mirror of
https://git.h3cjp.net/H3cJP/yuzu.git
synced 2024-11-15 18:33:00 +00:00
glsl: WIP var forward declaration
to fix Loop control flow.
This commit is contained in:
parent
2a71333716
commit
f6bbc76336
|
@ -29,7 +29,10 @@ std::string_view SamplerType(TextureType type) {
|
||||||
return "sampler2DArray";
|
return "sampler2DArray";
|
||||||
case TextureType::Color3D:
|
case TextureType::Color3D:
|
||||||
return "sampler3D";
|
return "sampler3D";
|
||||||
|
case TextureType::ColorCube:
|
||||||
|
return "samplerCube";
|
||||||
default:
|
default:
|
||||||
|
fmt::print("Texture type: {}", type);
|
||||||
throw NotImplementedException("Texture type: {}", type);
|
throw NotImplementedException("Texture type: {}", type);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
@ -39,7 +42,6 @@ std::string_view SamplerType(TextureType type) {
|
||||||
EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile& profile_,
|
EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile& profile_,
|
||||||
const RuntimeInfo& runtime_info_)
|
const RuntimeInfo& runtime_info_)
|
||||||
: info{program.info}, profile{profile_}, runtime_info{runtime_info_} {
|
: info{program.info}, profile{profile_}, runtime_info{runtime_info_} {
|
||||||
std::string header = "";
|
|
||||||
SetupExtensions(header);
|
SetupExtensions(header);
|
||||||
stage = program.stage;
|
stage = program.stage;
|
||||||
switch (program.stage) {
|
switch (program.stage) {
|
||||||
|
@ -67,12 +69,11 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile
|
||||||
program.workgroup_size[2]);
|
program.workgroup_size[2]);
|
||||||
break;
|
break;
|
||||||
}
|
}
|
||||||
code += header;
|
|
||||||
const std::string_view attr_stage{stage == Stage::Fragment ? "fragment" : "vertex"};
|
const std::string_view attr_stage{stage == Stage::Fragment ? "fragment" : "vertex"};
|
||||||
for (size_t index = 0; index < info.input_generics.size(); ++index) {
|
for (size_t index = 0; index < info.input_generics.size(); ++index) {
|
||||||
const auto& generic{info.input_generics[index]};
|
const auto& generic{info.input_generics[index]};
|
||||||
if (generic.used) {
|
if (generic.used) {
|
||||||
Add("layout(location={}) {} in vec4 in_attr{};", index,
|
header += fmt::format("layout(location={}) {} in vec4 in_attr{};", index,
|
||||||
InterpDecorator(generic.interpolation), index);
|
InterpDecorator(generic.interpolation), index);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
@ -80,11 +81,11 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile
|
||||||
if (!info.stores_frag_color[index]) {
|
if (!info.stores_frag_color[index]) {
|
||||||
continue;
|
continue;
|
||||||
}
|
}
|
||||||
Add("layout(location={})out vec4 frag_color{};", index, index);
|
header += fmt::format("layout(location={})out vec4 frag_color{};", index, index);
|
||||||
}
|
}
|
||||||
for (size_t index = 0; index < info.stores_generics.size(); ++index) {
|
for (size_t index = 0; index < info.stores_generics.size(); ++index) {
|
||||||
if (info.stores_generics[index]) {
|
if (info.stores_generics[index]) {
|
||||||
Add("layout(location={}) out vec4 out_attr{};", index, index);
|
header += fmt::format("layout(location={}) out vec4 out_attr{};", index, index);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
DefineConstantBuffers(bindings);
|
DefineConstantBuffers(bindings);
|
||||||
|
@ -92,14 +93,15 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile
|
||||||
SetupImages(bindings);
|
SetupImages(bindings);
|
||||||
DefineHelperFunctions();
|
DefineHelperFunctions();
|
||||||
|
|
||||||
Add("void main(){{");
|
header += "void main(){\n";
|
||||||
if (stage == Stage::VertexA || stage == Stage::VertexB) {
|
if (stage == Stage::VertexA || stage == Stage::VertexB) {
|
||||||
Add("gl_Position = vec4(0.0f, 0.0f, 0.0f, 1.0f);");
|
Add("gl_Position = vec4(0.0f, 0.0f, 0.0f, 1.0f);");
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitContext::SetupExtensions(std::string& header) {
|
void EmitContext::SetupExtensions(std::string&) {
|
||||||
header += "#extension GL_ARB_separate_shader_objects : enable\n";
|
header += "#extension GL_ARB_separate_shader_objects : enable\n";
|
||||||
|
// header += "#extension GL_ARB_texture_cube_map_array : enable\n";
|
||||||
if (info.uses_int64) {
|
if (info.uses_int64) {
|
||||||
header += "#extension GL_ARB_gpu_shader_int64 : enable\n";
|
header += "#extension GL_ARB_gpu_shader_int64 : enable\n";
|
||||||
}
|
}
|
||||||
|
@ -127,7 +129,8 @@ void EmitContext::DefineConstantBuffers(Bindings& bindings) {
|
||||||
return;
|
return;
|
||||||
}
|
}
|
||||||
for (const auto& desc : info.constant_buffer_descriptors) {
|
for (const auto& desc : info.constant_buffer_descriptors) {
|
||||||
Add("layout(std140,binding={}) uniform {}_cbuf_{}{{vec4 {}_cbuf{}[{}];}};",
|
header += fmt::format(
|
||||||
|
"layout(std140,binding={}) uniform {}_cbuf_{}{{vec4 {}_cbuf{}[{}];}};",
|
||||||
bindings.uniform_buffer, stage_name, desc.index, stage_name, desc.index, 4 * 1024);
|
bindings.uniform_buffer, stage_name, desc.index, stage_name, desc.index, 4 * 1024);
|
||||||
bindings.uniform_buffer += desc.count;
|
bindings.uniform_buffer += desc.count;
|
||||||
}
|
}
|
||||||
|
@ -138,53 +141,53 @@ void EmitContext::DefineStorageBuffers(Bindings& bindings) {
|
||||||
return;
|
return;
|
||||||
}
|
}
|
||||||
for (const auto& desc : info.storage_buffers_descriptors) {
|
for (const auto& desc : info.storage_buffers_descriptors) {
|
||||||
Add("layout(std430,binding={}) buffer ssbo_{}{{uint ssbo{}[];}};", bindings.storage_buffer,
|
header += fmt::format("layout(std430,binding={}) buffer ssbo_{}{{uint ssbo{}[];}};",
|
||||||
bindings.storage_buffer, desc.cbuf_index);
|
bindings.storage_buffer, bindings.storage_buffer, desc.cbuf_index);
|
||||||
bindings.storage_buffer += desc.count;
|
bindings.storage_buffer += desc.count;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitContext::DefineHelperFunctions() {
|
void EmitContext::DefineHelperFunctions() {
|
||||||
if (info.uses_global_increment) {
|
if (info.uses_global_increment) {
|
||||||
code += "uint CasIncrement(uint op_a,uint op_b){return(op_a>=op_b)?0u:(op_a+1u);}\n";
|
header += "uint CasIncrement(uint op_a,uint op_b){return(op_a>=op_b)?0u:(op_a+1u);}\n";
|
||||||
}
|
}
|
||||||
if (info.uses_global_decrement) {
|
if (info.uses_global_decrement) {
|
||||||
code +=
|
header +=
|
||||||
"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,float op_b){return "
|
header += "uint CasFloatAdd(uint op_a,float op_b){return "
|
||||||
"floatBitsToUint(uintBitsToFloat(op_a)+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,vec2 op_b){return "
|
header += "uint CasFloatAdd32x2(uint op_a,vec2 op_b){return "
|
||||||
"packHalf2x16(unpackHalf2x16(op_a)+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,vec2 op_b){return "
|
header += "uint CasFloatMin32x2(uint op_a,vec2 op_b){return "
|
||||||
"packHalf2x16(min(unpackHalf2x16(op_a),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,vec2 op_b){return "
|
header += "uint CasFloatMax32x2(uint op_a,vec2 op_b){return "
|
||||||
"packHalf2x16(max(unpackHalf2x16(op_a),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,f16vec2 op_b){return "
|
header += "uint CasFloatAdd16x2(uint op_a,f16vec2 op_b){return "
|
||||||
"packFloat2x16(unpackFloat2x16(op_a)+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,f16vec2 op_b){return "
|
header += "uint CasFloatMin16x2(uint op_a,f16vec2 op_b){return "
|
||||||
"packFloat2x16(min(unpackFloat2x16(op_a),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,f16vec2 op_b){return "
|
header += "uint CasFloatMax16x2(uint op_a,f16vec2 op_b){return "
|
||||||
"packFloat2x16(max(unpackFloat2x16(op_a),op_b));}\n";
|
"packFloat2x16(max(unpackFloat2x16(op_a),op_b));}\n";
|
||||||
}
|
}
|
||||||
if (info.uses_atomic_s32_min) {
|
if (info.uses_atomic_s32_min) {
|
||||||
code += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}";
|
header += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}";
|
||||||
}
|
}
|
||||||
if (info.uses_atomic_s32_max) {
|
if (info.uses_atomic_s32_max) {
|
||||||
code += "uint CasMaxS32(uint op_a,uint op_b){return uint(max(int(op_a),int(op_b)));}";
|
header += "uint CasMaxS32(uint op_a,uint op_b){return uint(max(int(op_a),int(op_b)));}";
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
@ -215,7 +218,8 @@ void EmitContext::SetupImages(Bindings& bindings) {
|
||||||
texture_bindings.push_back(bindings.texture);
|
texture_bindings.push_back(bindings.texture);
|
||||||
const auto indices{bindings.texture + desc.count};
|
const auto indices{bindings.texture + desc.count};
|
||||||
for (u32 index = bindings.texture; index < indices; ++index) {
|
for (u32 index = bindings.texture; index < indices; ++index) {
|
||||||
Add("layout(binding={}) uniform {} tex{};", bindings.texture, sampler_type, index);
|
header += fmt::format("layout(binding={}) uniform {} tex{};", bindings.texture,
|
||||||
|
sampler_type, index);
|
||||||
}
|
}
|
||||||
bindings.texture += desc.count;
|
bindings.texture += desc.count;
|
||||||
}
|
}
|
||||||
|
|
|
@ -119,6 +119,7 @@ public:
|
||||||
code += '\n';
|
code += '\n';
|
||||||
}
|
}
|
||||||
|
|
||||||
|
std::string header;
|
||||||
std::string code;
|
std::string code;
|
||||||
RegAlloc reg_alloc;
|
RegAlloc reg_alloc;
|
||||||
const Info& info;
|
const Info& info;
|
||||||
|
|
|
@ -83,6 +83,7 @@ void Invoke(EmitContext& ctx, IR::Inst* inst) {
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitInst(EmitContext& ctx, IR::Inst* inst) {
|
void EmitInst(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
// ctx.Add("/* {} */", inst->GetOpcode());
|
||||||
switch (inst->GetOpcode()) {
|
switch (inst->GetOpcode()) {
|
||||||
#define OPCODE(name, result_type, ...) \
|
#define OPCODE(name, result_type, ...) \
|
||||||
case IR::Opcode::name: \
|
case IR::Opcode::name: \
|
||||||
|
@ -108,12 +109,9 @@ void PrecolorInst(IR::Inst& phi) {
|
||||||
if (arg.IsImmediate()) {
|
if (arg.IsImmediate()) {
|
||||||
ir.PhiMove(phi, arg);
|
ir.PhiMove(phi, arg);
|
||||||
} else {
|
} else {
|
||||||
ir.PhiMove(phi, IR::Value{&*arg.InstRecursive()});
|
ir.PhiMove(phi, IR::Value{arg.InstRecursive()});
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
for (size_t i = 0; i < num_args; ++i) {
|
|
||||||
IR::IREmitter{*phi.PhiBlock(i)}.Reference(IR::Value{&phi});
|
|
||||||
}
|
|
||||||
}
|
}
|
||||||
|
|
||||||
void Precolor(const IR::Program& program) {
|
void Precolor(const IR::Program& program) {
|
||||||
|
@ -144,10 +142,7 @@ void EmitCode(EmitContext& ctx, const IR::Program& program) {
|
||||||
ctx.Add("break;");
|
ctx.Add("break;");
|
||||||
}
|
}
|
||||||
} else {
|
} else {
|
||||||
// TODO: implement this
|
ctx.Add("if({}){{break;}}", ctx.reg_alloc.Consume(node.data.break_node.cond));
|
||||||
ctx.Add("MOV.S.CC RC,{};"
|
|
||||||
"BRK (NE.x);",
|
|
||||||
0);
|
|
||||||
}
|
}
|
||||||
break;
|
break;
|
||||||
case IR::AbstractSyntaxNode::Type::Return:
|
case IR::AbstractSyntaxNode::Type::Return:
|
||||||
|
@ -155,10 +150,12 @@ void EmitCode(EmitContext& ctx, const IR::Program& program) {
|
||||||
ctx.Add("return;");
|
ctx.Add("return;");
|
||||||
break;
|
break;
|
||||||
case IR::AbstractSyntaxNode::Type::Loop:
|
case IR::AbstractSyntaxNode::Type::Loop:
|
||||||
ctx.Add("do{{");
|
ctx.Add("for(;;){{");
|
||||||
break;
|
break;
|
||||||
case IR::AbstractSyntaxNode::Type::Repeat:
|
case IR::AbstractSyntaxNode::Type::Repeat:
|
||||||
ctx.Add("}}while({});", ctx.reg_alloc.Consume(node.data.repeat.cond));
|
ctx.Add("if({}){{", ctx.reg_alloc.Consume(node.data.repeat.cond));
|
||||||
|
ctx.Add("continue;\n}}else{{");
|
||||||
|
ctx.Add("break;\n}}\n}}");
|
||||||
break;
|
break;
|
||||||
default:
|
default:
|
||||||
fmt::print("{}", node.type);
|
fmt::print("{}", node.type);
|
||||||
|
@ -182,7 +179,11 @@ std::string EmitGLSL(const Profile& profile, const RuntimeInfo& runtime_info, IR
|
||||||
Precolor(program);
|
Precolor(program);
|
||||||
EmitCode(ctx, program);
|
EmitCode(ctx, program);
|
||||||
const std::string version{fmt::format("#version 450{}\n", GlslVersionSpecifier(ctx))};
|
const std::string version{fmt::format("#version 450{}\n", GlslVersionSpecifier(ctx))};
|
||||||
ctx.code.insert(0, version);
|
ctx.header.insert(0, version);
|
||||||
|
for (size_t index = 0; index < ctx.reg_alloc.num_used_registers; ++index) {
|
||||||
|
ctx.header += fmt::format("{} R{};", ctx.reg_alloc.reg_types[index], index);
|
||||||
|
}
|
||||||
|
ctx.code.insert(0, ctx.header);
|
||||||
ctx.code += "}";
|
ctx.code += "}";
|
||||||
fmt::print("\n{}\n", ctx.code);
|
fmt::print("\n{}\n", ctx.code);
|
||||||
return ctx.code;
|
return ctx.code;
|
||||||
|
|
|
@ -28,8 +28,8 @@ void SetSignFlag(EmitContext& ctx, IR::Inst& inst, std::string_view result) {
|
||||||
}
|
}
|
||||||
} // Anonymous namespace
|
} // Anonymous namespace
|
||||||
void EmitIAdd32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b) {
|
void EmitIAdd32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b) {
|
||||||
const auto result{ctx.reg_alloc.Define(inst)};
|
const auto result{ctx.reg_alloc.Define(inst, Type::U32)};
|
||||||
ctx.Add("uint {}={}+{};", result, a, b);
|
ctx.Add("{}={}+{};", result, a, b);
|
||||||
SetZeroFlag(ctx, inst, result);
|
SetZeroFlag(ctx, inst, result);
|
||||||
SetSignFlag(ctx, inst, result);
|
SetSignFlag(ctx, inst, result);
|
||||||
}
|
}
|
||||||
|
@ -120,8 +120,8 @@ void EmitBitFieldSExtract(EmitContext& ctx, IR::Inst& inst, std::string_view bas
|
||||||
|
|
||||||
void EmitBitFieldUExtract(EmitContext& ctx, IR::Inst& inst, std::string_view base,
|
void EmitBitFieldUExtract(EmitContext& ctx, IR::Inst& inst, std::string_view base,
|
||||||
std::string_view offset, std::string_view count) {
|
std::string_view offset, std::string_view count) {
|
||||||
const auto result{ctx.reg_alloc.Define(inst)};
|
const auto result{ctx.reg_alloc.Define(inst, Type::U32)};
|
||||||
ctx.Add("uint {}=bitfieldExtract({},int({}),int({}));", result, base, offset, count);
|
ctx.Add("{}=bitfieldExtract({},int({}),int({}));", result, base, offset, count);
|
||||||
SetZeroFlag(ctx, inst, result);
|
SetZeroFlag(ctx, inst, result);
|
||||||
SetSignFlag(ctx, inst, result);
|
SetSignFlag(ctx, inst, result);
|
||||||
}
|
}
|
||||||
|
|
|
@ -74,7 +74,9 @@ std::string RegAlloc::Define(IR::Inst& inst, Type type) {
|
||||||
std::string type_str = "";
|
std::string type_str = "";
|
||||||
if (!register_defined[id.index]) {
|
if (!register_defined[id.index]) {
|
||||||
register_defined[id.index] = true;
|
register_defined[id.index] = true;
|
||||||
type_str = GetGlslType(type);
|
// type_str = GetGlslType(type);
|
||||||
|
reg_types.push_back(GetGlslType(type));
|
||||||
|
++num_used_registers;
|
||||||
}
|
}
|
||||||
inst.SetDefinition<Id>(id);
|
inst.SetDefinition<Id>(id);
|
||||||
return type_str + Representation(id);
|
return type_str + Representation(id);
|
||||||
|
|
|
@ -5,6 +5,7 @@
|
||||||
#pragma once
|
#pragma once
|
||||||
|
|
||||||
#include <bitset>
|
#include <bitset>
|
||||||
|
#include <vector>
|
||||||
|
|
||||||
#include "common/bit_field.h"
|
#include "common/bit_field.h"
|
||||||
#include "common/common_types.h"
|
#include "common/common_types.h"
|
||||||
|
@ -61,19 +62,21 @@ public:
|
||||||
std::string Define(IR::Inst& inst, IR::Type type);
|
std::string Define(IR::Inst& inst, IR::Type type);
|
||||||
|
|
||||||
std::string Consume(const IR::Value& value);
|
std::string Consume(const IR::Value& value);
|
||||||
|
std::string Consume(IR::Inst& inst);
|
||||||
|
|
||||||
std::string GetGlslType(Type type);
|
std::string GetGlslType(Type type);
|
||||||
std::string GetGlslType(IR::Type type);
|
std::string GetGlslType(IR::Type type);
|
||||||
|
|
||||||
|
size_t num_used_registers{};
|
||||||
|
std::vector<std::string> reg_types;
|
||||||
|
|
||||||
private:
|
private:
|
||||||
static constexpr size_t NUM_REGS = 4096;
|
static constexpr size_t NUM_REGS = 4096;
|
||||||
static constexpr size_t NUM_ELEMENTS = 4;
|
|
||||||
|
|
||||||
std::string Consume(IR::Inst& inst);
|
|
||||||
Type RegType(IR::Type type);
|
Type RegType(IR::Type type);
|
||||||
Id Alloc();
|
Id Alloc();
|
||||||
void Free(Id id);
|
void Free(Id id);
|
||||||
|
|
||||||
size_t num_used_registers{};
|
|
||||||
std::bitset<NUM_REGS> register_use{};
|
std::bitset<NUM_REGS> register_use{};
|
||||||
std::bitset<NUM_REGS> register_defined{};
|
std::bitset<NUM_REGS> register_defined{};
|
||||||
};
|
};
|
||||||
|
|
Loading…
Reference in a new issue