radv: calculate GFX9 GS and GFX10 NGG states before compiling shader variants
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Tue, 3 Sep 2019 10:46:55 +0000 (12:46 +0200)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Tue, 10 Sep 2019 07:26:58 +0000 (09:26 +0200)
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/vulkan/radv_pipeline.c

index 70f62725ac0cf70b5dcba754ac6105f30758cae5..1bcc165301fb82303f17c336e0a7143494945db0 100644 (file)
@@ -1496,16 +1496,18 @@ radv_pipeline_init_dynamic_state(struct radv_pipeline *pipeline,
 static void
 gfx9_get_gs_info(const VkGraphicsPipelineCreateInfo *pCreateInfo,
                  const struct radv_pipeline *pipeline,
+                nir_shader **nir,
+                struct radv_shader_info *infos,
                 struct gfx9_gs_info *out)
 {
-       struct radv_shader_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
+       struct radv_shader_info *gs_info = &infos[MESA_SHADER_GEOMETRY];
        struct radv_es_output_info *es_info;
        if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
-               es_info = radv_pipeline_has_tess(pipeline) ? &gs_info->tes.es_info : &gs_info->vs.es_info;
+               es_info = nir[MESA_SHADER_TESS_CTRL] ? &gs_info->tes.es_info : &gs_info->vs.es_info;
        else
-               es_info = radv_pipeline_has_tess(pipeline) ?
-                       &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.es_info :
-                       &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.es_info;
+               es_info = nir[MESA_SHADER_TESS_CTRL] ?
+                       &infos[MESA_SHADER_TESS_EVAL].tes.es_info :
+                       &infos[MESA_SHADER_VERTEX].vs.es_info;
 
        unsigned gs_num_invocations = MAX2(gs_info->gs.invocations, 1);
        bool uses_adjacency;
@@ -1623,21 +1625,20 @@ static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts
 }
 
 static unsigned
-radv_get_num_input_vertices(struct radv_pipeline *pipeline)
+radv_get_num_input_vertices(nir_shader **nir)
 {
-       if (radv_pipeline_has_gs(pipeline)) {
-               struct radv_shader_variant *gs =
-                       radv_get_shader(pipeline, MESA_SHADER_GEOMETRY);
+       if (nir[MESA_SHADER_GEOMETRY]) {
+               nir_shader *gs = nir[MESA_SHADER_GEOMETRY];
 
                return gs->info.gs.vertices_in;
        }
 
-       if (radv_pipeline_has_tess(pipeline)) {
-               struct radv_shader_variant *tes = radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL);
+       if (nir[MESA_SHADER_TESS_CTRL]) {
+               nir_shader *tes = nir[MESA_SHADER_TESS_EVAL];
 
-               if (tes->info.tes.point_mode)
+               if (tes->info.tess.point_mode)
                        return 1;
-               if (tes->info.tes.primitive_mode == GL_ISOLINES)
+               if (tes->info.tess.primitive_mode == GL_ISOLINES)
                        return 2;
                return 3;
        }
@@ -1648,16 +1649,18 @@ radv_get_num_input_vertices(struct radv_pipeline *pipeline)
 static void
 gfx10_get_ngg_info(const VkGraphicsPipelineCreateInfo *pCreateInfo,
                   struct radv_pipeline *pipeline,
+                  nir_shader **nir,
+                  struct radv_shader_info *infos,
                   struct gfx10_ngg_info *ngg)
 {
-       struct radv_shader_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
+       struct radv_shader_info *gs_info = &infos[MESA_SHADER_GEOMETRY];
        struct radv_es_output_info *es_info =
-               radv_pipeline_has_tess(pipeline) ? &gs_info->tes.es_info : &gs_info->vs.es_info;
-       unsigned gs_type = radv_pipeline_has_gs(pipeline) ? MESA_SHADER_GEOMETRY : MESA_SHADER_VERTEX;
-       unsigned max_verts_per_prim = radv_get_num_input_vertices(pipeline);
+               nir[MESA_SHADER_TESS_CTRL] ? &gs_info->tes.es_info : &gs_info->vs.es_info;
+       unsigned gs_type = nir[MESA_SHADER_GEOMETRY] ? MESA_SHADER_GEOMETRY : MESA_SHADER_VERTEX;
+       unsigned max_verts_per_prim = radv_get_num_input_vertices(nir);
        unsigned min_verts_per_prim =
                gs_type == MESA_SHADER_GEOMETRY ? max_verts_per_prim : 1;
-       unsigned gs_num_invocations = radv_pipeline_has_gs(pipeline) ? MAX2(gs_info->gs.invocations, 1) : 1;
+       unsigned gs_num_invocations = nir[MESA_SHADER_GEOMETRY] ? MAX2(gs_info->gs.invocations, 1) : 1;
        bool uses_adjacency;
        switch(pCreateInfo->pInputAssemblyState->topology) {
        case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
@@ -1738,8 +1741,8 @@ gfx10_get_ngg_info(const VkGraphicsPipelineCreateInfo *pCreateInfo,
                 * corresponding to the ES thread of the provoking vertex. All
                 * ES threads load and export PrimitiveID for their thread.
                 */
-               if (!radv_pipeline_has_tess(pipeline) &&
-                   pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.export_prim_id)
+               if (!nir[MESA_SHADER_TESS_CTRL] &&
+                   infos[MESA_SHADER_VERTEX].vs.outinfo.export_prim_id)
                        esvert_lds_size = MAX2(esvert_lds_size, 1);
        }
 
@@ -2562,6 +2565,7 @@ void radv_create_shaders(struct radv_pipeline *pipeline,
                          const struct radv_pipeline_key *key,
                          const VkPipelineShaderStageCreateInfo **pStages,
                          const VkPipelineCreateFlags flags,
+                         const VkGraphicsPipelineCreateInfo *pCreateInfo,
                          VkPipelineCreationFeedbackEXT *pipeline_feedback,
                          VkPipelineCreationFeedbackEXT **stage_feedbacks)
 {
@@ -2668,6 +2672,27 @@ void radv_create_shaders(struct radv_pipeline *pipeline,
 
        radv_fill_shader_info(pipeline, keys, infos, nir);
 
+       if ((nir[MESA_SHADER_VERTEX] &&
+            keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg) ||
+           (nir[MESA_SHADER_TESS_EVAL] &&
+            keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg)) {
+               struct gfx10_ngg_info *ngg_info;
+
+               if (nir[MESA_SHADER_GEOMETRY])
+                       ngg_info = &infos[MESA_SHADER_GEOMETRY].ngg_info;
+               else if (nir[MESA_SHADER_TESS_CTRL])
+                       ngg_info = &infos[MESA_SHADER_TESS_EVAL].ngg_info;
+               else
+                       ngg_info = &infos[MESA_SHADER_VERTEX].ngg_info;
+
+               gfx10_get_ngg_info(pCreateInfo, pipeline, nir, infos, ngg_info);
+       } else if (nir[MESA_SHADER_GEOMETRY]) {
+               struct gfx9_gs_info *gs_info =
+                       &infos[MESA_SHADER_GEOMETRY].gs_ring_info;
+
+               gfx9_get_gs_info(pCreateInfo, pipeline, nir, infos, gs_info);
+       }
+
        if (nir[MESA_SHADER_FRAGMENT]) {
                if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) {
                        radv_start_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT]);
@@ -4615,7 +4640,7 @@ radv_pipeline_init(struct radv_pipeline *pipeline,
        }
 
        struct radv_pipeline_key key = radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend, has_view_index);
-       radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
+       radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pCreateInfo, pipeline_feedback, stage_feedbacks);
 
        pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
        radv_pipeline_init_multisample_state(pipeline, &blend, pCreateInfo);
@@ -4680,22 +4705,10 @@ radv_pipeline_init(struct radv_pipeline *pipeline,
                }
        }
 
-       if (radv_pipeline_has_ngg(pipeline)) {
-               struct radv_shader_variant *ngg;
-
-               if (radv_pipeline_has_gs(pipeline))
-                       ngg = pipeline->shaders[MESA_SHADER_GEOMETRY];
-               else if (radv_pipeline_has_tess(pipeline))
-                       ngg = pipeline->shaders[MESA_SHADER_TESS_EVAL];
-               else
-                       ngg = pipeline->shaders[MESA_SHADER_VERTEX];
-
-               gfx10_get_ngg_info(pCreateInfo, pipeline, &ngg->info.ngg_info);
-       } else if (radv_pipeline_has_gs(pipeline)) {
+       if (radv_pipeline_has_gs(pipeline) && !radv_pipeline_has_ngg(pipeline)) {
                struct radv_shader_variant *gs =
                        pipeline->shaders[MESA_SHADER_GEOMETRY];
 
-               gfx9_get_gs_info(pCreateInfo, pipeline, &gs->info.gs_ring_info);
                calculate_gs_ring_sizes(pipeline, &gs->info.gs_ring_info);
        }
 
@@ -4882,7 +4895,7 @@ static VkResult radv_compute_pipeline_create(
                stage_feedbacks[MESA_SHADER_COMPUTE] = &creation_feedback->pPipelineStageCreationFeedbacks[0];
 
        pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage;
-       radv_create_shaders(pipeline, device, cache, &(struct radv_pipeline_key) {0}, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
+       radv_create_shaders(pipeline, device, cache, &(struct radv_pipeline_key) {0}, pStages, pCreateInfo->flags, NULL, pipeline_feedback, stage_feedbacks);
 
        pipeline->user_data_0[MESA_SHADER_COMPUTE] = radv_pipeline_stage_to_user_data_0(pipeline, MESA_SHADER_COMPUTE, device->physical_device->rad_info.chip_class);
        pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets;