v3dv: Implement VK_EXT_pipeline_creation_cache_control
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12381>
This commit is contained in:
@@ -534,7 +534,7 @@ Khronos extensions that are not part of any Vulkan version:
|
||||
VK_EXT_multi_draw DONE (anv, lvp, radv)
|
||||
VK_EXT_pci_bus_info DONE (anv, radv)
|
||||
VK_EXT_physical_device_drm DONE (anv, radv, v3dv)
|
||||
VK_EXT_pipeline_creation_cache_control DONE (anv, radv)
|
||||
VK_EXT_pipeline_creation_cache_control DONE (anv, radv, v3dv)
|
||||
VK_EXT_pipeline_creation_feedback DONE (anv, radv)
|
||||
VK_EXT_post_depth_coverage DONE (anv/gfx10+, lvp, radv)
|
||||
VK_EXT_private_data DONE (anv, lvp, radv, tu, v3dv)
|
||||
|
@@ -144,6 +144,7 @@ get_device_extensions(const struct v3dv_physical_device *device,
|
||||
.EXT_external_memory_dma_buf = true,
|
||||
.EXT_index_type_uint8 = true,
|
||||
.EXT_physical_device_drm = true,
|
||||
.EXT_pipeline_creation_cache_control = true,
|
||||
.EXT_private_data = true,
|
||||
};
|
||||
}
|
||||
@@ -1108,6 +1109,12 @@ v3dv_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
|
||||
break;
|
||||
}
|
||||
|
||||
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_CREATION_CACHE_CONTROL_FEATURES_EXT: {
|
||||
VkPhysicalDevicePipelineCreationCacheControlFeaturesEXT *features = (void *) ext;
|
||||
features->pipelineCreationCacheControl = true;
|
||||
break;
|
||||
}
|
||||
|
||||
/* Vulkan 1.1 */
|
||||
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES: {
|
||||
VkPhysicalDeviceVulkan11Features *features =
|
||||
@@ -1758,7 +1765,7 @@ v3dv_CreateDevice(VkPhysicalDevice physicalDevice,
|
||||
#endif
|
||||
init_device_meta(device);
|
||||
v3dv_bo_cache_init(device);
|
||||
v3dv_pipeline_cache_init(&device->default_pipeline_cache, device,
|
||||
v3dv_pipeline_cache_init(&device->default_pipeline_cache, device, 0,
|
||||
device->instance->default_pipeline_cache_enabled);
|
||||
device->default_attribute_float =
|
||||
v3dv_pipeline_create_default_attribute_values(device, NULL);
|
||||
|
@@ -2423,6 +2423,9 @@ pipeline_compile_graphics(struct v3dv_pipeline *pipeline,
|
||||
goto success;
|
||||
}
|
||||
|
||||
if (pCreateInfo->flags & VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT_EXT)
|
||||
return VK_PIPELINE_COMPILE_REQUIRED_EXT;
|
||||
|
||||
/* Otherwise we try to get the NIR shaders (either from the original SPIR-V
|
||||
* shader or the pipeline cache) and compile.
|
||||
*/
|
||||
@@ -2930,6 +2933,8 @@ graphics_pipeline_create(VkDevice _device,
|
||||
|
||||
if (result != VK_SUCCESS) {
|
||||
v3dv_destroy_pipeline(pipeline, device, pAllocator);
|
||||
if (result == VK_PIPELINE_COMPILE_REQUIRED_EXT)
|
||||
*pPipeline = VK_NULL_HANDLE;
|
||||
return result;
|
||||
}
|
||||
|
||||
@@ -2952,7 +2957,8 @@ v3dv_CreateGraphicsPipelines(VkDevice _device,
|
||||
if (unlikely(V3D_DEBUG & V3D_DEBUG_SHADERS))
|
||||
mtx_lock(&device->pdevice->mutex);
|
||||
|
||||
for (uint32_t i = 0; i < count; i++) {
|
||||
uint32_t i = 0;
|
||||
for (; i < count; i++) {
|
||||
VkResult local_result;
|
||||
|
||||
local_result = graphics_pipeline_create(_device,
|
||||
@@ -2964,9 +2970,16 @@ v3dv_CreateGraphicsPipelines(VkDevice _device,
|
||||
if (local_result != VK_SUCCESS) {
|
||||
result = local_result;
|
||||
pPipelines[i] = VK_NULL_HANDLE;
|
||||
|
||||
if (pCreateInfos[i].flags &
|
||||
VK_PIPELINE_CREATE_EARLY_RETURN_ON_FAILURE_BIT_EXT)
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
for (; i < count; i++)
|
||||
pPipelines[i] = VK_NULL_HANDLE;
|
||||
|
||||
if (unlikely(V3D_DEBUG & V3D_DEBUG_SHADERS))
|
||||
mtx_unlock(&device->pdevice->mutex);
|
||||
|
||||
@@ -3045,6 +3058,9 @@ pipeline_compile_compute(struct v3dv_pipeline *pipeline,
|
||||
goto success;
|
||||
}
|
||||
|
||||
if (info->flags & VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT_EXT)
|
||||
return VK_PIPELINE_COMPILE_REQUIRED_EXT;
|
||||
|
||||
pipeline->shared_data = v3dv_pipeline_shared_data_new_empty(pipeline_sha1,
|
||||
pipeline,
|
||||
false);
|
||||
@@ -3128,6 +3144,8 @@ compute_pipeline_create(VkDevice _device,
|
||||
pCreateInfo, pAllocator);
|
||||
if (result != VK_SUCCESS) {
|
||||
v3dv_destroy_pipeline(pipeline, device, pAllocator);
|
||||
if (result == VK_PIPELINE_COMPILE_REQUIRED_EXT)
|
||||
*pPipeline = VK_NULL_HANDLE;
|
||||
return result;
|
||||
}
|
||||
|
||||
@@ -3150,7 +3168,8 @@ v3dv_CreateComputePipelines(VkDevice _device,
|
||||
if (unlikely(V3D_DEBUG & V3D_DEBUG_SHADERS))
|
||||
mtx_lock(&device->pdevice->mutex);
|
||||
|
||||
for (uint32_t i = 0; i < createInfoCount; i++) {
|
||||
uint32_t i = 0;
|
||||
for (; i < createInfoCount; i++) {
|
||||
VkResult local_result;
|
||||
local_result = compute_pipeline_create(_device,
|
||||
pipelineCache,
|
||||
@@ -3161,9 +3180,16 @@ v3dv_CreateComputePipelines(VkDevice _device,
|
||||
if (local_result != VK_SUCCESS) {
|
||||
result = local_result;
|
||||
pPipelines[i] = VK_NULL_HANDLE;
|
||||
|
||||
if (pCreateInfos[i].flags &
|
||||
VK_PIPELINE_CREATE_EARLY_RETURN_ON_FAILURE_BIT_EXT)
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
for (; i < createInfoCount; i++)
|
||||
pPipelines[i] = VK_NULL_HANDLE;
|
||||
|
||||
if (unlikely(V3D_DEBUG & V3D_DEBUG_SHADERS))
|
||||
mtx_unlock(&device->pdevice->mutex);
|
||||
|
||||
|
@@ -63,6 +63,20 @@ cache_dump_stats(struct v3dv_pipeline_cache *cache)
|
||||
fprintf(stderr, " cache hit count: %d\n", cache->stats.hit);
|
||||
}
|
||||
|
||||
static void
|
||||
pipeline_cache_lock(struct v3dv_pipeline_cache *cache)
|
||||
{
|
||||
if (!cache->externally_synchronized)
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
}
|
||||
|
||||
static void
|
||||
pipeline_cache_unlock(struct v3dv_pipeline_cache *cache)
|
||||
{
|
||||
if (!cache->externally_synchronized)
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
}
|
||||
|
||||
void
|
||||
v3dv_pipeline_cache_upload_nir(struct v3dv_pipeline *pipeline,
|
||||
struct v3dv_pipeline_cache *cache,
|
||||
@@ -75,10 +89,10 @@ v3dv_pipeline_cache_upload_nir(struct v3dv_pipeline *pipeline,
|
||||
if (cache->nir_stats.count > V3DV_MAX_PIPELINE_CACHE_ENTRIES)
|
||||
return;
|
||||
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
pipeline_cache_lock(cache);
|
||||
struct hash_entry *entry =
|
||||
_mesa_hash_table_search(cache->nir_cache, sha1_key);
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
if (entry)
|
||||
return;
|
||||
|
||||
@@ -91,7 +105,7 @@ v3dv_pipeline_cache_upload_nir(struct v3dv_pipeline *pipeline,
|
||||
return;
|
||||
}
|
||||
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
pipeline_cache_lock(cache);
|
||||
/* Because ralloc isn't thread-safe, we have to do all this inside the
|
||||
* lock. We could unlock for the big memcpy but it's probably not worth
|
||||
* the hassle.
|
||||
@@ -99,7 +113,7 @@ v3dv_pipeline_cache_upload_nir(struct v3dv_pipeline *pipeline,
|
||||
entry = _mesa_hash_table_search(cache->nir_cache, sha1_key);
|
||||
if (entry) {
|
||||
blob_finish(&blob);
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
return;
|
||||
}
|
||||
|
||||
@@ -122,7 +136,7 @@ v3dv_pipeline_cache_upload_nir(struct v3dv_pipeline *pipeline,
|
||||
|
||||
_mesa_hash_table_insert(cache->nir_cache, snir->sha1_key, snir);
|
||||
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
}
|
||||
|
||||
nir_shader*
|
||||
@@ -143,12 +157,12 @@ v3dv_pipeline_cache_search_for_nir(struct v3dv_pipeline *pipeline,
|
||||
|
||||
const struct serialized_nir *snir = NULL;
|
||||
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
pipeline_cache_lock(cache);
|
||||
struct hash_entry *entry =
|
||||
_mesa_hash_table_search(cache->nir_cache, sha1_key);
|
||||
if (entry)
|
||||
snir = entry->data;
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
|
||||
if (snir) {
|
||||
struct blob_reader blob;
|
||||
@@ -185,6 +199,7 @@ v3dv_pipeline_cache_search_for_nir(struct v3dv_pipeline *pipeline,
|
||||
void
|
||||
v3dv_pipeline_cache_init(struct v3dv_pipeline_cache *cache,
|
||||
struct v3dv_device *device,
|
||||
VkPipelineCacheCreateFlags flags,
|
||||
bool cache_enabled)
|
||||
{
|
||||
cache->device = device;
|
||||
@@ -202,6 +217,9 @@ v3dv_pipeline_cache_init(struct v3dv_pipeline_cache *cache,
|
||||
cache->stats.miss = 0;
|
||||
cache->stats.hit = 0;
|
||||
cache->stats.count = 0;
|
||||
|
||||
cache->externally_synchronized = flags &
|
||||
VK_PIPELINE_CACHE_CREATE_EXTERNALLY_SYNCHRONIZED_BIT_EXT;
|
||||
} else {
|
||||
cache->nir_cache = NULL;
|
||||
cache->cache = NULL;
|
||||
@@ -241,7 +259,7 @@ v3dv_pipeline_cache_search_for_pipeline(struct v3dv_pipeline_cache *cache,
|
||||
fprintf(stderr, "pipeline cache %p, search pipeline with key %s\n", cache, sha1buf);
|
||||
}
|
||||
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
pipeline_cache_lock(cache);
|
||||
|
||||
struct hash_entry *entry =
|
||||
_mesa_hash_table_search(cache->cache, sha1_key);
|
||||
@@ -261,7 +279,7 @@ v3dv_pipeline_cache_search_for_pipeline(struct v3dv_pipeline_cache *cache,
|
||||
|
||||
v3dv_pipeline_shared_data_ref(cache_entry);
|
||||
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
|
||||
return cache_entry;
|
||||
}
|
||||
@@ -273,7 +291,7 @@ v3dv_pipeline_cache_search_for_pipeline(struct v3dv_pipeline_cache *cache,
|
||||
cache_dump_stats(cache);
|
||||
}
|
||||
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
|
||||
#ifdef ENABLE_SHADER_CACHE
|
||||
struct v3dv_device *device = cache->device;
|
||||
@@ -404,12 +422,12 @@ pipeline_cache_upload_shared_data(struct v3dv_pipeline_cache *cache,
|
||||
if (cache->stats.count > V3DV_MAX_PIPELINE_CACHE_ENTRIES)
|
||||
return;
|
||||
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
pipeline_cache_lock(cache);
|
||||
struct hash_entry *entry =
|
||||
_mesa_hash_table_search(cache->cache, shared_data->sha1_key);
|
||||
|
||||
if (entry) {
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
return;
|
||||
}
|
||||
|
||||
@@ -426,7 +444,7 @@ pipeline_cache_upload_shared_data(struct v3dv_pipeline_cache *cache,
|
||||
cache_dump_stats(cache);
|
||||
}
|
||||
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
|
||||
#ifdef ENABLE_SHADER_CACHE
|
||||
/* If we are being called from a on-disk-cache hit, we can skip writing to
|
||||
@@ -669,7 +687,6 @@ v3dv_CreatePipelineCache(VkDevice _device,
|
||||
struct v3dv_pipeline_cache *cache;
|
||||
|
||||
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
|
||||
assert(pCreateInfo->flags == 0);
|
||||
|
||||
cache = vk_object_zalloc(&device->vk, pAllocator,
|
||||
sizeof(*cache),
|
||||
@@ -678,7 +695,7 @@ v3dv_CreatePipelineCache(VkDevice _device,
|
||||
if (cache == NULL)
|
||||
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
|
||||
|
||||
v3dv_pipeline_cache_init(cache, device,
|
||||
v3dv_pipeline_cache_init(cache, device, pCreateInfo->flags,
|
||||
device->instance->pipeline_cache_enabled);
|
||||
|
||||
if (pCreateInfo->initialDataSize > 0) {
|
||||
@@ -918,7 +935,7 @@ v3dv_GetPipelineCacheData(VkDevice _device,
|
||||
struct v3dv_physical_device *pdevice = &device->instance->physicalDevice;
|
||||
VkResult result = VK_INCOMPLETE;
|
||||
|
||||
pthread_mutex_lock(&cache->mutex);
|
||||
pipeline_cache_lock(cache);
|
||||
|
||||
struct vk_pipeline_cache_header header = {
|
||||
.header_size = sizeof(struct vk_pipeline_cache_header),
|
||||
@@ -995,7 +1012,7 @@ v3dv_GetPipelineCacheData(VkDevice _device,
|
||||
done:
|
||||
blob_finish(&blob);
|
||||
|
||||
pthread_mutex_unlock(&cache->mutex);
|
||||
pipeline_cache_unlock(cache);
|
||||
|
||||
return result;
|
||||
}
|
||||
|
@@ -396,6 +396,9 @@ struct v3dv_pipeline_cache {
|
||||
|
||||
struct hash_table *cache;
|
||||
struct v3dv_pipeline_cache_stats stats;
|
||||
|
||||
/* For VK_EXT_pipeline_creation_cache_control. */
|
||||
bool externally_synchronized;
|
||||
};
|
||||
|
||||
struct v3dv_device {
|
||||
@@ -2017,6 +2020,7 @@ v3dv_immutable_samplers(const struct v3dv_descriptor_set_layout *set,
|
||||
|
||||
void v3dv_pipeline_cache_init(struct v3dv_pipeline_cache *cache,
|
||||
struct v3dv_device *device,
|
||||
VkPipelineCacheCreateFlags,
|
||||
bool cache_enabled);
|
||||
|
||||
void v3dv_pipeline_cache_finish(struct v3dv_pipeline_cache *cache);
|
||||
|
Reference in New Issue
Block a user