Support LUTs in single-function CFGs on Private storage class.

Fairly common pattern in unoptimized SPIR-V. Support this case as well.
This commit is contained in:
Hans-Kristian Arntzen 2019-02-06 10:38:18 +01:00
parent 84f56d0b43
commit 3e584f2c3f
18 changed files with 589 additions and 37 deletions

View File

@ -0,0 +1,62 @@
struct myType
{
float data;
};
static const myType _18 = { 0.0f };
static const myType _20 = { 1.0f };
static const myType _21[5] = { { 0.0f }, { 1.0f }, { 0.0f }, { 1.0f }, { 0.0f } };
static float4 gl_FragCoord;
static float4 o_color;
struct SPIRV_Cross_Input
{
float4 gl_FragCoord : SV_Position;
};
struct SPIRV_Cross_Output
{
float4 o_color : SV_Target0;
};
float mod(float x, float y)
{
return x - y * floor(x / y);
}
float2 mod(float2 x, float2 y)
{
return x - y * floor(x / y);
}
float3 mod(float3 x, float3 y)
{
return x - y * floor(x / y);
}
float4 mod(float4 x, float4 y)
{
return x - y * floor(x / y);
}
void frag_main()
{
if (_21[int(mod(gl_FragCoord.x, 4.0f))].data > 0.0f)
{
o_color = float4(0.0f, 1.0f, 0.0f, 1.0f);
}
else
{
o_color = float4(1.0f, 0.0f, 0.0f, 1.0f);
}
}
SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
{
gl_FragCoord = stage_input.gl_FragCoord;
frag_main();
SPIRV_Cross_Output stage_output;
stage_output.o_color = o_color;
return stage_output;
}

View File

@ -0,0 +1,40 @@
#pragma clang diagnostic ignored "-Wmissing-prototypes"
#include <metal_stdlib>
#include <simd/simd.h>
using namespace metal;
struct myType
{
float data;
};
constant myType _21[5] = { myType{ 0.0 }, myType{ 1.0 }, myType{ 0.0 }, myType{ 1.0 }, myType{ 0.0 } };
struct main0_out
{
float4 o_color [[color(0)]];
};
// Implementation of the GLSL mod() function, which is slightly different than Metal fmod()
template<typename Tx, typename Ty>
Tx mod(Tx x, Ty y)
{
return x - y * floor(x / y);
}
fragment main0_out main0(float4 gl_FragCoord [[position]])
{
main0_out out = {};
if (_21[int(mod(gl_FragCoord.x, 4.0))].data > 0.0)
{
out.o_color = float4(0.0, 1.0, 0.0, 1.0);
}
else
{
out.o_color = float4(1.0, 0.0, 0.0, 1.0);
}
return out;
}

View File

@ -0,0 +1,23 @@
#version 460
struct myType
{
float data;
};
const myType _21[5] = myType[](myType(0.0), myType(1.0), myType(0.0), myType(1.0), myType(0.0));
layout(location = 0) out vec4 o_color;
void main()
{
if (_21[int(mod(gl_FragCoord.x, 4.0))].data > 0.0)
{
o_color = vec4(0.0, 1.0, 0.0, 1.0);
}
else
{
o_color = vec4(1.0, 0.0, 0.0, 1.0);
}
}

View File

@ -0,0 +1,65 @@
struct myType
{
float data;
};
static const myType _18 = { 0.0f };
static const myType _20 = { 1.0f };
static const myType _21[5] = { { 0.0f }, { 1.0f }, { 0.0f }, { 1.0f }, { 0.0f } };
static float4 gl_FragCoord;
static float4 o_color;
struct SPIRV_Cross_Input
{
float4 gl_FragCoord : SV_Position;
};
struct SPIRV_Cross_Output
{
float4 o_color : SV_Target0;
};
float mod(float x, float y)
{
return x - y * floor(x / y);
}
float2 mod(float2 x, float2 y)
{
return x - y * floor(x / y);
}
float3 mod(float3 x, float3 y)
{
return x - y * floor(x / y);
}
float4 mod(float4 x, float4 y)
{
return x - y * floor(x / y);
}
void frag_main()
{
float2 uv = gl_FragCoord.xy;
int index = int(mod(uv.x, 4.0f));
myType elt = _21[index];
if (elt.data > 0.0f)
{
o_color = float4(0.0f, 1.0f, 0.0f, 1.0f);
}
else
{
o_color = float4(1.0f, 0.0f, 0.0f, 1.0f);
}
}
SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
{
gl_FragCoord = stage_input.gl_FragCoord;
frag_main();
SPIRV_Cross_Output stage_output;
stage_output.o_color = o_color;
return stage_output;
}

View File

@ -22,15 +22,10 @@ struct SPIRV_Cross_Output
float4 FragColor : SV_Target0;
};
static float lut[4];
static Foo foos[2];
void frag_main()
{
lut = _16;
foos = _28;
FragColor = lut[_line].xxxx;
FragColor += (foos[_line].a * foos[1 - _line].a).xxxx;
FragColor = _16[_line].xxxx;
FragColor += (_28[_line].a * _28[1 - _line].a).xxxx;
}
SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)

View File

@ -0,0 +1,56 @@
#pragma clang diagnostic ignored "-Wmissing-prototypes"
#include <metal_stdlib>
#include <simd/simd.h>
using namespace metal;
struct myType
{
float data;
};
constant myType _21[5] = { myType{ 0.0 }, myType{ 1.0 }, myType{ 0.0 }, myType{ 1.0 }, myType{ 0.0 } };
struct main0_out
{
float4 o_color [[color(0)]];
};
// Implementation of the GLSL mod() function, which is slightly different than Metal fmod()
template<typename Tx, typename Ty>
Tx mod(Tx x, Ty y)
{
return x - y * floor(x / y);
}
// Implementation of an array copy function to cover GLSL's ability to copy an array via assignment.
template<typename T, uint N>
void spvArrayCopyFromStack1(thread T (&dst)[N], thread const T (&src)[N])
{
for (uint i = 0; i < N; dst[i] = src[i], i++);
}
template<typename T, uint N>
void spvArrayCopyFromConstant1(thread T (&dst)[N], constant T (&src)[N])
{
for (uint i = 0; i < N; dst[i] = src[i], i++);
}
fragment main0_out main0(float4 gl_FragCoord [[position]])
{
main0_out out = {};
float2 uv = gl_FragCoord.xy;
int index = int(mod(uv.x, 4.0));
myType elt = _21[index];
if (elt.data > 0.0)
{
out.o_color = float4(0.0, 1.0, 0.0, 1.0);
}
else
{
out.o_color = float4(1.0, 0.0, 0.0, 1.0);
}
return out;
}

View File

@ -40,10 +40,8 @@ void spvArrayCopyFromConstant1(thread T (&dst)[N], constant T (&src)[N])
fragment main0_out main0(main0_in in [[stage_in]])
{
main0_out out = {};
float lut[4] = { 1.0, 4.0, 3.0, 2.0 };
Foo foos[2] = { Foo{ 10.0, 20.0 }, Foo{ 30.0, 40.0 } };
out.FragColor = float4(lut[in.line]);
out.FragColor += float4(foos[in.line].a * foos[1 - in.line].a);
out.FragColor = float4(_16[in.line]);
out.FragColor += float4(_28[in.line].a * _28[1 - in.line].a);
return out;
}

View File

@ -0,0 +1,26 @@
#version 460
struct myType
{
float data;
};
const myType _21[5] = myType[](myType(0.0), myType(1.0), myType(0.0), myType(1.0), myType(0.0));
layout(location = 0) out vec4 o_color;
void main()
{
vec2 uv = gl_FragCoord.xy;
int index = int(mod(uv.x, 4.0));
myType elt = _21[index];
if (elt.data > 0.0)
{
o_color = vec4(0.0, 1.0, 0.0, 1.0);
}
else
{
o_color = vec4(1.0, 0.0, 0.0, 1.0);
}
}

View File

@ -2,22 +2,22 @@
precision mediump float;
precision highp int;
const float _16[4] = float[](1.0, 4.0, 3.0, 2.0);
struct Foo
{
float a;
float b;
};
const Foo _28[2] = Foo[](Foo(10.0, 20.0), Foo(30.0, 40.0));
layout(location = 0) out vec4 FragColor;
layout(location = 0) flat in mediump int line;
float lut[4];
Foo foos[2];
void main()
{
lut = float[](1.0, 4.0, 3.0, 2.0);
foos = Foo[](Foo(10.0, 20.0), Foo(30.0, 40.0));
FragColor = vec4(lut[line]);
FragColor += vec4(foos[line].a * foos[1 - line].a);
FragColor = vec4(_16[line]);
FragColor += vec4(_28[line].a * _28[1 - line].a);
}

View File

@ -0,0 +1,86 @@
; SPIR-V
; Version: 1.3
; Generator: Khronos SPIR-V Tools Assembler; 0
; Bound: 54
; Schema: 0
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %gl_FragCoord %o_color
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 460
OpName %main "main"
OpName %myType "myType"
OpMemberName %myType 0 "data"
OpName %myData "myData"
OpName %uv "uv"
OpName %gl_FragCoord "gl_FragCoord"
OpName %index "index"
OpName %elt "elt"
OpName %o_color "o_color"
OpDecorate %gl_FragCoord BuiltIn FragCoord
OpDecorate %o_color Location 0
%void = OpTypeVoid
%11 = OpTypeFunction %void
%float = OpTypeFloat 32
%myType = OpTypeStruct %float
%uint = OpTypeInt 32 0
%uint_5 = OpConstant %uint 5
%_arr_myType_uint_5 = OpTypeArray %myType %uint_5
%_ptr_Private__arr_myType_uint_5 = OpTypePointer Private %_arr_myType_uint_5
%myData = OpVariable %_ptr_Private__arr_myType_uint_5 Private
%float_0 = OpConstant %float 0
%18 = OpConstantComposite %myType %float_0
%float_1 = OpConstant %float 1
%20 = OpConstantComposite %myType %float_1
%21 = OpConstantComposite %_arr_myType_uint_5 %18 %20 %18 %20 %18
%v2float = OpTypeVector %float 2
%_ptr_Function_v2float = OpTypePointer Function %v2float
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%gl_FragCoord = OpVariable %_ptr_Input_v4float Input
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%uint_0 = OpConstant %uint 0
%_ptr_Function_float = OpTypePointer Function %float
%float_4 = OpConstant %float 4
%_ptr_Function_myType = OpTypePointer Function %myType
%_ptr_Private_myType = OpTypePointer Private %myType
%int_0 = OpConstant %int 0
%bool = OpTypeBool
%_ptr_Output_v4float = OpTypePointer Output %v4float
%o_color = OpVariable %_ptr_Output_v4float Output
%36 = OpConstantComposite %v4float %float_0 %float_1 %float_0 %float_1
%37 = OpConstantComposite %v4float %float_1 %float_0 %float_0 %float_1
%main = OpFunction %void None %11
%38 = OpLabel
%uv = OpVariable %_ptr_Function_v2float Function
%index = OpVariable %_ptr_Function_int Function
%elt = OpVariable %_ptr_Function_myType Function
OpStore %myData %21
%39 = OpLoad %v4float %gl_FragCoord
%40 = OpVectorShuffle %v2float %39 %39 0 1
OpStore %uv %40
%41 = OpAccessChain %_ptr_Function_float %uv %uint_0
%42 = OpLoad %float %41
%43 = OpFMod %float %42 %float_4
%44 = OpConvertFToS %int %43
OpStore %index %44
%45 = OpLoad %int %index
%46 = OpAccessChain %_ptr_Private_myType %myData %45
%47 = OpLoad %myType %46
OpStore %elt %47
%48 = OpAccessChain %_ptr_Function_float %elt %int_0
%49 = OpLoad %float %48
%50 = OpFOrdGreaterThan %bool %49 %float_0
OpSelectionMerge %51 None
OpBranchConditional %50 %52 %53
%52 = OpLabel
OpStore %o_color %36
OpBranch %51
%53 = OpLabel
OpStore %o_color %37
OpBranch %51
%51 = OpLabel
OpReturn
OpFunctionEnd

View File

@ -0,0 +1,86 @@
; SPIR-V
; Version: 1.3
; Generator: Khronos SPIR-V Tools Assembler; 0
; Bound: 54
; Schema: 0
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %gl_FragCoord %o_color
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 460
OpName %main "main"
OpName %myType "myType"
OpMemberName %myType 0 "data"
OpName %myData "myData"
OpName %uv "uv"
OpName %gl_FragCoord "gl_FragCoord"
OpName %index "index"
OpName %elt "elt"
OpName %o_color "o_color"
OpDecorate %gl_FragCoord BuiltIn FragCoord
OpDecorate %o_color Location 0
%void = OpTypeVoid
%11 = OpTypeFunction %void
%float = OpTypeFloat 32
%myType = OpTypeStruct %float
%uint = OpTypeInt 32 0
%uint_5 = OpConstant %uint 5
%_arr_myType_uint_5 = OpTypeArray %myType %uint_5
%_ptr_Private__arr_myType_uint_5 = OpTypePointer Private %_arr_myType_uint_5
%myData = OpVariable %_ptr_Private__arr_myType_uint_5 Private
%float_0 = OpConstant %float 0
%18 = OpConstantComposite %myType %float_0
%float_1 = OpConstant %float 1
%20 = OpConstantComposite %myType %float_1
%21 = OpConstantComposite %_arr_myType_uint_5 %18 %20 %18 %20 %18
%v2float = OpTypeVector %float 2
%_ptr_Function_v2float = OpTypePointer Function %v2float
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%gl_FragCoord = OpVariable %_ptr_Input_v4float Input
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%uint_0 = OpConstant %uint 0
%_ptr_Function_float = OpTypePointer Function %float
%float_4 = OpConstant %float 4
%_ptr_Function_myType = OpTypePointer Function %myType
%_ptr_Private_myType = OpTypePointer Private %myType
%int_0 = OpConstant %int 0
%bool = OpTypeBool
%_ptr_Output_v4float = OpTypePointer Output %v4float
%o_color = OpVariable %_ptr_Output_v4float Output
%36 = OpConstantComposite %v4float %float_0 %float_1 %float_0 %float_1
%37 = OpConstantComposite %v4float %float_1 %float_0 %float_0 %float_1
%main = OpFunction %void None %11
%38 = OpLabel
%uv = OpVariable %_ptr_Function_v2float Function
%index = OpVariable %_ptr_Function_int Function
%elt = OpVariable %_ptr_Function_myType Function
OpStore %myData %21
%39 = OpLoad %v4float %gl_FragCoord
%40 = OpVectorShuffle %v2float %39 %39 0 1
OpStore %uv %40
%41 = OpAccessChain %_ptr_Function_float %uv %uint_0
%42 = OpLoad %float %41
%43 = OpFMod %float %42 %float_4
%44 = OpConvertFToS %int %43
OpStore %index %44
%45 = OpLoad %int %index
%46 = OpAccessChain %_ptr_Private_myType %myData %45
%47 = OpLoad %myType %46
OpStore %elt %47
%48 = OpAccessChain %_ptr_Function_float %elt %int_0
%49 = OpLoad %float %48
%50 = OpFOrdGreaterThan %bool %49 %float_0
OpSelectionMerge %51 None
OpBranchConditional %50 %52 %53
%52 = OpLabel
OpStore %o_color %36
OpBranch %51
%53 = OpLabel
OpStore %o_color %37
OpBranch %51
%51 = OpLabel
OpReturn
OpFunctionEnd

View File

@ -0,0 +1,86 @@
; SPIR-V
; Version: 1.3
; Generator: Khronos SPIR-V Tools Assembler; 0
; Bound: 54
; Schema: 0
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %gl_FragCoord %o_color
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 460
OpName %main "main"
OpName %myType "myType"
OpMemberName %myType 0 "data"
OpName %myData "myData"
OpName %uv "uv"
OpName %gl_FragCoord "gl_FragCoord"
OpName %index "index"
OpName %elt "elt"
OpName %o_color "o_color"
OpDecorate %gl_FragCoord BuiltIn FragCoord
OpDecorate %o_color Location 0
%void = OpTypeVoid
%11 = OpTypeFunction %void
%float = OpTypeFloat 32
%myType = OpTypeStruct %float
%uint = OpTypeInt 32 0
%uint_5 = OpConstant %uint 5
%_arr_myType_uint_5 = OpTypeArray %myType %uint_5
%_ptr_Private__arr_myType_uint_5 = OpTypePointer Private %_arr_myType_uint_5
%myData = OpVariable %_ptr_Private__arr_myType_uint_5 Private
%float_0 = OpConstant %float 0
%18 = OpConstantComposite %myType %float_0
%float_1 = OpConstant %float 1
%20 = OpConstantComposite %myType %float_1
%21 = OpConstantComposite %_arr_myType_uint_5 %18 %20 %18 %20 %18
%v2float = OpTypeVector %float 2
%_ptr_Function_v2float = OpTypePointer Function %v2float
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%gl_FragCoord = OpVariable %_ptr_Input_v4float Input
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%uint_0 = OpConstant %uint 0
%_ptr_Function_float = OpTypePointer Function %float
%float_4 = OpConstant %float 4
%_ptr_Function_myType = OpTypePointer Function %myType
%_ptr_Private_myType = OpTypePointer Private %myType
%int_0 = OpConstant %int 0
%bool = OpTypeBool
%_ptr_Output_v4float = OpTypePointer Output %v4float
%o_color = OpVariable %_ptr_Output_v4float Output
%36 = OpConstantComposite %v4float %float_0 %float_1 %float_0 %float_1
%37 = OpConstantComposite %v4float %float_1 %float_0 %float_0 %float_1
%main = OpFunction %void None %11
%38 = OpLabel
%uv = OpVariable %_ptr_Function_v2float Function
%index = OpVariable %_ptr_Function_int Function
%elt = OpVariable %_ptr_Function_myType Function
OpStore %myData %21
%39 = OpLoad %v4float %gl_FragCoord
%40 = OpVectorShuffle %v2float %39 %39 0 1
OpStore %uv %40
%41 = OpAccessChain %_ptr_Function_float %uv %uint_0
%42 = OpLoad %float %41
%43 = OpFMod %float %42 %float_4
%44 = OpConvertFToS %int %43
OpStore %index %44
%45 = OpLoad %int %index
%46 = OpAccessChain %_ptr_Private_myType %myData %45
%47 = OpLoad %myType %46
OpStore %elt %47
%48 = OpAccessChain %_ptr_Function_float %elt %int_0
%49 = OpLoad %float %48
%50 = OpFOrdGreaterThan %bool %49 %float_0
OpSelectionMerge %51 None
OpBranchConditional %50 %52 %53
%52 = OpLabel
OpStore %o_color %36
OpBranch %51
%53 = OpLabel
OpStore %o_color %37
OpBranch %51
%51 = OpLabel
OpReturn
OpFunctionEnd

View File

@ -3054,7 +3054,8 @@ bool Compiler::StaticExpressionAccessHandler::handle(spv::Op op, const uint32_t
return true;
}
void Compiler::find_function_local_luts(SPIRFunction &entry, const AnalyzeVariableScopeAccessHandler &handler)
void Compiler::find_function_local_luts(SPIRFunction &entry, const AnalyzeVariableScopeAccessHandler &handler,
bool single_function)
{
auto &cfg = *function_cfgs.find(entry.self)->second;
@ -3066,7 +3067,10 @@ void Compiler::find_function_local_luts(SPIRFunction &entry, const AnalyzeVariab
auto &type = expression_type(accessed_var.first);
// Only consider function local variables here.
if (var.storage != StorageClassFunction)
// If we only have a single function in our CFG, private storage is also fine,
// since it behaves like a function local variable.
bool allow_lut = var.storage == StorageClassFunction || (single_function && var.storage == StorageClassPrivate);
if (!allow_lut)
continue;
// We cannot be a phi variable.
@ -3651,13 +3655,14 @@ void Compiler::build_function_control_flow_graphs_and_analyze()
handler.function_cfgs[ir.default_entry_point].reset(new CFG(*this, get<SPIRFunction>(ir.default_entry_point)));
traverse_all_reachable_opcodes(get<SPIRFunction>(ir.default_entry_point), handler);
function_cfgs = move(handler.function_cfgs);
bool single_function = function_cfgs.size() <= 1;
for (auto &f : function_cfgs)
{
auto &func = get<SPIRFunction>(f.first);
AnalyzeVariableScopeAccessHandler scope_handler(*this, func);
analyze_variable_scope(func, scope_handler);
find_function_local_luts(func, scope_handler);
find_function_local_luts(func, scope_handler, single_function);
// Check if we can actually use the loop variables we found in analyze_variable_scope.
// To use multiple initializers, we need the same type and qualifiers.

View File

@ -938,7 +938,8 @@ protected:
};
void analyze_variable_scope(SPIRFunction &function, AnalyzeVariableScopeAccessHandler &handler);
void find_function_local_luts(SPIRFunction &function, const AnalyzeVariableScopeAccessHandler &handler);
void find_function_local_luts(SPIRFunction &function, const AnalyzeVariableScopeAccessHandler &handler,
bool single_function);
void make_constant_null(uint32_t id, uint32_t type);

View File

@ -2137,6 +2137,20 @@ void CompilerGLSL::declare_undefined_values()
statement("");
}
bool CompilerGLSL::variable_is_lut(const SPIRVariable &var) const
{
bool statically_assigned = var.statically_assigned && var.static_expression != 0 && var.remapped_variable;
if (statically_assigned)
{
auto *constant = maybe_get<SPIRConstant>(var.static_expression);
if (constant && constant->is_used_as_lut)
return true;
}
return false;
}
void CompilerGLSL::emit_resources()
{
auto &execution = get_entry_point();
@ -2347,9 +2361,12 @@ void CompilerGLSL::emit_resources()
auto &var = get<SPIRVariable>(global);
if (var.storage != StorageClassOutput)
{
add_resource_name(var.self);
statement(variable_decl(var), ";");
emitted = true;
if (!variable_is_lut(var))
{
add_resource_name(var.self);
statement(variable_decl(var), ";");
emitted = true;
}
}
}

View File

@ -638,6 +638,8 @@ protected:
uint32_t get_integer_width_for_instruction(const Instruction &instr) const;
bool variable_is_lut(const SPIRVariable &var) const;
private:
void init()
{

View File

@ -1327,21 +1327,24 @@ void CompilerHLSL::emit_resources()
auto &var = get<SPIRVariable>(global);
if (var.storage != StorageClassOutput)
{
add_resource_name(var.self);
const char *storage = nullptr;
switch (var.storage)
if (!variable_is_lut(var))
{
case StorageClassWorkgroup:
storage = "groupshared";
break;
add_resource_name(var.self);
default:
storage = "static";
break;
const char *storage = nullptr;
switch (var.storage)
{
case StorageClassWorkgroup:
storage = "groupshared";
break;
default:
storage = "static";
break;
}
statement(storage, " ", variable_decl(var), ";");
emitted = true;
}
statement(storage, " ", variable_decl(var), ";");
emitted = true;
}
}

View File

@ -534,7 +534,8 @@ void CompilerMSL::localize_global_variables()
auto &var = get<SPIRVariable>(v_id);
if (var.storage == StorageClassPrivate || var.storage == StorageClassWorkgroup)
{
entry_func.add_local_variable(v_id);
if (!variable_is_lut(var))
entry_func.add_local_variable(v_id);
iter = global_variables.erase(iter);
}
else