summaryrefslogtreecommitdiffstats
path: root/src
diff options
context:
space:
mode:
authorJason Ekstrand <[email protected]>2015-11-11 09:40:51 -0800
committerJason Ekstrand <[email protected]>2015-11-23 11:02:15 -0800
commit9cf108193b61c342c94c4cd980c4b403638e1051 (patch)
tree34d541e6fcd6917210f9696341ca12ea5a98dedb /src
parent9d703de85a7bdd6ae767aedd7690cb82794cb90a (diff)
i965/nir: Split shader optimization and lowering into three stages
At the moment, brw_create_nir just calls the three stages in sequence so there's not much difference. Soon, however, we will want to start doing variants in NIR at which point the postprocessing step will have to move from shader create time to codegen time. Reviewed-by: Iago Toral Quiroga <[email protected]>
Diffstat (limited to 'src')
-rw-r--r--src/mesa/drivers/dri/i965/brw_nir.c133
-rw-r--r--src/mesa/drivers/dri/i965/brw_nir.h9
2 files changed, 104 insertions, 38 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_nir.c b/src/mesa/drivers/dri/i965/brw_nir.c
index 973a20c0b4e..b9d523dd0b6 100644
--- a/src/mesa/drivers/dri/i965/brw_nir.c
+++ b/src/mesa/drivers/dri/i965/brw_nir.c
@@ -238,43 +238,33 @@ nir_optimize(nir_shader *nir, bool is_scalar)
return nir;
}
+/* Does some simple lowering and runs the standard suite of optimizations
+ *
+ * This is intended to be called more-or-less directly after you get the
+ * shader out of GLSL or some other source. While it is geared towards i965,
+ * it is not at all generator-specific except for the is_scalar flag. Even
+ * there, it is safe to call with is_scalar = false for a shader that is
+ * intended for the FS backend as long as nir_optimize is called again with
+ * is_scalar = true to scalarize everything prior to code gen.
+ */
nir_shader *
-brw_create_nir(struct brw_context *brw,
- const struct gl_shader_program *shader_prog,
- const struct gl_program *prog,
- gl_shader_stage stage,
- bool is_scalar)
+brw_preprocess_nir(nir_shader *nir, bool is_scalar)
{
- struct gl_context *ctx = &brw->ctx;
- const struct brw_device_info *devinfo = brw->intelScreen->devinfo;
- const nir_shader_compiler_options *options =
- ctx->Const.ShaderCompilerOptions[stage].NirOptions;
- static const nir_lower_tex_options tex_options = {
- .lower_txp = ~0,
- };
- bool debug_enabled = INTEL_DEBUG & intel_debug_flag_for_shader_stage(stage);
- bool progress = false;
- nir_shader *nir;
-
- /* First, lower the GLSL IR or Mesa IR to NIR */
- if (shader_prog) {
- nir = glsl_to_nir(shader_prog, stage, options);
- } else {
- nir = prog_to_nir(prog, options);
- OPT_V(nir_convert_to_ssa); /* turn registers into SSA */
- }
- nir_validate_shader(nir);
+ bool progress; /* Written by OPT and OPT_V */
+ (void)progress;
- if (stage == MESA_SHADER_GEOMETRY) {
+ if (nir->stage == MESA_SHADER_GEOMETRY)
OPT(nir_lower_gs_intrinsics);
- }
- OPT(nir_lower_global_vars_to_local);
+ static const nir_lower_tex_options tex_options = {
+ .lower_txp = ~0,
+ };
OPT_V(nir_lower_tex, &tex_options);
-
OPT(nir_normalize_cubemap_coords);
+ OPT(nir_lower_global_vars_to_local);
+
OPT(nir_split_var_copies);
nir = nir_optimize(nir, is_scalar);
@@ -285,6 +275,27 @@ brw_create_nir(struct brw_context *brw,
/* Get rid of split copies */
nir = nir_optimize(nir, is_scalar);
+ OPT(nir_remove_dead_variables);
+
+ return nir;
+}
+
+/* Lowers inputs, outputs, uniforms, and samplers for i965
+ *
+ * This function does all of the standard lowering prior to post-processing.
+ * The lowering done is highly gen, stage, and backend-specific. The
+ * shader_prog parameter is optional and is used only for lowering sampler
+ * derefs and atomics for GLSL shaders.
+ */
+nir_shader *
+brw_lower_nir(nir_shader *nir,
+ const struct brw_device_info *devinfo,
+ const struct gl_shader_program *shader_prog,
+ bool is_scalar)
+{
+ bool progress; /* Written by OPT and OPT_V */
+ (void)progress;
+
OPT_V(brw_nir_lower_inputs, devinfo, is_scalar);
OPT_V(brw_nir_lower_outputs, is_scalar);
nir_assign_var_locations(&nir->uniforms,
@@ -292,8 +303,6 @@ brw_create_nir(struct brw_context *brw,
is_scalar ? type_size_scalar : type_size_vec4);
OPT_V(nir_lower_io, nir_var_all, is_scalar ? type_size_scalar : type_size_vec4);
- OPT(nir_remove_dead_variables);
-
if (shader_prog) {
OPT_V(nir_lower_samplers, shader_prog);
}
@@ -304,9 +313,28 @@ brw_create_nir(struct brw_context *brw,
OPT_V(nir_lower_atomics, shader_prog);
}
- nir = nir_optimize(nir, is_scalar);
+ return nir_optimize(nir, is_scalar);
+}
+
+/* Prepare the given shader for codegen
+ *
+ * This function is intended to be called right before going into the actual
+ * backend and is highly backend-specific. Also, once this function has been
+ * called on a shader, it will no longer be in SSA form so most optimizations
+ * will not work.
+ */
+nir_shader *
+brw_postprocess_nir(nir_shader *nir,
+ const struct brw_device_info *devinfo,
+ bool is_scalar)
+{
+ bool debug_enabled =
+ (INTEL_DEBUG & intel_debug_flag_for_shader_stage(nir->stage));
+
+ bool progress; /* Written by OPT and OPT_V */
+ (void)progress;
- if (brw->gen >= 6) {
+ if (devinfo->gen >= 6) {
/* Try and fuse multiply-adds */
OPT(brw_nir_opt_peephole_ffma);
}
@@ -327,7 +355,7 @@ brw_create_nir(struct brw_context *brw,
}
fprintf(stderr, "NIR (SSA form) for %s shader:\n",
- _mesa_shader_stage_to_string(stage));
+ _mesa_shader_stage_to_string(nir->stage));
nir_print_shader(nir, stderr);
}
@@ -338,28 +366,57 @@ brw_create_nir(struct brw_context *brw,
OPT(nir_lower_vec_to_movs);
}
- /* Needed only so that OPT and OPT_V can set it */
- (void)progress;
-
/* This is the last pass we run before we start emitting stuff. It
* determines when we need to insert boolean resolves on Gen <= 5. We
* run it last because it stashes data in instr->pass_flags and we don't
* want that to be squashed by other NIR passes.
*/
- if (brw->gen <= 5)
+ if (devinfo->gen <= 5)
brw_nir_analyze_boolean_resolves(nir);
nir_sweep(nir);
if (unlikely(debug_enabled)) {
fprintf(stderr, "NIR (final form) for %s shader:\n",
- _mesa_shader_stage_to_string(stage));
+ _mesa_shader_stage_to_string(nir->stage));
nir_print_shader(nir, stderr);
}
return nir;
}
+nir_shader *
+brw_create_nir(struct brw_context *brw,
+ const struct gl_shader_program *shader_prog,
+ const struct gl_program *prog,
+ gl_shader_stage stage,
+ bool is_scalar)
+{
+ struct gl_context *ctx = &brw->ctx;
+ const struct brw_device_info *devinfo = brw->intelScreen->devinfo;
+ const nir_shader_compiler_options *options =
+ ctx->Const.ShaderCompilerOptions[stage].NirOptions;
+ bool progress;
+ nir_shader *nir;
+
+ /* First, lower the GLSL IR or Mesa IR to NIR */
+ if (shader_prog) {
+ nir = glsl_to_nir(shader_prog, stage, options);
+ } else {
+ nir = prog_to_nir(prog, options);
+ OPT_V(nir_convert_to_ssa); /* turn registers into SSA */
+ }
+ nir_validate_shader(nir);
+
+ (void)progress;
+
+ nir = brw_preprocess_nir(nir, is_scalar);
+ nir = brw_lower_nir(nir, devinfo, shader_prog, is_scalar);
+ nir = brw_postprocess_nir(nir, devinfo, is_scalar);
+
+ return nir;
+}
+
enum brw_reg_type
brw_type_for_nir_type(nir_alu_type type)
{
diff --git a/src/mesa/drivers/dri/i965/brw_nir.h b/src/mesa/drivers/dri/i965/brw_nir.h
index e7c93684fb3..baf2f137672 100644
--- a/src/mesa/drivers/dri/i965/brw_nir.h
+++ b/src/mesa/drivers/dri/i965/brw_nir.h
@@ -81,6 +81,15 @@ nir_shader *brw_create_nir(struct brw_context *brw,
gl_shader_stage stage,
bool is_scalar);
+nir_shader *brw_preprocess_nir(nir_shader *nir, bool is_scalar);
+nir_shader *brw_lower_nir(nir_shader *nir,
+ const struct brw_device_info *devinfo,
+ const struct gl_shader_program *shader_prog,
+ bool is_scalar);
+nir_shader *brw_postprocess_nir(nir_shader *nir,
+ const struct brw_device_info *devinfo,
+ bool is_scalar);
+
enum brw_reg_type brw_type_for_nir_type(nir_alu_type type);
enum glsl_base_type brw_glsl_base_type_for_nir_type(nir_alu_type type);