#include "ac_nir_to_llvm.h"
-struct cache_entry_variant_info {
- struct ac_shader_variant_info variant_info;
- struct ac_shader_config config;
- uint32_t rsrc1, rsrc2;
-};
-
struct cache_entry {
union {
unsigned char sha1[20];
uint32_t sha1_dw[5];
};
- uint32_t code_sizes[MESA_SHADER_STAGES];
+ uint32_t binary_sizes[MESA_SHADER_STAGES];
struct radv_shader_variant *variants[MESA_SHADER_STAGES];
char code[0];
};
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))
cache->table_size = 0;
{
size_t ret = sizeof(*entry);
for (int i = 0; i < MESA_SHADER_STAGES; ++i)
- if (entry->code_sizes[i])
- ret += sizeof(struct cache_entry_variant_info) + entry->code_sizes[i];
+ if (entry->binary_sizes[i])
+ ret += entry->binary_sizes[i];
return ret;
}
return entry;
}
-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)
- cache = device->mem_cache;
-
- pthread_mutex_lock(&cache->mutex);
-
- entry = radv_pipeline_cache_search_unlocked(cache, sha1);
-
- if (!entry) {
- if (!device->physical_device->disk_cache ||
- (device->instance->debug_flags & RADV_DEBUG_NO_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;
- }
- }
-
- 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) {
- pthread_mutex_unlock(&cache->mutex);
- 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;
- } else if (entry->code_sizes[i]) {
- p += sizeof(struct cache_entry_variant_info) + entry->code_sizes[i];
- }
-
- }
-
- 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));
- pthread_mutex_unlock(&cache->mutex);
- 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);
}
+static bool
+radv_is_cache_disabled(struct radv_device *device)
+{
+ /* 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);
+}
+
+/*
+ * Secure compiles cannot open files so we get the parent process to load the
+ * cache entry for us.
+ */
+static struct cache_entry *
+radv_sc_read_from_disk_cache(struct radv_device *device, uint8_t *disk_sha1)
+{
+ struct cache_entry *entry;
+ unsigned process = device->sc_state->secure_compile_thread_counter;
+ enum radv_secure_compile_type sc_type = RADV_SC_TYPE_READ_DISK_CACHE;
+
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output,
+ &sc_type, sizeof(enum radv_secure_compile_type));
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output,
+ disk_sha1, sizeof(uint8_t) * 20);
+
+ uint8_t found_cache_entry;
+ if (!radv_sc_read(device->sc_state->secure_compile_processes[process].fd_secure_input,
+ &found_cache_entry, sizeof(uint8_t), true))
+ return NULL;
+
+ if (found_cache_entry) {
+ size_t entry_size;
+ if (!radv_sc_read(device->sc_state->secure_compile_processes[process].fd_secure_input,
+ &entry_size, sizeof(size_t), true))
+ return NULL;
+
+ entry = malloc(entry_size);
+ if (!radv_sc_read(device->sc_state->secure_compile_processes[process].fd_secure_input,
+ entry, entry_size, true))
+ return NULL;
+
+ return entry;
+ }
+
+ return NULL;
+}
+
+/*
+ * Secure compiles cannot open files so we get the parent process to write to
+ * the disk cache for us.
+ */
+static void
+radv_sc_write_to_disk_cache(struct radv_device *device, uint8_t *disk_sha1,
+ struct cache_entry *entry)
+{
+ unsigned process = device->sc_state->secure_compile_thread_counter;
+ enum radv_secure_compile_type sc_type = RADV_SC_TYPE_WRITE_DISK_CACHE;
+
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output,
+ &sc_type, sizeof(enum radv_secure_compile_type));
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output,
+ disk_sha1, sizeof(uint8_t) * 20);
+
+ uint32_t size = entry_size(entry);
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output,
+ &size, sizeof(uint32_t));
+ write(device->sc_state->secure_compile_processes[process].fd_secure_output,
+ entry, size);
+}
+
+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);
+
+ entry = radv_pipeline_cache_search_unlocked(cache, sha1);
+
+ if (!entry) {
+ *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);
+
+ if (radv_device_use_secure_compile(device->instance)) {
+ entry = radv_sc_read_from_disk_cache(device, disk_sha1);
+ } else {
+ 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;
+
+ if (!(device->instance->debug_flags & RADV_DEBUG_NO_MEMORY_CACHE) ||
+ cache != device->mem_cache)
+ radv_pipeline_cache_add_entry(cache, new_entry);
+ }
+ }
+
+ char *p = entry->code;
+ for(int i = 0; i < MESA_SHADER_STAGES; ++i) {
+ if (!entry->variants[i] && entry->binary_sizes[i]) {
+ struct radv_shader_binary *binary = calloc(1, entry->binary_sizes[i]);
+ memcpy(binary, p, entry->binary_sizes[i]);
+ p += entry->binary_sizes[i];
+
+ entry->variants[i] = radv_shader_variant_create(device, binary, false);
+ free(binary);
+ } else if (entry->binary_sizes[i]) {
+ p += entry->binary_sizes[i];
+ }
+
+ }
+
+ memcpy(variants, entry->variants, sizeof(entry->variants));
+
+ if (device->instance->debug_flags & RADV_DEBUG_NO_MEMORY_CACHE &&
+ cache == device->mem_cache)
+ vk_free(&cache->alloc, entry);
+ else {
+ for (int i = 0; i < MESA_SHADER_STAGES; ++i)
+ if (entry->variants[i])
+ p_atomic_inc(&entry->variants[i]->ref_count);
+ }
+
+ pthread_mutex_unlock(&cache->mutex);
+ return true;
+}
+
void
radv_pipeline_cache_insert_shaders(struct radv_device *device,
struct radv_pipeline_cache *cache,
const unsigned char *sha1,
struct radv_shader_variant **variants,
- const void *const *codes,
- const unsigned *code_sizes)
+ struct radv_shader_binary *const *binaries)
{
if (!cache)
cache = device->mem_cache;
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])
- size += sizeof(struct cache_entry_variant_info) + code_sizes[i];
+ size += binaries[i]->total_size;
entry = vk_alloc(&cache->alloc, size, 8,
memcpy(entry->sha1, sha1, 20);
char* p = entry->code;
- struct cache_entry_variant_info info;
for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
if (!variants[i])
continue;
- entry->code_sizes[i] = code_sizes[i];
-
- info.config = variants[i]->config;
- info.variant_info = variants[i]->info;
- info.rsrc1 = variants[i]->rsrc1;
- info.rsrc2 = variants[i]->rsrc2;
- memcpy(p, &info, sizeof(struct cache_entry_variant_info));
- p += sizeof(struct cache_entry_variant_info);
+ entry->binary_sizes[i] = binaries[i]->total_size;
- memcpy(p, codes[i], code_sizes[i]);
- p += code_sizes[i];
+ memcpy(p, binaries[i], binaries[i]->total_size);
+ p += binaries[i]->total_size;
}
/* Always add cache items to disk. This will allow collection of
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);
+
+ /* Write the cache item out to the parent of this forked
+ * process.
+ */
+ if (radv_device_use_secure_compile(device->instance)) {
+ radv_sc_write_to_disk_cache(device, disk_sha1, entry);
+ } else {
+ disk_cache_put(device->physical_device->disk_cache,
+ disk_sha1, entry, entry_size(entry),
+ NULL);
+ }
+ }
+
+ if (device->instance->debug_flags & RADV_DEBUG_NO_MEMORY_CACHE &&
+ cache == device->mem_cache) {
+ vk_free2(&cache->alloc, NULL, entry);
+ pthread_mutex_unlock(&cache->mutex);
+ return;
}
/* We delay setting the variant so we have reproducible disk cache
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;