summaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/radeonsi
diff options
context:
space:
mode:
authorNicolai Hähnle <[email protected]>2017-06-24 17:48:22 +0200
committerNicolai Hähnle <[email protected]>2017-07-31 14:55:40 +0200
commit41d4016e061eea2a4be4f9c64930b818b9572f00 (patch)
tree41914e9b66de87d664b555a941b903c97bfdd624 /src/gallium/drivers/radeonsi
parentb49c2c9fa39f5a24d0c28863438d7e428aa564df (diff)
radeonsi/nir: perform radeonsi-specific lowering and optimization passes
Reviewed-by: Marek Olšák <[email protected]>
Diffstat (limited to 'src/gallium/drivers/radeonsi')
-rw-r--r--src/gallium/drivers/radeonsi/si_shader_nir.c41
1 files changed, 41 insertions, 0 deletions
diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index 2c311d3865d..d87f6b831d8 100644
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -337,6 +337,47 @@ si_lower_nir(struct si_shader_selector* sel)
variable->data.driver_location += 1;
}
}
+
+ /* Perform lowerings (and optimizations) of code.
+ *
+ * Performance considerations aside, we must:
+ * - lower certain ALU operations
+ * - ensure constant offsets for texture instructions are folded
+ * and copy-propagated
+ */
+ NIR_PASS_V(sel->nir, nir_lower_returns);
+ NIR_PASS_V(sel->nir, nir_lower_vars_to_ssa);
+ NIR_PASS_V(sel->nir, nir_lower_alu_to_scalar);
+ NIR_PASS_V(sel->nir, nir_lower_phis_to_scalar);
+
+ bool progress;
+ do {
+ progress = false;
+
+ /* (Constant) copy propagation is needed for txf with offsets. */
+ NIR_PASS(progress, sel->nir, nir_copy_prop);
+ NIR_PASS(progress, sel->nir, nir_opt_remove_phis);
+ NIR_PASS(progress, sel->nir, nir_opt_dce);
+ if (nir_opt_trivial_continues(sel->nir)) {
+ progress = true;
+ NIR_PASS(progress, sel->nir, nir_copy_prop);
+ NIR_PASS(progress, sel->nir, nir_opt_dce);
+ }
+ NIR_PASS(progress, sel->nir, nir_opt_if);
+ NIR_PASS(progress, sel->nir, nir_opt_dead_cf);
+ NIR_PASS(progress, sel->nir, nir_opt_cse);
+ NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8);
+
+ /* Needed for algebraic lowering */
+ NIR_PASS(progress, sel->nir, nir_opt_algebraic);
+ NIR_PASS(progress, sel->nir, nir_opt_constant_folding);
+
+ NIR_PASS(progress, sel->nir, nir_opt_undef);
+ NIR_PASS(progress, sel->nir, nir_opt_conditional_discard);
+ if (sel->nir->options->max_unroll_iterations) {
+ NIR_PASS(progress, sel->nir, nir_opt_loop_unroll, 0);
+ }
+ } while (progress);
}
static void declare_nir_input_vs(struct si_shader_context *ctx,