I don't understand this. llvm_test_verify_hash_program_output() only inserts a hashing step to the verification script, you still need the final fpcmp invocation (and the WORKDIR setting is separate for every step).

In the definition of llvm_test_verify_hash_program_output(), llvm_test_verify(WORKDIR ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_SOURCE_DIR}/HashProgramOutput.sh ${_file}) is called which sets the work directory to current binary directory. So we need not call llvm_test_verify with WORKDIR ${CMAKE_CURRENT_BINARY_DIR} argument after llvm_test_verify_hash_program_output, simply llvm_test_verify(${FPCMP} bilinear.reference_output bilinearOutput.txt) will work. Passing WORKDIR argument does not cause any harm though.

After applying this patch, the tests fail for me because it is taking the hash of stdout instead of the file mentioned in llvm_test_verify_hash_program_output macro.
Changing line 64 of cmake/modules/SingleMultiSource.cmake to ${CMAKE_SOURCE_DIR}/HashProgramOutput.sh ${file} will fix this.

Reduced test size
Total runtime is still there for about 5.5s (with benchmark library) because the benchmark library runs kernel for multiple inputs.

Thank you for the solution Santosh, It seems like my GitHub account was actually linked to some other account on Phabricator which I was not aware of, so I didn't get any notification for this patch on this account. I will try to get that account removed so that this won't happen in future.

made changes requested by @Meinersbur

Adding each kernel in separate patches so that it will be easy to review and also we can revert only that kernel which causes problems (if any).

Changes:

• Use glibc_compat_rand.c from Common folder of Rodinia
• Use float * instead of the pointer to VLA in main.c
• Changed CMakeLists.txt

changed CMakeLists.txt

Changes:
removed typedef int matrix[row][col]
compare the hash of output instead of checking all values

reverted back to int *arr from int (*arr)[][].

added some comments

changed array type from int (*) [][] to int * in main.cpp

Changes: update m along with image size when called with the benchmark library. For verification, only m=4 is used.

Changes: Used C99 VLAs in kernels.

Renamed glibc_compat_rand.cpp back to glibc_compat_rand.c and fixed the link erorr mentioned earlier in summary (when glibc_compat_rand.c was used) .

made changes requested by @Meinersbur

The kernel runs for 0.79 sec and the complete program runs for 0.81 sec.

It runs for 1.1 sec for normal problem size and 0.6 sec for smaller
The main kernel function takes around 0.7-0.8 sec of 1.1 sec.
Polly detects scops here but runtime is still same

Changed function parameters from "int image[HEIGHT][WIDTH]" to "int *image"

In D49339#1163029, @dberris wrote:

I'm not sure why these utilities are being added, is there some other patch that will be using this?

Yes. They are added as a separate patch and this patch will be mentioned if they are using this.

added LICENSE and fixed work directory issue.

Added -ffast-math flag... Polly recognizes entire kernel but still no performance improvement with Polly.

In D47675#1138707, @dberris wrote:

Thanks for making some of the changes. I'm still not clear on a couple of things.

Do you mind sharing some of the results with the new benchmark runs, with the different image sizes? Do we actually get the throughput numbers in there as well?

Formatted using clang format

Runtime on this differential

In D47675#1131201, @Meinersbur wrote:

Could you check whether llvm-lit correctly collects execution time, compile/link time, LLVM -stats, code size?

I don't know how to check LLVM -stats using lit.
Sizes matches the output of llvm-size, compile time and link time are also fine.

Updated input size, used malloc to allocate memory for the array.

In D47675#1129087, @Meinersbur wrote:

Looks great. Did you do a performance comparison with/without Polly?

Polly + O3 and only O3 are taking the same time. It seems like before code reaches Polly, It is already heavily optimized at O3 and Polly cannot find any further optimization possible on it even though it is in its SCoP.

Fixed "backprop.h" showing as exec.