Feature/oro 0 amdadvtech merge (#43)

* Add gitignore to the repository Signed-off-by: Chih-Chen Kao <[email protected]> * Fix missing CUDA properties. (#16) Signed-off-by: Chih-Chen Kao <[email protected]> * Feature/oro 0 radix sort (#19) * [ORO-0] Working 8 bit radix sort. * [ORO-0] Some optimization. * Create LICENSE * Update README.md (#15) * Feature/oro 0 raw get set (#19) * [ORO-0] Rename setter and getter. * [ORO-0] Fix when there is a dll but no device. * [ORO-0] Deletion function. * [ORO-0] Multi processor count. * [ORO-0] Extended the sort to more than 8 bits. Implemented tests. * [ORO-0] Moved temp buffer allocation out from the sort(). * [ORO-0] README. References. * [ORO-0] Debug flag. * Refactor the code to add the basic constructs to support selecting different scan algorithms. Add different implementation of the scan algorithm: CPU, single WG and all WG . Signed-off-by: Chih-Chen Kao <[email protected]> * Squashed commit of the following: commit 3f32bea2244653d59efb3c3eaa9433018dde5835 Author: takahiroharada <[email protected]> Date: Wed Apr 13 10:48:35 2022 -0700 [ORO-0] Fix nvrtc. * Optimization: Implement the single-pass kernel for GPU parallel scan. Fix a GPU memory bug. Signed-off-by: Chih-Chen Kao <[email protected]> * Feature/oro 0 kernel cache (#4) * [ORO-0] Cache kernel. * [ORO-0] Support newer HIP builds on windows (#22) * [ORO-0] Unit test. (#23) * Fix LDS scan bug. The previous implementation would lead to an error when the wavefront (wrap) size is not equal to the size of a workgroup (block). Since not all threads run simultaneously, for an input arrays larger than the wavefront size, the previous algorithm will not work because it performs the scan in-place on the input array. The results of one wavefront (wrap) will be overwritten by work items (threads) in another wavefront (wrap). Signed-off-by: Chih-Chen Kao <[email protected]> * Optimize the LDS scan algorithm. (#6) * Optimize the LDS scan algorithm. This version does not require a temp buffer and can support a LDS input size up to 2 times the workgroup size. Signed-off-by: Chih-Chen Kao <[email protected]> * Support an input array in LDS that is 2 times the WG size. Signed-off-by: Chih-Chen Kao <[email protected]> * Feature/oro 0 clean up (#7) * Squashed commit of the following: commit 3f32bea2244653d59efb3c3eaa9433018dde5835 Author: takahiroharada <[email protected]> Date: Wed Apr 13 10:48:35 2022 -0700 [ORO-0] Fix nvrtc. * [ORO-0] Clean up. * Feature/oro 0 clean up (#10) * Squashed commit of the following: commit 3f32bea2244653d59efb3c3eaa9433018dde5835 Author: takahiroharada <[email protected]> Date: Wed Apr 13 10:48:35 2022 -0700 [ORO-0] Fix nvrtc. * [ORO-0] Clean up. * [ORO-0] SortKernel1. Less complex. (#8) SortKernel (occupancy: 8) - vgpr: 128 - lds: 6704 SortKernel1 (occupancy: 9) - vgpr: 106 - lds 7720 * [ORO-0] Kernel execution time check. * Fix the memory access pattern and change it to coalesced memory access. (#11) Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Single kernel sort for small keys. (#12) * Optimize the Count kernel for less LDS usage to achieve full occupancy (#13) * Optimize the Count kernel to let it use less LDS and could achieve full occupancy. Signed-off-by: Chih-Chen Kao <[email protected]> * Remove __threadfence_block() Removes the boundary check in the inner loop. The upper bound is set only once before going into the loop. Signed-off-by: Chih-Chen Kao <[email protected]> * Introduce DRIVER and RTC APIs * Disable enum-variant * Improve paths * Add fields * Update Vulkan test * Define CUDA in terms of DRIVER and RTC * Optimize the sort kernel: single-pass 8bit sort & parallel scan in 4bit sort. (#14) * Fix a minor issue in CountKernel to make it more robust. Implement a single-pass 8-bit local sort. Implement a single-pass 8-bit local sort with shared bins. Signed-off-by: Chih-Chen Kao <[email protected]> * Fix nItemsPerWI and enable the version with shared LDS. Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Print driver version. * [ORO-0] Repro case. * Fix SORT_WG_SIZE. Fix stable sort order. Signed-off-by: Chih-Chen Kao <[email protected]> * Optimize sort kernel to remove inner boundary check. Adjust nItemsPerWI. Signed-off-by: Chih-Chen Kao <[email protected]> Co-authored-by: takahiroharada <[email protected]> * Merging another merge (#18) * Fix a minor issue in CountKernel to make it more robust. Implement a single-pass 8-bit local sort. Implement a single-pass 8-bit local sort with shared bins. Signed-off-by: Chih-Chen Kao <[email protected]> * Fix nItemsPerWI and enable the version with shared LDS. Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Print driver version. * [ORO-0] Repro case. * Fix SORT_WG_SIZE. Fix stable sort order. Signed-off-by: Chih-Chen Kao <[email protected]> * Optimize sort kernel to remove inner boundary check. Adjust nItemsPerWI. Signed-off-by: Chih-Chen Kao <[email protected]> * Calculate the number of WGs based on LDS and max-thread-per-WGP. (#15) * Calculate the number of WGs based on LDS and max-thread-per-WGP. Signed-off-by: Chih-Chen Kao <[email protected]> * Add a workaround for CUDA. Signed-off-by: Chih-Chen Kao <[email protected]> * Optimize the sort kernel: single-pass 8bit sort & parallel scan in 4bit sort. (#14) * Fix a minor issue in CountKernel to make it more robust. Implement a single-pass 8-bit local sort. Implement a single-pass 8-bit local sort with shared bins. Signed-off-by: Chih-Chen Kao <[email protected]> * Fix nItemsPerWI and enable the version with shared LDS. Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Print driver version. * [ORO-0] Repro case. * Fix SORT_WG_SIZE. Fix stable sort order. Signed-off-by: Chih-Chen Kao <[email protected]> * Optimize sort kernel to remove inner boundary check. Adjust nItemsPerWI. Signed-off-by: Chih-Chen Kao <[email protected]> Co-authored-by: takahiroharada <[email protected]> Co-authored-by: takahiroharada <[email protected]> Co-authored-by: Chih-Chen Kao <[email protected]> * Implement key-value pair sorting (#17) * Add gitignore to the repository Signed-off-by: Chih-Chen Kao <[email protected]> * Fix missing CUDA properties. (#16) Signed-off-by: Chih-Chen Kao <[email protected]> * Add basic structure for key-value pair sorting. Fix an error in single pass sort Signed-off-by: Chih-Chen Kao <[email protected]> * Add Value data in the test and sort it according to keys. Signed-off-by: Chih-Chen Kao <[email protected]> * Support Key only sorting. Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Make single pass kernel non compile time switch. * Support both Key-Only & Key-Value pair sort kernels Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Test change. * [ORO-0] A bug. * [ORO-0] NVIDIA occupancy computation fix. Test change. Tweak params to use single pass sort as much as possible. Co-authored-by: Takahiro Harada <[email protected]> Co-authored-by: takahiroharada <[email protected]> * [ORO-0] Revert demo code. * Fix missing CUDA properties. (#26) * Update Orochi.cpp * [ORO-0] Clean up. * [ORO-0] OroUtils. (#27) * [ORO-0] OroUtils. * [ORO-0] Linux build fix. * [ORO-0] Forgot to add. * [ORO-0] Linux build fix. * [ORO-0] Clean up. Co-authored-by: Chih-Chen Kao <[email protected]> Co-authored-by: Aaryaman Vasishta <[email protected]> Co-authored-by: Mehmet Oguz Derin <[email protected]> * Add kernel path and include dir to the functions. (#20) Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] BakeKernel. (#21) * [ORO-0] BakeKernel. * Update tools/genArgs.py commented code removal * Update tools/stringify.py commented code removal * Update tools/stringify.py commented code removal * Update tools/stringify.py commented code removal * Update tools/genArgs.py dead code removal * Update tools/stringify.py dead code removal * fix include Signed-off-by: Chih-Chen Kao <[email protected]> * fix script Signed-off-by: Chih-Chen Kao <[email protected]> * fix Signed-off-by: Chih-Chen Kao <[email protected]> Co-authored-by: Chih-Chen Kao <[email protected]> * Fix Orochi CUDA API (#23) Fix Orochi CUDA API Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Linux build fix. (#22) * [ORO-0] Linux build fix. * Fix Orochi CUDA API Signed-off-by: Chih-Chen Kao <[email protected]> Co-authored-by: Chih-Chen Kao <[email protected]> * Quick fix for old linux gcc which does not support std::exclusive_scan (#24) Quick fix for old linux gcc which does not support std::exclusive_scan Signed-off-by: Chih-Chen Kao <[email protected]> * Fix the kernel cache bug. (#25) Fix the kernel cache bug. The function should not return the oroFunctions that are created previously solely based on the names because they might be invalid. Signed-off-by: Chih-Chen Kao <[email protected]> * [ORO-0] Remove static variables. (#26) * [ORO-0] Remove static variables. * [ORO-0] Applied the suggestions. * [ORO-0] Linux regression fix. * Fix OrochiUtils::getFunctionFromString API (#27) Signed-off-by: Chih-Chen Kao <[email protected]> * Adding missing assert (#28) * Adding missing assert * Adding more asserts * Feature/oro 0 gpuopen merge (#31) * Fix oroGetDeviceProperties in cuda path. * Fix linux crash (#29) * [ORO-0] Added missing file. * [ORO-0] Remove printf from kernelExec and skip compilation of vulkan test on Linux (#31) * [ORO-0] Skip compilation of vulkan test on Linux * [ORO-0] Update kernelExec unit test - remove printf * [ORO-0] Remove cout * [ORO-0] Fix hipGetErrorString (#32) * [ORO-0] Fix hipGetErrorString It was incorrectly importing this API. Import the correct API in hipew. * [ORO-0] Remove printf from kernelExec and skip compilation of vulkan test on Linux (#31) * [ORO-0] Skip compilation of vulkan test on Linux * [ORO-0] Update kernelExec unit test - remove printf * [ORO-0] Remove cout * [ORO-0] Add Orochi error codes mapped to HIP/CUDA (#33) * Add missing path on Apple config. (#34) * [ORO-0] Adding hiprtc+comgr dlls to workaround the regression in 22.7.1 driver (#38) * [ORO-0] Adding hiprtc to workaround the regression in 22.7.1 driver released at 7/26/2022. * [ORO-0] Created win64 subdir. * [ORO-0] Add hiprtc.dll and comgr dll Co-authored-by: takahiroharada <[email protected]> * fix footnote markdown format (#39) * Fix orochi utils issue in unit tests Co-authored-by: Aaryaman Vasishta <[email protected]> Co-authored-by: Chih-Chen Kao <[email protected]> Co-authored-by: NevesLucas <[email protected]> Co-authored-by: PixelClear <[email protected]> Signed-off-by: Chih-Chen Kao <[email protected]> Co-authored-by: Chih-Chen Kao <[email protected]> Co-authored-by: Aaryaman Vasishta <[email protected]> Co-authored-by: Mehmet Oguz Derin <[email protected]> Co-authored-by: Daniel Meister <[email protected]> Co-authored-by: NevesLucas <[email protected]> Co-authored-by: PixelClear <[email protected]>
GPUOpen-LibrariesAndSDKs · Aug 18, 2022 · d78fb81 · d78fb81
1 parent 03c4676
commit d78fb81
Show file tree

Hide file tree

Showing 8 changed files with 357 additions and 18 deletions.
diff --git a/Orochi/OrochiUtils.cpp b/Orochi/OrochiUtils.cpp
@@ -155,7 +155,7 @@ struct OrochiUtilsImpl
 		return false;
 	}
 
-	static void getCacheFileName( oroDevice device, const char* moduleName, const char* functionName, const char* options, std::string& binFileName )
+	static void getCacheFileName( oroDevice device, const char* moduleName, const char* functionName, const char* options, std::string& binFileName, const std::string& cacheDirectory )
 	{
 		auto hashBin = []( const char* s, const size_t size )
 		{
@@ -220,7 +220,7 @@ struct OrochiUtilsImpl
 		using namespace std::string_literals;
 
 		deviceName = deviceName.substr( 0, deviceName.find( ":" ) );
-		binFileName = OrochiUtils::s_cacheDirectory + "/"s + moduleHash + "-"s + optionHash + ".v."s + deviceName + "."s + driverVersion + "_"s + std::to_string( 8 * sizeof( void* ) ) + ".bin"s;
+		binFileName = cacheDirectory + "/"s + moduleHash + "-"s + optionHash + ".v."s + deviceName + "."s + driverVersion + "_"s + std::to_string( 8 * sizeof( void* ) ) + ".bin"s;
 	}
 	static
 	bool isFileUpToDate( const char* binaryFileName, const char* srcFileName )
@@ -381,27 +381,47 @@ struct OrochiUtilsImpl
 	}
 };
 
-char* OrochiUtils::s_cacheDirectory = "./cache/";
-std::map<std::string, oroFunction> OrochiUtils::s_kernelMap;
+OrochiUtils::OrochiUtils() 
+{
+	m_cacheDirectory = "./cache/";
+}
+
+OrochiUtils::~OrochiUtils() 
+{
+}
 
 oroFunction OrochiUtils::getFunctionFromFile( oroDevice device, const char* path, const char* funcName, std::vector<const char*>* optsIn )
 { 
-	std::string cacheName = OrochiUtilsImpl::getCacheName( path, funcName );
-	if( s_kernelMap.find( cacheName.c_str() ) != s_kernelMap.end() )
+	const std::string cacheName = OrochiUtilsImpl::getCacheName( path, funcName );
+	if( m_kernelMap.find( cacheName.c_str() ) != m_kernelMap.end() )
 	{
-		return s_kernelMap[ cacheName ];
+		return m_kernelMap[ cacheName ];
 	}
 
 	std::string source;
 	if( !OrochiUtilsImpl::readSourceCode( path, source, 0 ) )
 		return 0;
 
 	oroFunction f = getFunction( device, source.c_str(), path, funcName, optsIn );
-	s_kernelMap[cacheName] = f;
+	m_kernelMap[cacheName] = f;
 	return f;
 }
 
-oroFunction OrochiUtils::getFunction( oroDevice device, const char* code, const char* path, const char* funcName, std::vector<const char*>* optsIn )
+oroFunction OrochiUtils::getFunctionFromString( oroDevice device, const char* source, const char* path, const char* funcName, std::vector<const char*>* optsIn, 
+	int numHeaders, const char** headers, const char** includeNames ) 
+{
+	const std::string cacheName = OrochiUtilsImpl::getCacheName( path, funcName );
+	if( m_kernelMap.find( cacheName.c_str() ) != m_kernelMap.end() )
+	{
+		return m_kernelMap[cacheName];
+	}
+	oroFunction f = getFunction( device, source, path, funcName, optsIn, numHeaders, headers, includeNames );
+	m_kernelMap[cacheName] = f;
+	return f;
+}
+
+oroFunction OrochiUtils::getFunction( oroDevice device, const char* code, const char* path, const char* funcName, std::vector<const char*>* optsIn, 
+	int numHeaders, const char** headers, const char** includeNames )
 {
 	std::vector<const char*> opts;
 	opts.push_back( "-std=c++17" );
@@ -422,7 +442,7 @@ oroFunction OrochiUtils::getFunction( oroDevice device, const char* code, const
 		std::string o;
 		for(int i=0; i<opts.size(); i++)
 			o.append( opts[i] );
-		OrochiUtilsImpl::getCacheFileName( device, path, funcName, o.c_str(), cacheFile );
+		OrochiUtilsImpl::getCacheFileName( device, path, funcName, o.c_str(), cacheFile, m_cacheDirectory );
 	}
 	if( OrochiUtilsImpl::isFileUpToDate( cacheFile.c_str(), path ) )
 	{
@@ -433,7 +453,8 @@ oroFunction OrochiUtils::getFunction( oroDevice device, const char* code, const
 	{
 		orortcProgram prog;
 		orortcResult e;
-		e = orortcCreateProgram( &prog, code, path, 0, 0, 0 );
+		e = orortcCreateProgram( &prog, code, path, numHeaders, headers, includeNames );
+		OROASSERT( e == ORORTC_SUCCESS, 0 );
 
 		e = orortcCompileProgram( prog, opts.size(), opts.data() );
 		if( e != ORORTC_SUCCESS )
@@ -449,18 +470,23 @@ oroFunction OrochiUtils::getFunction( oroDevice device, const char* code, const
 		}
 		size_t codeSize;
 		e = orortcGetCodeSize( prog, &codeSize );
+		OROASSERT( e == ORORTC_SUCCESS, 0 );
 
 		codec.resize( codeSize );
 		e = orortcGetCode( prog, codec.data() );
+		OROASSERT( e == ORORTC_SUCCESS, 0 );
 		e = orortcDestroyProgram( &prog );
+		OROASSERT( e == ORORTC_SUCCESS, 0 );
 
 		//store cache
-		OrochiUtilsImpl::createDirectory( s_cacheDirectory );
+		OrochiUtilsImpl::createDirectory( m_cacheDirectory.c_str() );
 		OrochiUtilsImpl::cacheBinaryToFile( codec, cacheFile );
 	}
 	oroModule module;
 	oroError ee = oroModuleLoadData( &module, codec.data() );
+	OROASSERT( ee == oroSuccess, 0 );
 	ee = oroModuleGetFunction( &function, module, funcName );
+	OROASSERT( ee == oroSuccess, 0 );
 
 	return function;
 }

diff --git a/Orochi/OrochiUtils.h b/Orochi/OrochiUtils.h
@@ -1,7 +1,7 @@
 #pragma once
 #include <Orochi/Orochi.h>
 #include <vector>
-#include <map>
+#include <unordered_map>
 #include <string>
 
 #if defined(_WIN32)
@@ -18,8 +18,14 @@ class OrochiUtils
 		int x, y, z, w;
 	};
 
-	static oroFunction getFunctionFromFile( oroDevice device, const char* path, const char* funcName, std::vector<const char*>* opts );
-	static oroFunction getFunction( oroDevice device, const char* code, const char* path, const char* funcName, std::vector<const char*>* opts );
+	OrochiUtils();
+	~OrochiUtils();
+
+	oroFunction getFunctionFromFile( oroDevice device, const char* path, const char* funcName, std::vector<const char*>* opts );
+	oroFunction getFunctionFromString( oroDevice device, const char* source, const char* path, const char* funcName, std::vector<const char*>* opts, 
+		int numHeaders, const char** headers, const char** includeNames );
+	oroFunction getFunction( oroDevice device, const char* code, const char* path, const char* funcName, std::vector<const char*>* opts, 
+		int numHeaders = 0, const char** headers = 0, const char** includeNames = 0 );
 
 	static void launch1D( oroFunction func, int nx, const void** args, int wgSize = 64, unsigned int sharedMemBytes = 0 );
 
@@ -64,6 +70,6 @@ class OrochiUtils
 	}
 
 public:
-	static char* s_cacheDirectory;
-	static std::map<std::string, oroFunction> s_kernelMap;
+	std::string m_cacheDirectory;
+	std::unordered_map<std::string, oroFunction> m_kernelMap;
 };
diff --git a/Test/Stopwatch.h b/Test/Stopwatch.h
@@ -0,0 +1,111 @@
+/*
+AMD copyrights (Copyright (c) 2011 Advanced Micro Devices, Inc. All rights reserved) 
+*/
+#pragma once
+
+#if defined(__WINDOWS__)
+	#define NOMINMAX
+	#include <windows.h>
+
+	#define TIME_TYPE LARGE_INTEGER
+	#define QUERY_FREQ(f) QueryPerformanceFrequency(&f)
+	#define RECORD(t) QueryPerformanceCounter(&t)
+	#define GET_TIME(t) (t).QuadPart*1000.0
+	#define GET_FREQ(f)	(f).QuadPart
+#else
+	#include <sys/time.h> 
+
+	#define TIME_TYPE timeval
+	#define QUERY_FREQ(f) f.tv_sec = 1
+	#define RECORD(t) gettimeofday(&t, 0)
+	#define GET_TIME(t) ((t).tv_sec*1000.0+(t).tv_usec/1000.0)
+	#define GET_FREQ(f)	1.0
+#endif
+
+
+class Stopwatch
+{
+	public:
+		__inline
+		Stopwatch();
+		__inline
+		void init();
+		__inline
+		void start();
+		__inline
+		void split();
+		__inline
+		float getCurrent();
+		__inline
+		void stop();
+		__inline
+		float getMs();
+		__inline
+		void getMs( float* times, int capacity );
+
+	private:
+		enum 
+		{
+			CAPACITY = 12,
+		};
+		int m_idx;
+
+		TIME_TYPE m_frequency;
+		TIME_TYPE m_t[CAPACITY];
+};
+
+__inline
+Stopwatch::Stopwatch()
+{
+//	QueryPerformanceFrequency( &m_frequency );
+	QUERY_FREQ( m_frequency );
+}
+
+__inline
+void Stopwatch::start()
+{
+	m_idx = 0;
+//	QueryPerformanceCounter(&m_t[m_idx++]);
+	RECORD( m_t[m_idx++] );
+}
+
+__inline
+void Stopwatch::split()
+{
+//	QueryPerformanceCounter(&m_t[m_idx++]);
+	RECORD( m_t[m_idx++] );
+}
+
+__inline
+float Stopwatch::getCurrent()
+{
+	TIME_TYPE t;
+	RECORD( t );
+	return (float)( GET_TIME(t) - GET_TIME(m_t[0]) )/GET_FREQ(m_frequency);
+}
+
+__inline
+void Stopwatch::stop()
+{
+	split();
+}
+
+__inline
+float Stopwatch::getMs()
+{
+//	return (float)(1000*(m_t[1].QuadPart - m_t[0].QuadPart))/m_frequency.QuadPart;
+	return (float)( GET_TIME( m_t[1] ) - GET_TIME( m_t[0] ) )/GET_FREQ( m_frequency );
+}
+
+__inline
+void Stopwatch::getMs(float* times, int capacity)
+{
+	for(int i=0; i<capacity; i++) times[i] = 0.f;
+
+	for(int i=0; i<std::min(capacity, m_idx-1); i++)
+	{
+//		times[i] = (float)(1000*(m_t[i+1].QuadPart - m_t[i].QuadPart))/m_frequency.QuadPart;
+		times[i] = (float)( GET_TIME( m_t[i+1] ) - GET_TIME( m_t[i] ) )/GET_FREQ( m_frequency );
+	}
+}
+
diff --git a/UnitTest/main.cpp b/UnitTest/main.cpp
@@ -2,6 +2,9 @@
 #include <Orochi/Orochi.h>
 #include <Orochi/OrochiUtils.h>
 
+#if defined( OROASSERT )
+	#undef OROASSERT
+#endif
 #define OROASSERT( x ) ASSERT_TRUE( x )
 #define OROCHECK( x ) { oroError e = x; OROASSERT( e == ORO_SUCCESS ); }
 
@@ -50,11 +53,12 @@ TEST_F( OroTestBase, deviceprops )
 
 TEST_F( OroTestBase, kernelExec ) 
 {
+	OrochiUtils o;
 	int a_host = -1;
 	int* a_device = nullptr;
 	OROCHECK( oroMalloc( (oroDeviceptr*)&a_device, sizeof( int ) ) );
 	OROCHECK( oroMemset( (oroDeviceptr)a_device, 0, sizeof( int ) ) );
-	oroFunction kernel = OrochiUtils::getFunctionFromFile( m_device, "../UnitTest/testKernel.h", "testKernel", 0 ); 
+	oroFunction kernel = o.getFunctionFromFile( m_device, "../UnitTest/testKernel.h", "testKernel", 0 ); 
 	const void* args[] = { &a_device };
 	OrochiUtils::launch1D( kernel, 64, args, 64 );
 	OrochiUtils::waitForCompletion();

diff --git a/tools/bakeKernel.bat b/tools/bakeKernel.bat
@@ -0,0 +1,6 @@
+echo // automatically generated, don't edit > ParallelPrimitives/cache/Kernels.h
+echo // automatically generated, don't edit > ParallelPrimitives/cache/KernelArgs.h
+python tools/stringify.py ./ParallelPrimitives/RadixSortKernels.h  >> ParallelPrimitives/cache/Kernels.h
+python tools/genArgs.py ./ParallelPrimitives/RadixSortKernels.h  >> ParallelPrimitives/cache/KernelArgs.h
+
+python tools/stringify.py ./ParallelPrimitives/RadixSortConfigs.h  >> ParallelPrimitives/cache/Kernels.h
diff --git a/tools/bakeKernel.sh b/tools/bakeKernel.sh
@@ -0,0 +1,7 @@
+# mkdir hiprt/cache/
+echo "// automatically generated, don't edit" > ParallelPrimitives/cache/Kernels.h
+echo "// automatically generated, don't edit" > ParallelPrimitives/cache/KernelArgs.h
+python tools/stringify.py ./ParallelPrimitives/RadixSortKernels.h  >> ParallelPrimitives/cache/Kernels.h
+python tools/genArgs.py ./ParallelPrimitives/RadixSortKernels.h  >> ParallelPrimitives/cache/KernelArgs.h
+
+python tools/stringify.py ./ParallelPrimitives/RadixSortConfigs.h  >> ParallelPrimitives/cache/Kernels.h
diff --git a/tools/genArgs.py b/tools/genArgs.py
@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+from __future__ import print_function
+
+import sys
+import os
+
+def genArgs( fileName, api, includes ):
+    with open(fileName) as f:
+        iName = os.path.basename( fileName ).split('.')[0]
+
+        print( '#if !defined(ORO_PP_LOAD_FROM_STRING)' )
+        print( '	static const char** '+iName+'Args = 0;' )
+        print( '#else' )
+        print( '	static const char* '+iName+'Args[] = {' )
+        includes += iName +'Includes[] = {'
+        for line in f.readlines():
+            a = line.strip('\r\n')
+            if a.find('#include') == -1:
+                continue
+            if a.find('#include') != -1 and a.find('inl.' + api) != -1:
+                continue
+            if (api == 'cl' or api == 'metal') and a.find('.cu') != -1:
+                continue
+            if (a.find('"') != -1 and a.find('#include') != -1):
+                continue
+
+            filename = os.path.basename(a.split('<')[1].split('>')[0])
+            includes += '"' + a.split('<')[1].split('>')[0] + '",'
+            name = filename.split('.' + api)[0]
+            name = name.split('.h')[0]
+            name = api + '_'+name
+            print ( name + ',' )
+        print( api + '_'+iName+'};' )
+        print( '#endif' )
+        return includes
+
+argvs = sys.argv
+
+files = []
+if len(argvs) >= 2:
+    files.append( argvs[1] )
+
+print( '#pragma once' )
+
+
+api = 'hip'
+
+# Visit each file
+print( 'namespace ' + api + ' {')
+
+includes = 'static const char* '
+for s in files:
+    includes = genArgs(s, api, includes)
+includes += '};'
+print( includes )
+print( '}\t//namespace ' + api)