*/
#include "util/mesa-sha1.h"
+#include "util/hash_table.h"
#include "util/debug.h"
#include "anv_private.h"
+static size_t
+anv_shader_bin_size(uint32_t prog_data_size, uint32_t key_size,
+ uint32_t surface_count, uint32_t sampler_count)
+{
+ const uint32_t binding_data_size =
+ (surface_count + sampler_count) * sizeof(struct anv_pipeline_binding);
+
+ return align_u32(sizeof(struct anv_shader_bin), 8) +
+ align_u32(prog_data_size, 8) +
+ align_u32(sizeof(uint32_t) + key_size, 8) +
+ align_u32(binding_data_size, 8);
+}
+
+struct anv_shader_bin *
+anv_shader_bin_create(struct anv_device *device,
+ const void *key_data, uint32_t key_size,
+ const void *kernel_data, uint32_t kernel_size,
+ const struct brw_stage_prog_data *prog_data,
+ uint32_t prog_data_size,
+ const struct anv_pipeline_bind_map *bind_map)
+{
+ const size_t size =
+ anv_shader_bin_size(prog_data_size, key_size,
+ bind_map->surface_count, bind_map->sampler_count);
+
+ struct anv_shader_bin *shader =
+ vk_alloc(&device->alloc, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+ if (!shader)
+ return NULL;
+
+ shader->ref_cnt = 1;
+
+ shader->kernel =
+ anv_state_pool_alloc(&device->instruction_state_pool, kernel_size, 64);
+ memcpy(shader->kernel.map, kernel_data, kernel_size);
+ shader->kernel_size = kernel_size;
+ shader->bind_map = *bind_map;
+ shader->prog_data_size = prog_data_size;
+
+ /* Now we fill out the floating data at the end */
+ void *data = shader;
+ data += align_u32(sizeof(struct anv_shader_bin), 8);
+
+ shader->prog_data = data;
+ memcpy(data, prog_data, prog_data_size);
+ data += align_u32(prog_data_size, 8);
+
+ shader->key = data;
+ struct anv_shader_bin_key *key = data;
+ key->size = key_size;
+ memcpy(key->data, key_data, key_size);
+ data += align_u32(sizeof(*key) + key_size, 8);
+
+ shader->bind_map.surface_to_descriptor = data;
+ memcpy(data, bind_map->surface_to_descriptor,
+ bind_map->surface_count * sizeof(struct anv_pipeline_binding));
+ data += bind_map->surface_count * sizeof(struct anv_pipeline_binding);
+
+ shader->bind_map.sampler_to_descriptor = data;
+ memcpy(data, bind_map->sampler_to_descriptor,
+ bind_map->sampler_count * sizeof(struct anv_pipeline_binding));
+
+ return shader;
+}
+
+void
+anv_shader_bin_destroy(struct anv_device *device,
+ struct anv_shader_bin *shader)
+{
+ assert(shader->ref_cnt == 0);
+ anv_state_pool_free(&device->instruction_state_pool, shader->kernel);
+ vk_free(&device->alloc, shader);
+}
+
+static size_t
+anv_shader_bin_data_size(const struct anv_shader_bin *shader)
+{
+ return anv_shader_bin_size(shader->prog_data_size,
+ shader->key->size,
+ shader->bind_map.surface_count,
+ shader->bind_map.sampler_count) +
+ align_u32(shader->kernel_size, 8);
+}
+
+static void
+anv_shader_bin_write_data(const struct anv_shader_bin *shader, void *data)
+{
+ size_t struct_size =
+ anv_shader_bin_size(shader->prog_data_size,
+ shader->key->size,
+ shader->bind_map.surface_count,
+ shader->bind_map.sampler_count);
+
+ memcpy(data, shader, struct_size);
+ data += struct_size;
+
+ memcpy(data, shader->kernel.map, shader->kernel_size);
+}
+
/* Remaining work:
*
* - Compact binding table layout so it's tight and not dependent on
* dual_src_blend.
*/
+static uint32_t
+shader_bin_key_hash_func(const void *void_key)
+{
+ const struct anv_shader_bin_key *key = void_key;
+ return _mesa_hash_data(key->data, key->size);
+}
+
+static bool
+shader_bin_key_compare_func(const void *void_a, const void *void_b)
+{
+ const struct anv_shader_bin_key *a = void_a, *b = void_b;
+ if (a->size != b->size)
+ return false;
+
+ return memcmp(a->data, b->data, a->size) == 0;
+}
+
void
anv_pipeline_cache_init(struct anv_pipeline_cache *cache,
- struct anv_device *device)
+ struct anv_device *device,
+ bool cache_enabled)
{
cache->device = device;
- anv_state_stream_init(&cache->program_stream,
- &device->instruction_block_pool);
pthread_mutex_init(&cache->mutex, NULL);
- cache->kernel_count = 0;
- cache->total_size = 0;
- cache->table_size = 1024;
- const size_t byte_size = cache->table_size * sizeof(cache->hash_table[0]);
- cache->hash_table = malloc(byte_size);
-
- /* We don't consider allocation failure fatal, we just start with a 0-sized
- * cache. */
- if (cache->hash_table == NULL)
- cache->table_size = 0;
- else
- memset(cache->hash_table, 0xff, byte_size);
+ if (cache_enabled) {
+ cache->cache = _mesa_hash_table_create(NULL, shader_bin_key_hash_func,
+ shader_bin_key_compare_func);
+ } else {
+ cache->cache = NULL;
+ }
}
void
anv_pipeline_cache_finish(struct anv_pipeline_cache *cache)
{
- anv_state_stream_finish(&cache->program_stream);
pthread_mutex_destroy(&cache->mutex);
- free(cache->hash_table);
-}
-
-struct cache_entry {
- unsigned char sha1[20];
- uint32_t prog_data_size;
- uint32_t kernel_size;
- char prog_data[0];
-
- /* kernel follows prog_data at next 64 byte aligned address */
-};
-static uint32_t
-entry_size(struct cache_entry *entry)
-{
- /* This returns the number of bytes needed to serialize an entry, which
- * doesn't include the alignment padding bytes.
- */
+ if (cache->cache) {
+ /* This is a bit unfortunate. In order to keep things from randomly
+ * going away, the shader cache has to hold a reference to all shader
+ * binaries it contains. We unref them when we destroy the cache.
+ */
+ struct hash_entry *entry;
+ hash_table_foreach(cache->cache, entry)
+ anv_shader_bin_unref(cache->device, entry->data);
- return sizeof(*entry) + entry->prog_data_size + entry->kernel_size;
+ _mesa_hash_table_destroy(cache->cache, NULL);
+ }
}
void
anv_hash_shader(unsigned char *hash, const void *key, size_t key_size,
struct anv_shader_module *module,
const char *entrypoint,
+ const struct anv_pipeline_layout *pipeline_layout,
const VkSpecializationInfo *spec_info)
{
struct mesa_sha1 *ctx;
_mesa_sha1_update(ctx, key, key_size);
_mesa_sha1_update(ctx, module->sha1, sizeof(module->sha1));
_mesa_sha1_update(ctx, entrypoint, strlen(entrypoint));
+ if (pipeline_layout) {
+ _mesa_sha1_update(ctx, pipeline_layout->sha1,
+ sizeof(pipeline_layout->sha1));
+ }
/* hash in shader stage, pipeline layout? */
if (spec_info) {
_mesa_sha1_update(ctx, spec_info->pMapEntries,
_mesa_sha1_final(ctx, hash);
}
-uint32_t
-anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
- const unsigned char *sha1, void *prog_data)
+static struct anv_shader_bin *
+anv_pipeline_cache_search_locked(struct anv_pipeline_cache *cache,
+ const void *key_data, uint32_t key_size)
{
- const uint32_t mask = cache->table_size - 1;
- const uint32_t start = (*(uint32_t *) sha1);
-
- for (uint32_t i = 0; i < cache->table_size; i++) {
- const uint32_t index = (start + i) & mask;
- const uint32_t offset = cache->hash_table[index];
-
- if (offset == ~0)
- return NO_KERNEL;
-
- struct cache_entry *entry =
- cache->program_stream.block_pool->map + offset;
- if (memcmp(entry->sha1, sha1, sizeof(entry->sha1)) == 0) {
- if (prog_data)
- memcpy(prog_data, entry->prog_data, entry->prog_data_size);
-
- const uint32_t preamble_size =
- align_u32(sizeof(*entry) + entry->prog_data_size, 64);
-
- return offset + preamble_size;
- }
- }
-
- return NO_KERNEL;
+ uint32_t vla[1 + DIV_ROUND_UP(key_size, sizeof(uint32_t))];
+ struct anv_shader_bin_key *key = (void *)vla;
+ key->size = key_size;
+ memcpy(key->data, key_data, key_size);
+
+ struct hash_entry *entry = _mesa_hash_table_search(cache->cache, key);
+ if (entry)
+ return entry->data;
+ else
+ return NULL;
}
-static void
-anv_pipeline_cache_add_entry(struct anv_pipeline_cache *cache,
- struct cache_entry *entry, uint32_t entry_offset)
+struct anv_shader_bin *
+anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
+ const void *key_data, uint32_t key_size)
{
- const uint32_t mask = cache->table_size - 1;
- const uint32_t start = (*(uint32_t *) entry->sha1);
+ if (!cache->cache)
+ return NULL;
- /* We'll always be able to insert when we get here. */
- assert(cache->kernel_count < cache->table_size / 2);
+ pthread_mutex_lock(&cache->mutex);
- for (uint32_t i = 0; i < cache->table_size; i++) {
- const uint32_t index = (start + i) & mask;
- if (cache->hash_table[index] == ~0) {
- cache->hash_table[index] = entry_offset;
- break;
- }
- }
+ struct anv_shader_bin *shader =
+ anv_pipeline_cache_search_locked(cache, key_data, key_size);
- cache->total_size += entry_size(entry);
- cache->kernel_count++;
+ pthread_mutex_unlock(&cache->mutex);
+
+ /* We increment refcount before handing it to the caller */
+ if (shader)
+ anv_shader_bin_ref(shader);
+
+ return shader;
}
-static VkResult
-anv_pipeline_cache_grow(struct anv_pipeline_cache *cache)
+static struct anv_shader_bin *
+anv_pipeline_cache_add_shader(struct anv_pipeline_cache *cache,
+ const void *key_data, uint32_t key_size,
+ const void *kernel_data, uint32_t kernel_size,
+ const void *prog_data, uint32_t prog_data_size,
+ const struct anv_pipeline_bind_map *bind_map)
{
- const uint32_t table_size = cache->table_size * 2;
- const uint32_t old_table_size = cache->table_size;
- const size_t byte_size = table_size * sizeof(cache->hash_table[0]);
- uint32_t *table;
- uint32_t *old_table = cache->hash_table;
-
- table = malloc(byte_size);
- if (table == NULL)
- return VK_ERROR_OUT_OF_HOST_MEMORY;
-
- cache->hash_table = table;
- cache->table_size = table_size;
- cache->kernel_count = 0;
- cache->total_size = 0;
-
- memset(cache->hash_table, 0xff, byte_size);
- for (uint32_t i = 0; i < old_table_size; i++) {
- const uint32_t offset = old_table[i];
- if (offset == ~0)
- continue;
+ struct anv_shader_bin *shader =
+ anv_pipeline_cache_search_locked(cache, key_data, key_size);
+ if (shader)
+ return shader;
- struct cache_entry *entry =
- cache->program_stream.block_pool->map + offset;
- anv_pipeline_cache_add_entry(cache, entry, offset);
- }
+ struct anv_shader_bin *bin =
+ anv_shader_bin_create(cache->device, key_data, key_size,
+ kernel_data, kernel_size,
+ prog_data, prog_data_size, bind_map);
+ if (!bin)
+ return NULL;
- free(old_table);
+ _mesa_hash_table_insert(cache->cache, bin->key, bin);
- return VK_SUCCESS;
+ return bin;
}
-uint32_t
+struct anv_shader_bin *
anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache,
- const unsigned char *sha1,
- const void *kernel, size_t kernel_size,
- const void *prog_data, size_t prog_data_size)
+ const void *key_data, uint32_t key_size,
+ const void *kernel_data, uint32_t kernel_size,
+ const struct brw_stage_prog_data *prog_data,
+ uint32_t prog_data_size,
+ const struct anv_pipeline_bind_map *bind_map)
{
- pthread_mutex_lock(&cache->mutex);
- struct cache_entry *entry;
-
- /* Meta pipelines don't have SPIR-V, so we can't hash them.
- * Consequentally, they just don't get cached.
- */
- const uint32_t preamble_size = sha1 ?
- align_u32(sizeof(*entry) + prog_data_size, 64) :
- 0;
-
- const uint32_t size = preamble_size + kernel_size;
-
- assert(size < cache->program_stream.block_pool->block_size);
- const struct anv_state state =
- anv_state_stream_alloc(&cache->program_stream, size, 64);
-
- if (sha1 && env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", false)) {
- assert(anv_pipeline_cache_search(cache, sha1, NULL) == NO_KERNEL);
- entry = state.map;
- memcpy(entry->sha1, sha1, sizeof(entry->sha1));
- entry->prog_data_size = prog_data_size;
- memcpy(entry->prog_data, prog_data, prog_data_size);
- entry->kernel_size = kernel_size;
-
- if (cache->kernel_count == cache->table_size / 2)
- anv_pipeline_cache_grow(cache);
-
- /* Failing to grow that hash table isn't fatal, but may mean we don't
- * have enough space to add this new kernel. Only add it if there's room.
- */
- if (cache->kernel_count < cache->table_size / 2)
- anv_pipeline_cache_add_entry(cache, entry, state.offset);
- }
+ if (cache->cache) {
+ pthread_mutex_lock(&cache->mutex);
- pthread_mutex_unlock(&cache->mutex);
+ struct anv_shader_bin *bin =
+ anv_pipeline_cache_add_shader(cache, key_data, key_size,
+ kernel_data, kernel_size,
+ prog_data, prog_data_size, bind_map);
- memcpy(state.map + preamble_size, kernel, kernel_size);
+ pthread_mutex_unlock(&cache->mutex);
- if (!cache->device->info.has_llc)
- anv_state_clflush(state);
+ /* We increment refcount before handing it to the caller */
+ anv_shader_bin_ref(bin);
- return state.offset + preamble_size;
+ return bin;
+ } else {
+ /* In this case, we're not caching it so the caller owns it entirely */
+ return anv_shader_bin_create(cache->device, key_data, key_size,
+ kernel_data, kernel_size,
+ prog_data, prog_data_size, bind_map);
+ }
}
struct cache_header {
struct cache_header header;
uint8_t uuid[VK_UUID_SIZE];
+ if (cache->cache == NULL)
+ return;
+
if (size < sizeof(header))
return;
memcpy(&header, data, sizeof(header));
const void *end = data + size;
const void *p = data + header.header_size;
- while (p < end) {
- /* The kernels aren't 64 byte aligned in the serialized format so
- * they're always right after the prog_data.
- */
- const struct cache_entry *entry = p;
- const void *kernel = &entry->prog_data[entry->prog_data_size];
+ /* Count is the total number of valid entries */
+ uint32_t count;
+ if (p + sizeof(count) >= end)
+ return;
+ memcpy(&count, p, sizeof(count));
+ p += align_u32(sizeof(count), 8);
+
+ for (uint32_t i = 0; i < count; i++) {
+ struct anv_shader_bin bin;
+ if (p + sizeof(bin) > end)
+ break;
+ memcpy(&bin, p, sizeof(bin));
+ p += align_u32(sizeof(struct anv_shader_bin), 8);
+
+ const void *prog_data = p;
+ p += align_u32(bin.prog_data_size, 8);
+
+ struct anv_shader_bin_key key;
+ if (p + sizeof(key) > end)
+ break;
+ memcpy(&key, p, sizeof(key));
+ const void *key_data = p + sizeof(key);
+ p += align_u32(sizeof(key) + key.size, 8);
+
+ /* We're going to memcpy this so getting rid of const is fine */
+ struct anv_pipeline_binding *bindings = (void *)p;
+ p += align_u32((bin.bind_map.surface_count + bin.bind_map.sampler_count) *
+ sizeof(struct anv_pipeline_binding), 8);
+ bin.bind_map.surface_to_descriptor = bindings;
+ bin.bind_map.sampler_to_descriptor = bindings + bin.bind_map.surface_count;
- anv_pipeline_cache_upload_kernel(cache, entry->sha1,
- kernel, entry->kernel_size,
- entry->prog_data, entry->prog_data_size);
- p = kernel + entry->kernel_size;
+ const void *kernel_data = p;
+ p += align_u32(bin.kernel_size, 8);
+
+ if (p > end)
+ break;
+
+ anv_pipeline_cache_add_shader(cache, key_data, key.size,
+ kernel_data, bin.kernel_size,
+ prog_data, bin.prog_data_size,
+ &bin.bind_map);
}
}
+static bool
+pipeline_cache_enabled()
+{
+ static int enabled = -1;
+ if (enabled < 0)
+ enabled = env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
+ return enabled;
+}
+
VkResult anv_CreatePipelineCache(
VkDevice _device,
const VkPipelineCacheCreateInfo* pCreateInfo,
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO);
assert(pCreateInfo->flags == 0);
- cache = anv_alloc2(&device->alloc, pAllocator,
+ cache = vk_alloc2(&device->alloc, pAllocator,
sizeof(*cache), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (cache == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- anv_pipeline_cache_init(cache, device);
+ anv_pipeline_cache_init(cache, device, pipeline_cache_enabled());
if (pCreateInfo->initialDataSize > 0)
anv_pipeline_cache_load(cache,
anv_pipeline_cache_finish(cache);
- anv_free2(&device->alloc, pAllocator, cache);
+ vk_free2(&device->alloc, pAllocator, cache);
}
VkResult anv_GetPipelineCacheData(
ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
struct cache_header *header;
- const size_t size = sizeof(*header) + cache->total_size;
-
if (pData == NULL) {
+ size_t size = align_u32(sizeof(*header), 8) +
+ align_u32(sizeof(uint32_t), 8);
+
+ if (cache->cache) {
+ struct hash_entry *entry;
+ hash_table_foreach(cache->cache, entry)
+ size += anv_shader_bin_data_size(entry->data);
+ }
+
*pDataSize = size;
return VK_SUCCESS;
}
header->vendor_id = 0x8086;
header->device_id = device->chipset_id;
anv_device_get_cache_uuid(header->uuid);
- p += header->header_size;
-
- struct cache_entry *entry;
- for (uint32_t i = 0; i < cache->table_size; i++) {
- if (cache->hash_table[i] == ~0)
- continue;
-
- entry = cache->program_stream.block_pool->map + cache->hash_table[i];
- if (end < p + entry_size(entry))
- break;
-
- memcpy(p, entry, sizeof(*entry) + entry->prog_data_size);
- p += sizeof(*entry) + entry->prog_data_size;
-
- void *kernel = (void *) entry +
- align_u32(sizeof(*entry) + entry->prog_data_size, 64);
-
- memcpy(p, kernel, entry->kernel_size);
- p += entry->kernel_size;
+ p += align_u32(header->header_size, 8);
+
+ uint32_t *count = p;
+ p += align_u32(sizeof(*count), 8);
+ *count = 0;
+
+ VkResult result = VK_SUCCESS;
+ if (cache->cache) {
+ struct hash_entry *entry;
+ hash_table_foreach(cache->cache, entry) {
+ struct anv_shader_bin *shader = entry->data;
+ size_t data_size = anv_shader_bin_data_size(entry->data);
+ if (p + data_size > end) {
+ result = VK_INCOMPLETE;
+ break;
+ }
+
+ anv_shader_bin_write_data(shader, p);
+ p += data_size;
+
+ (*count)++;
+ }
}
*pDataSize = p - pData;
- return VK_SUCCESS;
-}
-
-static void
-anv_pipeline_cache_merge(struct anv_pipeline_cache *dst,
- struct anv_pipeline_cache *src)
-{
- for (uint32_t i = 0; i < src->table_size; i++) {
- if (src->hash_table[i] == ~0)
- continue;
-
- struct cache_entry *entry =
- src->program_stream.block_pool->map + src->hash_table[i];
-
- if (anv_pipeline_cache_search(dst, entry->sha1, NULL) != NO_KERNEL)
- continue;
-
- const void *kernel = (void *) entry +
- align_u32(sizeof(*entry) + entry->prog_data_size, 64);
- anv_pipeline_cache_upload_kernel(dst, entry->sha1,
- kernel, entry->kernel_size,
- entry->prog_data, entry->prog_data_size);
- }
+ return result;
}
VkResult anv_MergePipelineCaches(
{
ANV_FROM_HANDLE(anv_pipeline_cache, dst, destCache);
+ if (!dst->cache)
+ return VK_SUCCESS;
+
for (uint32_t i = 0; i < srcCacheCount; i++) {
ANV_FROM_HANDLE(anv_pipeline_cache, src, pSrcCaches[i]);
+ if (!src->cache)
+ continue;
+
+ struct hash_entry *entry;
+ hash_table_foreach(src->cache, entry) {
+ struct anv_shader_bin *bin = entry->data;
+ if (_mesa_hash_table_search(dst->cache, bin->key))
+ continue;
- anv_pipeline_cache_merge(dst, src);
+ anv_shader_bin_ref(bin);
+ _mesa_hash_table_insert(dst->cache, bin->key, bin);
+ }
}
return VK_SUCCESS;
projects / mesa.git / blobdiff