1. They use SDWA + LLVM IR
2. Added these functions to test
3. Need to do exp, exp10, log, log10, rint
Change-Id: I06176acc6cb8bb054495310531777406a41b54e4
[ROCm/hip commit: 0e576295b4]
1. Added math functions for half precision
2. HRCP is not available due to device code linking errors, will be enabled once it is fixed
3. Added math functions to half test file
Change-Id: Ie317ce70ef518a4fc3f27142143d01e0327f5df3
[ROCm/hip commit: 8c978c210c]
1. Added SDWA implementation inside IR file
2. Added device functions to header + used them in test
Change-Id: Ib4e059a58eee201cc82438689e3e9bc5f9d26653
[ROCm/hip commit: 5ef8ef3bd7]
1. Removed HIP_EXPERIMENTAL env variable so that device code will be accessed from LLVM IR
2. Removed soft support from headers and moved to hip_fp16.cpp
3. Added LLVM IR + inline asm to hip_ir.ll
4. Added test for fp16
5. Added barriers for hcc 3.5 and hcc 4.0 for half support
a. Which means, hcc 4.0 can parse __fp16 but hcc 3.5 cant
b. HCC 4.0 code is implemented now, hcc 3.5 will be added later
Change-Id: Ic37859b2688ebb02e168bab643d1882bf4727952
[ROCm/hip commit: d180fdaae0]
All are marked as HIP_UNSUPPORTED.
IMPORTANT:
1. libraryPropertyType_t has no cuda prefix. => TO_DO: new matcher is needed.
2. all libraries (cublas, cufft, cusolver, cusparse, nvgraph) have started to use these types (since 8.0).
[ROCm/hip commit: 9a0780001b]
Includes some tricky manipulation of the locks for contexts and streams.
issue is that stealing a stream requires we lock the context to
walk the streams to find a victim. To avoid deadlock, we can't
have a stream locked when we lock the context. This implementation
releases the stream lock, then acquires the context and selects the
victim.
A more stable implemenation might be to copy the stream list
from a context so that a lock is not required to walk all streams.
Smart shared_ptr could be used to prevent the streams from being
deallocated during the walk.
[ROCm/hip commit: a3e0012567]
CUBLAS API 7.5 now is supported by hipify;
API calls, which are not yet supported by hcblas/hipblas, are listed as HIP_UNSUPPORTED.
[ROCm/hip commit: 6ceb85a03a]
The differences from the similar scripts for hipify.pl:
1. CSV file with extended statistics is produced.
2. scripts' arguments are changed a bit:
DIRNAME [hipify options] [--] [clang options]
where -- is a delimiter; all the arguments are optional, except DIRNAME.
Usage example:
./hipexamine2.sh ./tmp -o-stats ./tmp/stats.csv -- -I/usr/local/cuda-7.5/include -I/usr/local/hipify-clang/hipblas/include 2>&1 | tee log
[ROCm/hip commit: bcbbc32fa6]
+ REPLACED bytes, TOTAL bytes & CODE CHANGED are added to statistics.
+ -o-stats option for specifying the file with statistic.
[ROCm/hip commit: 6882057fd2]
https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/issues/53
+ Unsupported refs (by HIP) are now might be listed along with the supported ones.
+ Warnings are added for the unhandled (by HIPIFY) refs, for instance:
"warning: the following reference is not handled: 'cublasContext' [param decl ptr]."
+ Reflect unsupported CUDA API refs in statistics.
+ Occupancy API [HIP_UNSUPPORTED].
+ A few CUBLAS refs are listed as HIP_UNSUPPORTED.
TODO: Statistics in CSV file.
[ROCm/hip commit: 3dd32e969d]
- hipFunction_t is now returned by value. This eliminates dynamic
allocation / memory management complexity in the module. Removed
the kernel
name so the structure is just 16 bytes now.
- Moved the hsa_executable_load_module and hsa_executable_freeze
calls to the hipModuleLoad and hipModuleLoadData calls.
- Apply sharedMemBytes in hipModuleLaunchKernel to group segment
size (not private).
[ROCm/hip commit: 4d29885be3]
Aditya Atluri
Evgeny Mankov
Robert
Maneesh Gupta
Ben Sander
Rahul Garg
scchan