratch_waves = 0;
+
+ compute_scratch_size_per_wave = MAX2(compute_scratch_size_per_wave, queue->compute_scratch_size_per_wave);
+ if (compute_scratch_size_per_wave)
+ compute_scratch_waves = MIN2(compute_scratch_waves, UINT32_MAX / compute_scratch_size_per_wave);
+ else
+ compute_scratch_waves = 0;
+
+ if (scratch_size_per_wave <= queue->scratch_size_per_wave &&
+ scratch_waves <= queue->scratch_waves &&
+ compute_scratch_size_per_wave <= queue->compute_scratch_size_per_wave &&
+ compute_scratch_waves <= queue->compute_scratch_waves &&
esgs_ring_size <= queue->esgs_ring_size &&
gsvs_ring_size <= queue->gsvs_ring_size &&
- !add_tess_rings && !add_gds && !add_sample_positions &&
+ !add_tess_rings && !add_gds && !add_gds_oa && !add_sample_positions &&
queue->initial_preamble_cs) {
*initial_full_flush_preamble_cs = queue->initial_full_flush_preamble_cs;
*initial_preamble_cs = queue->initial_preamble_cs;
*continue_preamble_cs = queue->continue_preamble_cs;
- if (!scratch_size && !compute_scratch_size && !esgs_ring_size && !gsvs_ring_size)
+ if (!scratch_size_per_wave && !compute_scratch_size_per_wave &&
+ !esgs_ring_size && !gsvs_ring_size && !needs_tess_rings &&
+ !needs_gds && !needs_gds_oa && !needs_sample_positions)
*continue_preamble_cs = NULL;
return VK_SUCCESS;
}
- if (scratch_size > queue->scratch_size) {
+ uint32_t scratch_size = scratch_size_per_wave * scratch_waves;
+ uint32_t queue_scratch_size = queue->scratch_size_per_wave * queue->scratch_waves;
+ if (scratch_size > queue_scratch_size) {
scratch_bo = queue->device->ws->buffer_create(queue->device->ws,
scratch_size,
4096,
} else
scratch_bo = queue->scratch_bo;
- if (compute_scratch_size > queue->compute_scratch_size) {
+ uint32_t compute_scratch_size = compute_scratch_size_per_wave * compute_scratch_waves;
+ uint32_t compute_queue_scratch_size = queue->compute_scratch_size_per_wave * queue->compute_scratch_waves;
+ if (compute_scratch_size > compute_queue_scratch_size) {
compute_scratch_bo = queue->device->ws->buffer_create(queue->device->ws,
compute_scratch_size,
4096,
RADV_BO_PRIORITY_SCRATCH);
if (!gds_bo)
goto fail;
+ } else {
+ gds_bo = queue->gds_bo;
+ }
+
+ if (add_gds_oa) {
+ assert(queue->device->physical_device->rad_info.chip_class >= GFX10);
gds_oa_bo = queue->device->ws->buffer_create(queue->device->ws,
4, 1,
if (!gds_oa_bo)
goto fail;
} else {
- gds_bo = queue->gds_bo;
gds_oa_bo = queue->gds_oa_bo;
}
radv_emit_tess_factor_ring(queue, cs, hs_offchip_param,
tess_factor_ring_size, tess_rings_bo);
radv_emit_global_shader_pointers(queue, cs, descriptor_bo);
- radv_emit_compute_scratch(queue, cs, compute_scratch_bo);
+ radv_emit_compute_scratch(queue, cs, compute_scratch_size_per_wave,
+ compute_scratch_waves, compute_scratch_bo);
+ radv_emit_graphics_scratch(queue, cs, scratch_size_per_wave,
+ scratch_waves, scratch_bo);
if (gds_bo)
radv_cs_add_buffer(queue->device->ws, cs, gds_bo);
if (queue->scratch_bo)
queue->device->ws->buffer_destroy(queue->scratch_bo);
queue->scratch_bo = scratch_bo;
- queue->scratch_size = scratch_size;
}
+ queue->scratch_size_per_wave = scratch_size_per_wave;
+ queue->scratch_waves = scratch_waves;
if (compute_scratch_bo != queue->compute_scratch_bo) {
if (queue->compute_scratch_bo)
queue->device->ws->buffer_destroy(queue->compute_scratch_bo);
queue->compute_scratch_bo = compute_scratch_bo;
- queue->compute_scratch_size = compute_scratch_size;
}
+ queue->compute_scratch_size_per_wave = compute_scratch_size_per_wave;
+ queue->compute_scratch_waves = compute_scratch_waves;
if (esgs_ring_bo != queue->esgs_ring_bo) {
if (queue->esgs_ring_bo)
queue->has_gds = true;
}
- if (gds_oa_bo != queue->gds_oa_bo)
+ if (gds_oa_bo != queue->gds_oa_bo) {
queue->gds_oa_bo = gds_oa_bo;
+ queue->has_gds_oa = true;
+ }
if (descriptor_bo != queue->descriptor_bo) {
if (queue->descriptor_bo)
return vk_error(queue->device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
}
-static VkResult radv_alloc_sem_counts(struct radv_instance *instance,
+static VkResult radv_alloc_sem_counts(struct radv_device *device,
struct radv_winsys_sem_counts *counts,
int num_sems,
- const VkSemaphore *sems,
+ struct radv_semaphore_part **sems,
+ const uint64_t *timeline_values,
VkFence _fence,
- bool reset_temp)
+ bool is_signal)
{
int syncobj_idx = 0, sem_idx = 0;
return VK_SUCCESS;
for (uint32_t i = 0; i < num_sems; i++) {
- RADV_FROM_HANDLE(radv_semaphore, sem, sems[i]);
-
- if (sem->temp_syncobj || sem->syncobj)
+ switch(sems[i]->kind) {
+ case RADV_SEMAPHORE_SYNCOBJ:
counts->syncobj_count++;
- else
+ break;
+ case RADV_SEMAPHORE_WINSYS:
counts->sem_count++;
+ break;
+ case RADV_SEMAPHORE_NONE:
+ break;
+ case RADV_SEMAPHORE_TIMELINE:
+ counts->syncobj_count++;
+ break;
+ }
}
if (_fence != VK_NULL_HANDLE) {
if (counts->syncobj_count) {
counts->syncobj = (uint32_t *)malloc(sizeof(uint32_t) * counts->syncobj_count);
if (!counts->syncobj)
- return vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+ return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
if (counts->sem_count) {
counts->sem = (struct radeon_winsys_sem **)malloc(sizeof(struct radeon_winsys_sem *) * counts->sem_count);
if (!counts->sem) {
free(counts->syncobj);
- return vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+ return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
}
for (uint32_t i = 0; i < num_sems; i++) {
- RADV_FROM_HANDLE(radv_semaphore, sem, sems[i]);
+ switch(sems[i]->kind) {
+ case RADV_SEMAPHORE_NONE:
+ unreachable("Empty semaphore");
+ break;
+ case RADV_SEMAPHORE_SYNCOBJ:
+ counts->syncobj[syncobj_idx++] = sems[i]->syncobj;
+ break;
+ case RADV_SEMAPHORE_WINSYS:
+ counts->sem[sem_idx++] = sems[i]->ws_sem;
+ break;
+ case RADV_SEMAPHORE_TIMELINE: {
+ pthread_mutex_lock(&sems[i]->timeline.mutex);
+ struct radv_timeline_point *point = NULL;
+ if (is_signal) {
+ point = radv_timeline_add_point_locked(device, &sems[i]->timeline, timeline_values[i]);
+ } else {
+ point = radv_timeline_find_point_at_least_locked(device, &sems[i]->timeline, timeline_values[i]);
+ }
+
+ pthread_mutex_unlock(&sems[i]->timeline.mutex);
- if (sem->temp_syncobj) {
- counts->syncobj[syncobj_idx++] = sem->temp_syncobj;
+ if (point) {
+ counts->syncobj[syncobj_idx++] = point->syncobj;
+ } else {
+ /* Explicitly remove the semaphore so we might not find
+ * a point later post-submit. */
+ sems[i] = NULL;
+ }
+ break;
}
- else if (sem->syncobj)
- counts->syncobj[syncobj_idx++] = sem->syncobj;
- else {
- assert(sem->sem);
- counts->sem[sem_idx++] = sem->sem;
}
}
counts->syncobj[syncobj_idx++] = fence->syncobj;
}
+ assert(syncobj_idx <= counts->syncobj_count);
+ counts->syncobj_count = syncobj_idx;
+
return VK_SUCCESS;
}
static void radv_free_temp_syncobjs(struct radv_device *device,
int num_sems,
- const VkSemaphore *sems)
+ struct radv_semaphore_part *sems)
{
for (uint32_t i = 0; i < num_sems; i++) {
- RADV_FROM_HANDLE(radv_semaphore, sem, sems[i]);
-
- if (sem->temp_syncobj) {
- device->ws->destroy_syncobj(device->ws, sem->temp_syncobj);
- sem->temp_syncobj = 0;
- }
+ radv_destroy_semaphore_part(device, sems + i);
}
}
static VkResult
-radv_alloc_sem_info(struct radv_instance *instance,
+radv_alloc_sem_info(struct radv_device *device,
struct radv_winsys_sem_info *sem_info,
int num_wait_sems,
- const VkSemaphore *wait_sems,
+ struct radv_semaphore_part **wait_sems,
+ const uint64_t *wait_values,
int num_signal_sems,
- const VkSemaphore *signal_sems,
+ struct radv_semaphore_part **signal_sems,
+ const uint64_t *signal_values,
VkFence fence)
{
VkResult ret;
memset(sem_info, 0, sizeof(*sem_info));
- ret = radv_alloc_sem_counts(instance, &sem_info->wait, num_wait_sems, wait_sems, VK_NULL_HANDLE, true);
+ ret = radv_alloc_sem_counts(device, &sem_info->wait, num_wait_sems, wait_sems, wait_values, VK_NULL_HANDLE, false);
if (ret)
return ret;
- ret = radv_alloc_sem_counts(instance, &sem_info->signal, num_signal_sems, signal_sems, fence, false);
+ ret = radv_alloc_sem_counts(device, &sem_info->signal, num_signal_sems, signal_sems, signal_values, fence, true);
if (ret)
radv_free_sem_info(sem_info);
return ret;
}
-/* Signals fence as soon as all the work currently put on queue is done. */
-static VkResult radv_signal_fence(struct radv_queue *queue,
- struct radv_fence *fence)
+static void
+radv_finalize_timelines(struct radv_device *device,
+ uint32_t num_wait_sems,
+ struct radv_semaphore_part **wait_sems,
+ const uint64_t *wait_values,
+ uint32_t num_signal_sems,
+ struct radv_semaphore_part **signal_sems,
+ const uint64_t *signal_values,
+ struct list_head *processing_list)
+{
+ for (uint32_t i = 0; i < num_wait_sems; ++i) {
+ if (wait_sems[i] && wait_sems[i]->kind == RADV_SEMAPHORE_TIMELINE) {
+ pthread_mutex_lock(&wait_sems[i]->timeline.mutex);
+ struct radv_timeline_point *point =
+ radv_timeline_find_point_at_least_locked(device, &wait_sems[i]->timeline, wait_values[i]);
+ point->wait_count -= 2;
+ pthread_mutex_unlock(&wait_sems[i]->timeline.mutex);
+ }
+ }
+ for (uint32_t i = 0; i < num_signal_sems; ++i) {
+ if (signal_sems[i] && signal_sems[i]->kind == RADV_SEMAPHORE_TIMELINE) {
+ pthread_mutex_lock(&signal_sems[i]->timeline.mutex);
+ struct radv_timeline_point *point =
+ radv_timeline_find_point_at_least_locked(device, &signal_sems[i]->timeline, signal_values[i]);
+ signal_sems[i]->timeline.highest_submitted =
+ MAX2(signal_sems[i]->timeline.highest_submitted, point->value);
+ point->wait_count -= 2;
+ radv_timeline_trigger_waiters_locked(&signal_sems[i]->timeline, processing_list);
+ pthread_mutex_unlock(&signal_sems[i]->timeline.mutex);
+ }
+ }
+}
+
+static void
+radv_sparse_buffer_bind_memory(struct radv_device *device,
+ const VkSparseBufferMemoryBindInfo *bind)
{
- int ret;
- VkResult result;
- struct radv_winsys_sem_info sem_info;
+ RADV_FROM_HANDLE(radv_buffer, buffer, bind->buffer);
- result = radv_alloc_sem_info(queue->device->instance, &sem_info, 0, NULL, 0, NULL,
- radv_fence_to_handle(fence));
- if (result != VK_SUCCESS)
- return result;
+ for (uint32_t i = 0; i < bind->bindCount; ++i) {
+ struct radv_device_memory *mem = NULL;
- ret = queue->device->ws->cs_submit(queue->hw_ctx, queue->queue_idx,
- &queue->device->empty_cs[queue->queue_family_index],
- 1, NULL, NULL, &sem_info, NULL,
- false, fence->fence);
- radv_free_sem_info(&sem_info);
+ if (bind->pBinds[i].memory != VK_NULL_HANDLE)
+ mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
- if (ret)
- return vk_error(queue->device->instance, VK_ERROR_DEVICE_LOST);
+ device->ws->buffer_virtual_bind(buffer->bo,
+ bind->pBinds[i].resourceOffset,
+ bind->pBinds[i].size,
+ mem ? mem->bo : NULL,
+ bind->pBinds[i].memoryOffset);
+ }
+}
+
+static void
+radv_sparse_image_opaque_bind_memory(struct radv_device *device,
+ const VkSparseImageOpaqueMemoryBindInfo *bind)
+{
+ RADV_FROM_HANDLE(radv_image, image, bind->image);
+
+ for (uint32_t i = 0; i < bind->bindCount; ++i) {
+ struct radv_device_memory *mem = NULL;
+
+ if (bind->pBinds[i].memory != VK_NULL_HANDLE)
+ mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
+
+ device->ws->buffer_virtual_bind(image->bo,
+ bind->pBinds[i].resourceOffset,
+ bind->pBinds[i].size,
+ mem ? mem->bo : NULL,
+ bind->pBinds[i].memoryOffset);
+ }
+}
+static VkResult
+radv_get_preambles(struct radv_queue *queue,
+ const VkCommandBuffer *cmd_buffers,
+ uint32_t cmd_buffer_count,
+ struct radeon_cmdbuf **initial_full_flush_preamble_cs,
+ struct radeon_cmdbuf **initial_preamble_cs,
+ struct radeon_cmdbuf **continue_preamble_cs)
+{
+ uint32_t scratch_size_per_wave = 0, waves_wanted = 0;
+ uint32_t compute_scratch_size_per_wave = 0, compute_waves_wanted = 0;
+ uint32_t esgs_ring_size = 0, gsvs_ring_size = 0;
+ bool tess_rings_needed = false;
+ bool gds_needed = false;
+ bool gds_oa_needed = false;
+ bool sample_positions_needed = false;
+
+ for (uint32_t j = 0; j < cmd_buffer_count; j++) {
+ RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer,
+ cmd_buffers[j]);
+
+ scratch_size_per_wave = MAX2(scratch_size_per_wave, cmd_buffer->scratch_size_per_wave_needed);
+ waves_wanted = MAX2(waves_wanted, cmd_buffer->scratch_waves_wanted);
+ compute_scratch_size_per_wave = MAX2(compute_scratch_size_per_wave,
+ cmd_buffer->compute_scratch_size_per_wave_needed);
+ compute_waves_wanted = MAX2(compute_waves_wanted,
+ cmd_buffer->compute_scratch_waves_wanted);
+ esgs_ring_size = MAX2(esgs_ring_size, cmd_buffer->esgs_ring_size_needed);
+ gsvs_ring_size = MAX2(gsvs_ring_size, cmd_buffer->gsvs_ring_size_needed);
+ tess_rings_needed |= cmd_buffer->tess_rings_needed;
+ gds_needed |= cmd_buffer->gds_needed;
+ gds_oa_needed |= cmd_buffer->gds_oa_needed;
+ sample_positions_needed |= cmd_buffer->sample_positions_needed;
+ }
+
+ return radv_get_preamble_cs(queue, scratch_size_per_wave, waves_wanted,
+ compute_scratch_size_per_wave, compute_waves_wanted,
+ esgs_ring_size, gsvs_ring_size, tess_rings_needed,
+ gds_needed, gds_oa_needed, sample_positions_needed,
+ initial_full_flush_preamble_cs,
+ initial_preamble_cs, continue_preamble_cs);
+}
+
+struct radv_deferred_queue_submission {
+ struct radv_queue *queue;
+ VkCommandBuffer *cmd_buffers;
+ uint32_t cmd_buffer_count;
+
+ /* Sparse bindings that happen on a queue. */
+ VkSparseBufferMemoryBindInfo *buffer_binds;
+ uint32_t buffer_bind_count;
+ VkSparseImageOpaqueMemoryBindInfo *image_opaque_binds;
+ uint32_t image_opaque_bind_count;
+
+ bool flush_caches;
+ VkShaderStageFlags wait_dst_stage_mask;
+ struct radv_semaphore_part **wait_semaphores;
+ uint32_t wait_semaphore_count;
+ struct radv_semaphore_part **signal_semaphores;
+ uint32_t signal_semaphore_count;
+ VkFence fence;
+
+ uint64_t *wait_values;
+ uint64_t *signal_values;
+
+ struct radv_semaphore_part *temporary_semaphore_parts;
+ uint32_t temporary_semaphore_part_count;
+
+ struct list_head queue_pending_list;
+ uint32_t submission_wait_count;
+ struct radv_timeline_waiter *wait_nodes;
+
+ struct list_head processing_list;
+};
+
+struct radv_queue_submission {
+ const VkCommandBuffer *cmd_buffers;
+ uint32_t cmd_buffer_count;
+
+ /* Sparse bindings that happen on a queue. */
+ const VkSparseBufferMemoryBindInfo *buffer_binds;
+ uint32_t buffer_bind_count;
+ const VkSparseImageOpaqueMemoryBindInfo *image_opaque_binds;
+ uint32_t image_opaque_bind_count;
+
+ bool flush_caches;
+ VkPipelineStageFlags wait_dst_stage_mask;
+ const VkSemaphore *wait_semaphores;
+ uint32_t wait_semaphore_count;
+ const VkSemaphore *signal_semaphores;
+ uint32_t signal_semaphore_count;
+ VkFence fence;
+
+ const uint64_t *wait_values;
+ uint32_t wait_value_count;
+ const uint64_t *signal_values;
+ uint32_t signal_value_count;
+};
+
+static VkResult
+radv_create_deferred_submission(struct radv_queue *queue,
+ const struct radv_queue_submission *submission,
+ struct radv_deferred_queue_submission **out)
+{
+ struct radv_deferred_queue_submission *deferred = NULL;
+ size_t size = sizeof(struct radv_deferred_queue_submission);
+
+ uint32_t temporary_count = 0;
+ for (uint32_t i = 0; i < submission->wait_semaphore_count; ++i) {
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, submission->wait_semaphores[i]);
+ if (semaphore->temporary.kind != RADV_SEMAPHORE_NONE)
+ ++temporary_count;
+ }
+
+ size += submission->cmd_buffer_count * sizeof(VkCommandBuffer);
+ size += submission->buffer_bind_count * sizeof(VkSparseBufferMemoryBindInfo);
+ size += submission->image_opaque_bind_count * sizeof(VkSparseImageOpaqueMemoryBindInfo);
+ size += submission->wait_semaphore_count * sizeof(struct radv_semaphore_part *);
+ size += temporary_count * sizeof(struct radv_semaphore_part);
+ size += submission->signal_semaphore_count * sizeof(struct radv_semaphore_part *);
+ size += submission->wait_value_count * sizeof(uint64_t);
+ size += submission->signal_value_count * sizeof(uint64_t);
+ size += submission->wait_semaphore_count * sizeof(struct radv_timeline_waiter);
+
+ deferred = calloc(1, size);
+ if (!deferred)
+ return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+ deferred->queue = queue;
+
+ deferred->cmd_buffers = (void*)(deferred + 1);
+ deferred->cmd_buffer_count = submission->cmd_buffer_count;
+ memcpy(deferred->cmd_buffers, submission->cmd_buffers,
+ submission->cmd_buffer_count * sizeof(*deferred->cmd_buffers));
+
+ deferred->buffer_binds = (void*)(deferred->cmd_buffers + submission->cmd_buffer_count);
+ deferred->buffer_bind_count = submission->buffer_bind_count;
+ memcpy(deferred->buffer_binds, submission->buffer_binds,
+ submission->buffer_bind_count * sizeof(*deferred->buffer_binds));
+
+ deferred->image_opaque_binds = (void*)(deferred->buffer_binds + submission->buffer_bind_count);
+ deferred->image_opaque_bind_count = submission->image_opaque_bind_count;
+ memcpy(deferred->image_opaque_binds, submission->image_opaque_binds,
+ submission->image_opaque_bind_count * sizeof(*deferred->image_opaque_binds));
+
+ deferred->flush_caches = submission->flush_caches;
+ deferred->wait_dst_stage_mask = submission->wait_dst_stage_mask;
+
+ deferred->wait_semaphores = (void*)(deferred->image_opaque_binds + deferred->image_opaque_bind_count);
+ deferred->wait_semaphore_count = submission->wait_semaphore_count;
+
+ deferred->signal_semaphores = (void*)(deferred->wait_semaphores + deferred->wait_semaphore_count);
+ deferred->signal_semaphore_count = submission->signal_semaphore_count;
+
+ deferred->fence = submission->fence;
+
+ deferred->temporary_semaphore_parts = (void*)(deferred->signal_semaphores + deferred->signal_semaphore_count);
+ deferred->temporary_semaphore_part_count = temporary_count;
+
+ uint32_t temporary_idx = 0;
+ for (uint32_t i = 0; i < submission->wait_semaphore_count; ++i) {
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, submission->wait_semaphores[i]);
+ if (semaphore->temporary.kind != RADV_SEMAPHORE_NONE) {
+ deferred->wait_semaphores[i] = &deferred->temporary_semaphore_parts[temporary_idx];
+ deferred->temporary_semaphore_parts[temporary_idx] = semaphore->temporary;
+ semaphore->temporary.kind = RADV_SEMAPHORE_NONE;
+ ++temporary_idx;
+ } else
+ deferred->wait_semaphores[i] = &semaphore->permanent;
+ }
+
+ for (uint32_t i = 0; i < submission->signal_semaphore_count; ++i) {
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, submission->signal_semaphores[i]);
+ if (semaphore->temporary.kind != RADV_SEMAPHORE_NONE) {
+ deferred->signal_semaphores[i] = &semaphore->temporary;
+ } else {
+ deferred->signal_semaphores[i] = &semaphore->permanent;
+ }
+ }
+
+ deferred->wait_values = (void*)(deferred->temporary_semaphore_parts + temporary_count);
+ memcpy(deferred->wait_values, submission->wait_values, submission->wait_value_count * sizeof(uint64_t));
+ deferred->signal_values = deferred->wait_values + submission->wait_value_count;
+ memcpy(deferred->signal_values, submission->signal_values, submission->signal_value_count * sizeof(uint64_t));
+
+ deferred->wait_nodes = (void*)(deferred->signal_values + submission->signal_value_count);
+ /* This is worst-case. radv_queue_enqueue_submission will fill in further, but this
+ * ensure the submission is not accidentally triggered early when adding wait timelines. */
+ deferred->submission_wait_count = 1 + submission->wait_semaphore_count;
+
+ *out = deferred;
return VK_SUCCESS;
}
-VkResult radv_QueueSubmit(
- VkQueue _queue,
- uint32_t submitCount,
- const VkSubmitInfo* pSubmits,
- VkFence _fence)
+static void
+radv_queue_enqueue_submission(struct radv_deferred_queue_submission *submission,
+ struct list_head *processing_list)
+{
+ uint32_t wait_cnt = 0;
+ struct radv_timeline_waiter *waiter = submission->wait_nodes;
+ for (uint32_t i = 0; i < submission->wait_semaphore_count; ++i) {
+ if (submission->wait_semaphores[i]->kind == RADV_SEMAPHORE_TIMELINE) {
+ pthread_mutex_lock(&submission->wait_semaphores[i]->timeline.mutex);
+ if (submission->wait_semaphores[i]->timeline.highest_submitted < submission->wait_values[i]) {
+ ++wait_cnt;
+ waiter->value = submission->wait_values[i];
+ waiter->submission = submission;
+ list_addtail(&waiter->list, &submission->wait_semaphores[i]->timeline.waiters);
+ ++waiter;
+ }
+ pthread_mutex_unlock(&submission->wait_semaphores[i]->timeline.mutex);
+ }
+ }
+
+ pthread_mutex_lock(&submission->queue->pending_mutex);
+
+ bool is_first = list_is_empty(&submission->queue->pending_submissions);
+ list_addtail(&submission->queue_pending_list, &submission->queue->pending_submissions);
+
+ pthread_mutex_unlock(&submission->queue->pending_mutex);
+
+ /* If there is already a submission in the queue, that will decrement the counter by 1 when
+ * submitted, but if the queue was empty, we decrement ourselves as there is no previous
+ * submission. */
+ uint32_t decrement = submission->wait_semaphore_count - wait_cnt + (is_first ? 1 : 0);
+ if (__atomic_sub_fetch(&submission->submission_wait_count, decrement, __ATOMIC_ACQ_REL) == 0) {
+ list_addtail(&submission->processing_list, processing_list);
+ }
+}
+
+static void
+radv_queue_submission_update_queue(struct radv_deferred_queue_submission *submission,
+ struct list_head *processing_list)
{
- RADV_FROM_HANDLE(radv_queue, queue, _queue);
- RADV_FROM_HANDLE(radv_fence, fence, _fence);
- struct radeon_winsys_fence *base_fence = fence ? fence->fence : NULL;
+ pthread_mutex_lock(&submission->queue->pending_mutex);
+ list_del(&submission->queue_pending_list);
+
+ /* trigger the next submission in the queue. */
+ if (!list_is_empty(&submission->queue->pending_submissions)) {
+ struct radv_deferred_queue_submission *next_submission =
+ list_first_entry(&submission->queue->pending_submissions,
+ struct radv_deferred_queue_submission,
+ queue_pending_list);
+ if (p_atomic_dec_zero(&next_submission->submission_wait_count)) {
+ list_addtail(&next_submission->processing_list, processing_list);
+ }
+ }
+ pthread_mutex_unlock(&submission->queue->pending_mutex);
+
+ pthread_cond_broadcast(&submission->queue->device->timeline_cond);
+}
+
+static VkResult
+radv_queue_submit_deferred(struct radv_deferred_queue_submission *submission,
+ struct list_head *processing_list)
+{
+ RADV_FROM_HANDLE(radv_fence, fence, submission->fence);
+ struct radv_queue *queue = submission->queue;
struct radeon_winsys_ctx *ctx = queue->hw_ctx;
- int ret;
uint32_t max_cs_submission = queue->device->trace_bo ? 1 : RADV_MAX_IBS_PER_SUBMIT;
- uint32_t scratch_size = 0;
- uint32_t compute_scratch_size = 0;
- uint32_t esgs_ring_size = 0, gsvs_ring_size = 0;
- struct radeon_cmdbuf *initial_preamble_cs = NULL, *initial_flush_preamble_cs = NULL, *continue_preamble_cs = NULL;
+ struct radeon_winsys_fence *base_fence = fence ? fence->fence : NULL;
+ bool do_flush = submission->flush_caches || submission->wait_dst_stage_mask;
+ bool can_patch = true;
+ uint32_t advance;
+ struct radv_winsys_sem_info sem_info;
VkResult result;
- bool fence_emitted = false;
- bool tess_rings_needed = false;
- bool gds_needed = false;
- bool sample_positions_needed = false;
+ int ret;
+ struct radeon_cmdbuf *initial_preamble_cs = NULL;
+ struct radeon_cmdbuf *initial_flush_preamble_cs = NULL;
+ struct radeon_cmdbuf *continue_preamble_cs = NULL;
+
+ result = radv_get_preambles(queue, submission->cmd_buffers,
+ submission->cmd_buffer_count,
+ &initial_preamble_cs,
+ &initial_flush_preamble_cs,
+ &continue_preamble_cs);
+ if (result != VK_SUCCESS)
+ goto fail;
- /* Do this first so failing to allocate scratch buffers can't result in
- * partially executed submissions. */
- for (uint32_t i = 0; i < submitCount; i++) {
- for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
- RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer,
- pSubmits[i].pCommandBuffers[j]);
-
- scratch_size = MAX2(scratch_size, cmd_buffer->scratch_size_needed);
- compute_scratch_size = MAX2(compute_scratch_size,
- cmd_buffer->compute_scratch_size_needed);
- esgs_ring_size = MAX2(esgs_ring_size, cmd_buffer->esgs_ring_size_needed);
- gsvs_ring_size = MAX2(gsvs_ring_size, cmd_buffer->gsvs_ring_size_needed);
- tess_rings_needed |= cmd_buffer->tess_rings_needed;
- gds_needed |= cmd_buffer->gds_needed;
- sample_positions_needed |= cmd_buffer->sample_positions_needed;
- }
- }
-
- result = radv_get_preamble_cs(queue, scratch_size, compute_scratch_size,
- esgs_ring_size, gsvs_ring_size, tess_rings_needed,
- gds_needed, sample_positions_needed,
- &initial_flush_preamble_cs,
- &initial_preamble_cs, &continue_preamble_cs);
+ result = radv_alloc_sem_info(queue->device,
+ &sem_info,
+ submission->wait_semaphore_count,
+ submission->wait_semaphores,
+ submission->wait_values,
+ submission->signal_semaphore_count,
+ submission->signal_semaphores,
+ submission->signal_values,
+ submission->fence);
if (result != VK_SUCCESS)
- return result;
+ goto fail;
- for (uint32_t i = 0; i < submitCount; i++) {
- struct radeon_cmdbuf **cs_array;
- bool do_flush = !i || pSubmits[i].pWaitDstStageMask;
- bool can_patch = true;
- uint32_t advance;
- struct radv_winsys_sem_info sem_info;
-
- result = radv_alloc_sem_info(queue->device->instance,
- &sem_info,
- pSubmits[i].waitSemaphoreCount,
- pSubmits[i].pWaitSemaphores,
- pSubmits[i].signalSemaphoreCount,
- pSubmits[i].pSignalSemaphores,
- _fence);
- if (result != VK_SUCCESS)
- return result;
+ for (uint32_t i = 0; i < submission->buffer_bind_count; ++i) {
+ radv_sparse_buffer_bind_memory(queue->device,
+ submission->buffer_binds + i);
+ }
- if (!pSubmits[i].commandBufferCount) {
- if (pSubmits[i].waitSemaphoreCount || pSubmits[i].signalSemaphoreCount) {
- ret = queue->device->ws->cs_submit(ctx, queue->queue_idx,
- &queue->device->empty_cs[queue->queue_family_index],
- 1, NULL, NULL,
- &sem_info, NULL,
- false, base_fence);
- if (ret) {
- radv_loge("failed to submit CS %d\n", i);
- abort();
- }
- fence_emitted = true;
- }
- radv_free_sem_info(&sem_info);
- continue;
+ for (uint32_t i = 0; i < submission->image_opaque_bind_count; ++i) {
+ radv_sparse_image_opaque_bind_memory(queue->device,
+ submission->image_opaque_binds + i);
+ }
+
+ if (!submission->cmd_buffer_count) {
+ ret = queue->device->ws->cs_submit(ctx, queue->queue_idx,
+ &queue->device->empty_cs[queue->queue_family_index],
+ 1, NULL, NULL,
+ &sem_info, NULL,
+ false, base_fence);
+ if (ret) {
+ radv_loge("failed to submit CS\n");
+ abort();
}
- cs_array = malloc(sizeof(struct radeon_cmdbuf *) *
- (pSubmits[i].commandBufferCount));
+ goto success;
+ } else {
+ struct radeon_cmdbuf **cs_array = malloc(sizeof(struct radeon_cmdbuf *) *
+ (submission->cmd_buffer_count));
- for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
- RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer,
- pSubmits[i].pCommandBuffers[j]);
+ for (uint32_t j = 0; j < submission->cmd_buffer_count; j++) {
+ RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, submission->cmd_buffers[j]);
assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
cs_array[j] = cmd_buffer->cs;
cmd_buffer->status = RADV_CMD_BUFFER_STATUS_PENDING;
}
- for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j += advance) {
+ for (uint32_t j = 0; j < submission->cmd_buffer_count; j += advance) {
struct radeon_cmdbuf *initial_preamble = (do_flush && !j) ? initial_flush_preamble_cs : initial_preamble_cs;
const struct radv_winsys_bo_list *bo_list = NULL;
advance = MIN2(max_cs_submission,
- pSubmits[i].commandBufferCount - j);
+ submission->cmd_buffer_count - j);
if (queue->device->trace_bo)
*queue->device->trace_id_ptr = 0;
sem_info.cs_emit_wait = j == 0;
- sem_info.cs_emit_signal = j + advance == pSubmits[i].commandBufferCount;
+ sem_info.cs_emit_signal = j + advance == submission->cmd_buffer_count;
if (unlikely(queue->device->use_global_bo_list)) {
pthread_mutex_lock(&queue->device->bo_list.mutex);
}
ret = queue->device->ws->cs_submit(ctx, queue->queue_idx, cs_array + j,
- advance, initial_preamble, continue_preamble_cs,
- &sem_info, bo_list,
- can_patch, base_fence);
+ advance, initial_preamble, continue_preamble_cs,
+ &sem_info, bo_list,
+ can_patch, base_fence);
if (unlikely(queue->device->use_global_bo_list))
pthread_mutex_unlock(&queue->device->bo_list.mutex);
if (ret) {
- radv_loge("failed to submit CS %d\n", i);
+ radv_loge("failed to submit CS\n");
abort();
}
- fence_emitted = true;
if (queue->device->trace_bo) {
radv_check_gpu_hangs(queue, cs_array[j]);
}
}
- radv_free_temp_syncobjs(queue->device,
- pSubmits[i].waitSemaphoreCount,
- pSubmits[i].pWaitSemaphores);
- radv_free_sem_info(&sem_info);
- free(cs_array);
+ free(cs_array);
+ }
+
+success:
+ radv_free_temp_syncobjs(queue->device,
+ submission->temporary_semaphore_part_count,
+ submission->temporary_semaphore_parts);
+ radv_finalize_timelines(queue->device,
+ submission->wait_semaphore_count,
+ submission->wait_semaphores,
+ submission->wait_values,
+ submission->signal_semaphore_count,
+ submission->signal_semaphores,
+ submission->signal_values,
+ processing_list);
+ /* Has to happen after timeline finalization to make sure the
+ * condition variable is only triggered when timelines and queue have
+ * been updated. */
+ radv_queue_submission_update_queue(submission, processing_list);
+ radv_free_sem_info(&sem_info);
+ free(submission);
+ return VK_SUCCESS;
+
+fail:
+ radv_free_temp_syncobjs(queue->device,
+ submission->temporary_semaphore_part_count,
+ submission->temporary_semaphore_parts);
+ free(submission);
+ return VK_ERROR_DEVICE_LOST;
+}
+
+static VkResult
+radv_process_submissions(struct list_head *processing_list)
+{
+ while(!list_is_empty(processing_list)) {
+ struct radv_deferred_queue_submission *submission =
+ list_first_entry(processing_list, struct radv_deferred_queue_submission, processing_list);
+ list_del(&submission->processing_list);
+
+ VkResult result = radv_queue_submit_deferred(submission, processing_list);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+ return VK_SUCCESS;
+}
+
+static VkResult radv_queue_submit(struct radv_queue *queue,
+ const struct radv_queue_submission *submission)
+{
+ struct radv_deferred_queue_submission *deferred = NULL;
+
+ VkResult result = radv_create_deferred_submission(queue, submission, &deferred);
+ if (result != VK_SUCCESS)
+ return result;
+
+ struct list_head processing_list;
+ list_inithead(&processing_list);
+
+ radv_queue_enqueue_submission(deferred, &processing_list);
+ return radv_process_submissions(&processing_list);
+}
+
+/* Signals fence as soon as all the work currently put on queue is done. */
+static VkResult radv_signal_fence(struct radv_queue *queue,
+ VkFence fence)
+{
+ return radv_queue_submit(queue, &(struct radv_queue_submission) {
+ .fence = fence
+ });
+}
+
+static bool radv_submit_has_effects(const VkSubmitInfo *info)
+{
+ return info->commandBufferCount ||
+ info->waitSemaphoreCount ||
+ info->signalSemaphoreCount;
+}
+
+VkResult radv_QueueSubmit(
+ VkQueue _queue,
+ uint32_t submitCount,
+ const VkSubmitInfo* pSubmits,
+ VkFence fence)
+{
+ RADV_FROM_HANDLE(radv_queue, queue, _queue);
+ VkResult result;
+ uint32_t fence_idx = 0;
+ bool flushed_caches = false;
+
+ if (fence != VK_NULL_HANDLE) {
+ for (uint32_t i = 0; i < submitCount; ++i)
+ if (radv_submit_has_effects(pSubmits + i))
+ fence_idx = i;
+ } else
+ fence_idx = UINT32_MAX;
+
+ for (uint32_t i = 0; i < submitCount; i++) {
+ if (!radv_submit_has_effects(pSubmits + i) && fence_idx != i)
+ continue;
+
+ VkPipelineStageFlags wait_dst_stage_mask = 0;
+ for (unsigned j = 0; j < pSubmits[i].waitSemaphoreCount; ++j) {
+ wait_dst_stage_mask |= pSubmits[i].pWaitDstStageMask[j];
+ }
+
+ const VkTimelineSemaphoreSubmitInfo *timeline_info =
+ vk_find_struct_const(pSubmits[i].pNext, TIMELINE_SEMAPHORE_SUBMIT_INFO);
+
+ result = radv_queue_submit(queue, &(struct radv_queue_submission) {
+ .cmd_buffers = pSubmits[i].pCommandBuffers,
+ .cmd_buffer_count = pSubmits[i].commandBufferCount,
+ .wait_dst_stage_mask = wait_dst_stage_mask,
+ .flush_caches = !flushed_caches,
+ .wait_semaphores = pSubmits[i].pWaitSemaphores,
+ .wait_semaphore_count = pSubmits[i].waitSemaphoreCount,
+ .signal_semaphores = pSubmits[i].pSignalSemaphores,
+ .signal_semaphore_count = pSubmits[i].signalSemaphoreCount,
+ .fence = i == fence_idx ? fence : VK_NULL_HANDLE,
+ .wait_values = timeline_info ? timeline_info->pWaitSemaphoreValues : NULL,
+ .wait_value_count = timeline_info && timeline_info->pWaitSemaphoreValues ? timeline_info->waitSemaphoreValueCount : 0,
+ .signal_values = timeline_info ? timeline_info->pSignalSemaphoreValues : NULL,
+ .signal_value_count = timeline_info && timeline_info->pSignalSemaphoreValues ? timeline_info->signalSemaphoreValueCount : 0,
+ });
+ if (result != VK_SUCCESS)
+ return result;
+
+ flushed_caches = true;
}
- if (fence) {
- if (!fence_emitted) {
- result = radv_signal_fence(queue, fence);
- if (result != VK_SUCCESS)
- return result;
- }
+ if (fence != VK_NULL_HANDLE && !submitCount) {
+ result = radv_signal_fence(queue, fence);
+ if (result != VK_SUCCESS)
+ return result;
}
return VK_SUCCESS;
{
RADV_FROM_HANDLE(radv_queue, queue, _queue);
+ pthread_mutex_lock(&queue->pending_mutex);
+ while (!list_is_empty(&queue->pending_submissions)) {
+ pthread_cond_wait(&queue->device->timeline_cond, &queue->pending_mutex);
+ }
+ pthread_mutex_unlock(&queue->pending_mutex);
+
queue->device->ws->ctx_wait_idle(queue->hw_ctx,
radv_queue_family_to_ring(queue->queue_family_index),
queue->queue_idx);
struct radeon_bo_metadata metadata;
if (memory->image) {
- radv_init_metadata(device, memory->image, &metadata);
+ if (memory->image->tiling != VK_IMAGE_TILING_LINEAR)
+ radv_init_metadata(device, memory->image, &metadata);
device->ws->buffer_set_metadata(memory->bo, &metadata);
}
pFD);
}
+
+static void radv_free_memory(struct radv_device *device,
+ const VkAllocationCallbacks* pAllocator,
+ struct radv_device_memory *mem)
+{
+ if (mem == NULL)
+ return;
+
+#if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
+ if (mem->android_hardware_buffer)
+ AHardwareBuffer_release(mem->android_hardware_buffer);
+#endif
+
+ if (mem->bo) {
+ radv_bo_list_remove(device, mem->bo);
+ device->ws->buffer_destroy(mem->bo);
+ mem->bo = NULL;
+ }
+
+ vk_free2(&device->alloc, pAllocator, mem);
+}
+
static VkResult radv_alloc_memory(struct radv_device *device,
const VkMemoryAllocateInfo* pAllocateInfo,
const VkAllocationCallbacks* pAllocator,
assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
- if (pAllocateInfo->allocationSize == 0) {
- /* Apparently, this is allowed */
- *pMem = VK_NULL_HANDLE;
- return VK_SUCCESS;
- }
-
const VkImportMemoryFdInfoKHR *import_info =
vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
const VkMemoryDedicatedAllocateInfo *dedicate_info =
vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
const VkExportMemoryAllocateInfo *export_info =
vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
+ const struct VkImportAndroidHardwareBufferInfoANDROID *ahb_import_info =
+ vk_find_struct_const(pAllocateInfo->pNext,
+ IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
const VkImportMemoryHostPointerInfoEXT *host_ptr_info =
vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_HOST_POINTER_INFO_EXT);
const struct wsi_memory_allocate_info *wsi_info =
vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
- mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
+ if (pAllocateInfo->allocationSize == 0 && !ahb_import_info &&
+ !(export_info && (export_info->handleTypes & VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID))) {
+ /* Apparently, this is allowed */
+ *pMem = VK_NULL_HANDLE;
+ return VK_SUCCESS;
+ }
+
+ mem = vk_zalloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (mem == NULL)
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
(int)(priority_float * RADV_BO_PRIORITY_APPLICATION_MAX));
mem->user_ptr = NULL;
+ mem->bo = NULL;
- if (import_info) {
+#if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
+ mem->android_hardware_buffer = NULL;
+#endif
+
+ if (ahb_import_info) {
+ result = radv_import_ahb_memory(device, mem, priority, ahb_import_info);
+ if (result != VK_SUCCESS)
+ goto fail;
+ } else if(export_info && (export_info->handleTypes & VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)) {
+ result = radv_create_ahb_memory(device, mem, priority, pAllocateInfo);
+ if (result != VK_SUCCESS)
+ goto fail;
+ } else if (import_info) {
assert(import_info->handleType ==
VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
import_info->handleType ==
VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
mem->bo = device->ws->buffer_from_fd(device->ws, import_info->fd,
- priority, NULL, NULL);
+ priority, NULL);
if (!mem->bo) {
result = VK_ERROR_INVALID_EXTERNAL_HANDLE;
goto fail;
}
} else if (host_ptr_info) {
assert(host_ptr_info->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
- assert(mem_type_index == RADV_MEM_TYPE_GTT_CACHED);
+ assert(radv_is_mem_type_gtt_cached(mem_type_index));
mem->bo = device->ws->buffer_from_ptr(device->ws, host_ptr_info->pHostPointer,
pAllocateInfo->allocationSize,
priority);
}
} else {
uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
- if (mem_type_index == RADV_MEM_TYPE_GTT_WRITE_COMBINE ||
- mem_type_index == RADV_MEM_TYPE_GTT_CACHED)
+ if (radv_is_mem_type_gtt_wc(mem_type_index) ||
+ radv_is_mem_type_gtt_cached(mem_type_index))
domain = RADEON_DOMAIN_GTT;
else
domain = RADEON_DOMAIN_VRAM;
- if (mem_type_index == RADV_MEM_TYPE_VRAM)
+ if (radv_is_mem_type_vram(mem_type_index))
flags |= RADEON_FLAG_NO_CPU_ACCESS;
else
flags |= RADEON_FLAG_CPU_ACCESS;
- if (mem_type_index == RADV_MEM_TYPE_GTT_WRITE_COMBINE)
+ if (radv_is_mem_type_gtt_wc(mem_type_index))
flags |= RADEON_FLAG_GTT_WC;
if (!dedicate_info && !import_info && (!export_info || !export_info->handleTypes)) {
}
}
+ if (radv_is_mem_type_uncached(mem_type_index)) {
+ assert(device->physical_device->rad_info.has_l2_uncached);
+ flags |= RADEON_FLAG_VA_UNCACHED;
+ }
+
mem->bo = device->ws->buffer_create(device->ws, alloc_size, device->physical_device->rad_info.max_alignment,
domain, flags, priority);
result = radv_bo_list_add(device, mem->bo);
if (result != VK_SUCCESS)
- goto fail_bo;
+ goto fail;
*pMem = radv_device_memory_to_handle(mem);
return VK_SUCCESS;
-fail_bo:
- device->ws->buffer_destroy(mem->bo);
fail:
- vk_free2(&device->alloc, pAllocator, mem);
+ radv_free_memory(device, pAllocator,mem);
return result;
}
RADV_FROM_HANDLE(radv_device, device, _device);
RADV_FROM_HANDLE(radv_device_memory, mem, _mem);
- if (mem == NULL)
- return;
-
- radv_bo_list_remove(device, mem->bo);
- device->ws->buffer_destroy(mem->bo);
- mem->bo = NULL;
-
- vk_free2(&device->alloc, pAllocator, mem);
+ radv_free_memory(device, pAllocator, mem);
}
VkResult radv_MapMemory(
{
radv_GetBufferMemoryRequirements(device, pInfo->buffer,
&pMemoryRequirements->memoryRequirements);
- RADV_FROM_HANDLE(radv_buffer, buffer, pInfo->buffer);
vk_foreach_struct(ext, pMemoryRequirements->pNext) {
switch (ext->sType) {
case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
VkMemoryDedicatedRequirements *req =
(VkMemoryDedicatedRequirements *) ext;
- req->requiresDedicatedAllocation = buffer->shareable;
+ req->requiresDedicatedAllocation = false;
req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
break;
}
case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
VkMemoryDedicatedRequirements *req =
(VkMemoryDedicatedRequirements *) ext;
- req->requiresDedicatedAllocation = image->shareable;
+ req->requiresDedicatedAllocation = image->shareable &&
+ image->tiling != VK_IMAGE_TILING_LINEAR;
req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
break;
}
return radv_BindImageMemory2(device, 1, &info);
}
-
-static void
-radv_sparse_buffer_bind_memory(struct radv_device *device,
- const VkSparseBufferMemoryBindInfo *bind)
-{
- RADV_FROM_HANDLE(radv_buffer, buffer, bind->buffer);
-
- for (uint32_t i = 0; i < bind->bindCount; ++i) {
- struct radv_device_memory *mem = NULL;
-
- if (bind->pBinds[i].memory != VK_NULL_HANDLE)
- mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
-
- device->ws->buffer_virtual_bind(buffer->bo,
- bind->pBinds[i].resourceOffset,
- bind->pBinds[i].size,
- mem ? mem->bo : NULL,
- bind->pBinds[i].memoryOffset);
- }
-}
-
-static void
-radv_sparse_image_opaque_bind_memory(struct radv_device *device,
- const VkSparseImageOpaqueMemoryBindInfo *bind)
+static bool radv_sparse_bind_has_effects(const VkBindSparseInfo *info)
{
- RADV_FROM_HANDLE(radv_image, image, bind->image);
-
- for (uint32_t i = 0; i < bind->bindCount; ++i) {
- struct radv_device_memory *mem = NULL;
-
- if (bind->pBinds[i].memory != VK_NULL_HANDLE)
- mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
-
- device->ws->buffer_virtual_bind(image->bo,
- bind->pBinds[i].resourceOffset,
- bind->pBinds[i].size,
- mem ? mem->bo : NULL,
- bind->pBinds[i].memoryOffset);
- }
+ return info->bufferBindCount ||
+ info->imageOpaqueBindCount ||
+ info->imageBindCount ||
+ info->waitSemaphoreCount ||
+ info->signalSemaphoreCount;
}
VkResult radv_QueueBindSparse(
VkQueue _queue,
uint32_t bindInfoCount,
const VkBindSparseInfo* pBindInfo,
- VkFence _fence)
+ VkFence fence)
{
- RADV_FROM_HANDLE(radv_fence, fence, _fence);
RADV_FROM_HANDLE(radv_queue, queue, _queue);
- struct radeon_winsys_fence *base_fence = fence ? fence->fence : NULL;
- bool fence_emitted = false;
VkResult result;
- int ret;
+ uint32_t fence_idx = 0;
+
+ if (fence != VK_NULL_HANDLE) {
+ for (uint32_t i = 0; i < bindInfoCount; ++i)
+ if (radv_sparse_bind_has_effects(pBindInfo + i))
+ fence_idx = i;
+ } else
+ fence_idx = UINT32_MAX;
for (uint32_t i = 0; i < bindInfoCount; ++i) {
- struct radv_winsys_sem_info sem_info;
- for (uint32_t j = 0; j < pBindInfo[i].bufferBindCount; ++j) {
- radv_sparse_buffer_bind_memory(queue->device,
- pBindInfo[i].pBufferBinds + j);
- }
+ if (i != fence_idx && !radv_sparse_bind_has_effects(pBindInfo + i))
+ continue;
- for (uint32_t j = 0; j < pBindInfo[i].imageOpaqueBindCount; ++j) {
- radv_sparse_image_opaque_bind_memory(queue->device,
- pBindInfo[i].pImageOpaqueBinds + j);
- }
+ const VkTimelineSemaphoreSubmitInfo *timeline_info =
+ vk_find_struct_const(pBindInfo[i].pNext, TIMELINE_SEMAPHORE_SUBMIT_INFO);
+
+ VkResult result = radv_queue_submit(queue, &(struct radv_queue_submission) {
+ .buffer_binds = pBindInfo[i].pBufferBinds,
+ .buffer_bind_count = pBindInfo[i].bufferBindCount,
+ .image_opaque_binds = pBindInfo[i].pImageOpaqueBinds,
+ .image_opaque_bind_count = pBindInfo[i].imageOpaqueBindCount,
+ .wait_semaphores = pBindInfo[i].pWaitSemaphores,
+ .wait_semaphore_count = pBindInfo[i].waitSemaphoreCount,
+ .signal_semaphores = pBindInfo[i].pSignalSemaphores,
+ .signal_semaphore_count = pBindInfo[i].signalSemaphoreCount,
+ .fence = i == fence_idx ? fence : VK_NULL_HANDLE,
+ .wait_values = timeline_info ? timeline_info->pWaitSemaphoreValues : NULL,
+ .wait_value_count = timeline_info && timeline_info->pWaitSemaphoreValues ? timeline_info->waitSemaphoreValueCount : 0,
+ .signal_values = timeline_info ? timeline_info->pSignalSemaphoreValues : NULL,
+ .signal_value_count = timeline_info && timeline_info->pSignalSemaphoreValues ? timeline_info->signalSemaphoreValueCount : 0,
+ });
- VkResult result;
- result = radv_alloc_sem_info(queue->device->instance,
- &sem_info,
- pBindInfo[i].waitSemaphoreCount,
- pBindInfo[i].pWaitSemaphores,
- pBindInfo[i].signalSemaphoreCount,
- pBindInfo[i].pSignalSemaphores,
- _fence);
if (result != VK_SUCCESS)
return result;
-
- if (pBindInfo[i].waitSemaphoreCount || pBindInfo[i].signalSemaphoreCount) {
- ret = queue->device->ws->cs_submit(queue->hw_ctx, queue->queue_idx,
- &queue->device->empty_cs[queue->queue_family_index],
- 1, NULL, NULL,
- &sem_info, NULL,
- false, base_fence);
- if (ret) {
- radv_loge("failed to submit CS %d\n", i);
- abort();
- }
-
- fence_emitted = true;
- }
-
- radv_free_sem_info(&sem_info);
-
}
- if (fence) {
- if (!fence_emitted) {
- result = radv_signal_fence(queue, fence);
- if (result != VK_SUCCESS)
- return result;
- }
+ if (fence != VK_NULL_HANDLE && !bindInfoCount) {
+ result = radv_signal_fence(queue, fence);
+ if (result != VK_SUCCESS)
+ return result;
}
return VK_SUCCESS;
// Queue semaphore functions
+static void
+radv_create_timeline(struct radv_timeline *timeline, uint64_t value)
+{
+ timeline->highest_signaled = value;
+ timeline->highest_submitted = value;
+ list_inithead(&timeline->points);
+ list_inithead(&timeline->free_points);
+ list_inithead(&timeline->waiters);
+ pthread_mutex_init(&timeline->mutex, NULL);
+}
+
+static void
+radv_destroy_timeline(struct radv_device *device,
+ struct radv_timeline *timeline)
+{
+ list_for_each_entry_safe(struct radv_timeline_point, point,
+ &timeline->free_points, list) {
+ list_del(&point->list);
+ device->ws->destroy_syncobj(device->ws, point->syncobj);
+ free(point);
+ }
+ list_for_each_entry_safe(struct radv_timeline_point, point,
+ &timeline->points, list) {
+ list_del(&point->list);
+ device->ws->destroy_syncobj(device->ws, point->syncobj);
+ free(point);
+ }
+ pthread_mutex_destroy(&timeline->mutex);
+}
+
+static void
+radv_timeline_gc_locked(struct radv_device *device,
+ struct radv_timeline *timeline)
+{
+ list_for_each_entry_safe(struct radv_timeline_point, point,
+ &timeline->points, list) {
+ if (point->wait_count || point->value > timeline->highest_submitted)
+ return;
+
+ if (device->ws->wait_syncobj(device->ws, &point->syncobj, 1, true, 0)) {
+ timeline->highest_signaled = point->value;
+ list_del(&point->list);
+ list_add(&point->list, &timeline->free_points);
+ }
+ }
+}
+
+static struct radv_timeline_point *
+radv_timeline_find_point_at_least_locked(struct radv_device *device,
+ struct radv_timeline *timeline,
+ uint64_t p)
+{
+ radv_timeline_gc_locked(device, timeline);
+
+ if (p <= timeline->highest_signaled)
+ return NULL;
+
+ list_for_each_entry(struct radv_timeline_point, point,
+ &timeline->points, list) {
+ if (point->value >= p) {
+ ++point->wait_count;
+ return point;
+ }
+ }
+ return NULL;
+}
+
+static struct radv_timeline_point *
+radv_timeline_add_point_locked(struct radv_device *device,
+ struct radv_timeline *timeline,
+ uint64_t p)
+{
+ radv_timeline_gc_locked(device, timeline);
+
+ struct radv_timeline_point *ret = NULL;
+ struct radv_timeline_point *prev = NULL;
+
+ if (p <= timeline->highest_signaled)
+ return NULL;
+
+ list_for_each_entry(struct radv_timeline_point, point,
+ &timeline->points, list) {
+ if (point->value == p) {
+ return NULL;
+ }
+
+ if (point->value < p)
+ prev = point;
+ }
+
+ if (list_is_empty(&timeline->free_points)) {
+ ret = malloc(sizeof(struct radv_timeline_point));
+ device->ws->create_syncobj(device->ws, &ret->syncobj);
+ } else {
+ ret = list_first_entry(&timeline->free_points, struct radv_timeline_point, list);
+ list_del(&ret->list);
+
+ device->ws->reset_syncobj(device->ws, ret->syncobj);
+ }
+
+ ret->value = p;
+ ret->wait_count = 1;
+
+ if (prev) {
+ list_add(&ret->list, &prev->list);
+ } else {
+ list_addtail(&ret->list, &timeline->points);
+ }
+ return ret;
+}
+
+
+static VkResult
+radv_timeline_wait_locked(struct radv_device *device,
+ struct radv_timeline *timeline,
+ uint64_t value,
+ uint64_t abs_timeout)
+{
+ while(timeline->highest_submitted < value) {
+ struct timespec abstime;
+ timespec_from_nsec(&abstime, abs_timeout);
+
+ pthread_cond_timedwait(&device->timeline_cond, &timeline->mutex, &abstime);
+
+ if (radv_get_current_time() >= abs_timeout && timeline->highest_submitted < value)
+ return VK_TIMEOUT;
+ }
+
+ struct radv_timeline_point *point = radv_timeline_find_point_at_least_locked(device, timeline, value);
+ if (!point)
+ return VK_SUCCESS;
+
+ pthread_mutex_unlock(&timeline->mutex);
+
+ bool success = device->ws->wait_syncobj(device->ws, &point->syncobj, 1, true, abs_timeout);
+
+ pthread_mutex_lock(&timeline->mutex);
+ point->wait_count--;
+ return success ? VK_SUCCESS : VK_TIMEOUT;
+}
+
+static void
+radv_timeline_trigger_waiters_locked(struct radv_timeline *timeline,
+ struct list_head *processing_list)
+{
+ list_for_each_entry_safe(struct radv_timeline_waiter, waiter,
+ &timeline->waiters, list) {
+ if (waiter->value > timeline->highest_submitted)
+ continue;
+
+ if (p_atomic_dec_zero(&waiter->submission->submission_wait_count)) {
+ list_addtail(&waiter->submission->processing_list, processing_list);
+ }
+ list_del(&waiter->list);
+ }
+}
+
+static
+void radv_destroy_semaphore_part(struct radv_device *device,
+ struct radv_semaphore_part *part)
+{
+ switch(part->kind) {
+ case RADV_SEMAPHORE_NONE:
+ break;
+ case RADV_SEMAPHORE_WINSYS:
+ device->ws->destroy_sem(part->ws_sem);
+ break;
+ case RADV_SEMAPHORE_TIMELINE:
+ radv_destroy_timeline(device, &part->timeline);
+ break;
+ case RADV_SEMAPHORE_SYNCOBJ:
+ device->ws->destroy_syncobj(device->ws, part->syncobj);
+ break;
+ }
+ part->kind = RADV_SEMAPHORE_NONE;
+}
+
+static VkSemaphoreTypeKHR
+radv_get_semaphore_type(const void *pNext, uint64_t *initial_value)
+{
+ const VkSemaphoreTypeCreateInfo *type_info =
+ vk_find_struct_const(pNext, SEMAPHORE_TYPE_CREATE_INFO);
+
+ if (!type_info)
+ return VK_SEMAPHORE_TYPE_BINARY;
+
+ if (initial_value)
+ *initial_value = type_info->initialValue;
+ return type_info->semaphoreType;
+}
+
VkResult radv_CreateSemaphore(
VkDevice _device,
const VkSemaphoreCreateInfo* pCreateInfo,
vk_find_struct_const(pCreateInfo->pNext, EXPORT_SEMAPHORE_CREATE_INFO);
VkExternalSemaphoreHandleTypeFlags handleTypes =
export ? export->handleTypes : 0;
+ uint64_t initial_value = 0;
+ VkSemaphoreTypeKHR type = radv_get_semaphore_type(pCreateInfo->pNext, &initial_value);
struct radv_semaphore *sem = vk_alloc2(&device->alloc, pAllocator,
sizeof(*sem), 8,
if (!sem)
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
- sem->temp_syncobj = 0;
- /* create a syncobject if we are going to export this semaphore */
- if (device->always_use_syncobj || handleTypes) {
+ sem->temporary.kind = RADV_SEMAPHORE_NONE;
+ sem->permanent.kind = RADV_SEMAPHORE_NONE;
+
+ if (type == VK_SEMAPHORE_TYPE_TIMELINE) {
+ radv_create_timeline(&sem->permanent.timeline, initial_value);
+ sem->permanent.kind = RADV_SEMAPHORE_TIMELINE;
+ } else if (device->always_use_syncobj || handleTypes) {
assert (device->physical_device->rad_info.has_syncobj);
- int ret = device->ws->create_syncobj(device->ws, &sem->syncobj);
+ int ret = device->ws->create_syncobj(device->ws, &sem->permanent.syncobj);
if (ret) {
vk_free2(&device->alloc, pAllocator, sem);
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
- sem->sem = NULL;
+ sem->permanent.kind = RADV_SEMAPHORE_SYNCOBJ;
} else {
- sem->sem = device->ws->create_sem(device->ws);
- if (!sem->sem) {
+ sem->permanent.ws_sem = device->ws->create_sem(device->ws);
+ if (!sem->permanent.ws_sem) {
vk_free2(&device->alloc, pAllocator, sem);
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
- sem->syncobj = 0;
+ sem->permanent.kind = RADV_SEMAPHORE_WINSYS;
}
*pSemaphore = radv_semaphore_to_handle(sem);
if (!_semaphore)
return;
- if (sem->syncobj)
- device->ws->destroy_syncobj(device->ws, sem->syncobj);
- else
- device->ws->destroy_sem(sem->sem);
+ radv_destroy_semaphore_part(device, &sem->temporary);
+ radv_destroy_semaphore_part(device, &sem->permanent);
vk_free2(&device->alloc, pAllocator, sem);
}
+VkResult
+radv_GetSemaphoreCounterValue(VkDevice _device,
+ VkSemaphore _semaphore,
+ uint64_t* pValue)
+{
+ RADV_FROM_HANDLE(radv_device, device, _device);
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, _semaphore);
+
+ struct radv_semaphore_part *part =
+ semaphore->temporary.kind != RADV_SEMAPHORE_NONE ? &semaphore->temporary : &semaphore->permanent;
+
+ switch (part->kind) {
+ case RADV_SEMAPHORE_TIMELINE: {
+ pthread_mutex_lock(&part->timeline.mutex);
+ radv_timeline_gc_locked(device, &part->timeline);
+ *pValue = part->timeline.highest_signaled;
+ pthread_mutex_unlock(&part->timeline.mutex);
+ return VK_SUCCESS;
+ }
+ case RADV_SEMAPHORE_NONE:
+ case RADV_SEMAPHORE_SYNCOBJ:
+ case RADV_SEMAPHORE_WINSYS:
+ unreachable("Invalid semaphore type");
+ }
+ unreachable("Unhandled semaphore type");
+}
+
+
+static VkResult
+radv_wait_timelines(struct radv_device *device,
+ const VkSemaphoreWaitInfo* pWaitInfo,
+ uint64_t abs_timeout)
+{
+ if ((pWaitInfo->flags & VK_SEMAPHORE_WAIT_ANY_BIT_KHR) && pWaitInfo->semaphoreCount > 1) {
+ for (;;) {
+ for(uint32_t i = 0; i < pWaitInfo->semaphoreCount; ++i) {
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, pWaitInfo->pSemaphores[i]);
+ pthread_mutex_lock(&semaphore->permanent.timeline.mutex);
+ VkResult result = radv_timeline_wait_locked(device, &semaphore->permanent.timeline, pWaitInfo->pValues[i], 0);
+ pthread_mutex_unlock(&semaphore->permanent.timeline.mutex);
+
+ if (result == VK_SUCCESS)
+ return VK_SUCCESS;
+ }
+ if (radv_get_current_time() > abs_timeout)
+ return VK_TIMEOUT;
+ }
+ }
+
+ for(uint32_t i = 0; i < pWaitInfo->semaphoreCount; ++i) {
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, pWaitInfo->pSemaphores[i]);
+ pthread_mutex_lock(&semaphore->permanent.timeline.mutex);
+ VkResult result = radv_timeline_wait_locked(device, &semaphore->permanent.timeline, pWaitInfo->pValues[i], abs_timeout);
+ pthread_mutex_unlock(&semaphore->permanent.timeline.mutex);
+
+ if (result != VK_SUCCESS)
+ return result;
+ }
+ return VK_SUCCESS;
+}
+VkResult
+radv_WaitSemaphores(VkDevice _device,
+ const VkSemaphoreWaitInfo* pWaitInfo,
+ uint64_t timeout)
+{
+ RADV_FROM_HANDLE(radv_device, device, _device);
+ uint64_t abs_timeout = radv_get_absolute_timeout(timeout);
+ return radv_wait_timelines(device, pWaitInfo, abs_timeout);
+}
+
+VkResult
+radv_SignalSemaphore(VkDevice _device,
+ const VkSemaphoreSignalInfo* pSignalInfo)
+{
+ RADV_FROM_HANDLE(radv_device, device, _device);
+ RADV_FROM_HANDLE(radv_semaphore, semaphore, pSignalInfo->semaphore);
+
+ struct radv_semaphore_part *part =
+ semaphore->temporary.kind != RADV_SEMAPHORE_NONE ? &semaphore->temporary : &semaphore->permanent;
+
+ switch(part->kind) {
+ case RADV_SEMAPHORE_TIMELINE: {
+ pthread_mutex_lock(&part->timeline.mutex);
+ radv_timeline_gc_locked(device, &part->timeline);
+ part->timeline.highest_submitted = MAX2(part->timeline.highest_submitted, pSignalInfo->value);
+ part->timeline.highest_signaled = MAX2(part->timeline.highest_signaled, pSignalInfo->value);
+
+ struct list_head processing_list;
+ list_inithead(&processing_list);
+ radv_timeline_trigger_waiters_locked(&part->timeline, &processing_list);
+ pthread_mutex_unlock(&part->timeline.mutex);
+
+ return radv_process_submissions(&processing_list);
+ }
+ case RADV_SEMAPHORE_NONE:
+ case RADV_SEMAPHORE_SYNCOBJ:
+ case RADV_SEMAPHORE_WINSYS:
+ unreachable("Invalid semaphore type");
+ }
+ return VK_SUCCESS;
+}
+
+
+
VkResult radv_CreateEvent(
VkDevice _device,
const VkEventCreateInfo* pCreateInfo,
vk_free2(&device->alloc, pAllocator, buffer);
}
-VkDeviceAddress radv_GetBufferDeviceAddressEXT(
+VkDeviceAddress radv_GetBufferDeviceAddress(
VkDevice device,
- const VkBufferDeviceAddressInfoEXT* pInfo)
+ const VkBufferDeviceAddressInfo* pInfo)
{
RADV_FROM_HANDLE(radv_buffer, buffer, pInfo->buffer);
return radv_buffer_get_va(buffer->bo) + buffer->offset;
}
+uint64_t radv_GetBufferOpaqueCaptureAddress(VkDevice device,
+ const VkBufferDeviceAddressInfo* pInfo)
+{
+ return 0;
+}
+
+uint64_t radv_GetDeviceMemoryOpaqueCaptureAddress(VkDevice device,
+ const VkDeviceMemoryOpaqueCaptureAddressInfo* pInfo)
+{
+ return 0;
+}
+
static inline unsigned
si_tile_mode_index(const struct radv_image_plane *plane, unsigned level, bool stencil)
{
{
RADV_FROM_HANDLE(radv_device, device, _device);
struct radv_framebuffer *framebuffer;
- const VkFramebufferAttachmentsCreateInfoKHR *imageless_create_info =
+ const VkFramebufferAttachmentsCreateInfo *imageless_create_info =
vk_find_struct_const(pCreateInfo->pNext,
- FRAMEBUFFER_ATTACHMENTS_CREATE_INFO_KHR);
+ FRAMEBUFFER_ATTACHMENTS_CREATE_INFO);
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
framebuffer->layers = pCreateInfo->layers;
if (imageless_create_info) {
for (unsigned i = 0; i < imageless_create_info->attachmentImageInfoCount; ++i) {
- const VkFramebufferAttachmentImageInfoKHR *attachment =
+ const VkFramebufferAttachmentImageInfo *attachment =
imageless_create_info->pAttachmentImageInfos + i;
framebuffer->width = MIN2(framebuffer->width, attachment->width);
framebuffer->height = MIN2(framebuffer->height, attachment->height);
}
static unsigned
-radv_tex_filter_mode(VkSamplerReductionModeEXT mode)
+radv_tex_filter_mode(VkSamplerReductionMode mode)
{
switch (mode) {
case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT:
bool compat_mode = device->physical_device->rad_info.chip_class == GFX8 ||
device->physical_device->rad_info.chip_class == GFX9;
unsigned filter_mode = V_008F30_SQ_IMG_FILTER_MODE_BLEND;
+ unsigned depth_compare_func = V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
- const struct VkSamplerReductionModeCreateInfoEXT *sampler_reduction =
+ const struct VkSamplerReductionModeCreateInfo *sampler_reduction =
vk_find_struct_const(pCreateInfo->pNext,
- SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT);
+ SAMPLER_REDUCTION_MODE_CREATE_INFO);
if (sampler_reduction)
filter_mode = radv_tex_filter_mode(sampler_reduction->reductionMode);
+ if (pCreateInfo->compareEnable)
+ depth_compare_func = radv_tex_compare(pCreateInfo->compareOp);
+
sampler->state[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo->addressModeU)) |
S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo->addressModeV)) |
S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo->addressModeW)) |
S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
- S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo->compareOp)) |
+ S_008F30_DEPTH_COMPARE_FUNC(depth_compare_func) |
S_008F30_FORCE_UNNORMALIZED(pCreateInfo->unnormalizedCoordinates ? 1 : 0) |
S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) |
S_008F30_ANISO_BIAS(max_aniso_ratio) |
{
RADV_FROM_HANDLE(radv_device, device, _device);
RADV_FROM_HANDLE(radv_semaphore, sem, pImportSemaphoreFdInfo->semaphore);
- uint32_t *syncobj_dst = NULL;
+ VkResult result;
+ struct radv_semaphore_part *dst = NULL;
if (pImportSemaphoreFdInfo->flags & VK_SEMAPHORE_IMPORT_TEMPORARY_BIT) {
- syncobj_dst = &sem->temp_syncobj;
+ dst = &sem->temporary;
} else {
- syncobj_dst = &sem->syncobj;
+ dst = &sem->permanent;
}
+ uint32_t syncobj = dst->kind == RADV_SEMAPHORE_SYNCOBJ ? dst->syncobj : 0;
+
switch(pImportSemaphoreFdInfo->handleType) {
case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT:
- return radv_import_opaque_fd(device, pImportSemaphoreFdInfo->fd, syncobj_dst);
+ result = radv_import_opaque_fd(device, pImportSemaphoreFdInfo->fd, &syncobj);
+ break;
case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT:
- return radv_import_sync_fd(device, pImportSemaphoreFdInfo->fd, syncobj_dst);
+ result = radv_import_sync_fd(device, pImportSemaphoreFdInfo->fd, &syncobj);
+ break;
default:
unreachable("Unhandled semaphore handle type");
}
+
+ if (result == VK_SUCCESS) {
+ dst->syncobj = syncobj;
+ dst->kind = RADV_SEMAPHORE_SYNCOBJ;
+ }
+
+ return result;
}
VkResult radv_GetSemaphoreFdKHR(VkDevice _device,
int ret;
uint32_t syncobj_handle;
- if (sem->temp_syncobj)
- syncobj_handle = sem->temp_syncobj;
- else
- syncobj_handle = sem->syncobj;
+ if (sem->temporary.kind != RADV_SEMAPHORE_NONE) {
+ assert(sem->temporary.kind == RADV_SEMAPHORE_SYNCOBJ);
+ syncobj_handle = sem->temporary.syncobj;
+ } else {
+ assert(sem->permanent.kind == RADV_SEMAPHORE_SYNCOBJ);
+ syncobj_handle = sem->permanent.syncobj;
+ }
switch(pGetFdInfo->handleType) {
case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT:
case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT:
ret = device->ws->export_syncobj_to_sync_file(device->ws, syncobj_handle, pFd);
if (!ret) {
- if (sem->temp_syncobj) {
- close (sem->temp_syncobj);
- sem->temp_syncobj = 0;
+ if (sem->temporary.kind != RADV_SEMAPHORE_NONE) {
+ radv_destroy_semaphore_part(device, &sem->temporary);
} else {
device->ws->reset_syncobj(device->ws, syncobj_handle);
}
VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
{
RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
+ VkSemaphoreTypeKHR type = radv_get_semaphore_type(pExternalSemaphoreInfo->pNext, NULL);
+
+ if (type == VK_SEMAPHORE_TYPE_TIMELINE) {
+ pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
+ pExternalSemaphoreProperties->compatibleHandleTypes = 0;
+ pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
/* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
- if (pdevice->rad_info.has_syncobj_wait_for_submit &&
- (pExternalSemaphoreInfo->handleType == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT ||
- pExternalSemaphoreInfo->handleType == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT)) {
+ } else if (pdevice->rad_info.has_syncobj_wait_for_submit &&
+ (pExternalSemaphoreInfo->handleType == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT ||
+ pExternalSemaphoreInfo->handleType == VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT)) {
pExternalSemaphoreProperties->exportFromImportedHandleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT | VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
pExternalSemaphoreProperties->compatibleHandleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT | VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT;
pExternalSemaphoreProperties->externalSemaphoreFeatures = VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT |