Implement IMAGE_ATOMIC_FMIN and IMAGE_ATOMIC_FMAX for 32bit floats (#2820)

* Implement IMAGE_ATOMIC_FMIN and IMAGE_ATOMIC_FMAX for 32bit floats

* Handle missing VK_EXT_shader_atomic_float2
This commit is contained in:
Marcin Mikołajczyk 2025-04-30 20:42:08 +02:00 committed by GitHub
parent a3bbf2274f
commit c08f92aca1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 106 additions and 2 deletions

View file

@ -270,6 +270,10 @@ void SetupCapabilities(const Info& info, const Profile& profile, EmitContext& ct
if (info.has_image_query) {
ctx.AddCapability(spv::Capability::ImageQuery);
}
if (info.uses_atomic_float_min_max) {
ctx.AddExtension("SPV_EXT_shader_atomic_float_min_max");
ctx.AddCapability(spv::Capability::AtomicFloat32MinMaxEXT);
}
if (info.uses_lane_id) {
ctx.AddCapability(spv::Capability::GroupNonUniform);
}

View file

@ -75,6 +75,14 @@ Id ImageAtomicU32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id va
const auto [scope, semantics]{AtomicArgs(ctx)};
return (ctx.*atomic_func)(ctx.U32[1], pointer, scope, semantics, value);
}
Id ImageAtomicF32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value,
Id (Sirit::Module::*atomic_func)(Id, Id, Id, Id, Id)) {
const auto& texture = ctx.images[handle & 0xFFFF];
const Id pointer{ctx.OpImageTexelPointer(ctx.image_f32, texture.id, coords, ctx.ConstU32(0U))};
const auto [scope, semantics]{AtomicArgs(ctx)};
return (ctx.*atomic_func)(ctx.F32[1], pointer, scope, semantics, value);
}
} // Anonymous namespace
Id EmitSharedAtomicIAdd32(EmitContext& ctx, Id offset, Id value) {
@ -187,6 +195,40 @@ Id EmitImageAtomicUMax32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords
return ImageAtomicU32(ctx, inst, handle, coords, value, &Sirit::Module::OpAtomicUMax);
}
Id EmitImageAtomicFMax32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value) {
if (ctx.profile.supports_image_fp32_atomic_min_max) {
return ImageAtomicF32(ctx, inst, handle, coords, value, &Sirit::Module::OpAtomicFMax);
}
const auto u32_value = ctx.OpBitcast(ctx.U32[1], value);
const auto sign_bit_set =
ctx.OpBitFieldUExtract(ctx.U32[1], u32_value, ctx.ConstU32(31u), ctx.ConstU32(1u));
const auto result = ctx.OpSelect(
ctx.F32[1], sign_bit_set,
EmitBitCastF32U32(ctx, EmitImageAtomicUMin32(ctx, inst, handle, coords, u32_value)),
EmitBitCastF32U32(ctx, EmitImageAtomicSMax32(ctx, inst, handle, coords, u32_value)));
return result;
}
Id EmitImageAtomicFMin32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value) {
if (ctx.profile.supports_image_fp32_atomic_min_max) {
return ImageAtomicF32(ctx, inst, handle, coords, value, &Sirit::Module::OpAtomicFMin);
}
const auto u32_value = ctx.OpBitcast(ctx.U32[1], value);
const auto sign_bit_set =
ctx.OpBitFieldUExtract(ctx.U32[1], u32_value, ctx.ConstU32(31u), ctx.ConstU32(1u));
const auto result = ctx.OpSelect(
ctx.F32[1], sign_bit_set,
EmitBitCastF32U32(ctx, EmitImageAtomicUMax32(ctx, inst, handle, coords, u32_value)),
EmitBitCastF32U32(ctx, EmitImageAtomicSMin32(ctx, inst, handle, coords, u32_value)));
return result;
}
Id EmitImageAtomicInc32(EmitContext&, IR::Inst*, u32, Id, Id) {
// TODO: This is not yet implemented
throw NotImplementedException("SPIR-V Instruction");

View file

@ -482,6 +482,8 @@ Id EmitImageAtomicSMin32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords
Id EmitImageAtomicUMin32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicSMax32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicUMax32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicFMax32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicFMin32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicInc32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicDec32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);
Id EmitImageAtomicAnd32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id value);

View file

@ -869,6 +869,7 @@ void EmitContext::DefineImagesAndSamplers() {
}
if (std::ranges::any_of(info.images, &ImageResource::is_atomic)) {
image_u32 = TypePointer(spv::StorageClass::Image, U32[1]);
image_f32 = TypePointer(spv::StorageClass::Image, F32[1]);
}
if (info.samplers.empty()) {
return;

View file

@ -207,6 +207,7 @@ public:
Id invocation_id{};
Id subgroup_local_invocation_id{};
Id image_u32{};
Id image_f32{};
Id shared_memory_u8{};
Id shared_memory_u16{};