fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-03 05:38:16 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

spirv2dxil: Lower large temps to scratch

Browse source 
WARP has a temp register limit, and the control flow needed to convert
indirect to direct accesses on large temps ends up bloating shaders massively.
We can just go ahead and spill these large temps to scratch, which maps
to an alloca in DXIL.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22787>
This commit is contained in:
Jesse Natalie 2023-05-01 16:18:22 -07:00 • committed by Marge Bot
parent a4ce095bad
commit a6ac1f5fbd
1 changed files with 30 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

30
src/microsoft/spirv_to_dxil/dxil_spirv_nir.c
View file

@ -129,6 +129,34 @@ shared_var_info(const struct glsl_type* type, unsigned* size, unsigned* align)
*align = comp_size;
}
static void
temp_var_info(const struct glsl_type* type, unsigned* size, unsigned* align)
{
uint32_t base_size, base_align;
switch (glsl_get_base_type(type)) {
case GLSL_TYPE_ARRAY:
temp_var_info(glsl_get_array_element(type), &base_size, align);
*size = base_size * glsl_array_size(type);
break;
case GLSL_TYPE_STRUCT:
case GLSL_TYPE_INTERFACE:
*size = 0;
*align = 0;
for (uint32_t i = 0; i < glsl_get_length(type); ++i) {
temp_var_info(glsl_get_struct_field(type, i), &base_size, &base_align);
*size = ALIGN_POT(*size, base_align) + base_size;
*align = MAX2(*align, base_align);
}
break;
default:
glsl_get_natural_size_align_bytes(type, &base_size, &base_align);
*align = MAX2(base_align, 4);
*size = ALIGN_POT(base_size, *align);
break;
}
}
static nir_variable *
add_runtime_data_var(nir_shader *nir, unsigned desc_set, unsigned binding)
{
@ -1037,6 +1065,8 @@ dxil_spirv_nir_passes(nir_shader *nir,
shared_var_info);
}
NIR_PASS_V(nir, dxil_nir_split_unaligned_loads_stores, nir_var_mem_shared);
NIR_PASS_V(nir, nir_lower_vars_to_scratch, nir_var_function_temp | nir_var_shader_temp,
256 /* arbitrary */, temp_var_info);
NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_shared,
nir_address_format_32bit_offset);