tu: Rework passing shared consts
authorConnor Abbott <cwabbott0@gmail.com>
Thu, 7 Sep 2023 14:42:44 +0000 (16:42 +0200)
committerMarge Bot <emma+marge@anholt.net>
Mon, 25 Sep 2023 19:03:56 +0000 (19:03 +0000)
The way this works now is awkward to map to shader objects. We don't
have the pipeline layout when "linking" shaders at draw time, so we have
to piece it together from the shaders. Store the information we need in
the shaders and piece it together.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25276>

src/freedreno/vulkan/tu_cmd_buffer.cc
src/freedreno/vulkan/tu_pipeline.cc
src/freedreno/vulkan/tu_pipeline.h
src/freedreno/vulkan/tu_shader.cc
src/freedreno/vulkan/tu_shader.h

index fb9b42d..7370abc 100644 (file)
@@ -4309,10 +4309,10 @@ tu6_emit_shared_consts(struct tu_cs *cs,
                        uint32_t *push_constants,
                        bool compute)
 {
-   if (pipeline->shared_consts.dwords > 0) {
+   if (pipeline->program.shared_consts.dwords > 0) {
       /* Offset and num_units for shared consts are in units of dwords. */
-      unsigned num_units = pipeline->shared_consts.dwords;
-      unsigned offset = pipeline->shared_consts.lo;
+      unsigned num_units = pipeline->program.shared_consts.dwords;
+      unsigned offset = pipeline->program.shared_consts.lo;
 
       enum a6xx_state_type st = compute ? ST6_UBO : ST6_CONSTANTS;
       uint32_t cp_load_state = compute ? CP_LOAD_STATE6_FRAG : CP_LOAD_STATE6;
@@ -4338,8 +4338,8 @@ tu6_const_size(struct tu_cmd_buffer *cmd,
 {
    uint32_t dwords = 0;
 
-   if (pipeline->shared_consts.dwords > 0) {
-      dwords += pipeline->shared_consts.dwords + 4;
+   if (pipeline->program.shared_consts.dwords > 0) {
+      dwords += pipeline->program.shared_consts.dwords + 4;
    }
 
    if (compute) {
@@ -4367,7 +4367,7 @@ tu6_emit_consts(struct tu_cmd_buffer *cmd,
    struct tu_cs cs;
    tu_cs_begin_sub_stream(&cmd->sub_cs, dwords, &cs);
 
-   if (pipeline->shared_consts.dwords > 0) {
+   if (pipeline->program.shared_consts.dwords > 0) {
       tu6_emit_shared_consts(&cs, pipeline, cmd->push_constants, compute);
 
       for (uint32_t i = 0; i < ARRAY_SIZE(pipeline->program.link); i++) {
index 2d69cb7..20a1ce9 100644 (file)
@@ -1760,13 +1760,6 @@ tu_pipeline_builder_compile_shaders(struct tu_pipeline_builder *builder,
       must_compile = true;
    }
 
-   if (tu6_shared_constants_enable(&builder->layout, builder->device->compiler)) {
-      pipeline->shared_consts = (struct tu_push_constant_range) {
-         .lo = 0,
-         .dwords = builder->layout.push_constant_size / 4,
-      };
-   }
-
    /* Forward declare everything due to the goto usage */
    nir_shader *nir[ARRAY_SIZE(stage_infos)] = { NULL };
    struct tu_shader *shaders[ARRAY_SIZE(stage_infos)] = { NULL };
@@ -2164,18 +2157,6 @@ tu_pipeline_builder_parse_libraries(struct tu_pipeline_builder *builder,
          tu_pipeline_to_graphics_lib(pipeline)->state |= library->state;
 
       if (library->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_VERTEX_INPUT_INTERFACE_BIT_EXT) {
-         pipeline->shared_consts = library->base.shared_consts;
-      }
-
-      if (library->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) {
-         pipeline->ds = library->base.ds;
-         pipeline->lrz.lrz_status |= library->base.lrz.lrz_status;
-         pipeline->shared_consts = library->base.shared_consts;
-      }
-
-      if (library->state &
           VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT) {
          pipeline->output = library->base.output;
          pipeline->lrz.lrz_status |= library->base.lrz.lrz_status;
@@ -2337,6 +2318,11 @@ tu_pipeline_builder_parse_shader_stages(struct tu_pipeline_builder *builder,
       tu_pipeline_set_linkage(&pipeline->program.link[i],
                               &pipeline->shaders[i]->const_state,
                               variants[i]);
+
+      if (pipeline->shaders[i]->shared_consts.dwords != 0) {
+         pipeline->program.shared_consts =
+            pipeline->shaders[i]->shared_consts;
+      }
    }
 
    const struct ir3_shader_variant *vs = variants[MESA_SHADER_VERTEX];
@@ -4297,13 +4283,6 @@ tu_compute_pipeline_create(VkDevice device,
          VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT;
    }
 
-   if (tu6_shared_constants_enable(layout, dev->compiler)) {
-      pipeline->base.shared_consts = (struct tu_push_constant_range) {
-         .lo = 0,
-         .dwords = layout->push_constant_size / 4,
-      };
-   }
-
    char *nir_initial_disasm = NULL;
 
    if (!shader) {
@@ -4339,6 +4318,8 @@ tu_compute_pipeline_create(VkDevice device,
       creation_feedback->pPipelineStageCreationFeedbacks[0] = pipeline_feedback;
    }
 
+   pipeline->base.program.shared_consts = shader->shared_consts;
+
    pipeline->base.active_desc_sets = shader->active_desc_sets;
 
    v = shader->variant;
index 7d6b2d4..e586e9b 100644 (file)
@@ -135,8 +135,6 @@ struct tu_pipeline
    /* draw states for the pipeline */
    struct tu_draw_state load_state;
 
-   struct tu_push_constant_range shared_consts;
-
    struct tu_shader *shaders[MESA_SHADER_STAGES];
 
    struct
@@ -151,6 +149,8 @@ struct tu_pipeline
 
       uint32_t hs_param_dwords;
 
+      struct tu_push_constant_range shared_consts;
+
       struct tu_program_descriptor_linkage link[MESA_SHADER_STAGES];
 
       bool per_view_viewport;
index b24b17d..78173a7 100644 (file)
@@ -2084,6 +2084,7 @@ tu_shader_serialize(struct vk_pipeline_cache_object *object,
       container_of(object, struct tu_shader, base);
 
    blob_write_bytes(blob, &shader->const_state, sizeof(shader->const_state));
+   blob_write_bytes(blob, &shader->shared_consts, sizeof(shader->shared_consts));
    blob_write_uint32(blob, shader->view_mask);
    blob_write_uint8(blob, shader->active_desc_sets);
 
@@ -2125,6 +2126,7 @@ tu_shader_deserialize(struct vk_pipeline_cache *cache,
       return NULL;
 
    blob_copy_bytes(blob, &shader->const_state, sizeof(shader->const_state));
+   blob_copy_bytes(blob, &shader->shared_consts, sizeof(shader->shared_consts));
    shader->view_mask = blob_read_uint32(blob);
    shader->active_desc_sets = blob_read_uint8(blob);
 
@@ -2276,8 +2278,13 @@ tu_shader_create(struct tu_device *dev,
    ir3_finalize_nir(dev->compiler, nir);
 
    bool shared_consts_enable = tu6_shared_constants_enable(layout, dev->compiler);
-   if (shared_consts_enable)
+   if (shared_consts_enable) {
       assert(!shader->const_state.push_consts.dwords);
+      shader->shared_consts = (struct tu_push_constant_range) {
+         .lo = 0,
+         .dwords = layout->push_constant_size / 4,
+      };
+   }
 
    const struct ir3_shader_options options = {
       .reserved_user_consts = reserved_consts_vec4,
index b086c27..69f4022 100644 (file)
@@ -63,6 +63,11 @@ struct tu_shader
    uint32_t view_mask;
    uint8_t active_desc_sets;
 
+   /* This is the range of shared consts used by all shaders. It must be the
+    * same between shaders.
+    */
+   struct tu_push_constant_range shared_consts;
+
    union {
       struct {
          unsigned patch_type;