#include "ac_nir_to_llvm.h"
struct cache_entry_variant_info {
- struct ac_shader_variant_info variant_info;
+ struct radv_shader_variant_info variant_info;
struct ac_shader_config config;
uint32_t rsrc1, rsrc2;
};
cache->hash_table = malloc(byte_size);
/* We don't consider allocation failure fatal, we just start with a 0-sized
- * cache. */
+ * cache. Disable caching when we want to keep shader debug info, since
+ * we don't get the debug info on cached shaders. */
if (cache->hash_table == NULL ||
- (device->instance->debug_flags & RADV_DEBUG_NO_CACHE))
+ (device->instance->debug_flags & RADV_DEBUG_NO_CACHE) ||
+ device->keep_shader_info)
cache->table_size = 0;
else
memset(cache->hash_table, 0, byte_size);
return ret;
}
-void
-radv_hash_shader(unsigned char *hash, struct radv_shader_module *module,
- const char *entrypoint,
- const VkSpecializationInfo *spec_info,
- const struct radv_pipeline_layout *layout,
- const struct ac_shader_variant_key *key,
- uint32_t flags)
-{
- struct mesa_sha1 ctx;
-
- _mesa_sha1_init(&ctx);
- if (key)
- _mesa_sha1_update(&ctx, key, sizeof(*key));
- _mesa_sha1_update(&ctx, module->sha1, sizeof(module->sha1));
- _mesa_sha1_update(&ctx, entrypoint, strlen(entrypoint));
- if (layout)
- _mesa_sha1_update(&ctx, layout->sha1, sizeof(layout->sha1));
- if (spec_info) {
- _mesa_sha1_update(&ctx, spec_info->pMapEntries,
- spec_info->mapEntryCount * sizeof spec_info->pMapEntries[0]);
- _mesa_sha1_update(&ctx, spec_info->pData, spec_info->dataSize);
- }
- _mesa_sha1_update(&ctx, &flags, 4);
- _mesa_sha1_final(&ctx, hash);
-}
-
void
radv_hash_shaders(unsigned char *hash,
const VkPipelineShaderStageCreateInfo **stages,
const struct radv_pipeline_layout *layout,
- const struct ac_shader_variant_key *keys,
+ const struct radv_pipeline_key *key,
uint32_t flags)
{
struct mesa_sha1 ctx;
_mesa_sha1_init(&ctx);
- if (keys)
- _mesa_sha1_update(&ctx, keys, sizeof(*keys) * MESA_SHADER_STAGES);
+ if (key)
+ _mesa_sha1_update(&ctx, key, sizeof(*key));
if (layout)
_mesa_sha1_update(&ctx, layout->sha1, sizeof(layout->sha1));
return entry;
}
-struct radv_shader_variant *
-radv_create_shader_variant_from_pipeline_cache(struct radv_device *device,
- struct radv_pipeline_cache *cache,
- const unsigned char *sha1)
-{
- struct cache_entry *entry = NULL;
-
- if (cache)
- entry = radv_pipeline_cache_search(cache, sha1);
- else
- entry = radv_pipeline_cache_search(device->mem_cache, sha1);
-
- if (!entry) {
- if (!device->physical_device->disk_cache)
- return NULL;
- uint8_t disk_sha1[20];
- disk_cache_compute_key(device->physical_device->disk_cache,
- sha1, 20, disk_sha1);
- entry = (struct cache_entry *)
- disk_cache_get(device->physical_device->disk_cache,
- disk_sha1, NULL);
- if (!entry)
- return NULL;
- }
-
- if (!entry->variants[0]) {
- struct radv_shader_variant *variant;
- char *p = entry->code;
- struct cache_entry_variant_info info;
-
- variant = calloc(1, sizeof(struct radv_shader_variant));
- if (!variant)
- return NULL;
-
- memcpy(&info, p, sizeof(struct cache_entry_variant_info));
- p += sizeof(struct cache_entry_variant_info);
-
- variant->code_size = entry->code_sizes[0];
- variant->config = info.config;
- variant->info = info.variant_info;
- variant->rsrc1 = info.rsrc1;
- variant->rsrc2 = info.rsrc2;
- variant->ref_count = 1;
-
- void *ptr = radv_alloc_shader_memory(device, variant);
- memcpy(ptr, p, entry->code_sizes[0]);
-
- entry->variants[0] = variant;
- }
-
- p_atomic_inc(&entry->variants[0]->ref_count);
- return entry->variants[0];
-}
-
-bool
-radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
- struct radv_pipeline_cache *cache,
- const unsigned char *sha1,
- struct radv_shader_variant **variants)
-{
- struct cache_entry *entry;
- if (cache)
- entry = radv_pipeline_cache_search(cache, sha1);
- else
- entry = radv_pipeline_cache_search(device->mem_cache, sha1);
-
- if (!entry) {
- if (!device->physical_device->disk_cache)
- return false;
-
- uint8_t disk_sha1[20];
- disk_cache_compute_key(device->physical_device->disk_cache,
- sha1, 20, disk_sha1);
- entry = (struct cache_entry *)
- disk_cache_get(device->physical_device->disk_cache,
- disk_sha1, NULL);
- if (!entry)
- return false;
- }
-
- char *p = entry->code;
- for(int i = 0; i < MESA_SHADER_STAGES; ++i) {
- if (!entry->variants[i] && entry->code_sizes[i]) {
- struct radv_shader_variant *variant;
- struct cache_entry_variant_info info;
-
- variant = calloc(1, sizeof(struct radv_shader_variant));
- if (!variant)
- return false;
-
- memcpy(&info, p, sizeof(struct cache_entry_variant_info));
- p += sizeof(struct cache_entry_variant_info);
-
- variant->config = info.config;
- variant->info = info.variant_info;
- variant->rsrc1 = info.rsrc1;
- variant->rsrc2 = info.rsrc2;
- variant->code_size = entry->code_sizes[i];
- variant->ref_count = 1;
-
- void *ptr = radv_alloc_shader_memory(device, variant);
- memcpy(ptr, p, entry->code_sizes[i]);
- p += entry->code_sizes[i];
-
- entry->variants[i] = variant;
- }
-
- }
-
- for (int i = 0; i < MESA_SHADER_STAGES; ++i)
- if (entry->variants[i])
- p_atomic_inc(&entry->variants[i]->ref_count);
-
- memcpy(variants, entry->variants, sizeof(entry->variants));
- return true;
-}
-
-
static void
radv_pipeline_cache_set_entry(struct radv_pipeline_cache *cache,
struct cache_entry *entry)
table = malloc(byte_size);
if (table == NULL)
- return VK_ERROR_OUT_OF_HOST_MEMORY;
+ return vk_error(cache->device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
cache->hash_table = table;
cache->table_size = table_size;
radv_pipeline_cache_set_entry(cache, entry);
}
-struct radv_shader_variant *
-radv_pipeline_cache_insert_shader(struct radv_device *device,
- struct radv_pipeline_cache *cache,
- const unsigned char *sha1,
- struct radv_shader_variant *variant,
- const void *code, unsigned code_size)
+static bool
+radv_is_cache_disabled(struct radv_device *device)
{
- if (!cache)
+ /* Pipeline caches can be disabled with RADV_DEBUG=nocache, with
+ * MESA_GLSL_CACHE_DISABLE=1, and when VK_AMD_shader_info is requested.
+ */
+ return (device->instance->debug_flags & RADV_DEBUG_NO_CACHE) ||
+ device->keep_shader_info;
+}
+
+bool
+radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
+ struct radv_pipeline_cache *cache,
+ const unsigned char *sha1,
+ struct radv_shader_variant **variants,
+ bool *found_in_application_cache)
+{
+ struct cache_entry *entry;
+
+ if (!cache) {
cache = device->mem_cache;
+ *found_in_application_cache = false;
+ }
pthread_mutex_lock(&cache->mutex);
- struct cache_entry *entry = radv_pipeline_cache_search_unlocked(cache, sha1);
- if (entry) {
- if (entry->variants[0]) {
- radv_shader_variant_destroy(cache->device, variant);
- variant = entry->variants[0];
- } else {
- entry->variants[0] = variant;
- }
- p_atomic_inc(&variant->ref_count);
- pthread_mutex_unlock(&cache->mutex);
- return variant;
- }
- entry = vk_alloc(&cache->alloc, sizeof(*entry) + sizeof(struct cache_entry_variant_info) + code_size, 8,
- VK_SYSTEM_ALLOCATION_SCOPE_CACHE);
+ entry = radv_pipeline_cache_search_unlocked(cache, sha1);
+
if (!entry) {
- pthread_mutex_unlock(&cache->mutex);
- return variant;
+ *found_in_application_cache = false;
+
+ /* Don't cache when we want debug info, since this isn't
+ * present in the cache.
+ */
+ if (radv_is_cache_disabled(device) || !device->physical_device->disk_cache) {
+ pthread_mutex_unlock(&cache->mutex);
+ return false;
+ }
+
+ uint8_t disk_sha1[20];
+ disk_cache_compute_key(device->physical_device->disk_cache,
+ sha1, 20, disk_sha1);
+ entry = (struct cache_entry *)
+ disk_cache_get(device->physical_device->disk_cache,
+ disk_sha1, NULL);
+ if (!entry) {
+ pthread_mutex_unlock(&cache->mutex);
+ return false;
+ } else {
+ size_t size = entry_size(entry);
+ struct cache_entry *new_entry = vk_alloc(&cache->alloc, size, 8,
+ VK_SYSTEM_ALLOCATION_SCOPE_CACHE);
+ if (!new_entry) {
+ free(entry);
+ pthread_mutex_unlock(&cache->mutex);
+ return false;
+ }
+
+ memcpy(new_entry, entry, entry_size(entry));
+ free(entry);
+ entry = new_entry;
+
+ radv_pipeline_cache_add_entry(cache, new_entry);
+ }
}
- memset(entry, 0, sizeof(*entry));
+ char *p = entry->code;
+ for(int i = 0; i < MESA_SHADER_STAGES; ++i) {
+ if (!entry->variants[i] && entry->code_sizes[i]) {
+ struct radv_shader_variant *variant;
+ struct cache_entry_variant_info info;
- char* p = entry->code;
- struct cache_entry_variant_info info;
+ variant = calloc(1, sizeof(struct radv_shader_variant));
+ if (!variant) {
+ pthread_mutex_unlock(&cache->mutex);
+ return false;
+ }
- info.config = variant->config;
- info.variant_info = variant->info;
- info.rsrc1 = variant->rsrc1;
- info.rsrc2 = variant->rsrc2;
- memcpy(p, &info, sizeof(struct cache_entry_variant_info));
- p += sizeof(struct cache_entry_variant_info);
+ memcpy(&info, p, sizeof(struct cache_entry_variant_info));
+ p += sizeof(struct cache_entry_variant_info);
- memcpy(entry->sha1, sha1, 20);
- memcpy(p, code, code_size);
+ variant->config = info.config;
+ variant->info = info.variant_info;
+ variant->rsrc1 = info.rsrc1;
+ variant->rsrc2 = info.rsrc2;
+ variant->code_size = entry->code_sizes[i];
+ variant->ref_count = 1;
- entry->code_sizes[0] = code_size;
+ void *ptr = radv_alloc_shader_memory(device, variant);
+ memcpy(ptr, p, entry->code_sizes[i]);
+ p += entry->code_sizes[i];
- /* Set variant to NULL so we have reproducible cache items */
- entry->variants[0] = NULL;
+ entry->variants[i] = variant;
+ } else if (entry->code_sizes[i]) {
+ p += sizeof(struct cache_entry_variant_info) + entry->code_sizes[i];
+ }
- /* Always add cache items to disk. This will allow collection of
- * compiled shaders by third parties such as steam, even if the app
- * implements its own pipeline cache.
- */
- if (device->physical_device->disk_cache) {
- uint8_t disk_sha1[20];
- disk_cache_compute_key(device->physical_device->disk_cache, sha1, 20,
- disk_sha1);
- disk_cache_put(device->physical_device->disk_cache,
- disk_sha1, entry, entry_size(entry), NULL);
}
- entry->variants[0] = variant;
- p_atomic_inc(&variant->ref_count);
-
- radv_pipeline_cache_add_entry(cache, entry);
+ for (int i = 0; i < MESA_SHADER_STAGES; ++i)
+ if (entry->variants[i])
+ p_atomic_inc(&entry->variants[i]->ref_count);
- cache->modified = true;
+ memcpy(variants, entry->variants, sizeof(entry->variants));
pthread_mutex_unlock(&cache->mutex);
- return variant;
+ return true;
}
void
pthread_mutex_unlock(&cache->mutex);
return;
}
+
+ /* Don't cache when we want debug info, since this isn't
+ * present in the cache.
+ */
+ if (radv_is_cache_disabled(device)) {
+ pthread_mutex_unlock(&cache->mutex);
+ return;
+ }
+
size_t size = sizeof(*entry);
for (int i = 0; i < MESA_SHADER_STAGES; ++i)
if (variants[i])
char* p = entry->code;
struct cache_entry_variant_info info;
+ memset(&info, 0, sizeof(info));
for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
if (!variants[i])
uint8_t uuid[VK_UUID_SIZE];
};
-void
+bool
radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
const void *data, size_t size)
{
struct cache_header header;
if (size < sizeof(header))
- return;
+ return false;
memcpy(&header, data, sizeof(header));
if (header.header_size < sizeof(header))
- return;
+ return false;
if (header.header_version != VK_PIPELINE_CACHE_HEADER_VERSION_ONE)
- return;
+ return false;
if (header.vendor_id != ATI_VENDOR_ID)
- return;
+ return false;
if (header.device_id != device->physical_device->rad_info.pci_id)
- return;
+ return false;
if (memcmp(header.uuid, device->physical_device->cache_uuid, VK_UUID_SIZE) != 0)
- return;
+ return false;
char *end = (void *) data + size;
char *p = (void *) data + header.header_size;
}
p += size;
}
+
+ return true;
}
VkResult radv_CreatePipelineCache(
sizeof(*cache), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (cache == NULL)
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
if (pAllocator)
cache->alloc = *pAllocator;
RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
struct cache_header *header;
VkResult result = VK_SUCCESS;
+
+ pthread_mutex_lock(&cache->mutex);
+
const size_t size = sizeof(*header) + cache->total_size;
if (pData == NULL) {
+ pthread_mutex_unlock(&cache->mutex);
*pDataSize = size;
return VK_SUCCESS;
}
if (*pDataSize < sizeof(*header)) {
+ pthread_mutex_unlock(&cache->mutex);
*pDataSize = 0;
return VK_INCOMPLETE;
}
}
*pDataSize = p - pData;
+ pthread_mutex_unlock(&cache->mutex);
return result;
}