yuzu-emu
/
yuzu-mainline
Archived
1
0
Fork 0
Code

glsl: Simply FP storage atomics

This commit is contained in:
ameerj 2021-05-25 02:01:32 -04:00
parent 9cc1b8a873
commit 3482df1176
2 changed files with 28 additions and 48 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
src/shader_recompiler/backend/glsl/emit_context.cpp
View File

@ -78,32 +78,32 @@ void EmitContext::DefineHelperFunctions() {
"uint CasDecrement(uint op_a,uint op_b){return(op_a==0||op_a>op_b)?op_b:(op_a-1u);}\n"; "uint CasDecrement(uint op_a,uint op_b){return(op_a==0||op_a>op_b)?op_b:(op_a-1u);}\n";
} }
if (info.uses_atomic_f32_add) { if (info.uses_atomic_f32_add) {
code += "uint CasFloatAdd(uint op_a,uint op_b){return " code += "uint CasFloatAdd(uint op_a,float op_b){return "
"floatBitsToUint(uintBitsToFloat(op_a)+uintBitsToFloat(op_b));}\n"; "floatBitsToUint(uintBitsToFloat(op_a)+op_b);}\n";
} }
if (info.uses_atomic_f32x2_add) { if (info.uses_atomic_f32x2_add) {
code += "uint CasFloatAdd32x2(uint op_a,uint op_b){return " code += "uint CasFloatAdd32x2(uint op_a,vec2 op_b){return "
"packHalf2x16(unpackHalf2x16(op_a)+unpackHalf2x16(op_b));}\n"; "packHalf2x16(unpackHalf2x16(op_a)+op_b);}\n";
} }
if (info.uses_atomic_f32x2_min) { if (info.uses_atomic_f32x2_min) {
code += "uint CasFloatMin32x2(uint op_a,uint op_b){return " code += "uint CasFloatMin32x2(uint op_a,vec2 op_b){return "
"packHalf2x16(min(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n"; "packHalf2x16(min(unpackHalf2x16(op_a),op_b));}\n";
} }
if (info.uses_atomic_f32x2_max) { if (info.uses_atomic_f32x2_max) {
code += "uint CasFloatMax32x2(uint op_a,uint op_b){return " code += "uint CasFloatMax32x2(uint op_a,vec2 op_b){return "
"packHalf2x16(max(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n"; "packHalf2x16(max(unpackHalf2x16(op_a),op_b));}\n";
} }
if (info.uses_atomic_f16x2_add) { if (info.uses_atomic_f16x2_add) {
code += "uint CasFloatAdd16x2(uint op_a,uint op_b){return " code += "uint CasFloatAdd16x2(uint op_a,f16vec2 op_b){return "
"packFloat2x16(unpackFloat2x16(op_a)+unpackFloat2x16(op_b));}\n"; "packFloat2x16(unpackFloat2x16(op_a)+op_b);}\n";
} }
if (info.uses_atomic_f16x2_min) { if (info.uses_atomic_f16x2_min) {
code += "uint CasFloatMin16x2(uint op_a,uint op_b){return " code += "uint CasFloatMin16x2(uint op_a,f16vec2 op_b){return "
"packFloat2x16(min(unpackFloat2x16(op_a),unpackFloat2x16(op_b)));}\n"; "packFloat2x16(min(unpackFloat2x16(op_a),op_b));}\n";
} }
if (info.uses_atomic_f16x2_max) { if (info.uses_atomic_f16x2_max) {
code += "uint CasFloatMax16x2(uint op_a,uint op_b){return " code += "uint CasFloatMax16x2(uint op_a,f16vec2 op_b){return "
"packFloat2x16(max(unpackFloat2x16(op_a),unpackFloat2x16(op_b)));}\n"; "packFloat2x16(max(unpackFloat2x16(op_a),op_b));}\n";
} }
// TODO: Track this usage // TODO: Track this usage
code += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}"; code += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}";

48
src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
View File

@ -19,16 +19,11 @@ for (;;){{
if ({}==old_value){{break;}} if ({}==old_value){{break;}}
}})"}; }})"};
void CasFunction(EmitContext& ctx, std::string_view ret, std::string_view ssbo, void CasFunction(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
std::string_view value, std::string_view function) { const IR::Value& offset, std::string_view value, std::string_view function) {
ctx.Add(cas_loop.data(), ret, ssbo, ret, ssbo, function, ssbo, value, ret);
}
void CasFunctionInt32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) {
const auto ret{ctx.reg_alloc.Define(inst)}; const auto ret{ctx.reg_alloc.Define(inst)};
const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())}; const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
CasFunction(ctx, ret, ssbo, value, function); ctx.Add(cas_loop.data(), ret, ssbo, ret, ssbo, function, ssbo, value, ret);
} }
void CasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void CasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -37,25 +32,10 @@ void CasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const std::string u32_value{fmt::format("floatBitsToUint({})", value)}; const std::string u32_value{fmt::format("floatBitsToUint({})", value)};
const auto ret{ctx.reg_alloc.Define(inst)}; const auto ret{ctx.reg_alloc.Define(inst)};
const auto ret_32{ret + "_u32"}; const auto ret_32{ret + "_u32"};
CasFunction(ctx, ret_32, ssbo, u32_value, function); ctx.Add(cas_loop.data(), ret_32, ssbo, ret_32, ssbo, function, ssbo, value, ret_32);
ctx.Add("float {}=uintBitsToFloat({});", ret, ret_32); ctx.Add("float {}=uintBitsToFloat({});", ret, ret_32);
} }
void CasFunctionF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) {
const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
const std::string u32_value{fmt::format("packHalf2x16({})", value)};
const auto ret{ctx.reg_alloc.Define(inst)};
CasFunction(ctx, ret, ssbo, u32_value, function);
}
void CasFunctionF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) {
const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
const std::string u32_value{fmt::format("packFloat2x16({})", value)};
const auto ret{ctx.reg_alloc.Define(inst)};
CasFunction(ctx, ret, ssbo, u32_value, function);
}
} // namespace } // namespace
void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -66,7 +46,7 @@ void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
const std::string u32_value{fmt::format("uint({})", value)}; const std::string u32_value{fmt::format("uint({})", value)};
CasFunctionInt32(ctx, inst, binding, offset, u32_value, "CasMinS32"); CasFunction(ctx, inst, binding, offset, u32_value, "CasMinS32");
} }
void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -77,7 +57,7 @@ void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
const std::string u32_value{fmt::format("uint({})", value)}; const std::string u32_value{fmt::format("uint({})", value)};
CasFunctionInt32(ctx, inst, binding, offset, u32_value, "CasMaxS32"); CasFunction(ctx, inst, binding, offset, u32_value, "CasMaxS32");
} }
void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -87,12 +67,12 @@ void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
void EmitStorageAtomicInc32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicInc32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionInt32(ctx, inst, binding, offset, value, "CasIncrement"); CasFunction(ctx, inst, binding, offset, value, "CasIncrement");
} }
void EmitStorageAtomicDec32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicDec32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionInt32(ctx, inst, binding, offset, value, "CasDecrement"); CasFunction(ctx, inst, binding, offset, value, "CasDecrement");
} }
void EmitStorageAtomicAnd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAnd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -199,32 +179,32 @@ void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
void EmitStorageAtomicAddF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAddF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatAdd16x2"); CasFunction(ctx, inst, binding, offset, value, "CasFloatAdd16x2");
} }
void EmitStorageAtomicAddF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAddF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatAdd32x2"); CasFunction(ctx, inst, binding, offset, value, "CasFloatAdd32x2");
} }
void EmitStorageAtomicMinF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicMinF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatMin16x2"); CasFunction(ctx, inst, binding, offset, value, "CasFloatMin16x2");
} }
void EmitStorageAtomicMinF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicMinF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMin32x2"); CasFunction(ctx, inst, binding, offset, value, "CasFloatMin32x2");
} }
void EmitStorageAtomicMaxF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicMaxF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatMax16x2"); CasFunction(ctx, inst, binding, offset, value, "CasFloatMax16x2");
} }
void EmitStorageAtomicMaxF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicMaxF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMax32x2"); CasFunction(ctx, inst, binding, offset, value, "CasFloatMax32x2");
} }
void EmitGlobalAtomicIAdd32(EmitContext&) { void EmitGlobalAtomicIAdd32(EmitContext&) {