wine/vkd3d
0
0
Fork 0
mirror of https://gitlab.winehq.org/wine/vkd3d.git synced 2024-11-21 16:46:41 -08:00
Code Issues Packages Projects Releases Wiki Activity

vkd3d-shader/hlsl: Allow prioritizing smaller writemasks when allocating signature elements.

Browse Source 
For now this doesn't have effect yet, until semantic allocation is
optimized.
This commit is contained in:
Francisco Casas 2024-09-26 20:03:52 -03:00 committed by Henri Verbeet
parent d562b03c43
commit 88dd082160
Notes: Henri Verbeet 2024-10-22 20:54:15 +02:00 
Approved-by: Elizabeth Figura (@zfigura)
Approved-by: Henri Verbeet (@hverbeet)
Merge-Request: https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/1193
1 changed files with 25 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
libs/vkd3d-shader/hlsl_codegen.c
View File

@ -4459,6 +4459,13 @@ struct register_allocator
/* Total number of registers allocated so far. Used to declare sm4 temp count. */
uint32_t reg_count;
/* Special flag so allocations that can share registers prioritize those
* that will result in smaller writemasks.
* For instance, a single-register allocation would prefer to share a register
* whose .xy components are already allocated (becoming .z) instead of a
* register whose .xyz components are already allocated (becoming .w). */
bool prioritize_smaller_writemasks;
};
static unsigned int get_available_writemask(const struct register_allocator *allocator,
@ -4517,23 +4524,31 @@ static struct hlsl_reg allocate_register(struct hlsl_ctx *ctx, struct register_a
unsigned int component_count, int mode, bool force_align)
{
unsigned int required_size = force_align ? 4 : reg_size;
unsigned int writemask = 0, pref;
struct hlsl_reg ret = {0};
unsigned int writemask;
uint32_t reg_idx;
VKD3D_ASSERT(component_count <= reg_size);
for (reg_idx = 0;; ++reg_idx)
pref = allocator->prioritize_smaller_writemasks ? 4 : required_size;
for (; pref >= required_size; --pref)
{
writemask = get_available_writemask(allocator, first_write, last_read, reg_idx, mode);
if (vkd3d_popcount(writemask) >= required_size)
for (reg_idx = 0; pref == required_size || reg_idx < allocator->reg_count; ++reg_idx)
{
writemask = hlsl_combine_writemasks(writemask, (1u << reg_size) - 1);
break;
unsigned int available_writemask = get_available_writemask(allocator,
first_write, last_read, reg_idx, mode);
if (vkd3d_popcount(available_writemask) >= pref)
{
writemask = hlsl_combine_writemasks(available_writemask, (1u << reg_size) - 1);
break;
}
}
if (writemask)
break;
}
VKD3D_ASSERT(vkd3d_popcount(writemask) == reg_size);
record_allocation(ctx, allocator, reg_idx, writemask, first_write, last_read, mode);
ret.id = reg_idx;
@ -5323,6 +5338,9 @@ static void allocate_semantic_registers(struct hlsl_ctx *ctx, struct hlsl_ir_fun
bool is_patch_constant_func = entry_func == ctx->patch_constant_func;
struct hlsl_ir_var *var;
input_allocator.prioritize_smaller_writemasks = true;
output_allocator.prioritize_smaller_writemasks = true;
LIST_FOR_EACH_ENTRY(var, &entry_func->extern_vars, struct hlsl_ir_var, extern_entry)
{
if (var->is_input_semantic)