wine/vkd3d
0
0
Fork 0
mirror of https://gitlab.winehq.org/wine/vkd3d.git synced 2025-04-13 05:43:18 -07:00
Code Issues Packages Projects Releases Wiki Activity

vkd3d-shader/hlsl: Introduce a compiler pass to vectorize stores.

Browse Source
This commit is contained in:
Elizabeth Figura
2024-10-30 11:51:07 -05:00
committed by Henri Verbeet
parent 1a999f74fc
commit f576ecc992
Notes: Henri Verbeet 2025-04-03 20:34:10 +02:00 
Approved-by: Francisco Casas (@fcasas)
Approved-by: Henri Verbeet (@hverbeet)
Merge-Request: https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/1443
3 changed files with 364 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
libs/vkd3d-shader/hlsl.c
View File

@@ -1588,6 +1588,43 @@ void hlsl_block_add_simple_store(struct hlsl_ctx *ctx, struct hlsl_block *block,
hlsl_block_add_store_index(ctx, block, &lhs_deref, NULL, rhs, 0, &rhs->loc);
}
static struct hlsl_ir_node *hlsl_new_store_parent(struct hlsl_ctx *ctx,
const struct hlsl_deref *lhs, unsigned int path_len, struct hlsl_ir_node *rhs,
unsigned int writemask, const struct vkd3d_shader_location *loc)
{
struct hlsl_ir_store *store;
VKD3D_ASSERT(!hlsl_deref_is_lowered(lhs));
VKD3D_ASSERT(lhs->path_len >= path_len);
if (!(store = hlsl_alloc(ctx, sizeof(*store))))
return NULL;
init_node(&store->node, HLSL_IR_STORE, NULL, loc);
if (!hlsl_init_deref(ctx, &store->lhs, lhs->var, path_len))
{
vkd3d_free(store);
return NULL;
}
for (unsigned int i = 0; i < path_len; ++i)
hlsl_src_from_node(&store->lhs.path[i], lhs->path[i].node);
hlsl_src_from_node(&store->rhs, rhs);
if (!writemask && type_is_single_reg(rhs->data_type))
writemask = (1 << rhs->data_type->e.numeric.dimx) - 1;
store->writemask = writemask;
return &store->node;
}
void hlsl_block_add_store_parent(struct hlsl_ctx *ctx, struct hlsl_block *block,
const struct hlsl_deref *lhs, unsigned int path_len, struct hlsl_ir_node *rhs,
unsigned int writemask, const struct vkd3d_shader_location *loc)
{
append_new_instr(ctx, block, hlsl_new_store_parent(ctx, lhs, path_len, rhs, writemask, loc));
}
void hlsl_block_add_store_component(struct hlsl_ctx *ctx, struct hlsl_block *block,
const struct hlsl_deref *lhs, unsigned int comp, struct hlsl_ir_node *rhs)
{

3
libs/vkd3d-shader/hlsl.h
View File

@@ -1565,6 +1565,9 @@ void hlsl_block_add_store_component(struct hlsl_ctx *ctx, struct hlsl_block *blo
void hlsl_block_add_store_index(struct hlsl_ctx *ctx, struct hlsl_block *block,
const struct hlsl_deref *lhs, struct hlsl_ir_node *idx, struct hlsl_ir_node *rhs,
unsigned int writemask, const struct vkd3d_shader_location *loc);
void hlsl_block_add_store_parent(struct hlsl_ctx *ctx, struct hlsl_block *block,
const struct hlsl_deref *lhs, unsigned int path_len, struct hlsl_ir_node *rhs,
unsigned int writemask, const struct vkd3d_shader_location *loc);
struct hlsl_ir_node *hlsl_block_add_swizzle(struct hlsl_ctx *ctx, struct hlsl_block *block, uint32_t s,
unsigned int width, struct hlsl_ir_node *val, const struct vkd3d_shader_location *loc);
struct hlsl_ir_node *hlsl_block_add_uint_constant(struct hlsl_ctx *ctx, struct hlsl_block *block,

324
libs/vkd3d-shader/hlsl_codegen.c
View File

@@ -2716,6 +2716,329 @@ out:
return progress;
}
struct vectorize_stores_state
{
struct vectorizable_stores_group
{
struct hlsl_block *block;
/* We handle overlapping stores, because it's not really easier not to.
* In theory, then, we could collect an arbitrary number of stores here.
*
* In practice, overlapping stores are unlikely, and of course at most
* 4 stores can appear without overlap. Therefore, for simplicity, we
* just use a fixed array of 4.
*
* Since computing the writemask requires traversing the deref, and we
* need to do that anyway, we store it here for convenience. */
struct hlsl_ir_store *stores[4];
unsigned int path_len;
uint8_t writemasks[4];
uint8_t store_count;
bool dirty;
} *groups;
size_t count, capacity;
};
/* This must be a store to a subsection of a vector.
* In theory we can also vectorize stores to packed struct fields,
* but this requires target-specific knowledge and is probably best left
* to a VSIR pass. */
static bool can_vectorize_store(struct hlsl_ctx *ctx, struct hlsl_ir_store *store,
unsigned int *path_len, uint8_t *writemask)
{
struct hlsl_type *type = store->lhs.var->data_type;
unsigned int i;
if (store->rhs.node->data_type->class > HLSL_CLASS_VECTOR)
return false;
if (type->class == HLSL_CLASS_SCALAR)
return false;
for (i = 0; type->class != HLSL_CLASS_VECTOR && i < store->lhs.path_len; ++i)
type = hlsl_get_element_type_from_path_index(ctx, type, store->lhs.path[i].node);
if (type->class != HLSL_CLASS_VECTOR)
return false;
*path_len = i;
if (i < store->lhs.path_len)
{
struct hlsl_ir_constant *c;
/* This is a store to a scalar component of a vector, achieved via
* indexing. */
if (store->lhs.path[i].node->type != HLSL_IR_CONSTANT)
return false;
c = hlsl_ir_constant(store->lhs.path[i].node);
*writemask = (1u << c->value.u[0].u);
}
else
{
*writemask = store->writemask;
}
return true;
}
static bool derefs_are_same_vector(struct hlsl_ctx *ctx, const struct hlsl_deref *a, const struct hlsl_deref *b)
{
struct hlsl_type *type = a->var->data_type;
if (a->var != b->var)
return false;
for (unsigned int i = 0; type->class != HLSL_CLASS_VECTOR && i < a->path_len && i < b->path_len; ++i)
{
if (a->path[i].node != b->path[i].node)
return false;
type = hlsl_get_element_type_from_path_index(ctx, type, a->path[i].node);
}
return true;
}
static void record_vectorizable_store(struct hlsl_ctx *ctx, struct hlsl_block *block,
struct hlsl_ir_store *store, struct vectorize_stores_state *state)
{
unsigned int path_len;
uint8_t writemask;
if (!can_vectorize_store(ctx, store, &path_len, &writemask))
{
/* In the case of a dynamically indexed vector, we must invalidate
* any groups that statically index the same vector.
* For the sake of expediency, we go one step further and invalidate
* any groups that store to the same variable.
* (We also don't check that that was the reason why this store isn't
* vectorizable.)
* We could be more granular, but we'll defer that until it comes
* up in practice. */
for (size_t i = 0; i < state->count; ++i)
{
if (state->groups[i].stores[0]->lhs.var == store->lhs.var)
state->groups[i].dirty = true;
}
return;
}
for (size_t i = 0; i < state->count; ++i)
{
struct vectorizable_stores_group *group = &state->groups[i];
struct hlsl_ir_store *other = group->stores[0];
if (group->dirty)
continue;
if (derefs_are_same_vector(ctx, &store->lhs, &other->lhs))
{
/* Stores must be in the same CFG block. If they're not,
* they're not executed in exactly the same flow, and
* therefore can't be vectorized. */
if (group->block == block
&& is_same_vectorizable_source(store->rhs.node, other->rhs.node))
{
if (group->store_count < ARRAY_SIZE(group->stores))
{
group->stores[group->store_count] = store;
group->writemasks[group->store_count] = writemask;
++group->store_count;
return;
}
}
else
{
/* A store to the same vector with a different source, or in
* a different CFG block, invalidates any earlier store.
*
* A store to a component which *contains* the vector in
* question would also invalidate, but we should have split all
* of those by the time we get here. */
group->dirty = true;
/* Note that we do exit this loop early if we find a store A we
* can vectorize with, but that's fine. If there was a store B
* also in the state that we can't vectorize with, it would
* already have invalidated A. */
}
}
else
{
/* This could still be a store to the same vector, if e.g. the
* vector is part of a dynamically indexed array, or the path has
* two equivalent instructions which refer to the same component.
* [CSE may help with the latter, but we don't have it yet,
* and we shouldn't depend on it anyway.]
* For the sake of expediency, we just invalidate it if it refers
* to the same variable at all.
* As above, we could be more granular, but we'll defer that until
* it comes up in practice. */
if (store->lhs.var == other->lhs.var)
group->dirty = true;
/* As above, we don't need to worry about exiting the loop early. */
}
}
if (!hlsl_array_reserve(ctx, (void **)&state->groups,
&state->capacity, state->count + 1, sizeof(*state->groups)))
return;
state->groups[state->count].block = block;
state->groups[state->count].stores[0] = store;
state->groups[state->count].path_len = path_len;
state->groups[state->count].writemasks[0] = writemask;
state->groups[state->count].store_count = 1;
state->groups[state->count].dirty = false;
++state->count;
}
static void find_vectorizable_store_groups(struct hlsl_ctx *ctx, struct hlsl_block *block,
struct vectorize_stores_state *state)
{
struct hlsl_ir_node *instr;
LIST_FOR_EACH_ENTRY(instr, &block->instrs, struct hlsl_ir_node, entry)
{
if (instr->type == HLSL_IR_STORE)
{
record_vectorizable_store(ctx, block, hlsl_ir_store(instr), state);
}
else if (instr->type == HLSL_IR_LOAD)
{
struct hlsl_ir_var *var = hlsl_ir_load(instr)->src.var;
/* By vectorizing store A with store B, we are effectively moving
* store A down to happen at the same time as store B.
* If there was a load of the same variable between the two, this
* would be incorrect.
* Therefore invalidate all stores to this variable. As above, we
* could be more granular if necessary. */
for (unsigned int i = 0; i < state->count; ++i)
{
if (state->groups[i].stores[0]->lhs.var == var)
state->groups[i].dirty = true;
}
}
else if (instr->type == HLSL_IR_IF)
{
struct hlsl_ir_if *iff = hlsl_ir_if(instr);
find_vectorizable_store_groups(ctx, &iff->then_block, state);
find_vectorizable_store_groups(ctx, &iff->else_block, state);
}
else if (instr->type == HLSL_IR_LOOP)
{
find_vectorizable_store_groups(ctx, &hlsl_ir_loop(instr)->body, state);
}
else if (instr->type == HLSL_IR_SWITCH)
{
struct hlsl_ir_switch *s = hlsl_ir_switch(instr);
struct hlsl_ir_switch_case *c;
LIST_FOR_EACH_ENTRY(c, &s->cases, struct hlsl_ir_switch_case, entry)
find_vectorizable_store_groups(ctx, &c->body, state);
}
}
}
/* Combine sequences like
*
* 2: @1.yw
* 3: @1.zy
* 4: var.xy = @2
* 5: var.yw = @3
*
* to
*
* 2: @1.yzy
* 5: var.xyw = @2
*
* There are a lot of gotchas here. We need to make sure the two stores are to
* the same vector (which may be embedded in a complex variable), that they're
* always executed in the same control flow, and that there aren't any other
* stores or loads on the same vector in the middle. */
static bool vectorize_stores(struct hlsl_ctx *ctx, struct hlsl_block *block)
{
struct vectorize_stores_state state = {0};
bool progress = false;
find_vectorizable_store_groups(ctx, block, &state);
for (unsigned int i = 0; i < state.count; ++i)
{
struct vectorizable_stores_group *group = &state.groups[i];
uint32_t new_swizzle = 0, new_writemask = 0;
struct hlsl_ir_node *new_rhs, *value;
uint32_t swizzle_components[4];
unsigned int component_count;
struct hlsl_ir_store *store;
struct hlsl_block new_block;
if (group->store_count == 1)
continue;
hlsl_block_init(&new_block);
/* Compute the swizzle components. */
for (unsigned int j = 0; j < group->store_count; ++j)
{
unsigned int writemask = group->writemasks[j];
uint32_t rhs_swizzle;
store = group->stores[j];
if (store->rhs.node->type == HLSL_IR_SWIZZLE)
rhs_swizzle = hlsl_ir_swizzle(store->rhs.node)->u.vector;
else
rhs_swizzle = HLSL_SWIZZLE(X, Y, Z, W);
component_count = 0;
for (unsigned int k = 0; k < 4; ++k)
{
if (writemask & (1u << k))
swizzle_components[k] = hlsl_swizzle_get_component(rhs_swizzle, component_count++);
}
new_writemask |= writemask;
}
/* Construct the new swizzle. */
component_count = 0;
for (unsigned int k = 0; k < 4; ++k)
{
if (new_writemask & (1u << k))
hlsl_swizzle_set_component(&new_swizzle, component_count++, swizzle_components[k]);
}
store = group->stores[0];
value = store->rhs.node;
if (value->type == HLSL_IR_SWIZZLE)
value = hlsl_ir_swizzle(value)->val.node;
new_rhs = hlsl_block_add_swizzle(ctx, &new_block, new_swizzle, component_count, value, &value->loc);
hlsl_block_add_store_parent(ctx, &new_block, &store->lhs,
group->path_len, new_rhs, new_writemask, &store->node.loc);
TRACE("Combining %u stores to %s.\n", group->store_count, store->lhs.var->name);
list_move_before(&group->stores[group->store_count - 1]->node.entry, &new_block.instrs);
for (unsigned int j = 0; j < group->store_count; ++j)
{
list_remove(&group->stores[j]->node.entry);
hlsl_free_instr(&group->stores[j]->node);
}
progress = true;
}
vkd3d_free(state.groups);
return progress;
}
static enum validation_result validate_component_index_range_from_deref(struct hlsl_ctx *ctx,
const struct hlsl_deref *deref)
{
@@ -13017,6 +13340,7 @@ static void process_entry_function(struct hlsl_ctx *ctx,
progress |= hlsl_transform_ir(ctx, dce, body, NULL);
progress |= hlsl_transform_ir(ctx, fold_swizzle_chains, body, NULL);
progress |= hlsl_transform_ir(ctx, remove_trivial_swizzles, body, NULL);
progress |= vectorize_stores(ctx, body);
} while (progress);
hlsl_transform_ir(ctx, track_object_components_sampler_dim, body, NULL);