AuroraMiddleware/glslang
1
0
Fork 0
mirror of https://github.com/KhronosGroup/glslang synced 2024-11-08 11:30:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

If a shader's only use of float16 is for FConvert, add the Float16 capability.

Browse Source 
When visiting instructions, check that there are no 16-bit storage capabilities
and add the Float16 capability. Same for int8/int16.
This commit is contained in:
Jeff Bolz 2019-05-01 11:45:36 -05:00
parent 86c72c9486
commit faac86e5d6
6 changed files with 157 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
SPIRV/SpvPostProcess.cpp
View File

@ -118,9 +118,46 @@ void Builder::postProcessType(const Instruction& inst, Id typeId)
case OpAccessChain:
case OpPtrAccessChain:
case OpCopyObject:
break;
case OpFConvert:
case OpSConvert:
case OpUConvert:
// Look for any 8/16-bit storage capabilities. If there are none, assume that
// the convert instruction requires the Float16/Int8/16 capability.
if (containsType(typeId, OpTypeFloat, 16) || containsType(typeId, OpTypeInt, 16)) {
bool foundStorage = false;
for (auto it = capabilities.begin(); it != capabilities.end(); ++it) {
spv::Capability cap = *it;
if (cap == spv::CapabilityStorageInputOutput16 ||
cap == spv::CapabilityStoragePushConstant16 ||
cap == spv::CapabilityStorageUniformBufferBlock16 ||
cap == spv::CapabilityStorageUniform16) {
foundStorage = true;
break;
}
}
if (!foundStorage) {
if (containsType(typeId, OpTypeFloat, 16))
addCapability(CapabilityFloat16);
if (containsType(typeId, OpTypeInt, 16))
addCapability(CapabilityInt16);
}
}
if (containsType(typeId, OpTypeInt, 8)) {
bool foundStorage = false;
for (auto it = capabilities.begin(); it != capabilities.end(); ++it) {
spv::Capability cap = *it;
if (cap == spv::CapabilityStoragePushConstant8 ||
cap == spv::CapabilityUniformAndStorageBuffer8BitAccess ||
cap == spv::CapabilityStorageBuffer8BitAccess) {
foundStorage = true;
break;
}
}
if (!foundStorage) {
addCapability(CapabilityInt8);
}
}
break;
case OpExtInst:
#if AMD_EXTENSIONS
@ -327,6 +364,24 @@ void Builder::postProcess()
// Add per-instruction capabilities, extensions, etc.,
// Look for any 8/16 bit type in physical storage buffer class, and set the
// appropriate capability. This happens in createSpvVariable for other storage
// classes, but there isn't always a variable for physical storage buffer.
for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) {
Instruction* type = groupedTypes[OpTypePointer][t];
if (type->getImmediateOperand(0) == (unsigned)StorageClassPhysicalStorageBufferEXT) {
if (containsType(type->getIdOperand(1), OpTypeInt, 8)) {
addExtension(spv::E_SPV_KHR_8bit_storage);
addCapability(spv::CapabilityStorageBuffer8BitAccess);
}
if (containsType(type->getIdOperand(1), OpTypeInt, 16) ||
containsType(type->getIdOperand(1), OpTypeFloat, 16)) {
addExtension(spv::E_SPV_KHR_16bit_storage);
addCapability(spv::CapabilityStorageBuffer16BitAccess);
}
}
}
// process all reachable instructions...
for (auto bi = reachableBlocks.cbegin(); bi != reachableBlocks.cend(); ++bi) {
const Block* block = *bi;
@ -366,24 +421,6 @@ void Builder::postProcess()
}
}
}
// Look for any 8/16 bit type in physical storage buffer class, and set the
// appropriate capability. This happens in createSpvVariable for other storage
// classes, but there isn't always a variable for physical storage buffer.
for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) {
Instruction* type = groupedTypes[OpTypePointer][t];
if (type->getImmediateOperand(0) == (unsigned)StorageClassPhysicalStorageBufferEXT) {
if (containsType(type->getIdOperand(1), OpTypeInt, 8)) {
addExtension(spv::E_SPV_KHR_8bit_storage);
addCapability(spv::CapabilityStorageBuffer8BitAccess);
}
if (containsType(type->getIdOperand(1), OpTypeInt, 16) ||
containsType(type->getIdOperand(1), OpTypeFloat, 16)) {
addExtension(spv::E_SPV_KHR_16bit_storage);
addCapability(spv::CapabilityStorageBuffer16BitAccess);
}
}
}
}
}; // end spv namespace

39
Test/baseResults/spv.float16convertonlyarith.comp.out Normal file
View File

@ -0,0 +1,39 @@
spv.float16convertonlyarith.comp
// Module Version 10000
// Generated by (magic number): 80007
// Id's are bound by 22
Capability Shader
Capability Float16
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint GLCompute 4 "main"
ExecutionMode 4 LocalSize 16 16 1
Source GLSL 450
SourceExtension "GL_EXT_shader_explicit_arithmetic_types_float16"
Name 4 "main"
Name 9 "v"
Decorate 21 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeFloat 32
7: TypeVector 6(float) 4
8: TypePointer Function 7(fvec4)
10: 6(float) Constant 0
11: 7(fvec4) ConstantComposite 10 10 10 10
13: TypeFloat 16
14: TypeVector 13(float16_t) 4
17: TypeInt 32 0
18: TypeVector 17(int) 3
19: 17(int) Constant 16
20: 17(int) Constant 1
21: 18(ivec3) ConstantComposite 19 19 20
4(main): 2 Function None 3
5: Label
9(v): 8(ptr) Variable Function
Store 9(v) 11
12: 7(fvec4) Load 9(v)
15: 14(f16vec4) FConvert 12
16: 7(fvec4) FConvert 15
Return
FunctionEnd

39
Test/baseResults/spv.float16convertonlystorage.comp.out Normal file
View File

@ -0,0 +1,39 @@
spv.float16convertonlystorage.comp
// Module Version 10000
// Generated by (magic number): 80007
// Id's are bound by 22
Capability Shader
Capability Float16
1: ExtInstImport "GLSL.std.450"
MemoryModel Logical GLSL450
EntryPoint GLCompute 4 "main"
ExecutionMode 4 LocalSize 16 16 1
Source GLSL 450
SourceExtension "GL_EXT_shader_16bit_storage"
Name 4 "main"
Name 9 "v"
Decorate 21 BuiltIn WorkgroupSize
2: TypeVoid
3: TypeFunction 2
6: TypeFloat 32
7: TypeVector 6(float) 4
8: TypePointer Function 7(fvec4)
10: 6(float) Constant 0
11: 7(fvec4) ConstantComposite 10 10 10 10
13: TypeFloat 16
14: TypeVector 13(float16_t) 4
17: TypeInt 32 0
18: TypeVector 17(int) 3
19: 17(int) Constant 16
20: 17(int) Constant 1
21: 18(ivec3) ConstantComposite 19 19 20
4(main): 2 Function None 3
5: Label
9(v): 8(ptr) Variable Function
Store 9(v) 11
12: 7(fvec4) Load 9(v)
15: 14(f16vec4) FConvert 12
16: 7(fvec4) FConvert 15
Return
FunctionEnd

11
Test/spv.float16convertonlyarith.comp Normal file
View File

@ -0,0 +1,11 @@
#version 450 core
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;
void main()
{
vec4 v = vec4(0.0);
vec4(f16vec4(v));
}

11
Test/spv.float16convertonlystorage.comp Normal file
View File

@ -0,0 +1,11 @@
#version 450 core
#extension GL_EXT_shader_16bit_storage : require
layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;
void main()
{
vec4 v = vec4(0.0);
vec4(f16vec4(v));
}

2
gtests/Spv.FromFile.cpp
View File

@ -300,6 +300,8 @@ INSTANTIATE_TEST_CASE_P(
"spv.earlyReturnDiscard.frag",
"spv.extPostDepthCoverage.frag",
"spv.extPostDepthCoverage_Error.frag",
"spv.float16convertonlyarith.comp",
"spv.float16convertonlystorage.comp",
"spv.flowControl.frag",
"spv.forLoop.frag",
"spv.forwardFun.frag",