radv: Pre-compute descriptor set layout hash.

While analyzing cache loading performance, hashing the pipeline layout was
surprisingly consuming around 4% of time, sometimes close to the cost of
hashing shader modules.

Turns out we were hashing the pipeline layout on every pipeline creation.
Considering that pipeline layouts are usually deduplicated by the
application, this was amplifying the hashing cost by a big margin.

With Graphics Pipeline Library, we do need to rebuild the pipeline layout
by combining those from each library, but we can memoize the hash of the
descriptor set layout. The cost of re-hashing hashes is negligible since
each descriptor set layout can amount to 1–2KB in size.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22254>

diff --git a/src/amd/vulkan/radv_descriptor_set.c b/src/amd/vulkan/radv_descriptor_set.c
index 082ae47..c35cec1 100644 (file)
--- a/src/amd/vulkan/radv_descriptor_set.c
+++ b/src/amd/vulkan/radv_descriptor_set.c
@@ -183,7 +183,6 @@ radv_CreateDescriptorSetLayout(VkDevice _device, const VkDescriptorSetLayoutCrea
       return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    set_layout->flags = pCreateInfo->flags;
-   set_layout->layout_size = size;
 
    /* We just allocate all the samplers at the end of the struct */
    uint32_t *samplers = (uint32_t *)&set_layout->binding[num_bindings];
@@ -390,6 +389,14 @@ radv_CreateDescriptorSetLayout(VkDevice _device, const VkDescriptorSetLayoutCrea
    set_layout->buffer_count = buffer_count;
    set_layout->dynamic_offset_count = dynamic_offset_count;
 
+   /* Hash the entire set layout except vk_descriptor_set_layout. The rest of the set layout is
+    * carefully constructed to not have pointers so a full hash instead of a per-field hash
+    * should be ok.
+    */
+   uint32_t hash_offset =
+      offsetof(struct radv_descriptor_set_layout, hash) + sizeof(set_layout->hash);
+   _mesa_sha1_compute((const char *)set_layout + hash_offset, size - hash_offset, set_layout->hash);
+
    *pSetLayout = radv_descriptor_set_layout_to_handle(set_layout);
 
    return VK_SUCCESS;
@@ -567,13 +574,7 @@ radv_pipeline_layout_hash(struct radv_pipeline_layout *layout)
       if (!set_layout)
          continue;
 
-      /* Hash the entire set layout except vk_descriptor_set_layout. The rest of the set layout is
-       * carefully constructed to not have pointers so a full hash instead of a per-field hash
-       * should be ok.
-       */
-      uint32_t hash_offset = sizeof(struct vk_descriptor_set_layout);
-      _mesa_sha1_update(&ctx, (const char *)set_layout + hash_offset,
-                        set_layout->layout_size - hash_offset);
+      _mesa_sha1_update(&ctx, set_layout->hash, sizeof(set_layout->hash));
    }
    _mesa_sha1_update(&ctx, &layout->push_constant_size, sizeof(layout->push_constant_size));
    _mesa_sha1_final(&ctx, layout->sha1);

diff --git a/src/amd/vulkan/radv_descriptor_set.h b/src/amd/vulkan/radv_descriptor_set.h
index 9302ec4..77a3b3c 100644 (file)
--- a/src/amd/vulkan/radv_descriptor_set.h
+++ b/src/amd/vulkan/radv_descriptor_set.h
@@ -54,6 +54,9 @@ struct radv_descriptor_set_binding_layout {
 struct radv_descriptor_set_layout {
    struct vk_descriptor_set_layout vk;
 
+   /* Hash of all fields below */
+   uint8_t hash[SHA1_DIGEST_LENGTH];
+
    /* Everything below is hashed and shouldn't contain any pointers. Be careful when modifying this
     * structure.
     */
@@ -67,9 +70,6 @@ struct radv_descriptor_set_layout {
    /* Total size of the descriptor set with room for all array entries */
    uint32_t size;
 
-   /* CPU size of this struct + all associated data, for hashing. */
-   uint32_t layout_size;
-
    /* Shader stages affected by this descriptor set */
    uint16_t shader_stages;
    uint16_t dynamic_shader_stages;