shader: share formatted stores for GPU-selected signed byte records

This commit is contained in:
nmzik
2026-10-14 17:10:02 +02:00
parent 04cb4c6553
commit 6d6da57a95
5 changed files with 329 additions and 38 deletions
@@ -249,11 +249,18 @@ bool NormalizeIndirectBuffer(DescriptorValue& value, bool indexed, const BufferR
descriptor.OutOfBounds() != (indexed ? 0u : 3u)) {
return SpecializationFail("indirect buffers require exact linear index or byte bounds");
}
if (formatted_read) {
if (resource.formatted) {
const auto format = Format::GetFormatInfo(descriptor.Format());
if (!indexed || format.type == Format::ComponentType::Unknown)
return SpecializationFail("indirect formatted loads require a supported indexed format");
if (descriptor.Base48() == 0u || descriptor.Stride() == 0u || descriptor.NumRecords() == 0u) {
return SpecializationFail("indirect formatted buffers require a supported indexed format");
if (!formatted_read &&
((descriptor.Format() != Prospero::BufferFormat::k32UInt &&
descriptor.Format() != Prospero::BufferFormat::k8SInt) ||
descriptor.DstSelXYZW() != DstSel(4, 0, 0, 0) ||
descriptor.Stride() != format.byte_size))
return SpecializationFail("indirect formatted X stores require tight SInt8 or UInt32 identity records");
if (formatted_read &&
(descriptor.Base48() == 0u || descriptor.Stride() == 0u || descriptor.NumRecords() == 0u)) {
// Every index is out of bounds. Preserve the format and swizzle constants.
value.dwords[1] &= ~(0x3fffu << 16u);
value.dwords[2] = 0u;
@@ -263,14 +270,10 @@ bool NormalizeIndirectBuffer(DescriptorValue& value, bool indexed, const BufferR
if (descriptor.Stride() < format.byte_size ||
(descriptor.Stride() & (alignment - 1u)) != 0u ||
(descriptor.Base48() & (alignment - 1u)) != 0u)
return SpecializationFail("indirect formatted loads require aligned full-format records");
return SpecializationFail("indirect formatted buffers require aligned full-format records");
} else if ((!indexed && descriptor.Stride() != 1u) ||
(indexed && resource.max_byte_extent != 1u && descriptor.Stride() != 4u)) {
return SpecializationFail("indirect buffer stores require exact linear DWORD or byte bounds");
} else if (resource.formatted && (descriptor.Format() != Prospero::BufferFormat::k32UInt ||
descriptor.DstSelXYZW() != DstSel(4, 0, 0, 0) ||
(descriptor.Base48() & 3u) != 0u)) {
return SpecializationFail("indirect formatted X stores require an aligned 32-bit UInt identity mapping");
}
return true;
}
@@ -1614,16 +1617,6 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
inst.SetFlags(image_remap[inst.Flags<uint32_t>()]);
continue;
}
if (inst.GetOpcode() == ValueOpcode::StoreBufferU32) {
auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
if (memory.kind == ResourceKind::Buffer && memory.formatted &&
program.descriptor_sources[buffers[memory.resource].source].indirect_descriptor) {
// Candidate validation proved an index-only k32UInt identity store.
// Reuse finite raw-store dispatch and its original-index bounds check.
memory.formatted = false;
}
continue;
}
if (BufferAccessOf(inst.GetOpcode()) == BufferAccess::Read) {
const auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
const bool native_oob = memory.kind == ResourceKind::Buffer &&