fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-25 06:30:10 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

treewide: don't lower to LCSSA before calling nir_divergence_analysis()

Browse source 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30787>
This commit is contained in:
Daniel Schürmann 2024-08-21 15:05:55 +02:00 • committed by Marge Bot
parent 95ed72922e
commit 87cb42f953
10 changed files with 4 additions and 41 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/amd/common/ac_nir_lower_ngg.c
View file

@ -1120,8 +1120,6 @@ analyze_shader_before_culling_walk(nir_def *ssa,
static void
analyze_shader_before_culling(nir_shader *shader, lower_ngg_nogs_state *s)
{
/* LCSSA is needed to get correct results from divergence analysis. */
nir_convert_to_lcssa(shader, true, true);
/* We need divergence info for culling shaders. */
nir_divergence_analysis(shader);

14
src/amd/vulkan/radv_pipeline.c
View file

@ -446,10 +446,9 @@ radv_postprocess_nir(struct radv_device *device, const struct radv_graphics_stat
bool fix_derivs_in_divergent_cf =
stage->stage == MESA_SHADER_FRAGMENT && !radv_use_llvm_for_stage(pdev, stage->stage);
if (fix_derivs_in_divergent_cf) {
NIR_PASS(_, stage->nir, nir_convert_to_lcssa, true, true);
if (fix_derivs_in_divergent_cf)
nir_divergence_analysis(stage->nir);
}
NIR_PASS(_, stage->nir, ac_nir_lower_tex,
&(ac_nir_lower_tex_options){
.gfx_level = gfx_level,
@ -457,8 +456,6 @@ radv_postprocess_nir(struct radv_device *device, const struct radv_graphics_stat
.fix_derivs_in_divergent_cf = fix_derivs_in_divergent_cf,
.max_wqm_vgprs = 64, // TODO: improve spiller and RA support for linear VGPRs
});
if (fix_derivs_in_divergent_cf)
NIR_PASS(_, stage->nir, nir_opt_remove_phis); /* cleanup LCSSA phis */
if (stage->nir->info.uses_resource_info_query)
NIR_PASS(_, stage->nir, ac_nir_lower_resinfo, gfx_level);
@ -577,17 +574,12 @@ radv_postprocess_nir(struct radv_device *device, const struct radv_graphics_stat
NIR_PASS(_, stage->nir, nir_lower_fp16_casts, nir_lower_fp16_split_fp64);
if (stage->nir->info.bit_sizes_int & (8 | 16)) {
if (gfx_level >= GFX8) {
NIR_PASS(_, stage->nir, nir_convert_to_lcssa, true, true);
if (gfx_level >= GFX8)
nir_divergence_analysis(stage->nir);
}
if (nir_lower_bit_size(stage->nir, lower_bit_size_callback, device)) {
NIR_PASS(_, stage->nir, nir_opt_constant_folding);
}
if (gfx_level >= GFX8)
NIR_PASS(_, stage->nir, nir_opt_remove_phis); /* cleanup LCSSA phis */
}
if (gfx_level >= GFX9) {
bool separate_g16 = gfx_level >= GFX10;

1
src/asahi/compiler/agx_compile.c
View file

@ -2969,7 +2969,6 @@ agx_optimize_nir(nir_shader *nir, bool soft_fault, unsigned *preamble_size)
});
NIR_PASS(_, nir, nir_lower_pack);
nir_convert_to_lcssa(nir, true, true);
NIR_PASS_V(nir, nir_divergence_analysis);
bool progress = false;

3
src/compiler/nir/nir_opt_non_uniform_access.c
View file

@ -231,14 +231,11 @@ nir_opt_non_uniform_access_instr(nir_builder *b, nir_instr *instr, UNUSED void *
bool
nir_opt_non_uniform_access(nir_shader *shader)
{
NIR_PASS(_, shader, nir_convert_to_lcssa, true, true);
nir_divergence_analysis(shader);
bool progress = nir_shader_instructions_pass(shader,
nir_opt_non_uniform_access_instr,
nir_metadata_all, NULL);
NIR_PASS(_, shader, nir_opt_remove_phis); /* cleanup LCSSA phis */
return progress;
}

2
src/compiler/nir/nir_opt_varyings.c
View file

@ -4328,8 +4328,6 @@ nir_opt_varyings(nir_shader *producer, nir_shader *consumer, bool spirv,
* divergence information.
*/
if (consumer->info.stage == MESA_SHADER_FRAGMENT) {
/* Required by the divergence analysis. */
NIR_PASS(_, producer, nir_convert_to_lcssa, true, true);
nir_vertex_divergence_analysis(producer);
}

1
src/gallium/drivers/radeonsi/si_shader_nir.c
View file

@ -480,7 +480,6 @@ char *si_finalize_nir(struct pipe_screen *screen, void *nirptr)
if (progress)
si_nir_opts(sscreen, nir, false);
NIR_PASS_V(nir, nir_convert_to_lcssa, true, true); /* required by divergence analysis */
NIR_PASS_V(nir, nir_divergence_analysis); /* to find divergent loops */
/* Must be after divergence analysis. */

9
src/intel/compiler/brw_nir.c
View file

@ -1596,9 +1596,6 @@ brw_vectorize_lower_mem_access(nir_shader *nir,
OPT(nir_opt_load_store_vectorize, &options);
/* Required for nir_divergence_analysis() */
OPT(nir_convert_to_lcssa, true, true);
/* When HW supports block loads, using the divergence analysis, try
* to find uniform SSBO loads and turn them into block loads.
*
@ -1612,7 +1609,6 @@ brw_vectorize_lower_mem_access(nir_shader *nir,
nir_divergence_analysis(nir);
if (OPT(intel_nir_blockify_uniform_loads, compiler->devinfo))
OPT(nir_opt_load_store_vectorize, &options);
OPT(nir_opt_remove_phis);
nir_lower_mem_access_bit_sizes_options mem_access_options = {
.modes = nir_var_mem_ssbo |
@ -1783,7 +1779,6 @@ brw_postprocess_nir(nir_shader *nir, const struct brw_compiler *compiler,
OPT(nir_opt_dead_cf);
bool divergence_analysis_dirty = false;
NIR_PASS(_, nir, nir_convert_to_lcssa, true, true);
NIR_PASS_V(nir, nir_divergence_analysis);
static const nir_lower_subgroups_options subgroups_options = {
@ -1833,16 +1828,12 @@ brw_postprocess_nir(nir_shader *nir, const struct brw_compiler *compiler,
/* Do this only after the last opt_gcm. GCM will undo this lowering. */
if (nir->info.stage == MESA_SHADER_FRAGMENT) {
if (divergence_analysis_dirty) {
NIR_PASS(_, nir, nir_convert_to_lcssa, true, true);
NIR_PASS_V(nir, nir_divergence_analysis);
}
OPT(intel_nir_lower_non_uniform_barycentric_at_sample);
}
/* Clean up LCSSA phis */
OPT(nir_opt_remove_phis);
OPT(nir_lower_bool_to_int32);
OPT(nir_copy_prop);
OPT(nir_opt_dce);

5
src/intel/compiler/elk/elk_nir.c
View file

@ -1464,7 +1464,6 @@ elk_postprocess_nir(nir_shader *nir, const struct elk_compiler *compiler,
OPT(nir_opt_dead_cf);
bool divergence_analysis_dirty = false;
NIR_PASS(_, nir, nir_convert_to_lcssa, true, true);
NIR_PASS_V(nir, nir_divergence_analysis);
/* TODO: Enable nir_opt_uniform_atomics on Gfx7.x too.
@ -1489,16 +1488,12 @@ elk_postprocess_nir(nir_shader *nir, const struct elk_compiler *compiler,
/* Do this only after the last opt_gcm. GCM will undo this lowering. */
if (nir->info.stage == MESA_SHADER_FRAGMENT) {
if (divergence_analysis_dirty) {
NIR_PASS(_, nir, nir_convert_to_lcssa, true, true);
NIR_PASS_V(nir, nir_divergence_analysis);
}
OPT(intel_nir_lower_non_uniform_barycentric_at_sample);
}
/* Clean up LCSSA phis */
OPT(nir_opt_remove_phis);
OPT(nir_lower_bool_to_int32);
OPT(nir_copy_prop);
OPT(nir_opt_dce);

7
src/intel/vulkan/anv_pipeline.c
View file

@ -1113,16 +1113,11 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
NIR_PASS(progress, nir, nir_opt_dce);
} while (progress);
/* Required for nir_divergence_analysis() which is needed for
* anv_nir_lower_ubo_loads.
*/
NIR_PASS(_, nir, nir_convert_to_lcssa, true, true);
/* Needed for anv_nir_lower_ubo_loads. */
nir_divergence_analysis(nir);
NIR_PASS(_, nir, anv_nir_lower_ubo_loads);
NIR_PASS(_, nir, nir_opt_remove_phis);
enum nir_lower_non_uniform_access_type lower_non_uniform_access_types =
nir_lower_non_uniform_texture_access |
nir_lower_non_uniform_image_access |

1
src/panfrost/compiler/bifrost_compile.c
View file

@ -4816,7 +4816,6 @@ bi_optimize_nir(nir_shader *nir, unsigned gpu_id, bool is_blend)
nir->info.images_used[0];
if (any_indirects) {
nir_convert_to_lcssa(nir, true, true);
NIR_PASS_V(nir, nir_divergence_analysis);
NIR_PASS_V(nir, bi_lower_divergent_indirects,
pan_subgroup_size(pan_arch(gpu_id)));