OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

vtn: handle struct kernel arguments passed by value

Browse Source 
Due to LLVM ABI reasons the SPIRV-LLVM-Translator always uses pointers to
private memory for struct function parameters. This includes kernel entry
points.

However technically it's also legal to pass those parameters by value
according to the OpenCL SPIR-V Env spec.

One compiler making use of this is e.g. artic based on Thorin.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12149
Cc: mesa-stable
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
(cherry picked from commit d0560f59ce)

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32261>
This commit is contained in:
Karol Herbst
2024-11-14 13:14:10 +01:00
committed by Dylan Baker
parent f4d83eb508
commit 9728a9075c
2 changed files with 40 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.pick_status.json
View File

@@ -134,7 +134,7 @@
"description": "vtn: handle struct kernel arguments passed by value",
"nominated": true,
"nomination_type": 1,
"resolution": 0,
"resolution": 1,
"main_sha": null,
"because_sha": null,
"notes": null

43
src/compiler/spirv/spirv_to_nir.c
View File

@@ -6767,6 +6767,27 @@ vtn_create_builder(const uint32_t *words, size_t word_count,
return NULL;
}
/* See glsl_type_add_to_function_params and vtn_ssa_value_add_to_call_params */
static void
vtn_emit_kernel_entry_point_wrapper_struct_param(struct nir_builder *b,
nir_deref_instr *deref,
nir_call_instr *call,
unsigned *idx)
{
if (glsl_type_is_vector_or_scalar(deref->type)) {
call->params[(*idx)++] = nir_src_for_ssa(nir_load_deref(b, deref));
} else {
unsigned elems = glsl_get_length(deref->type);
for (unsigned i = 0; i < elems; i++) {
nir_deref_instr *child_deref = glsl_type_is_struct(deref->type)
? nir_build_deref_struct(b, deref, i)
: nir_build_deref_array_imm(b, deref, i);
vtn_emit_kernel_entry_point_wrapper_struct_param(b, child_deref, call,
idx);
}
}
}
static nir_function *
vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b,
nir_function *entry_point)
@@ -6785,7 +6806,8 @@ vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b,
nir_call_instr *call = nir_call_instr_create(b->nb.shader, entry_point);
for (unsigned i = 0; i < entry_point->num_params; ++i) {
unsigned call_idx = 0;
for (unsigned i = 0; i < b->entry_point->func->type->length; ++i) {
struct vtn_type *param_type = b->entry_point->func->type->params[i];
b->shader->info.cs.has_variable_shared_mem |=
@@ -6826,17 +6848,30 @@ vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b,
nir_variable *copy_var =
nir_local_variable_create(impl, in_var->type, "copy_in");
nir_copy_var(&b->nb, copy_var, in_var);
call->params[i] =
call->params[call_idx++] =
nir_src_for_ssa(&nir_build_deref_var(&b->nb, copy_var)->def);
} else if (param_type->base_type == vtn_base_type_image ||
param_type->base_type == vtn_base_type_sampler) {
/* Don't load the var, just pass a deref of it */
call->params[i] = nir_src_for_ssa(&nir_build_deref_var(&b->nb, in_var)->def);
call->params[call_idx++] =
nir_src_for_ssa(&nir_build_deref_var(&b->nb, in_var)->def);
} else if (param_type->base_type == vtn_base_type_struct) {
/* We decompose struct and array parameters in vtn, so we'll need to
* handle it here explicitly.
* We have to keep the arguments on the actual entry point intact,
* because the runtimes rely on it to match the SPIR-V.
*/
nir_deref_instr *deref = nir_build_deref_var(&b->nb, in_var);
vtn_emit_kernel_entry_point_wrapper_struct_param(&b->nb, deref, call,
&call_idx);
} else {
call->params[i] = nir_src_for_ssa(nir_load_var(&b->nb, in_var));
call->params[call_idx++] =
nir_src_for_ssa(nir_load_var(&b->nb, in_var));
}
}
assert(call_idx == entry_point->num_params);
nir_builder_instr_insert(&b->nb, &call->instr);
return main_entry_point;