EPR #414091 - Compiler Lib/HSAIL: Fix for -fno-bin-cg on multiple kernels
[Problem] If CL or BRIG contains multiple kernels, compilation with option -fno-bin-cg leads to finalizer error.
[Solution] Check that all kernels are finalized before deleting BRIG from BIF. Actual deletion occurs after finalizing the last kernel.
[Misc]
+ fix erroneous successful exit from finalizing BRIG -> ISA function when BRIG is not extracted successfully
+ string constants for symbol prefixes replaced on using findBIF30SymStruct
+ obsolete __AMP_ is removed from code
+ cosmetic changes in hsail_be
+ test for the fix is added: ursa.pl -t complib -M-fno-bin-cg-g_multiple_kernels
[Tests] pre check-in, make smoke, ursa complib, bingen test (from Benjamin Coquelle)
[Reviewers] Brian Sumner, Artem Tamazov, Matthew Arsenault
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#60 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.cpp#37 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.hpp#13 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-cg-g/HelloWorld_Kernel_cl.cl#2 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-cg/HelloWorld_Kernel_cl.cl#2 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg-g/HelloWorld_Kernel_cl.cl#2 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg-g_multiple_kernels/MM_Kernels.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg/HelloWorld_Kernel_cl.cl#2 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/tlst/complib.tlst#6 edit
ECR #333753 - Compiler Lib/ORCA RT/Performance: BRIG in BIF is not needed anymore for finalizing & executing ISA (except cases with program scope global variables).
[Description] The feature is intended to minimize the binary size for execution.
This is half-hearted solution: If -fno-bin-cg is set, then after ISA finalization all BRIG sections are removed from the binary, but if there are program scope global variables in code, the option is ignored and all BRIG sections retain in binary for further globals allocation & initialization. The complete solution awaits Code Objects implementation.
+ Additionally change fixes Bug 10478.
+ Recompilation steps determination is changed in RT.
+ symBRIGLoaderMap is added to bif_section_labels.
+ RT_CONTAINS_LOADER_MAP is added for aclQueryInfo in order to check symBRIGLoaderMap existance in binary.
+ complib tests are added on -fbin-cg/-fno-bin-cg.
+ ocltst -t complib -M CLEnumCheck is updated.
[Side effects] performance improvement, memory consumption reduction
[TODO] Do the same on .hsa bits.
[Testing] pre check-in, make smoke, complib, ocltst: complib, compiler, runtime, binary
[Reviewer] German Andryeyev
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#59 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/brig_loader.cpp#17 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.cpp#36 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.hpp#12 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/scwrapper/scClientAPI.cpp#21 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclEnums.h#16 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/bif_section_labels.hpp#19 edit
... //depot/stg/opencl/drivers/opencl/compiler/loader/libloader/loader.cpp#12 edit
... //depot/stg/opencl/drivers/opencl/compiler/tools/aoc2/aoc2.cpp#64 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuprogram.cpp#188 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-cg-g/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-cg-g_globals/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-cg/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-cg_globals/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg-g/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg-g_globals/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-cg_globals/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/tlst/complib.tlst#5 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLEnumCheck.cpp#40 edit
ECR #333753 - Partial fix for Bug 10478 "Fix -fno-bin-llvmir/-fno-bin-hsail options"
If option -fno-bin-llvmi is set, .llvmir section is deleted from BIF on CG phase instead of FE. Both HSA & AMDIL are affected.
[Fixed] -fno-bin-llvm option causes clBuildProgram fail with error -11.
Took place only if compiled from OpenCL
[TODO] If possible -fno-bin-hsail should avoid putting HSAIL binary (BRIG) into BIF.
[Tests] pre check-in, make smoke, complib
[Reviewers] Brian Sumner, Nikolay Haustov
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/frontend.cpp#31 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/frontend_clang.cpp#17 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#58 edit
... //depot/stg/opencl/drivers/opencl/compiler/tools/aoc2/aoc2.cpp#63 edit
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fbin-llvmir/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/src/complib/options/-fno-bin-llvmir/HelloWorld_Kernel_cl.cl#1 add
... //depot/stg/opencl/drivers/opencl/tests/hsa/tlst/complib.tlst#3 edit
ECR #333753 - hsa_foundation RT/Compiler Lib: recompilation algorithm rework
1. Recompilation algorithm rework in order to avoid superfluous recompilations.
2. Replace aclExtractSymbol/Section with aclQueryInfo for symbol/section detection.
The replaced calls in RT previously performed actual extraction of the sections from the BIF with memory allocation and copying. But what is needed in fact is only to determine whether the section exists in BIF or not to make a further decision on needed recompilations. With aclQueryInfo and new added enums RT_CONTAINS_LLVMIR, RT_CONTAINS_OPTIONS, RT_CONTAINS_BRIG, RT_CONTAINS_HSAIL, RT_CONTAINS_ISA Runtime starts querying not the whole sections but the bool flag which indicates the existence of the corresponding section(s) without any memory allocations. Every compilation on RT starting from LLVMIR is affected by the change including compilation of blit kernels.
3. Fix in Compiler Lib for correct ACL_INVALID_ARG detection (for wrong/unsupported compilations).
[Side Effects] performance improvement, memory consumption reduction
[ToDo] Do not finalize program if ISA is already provided in BIF and options are unchanged.
[Testing] pre check-in, ocltst complib, ocl conformance 2.0 compiler & api
[Reviewers] German Andryeyev, Brian Sumner
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#56 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsaprogram.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsaprogram.hpp#3 edit
ECR #333753 - Performance: Stop obligatory BRIG disassembling to HSAIL
[Important]: HSAIL is not being disassembled from BRIG and not being inserted into BIF anymore by default.
Testing: pre check-in, smoke_clang
Reviewers: Stanislav Mekhanoshin, Brian Sumner, Artem Tamazov
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#55 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.cpp#33 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.hpp#11 edit
... //depot/stg/opencl/drivers/opencl/compiler/tools/aoc2/aoc2.cpp#60 edit
ECR #333753 - new tests are added to ocltst -m oclcomplib -t OCLRTGetInfo
tests on aclQueryInfo for:
RT_KERNEL_NAMES, RT_CONTAINS_LLVMIR, RT_CONTAINS_OPTIONS, RT_CONTAINS_BRIG, RT_CONTAINS_HSAIL, RT_CONTAINS_ISA
+ query for RT_CONTAINS_HSAIL is fixed in Compiler Lib: looking for symbol symHSAILText instead of section aclCODEGEN, because aclCODEGEN section may contain also symOpenclMeta, symOpenclKernel, symOpenclStub besides symHSAILText.
Testing: pre check-in, ocltst -m oclcomplib
Reviewer: Brian Sumner
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#53 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/oclrtGetInfo.cpp#16 edit
ECR #333753 - Compiler Lib/RT/Performance: Replace aclExtractSymbol/Section with aclQueryInfo for symbol/section detection.
The replaced calls in RT previously performed actual extraction of the sections from the BIF with memory allocation and copying. But what is needed in fact is only to determine whether the section exists in BIF or not to make a further decision on needed recompilations. With aclQueryInfo and new added enums RT_CONTAINS_LLVMIR, RT_CONTAINS_OPTIONS, RT_CONTAINS_BRIG, RT_CONTAINS_HSAIL, RT_CONTAINS_ISA Runtime starts querying not the whole sections but the bool flag which indicates the existance of the corresponding section(s) without any memory allocations. Every compilation on RT starting from LLVMIR is affected by the change including compilation of blit kernels.
Side Effects: performance improvement, memory consumption reduction
Testing: pre check-in, ocl conformance (api, basic, compiler), ocltst complib
Reviewers: Brian Sumner, German Andryeyev, Artem Tamazov
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/api/v0_8/acl.cpp#21 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#52 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclEnums.h#14 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuprogram.cpp#180 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLEnumCheck.cpp#38 edit
ECR #333753 - HSA RT/Compiler Lib/Performance: Elimination of HSAIL text usage in RT
Extracting HSAIL from the binary and parsing it for the kernel names in RT were replaced with aclQueryInfo call for RT_KERNEL_NAMES.
Kernel names are obtained now from the corresponding metadata symbols names, which are already presented in BIF at kernel finalization stage.
Side effect: performance improvement
Next Step: Performance: elimination of BRIG disassembling to HSAIL as obligatory stage in Compiler Lib (previously was needed only by RT).
Testing: pre check-in, ocl conformnace 2.0 (basic, api, compiler, workgroups, device_execution)
Reviewers: Stanislav Mekhanoshin, German Andryeyev, Brian Sumner
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/api/v0_8/acl.cpp#20 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#51 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclEnums.h#13 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/loaders/bif/bifbase.cpp#50 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/loaders/bif/bifbase.hpp#22 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/v0_8/libUtils.h#11 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#266 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuprogram.cpp#179 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLEnumCheck.cpp#37 edit
ECR #333753 - Compiler Lib/RT: Metadata related code refactor, annotation, minor fixes & additional checks
+ refactor if_aclQueryInfo() in order to simplify code and to avoid direct usage of aclMetadata struct members types
+ annotation on why we need to use deserializeCLMetadata on "serialized" (to NULL) pointers
+ erroneously forgotten RT_KERNEL_NAME was added to aclQueryType enum
+ OCLRTGetInfo, CLEnumCheck tests from ocltst oclcomplib was updated to use RT_KERNEL_NAME
+ testing of printf is added to OCLRTGetInfo
+ minor fixes and additional checks
tests: pre check-in, ocltst -m oclcomplib
Reviewers: Artem Tamazov, Brian Sumner, German Andryeyev
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#49 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclEnums.h#12 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclStructs.h#12 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#265 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLEnumCheck.cpp#36 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/clSourceShaders.h#5 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/oclrtGetInfo.cpp#14 edit
ECR #333753 - Compiler Lib: improve & refactor HSAIL text routines
+ HSAIL text is always being inserted into BIF now in one place of Codegen phase
+ AMDIL & HSAIL paths are unified at Codegen phase
+ Error handling is improved
Testing: make smoke_clang, pre check-in
Reviewers: Brian Sumner, Nikolay Haustov
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#47 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.cpp#32 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.hpp#10 edit
ECR #333753 - RS compilation path.
RS compilation will be in 2 stages, first generates BRIG and the next is done via HSA Finalize API (that involves a load step).
Existing code in compiler/lib has a bug in that when the final output expected is HSAIL_BINARY, compilation should stop with invoking the llvm compiler (and the built-in assembler), not go all the way to ISA.
Tests: precheckin, hsa smoke
hsa/tests/RS/ test harness will be changed in a separate changelist.
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#44 edit
... //depot/stg/opencl/drivers/opencl/compiler/tools/aoc2/aoc2.cpp#57 edit
ECR #333753 - Compiler Lib: Error if wrong combination of march & OpenCL version
HSAIL doesn't support OpenCL version < 2.0
AMDIL doesn't support OpenCL version >= 2.0
Affects only ORCA builds (both 1.2 & 2.0), doesn't affect .hsa build.
Testing: smoke_clang, pre check-in
Reviewer: Stanislav Mekhanoshin
Affected files ...
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#38 edit
foreman