turnip: Only emit descriptor loads for active stages in the pipeline.

zink has a push descriptor template layout that has every possible stage,
which gets used regardless of what stages are in the pipeline.  By
skipping over the unused stages, we cut the CP overhead.

Improves TU_DEBUG=sysmem gfxbench gl_driver2 on zink by 6.57% +/-
0.331143% (n=5).

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18562>

diff --git a/src/freedreno/vulkan/tu_pipeline.c b/src/freedreno/vulkan/tu_pipeline.c
index ba44b74..e5d15bb 100644 (file)
--- a/src/freedreno/vulkan/tu_pipeline.c
+++ b/src/freedreno/vulkan/tu_pipeline.c
@@ -53,7 +53,7 @@ emit_load_state(struct tu_cs *cs, unsigned opcode, enum a6xx_state_type st,
 
 static unsigned
 tu6_load_state_size(struct tu_pipeline *pipeline,
-                    struct tu_pipeline_layout *layout, bool compute)
+                    struct tu_pipeline_layout *layout)
 {
    const unsigned load_state_size = 4;
    unsigned size = 0;
@@ -65,13 +65,8 @@ tu6_load_state_size(struct tu_pipeline *pipeline,
       for (unsigned j = 0; j < set_layout->binding_count; j++) {
          struct tu_descriptor_set_binding_layout *binding = &set_layout->binding[j];
          unsigned count = 0;
-         /* Note: some users, like amber for example, pass in
-          * VK_SHADER_STAGE_ALL which includes a bunch of extra bits, so
-          * filter these out by using VK_SHADER_STAGE_ALL_GRAPHICS explicitly.
-          */
-         VkShaderStageFlags stages = compute ?
-            binding->shader_stages & VK_SHADER_STAGE_COMPUTE_BIT :
-            binding->shader_stages & VK_SHADER_STAGE_ALL_GRAPHICS;
+         /* See comment in tu6_emit_load_state(). */
+         VkShaderStageFlags stages = pipeline->active_stages & binding->shader_stages;
          unsigned stage_count = util_bitcount(stages);
 
          if (!binding->array_size)
@@ -83,9 +78,7 @@ tu6_load_state_size(struct tu_pipeline *pipeline,
          case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
          case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
             /* IBO-backed resources only need one packet for all graphics stages */
-            if (stages & ~VK_SHADER_STAGE_COMPUTE_BIT)
-               count += 1;
-            if (stages & VK_SHADER_STAGE_COMPUTE_BIT)
+            if (stage_count)
                count += 1;
             break;
          case VK_DESCRIPTOR_TYPE_SAMPLER:
@@ -116,9 +109,9 @@ tu6_load_state_size(struct tu_pipeline *pipeline,
 
 static void
 tu6_emit_load_state(struct tu_pipeline *pipeline,
-                    struct tu_pipeline_layout *layout, bool compute)
+                    struct tu_pipeline_layout *layout)
 {
-   unsigned size = tu6_load_state_size(pipeline, layout, compute);
+   unsigned size = tu6_load_state_size(pipeline, layout);
    if (size == 0)
       return;
 
@@ -150,13 +143,12 @@ tu6_emit_load_state(struct tu_pipeline *pipeline,
          struct tu_descriptor_set_binding_layout *binding = &set_layout->binding[j];
          unsigned base = i;
          unsigned offset = binding->offset / 4;
-         /* Note: some users, like amber for example, pass in
-          * VK_SHADER_STAGE_ALL which includes a bunch of extra bits, so
-          * filter these out by using VK_SHADER_STAGE_ALL_GRAPHICS explicitly.
+         /* Note: amber sets VK_SHADER_STAGE_ALL for its descriptor layout, and
+          * zink has descriptors for each stage in the push layout even if some
+          * stages aren't present in a used pipeline.  We don't want to emit
+          * loads for unused descriptors.
           */
-         VkShaderStageFlags stages = compute ?
-            binding->shader_stages & VK_SHADER_STAGE_COMPUTE_BIT :
-            binding->shader_stages & VK_SHADER_STAGE_ALL_GRAPHICS;
+         VkShaderStageFlags stages = pipeline->active_stages & binding->shader_stages;
          unsigned count = binding->array_size;
          if (count == 0 || stages == 0)
             continue;
@@ -2395,7 +2387,7 @@ tu_pipeline_allocate_cs(struct tu_device *dev,
                         struct tu_pipeline_builder *builder,
                         struct ir3_shader_variant *compute)
 {
-   uint32_t size = 1024 + tu6_load_state_size(pipeline, layout, compute);
+   uint32_t size = 1024 + tu6_load_state_size(pipeline, layout);
 
    /* graphics case: */
    if (builder) {
@@ -2861,6 +2853,8 @@ tu_pipeline_builder_compile_shaders(struct tu_pipeline_builder *builder,
       gl_shader_stage stage =
          vk_to_mesa_shader_stage(builder->create_info->pStages[i].stage);
       stage_infos[stage] = &builder->create_info->pStages[i];
+
+      pipeline->active_stages |= builder->create_info->pStages[i].stage;
    }
 
    if (tu6_shared_constants_enable(builder->layout, builder->device->compiler)) {
@@ -3303,12 +3297,6 @@ tu_pipeline_builder_parse_shader_stages(struct tu_pipeline_builder *builder,
    tu6_emit_program(&prog_cs, builder, true, pipeline);
    pipeline->program.binning_state = tu_cs_end_draw_state(&pipeline->cs, &prog_cs);
 
-   VkShaderStageFlags stages = 0;
-   for (unsigned i = 0; i < builder->create_info->stageCount; i++) {
-      stages |= builder->create_info->pStages[i].stage;
-   }
-   pipeline->active_stages = stages;
-
    for (unsigned i = 0; i < ARRAY_SIZE(builder->shaders->variants); i++) {
       if (!builder->shaders->variants[i])
          continue;
@@ -3960,7 +3948,7 @@ tu_pipeline_builder_build(struct tu_pipeline_builder *builder,
    tu_pipeline_builder_parse_depth_stencil(builder, *pipeline);
    tu_pipeline_builder_parse_multisample_and_color_blend(builder, *pipeline);
    tu_pipeline_builder_parse_rasterization_order(builder, *pipeline);
-   tu6_emit_load_state(*pipeline, builder->layout, false);
+   tu6_emit_load_state(*pipeline, builder->layout);
 
    return VK_SUCCESS;
 }
@@ -4218,6 +4206,7 @@ tu_compute_pipeline_create(VkDevice device,
 
    pipeline->executables_mem_ctx = ralloc_context(NULL);
    util_dynarray_init(&pipeline->executables, pipeline->executables_mem_ctx);
+   pipeline->active_stages = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
 
    struct tu_shader_key key = { };
    tu_shader_key_init(&key, stage_info, dev);
@@ -4336,7 +4325,7 @@ tu_compute_pipeline_create(VkDevice device,
    tu6_emit_cs_config(&prog_cs, v, &pvtmem, shader_iova);
    pipeline->program.state = tu_cs_end_draw_state(&pipeline->cs, &prog_cs);
 
-   tu6_emit_load_state(pipeline, layout, true);
+   tu6_emit_load_state(pipeline, layout);
 
    tu_append_executable(pipeline, v, nir_initial_disasm);