Mesa (main): nir: add support for forced sampler indirect loop unrolling
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Tue May 17 02:41:46 UTC 2022
Module: Mesa
Branch: main
Commit: ff8ddcb23e406a986434a39c3a51857f432a131f
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=ff8ddcb23e406a986434a39c3a51857f432a131f
Author: Timothy Arceri <tarceri at itsqueeze.com>
Date: Wed May 11 20:12:56 2022 +1000
nir: add support for forced sampler indirect loop unrolling
Some drivers don't support these indirects and therefore require
loop unrolling if a shader uses a loop induction variable to
access a sampler array.
Here we add a new nir shader compiler option that drivers can set,
this will be the equivalent of the EmitNoIndirectSampler setting
used in the GLSL IR unrolling pass.
Reviewed-by: Emma Anholt <emma at anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16543>
---
src/compiler/nir/nir.h | 8 +++++-
src/compiler/nir/nir_inline_uniforms.c | 3 ++-
src/compiler/nir/nir_loop_analyze.c | 45 +++++++++++++++++++++++++++-------
src/compiler/nir/nir_metadata.c | 8 +++++-
src/compiler/nir/nir_opt_gcm.c | 3 ++-
src/compiler/nir/nir_opt_loop_unroll.c | 10 +++++---
6 files changed, 61 insertions(+), 16 deletions(-)
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 985f07c7d3c..968aa78ade3 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -3578,6 +3578,11 @@ typedef struct nir_shader_compiler_options {
* vectorized IO can pack more varyings when linking. */
bool linker_ignore_precision;
+ /* Specifies if indirect sampler array access will trigger forced loop
+ * unrolling.
+ */
+ bool force_indirect_unrolling_sampler;
+
/**
* Specifies which type of indirectly accessed variables should force
* loop unrolling.
@@ -5312,7 +5317,8 @@ void nir_live_ssa_defs_impl(nir_function_impl *impl);
const BITSET_WORD *nir_get_live_ssa_defs(nir_cursor cursor, void *mem_ctx);
void nir_loop_analyze_impl(nir_function_impl *impl,
- nir_variable_mode indirect_mask);
+ nir_variable_mode indirect_mask,
+ bool force_unroll_sampler_indirect);
bool nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b);
diff --git a/src/compiler/nir/nir_inline_uniforms.c b/src/compiler/nir/nir_inline_uniforms.c
index 652b9534691..57c29fc07bd 100644
--- a/src/compiler/nir/nir_inline_uniforms.c
+++ b/src/compiler/nir/nir_inline_uniforms.c
@@ -321,7 +321,8 @@ nir_find_inlinable_uniforms(nir_shader *shader)
nir_foreach_function(function, shader) {
if (function->impl) {
- nir_metadata_require(function->impl, nir_metadata_loop_analysis, nir_var_all);
+ nir_metadata_require(function->impl, nir_metadata_loop_analysis,
+ nir_var_all, false);
foreach_list_typed(nir_cf_node, node, node, &function->impl->body)
process_node(node, NULL, uni_offsets, &num_offsets);
diff --git a/src/compiler/nir/nir_loop_analyze.c b/src/compiler/nir/nir_loop_analyze.c
index c35a129d428..3b54a068429 100644
--- a/src/compiler/nir/nir_loop_analyze.c
+++ b/src/compiler/nir/nir_loop_analyze.c
@@ -77,6 +77,7 @@ typedef struct {
nir_variable_mode indirect_mask;
+ bool force_unroll_sampler_indirect;
} loop_info_state;
static nir_loop_variable *
@@ -601,6 +602,7 @@ find_array_access_via_induction(loop_info_state *state,
*array_index_out = array_index;
nir_deref_instr *parent = nir_deref_instr_parent(d);
+
if (glsl_type_is_array_or_matrix(parent->type)) {
return glsl_get_length(parent->type);
} else {
@@ -1208,7 +1210,8 @@ find_trip_count(loop_info_state *state, unsigned execution_mode)
}
static bool
-force_unroll_array_access(loop_info_state *state, nir_deref_instr *deref)
+force_unroll_array_access(loop_info_state *state, nir_deref_instr *deref,
+ bool contains_sampler)
{
unsigned array_size = find_array_access_via_induction(state, deref, NULL);
if (array_size) {
@@ -1221,6 +1224,9 @@ force_unroll_array_access(loop_info_state *state, nir_deref_instr *deref)
if (nir_deref_mode_must_be(deref, state->indirect_mask))
return true;
+
+ if (contains_sampler && state->force_unroll_sampler_indirect)
+ return true;
}
return false;
@@ -1230,6 +1236,22 @@ static bool
force_unroll_heuristics(loop_info_state *state, nir_block *block)
{
nir_foreach_instr(instr, block) {
+ if (instr->type == nir_instr_type_tex) {
+ nir_tex_instr *tex_instr = nir_instr_as_tex(instr);
+ int sampler_idx =
+ nir_tex_instr_src_index(tex_instr,
+ nir_tex_src_sampler_deref);
+
+
+ if (sampler_idx >= 0) {
+ nir_deref_instr *deref =
+ nir_instr_as_deref(tex_instr->src[sampler_idx].src.ssa->parent_instr);
+ if (force_unroll_array_access(state, deref, true))
+ return true;
+ }
+ }
+
+
if (instr->type != nir_instr_type_intrinsic)
continue;
@@ -1242,12 +1264,14 @@ force_unroll_heuristics(loop_info_state *state, nir_block *block)
intrin->intrinsic == nir_intrinsic_store_deref ||
intrin->intrinsic == nir_intrinsic_copy_deref) {
if (force_unroll_array_access(state,
- nir_src_as_deref(intrin->src[0])))
+ nir_src_as_deref(intrin->src[0]),
+ false))
return true;
if (intrin->intrinsic == nir_intrinsic_copy_deref &&
force_unroll_array_access(state,
- nir_src_as_deref(intrin->src[1])))
+ nir_src_as_deref(intrin->src[1]),
+ false))
return true;
}
}
@@ -1343,7 +1367,8 @@ initialize_loop_info_state(nir_loop *loop, void *mem_ctx,
}
static void
-process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask)
+process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask,
+ bool force_unroll_sampler_indirect)
{
switch (cf_node->type) {
case nir_cf_node_block:
@@ -1351,15 +1376,15 @@ process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask)
case nir_cf_node_if: {
nir_if *if_stmt = nir_cf_node_as_if(cf_node);
foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->then_list)
- process_loops(nested_node, indirect_mask);
+ process_loops(nested_node, indirect_mask, force_unroll_sampler_indirect);
foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->else_list)
- process_loops(nested_node, indirect_mask);
+ process_loops(nested_node, indirect_mask, force_unroll_sampler_indirect);
return;
}
case nir_cf_node_loop: {
nir_loop *loop = nir_cf_node_as_loop(cf_node);
foreach_list_typed(nir_cf_node, nested_node, node, &loop->body)
- process_loops(nested_node, indirect_mask);
+ process_loops(nested_node, indirect_mask, force_unroll_sampler_indirect);
break;
}
default:
@@ -1372,6 +1397,7 @@ process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask)
loop_info_state *state = initialize_loop_info_state(loop, mem_ctx, impl);
state->indirect_mask = indirect_mask;
+ state->force_unroll_sampler_indirect = force_unroll_sampler_indirect;
get_loop_info(state, impl);
@@ -1380,9 +1406,10 @@ process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask)
void
nir_loop_analyze_impl(nir_function_impl *impl,
- nir_variable_mode indirect_mask)
+ nir_variable_mode indirect_mask,
+ bool force_unroll_sampler_indirect)
{
nir_index_ssa_defs(impl);
foreach_list_typed(nir_cf_node, node, node, &impl->body)
- process_loops(node, indirect_mask);
+ process_loops(node, indirect_mask, force_unroll_sampler_indirect);
}
diff --git a/src/compiler/nir/nir_metadata.c b/src/compiler/nir/nir_metadata.c
index 0cc8f0624d0..4c47ef1efb4 100644
--- a/src/compiler/nir/nir_metadata.c
+++ b/src/compiler/nir/nir_metadata.c
@@ -46,7 +46,13 @@ nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...)
if (NEEDS_UPDATE(nir_metadata_loop_analysis)) {
va_list ap;
va_start(ap, required);
- nir_loop_analyze_impl(impl, va_arg(ap, nir_variable_mode));
+ /* !! Warning !! Do not move these va_arg() call directly to
+ * nir_loop_analyze_impl() as parameters because the execution order will
+ * become undefined.
+ */
+ nir_variable_mode mode = va_arg(ap, nir_variable_mode);
+ int force_unroll_sampler_indirect = va_arg(ap, int);
+ nir_loop_analyze_impl(impl, mode, force_unroll_sampler_indirect);
va_end(ap);
}
diff --git a/src/compiler/nir/nir_opt_gcm.c b/src/compiler/nir/nir_opt_gcm.c
index 529ca6439e5..694ee641e42 100644
--- a/src/compiler/nir/nir_opt_gcm.c
+++ b/src/compiler/nir/nir_opt_gcm.c
@@ -735,7 +735,8 @@ opt_gcm_impl(nir_shader *shader, nir_function_impl *impl, bool value_number)
nir_metadata_require(impl, nir_metadata_block_index |
nir_metadata_dominance);
nir_metadata_require(impl, nir_metadata_loop_analysis,
- shader->options->force_indirect_unrolling);
+ shader->options->force_indirect_unrolling,
+ shader->options->force_indirect_unrolling_sampler);
/* A previous pass may have left pass_flags dirty, so clear it all out. */
nir_foreach_block(block, impl)
diff --git a/src/compiler/nir/nir_opt_loop_unroll.c b/src/compiler/nir/nir_opt_loop_unroll.c
index 49c696dc514..d101d74c5d7 100644
--- a/src/compiler/nir/nir_opt_loop_unroll.c
+++ b/src/compiler/nir/nir_opt_loop_unroll.c
@@ -1090,10 +1090,12 @@ exit:
static bool
nir_opt_loop_unroll_impl(nir_function_impl *impl,
- nir_variable_mode indirect_mask)
+ nir_variable_mode indirect_mask,
+ bool force_unroll_sampler_indirect)
{
bool progress = false;
- nir_metadata_require(impl, nir_metadata_loop_analysis, indirect_mask);
+ nir_metadata_require(impl, nir_metadata_loop_analysis, indirect_mask,
+ (int) force_unroll_sampler_indirect);
nir_metadata_require(impl, nir_metadata_block_index);
bool has_nested_loop = false;
@@ -1119,10 +1121,12 @@ nir_opt_loop_unroll(nir_shader *shader)
{
bool progress = false;
+ bool force_unroll_sampler_indirect = shader->options->force_indirect_unrolling_sampler;
nir_variable_mode indirect_mask = shader->options->force_indirect_unrolling;
nir_foreach_function(function, shader) {
if (function->impl) {
- progress |= nir_opt_loop_unroll_impl(function->impl, indirect_mask);
+ progress |= nir_opt_loop_unroll_impl(function->impl, indirect_mask,
+ force_unroll_sampler_indirect);
}
}
return progress;
More information about the mesa-commit
mailing list