Merge pull request #121 from sampath1117/sr/num_threads_update

Default value for numThreads used for OpenMp

Readme.md

@@ -139,7 +139,7 @@ AMD ROCm Performance Primitives (**RPP**) library is a comprehensive high-perfor
   sudo make install
   ```
 
-## Build & Install RPP 
+## Build & Install RPP
 
 The ROCm Performance Primitives (RPP) library has support for three backends: HIP, OpenCL, and CPU:
 
@@ -221,6 +221,9 @@ $ sudo make install
 
     // Create handle
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

include/rpp.h

@@ -103,7 +103,7 @@ extern "C" SHARED_PUBLIC rppStatus_t rppCreate(rppHandle_t* handle);
 // *param[in] nBatchSize Batch size
 // *param[in] numThreads number of threads to be used for OpenMP pragma
 // *returns a rppStatus_t enumeration.
-extern "C" SHARED_PUBLIC rppStatus_t rppCreateWithBatchSize(rppHandle_t* handle, size_t nBatchSize, Rpp32u numThreads);
+extern "C" SHARED_PUBLIC rppStatus_t rppCreateWithBatchSize(rppHandle_t* handle, size_t nBatchSize, Rpp32u numThreads = 0);
 
 /******************** rppDestroy ********************/
 

src/include/common/rpp/handle.hpp

@@ -66,7 +66,7 @@ using rocblas_handle_ptr = RPP_MANAGE_PTR(rocblas_handle, rocblas_destroy_handle
 struct Handle : rppHandle
 {
     Handle();
-    Handle(size_t nBatchSize, Rpp32u numThreads);
+    Handle(size_t nBatchSize, Rpp32u numThreads = 0);
     Handle(Handle&&) noexcept;
     ~Handle();
 

src/modules/handlehost.cpp

@@ -29,6 +29,7 @@
 #include <unistd.h>
 #endif
 
+#include<thread>
 #include "config.h"
 #include "rpp/logger.hpp"
 #include "rpp/handle.hpp"
@@ -46,9 +47,6 @@ struct HandleImpl
     void PreInitializeBufferCPU()
     {
         this->initHandle = new InitHandle();
-        if(this->numThreads == 0)
-            this->numThreads = this->nBatchSize;
-
         this->initHandle->nbatchSize = this->nBatchSize;
         this->initHandle->mem.mcpu.maxSrcSize = (RppiSize *)malloc(sizeof(RppiSize) * this->nBatchSize);
         this->initHandle->mem.mcpu.maxDstSize = (RppiSize *)malloc(sizeof(RppiSize) * this->nBatchSize);
@@ -60,6 +58,7 @@ struct HandleImpl
 Handle::Handle(size_t batchSize, Rpp32u numThreads) : impl(new HandleImpl())
 {
     impl->nBatchSize = batchSize;
+    numThreads = std::min(numThreads, std::thread::hardware_concurrency());
     if(numThreads == 0)
         numThreads = batchSize;
     impl->numThreads = numThreads;
@@ -69,6 +68,9 @@ Handle::Handle(size_t batchSize, Rpp32u numThreads) : impl(new HandleImpl())
 Handle::Handle() : impl(new HandleImpl())
 {
     impl->PreInitializeBufferCPU();
+    impl->numThreads = std::min(impl->numThreads, std::thread::hardware_concurrency());
+    if(impl->numThreads == 0)
+        impl->numThreads = impl->nBatchSize;
     RPP_LOG_I(*this);
 }
 

src/modules/hip/handlehip.cpp

@@ -29,6 +29,7 @@
 #include <unistd.h>
 #endif
 
+#include<thread>
 #include "config.h"
 #include "rpp/device_name.hpp"
 #include "rpp/errors.hpp"
@@ -172,8 +173,6 @@ struct HandleImpl
     void PreInitializeBufferCPU()
     {
         this->initHandle = new InitHandle();
-        if(this->numThreads == 0)
-            this->numThreads = this->nBatchSize;
 
         this->initHandle->nbatchSize = this->nBatchSize;
         this->initHandle->mem.mcpu.srcSize = (RppiSize *)malloc(sizeof(RppiSize) * this->nBatchSize);
@@ -283,6 +282,7 @@ Handle::Handle(rppAcceleratorQueue_t stream) : impl(new HandleImpl())
 Handle::Handle(size_t batchSize, Rpp32u numThreads) : impl(new HandleImpl())
 {
     impl->nBatchSize = batchSize;
+    numThreads = std::min(numThreads, std::thread::hardware_concurrency());
     if(numThreads == 0)
         numThreads = batchSize;
     impl->numThreads = numThreads;
@@ -303,6 +303,9 @@ Handle::Handle() : impl(new HandleImpl())
 #endif
     this->SetAllocator(nullptr, nullptr, nullptr);
     impl->PreInitializeBuffer();
+    impl->numThreads = std::min(impl->numThreads, std::thread::hardware_concurrency());
+    if(impl->numThreads == 0)
+        impl->numThreads = impl->nBatchSize;
     RPP_LOG_I(*this);
 }
 

utilities/rpp-performancetests/HIP_NEW/Single_host.cpp

@@ -176,6 +176,9 @@ int main(int argc, char **argv)
 
     closedir(dr2);
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-performancetests/HOST_NEW/BatchPD_host_pkd3.cpp

@@ -645,6 +645,9 @@ int main(int argc, char **argv)
     }
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-performancetests/HOST_NEW/BatchPD_host_pln1.cpp

@@ -646,6 +646,9 @@ int main(int argc, char **argv)
     }
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-performancetests/HOST_NEW/BatchPD_host_pln3.cpp

@@ -748,6 +748,9 @@ int main(int argc, char **argv)
     }
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-performancetests/HOST_NEW/Single_host.cpp

@@ -176,6 +176,9 @@ int main(int argc, char **argv)
 
     closedir(dr2);
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-performancetests/HOST_NEW/Tensor_host_pkd3.cpp

@@ -595,6 +595,9 @@ int main(int argc, char **argv)
     // Run case-wise RPP API and measure time
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-performancetests/HOST_NEW/Tensor_host_pln1.cpp

@@ -586,6 +586,9 @@ int main(int argc, char **argv)
     // Run case-wise RPP API and measure time
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-performancetests/HOST_NEW/Tensor_host_pln3.cpp

@@ -672,6 +672,9 @@ int main(int argc, char **argv)
     // Run case-wise RPP API and measure time
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-performancetests/OCL_NEW/Single_host.cpp

@@ -176,6 +176,9 @@ int main(int argc, char **argv)
 
     closedir(dr2);
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-unittests/HIP_NEW/Single_host.cpp

@@ -176,6 +176,9 @@ int main(int argc, char **argv)
 
     closedir(dr2);
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-unittests/HOST_NEW/BatchPD_host_pkd3.cpp

@@ -647,6 +647,9 @@ int main(int argc, char **argv)
     }
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-unittests/HOST_NEW/BatchPD_host_pln1.cpp

@@ -648,6 +648,9 @@ int main(int argc, char **argv)
     }
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-unittests/HOST_NEW/BatchPD_host_pln3.cpp

@@ -751,6 +751,9 @@ int main(int argc, char **argv)
     }
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-unittests/HOST_NEW/Single_host.cpp

@@ -176,6 +176,9 @@ int main(int argc, char **argv)
 
     closedir(dr2);
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/rpp-unittests/HOST_NEW/Tensor_host_pkd3.cpp

@@ -609,6 +609,9 @@ int main(int argc, char **argv)
     // Run case-wise RPP API and measure time
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-unittests/HOST_NEW/Tensor_host_pln1.cpp

@@ -599,6 +599,9 @@ int main(int argc, char **argv)
     // Run case-wise RPP API and measure time
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-unittests/HOST_NEW/Tensor_host_pln3.cpp

@@ -685,6 +685,9 @@ int main(int argc, char **argv)
     // Run case-wise RPP API and measure time
 
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
     clock_t start, end;

utilities/rpp-unittests/OCL_NEW/Single_host.cpp

@@ -176,6 +176,9 @@ int main(int argc, char **argv)
 
     closedir(dr2);
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);
 

utilities/test_suite/HOST/Tensor_host.cpp

@@ -380,6 +380,9 @@ int main(int argc, char **argv)
 
     // Run case-wise RPP API and measure time
     rppHandle_t handle;
+
+    // Set the number of threads to be used by OpenMP pragma for RPP batch processing on host.
+    // If numThreads value passed is 0, number of OpenMP threads used by RPP will be set to batch size
     Rpp32u numThreads = 0;
     rppCreateWithBatchSize(&handle, noOfImages, numThreads);