i965/gen9: Optimize slice and subslice load balancing behavior.

[mesa.git] / src / mesa / drivers / dri / i965 / brw_tcs.c

diff --git a/src/mesa/drivers/dri/i965/brw_tcs.c b/src/mesa/drivers/dri/i965/brw_tcs.c
index 17f4130c0955726d5e6e339fdba0400ebc801522..b5290ec50688d0d8f492bfae9221b6767169694b 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_tcs.c
+++ b/src/mesa/drivers/dri/i965/brw_tcs.c
@@ -34,41 +34,6 @@
 #include "program/prog_parameter.h"
 #include "nir_builder.h"
 
-static void
-brw_tcs_debug_recompile(struct brw_context *brw, struct gl_program *prog,
-                       const struct brw_tcs_prog_key *key)
-{
-   perf_debug("Recompiling tessellation control shader for program %d\n",
-              prog->Id);
-
-   bool found = false;
-   const struct brw_tcs_prog_key *old_key =
-      brw_find_previous_compile(&brw->cache, BRW_CACHE_TCS_PROG,
-                                key->program_string_id);
-
-   if (!old_key) {
-      perf_debug("  Didn't find previous compile in the shader cache for "
-                 "debug\n");
-      return;
-   }
-
-   found |= key_debug(brw, "input vertices", old_key->input_vertices,
-                      key->input_vertices);
-   found |= key_debug(brw, "outputs written", old_key->outputs_written,
-                      key->outputs_written);
-   found |= key_debug(brw, "patch outputs written", old_key->patch_outputs_written,
-                      key->patch_outputs_written);
-   found |= key_debug(brw, "TES primitive mode", old_key->tes_primitive_mode,
-                      key->tes_primitive_mode);
-   found |= key_debug(brw, "quads and equal_spacing workaround",
-                      old_key->quads_workaround, key->quads_workaround);
-   found |= brw_debug_recompile_sampler_key(brw, &old_key->tex, &key->tex);
-
-   if (!found) {
-      perf_debug("  Something else\n");
-   }
-}
-
 static bool
 brw_codegen_tcs_prog(struct brw_context *brw, struct brw_program *tcp,
                      struct brw_program *tep, struct brw_tcs_prog_key *key)
@@ -84,7 +49,7 @@ brw_codegen_tcs_prog(struct brw_context *brw, struct brw_program *tcp,
 
    void *mem_ctx = ralloc_context(NULL);
    if (tcp) {
-      nir = tcp->program.nir;
+      nir = nir_shader_clone(mem_ctx, tcp->program.nir);
    } else {
       const nir_shader_compiler_options *options =
          ctx->Const.ShaderCompilerOptions[MESA_SHADER_TESS_CTRL].NirOptions;
@@ -100,7 +65,7 @@ brw_codegen_tcs_prog(struct brw_context *brw, struct brw_program *tcp,
       brw_nir_setup_glsl_uniforms(mem_ctx, nir, &tcp->program,
                                   &prog_data.base.base,
                                   compiler->scalar_stage[MESA_SHADER_TESS_CTRL]);
-      brw_nir_analyze_ubo_ranges(compiler, tcp->program.nir, NULL,
+      brw_nir_analyze_ubo_ranges(compiler, nir, NULL,
                                  prog_data.base.base.ubo_ranges);
    } else {
       /* Upload the Patch URB Header as the first two uniforms.
@@ -161,7 +126,8 @@ brw_codegen_tcs_prog(struct brw_context *brw, struct brw_program *tcp,
    if (unlikely(brw->perf_debug)) {
       if (tcp) {
          if (tcp->compiled_once) {
-            brw_tcs_debug_recompile(brw, &tcp->program, key);
+            brw_debug_recompile(brw, MESA_SHADER_TESS_CTRL, tcp->program.Id,
+                                &key->base);
          }
          tcp->compiled_once = true;
       }
@@ -194,6 +160,7 @@ brw_tcs_populate_key(struct brw_context *brw,
                      struct brw_tcs_prog_key *key)
 {
    const struct gen_device_info *devinfo = &brw->screen->devinfo;
+   const struct brw_compiler *compiler = brw->screen->compiler;
    struct brw_program *tcp =
       (struct brw_program *) brw->programs[MESA_SHADER_TESS_CTRL];
    struct brw_program *tep =
@@ -211,7 +178,7 @@ brw_tcs_populate_key(struct brw_context *brw,
       per_patch_slots |= prog->info.patch_outputs_written;
    }
 
-   if (devinfo->gen < 8 || !tcp)
+   if (devinfo->gen < 8 || !tcp || compiler->use_tcs_8_patch)
       key->input_vertices = brw->ctx.TessCtrlProgram.patch_vertices;
    key->outputs_written = per_vertex_slots;
    key->patch_outputs_written = per_patch_slots;
@@ -225,10 +192,8 @@ brw_tcs_populate_key(struct brw_context *brw,
                            tep->program.info.tess.spacing == TESS_SPACING_EQUAL;
 
    if (tcp) {
-      key->program_string_id = tcp->id;
-
       /* _NEW_TEXTURE */
-      brw_populate_sampler_prog_key_data(&brw->ctx, &tcp->program, &key->tex);
+      brw_populate_base_prog_key(&brw->ctx, tcp, &key->base);
    }
 }
 
@@ -240,7 +205,7 @@ brw_upload_tcs_prog(struct brw_context *brw)
    /* BRW_NEW_TESS_PROGRAMS */
    struct brw_program *tcp =
       (struct brw_program *) brw->programs[MESA_SHADER_TESS_CTRL];
-   MAYBE_UNUSED struct brw_program *tep =
+   ASSERTED struct brw_program *tep =
       (struct brw_program *) brw->programs[MESA_SHADER_TESS_EVAL];
    assert(tep);
 
@@ -262,29 +227,29 @@ brw_upload_tcs_prog(struct brw_context *brw)
 
    tcp = (struct brw_program *) brw->programs[MESA_SHADER_TESS_CTRL];
    if (tcp)
-      tcp->id = key.program_string_id;
+      tcp->id = key.base.program_string_id;
 
-   MAYBE_UNUSED bool success = brw_codegen_tcs_prog(brw, tcp, tep, &key);
+   ASSERTED bool success = brw_codegen_tcs_prog(brw, tcp, tep, &key);
    assert(success);
 }
 
 void
-brw_tcs_populate_default_key(const struct gen_device_info *devinfo,
+brw_tcs_populate_default_key(const struct brw_compiler *compiler,
                              struct brw_tcs_prog_key *key,
                              struct gl_shader_program *sh_prog,
                              struct gl_program *prog)
 {
+   const struct gen_device_info *devinfo = compiler->devinfo;
    struct brw_program *btcp = brw_program(prog);
    const struct gl_linked_shader *tes =
       sh_prog->_LinkedShaders[MESA_SHADER_TESS_EVAL];
 
    memset(key, 0, sizeof(*key));
 
-   key->program_string_id = btcp->id;
-   brw_setup_tex_for_precompile(devinfo, &key->tex, prog);
+   brw_populate_default_base_prog_key(devinfo, btcp, &key->base);
 
    /* Guess that the input and output patches have the same dimensionality. */
-   if (devinfo->gen < 8)
+   if (devinfo->gen < 8 || compiler->use_tcs_8_patch)
       key->input_vertices = prog->info.tess.tcs_vertices_out;
 
    if (tes) {
@@ -306,6 +271,7 @@ brw_tcs_precompile(struct gl_context *ctx,
                    struct gl_program *prog)
 {
    struct brw_context *brw = brw_context(ctx);
+   const struct brw_compiler *compiler = brw->screen->compiler;
    struct brw_tcs_prog_key key;
    uint32_t old_prog_offset = brw->tcs.base.prog_offset;
    struct brw_stage_prog_data *old_prog_data = brw->tcs.base.prog_data;
@@ -316,7 +282,7 @@ brw_tcs_precompile(struct gl_context *ctx,
       shader_prog->_LinkedShaders[MESA_SHADER_TESS_EVAL];
    struct brw_program *btep = tes ? brw_program(tes->Program) : NULL;
 
-   brw_tcs_populate_default_key(&brw->screen->devinfo, &key, shader_prog, prog);
+   brw_tcs_populate_default_key(compiler, &key, shader_prog, prog);
 
    success = brw_codegen_tcs_prog(brw, btcp, btep, &key);