AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge "[hip-vdi]Fix for TF build failure [SWDEV-225827]" into amd-master-next

Browse Source 
[ROCm/clr commit: 4cd02f20cf]
This commit is contained in:
Sarbojit Sarkar
2020-03-18 11:52:46 -04:00
committed by Gerrit Code Review
parent b77ba57b22 7cac7ff3b5
commit bfc7d1c663
5 changed files with 46 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/include/hip/hcc_detail/hip_runtime_api.h
+7
View File
@@ -3437,6 +3437,13 @@ hipError_t hipBindTextureToMipmappedArray(const texture<T, dim, readMode>& tex,
}
#if __HIP_VDI__ && !defined(__HCC__)
template <typename F>
inline hipError_t hipOccupancyMaxPotentialBlockSize(uint32_t* gridSize, uint32_t* blockSize,
F kernel, size_t dynSharedMemPerBlk, uint32_t blockSizeLimit) {
return hipOccupancyMaxPotentialBlockSize(gridSize, blockSize,(hipFunction_t)kernel, dynSharedMemPerBlk, blockSizeLimit);
}
template <class T>
inline hipError_t hipLaunchCooperativeKernel(T f, dim3 gridDim, dim3 blockDim,
void** kernelParams, unsigned int sharedMemBytes, hipStream_t stream) {
projects/clr/hipamd/tests/src/runtimeApi/module/hipOccupancyMaxPotentialBlockSize.cpp
+1 -3
View File
@@ -22,7 +22,7 @@ THE SOFTWARE.
// Test the Grid_Launch syntax.
/* HIT_START
* BUILD: %t %s ../../test_common.cpp EXCLUDE_HIP_PLATFORM nvcc vdi
* BUILD: %t %s ../../test_common.cpp EXCLUDE_HIP_PLATFORM nvcc
* TEST: %t
* HIT_END
*/
@@ -46,10 +46,8 @@ int main(int argc, char* argv[]) {
// test case for using kernel function pointer
uint32_t gridSize = 0;
uint32_t blockSize = 0;
#if defined(__HIP_PLATFORM_HCC__) && GENERIC_GRID_LAUNCH == 1 && defined(__HCC__)
hipOccupancyMaxPotentialBlockSize(&gridSize, &blockSize, f1, 0, 0);
assert(gridSize != 0 && blockSize != 0);
#endif
// test case for using kernel function pointer with template
gridSize = 0;
projects/clr/hipamd/vdi/hip_hcc.def.in
+1
View File
@@ -135,6 +135,7 @@ hipModuleLoad
hipModuleLoadData
hipModuleLoadDataEx
hipModuleUnload
hipOccupancyMaxPotentialBlockSize
hipOccupancyMaxActiveBlocksPerMultiprocessor
hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags
hipFuncGetAttributes
projects/clr/hipamd/vdi/hip_hcc.map.in
+1
View File
@@ -135,6 +135,7 @@ global:
hipModuleLoadData;
hipModuleLoadDataEx;
hipModuleUnload;
hipOccupancyMaxPotentialBlockSize;
hipOccupancyMaxActiveBlocksPerMultiprocessor;
hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags;
hipFuncGetAttributes;
projects/clr/hipamd/vdi/hip_platform.cpp
+36 -12
View File
@@ -604,13 +604,12 @@ hipError_t ihipCreateGlobalVarObj(const char* name, hipModule_t hmod, amd::Memor
namespace hip_impl {
hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(int* numBlocks,
hipFunction_t f,
int blockSize,
size_t dynamicSMemSize)
hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(int* numBlocks, int* numGrids,
hipFunction_t f, int blockSize,
size_t dynamicSMemSize, bool bCalcPotentialBlkSz)
{
HIP_INIT_API(NONE, f, blockSize, dynamicSMemSize);
HIP_INIT_API(NONE, f, blockSize, dynamicSMemSize, bCalcPotentialBlkSz);
if(numBlocks == nullptr){HIP_RETURN(hipErrorInvalidValue);}
int deviceId = ihipGetDevice();
// FIXME: Function may not be a device function and may have been obtaiend via
// hipModuleGetFunction and thus not in the functions_ map. Check the map
@@ -627,12 +626,16 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(int* numBlocks,
if (!kernel) {
HIP_RETURN(hipErrorOutOfMemory);
}
if (blockSize == 0) {
HIP_RETURN(hipErrorInvalidValue);
}
amd::Device* device = hip::getCurrentDevice()->devices()[0];
const device::Kernel::WorkGroupInfo* wrkGrpInfo = kernel->getDeviceKernel(*device)->workGroupInfo();
if (blockSize == 0) {
if (bCalcPotentialBlkSz == false){
HIP_RETURN(hipErrorInvalidValue);
}
else {
blockSize = device->info().maxWorkGroupSize_; // maxwavefrontperblock
}
}
// Find threads accupancy per CU => simd_per_cu * GPR usage
constexpr size_t MaxWavesPerSimd = 8; // Limited by SPI 32 per CU, hence 8 per SIMD
size_t VgprWaves = wrkGrpInfo->availableVGPRs_ / amd::alignUp(wrkGrpInfo->usedVGPRs_, 4);
@@ -658,12 +661,33 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(int* numBlocks,
int lds_occupancy = static_cast<int>(device->info().localMemSize_ / total_used_lds);
*numBlocks = std::min(*numBlocks, lds_occupancy);
}
if (bCalcPotentialBlkSz){
if (numGrids == nullptr){
HIP_RETURN(hipErrorInvalidValue);
}
*numGrids = *numBlocks * device->info().numRTCUs_;
}
HIP_RETURN(hipSuccess);
}
}
extern "C" {
// FIXME: Need to replace `uint32_t` with `int` finally.
hipError_t hipOccupancyMaxPotentialBlockSize(uint32_t* gridSize, uint32_t* blockSize,
hipFunction_t f, size_t dynSharedMemPerBlk,
uint32_t blockSizeLimit)
{
int numGrids = 0;
int numBlocks = 0;
hipError_t Ret = hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(&numBlocks, &numGrids, f, 0, dynSharedMemPerBlk,true);
if (Ret == hipSuccess){
*blockSize = numBlocks;
*gridSize = numGrids;
}
HIP_RETURN(Ret);
}
// FIXME: Need to replace `uint32_t` with `int` finally.
hipError_t hipOccupancyMaxActiveBlocksPerMultiprocessor(uint32_t* numBlocks,
hipFunction_t f,
@@ -671,7 +695,7 @@ hipError_t hipOccupancyMaxActiveBlocksPerMultiprocessor(uint32_t* numBlocks,
size_t dynamicSMemSize)
{
int NB;
hipError_t Ret = hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(&NB, f, blockSize, dynamicSMemSize);
hipError_t Ret = hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(&NB, nullptr, f, blockSize, dynamicSMemSize, false);
*numBlocks = NB;
HIP_RETURN(Ret);
}
@@ -684,7 +708,7 @@ hipError_t hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(uint32_t* numBl
unsigned int flags)
{
int NB;
hipError_t Ret = hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(&NB, f, blockSize, dynamicSMemSize);
hipError_t Ret = hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(&NB, nullptr, f, blockSize, dynamicSMemSize, false);
*numBlocks = NB;
HIP_RETURN(Ret);
}