mirror of
https://github.com/yuzu-emu/yuzu-mainline.git
synced 2024-12-28 15:25:30 +00:00
shader: Mark SSBOs as written when they are
This commit is contained in:
parent
d819ba4489
commit
12f5f32098
|
@ -6,6 +6,7 @@
|
|||
#include <compare>
|
||||
#include <optional>
|
||||
#include <ranges>
|
||||
#include <map>
|
||||
|
||||
#include <boost/container/flat_set.hpp>
|
||||
#include <boost/container/small_vector.hpp>
|
||||
|
@ -45,6 +46,7 @@ using StorageBufferSet =
|
|||
using StorageInstVector = boost::container::small_vector<StorageInst, 24>;
|
||||
using VisitedBlocks = boost::container::flat_set<IR::Block*, std::less<IR::Block*>,
|
||||
boost::container::small_vector<IR::Block*, 4>>;
|
||||
using StorageWritesMap = std::map<StorageBufferAddr, bool>;
|
||||
|
||||
/// Returns true when the instruction is a global memory instruction
|
||||
bool IsGlobalMemory(const IR::Inst& inst) {
|
||||
|
@ -69,6 +71,22 @@ bool IsGlobalMemory(const IR::Inst& inst) {
|
|||
}
|
||||
}
|
||||
|
||||
/// Returns true when the instruction is a global memory instruction
|
||||
bool IsGlobalMemoryWrite(const IR::Inst& inst) {
|
||||
switch (inst.Opcode()) {
|
||||
case IR::Opcode::WriteGlobalS8:
|
||||
case IR::Opcode::WriteGlobalU8:
|
||||
case IR::Opcode::WriteGlobalS16:
|
||||
case IR::Opcode::WriteGlobalU16:
|
||||
case IR::Opcode::WriteGlobal32:
|
||||
case IR::Opcode::WriteGlobal64:
|
||||
case IR::Opcode::WriteGlobal128:
|
||||
return true;
|
||||
default:
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
/// Converts a global memory opcode to its storage buffer equivalent
|
||||
IR::Opcode GlobalToStorage(IR::Opcode opcode) {
|
||||
switch (opcode) {
|
||||
|
@ -248,7 +266,7 @@ std::optional<StorageBufferAddr> Track(IR::Block* block, const IR::Value& value,
|
|||
|
||||
/// Collects the storage buffer used by a global memory instruction and the instruction itself
|
||||
void CollectStorageBuffers(IR::Block& block, IR::Inst& inst, StorageBufferSet& storage_buffer_set,
|
||||
StorageInstVector& to_replace) {
|
||||
StorageInstVector& to_replace, StorageWritesMap& writes_map) {
|
||||
// NVN puts storage buffers in a specific range, we have to bias towards these addresses to
|
||||
// avoid getting false positives
|
||||
static constexpr Bias nvn_bias{
|
||||
|
@ -277,6 +295,13 @@ void CollectStorageBuffers(IR::Block& block, IR::Inst& inst, StorageBufferSet& s
|
|||
}
|
||||
}
|
||||
// Collect storage buffer and the instruction
|
||||
const bool is_a_write = IsGlobalMemoryWrite(inst);
|
||||
auto it = writes_map.find(*storage_buffer);
|
||||
if (it == writes_map.end()) {
|
||||
writes_map[*storage_buffer] = is_a_write;
|
||||
} else {
|
||||
it->second = it->second || is_a_write;
|
||||
}
|
||||
storage_buffer_set.insert(*storage_buffer);
|
||||
to_replace.push_back(StorageInst{
|
||||
.storage_buffer{*storage_buffer},
|
||||
|
@ -350,13 +375,14 @@ void Replace(IR::Block& block, IR::Inst& inst, const IR::U32& storage_index,
|
|||
void GlobalMemoryToStorageBufferPass(IR::Program& program) {
|
||||
StorageBufferSet storage_buffers;
|
||||
StorageInstVector to_replace;
|
||||
StorageWritesMap writes_map;
|
||||
|
||||
for (IR::Block* const block : program.post_order_blocks) {
|
||||
for (IR::Inst& inst : block->Instructions()) {
|
||||
if (!IsGlobalMemory(inst)) {
|
||||
continue;
|
||||
}
|
||||
CollectStorageBuffers(*block, inst, storage_buffers, to_replace);
|
||||
CollectStorageBuffers(*block, inst, storage_buffers, to_replace, writes_map);
|
||||
}
|
||||
}
|
||||
Info& info{program.info};
|
||||
|
@ -366,6 +392,7 @@ void GlobalMemoryToStorageBufferPass(IR::Program& program) {
|
|||
.cbuf_index{storage_buffer.index},
|
||||
.cbuf_offset{storage_buffer.offset},
|
||||
.count{1},
|
||||
.is_written{writes_map[storage_buffer]},
|
||||
});
|
||||
++storage_index;
|
||||
}
|
||||
|
|
|
@ -59,6 +59,7 @@ struct StorageBufferDescriptor {
|
|||
u32 cbuf_index;
|
||||
u32 cbuf_offset;
|
||||
u32 count;
|
||||
bool is_written;
|
||||
};
|
||||
|
||||
struct Info {
|
||||
|
|
|
@ -75,7 +75,7 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
|
|||
size_t ssbo_index{};
|
||||
for (const auto& desc : info.storage_buffers_descriptors) {
|
||||
ASSERT(desc.count == 1);
|
||||
buffer_cache.BindComputeStorageBuffer(ssbo_index, desc.cbuf_index, desc.cbuf_offset, true);
|
||||
buffer_cache.BindComputeStorageBuffer(ssbo_index, desc.cbuf_index, desc.cbuf_offset, desc.is_written);
|
||||
++ssbo_index;
|
||||
}
|
||||
buffer_cache.UpdateComputeBuffers();
|
||||
|
|
|
@ -163,7 +163,7 @@ void GraphicsPipeline::Configure(bool is_indexed) {
|
|||
for (const auto& desc : info.storage_buffers_descriptors) {
|
||||
ASSERT(desc.count == 1);
|
||||
buffer_cache.BindGraphicsStorageBuffer(stage, index, desc.cbuf_index, desc.cbuf_offset,
|
||||
true);
|
||||
desc.is_written);
|
||||
++index;
|
||||
}
|
||||
const auto& cbufs{maxwell3d.state.shader_stages[stage].const_buffers};
|
||||
|
|
Loading…
Reference in a new issue