anv: add direct descriptor support to apply_layout

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21645>

src/intel/genxml/meson.build

@@ -74,6 +74,10 @@ genX_bits_included_symbols = [
   'RENDER_SURFACE_STATE::Green Clear Color',
   'RENDER_SURFACE_STATE::Blue Clear Color',
   'RENDER_SURFACE_STATE::Alpha Clear Color',
+  'RENDER_SURFACE_STATE::Width',
+  'RENDER_SURFACE_STATE::Height',
+  'RENDER_SURFACE_STATE::Depth',
+  'RENDER_SURFACE_STATE::Surface Type',
   'CLEAR_COLOR',
   'VERTEX_BUFFER_STATE::Buffer Starting Address',
   'CPS_STATE',

src/intel/vulkan/anv_nir.h

@@ -31,6 +31,18 @@
 extern "C" {
 #endif
 
+/* This map is represent a mapping where the key is the NIR
+ * nir_intrinsic_resource_intel::block index. It allows mapping bindless UBOs
+ * accesses to descriptor entry.
+ *
+ * This map only temporary lives between the anv_nir_apply_pipeline_layout()
+ * and anv_nir_compute_push_layout() passes.
+ */
+struct anv_pipeline_push_map {
+   uint32_t                     block_count;
+   struct anv_pipeline_binding *block_to_descriptor;
+};
+
 bool anv_check_for_primitive_replication(struct anv_device *device,
                                          VkShaderStageFlags stages,
                                          nir_shader **shaders,
@@ -71,7 +83,9 @@ void anv_nir_apply_pipeline_layout(nir_shader *shader,
                                    bool robust_buffer_access,
                                    bool independent_sets,
                                    const struct anv_pipeline_sets_layout *layout,
-                                   struct anv_pipeline_bind_map *map);
+                                   struct anv_pipeline_bind_map *map,
+                                   struct anv_pipeline_push_map *push_map,
+                                   void *push_map_mem_ctx);
 
 void anv_nir_compute_push_layout(nir_shader *nir,
                                  const struct anv_physical_device *pdevice,
@@ -79,6 +93,7 @@ void anv_nir_compute_push_layout(nir_shader *nir,
                                  bool fragment_dynamic,
                                  struct brw_stage_prog_data *prog_data,
                                  struct anv_pipeline_bind_map *map,
+                                 const struct anv_pipeline_push_map *push_map,
                                  void *mem_ctx);
 
 void anv_nir_validate_push_layout(struct brw_stage_prog_data *prog_data,

src/intel/vulkan/anv_nir_compute_push_layout.c

@@ -35,6 +35,7 @@ anv_nir_compute_push_layout(nir_shader *nir,
                             bool fragment_dynamic,
                             struct brw_stage_prog_data *prog_data,
                             struct anv_pipeline_bind_map *map,
+                            const struct anv_pipeline_push_map *push_map,
                             void *mem_ctx)
 {
    const struct brw_compiler *compiler = pdevice->compiler;
@@ -257,8 +258,9 @@ anv_nir_compute_push_layout(nir_shader *nir,
             continue;
          }
 
+         assert(ubo_range->block < push_map->block_count);
          const struct anv_pipeline_binding *binding =
-            &map->surface_to_descriptor[ubo_range->block];
+            &push_map->block_to_descriptor[ubo_range->block];
 
          map->push_ranges[n++] = (struct anv_push_range) {
             .set = binding->set,
@@ -299,6 +301,16 @@ anv_nir_compute_push_layout(nir_shader *nir,
          (fs_msaa_flags_offset - push_start) / 4;
    }
 
+#if 0
+   fprintf(stderr, "stage=%s push ranges:\n", gl_shader_stage_name(nir->info.stage));
+   for (unsigned i = 0; i < ARRAY_SIZE(map->push_ranges); i++)
+      fprintf(stderr, "   range%i: %03u-%03u set=%u index=%u\n", i,
+              map->push_ranges[i].start,
+              map->push_ranges[i].length,
+              map->push_ranges[i].set,
+              map->push_ranges[i].index);
+#endif
+
    /* Now that we're done computing the push constant portion of the
     * bind map, hash it.  This lets us quickly determine if the actual
     * mapping has changed and not just a no-op pipeline change.

src/intel/vulkan/anv_pipeline.c

@@ -978,11 +978,13 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
    stage->push_desc_info.used_descriptors =
       anv_nir_compute_used_push_descriptors(nir, layout);
 
+   struct anv_pipeline_push_map push_map = {};
+
    /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
    NIR_PASS_V(nir, anv_nir_apply_pipeline_layout,
               pdevice, pipeline->device->robust_buffer_access,
               layout->independent_sets,
-              layout, &stage->bind_map);
+              layout, &stage->bind_map, &push_map, mem_ctx);
 
    NIR_PASS(_, nir, nir_lower_explicit_io, nir_var_mem_ubo,
             anv_nir_ubo_addr_format(pdevice, pipeline->device->robust_buffer_access));
@@ -993,8 +995,14 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
     * calculations often create and then constant-fold so that, when we
     * get to anv_nir_lower_ubo_loads, we can detect constant offsets.
     */
-   NIR_PASS(_, nir, nir_copy_prop);
+   bool progress;
-   NIR_PASS(_, nir, nir_opt_constant_folding);
+   do {
+      progress = false;
+      NIR_PASS(progress, nir, nir_opt_algebraic);
+      NIR_PASS(progress, nir, nir_copy_prop);
+      NIR_PASS(progress, nir, nir_opt_constant_folding);
+      NIR_PASS(progress, nir, nir_opt_dce);
+   } while (progress);
 
    /* Required for nir_divergence_analysis() which is needed for
     * anv_nir_lower_ubo_loads.
@@ -1007,7 +1015,9 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
    NIR_PASS(_, nir, nir_opt_remove_phis);
 
    enum nir_lower_non_uniform_access_type lower_non_uniform_access_types =
-      nir_lower_non_uniform_texture_access | nir_lower_non_uniform_image_access;
+      nir_lower_non_uniform_texture_access |
+      nir_lower_non_uniform_image_access |
+      nir_lower_non_uniform_get_ssbo_size;
 
    /* In practice, most shaders do not have non-uniform-qualified
     * accesses (see
@@ -1038,7 +1048,7 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
    NIR_PASS_V(nir, anv_nir_compute_push_layout,
               pdevice, pipeline->device->robust_buffer_access,
               anv_graphics_pipeline_stage_fragment_dynamic(stage),
-              prog_data, &stage->bind_map, mem_ctx);
+              prog_data, &stage->bind_map, &push_map, mem_ctx);
 
    NIR_PASS_V(nir, anv_nir_lower_resource_intel, pdevice,
               pipeline->layout.type);
@@ -3193,6 +3203,15 @@ VkResult anv_CreateGraphicsPipelines(
    return result;
 }
 
+static bool
+should_remat_cb(nir_instr *instr, void *data)
+{
+   if (instr->type != nir_instr_type_intrinsic)
+      return false;
+
+   return nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_resource_intel;
+}
+
 static VkResult
 compile_upload_rt_shader(struct anv_ray_tracing_pipeline *pipeline,
                          struct vk_pipeline_cache *cache,
@@ -3214,6 +3233,7 @@ compile_upload_rt_shader(struct anv_ray_tracing_pipeline *pipeline,
          .localized_loads = true,
          .vectorizer_callback = brw_nir_should_vectorize_mem,
          .vectorizer_data = NULL,
+         .should_remat_callback = should_remat_cb,
       };
 
       NIR_PASS(_, nir, nir_lower_shader_calls, &opts,