2
1
Fork 0
mirror of https://github.com/yuzu-emu/yuzu.git synced 2024-07-04 23:31:19 +01:00

shader: Implement MEMBAR

This commit is contained in:
FernandoS27 2021-04-02 19:27:30 +02:00 committed by ameerj
parent ecb30c9072
commit 655f7a570a
9 changed files with 121 additions and 11 deletions

View file

@ -3,6 +3,7 @@ add_library(shader_recompiler STATIC
backend/spirv/emit_context.h
backend/spirv/emit_spirv.cpp
backend/spirv/emit_spirv.h
backend/spirv/emit_spirv_barriers.cpp
backend/spirv/emit_spirv_bitwise_conversion.cpp
backend/spirv/emit_spirv_composite.cpp
backend/spirv/emit_spirv_context_get_set.cpp
@ -63,6 +64,7 @@ add_library(shader_recompiler STATIC
frontend/maxwell/program.h
frontend/maxwell/structured_control_flow.cpp
frontend/maxwell/structured_control_flow.h
frontend/maxwell/translate/impl/barrier_operations.cpp
frontend/maxwell/translate/impl/bitfield_extract.cpp
frontend/maxwell/translate/impl/bitfield_insert.cpp
frontend/maxwell/translate/impl/branch_indirect.cpp

View file

@ -28,6 +28,7 @@ void EmitSelectionMerge(EmitContext& ctx, Id merge_label);
void EmitReturn(EmitContext& ctx);
void EmitUnreachable(EmitContext& ctx);
void EmitDemoteToHelperInvocation(EmitContext& ctx, Id continue_label);
void EmitMemoryBarrier(EmitContext& ctx, IR::Inst* inst);
void EmitPrologue(EmitContext& ctx);
void EmitEpilogue(EmitContext& ctx);
void EmitGetRegister(EmitContext& ctx);

View file

@ -0,0 +1,40 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "shader_recompiler/backend/spirv/emit_spirv.h"
#include "shader_recompiler/frontend/ir/modifiers.h"
namespace Shader::Backend::SPIRV {
namespace {
spv::Scope MemoryScopeToSpirVScope(IR::MemoryScope scope) {
switch (scope) {
case IR::MemoryScope::Warp:
return spv::Scope::Subgroup;
case IR::MemoryScope::Workgroup:
return spv::Scope::Workgroup;
case IR::MemoryScope::Device:
return spv::Scope::Device;
case IR::MemoryScope::System:
return spv::Scope::CrossDevice;
case IR::MemoryScope::DontCare:
return spv::Scope::Invocation;
default:
throw NotImplementedException("Unknown memory scope!");
}
}
} // namespace
void EmitMemoryBarrier(EmitContext& ctx, IR::Inst* inst) {
const auto info{inst->Flags<IR::BarrierInstInfo>()};
const auto semantics =
spv::MemorySemanticsMask::AcquireRelease | spv::MemorySemanticsMask::UniformMemory |
spv::MemorySemanticsMask::WorkgroupMemory | spv::MemorySemanticsMask::AtomicCounterMemory |
spv::MemorySemanticsMask::ImageMemory;
const auto scope = MemoryScopeToSpirVScope(info.scope);
ctx.OpMemoryBarrier(ctx.Constant(ctx.U32[1], static_cast<u32>(scope)),
ctx.Constant(ctx.U32[1], static_cast<u32>(semantics)));
}
} // namespace Shader::Backend::SPIRV

View file

@ -82,6 +82,10 @@ void IREmitter::SelectionMerge(Block* merge_block) {
Inst(Opcode::SelectionMerge, merge_block);
}
void IREmitter::MemoryBarrier(BarrierInstInfo info) {
Inst(Opcode::MemoryBarrier, Flags{info});
}
void IREmitter::Return() {
block->SetReturn();
Inst(Opcode::Return);

View file

@ -136,6 +136,8 @@ public:
[[nodiscard]] Value Select(const U1& condition, const Value& true_value,
const Value& false_value);
[[nodiscard]] void MemoryBarrier(BarrierInstInfo info);
template <typename Dest, typename Source>
[[nodiscard]] Dest BitCast(const Source& value);

View file

@ -25,6 +25,14 @@ enum class FpRounding : u8 {
RZ, // Round towards zero
};
enum class MemoryScope : u32 {
DontCare,
Warp,
Workgroup,
Device,
System
};
struct FpControl {
bool no_contraction{false};
FpRounding rounding{FpRounding::DontCare};
@ -32,6 +40,11 @@ struct FpControl {
};
static_assert(sizeof(FpControl) <= sizeof(u32));
union BarrierInstInfo {
u32 raw;
BitField<0, 3, MemoryScope> scope;
};
union TextureInstInfo {
u32 raw;
BitField<0, 8, TextureType> type;

View file

@ -16,6 +16,9 @@ OPCODE(Return, Void,
OPCODE(Unreachable, Void, )
OPCODE(DemoteToHelperInvocation, Void, Label, )
// Barriers
OPCODE(MemoryBarrier, Void, )
// Special operations
OPCODE(Prologue, Void, )
OPCODE(Epilogue, Void, )

View file

@ -0,0 +1,56 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "common/bit_field.h"
#include "common/common_types.h"
#include "shader_recompiler/frontend/ir/modifiers.h"
#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
#include "shader_recompiler/frontend/maxwell/opcodes.h"
namespace Shader::Maxwell {
namespace {
// Seems to be in CUDA terminology.
enum class LocalScope : u64 {
CTG = 0,
GL = 1,
SYS = 2,
VC = 3,
};
IR::MemoryScope LocalScopeToMemoryScope(LocalScope scope) {
switch (scope) {
case LocalScope::CTG:
return IR::MemoryScope::Warp;
case LocalScope::GL:
return IR::MemoryScope::Device;
case LocalScope::SYS:
return IR::MemoryScope::System;
case LocalScope::VC:
return IR::MemoryScope::Workgroup; // or should be device?
default:
throw NotImplementedException("Unimplemented Local Scope {}", scope);
}
}
} // namespace
void TranslatorVisitor::MEMBAR(u64 inst) {
union {
u64 raw;
BitField<8, 2, LocalScope> scope;
} membar{inst};
IR::BarrierInstInfo info{};
info.scope.Assign(LocalScopeToMemoryScope(membar.scope));
ir.MemoryBarrier(info);
}
void TranslatorVisitor::DEPBAR() {
// DEPBAR is a no-op
}
void TranslatorVisitor::BAR(u64) {
throw NotImplementedException("Instruction {} is not implemented", Opcode::BAR);
}
} // namespace Shader::Maxwell

View file

@ -37,10 +37,6 @@ void TranslatorVisitor::B2R(u64) {
ThrowNotImplemented(Opcode::B2R);
}
void TranslatorVisitor::BAR(u64) {
ThrowNotImplemented(Opcode::BAR);
}
void TranslatorVisitor::BPT(u64) {
ThrowNotImplemented(Opcode::BPT);
}
@ -73,9 +69,6 @@ void TranslatorVisitor::CS2R(u64) {
ThrowNotImplemented(Opcode::CS2R);
}
void TranslatorVisitor::DEPBAR() {
// DEPBAR is a no-op
}
void TranslatorVisitor::FCHK_reg(u64) {
ThrowNotImplemented(Opcode::FCHK_reg);
@ -189,10 +182,6 @@ void TranslatorVisitor::LONGJMP(u64) {
ThrowNotImplemented(Opcode::LONGJMP);
}
void TranslatorVisitor::MEMBAR(u64) {
ThrowNotImplemented(Opcode::MEMBAR);
}
void TranslatorVisitor::NOP(u64) {
ThrowNotImplemented(Opcode::NOP);
}