NV workaround. Using the command queue to get the device id

AMD-FirePro · AMD-FirePro · commit cdb291951cde · 2013-10-14T10:23:35.000+01:00
diff --git a/src/library/generator.copy.cpp b/src/library/generator.copy.cpp
@@ -445,7 +445,7 @@ clfftStatus FFTPlan::GetMax1DLengthPvt<Copy> (size_t * longest) const
 using namespace CopyGenerator;
 
 template<>
-clfftStatus FFTPlan::GenerateKernelPvt<Copy>(FFTRepo& fftRepo ) const
+clfftStatus FFTPlan::GenerateKernelPvt<Copy>(FFTRepo& fftRepo, const cl_command_queue commQueueFFT ) const
 {
     FFTKernelGenKeyParams params;
     OPENCL_V( this->GetKernelGenKeyPvt<Copy> (params), _T("GetKernelGenKey() failed!") );
diff --git a/src/library/generator.stockham.cpp b/src/library/generator.stockham.cpp
@@ -3229,24 +3229,30 @@ clfftStatus FFTPlan::GetMax1DLengthPvt<Stockham> (size_t * longest) const
 }
 
 template<>
-clfftStatus FFTPlan::GenerateKernelPvt<Stockham>(FFTRepo& fftRepo ) const
+clfftStatus FFTPlan::GenerateKernelPvt<Stockham>(FFTRepo& fftRepo, const cl_command_queue commQueueFFT ) const
 {
     FFTKernelGenKeyParams params;
     OPENCL_V( this->GetKernelGenKeyPvt<Stockham> (params), _T("GetKernelGenKey() failed!") );
 
+    cl_int status = CL_SUCCESS;
+    cl_device_id Device = NULL;
+    status = clGetCommandQueueInfo(commQueueFFT, CL_QUEUE_DEVICE, sizeof(cl_device_id), &Device, NULL);
+
+    OPENCL_V( status, _T( "clGetCommandQueueInfo failed" ) );
+
 	std::string programCode;
 	Precision pr = (params.fft_precision == CLFFT_SINGLE) ? P_SINGLE : P_DOUBLE;
 	switch(pr)
 	{
 	case P_SINGLE:
 		{
 			Kernel<P_SINGLE> kernel(params);
-			kernel.GenerateKernel(programCode, devices[0]);
+			kernel.GenerateKernel(programCode, Device);
 		} break;
 	case P_DOUBLE:
 		{
 			Kernel<P_DOUBLE> kernel(params);
-			kernel.GenerateKernel(programCode, devices[0]);
+			kernel.GenerateKernel(programCode, Device);
 		} break;
 	}
 
diff --git a/src/library/generator.transpose.cpp b/src/library/generator.transpose.cpp
@@ -822,7 +822,7 @@ clfftStatus FFTPlan::GetWorkSizesPvt<Transpose> (std::vector<size_t> & globalWS,
 //	OpenCL does not take unicode strings as input, so this routine returns only ASCII strings
 //	Feed this generator the FFTPlan, and it returns the generated program as a string
 template<>
-clfftStatus FFTPlan::GenerateKernelPvt<Transpose> ( FFTRepo& fftRepo ) const
+clfftStatus FFTPlan::GenerateKernelPvt<Transpose> ( FFTRepo& fftRepo, const cl_command_queue commQueueFFT ) const
 {
 	FFTKernelGenKeyParams params;
 	OPENCL_V( this->GetKernelGenKeyPvt<Transpose> (params), _T("GetKernelGenKey() failed!") );
diff --git a/src/library/plan.cpp b/src/library/plan.cpp
@@ -543,7 +543,7 @@ clfftStatus	clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma
 
 	if(fftPlan->gen == Copy)
 	{
-		OPENCL_V( fftPlan->GenerateKernel( fftRepo ), _T( "GenerateKernel() failed" ) );
+		OPENCL_V( fftPlan->GenerateKernel( fftRepo, *commQueueFFT ), _T( "GenerateKernel() failed" ) );
 		OPENCL_V( CompileKernels( *commQueueFFT, plHandle, fftPlan->gen, fftPlan ), _T( "CompileKernels() failed" ) );
 		fftPlan->baked		= true;
 		return	CLFFT_SUCCESS;
@@ -1505,7 +1505,7 @@ clfftStatus	clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma
 				//break;
 				if (fftPlan->transflag) //Transpose for 2D
 				{
-					OPENCL_V( fftPlan->GenerateKernel( fftRepo ), _T( "GenerateTransposeProgram() failed" ) );
+					OPENCL_V( fftPlan->GenerateKernel( fftRepo, *commQueueFFT ), _T( "GenerateTransposeProgram() failed" ) );
 					OPENCL_V( CompileKernels( *commQueueFFT, plHandle, fftPlan->gen, fftPlan ), _T( "CompileKernels() failed" ) );
 
 					fftPlan->baked		= true;
@@ -2445,7 +2445,7 @@ clfftStatus	clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma
 	}
 
 	//	For the radices that we have factored, we need to load/compile and build the appropriate OpenCL kernels
-	OPENCL_V( fftPlan->GenerateKernel( fftRepo ), _T( "GenerateKernel() failed" ) );
+	OPENCL_V( fftPlan->GenerateKernel( fftRepo, *commQueueFFT ), _T( "GenerateKernel() failed" ) );
 
 	//	For the radices that we have factored, we need to load/compile and build the appropriate OpenCL kernels
 	OPENCL_V( CompileKernels( *commQueueFFT, plHandle, fftPlan->gen, fftPlan ), _T( "CompileKernels() failed" ) );
@@ -3265,13 +3265,13 @@ clfftStatus  FFTPlan::GetKernelGenKey (FFTKernelGenKeyParams & params) const
 	}
 }
 
-clfftStatus  FFTPlan::GenerateKernel (FFTRepo & fftRepo) const
+clfftStatus  FFTPlan::GenerateKernel (FFTRepo & fftRepo, const cl_command_queue commQueueFFT) const
 {
 	switch(gen)
 	{
-	case Stockham:		return GenerateKernelPvt<Stockham>(fftRepo);
-	case Transpose:		return GenerateKernelPvt<Transpose>(fftRepo);
-	case Copy:			return GenerateKernelPvt<Copy>(fftRepo);
+	case Stockham:		return GenerateKernelPvt<Stockham>(fftRepo, commQueueFFT);
+	case Transpose:		return GenerateKernelPvt<Transpose>(fftRepo, commQueueFFT);
+	case Copy:			return GenerateKernelPvt<Copy>(fftRepo, commQueueFFT);
 	default:			assert(false); return CLFFT_NOTIMPLEMENTED;
 	}
 }
diff --git a/src/library/plan.h b/src/library/plan.h
@@ -202,7 +202,7 @@ class	FFTPlan
 	clfftStatus GetKernelGenKeyPvt (FFTKernelGenKeyParams & params) const;
 
 	template <clfftGenerators G>
-	clfftStatus GenerateKernelPvt (FFTRepo& fftRepo) const;
+	clfftStatus GenerateKernelPvt (FFTRepo& fftRepo,  const cl_command_queue commQueueFFT ) const;
 
 	template <clfftGenerators G>
 	clfftStatus GetMax1DLengthPvt (size_t *longest ) const;
@@ -338,7 +338,7 @@ class	FFTPlan
 
 	clfftStatus GetWorkSizes (std::vector<size_t> & globalws, std::vector<size_t> & localws) const;
 	clfftStatus GetKernelGenKey (FFTKernelGenKeyParams & params) const;
-	clfftStatus GenerateKernel (FFTRepo & fftRepo) const;
+	clfftStatus GenerateKernel (FFTRepo & fftRepo, const cl_command_queue commQueueFFT) const;
 	clfftStatus GetMax1DLength (size_t *longest ) const;
 
 	void ResetBinarySizes();

Original file line number	Diff line number	Diff line change
`@@ -445,7 +445,7 @@ clfftStatus FFTPlan::GetMax1DLengthPvt<Copy> (size_t * longest) const`
`445`	`445`	`using namespace CopyGenerator;`
`446`	`446`
`447`	`447`	`template<>`
`448`		`-clfftStatus FFTPlan::GenerateKernelPvt<Copy>(FFTRepo& fftRepo ) const`
	`448`	`+clfftStatus FFTPlan::GenerateKernelPvt<Copy>(FFTRepo& fftRepo, const cl_command_queue commQueueFFT ) const`
`449`	`449`	`{`
`450`	`450`	`FFTKernelGenKeyParams params;`
`451`	`451`	`OPENCL_V( this->GetKernelGenKeyPvt<Copy> (params), _T("GetKernelGenKey() failed!") );`
Original file line number	Diff line number	Diff line change
`@@ -3229,24 +3229,30 @@ clfftStatus FFTPlan::GetMax1DLengthPvt<Stockham> (size_t * longest) const`
`3229`	`3229`	`}`
`3230`	`3230`
`3231`	`3231`	`template<>`
`3232`		`-clfftStatus FFTPlan::GenerateKernelPvt<Stockham>(FFTRepo& fftRepo ) const`
	`3232`	`+clfftStatus FFTPlan::GenerateKernelPvt<Stockham>(FFTRepo& fftRepo, const cl_command_queue commQueueFFT ) const`
`3233`	`3233`	`{`
`3234`	`3234`	`FFTKernelGenKeyParams params;`
`3235`	`3235`	`OPENCL_V( this->GetKernelGenKeyPvt<Stockham> (params), _T("GetKernelGenKey() failed!") );`
`3236`	`3236`
	`3237`	`+ cl_int status = CL_SUCCESS;`
	`3238`	`+ cl_device_id Device = NULL;`
	`3239`	`+ status = clGetCommandQueueInfo(commQueueFFT, CL_QUEUE_DEVICE, sizeof(cl_device_id), &Device, NULL);`
	`3240`	`+`
	`3241`	`+ OPENCL_V( status, _T( "clGetCommandQueueInfo failed" ) );`
	`3242`	`+`
`3237`	`3243`	`std::string programCode;`
`3238`	`3244`	`Precision pr = (params.fft_precision == CLFFT_SINGLE) ? P_SINGLE : P_DOUBLE;`
`3239`	`3245`	`switch(pr)`
`3240`	`3246`	`{`
`3241`	`3247`	`case P_SINGLE:`
`3242`	`3248`	`{`
`3243`	`3249`	`Kernel<P_SINGLE> kernel(params);`
`3244`		`- kernel.GenerateKernel(programCode, devices[0]);`
	`3250`	`+ kernel.GenerateKernel(programCode, Device);`
`3245`	`3251`	`} break;`
`3246`	`3252`	`case P_DOUBLE:`
`3247`	`3253`	`{`
`3248`	`3254`	`Kernel<P_DOUBLE> kernel(params);`
`3249`		`- kernel.GenerateKernel(programCode, devices[0]);`
	`3255`	`+ kernel.GenerateKernel(programCode, Device);`
`3250`	`3256`	`} break;`
`3251`	`3257`	`}`
`3252`	`3258`
Original file line number	Diff line number	Diff line change
`@@ -822,7 +822,7 @@ clfftStatus FFTPlan::GetWorkSizesPvt<Transpose> (std::vector<size_t> & globalWS,`
`822`	`822`	`// OpenCL does not take unicode strings as input, so this routine returns only ASCII strings`
`823`	`823`	`// Feed this generator the FFTPlan, and it returns the generated program as a string`
`824`	`824`	`template<>`
`825`		`-clfftStatus FFTPlan::GenerateKernelPvt<Transpose> ( FFTRepo& fftRepo ) const`
	`825`	`+clfftStatus FFTPlan::GenerateKernelPvt<Transpose> ( FFTRepo& fftRepo, const cl_command_queue commQueueFFT ) const`
`826`	`826`	`{`
`827`	`827`	`FFTKernelGenKeyParams params;`
`828`	`828`	`OPENCL_V( this->GetKernelGenKeyPvt<Transpose> (params), _T("GetKernelGenKey() failed!") );`
Original file line number	Diff line number	Diff line change
`@@ -543,7 +543,7 @@ clfftStatus clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma`
`543`	`543`
`544`	`544`	`if(fftPlan->gen == Copy)`
`545`	`545`	`{`
`546`		`- OPENCL_V( fftPlan->GenerateKernel( fftRepo ), _T( "GenerateKernel() failed" ) );`
	`546`	`+ OPENCL_V( fftPlan->GenerateKernel( fftRepo, *commQueueFFT ), _T( "GenerateKernel() failed" ) );`
`547`	`547`	`OPENCL_V( CompileKernels( *commQueueFFT, plHandle, fftPlan->gen, fftPlan ), _T( "CompileKernels() failed" ) );`
`548`	`548`	`fftPlan->baked = true;`
`549`	`549`	`return CLFFT_SUCCESS;`
`@@ -1505,7 +1505,7 @@ clfftStatus clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma`
`1505`	`1505`	`//break;`
`1506`	`1506`	`if (fftPlan->transflag) //Transpose for 2D`
`1507`	`1507`	`{`
`1508`		`- OPENCL_V( fftPlan->GenerateKernel( fftRepo ), _T( "GenerateTransposeProgram() failed" ) );`
	`1508`	`+ OPENCL_V( fftPlan->GenerateKernel( fftRepo, *commQueueFFT ), _T( "GenerateTransposeProgram() failed" ) );`
`1509`	`1509`	`OPENCL_V( CompileKernels( *commQueueFFT, plHandle, fftPlan->gen, fftPlan ), _T( "CompileKernels() failed" ) );`
`1510`	`1510`
`1511`	`1511`	`fftPlan->baked = true;`
`@@ -2445,7 +2445,7 @@ clfftStatus clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma`
`2445`	`2445`	`}`
`2446`	`2446`
`2447`	`2447`	`// For the radices that we have factored, we need to load/compile and build the appropriate OpenCL kernels`
`2448`		`- OPENCL_V( fftPlan->GenerateKernel( fftRepo ), _T( "GenerateKernel() failed" ) );`
	`2448`	`+ OPENCL_V( fftPlan->GenerateKernel( fftRepo, *commQueueFFT ), _T( "GenerateKernel() failed" ) );`
`2449`	`2449`
`2450`	`2450`	`// For the radices that we have factored, we need to load/compile and build the appropriate OpenCL kernels`
`2451`	`2451`	`OPENCL_V( CompileKernels( *commQueueFFT, plHandle, fftPlan->gen, fftPlan ), _T( "CompileKernels() failed" ) );`
`@@ -3265,13 +3265,13 @@ clfftStatus FFTPlan::GetKernelGenKey (FFTKernelGenKeyParams & params) const`
`3265`	`3265`	`}`
`3266`	`3266`	`}`
`3267`	`3267`
`3268`		`-clfftStatus FFTPlan::GenerateKernel (FFTRepo & fftRepo) const`
	`3268`	`+clfftStatus FFTPlan::GenerateKernel (FFTRepo & fftRepo, const cl_command_queue commQueueFFT) const`
`3269`	`3269`	`{`
`3270`	`3270`	`switch(gen)`
`3271`	`3271`	`{`
`3272`		`- case Stockham: return GenerateKernelPvt<Stockham>(fftRepo);`
`3273`		`- case Transpose: return GenerateKernelPvt<Transpose>(fftRepo);`
`3274`		`- case Copy: return GenerateKernelPvt<Copy>(fftRepo);`
	`3272`	`+ case Stockham: return GenerateKernelPvt<Stockham>(fftRepo, commQueueFFT);`
	`3273`	`+ case Transpose: return GenerateKernelPvt<Transpose>(fftRepo, commQueueFFT);`
	`3274`	`+ case Copy: return GenerateKernelPvt<Copy>(fftRepo, commQueueFFT);`
`3275`	`3275`	`default: assert(false); return CLFFT_NOTIMPLEMENTED;`
`3276`	`3276`	`}`
`3277`	`3277`	`}`