X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fintel%2Fvulkan%2Fanv_pipeline_cache.c;h=e1d48b879b07effede1be87a0355a76b0e36a120;hb=f118ca20758c85da1aaf1792e61aadb298b32a47;hp=f782d04fbe69ffda30edb32cc1abfba6e278b27c;hpb=b8ab78d1afd50cd10bf9be7046f27e4aa22d57ac;p=mesa.git diff --git a/src/intel/vulkan/anv_pipeline_cache.c b/src/intel/vulkan/anv_pipeline_cache.c index f782d04fbe6..e1d48b879b0 100644 --- a/src/intel/vulkan/anv_pipeline_cache.c +++ b/src/intel/vulkan/anv_pipeline_cache.c @@ -21,23 +21,31 @@ * IN THE SOFTWARE. */ -#include "compiler/blob.h" +#include "util/blob.h" #include "util/hash_table.h" #include "util/debug.h" +#include "util/disk_cache.h" +#include "util/mesa-sha1.h" +#include "nir/nir_serialize.h" #include "anv_private.h" +#include "nir/nir_xfb_info.h" 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 void *constant_data, uint32_t constant_data_size, const struct brw_stage_prog_data *prog_data_in, uint32_t prog_data_size, const void *prog_data_param_in, + const struct brw_compile_stats *stats, uint32_t num_stats, + const nir_xfb_info *xfb_info_in, const struct anv_pipeline_bind_map *bind_map) { struct anv_shader_bin *shader; struct anv_shader_bin_key *key; struct brw_stage_prog_data *prog_data; - const union gl_constant_value **prog_data_param; + uint32_t *prog_data_param; + nir_xfb_info *xfb_info; struct anv_pipeline_binding *surface_to_descriptor, *sampler_to_descriptor; ANV_MULTIALLOC(ma); @@ -45,6 +53,10 @@ anv_shader_bin_create(struct anv_device *device, anv_multialloc_add_size(&ma, &key, sizeof(*key) + key_size); anv_multialloc_add_size(&ma, &prog_data, prog_data_size); anv_multialloc_add(&ma, &prog_data_param, prog_data_in->nr_params); + if (xfb_info_in) { + uint32_t xfb_info_size = nir_xfb_info_size(xfb_info_in->output_count); + anv_multialloc_add_size(&ma, &xfb_info, xfb_info_size); + } anv_multialloc_add(&ma, &surface_to_descriptor, bind_map->surface_count); anv_multialloc_add(&ma, &sampler_to_descriptor, @@ -65,6 +77,16 @@ anv_shader_bin_create(struct anv_device *device, memcpy(shader->kernel.map, kernel_data, kernel_size); shader->kernel_size = kernel_size; + if (constant_data_size) { + shader->constant_data = + anv_state_pool_alloc(&device->dynamic_state_pool, + constant_data_size, 32); + memcpy(shader->constant_data.map, constant_data, constant_data_size); + } else { + shader->constant_data = ANV_STATE_NULL; + } + shader->constant_data_size = constant_data_size; + memcpy(prog_data, prog_data_in, prog_data_size); memcpy(prog_data_param, prog_data_param_in, prog_data->nr_params * sizeof(*prog_data_param)); @@ -72,6 +94,19 @@ anv_shader_bin_create(struct anv_device *device, shader->prog_data = prog_data; shader->prog_data_size = prog_data_size; + assert(num_stats <= ARRAY_SIZE(shader->stats)); + typed_memcpy(shader->stats, stats, num_stats); + shader->num_stats = num_stats; + + if (xfb_info_in) { + *xfb_info = *xfb_info_in; + typed_memcpy(xfb_info->outputs, xfb_info_in->outputs, + xfb_info_in->output_count); + shader->xfb_info = xfb_info; + } else { + shader->xfb_info = NULL; + } + shader->bind_map = *bind_map; typed_memcpy(surface_to_descriptor, bind_map->surface_to_descriptor, bind_map->surface_count); @@ -89,6 +124,7 @@ anv_shader_bin_destroy(struct anv_device *device, { assert(shader->ref_cnt == 0); anv_state_pool_free(&device->instruction_state_pool, shader->kernel); + anv_state_pool_free(&device->dynamic_state_pool, shader->constant_data); vk_free(&device->alloc, shader); } @@ -96,31 +132,45 @@ static bool anv_shader_bin_write_to_blob(const struct anv_shader_bin *shader, struct blob *blob) { - bool ok; - - ok = blob_write_uint32(blob, shader->key->size); - ok = blob_write_bytes(blob, shader->key->data, shader->key->size); - - ok = blob_write_uint32(blob, shader->kernel_size); - ok = blob_write_bytes(blob, shader->kernel.map, shader->kernel_size); - - ok = blob_write_uint32(blob, shader->prog_data_size); - ok = blob_write_bytes(blob, shader->prog_data, shader->prog_data_size); - ok = blob_write_bytes(blob, shader->prog_data->param, - shader->prog_data->nr_params * - sizeof(*shader->prog_data->param)); - - ok = blob_write_uint32(blob, shader->bind_map.surface_count); - ok = blob_write_uint32(blob, shader->bind_map.sampler_count); - ok = blob_write_uint32(blob, shader->bind_map.image_count); - ok = blob_write_bytes(blob, shader->bind_map.surface_to_descriptor, - shader->bind_map.surface_count * - sizeof(*shader->bind_map.surface_to_descriptor)); - ok = blob_write_bytes(blob, shader->bind_map.sampler_to_descriptor, - shader->bind_map.sampler_count * - sizeof(*shader->bind_map.sampler_to_descriptor)); - - return ok; + blob_write_uint32(blob, shader->key->size); + blob_write_bytes(blob, shader->key->data, shader->key->size); + + blob_write_uint32(blob, shader->kernel_size); + blob_write_bytes(blob, shader->kernel.map, shader->kernel_size); + + blob_write_uint32(blob, shader->constant_data_size); + blob_write_bytes(blob, shader->constant_data.map, + shader->constant_data_size); + + blob_write_uint32(blob, shader->prog_data_size); + blob_write_bytes(blob, shader->prog_data, shader->prog_data_size); + blob_write_bytes(blob, shader->prog_data->param, + shader->prog_data->nr_params * + sizeof(*shader->prog_data->param)); + + blob_write_uint32(blob, shader->num_stats); + blob_write_bytes(blob, shader->stats, + shader->num_stats * sizeof(shader->stats[0])); + + if (shader->xfb_info) { + uint32_t xfb_info_size = + nir_xfb_info_size(shader->xfb_info->output_count); + blob_write_uint32(blob, xfb_info_size); + blob_write_bytes(blob, shader->xfb_info, xfb_info_size); + } else { + blob_write_uint32(blob, 0); + } + + blob_write_uint32(blob, shader->bind_map.surface_count); + blob_write_uint32(blob, shader->bind_map.sampler_count); + blob_write_bytes(blob, shader->bind_map.surface_to_descriptor, + shader->bind_map.surface_count * + sizeof(*shader->bind_map.surface_to_descriptor)); + blob_write_bytes(blob, shader->bind_map.sampler_to_descriptor, + shader->bind_map.sampler_count * + sizeof(*shader->bind_map.sampler_to_descriptor)); + + return !blob->out_of_memory; } static struct anv_shader_bin * @@ -133,6 +183,9 @@ anv_shader_bin_create_from_blob(struct anv_device *device, uint32_t kernel_size = blob_read_uint32(blob); const void *kernel_data = blob_read_bytes(blob, kernel_size); + uint32_t constant_data_size = blob_read_uint32(blob); + const void *constant_data = blob_read_bytes(blob, constant_data_size); + uint32_t prog_data_size = blob_read_uint32(blob); const struct brw_stage_prog_data *prog_data = blob_read_bytes(blob, prog_data_size); @@ -141,10 +194,18 @@ anv_shader_bin_create_from_blob(struct anv_device *device, const void *prog_data_param = blob_read_bytes(blob, prog_data->nr_params * sizeof(*prog_data->param)); + uint32_t num_stats = blob_read_uint32(blob); + const struct brw_compile_stats *stats = + blob_read_bytes(blob, num_stats * sizeof(stats[0])); + + const nir_xfb_info *xfb_info = NULL; + uint32_t xfb_size = blob_read_uint32(blob); + if (xfb_size) + xfb_info = blob_read_bytes(blob, xfb_size); + struct anv_pipeline_bind_map bind_map; bind_map.surface_count = blob_read_uint32(blob); bind_map.sampler_count = blob_read_uint32(blob); - bind_map.image_count = blob_read_uint32(blob); bind_map.surface_to_descriptor = (void *) blob_read_bytes(blob, bind_map.surface_count * sizeof(*bind_map.surface_to_descriptor)); @@ -158,8 +219,9 @@ anv_shader_bin_create_from_blob(struct anv_device *device, return anv_shader_bin_create(device, key_data, key_size, kernel_data, kernel_size, + constant_data, constant_data_size, prog_data, prog_data_size, prog_data_param, - &bind_map); + stats, num_stats, xfb_info, &bind_map); } /* Remaining work: @@ -169,9 +231,7 @@ anv_shader_bin_create_from_blob(struct anv_device *device, * * - Review prog_data struct for size and cacheability: struct * brw_stage_prog_data has binding_table which uses a lot of uint32_t for 8 - * bit quantities etc; param, pull_param, and image_params are pointers, we - * just need the compation map. use bit fields for all bools, eg - * dual_src_blend. + * bit quantities etc; use bit fields for all bools, eg dual_src_blend. */ static uint32_t @@ -191,6 +251,18 @@ shader_bin_key_compare_func(const void *void_a, const void *void_b) return memcmp(a->data, b->data, a->size) == 0; } +static uint32_t +sha1_hash_func(const void *sha1) +{ + return _mesa_hash_data(sha1, 20); +} + +static bool +sha1_compare_func(const void *sha1_a, const void *sha1_b) +{ + return memcmp(sha1_a, sha1_b, 20) == 0; +} + void anv_pipeline_cache_init(struct anv_pipeline_cache *cache, struct anv_device *device, @@ -202,8 +274,11 @@ anv_pipeline_cache_init(struct anv_pipeline_cache *cache, if (cache_enabled) { cache->cache = _mesa_hash_table_create(NULL, shader_bin_key_hash_func, shader_bin_key_compare_func); + cache->nir_cache = _mesa_hash_table_create(NULL, sha1_hash_func, + sha1_compare_func); } else { cache->cache = NULL; + cache->nir_cache = NULL; } } @@ -217,12 +292,18 @@ anv_pipeline_cache_finish(struct anv_pipeline_cache *cache) * 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); _mesa_hash_table_destroy(cache->cache, NULL); } + + if (cache->nir_cache) { + hash_table_foreach(cache->nir_cache, entry) + ralloc_free(entry->data); + + _mesa_hash_table_destroy(cache->nir_cache, NULL); + } } static struct anv_shader_bin * @@ -262,14 +343,39 @@ anv_pipeline_cache_search(struct anv_pipeline_cache *cache, return shader; } +static void +anv_pipeline_cache_add_shader_bin(struct anv_pipeline_cache *cache, + struct anv_shader_bin *bin) +{ + if (!cache->cache) + return; + + pthread_mutex_lock(&cache->mutex); + + struct hash_entry *entry = _mesa_hash_table_search(cache->cache, bin->key); + if (entry == NULL) { + /* Take a reference for the cache */ + anv_shader_bin_ref(bin); + _mesa_hash_table_insert(cache->cache, bin->key, bin); + } + + pthread_mutex_unlock(&cache->mutex); +} + 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 struct brw_stage_prog_data *prog_data, - uint32_t prog_data_size, - const void *prog_data_param, - const struct anv_pipeline_bind_map *bind_map) +anv_pipeline_cache_add_shader_locked(struct anv_pipeline_cache *cache, + const void *key_data, uint32_t key_size, + const void *kernel_data, + uint32_t kernel_size, + const void *constant_data, + uint32_t constant_data_size, + const struct brw_stage_prog_data *prog_data, + uint32_t prog_data_size, + const void *prog_data_param, + const struct brw_compile_stats *stats, + uint32_t num_stats, + const nir_xfb_info *xfb_info, + const struct anv_pipeline_bind_map *bind_map) { struct anv_shader_bin *shader = anv_pipeline_cache_search_locked(cache, key_data, key_size); @@ -279,8 +385,9 @@ anv_pipeline_cache_add_shader(struct anv_pipeline_cache *cache, struct anv_shader_bin *bin = anv_shader_bin_create(cache->device, key_data, key_size, kernel_data, kernel_size, + constant_data, constant_data_size, prog_data, prog_data_size, prog_data_param, - bind_map); + stats, num_stats, xfb_info, bind_map); if (!bin) return NULL; @@ -293,18 +400,26 @@ struct anv_shader_bin * anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache, const void *key_data, uint32_t key_size, const void *kernel_data, uint32_t kernel_size, + const void *constant_data, + uint32_t constant_data_size, const struct brw_stage_prog_data *prog_data, uint32_t prog_data_size, + const struct brw_compile_stats *stats, + uint32_t num_stats, + const nir_xfb_info *xfb_info, const struct anv_pipeline_bind_map *bind_map) { if (cache->cache) { pthread_mutex_lock(&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, - prog_data->param, bind_map); + anv_pipeline_cache_add_shader_locked(cache, key_data, key_size, + kernel_data, kernel_size, + constant_data, constant_data_size, + prog_data, prog_data_size, + prog_data->param, + stats, num_stats, + xfb_info, bind_map); pthread_mutex_unlock(&cache->mutex); @@ -317,8 +432,11 @@ anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache, /* 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, + constant_data, constant_data_size, prog_data, prog_data_size, - prog_data->param, bind_map); + prog_data->param, + stats, num_stats, + xfb_info, bind_map); } } @@ -369,15 +487,6 @@ anv_pipeline_cache_load(struct anv_pipeline_cache *cache, } } -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, @@ -396,7 +505,8 @@ VkResult anv_CreatePipelineCache( if (cache == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - anv_pipeline_cache_init(cache, device, pipeline_cache_enabled()); + anv_pipeline_cache_init(cache, device, + device->instance->pipeline_cache_enabled); if (pCreateInfo->initialDataSize > 0) anv_pipeline_cache_load(cache, @@ -451,7 +561,7 @@ VkResult anv_GetPipelineCacheData( blob_write_bytes(&blob, &header, sizeof(header)); uint32_t count = 0; - ssize_t count_offset = blob_reserve_uint32(&blob); + intptr_t count_offset = blob_reserve_uint32(&blob); if (count_offset < 0) { *pDataSize = 0; blob_finish(&blob); @@ -460,7 +570,6 @@ VkResult anv_GetPipelineCacheData( VkResult result = VK_SUCCESS; if (cache->cache) { - struct hash_entry *entry; hash_table_foreach(cache->cache, entry) { struct anv_shader_bin *shader = entry->data; @@ -501,7 +610,6 @@ VkResult anv_MergePipelineCaches( if (!src->cache) continue; - struct hash_entry *entry; hash_table_foreach(src->cache, entry) { struct anv_shader_bin *bin = entry->data; assert(bin); @@ -516,3 +624,189 @@ VkResult anv_MergePipelineCaches( return VK_SUCCESS; } + +struct anv_shader_bin * +anv_device_search_for_kernel(struct anv_device *device, + struct anv_pipeline_cache *cache, + const void *key_data, uint32_t key_size, + bool *user_cache_hit) +{ + struct anv_shader_bin *bin; + + *user_cache_hit = false; + + if (cache) { + bin = anv_pipeline_cache_search(cache, key_data, key_size); + if (bin) { + *user_cache_hit = cache != &device->default_pipeline_cache; + return bin; + } + } + +#ifdef ENABLE_SHADER_CACHE + struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache; + if (disk_cache && device->instance->pipeline_cache_enabled) { + cache_key cache_key; + disk_cache_compute_key(disk_cache, key_data, key_size, cache_key); + + size_t buffer_size; + uint8_t *buffer = disk_cache_get(disk_cache, cache_key, &buffer_size); + if (buffer) { + struct blob_reader blob; + blob_reader_init(&blob, buffer, buffer_size); + bin = anv_shader_bin_create_from_blob(device, &blob); + free(buffer); + + if (bin) { + if (cache) + anv_pipeline_cache_add_shader_bin(cache, bin); + return bin; + } + } + } +#endif + + return NULL; +} + +struct anv_shader_bin * +anv_device_upload_kernel(struct anv_device *device, + struct anv_pipeline_cache *cache, + const void *key_data, uint32_t key_size, + const void *kernel_data, uint32_t kernel_size, + const void *constant_data, + uint32_t constant_data_size, + const struct brw_stage_prog_data *prog_data, + uint32_t prog_data_size, + const struct brw_compile_stats *stats, + uint32_t num_stats, + const nir_xfb_info *xfb_info, + const struct anv_pipeline_bind_map *bind_map) +{ + struct anv_shader_bin *bin; + if (cache) { + bin = anv_pipeline_cache_upload_kernel(cache, key_data, key_size, + kernel_data, kernel_size, + constant_data, constant_data_size, + prog_data, prog_data_size, + stats, num_stats, + xfb_info, bind_map); + } else { + bin = anv_shader_bin_create(device, key_data, key_size, + kernel_data, kernel_size, + constant_data, constant_data_size, + prog_data, prog_data_size, + prog_data->param, + stats, num_stats, + xfb_info, bind_map); + } + + if (bin == NULL) + return NULL; + +#ifdef ENABLE_SHADER_CACHE + struct disk_cache *disk_cache = device->instance->physicalDevice.disk_cache; + if (disk_cache) { + struct blob binary; + blob_init(&binary); + if (anv_shader_bin_write_to_blob(bin, &binary)) { + cache_key cache_key; + disk_cache_compute_key(disk_cache, key_data, key_size, cache_key); + + disk_cache_put(disk_cache, cache_key, binary.data, binary.size, NULL); + } + + blob_finish(&binary); + } +#endif + + return bin; +} + +struct serialized_nir { + unsigned char sha1_key[20]; + size_t size; + char data[0]; +}; + +struct nir_shader * +anv_device_search_for_nir(struct anv_device *device, + struct anv_pipeline_cache *cache, + const nir_shader_compiler_options *nir_options, + unsigned char sha1_key[20], + void *mem_ctx) +{ + if (cache && cache->nir_cache) { + const struct serialized_nir *snir = NULL; + + pthread_mutex_lock(&cache->mutex); + struct hash_entry *entry = + _mesa_hash_table_search(cache->nir_cache, sha1_key); + if (entry) + snir = entry->data; + pthread_mutex_unlock(&cache->mutex); + + if (snir) { + struct blob_reader blob; + blob_reader_init(&blob, snir->data, snir->size); + + nir_shader *nir = nir_deserialize(mem_ctx, nir_options, &blob); + if (blob.overrun) { + ralloc_free(nir); + } else { + return nir; + } + } + } + + return NULL; +} + +void +anv_device_upload_nir(struct anv_device *device, + struct anv_pipeline_cache *cache, + const struct nir_shader *nir, + unsigned char sha1_key[20]) +{ + if (cache && cache->nir_cache) { + pthread_mutex_lock(&cache->mutex); + struct hash_entry *entry = + _mesa_hash_table_search(cache->nir_cache, sha1_key); + pthread_mutex_unlock(&cache->mutex); + if (entry) + return; + + struct blob blob; + blob_init(&blob); + + nir_serialize(&blob, nir, false); + if (blob.out_of_memory) { + blob_finish(&blob); + return; + } + + pthread_mutex_lock(&cache->mutex); + /* 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. + */ + entry = _mesa_hash_table_search(cache->nir_cache, sha1_key); + if (entry) { + blob_finish(&blob); + pthread_mutex_unlock(&cache->mutex); + return; + } + + struct serialized_nir *snir = + ralloc_size(cache->nir_cache, sizeof(*snir) + blob.size); + memcpy(snir->sha1_key, sha1_key, 20); + snir->size = blob.size; + memcpy(snir->data, blob.data, blob.size); + + blob_finish(&blob); + + _mesa_hash_table_insert(cache->nir_cache, snir->sha1_key, snir); + + pthread_mutex_unlock(&cache->mutex); + } +}