Skip to content

Commit 198c08e

Browse files
pkwasnie-inteligcbot
pkwasnie-intel
authored and
igcbot
committed
rename ShortImplicitPayloadHeader to GlobalOffset
Description: Compute workloads add following implicit arguments: 1. payloadHeader - 8 x i32 packing global_id_offset (3 x i32), local_size (3 x i32) and 2 x i32 reserved. 2. enqueued_local_size - 3 x i32 local_size is never used in favour of enqueued_local_size. In the end, payloadHeader has unused 20 bytes. shortPayloadHeader was introduced to reduce the size. It incorrectly used the same name "payloadHeader". To fix the issue, argument must change name. Instead of "shortPayloadHeader", it is renamed to "globalOffset", as this is exactly what it is. ---------------------------
1 parent 2286a31 commit 198c08e

File tree

14 files changed

+37
-17
lines changed

14 files changed

+37
-17
lines changed

‎IGC/AdaptorCommon/ImplicitArgs.cpp

+1-1
Original file line numberDiff line numberDiff line change
@@ -25,7 +25,7 @@ static const std::vector<ImplicitArg> IMPLICIT_ARGS = {
2525
ImplicitArg(ImplicitArg::R0, "r0", ImplicitArg::INT, WIAnalysis::UNIFORM_THREAD, 8, ImplicitArg::ALIGN_GRF, false, GenISAIntrinsic::GenISA_getR0),
2626

2727
ImplicitArg(ImplicitArg::PAYLOAD_HEADER, "payloadHeader", ImplicitArg::INT, WIAnalysis::UNIFORM_WORKGROUP, 8, ImplicitArg::ALIGN_GRF, true, GenISAIntrinsic::GenISA_getPayloadHeader),
28-
ImplicitArg(ImplicitArg::PAYLOAD_HEADER_SHORT, "payloadHeader", ImplicitArg::INT, WIAnalysis::UNIFORM_WORKGROUP, 3, ImplicitArg::ALIGN_DWORD, true, GenISAIntrinsic::GenISA_getPayloadHeader),
28+
ImplicitArg(ImplicitArg::GLOBAL_OFFSET, "globalOffset", ImplicitArg::INT, WIAnalysis::UNIFORM_WORKGROUP, 3, ImplicitArg::ALIGN_DWORD, true, GenISAIntrinsic::GenISA_getGlobalOffset),
2929
ImplicitArg(ImplicitArg::WORK_DIM, "workDim", ImplicitArg::INT, WIAnalysis::UNIFORM_GLOBAL, 1, ImplicitArg::ALIGN_DWORD, true, GenISAIntrinsic::GenISA_getWorkDim),
3030

3131
ImplicitArg(ImplicitArg::NUM_GROUPS, "numWorkGroups", ImplicitArg::INT, WIAnalysis::UNIFORM_GLOBAL, 3, ImplicitArg::ALIGN_DWORD, true, GenISAIntrinsic::GenISA_getNumWorkGroups),

‎IGC/AdaptorCommon/ImplicitArgs.hpp

+1-1
Original file line numberDiff line numberDiff line change
@@ -38,7 +38,7 @@ namespace IGC
3838
R0,
3939
START_ID = R0,
4040
PAYLOAD_HEADER,
41-
PAYLOAD_HEADER_SHORT, // reduces PayloadHeader to 3xi32
41+
GLOBAL_OFFSET, // previously packed in 8xi32 PayloadHeader
4242

4343
// WI information
4444
WORK_DIM,

‎IGC/Compiler/CISACodeGen/AdvCodeMotion.cpp

+2
Original file line numberDiff line numberDiff line change
@@ -168,6 +168,8 @@ void WorkItemSetting::collect(Function* F) {
168168
GroupId = getXYZ(&*AI, 1, 6, 7);
169169
else if (Name == "payloadHeader")
170170
GlobalOffset = getXYZ(&*AI, 0, 1, 2);
171+
else if (Name == "globalOffset")
172+
GlobalOffset = getXYZ(&*AI, 0, 1, 2);
171173
else if (Name == "globalSize")
172174
GlobalSize = getXYZ(&*AI, 0, 1, 2);
173175
else if (Name == "globalSize1")

‎IGC/Compiler/CISACodeGen/CShader.cpp

+1-1
Original file line numberDiff line numberDiff line change
@@ -2841,7 +2841,7 @@ CVariable* CShader::getOrCreateArgumentSymbol(
28412841
// optimization, with some advanced analysis.
28422842
if (ArgType == ImplicitArg::ArgType::R0 ||
28432843
ArgType == ImplicitArg::ArgType::PAYLOAD_HEADER ||
2844-
ArgType == ImplicitArg::ArgType::PAYLOAD_HEADER_SHORT ||
2844+
ArgType == ImplicitArg::ArgType::GLOBAL_OFFSET ||
28452845
ArgType == ImplicitArg::ArgType::WORK_DIM ||
28462846
ArgType == ImplicitArg::ArgType::NUM_GROUPS ||
28472847
ArgType == ImplicitArg::ArgType::GLOBAL_SIZE ||

‎IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

+2-1
Original file line numberDiff line numberDiff line change
@@ -9557,6 +9557,7 @@ void EmitPass::EmitGenIntrinsicMessage(llvm::GenIntrinsicInst* inst)
95579557
break;
95589558
case GenISAIntrinsic::GenISA_getR0:
95599559
case GenISAIntrinsic::GenISA_getPayloadHeader:
9560+
case GenISAIntrinsic::GenISA_getGlobalOffset:
95609561
case GenISAIntrinsic::GenISA_getWorkDim:
95619562
case GenISAIntrinsic::GenISA_getNumWorkGroups:
95629563
case GenISAIntrinsic::GenISA_getLocalSize:
@@ -22921,7 +22922,7 @@ void EmitPass::emitImplicitArgIntrinsic(llvm::GenIntrinsicInst* I)
2292122922

2292222923
if (IAtype == ImplicitArg::ArgType::R0 ||
2292322924
IAtype == ImplicitArg::ArgType::PAYLOAD_HEADER ||
22924-
IAtype == ImplicitArg::ArgType::PAYLOAD_HEADER_SHORT ||
22925+
IAtype == ImplicitArg::ArgType::GLOBAL_OFFSET ||
2292522926
IAtype == ImplicitArg::ArgType::WORK_DIM ||
2292622927
IAtype == ImplicitArg::ArgType::NUM_GROUPS ||
2292722928
IAtype == ImplicitArg::ArgType::GLOBAL_SIZE ||

‎IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp

+3-5
Original file line numberDiff line numberDiff line change
@@ -919,9 +919,7 @@ namespace IGC
919919
zebin::PreDefinedAttrGetter::ArgType::local_size, cur_pos, size);
920920
break;
921921
}
922-
case KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER_SHORT: {
923-
// PayloadHeader contains global work offset x,y,z
924-
// global work offset size is int32x3
922+
case KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET: {
925923
uint32_t size = iOpenCL::DATA_PARAMETER_DATA_SIZE * 3;
926924
zebin::ZEInfoBuilder::addPayloadArgumentImplicit(m_kernelInfo.m_zePayloadArgs,
927925
zebin::PreDefinedAttrGetter::ArgType::global_id_offset, payloadPosition, size);
@@ -1520,7 +1518,7 @@ namespace IGC
15201518
break;
15211519

15221520
case KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER:
1523-
case KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER_SHORT:
1521+
case KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET:
15241522
// PayloadHeader contains global work offset x,y,z and local size x,y,z -->
15251523
// total of 6 annotations, 3 of each type
15261524
// Short PayloadHeader reduces it to only global work offset
@@ -2458,7 +2456,7 @@ namespace IGC
24582456
{
24592457
IsUnusedArg |=
24602458
(arg.getArgType() == KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER || // contains global_id_offset
2461-
arg.getArgType() == KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER_SHORT ||
2459+
arg.getArgType() == KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET ||
24622460
arg.getArgType() == KernelArg::ArgType::IMPLICIT_ENQUEUED_LOCAL_WORK_SIZE) &&
24632461
arg.getArg()->use_empty();
24642462
}

‎IGC/Compiler/CISACodeGen/Platform.hpp

+6
Original file line numberDiff line numberDiff line change
@@ -1897,11 +1897,17 @@ bool allowProceedBasedApproachForRayQueryDynamicRayManagementMechanism() const
18971897
return IGC_IS_FLAG_DISABLED(DisableProceedBasedApproachForRayQueryDynamicRayManagementMechanism);
18981898
}
18991899

1900+
// Payload header is 8xi32 kernel argument packing 3xi32 global offset.
1901+
// This function controls if payload header can be replaced with direct
1902+
// use of global offset.
19001903
bool allowShortImplicitPayloadHeader() const
19011904
{
19021905
if (IGC_IS_FLAG_SET(ShortImplicitPayloadHeader))
19031906
return IGC_IS_FLAG_ENABLED(ShortImplicitPayloadHeader);
19041907

1908+
if (!supportsZEBin())
1909+
return false;
1910+
19051911
return false;
19061912
}
19071913

‎IGC/Compiler/CISACodeGen/WIAnalysis.cpp

+2
Original file line numberDiff line numberDiff line change
@@ -1481,6 +1481,7 @@ WIAnalysis::WIDependancy WIAnalysisRunner::calculate_dep(const CallInst* inst)
14811481
GII_id == GenISAIntrinsic::GenISA_vectorUniform ||
14821482
GII_id == GenISAIntrinsic::GenISA_getR0 ||
14831483
GII_id == GenISAIntrinsic::GenISA_getPayloadHeader ||
1484+
GII_id == GenISAIntrinsic::GenISA_getGlobalOffset ||
14841485
GII_id == GenISAIntrinsic::GenISA_getWorkDim ||
14851486
GII_id == GenISAIntrinsic::GenISA_getNumWorkGroups ||
14861487
GII_id == GenISAIntrinsic::GenISA_getLocalSize ||
@@ -1535,6 +1536,7 @@ WIAnalysis::WIDependancy WIAnalysisRunner::calculate_dep(const CallInst* inst)
15351536
return WIAnalysis::UNIFORM_THREAD;
15361537
case GenISAIntrinsic::GenISA_getR0:
15371538
case GenISAIntrinsic::GenISA_getPayloadHeader:
1539+
case GenISAIntrinsic::GenISA_getGlobalOffset:
15381540
case GenISAIntrinsic::GenISA_getWorkDim:
15391541
case GenISAIntrinsic::GenISA_getNumWorkGroups:
15401542
case GenISAIntrinsic::GenISA_getLocalSize:

‎IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs/KernelArgs.cpp

+4-4
Original file line numberDiff line numberDiff line change
@@ -253,8 +253,8 @@ KernelArg::ArgType KernelArg::calcArgType(const ImplicitArg& arg) const
253253
return KernelArg::ArgType::IMPLICIT_R0;
254254
case ImplicitArg::PAYLOAD_HEADER:
255255
return KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER;
256-
case ImplicitArg::PAYLOAD_HEADER_SHORT:
257-
return KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER_SHORT;
256+
case ImplicitArg::GLOBAL_OFFSET:
257+
return KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET;
258258
case ImplicitArg::PRIVATE_BASE:
259259
return KernelArg::ArgType::IMPLICIT_PRIVATE_BASE;
260260
case ImplicitArg::CONSTANT_BASE:
@@ -809,7 +809,7 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
809809

810810
KernelArg::ArgType::RUNTIME_VALUE,
811811
KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER,
812-
KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER_SHORT,
812+
KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET,
813813

814814
KernelArg::ArgType::PTR_LOCAL,
815815
KernelArg::ArgType::PTR_GLOBAL,
@@ -936,7 +936,7 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
936936

937937
KernelArg::ArgType::RUNTIME_VALUE,
938938
KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER,
939-
KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER_SHORT,
939+
KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET,
940940
KernelArg::ArgType::PTR_LOCAL,
941941
KernelArg::ArgType::PTR_GLOBAL,
942942
KernelArg::ArgType::PTR_CONSTANT,

‎IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs/KernelArgs.hpp

+1-1
Original file line numberDiff line numberDiff line change
@@ -44,7 +44,7 @@ namespace IGC
4444
R1,
4545

4646
IMPLICIT_PAYLOAD_HEADER, // known as INPUT_HEADER in USC
47-
IMPLICIT_PAYLOAD_HEADER_SHORT, // payload header reduced to 3xi32
47+
IMPLICIT_GLOBAL_OFFSET, // previously packed in payload header
4848

4949
PTR_LOCAL,
5050
PTR_GLOBAL,

‎IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.cpp

+2-1
Original file line numberDiff line numberDiff line change
@@ -513,7 +513,7 @@ Value* WIFuncResolution::getGlobalOffset(CallInst& CI)
513513
// %globalOffset = extractelement <8 x i32> %payloadHeader, i32 %dim
514514

515515
auto Ty = getAnalysis<CodeGenContextWrapper>().getCodeGenContext()->platform.allowShortImplicitPayloadHeader()
516-
? ImplicitArg::PAYLOAD_HEADER_SHORT : ImplicitArg::PAYLOAD_HEADER;
516+
? ImplicitArg::GLOBAL_OFFSET : ImplicitArg::PAYLOAD_HEADER;
517517
auto F = CI.getFunction();
518518
Value* V = m_implicitArgs.getImplicitArgValue(*F, Ty, m_pMdUtils);
519519
IGC_ASSERT(V != nullptr);
@@ -870,6 +870,7 @@ void LowerImplicitArgIntrinsics::visitCallInst(CallInst& CI)
870870
break;
871871
}
872872
case GenISAIntrinsic::GenISA_getPayloadHeader:
873+
case GenISAIntrinsic::GenISA_getGlobalOffset:
873874
{
874875
// global_offset is loaded from PayloadHeader[0:2]
875876
// currently there are no other uses for payload header.

‎IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncsAnalysis.cpp

+2-1
Original file line numberDiff line numberDiff line change
@@ -97,7 +97,8 @@ bool WIFuncsAnalysis::runOnFunction(Function& F)
9797
SmallVector<ImplicitArg::ArgType, ImplicitArg::NUM_IMPLICIT_ARGS> implicitArgs;
9898

9999
const bool RequirePayloadHeader = m_ctx->m_DriverInfo.RequirePayloadHeader();
100-
const auto PayloadHeaderType = m_ctx->platform.allowShortImplicitPayloadHeader() ? ImplicitArg::PAYLOAD_HEADER_SHORT : ImplicitArg::PAYLOAD_HEADER;
100+
// 8xi32 payload header packs 3xi32 global offset. If possible, use global offset directly.
101+
const auto PayloadHeaderType = m_ctx->platform.allowShortImplicitPayloadHeader() ? ImplicitArg::GLOBAL_OFFSET : ImplicitArg::PAYLOAD_HEADER;
101102

102103
// All OpenCL kernels receive R0 and Payload Header implicitly
103104
if (isEntryFunc(m_pMDUtils, &F))

‎IGC/GenISAIntrinsics/generator/input/Intrinsic_definitions.yml

+9
Original file line numberDiff line numberDiff line change
@@ -5635,6 +5635,15 @@ intrinsics:
56355635
arguments: []
56365636
attributes:
56375637
- !AttributeID "NoUnwind"
5638+
- !<IntrinsicDefinition>
5639+
name: "GenISA_getGlobalOffset"
5640+
comment: ""
5641+
return_definition: !<ReturnDefinition>
5642+
type_definition: *v_any_
5643+
comment: "result"
5644+
arguments: []
5645+
attributes:
5646+
- !AttributeID "NoUnwind"
56385647
- !<IntrinsicDefinition>
56395648
name: "GenISA_getWorkDim"
56405649
comment: ""

‎IGC/common/igc_flags.h

+1-1
Original file line numberDiff line numberDiff line change
@@ -942,7 +942,7 @@ DECLARE_IGC_REGKEY_ENUM(SupportUniformPrivateMemorySpace, -1, \
942942
" 0 - force disabled" \
943943
" 1 - force enabled", \
944944
TRIBOOL_OPTIONS, true)
945-
DECLARE_IGC_REGKEY(bool, ShortImplicitPayloadHeader, false, "Reduces implicit payload header arg from 8xi32 to 3xi32", true)
945+
DECLARE_IGC_REGKEY(bool, ShortImplicitPayloadHeader, false, "Replaces implicit kernel argument 8xi32 payloadHeader with 3xi32 globalOffset", true)
946946
DECLARE_IGC_REGKEY(bool, RemoveUnusedIdImplicitArguments, false, "Remove implicit arguments: global_id_offset (payloadHeader) and/or enqueued_local_size if unused. " \
947947
"Useful if kernel doesn't use global id.", true)
948948

0 commit comments

Comments
 (0)