i965/gen9: Optimize slice and subslice load balancing behavior.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_gs.c
index c040665a3b611d98f11f0434e2bdb5db7b8be680..07ca98fda8f69f8f6a7082c32f58d266630651b4 100644 (file)
 #include "brw_program.h"
 #include "compiler/glsl/ir_uniform.h"
 
-static void
-brw_gs_debug_recompile(struct brw_context *brw, struct gl_program *prog,
-                       const struct brw_gs_prog_key *key)
-{
-   perf_debug("Recompiling geometry shader for program %d\n", prog->Id);
-
-   bool found = false;
-   const struct brw_gs_prog_key *old_key =
-      brw_find_previous_compile(&brw->cache, BRW_CACHE_GS_PROG,
-                                key->program_string_id);
-
-   if (!old_key) {
-      perf_debug("  Didn't find previous compile in the shader cache for "
-                 "debug\n");
-      return;
-   }
-
-   found |= brw_debug_recompile_sampler_key(brw, &old_key->tex, &key->tex);
-
-   if (!found) {
-      perf_debug("  Something else\n");
-   }
-}
-
 static void
 assign_gs_binding_table_offsets(const struct gen_device_info *devinfo,
                                 const struct gl_program *prog,
@@ -87,33 +63,19 @@ brw_codegen_gs_prog(struct brw_context *brw,
 
    memset(&prog_data, 0, sizeof(prog_data));
 
-   assign_gs_binding_table_offsets(devinfo, &gp->program, &prog_data);
+   void *mem_ctx = ralloc_context(NULL);
 
-   /* Allocate the references to the uniforms that will end up in the
-    * prog_data associated with the compiled program, and which will be freed
-    * by the state cache.
-    *
-    * Note: param_count needs to be num_uniform_components * 4, since we add
-    * padding around uniform values below vec4 size, so the worst case is that
-    * every uniform is a float which gets padded to the size of a vec4.
-    */
-   int param_count = gp->program.nir->num_uniforms / 4;
+   nir_shader *nir = nir_shader_clone(mem_ctx, gp->program.nir);
 
-   prog_data.base.base.param = rzalloc_array(NULL, uint32_t, param_count);
-   prog_data.base.base.pull_param = rzalloc_array(NULL, uint32_t, param_count);
-   prog_data.base.base.image_param =
-      rzalloc_array(NULL, struct brw_image_param,
-                    gp->program.info.num_images);
-   prog_data.base.base.nr_params = param_count;
-   prog_data.base.base.nr_image_params = gp->program.info.num_images;
+   assign_gs_binding_table_offsets(devinfo, &gp->program, &prog_data);
 
-   brw_nir_setup_glsl_uniforms(gp->program.nir, &gp->program,
+   brw_nir_setup_glsl_uniforms(mem_ctx, nir, &gp->program,
                                &prog_data.base.base,
                                compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
-   brw_nir_analyze_ubo_ranges(compiler, gp->program.nir,
+   brw_nir_analyze_ubo_ranges(compiler, nir, NULL,
                               prog_data.base.base.ubo_ranges);
 
-   uint64_t outputs_written = gp->program.nir->info.outputs_written;
+   uint64_t outputs_written = nir->info.outputs_written;
 
    brw_compute_vue_map(devinfo,
                        &prog_data.base.vue_map, outputs_written,
@@ -128,13 +90,10 @@ brw_codegen_gs_prog(struct brw_context *brw,
       start_time = get_time();
    }
 
-   void *mem_ctx = ralloc_context(NULL);
-   unsigned program_size;
    char *error_str;
    const unsigned *program =
       brw_compile_gs(brw->screen->compiler, brw, mem_ctx, key,
-                     &prog_data, gp->program.nir, &gp->program,
-                     st_index, &program_size, &error_str);
+                     &prog_data, nir, &gp->program, st_index, &error_str);
    if (program == NULL) {
       ralloc_strcat(&gp->program.sh.data->InfoLog, error_str);
       _mesa_problem(NULL, "Failed to compile geometry shader: %s\n", error_str);
@@ -145,7 +104,8 @@ brw_codegen_gs_prog(struct brw_context *brw,
 
    if (unlikely(brw->perf_debug)) {
       if (gp->compiled_once) {
-         brw_gs_debug_recompile(brw, &gp->program, key);
+         brw_debug_recompile(brw, MESA_SHADER_GEOMETRY, gp->program.Id,
+                             &key->base);
       }
       if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
          perf_debug("GS compile took %.03f ms and stalled the GPU\n",
@@ -156,12 +116,14 @@ brw_codegen_gs_prog(struct brw_context *brw,
 
    /* Scratch space is used for register spilling */
    brw_alloc_stage_scratch(brw, stage_state,
-                           prog_data.base.base.total_scratch,
-                           devinfo->max_gs_threads);
+                           prog_data.base.base.total_scratch);
 
+   /* The param and pull_param arrays will be freed by the shader cache. */
+   ralloc_steal(NULL, prog_data.base.base.param);
+   ralloc_steal(NULL, prog_data.base.base.pull_param);
    brw_upload_cache(&brw->cache, BRW_CACHE_GS_PROG,
                     key, sizeof(*key),
-                    program, program_size,
+                    program, prog_data.base.base.program_size,
                     &prog_data, sizeof(prog_data),
                     &stage_state->prog_offset, &brw->gs.base.prog_data);
    ralloc_free(mem_ctx);
@@ -188,10 +150,7 @@ brw_gs_populate_key(struct brw_context *brw,
 
    memset(key, 0, sizeof(*key));
 
-   key->program_string_id = gp->id;
-
-   /* _NEW_TEXTURE */
-   brw_populate_sampler_prog_key_data(ctx, &gp->program, &key->tex);
+   brw_populate_base_prog_key(ctx, gp, &key->base);
 }
 
 void
@@ -208,14 +167,32 @@ brw_upload_gs_prog(struct brw_context *brw)
 
    brw_gs_populate_key(brw, &key);
 
-   if (!brw_search_cache(&brw->cache, BRW_CACHE_GS_PROG,
-                         &key, sizeof(key),
-                         &stage_state->prog_offset,
-                         &brw->gs.base.prog_data)) {
-      bool success = brw_codegen_gs_prog(brw, gp, &key);
-      assert(success);
-      (void)success;
-   }
+   if (brw_search_cache(&brw->cache, BRW_CACHE_GS_PROG, &key, sizeof(key),
+                        &stage_state->prog_offset, &brw->gs.base.prog_data,
+                        true))
+      return;
+
+   if (brw_disk_cache_upload_program(brw, MESA_SHADER_GEOMETRY))
+      return;
+
+   gp = (struct brw_program *) brw->programs[MESA_SHADER_GEOMETRY];
+   gp->id = key.base.program_string_id;
+
+   ASSERTED bool success = brw_codegen_gs_prog(brw, gp, &key);
+   assert(success);
+}
+
+void
+brw_gs_populate_default_key(const struct brw_compiler *compiler,
+                            struct brw_gs_prog_key *key,
+                            struct gl_program *prog)
+{
+   const struct gen_device_info *devinfo = compiler->devinfo;
+
+   memset(key, 0, sizeof(*key));
+
+   brw_populate_default_base_prog_key(devinfo, brw_program(prog),
+                                      &key->base);
 }
 
 bool
@@ -229,10 +206,7 @@ brw_gs_precompile(struct gl_context *ctx, struct gl_program *prog)
 
    struct brw_program *bgp = brw_program(prog);
 
-   memset(&key, 0, sizeof(key));
-
-   brw_setup_tex_for_precompile(brw, &key.tex, prog);
-   key.program_string_id = bgp->id;
+   brw_gs_populate_default_key(brw->screen->compiler, &key, prog);
 
    success = brw_codegen_gs_prog(brw, bgp, &key);