i965: Reduce cross-pollination between the DRI driver and compiler
[mesa.git] / src / mesa / drivers / dri / i965 / brw_gs.c
index dc59b06bac10ebc16eb6afdfabe216903d83a910..412ea97dc09d87d95d04224071d9e790b4450e9f 100644 (file)
 
 #include "brw_gs.h"
 #include "brw_context.h"
-#include "brw_vec4_gs_visitor.h"
 #include "brw_state.h"
 #include "brw_ff_gs.h"
 #include "brw_nir.h"
+#include "brw_program.h"
+#include "compiler/glsl/ir_uniform.h"
 
 static void
-assign_gs_binding_table_offsets(const struct brw_device_info *devinfo,
-                                const struct gl_shader_program *shader_prog,
+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,
                                 struct brw_gs_prog_data *prog_data)
 {
@@ -45,25 +69,25 @@ assign_gs_binding_table_offsets(const struct brw_device_info *devinfo,
     */
    uint32_t reserved = devinfo->gen == 6 ? BRW_MAX_SOL_BINDINGS : 0;
 
-   brw_assign_common_binding_table_offsets(MESA_SHADER_GEOMETRY, devinfo,
-                                           shader_prog, prog,
-                                           &prog_data->base.base,
-                                           reserved);
+   brw_assign_common_binding_table_offsets(devinfo, prog,
+                                           &prog_data->base.base, reserved);
 }
 
-bool
+static bool
 brw_codegen_gs_prog(struct brw_context *brw,
-                    struct gl_shader_program *prog,
-                    struct brw_geometry_program *gp,
+                    struct brw_program *gp,
                     struct brw_gs_prog_key *key)
 {
-   struct gl_shader *shader = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
+   struct brw_compiler *compiler = brw->screen->compiler;
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
    struct brw_stage_state *stage_state = &brw->gs.base;
    struct brw_gs_prog_data prog_data;
+   bool start_busy = false;
+   double start_time = 0;
+
    memset(&prog_data, 0, sizeof(prog_data));
 
-   assign_gs_binding_table_offsets(brw->intelScreen->devinfo, prog,
-                                   &gp->program.Base, &prog_data);
+   assign_gs_binding_table_offsets(devinfo, &gp->program, &prog_data);
 
    /* Allocate the references to the uniforms that will end up in the
     * prog_data associated with the compiled program, and which will be freed
@@ -73,65 +97,80 @@ brw_codegen_gs_prog(struct brw_context *brw,
     * 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.
     */
-   struct gl_shader *gs = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
-   int param_count = gp->program.Base.nir->num_uniforms * 4;
+   int param_count = gp->program.nir->num_uniforms / 4;
 
    prog_data.base.base.param =
       rzalloc_array(NULL, const gl_constant_value *, param_count);
    prog_data.base.base.pull_param =
       rzalloc_array(NULL, const gl_constant_value *, param_count);
    prog_data.base.base.image_param =
-      rzalloc_array(NULL, struct brw_image_param, gs->NumImages);
+      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 = gs->NumImages;
+   prog_data.base.base.nr_image_params = gp->program.info.num_images;
 
-   brw_nir_setup_glsl_uniforms(gp->program.Base.nir, prog, &gp->program.Base,
-                               &prog_data.base.base, false);
+   brw_nir_setup_glsl_uniforms(gp->program.nir, &gp->program,
+                               &prog_data.base.base,
+                               compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
 
-   GLbitfield64 outputs_written = gp->program.Base.OutputsWritten;
+   uint64_t outputs_written = gp->program.info.outputs_written;
 
-   brw_compute_vue_map(brw->intelScreen->devinfo,
+   brw_compute_vue_map(devinfo,
                        &prog_data.base.vue_map, outputs_written,
-                       prog ? prog->SeparateShader : false);
-
-   if (unlikely(INTEL_DEBUG & DEBUG_GS))
-      brw_dump_ir("geometry", prog, gs, NULL);
+                       gp->program.info.separate_shader);
 
    int st_index = -1;
    if (INTEL_DEBUG & DEBUG_SHADER_TIME)
-      st_index = brw_get_shader_time_index(brw, prog, NULL, ST_GS);
+      st_index = brw_get_shader_time_index(brw, &gp->program, ST_GS, true);
+
+   if (unlikely(brw->perf_debug)) {
+      start_busy = brw->batch.last_bo && drm_intel_bo_busy(brw->batch.last_bo);
+      start_time = get_time();
+   }
 
    void *mem_ctx = ralloc_context(NULL);
    unsigned program_size;
    char *error_str;
    const unsigned *program =
-      brw_compile_gs(brw->intelScreen->compiler, brw, mem_ctx, key,
-                     &prog_data, shader->Program->nir, prog,
+      brw_compile_gs(brw->screen->compiler, brw, mem_ctx, key,
+                     &prog_data, gp->program.nir, &gp->program,
                      st_index, &program_size, &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);
+
       ralloc_free(mem_ctx);
       return false;
    }
 
-   /* Scratch space is used for register spilling */
-   if (prog_data.base.base.total_scratch) {
-      brw_get_scratch_bo(brw, &stage_state->scratch_bo,
-                        prog_data.base.base.total_scratch *
-                         brw->max_gs_threads);
+   if (unlikely(brw->perf_debug)) {
+      if (gp->compiled_once) {
+         brw_gs_debug_recompile(brw, &gp->program, key);
+      }
+      if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
+         perf_debug("GS compile took %.03f ms and stalled the GPU\n",
+                    (get_time() - start_time) * 1000);
+      }
+      gp->compiled_once = true;
    }
 
+   /* Scratch space is used for register spilling */
+   brw_alloc_stage_scratch(brw, stage_state,
+                           prog_data.base.base.total_scratch,
+                           devinfo->max_gs_threads);
+
    brw_upload_cache(&brw->cache, BRW_CACHE_GS_PROG,
                     key, sizeof(*key),
                     program, program_size,
                     &prog_data, sizeof(prog_data),
-                    &stage_state->prog_offset, &brw->gs.prog_data);
+                    &stage_state->prog_offset, &brw->gs.base.prog_data);
    ralloc_free(mem_ctx);
 
    return true;
 }
 
 static bool
-brw_gs_state_dirty(struct brw_context *brw)
+brw_gs_state_dirty(const struct brw_context *brw)
 {
    return brw_state_dirty(brw,
                           _NEW_TEXTURE,
@@ -139,35 +178,28 @@ brw_gs_state_dirty(struct brw_context *brw)
                           BRW_NEW_TRANSFORM_FEEDBACK);
 }
 
-static void
+void
 brw_gs_populate_key(struct brw_context *brw,
                     struct brw_gs_prog_key *key)
 {
    struct gl_context *ctx = &brw->ctx;
-   struct brw_stage_state *stage_state = &brw->gs.base;
-   struct brw_geometry_program *gp =
-      (struct brw_geometry_program *) brw->geometry_program;
-   struct gl_program *prog = &gp->program.Base;
+   struct brw_program *gp = (struct brw_program *) brw->geometry_program;
 
    memset(key, 0, sizeof(*key));
 
    key->program_string_id = gp->id;
 
    /* _NEW_TEXTURE */
-   brw_populate_sampler_prog_key_data(ctx, prog, stage_state->sampler_count,
-                                      &key->tex);
+   brw_populate_sampler_prog_key_data(ctx, &gp->program, &key->tex);
 }
 
 void
 brw_upload_gs_prog(struct brw_context *brw)
 {
-   struct gl_context *ctx = &brw->ctx;
-   struct gl_shader_program **current = ctx->_Shader->CurrentProgram;
    struct brw_stage_state *stage_state = &brw->gs.base;
    struct brw_gs_prog_key key;
    /* BRW_NEW_GEOMETRY_PROGRAM */
-   struct brw_geometry_program *gp =
-      (struct brw_geometry_program *) brw->geometry_program;
+   struct brw_program *gp = (struct brw_program *) brw->geometry_program;
 
    if (!brw_gs_state_dirty(brw))
       return;
@@ -183,7 +215,6 @@ brw_upload_gs_prog(struct brw_context *brw)
       /* Other state atoms had better not try to access prog_data, since
        * there's no GS program.
        */
-      brw->gs.prog_data = NULL;
       brw->gs.base.prog_data = NULL;
 
       return;
@@ -193,38 +224,34 @@ brw_upload_gs_prog(struct brw_context *brw)
 
    if (!brw_search_cache(&brw->cache, BRW_CACHE_GS_PROG,
                          &key, sizeof(key),
-                         &stage_state->prog_offset, &brw->gs.prog_data)) {
-      bool success = brw_codegen_gs_prog(brw, current[MESA_SHADER_GEOMETRY],
-                                         gp, &key);
+                         &stage_state->prog_offset,
+                         &brw->gs.base.prog_data)) {
+      bool success = brw_codegen_gs_prog(brw, gp, &key);
       assert(success);
       (void)success;
    }
-   brw->gs.base.prog_data = &brw->gs.prog_data->base.base;
 }
 
 bool
-brw_gs_precompile(struct gl_context *ctx,
-                  struct gl_shader_program *shader_prog,
-                  struct gl_program *prog)
+brw_gs_precompile(struct gl_context *ctx, struct gl_program *prog)
 {
    struct brw_context *brw = brw_context(ctx);
    struct brw_gs_prog_key key;
    uint32_t old_prog_offset = brw->gs.base.prog_offset;
-   struct brw_gs_prog_data *old_prog_data = brw->gs.prog_data;
+   struct brw_stage_prog_data *old_prog_data = brw->gs.base.prog_data;
    bool success;
 
-   struct gl_geometry_program *gp = (struct gl_geometry_program *) prog;
-   struct brw_geometry_program *bgp = brw_geometry_program(gp);
+   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;
 
-   success = brw_codegen_gs_prog(brw, shader_prog, bgp, &key);
+   success = brw_codegen_gs_prog(brw, bgp, &key);
 
    brw->gs.base.prog_offset = old_prog_offset;
-   brw->gs.prog_data = old_prog_data;
+   brw->gs.base.prog_data = old_prog_data;
 
    return success;
 }