[Mesa-dev] [PATCH 12/12] i965: use nir unrolling for scalar backend

[Timothy Arceri]

---
 src/compiler/glsl/glsl_parser_extras.cpp | 12 +++++++-----
 src/compiler/nir/nir_metadata.c          |  2 ++
 src/mesa/drivers/dri/i965/brw_compiler.c |  6 +++++-
 src/mesa/drivers/dri/i965/brw_nir.c      |  4 ++++
 4 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/src/compiler/glsl/glsl_parser_extras.cpp b/src/compiler/glsl/glsl_parser_extras.cpp
index a185759..8d2807f 100644
--- a/src/compiler/glsl/glsl_parser_extras.cpp
+++ b/src/compiler/glsl/glsl_parser_extras.cpp
@@ -2047,12 +2047,14 @@ do_common_optimization(exec_list *ir, bool linked,
    OPT(optimize_split_arrays, ir, linked);
    OPT(optimize_redundant_jumps, ir);
 
-   loop_state *ls = analyze_loop_variables(ir);
-   if (ls->loop_found) {
-      OPT(set_loop_controls, ir, ls);
-      OPT(unroll_loops, ir, ls, options);
+   if (options->MaxUnrollIterations != 0) {
+      loop_state *ls = analyze_loop_variables(ir);
+      if (ls->loop_found) {
+         OPT(set_loop_controls, ir, ls);
+         OPT(unroll_loops, ir, ls, options);
+      }
+      delete ls;
    }
-   delete ls;
 
 #undef OPT
 
diff --git a/src/compiler/nir/nir_metadata.c b/src/compiler/nir/nir_metadata.c
index 9e1cff5..b163377 100644
--- a/src/compiler/nir/nir_metadata.c
+++ b/src/compiler/nir/nir_metadata.c
@@ -41,6 +41,8 @@ nir_metadata_require(nir_function_impl *impl, nir_metadata required)
       nir_calc_dominance_impl(impl);
    if (NEEDS_UPDATE(nir_metadata_live_ssa_defs))
       nir_live_ssa_defs_impl(impl);
+   if (NEEDS_UPDATE(nir_metadata_loop_analysis))
+      nir_loop_analyze_impl(impl);
 
 #undef NEEDS_UPDATE
 
diff --git a/src/mesa/drivers/dri/i965/brw_compiler.c b/src/mesa/drivers/dri/i965/brw_compiler.c
index afb70a1..df570bf 100644
--- a/src/mesa/drivers/dri/i965/brw_compiler.c
+++ b/src/mesa/drivers/dri/i965/brw_compiler.c
@@ -55,6 +55,7 @@ static const struct nir_shader_compiler_options scalar_nir_options = {
    .lower_unpack_snorm_4x8 = true,
    .lower_unpack_unorm_2x16 = true,
    .lower_unpack_unorm_4x8 = true,
+   .max_unroll_iterations = 32,
 };
 
 static const struct nir_shader_compiler_options vector_nir_options = {
@@ -75,6 +76,7 @@ static const struct nir_shader_compiler_options vector_nir_options = {
    .lower_unpack_unorm_2x16 = true,
    .lower_extract_byte = true,
    .lower_extract_word = true,
+   .max_unroll_iterations = 0,
 };
 
 static const struct nir_shader_compiler_options vector_nir_options_gen6 = {
@@ -92,6 +94,7 @@ static const struct nir_shader_compiler_options vector_nir_options_gen6 = {
    .lower_unpack_unorm_2x16 = true,
    .lower_extract_byte = true,
    .lower_extract_word = true,
+   .max_unroll_iterations = 0,
 };
 
 struct brw_compiler *
@@ -119,7 +122,6 @@ brw_compiler_create(void *mem_ctx, const struct brw_device_info *devinfo)
 
    /* We want the GLSL compiler to emit code that uses condition codes */
    for (int i = 0; i < MESA_SHADER_STAGES; i++) {
-      compiler->glsl_compiler_options[i].MaxUnrollIterations = 32;
       compiler->glsl_compiler_options[i].MaxIfDepth =
          devinfo->gen < 6 ? 16 : UINT_MAX;
 
@@ -140,8 +142,10 @@ brw_compiler_create(void *mem_ctx, const struct brw_device_info *devinfo)
          compiler->glsl_compiler_options[i].EmitNoIndirectSampler = true;
 
       if (is_scalar) {
+         compiler->glsl_compiler_options[i].MaxUnrollIterations = 0;
          compiler->glsl_compiler_options[i].NirOptions = &scalar_nir_options;
       } else {
+         compiler->glsl_compiler_options[i].MaxUnrollIterations = 32;
          compiler->glsl_compiler_options[i].NirOptions =
             devinfo->gen < 6 ? &vector_nir_options : &vector_nir_options_gen6;
       }
diff --git a/src/mesa/drivers/dri/i965/brw_nir.c b/src/mesa/drivers/dri/i965/brw_nir.c
index 6122c2a..e63e299 100644
--- a/src/mesa/drivers/dri/i965/brw_nir.c
+++ b/src/mesa/drivers/dri/i965/brw_nir.c
@@ -387,6 +387,10 @@ nir_optimize(nir_shader *nir, bool is_scalar)
       OPT(nir_opt_algebraic);
       OPT(nir_opt_constant_folding);
       OPT(nir_opt_dead_cf);
+      if (nir->options->max_unroll_iterations != 0) {
+         OPT_V(nir_form_LCSSA);
+         OPT(nir_opt_loop_unroll);
+      }
       OPT(nir_opt_remove_phis);
       OPT(nir_opt_undef);
       OPT_V(nir_lower_doubles, nir_lower_drcp |
-- 
2.7.4