diff options
50 files changed, 442 insertions, 677 deletions
diff --git a/scripts/generator/generator/cpp.py b/scripts/generator/generator/cpp.py index 79d6b2a1..17e418e3 100644 --- a/scripts/generator/generator/cpp.py +++ b/scripts/generator/generator/cpp.py @@ -318,11 +318,9 @@ def wrapper_cublas(routine): result += " " + scalar + "_cuda.y = " + scalar + ".imag();" + NL # Calls the cuBLAS routine - result += " cublasHandle_t handle;" + NL - result += " if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; }" + NL result += " auto status = cublas" + flavour.name_cublas() + routine.name + "(handle, " result += ("," + NL + indent).join([a for a in arguments]) + ");" + NL - result += " cublasDestroy(handle);" + NL + result += " cudaDeviceSynchronize();" + NL result += " return status;" # There is no cuBLAS available, forward the call to one of the available functions @@ -335,11 +333,10 @@ def wrapper_cublas(routine): # result += " auto " + buf + "_buffer_bis = HalfToFloatBuffer(" + buf + "_buffer, queues[0]);" + NL # # Call the float routine - # result += " cublasHandle_t handle;" + NL - # result += " auto status = cublasX" + routine.name + "(handle," + # result += " return cublasX" + routine.name + "(handle," # result += ("," + NL + indent).join([a for a in routine.arguments_half()]) + ");" + NL - # result += " cublasDestroy(handle);" + NL - # result += " return status;" + NL + # result += " cudaDeviceSynchronize();" + NL + # result += " return status;" # # Convert back to half # for buf in routine.outputs: diff --git a/scripts/generator/generator/routine.py b/scripts/generator/generator/routine.py index a7abfde5..1c534611 100644 --- a/scripts/generator/generator/routine.py +++ b/scripts/generator/generator/routine.py @@ -884,6 +884,6 @@ class Routine: if def_only: result += flavour.name result += ">\n" - result += "cublasStatus_t cublasX" + self.name + template + "(" + result += "cublasStatus_t cublasX" + self.name + template + "(cublasHandle_t handle, " result += (",\n" + indent).join([a for a in self.arguments_def_wrapper_cublas(flavour)]) + ")" return result diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp index 7aadb983..b40ec541 100644 --- a/src/utilities/utilities.hpp +++ b/src/utilities/utilities.hpp @@ -193,6 +193,9 @@ struct Arguments { size_t step = 1; size_t num_steps = 0; size_t num_runs = 10; + #ifdef CLBLAST_REF_CUBLAS + void* cublas_handle; // cublasHandle_t + #endif // Common arguments size_t platform_id = 0; size_t device_id = 0; diff --git a/test/performance/client.cpp b/test/performance/client.cpp index a2f0f9f4..dc98ffbd 100644 --- a/test/performance/client.cpp +++ b/test/performance/client.cpp @@ -183,7 +183,7 @@ void Client<T,U>::PerformanceTest(Arguments<U> &args, const SetMetric set_sizes) if (args.compare_clblas) { clblasSetup(); } #endif #ifdef CLBLAST_REF_CUBLAS - cudaSetDevice(static_cast<int>(args.device_id)); + if (args.compare_cublas) { cublasSetup(args); } #endif // Iterates over all "num_step" values jumping by "step" each time @@ -272,6 +272,9 @@ void Client<T,U>::PerformanceTest(Arguments<U> &args, const SetMetric set_sizes) #ifdef CLBLAST_REF_CLBLAS if (args.compare_clblas) { clblasTeardown(); } #endif + #ifdef CLBLAST_REF_CUBLAS + if (args.compare_cublas) { cublasTeardown(args); } + #endif } // ================================================================================================= diff --git a/test/routines/level1/xamax.hpp b/test/routines/level1/xamax.hpp index dcd48a47..04bdaa3d 100644 --- a/test/routines/level1/xamax.hpp +++ b/test/routines/level1/xamax.hpp @@ -106,7 +106,7 @@ class TestXamax { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXamax(args.n, + auto status = cublasXamax(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.scalar, args.imax_offset, buffers.x_vec, args.x_offset, args.x_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } diff --git a/test/routines/level1/xasum.hpp b/test/routines/level1/xasum.hpp index e7e41fe5..6add9c64 100644 --- a/test/routines/level1/xasum.hpp +++ b/test/routines/level1/xasum.hpp @@ -106,7 +106,7 @@ class TestXasum { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXasum(args.n, + auto status = cublasXasum(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.scalar, args.asum_offset, buffers.x_vec, args.x_offset, args.x_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } diff --git a/test/routines/level1/xaxpy.hpp b/test/routines/level1/xaxpy.hpp index 98f0e380..17cae6ad 100644 --- a/test/routines/level1/xaxpy.hpp +++ b/test/routines/level1/xaxpy.hpp @@ -107,7 +107,7 @@ class TestXaxpy { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXaxpy(args.n, args.alpha, + auto status = cublasXaxpy(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } diff --git a/test/routines/level1/xcopy.hpp b/test/routines/level1/xcopy.hpp index 65e498ee..7a5c99b8 100644 --- a/test/routines/level1/xcopy.hpp +++ b/test/routines/level1/xcopy.hpp @@ -106,7 +106,7 @@ class TestXcopy { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXcopy(args.n, + auto status = cublasXcopy(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } diff --git a/test/routines/level1/xdot.hpp b/test/routines/level1/xdot.hpp index c95b16ef..1ea25994 100644 --- a/test/routines/level1/xdot.hpp +++ b/test/routines/level1/xdot.hpp @@ -113,7 +113,7 @@ class TestXdot { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXdot(args.n, + auto status = cublasXdot(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.scalar, args.dot_offset, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc); diff --git a/test/routines/level1/xdotc.hpp b/test/routines/level1/xdotc.hpp index 0c99be25..c800c1f5 100644 --- a/test/routines/level1/xdotc.hpp +++ b/test/routines/level1/xdotc.hpp @@ -113,7 +113,7 @@ class TestXdotc { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXdotc(args.n, + auto status = cublasXdotc(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.scalar, args.dot_offset, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc); diff --git a/test/routines/level1/xdotu.hpp b/test/routines/level1/xdotu.hpp index bf6bcd80..3545a3a6 100644 --- a/test/routines/level1/xdotu.hpp +++ b/test/routines/level1/xdotu.hpp @@ -113,7 +113,7 @@ class TestXdotu { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXdotu(args.n, + auto status = cublasXdotu(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.scalar, args.dot_offset, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc); diff --git a/test/routines/level1/xnrm2.hpp b/test/routines/level1/xnrm2.hpp index 096604d1..1db70537 100644 --- a/test/routines/level1/xnrm2.hpp +++ b/test/routines/level1/xnrm2.hpp @@ -106,7 +106,7 @@ class TestXnrm2 { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXnrm2(args.n, + auto status = cublasXnrm2(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.scalar, args.nrm2_offset, buffers.x_vec, args.x_offset, args.x_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } diff --git a/test/routines/level1/xscal.hpp b/test/routines/level1/xscal.hpp index 09b53839..efa0988d 100644 --- a/test/routines/level1/xscal.hpp +++ b/test/routines/level1/xscal.hpp @@ -100,7 +100,7 @@ class TestXscal { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXscal(args.n, args.alpha, + auto status = cublasXscal(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } } diff --git a/test/routines/level1/xswap.hpp b/test/routines/level1/xswap.hpp index 0d6fe451..d778cc23 100644 --- a/test/routines/level1/xswap.hpp +++ b/test/routines/level1/xswap.hpp @@ -106,7 +106,7 @@ class TestXswap { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXswap(args.n, + auto status = cublasXswap(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc); if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; } diff --git a/test/routines/level2/xgbmv.hpp b/test/routines/level2/xgbmv.hpp index 77abcfff..23138c77 100644 --- a/test/routines/level2/xgbmv.hpp +++ b/test/routines/level2/xgbmv.hpp @@ -126,7 +126,7 @@ class TestXgbmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXgbmv(args.layout, + auto status = cublasXgbmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.a_transpose), args.m, args.n, args.kl, args.ku, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, diff --git a/test/routines/level2/xgemv.hpp b/test/routines/level2/xgemv.hpp index c0c59152..0ee53b80 100644 --- a/test/routines/level2/xgemv.hpp +++ b/test/routines/level2/xgemv.hpp @@ -126,7 +126,7 @@ class TestXgemv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXgemv(args.layout, + auto status = cublasXgemv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.a_transpose), args.m, args.n, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, diff --git a/test/routines/level2/xger.hpp b/test/routines/level2/xger.hpp index 7fe37e1a..92a1a2ae 100644 --- a/test/routines/level2/xger.hpp +++ b/test/routines/level2/xger.hpp @@ -120,7 +120,7 @@ class TestXger { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXger(args.layout, + auto status = cublasXger(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, args.m, args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc, diff --git a/test/routines/level2/xgerc.hpp b/test/routines/level2/xgerc.hpp index b50cf672..5d899398 100644 --- a/test/routines/level2/xgerc.hpp +++ b/test/routines/level2/xgerc.hpp @@ -120,7 +120,7 @@ class TestXgerc { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXgerc(args.layout, + auto status = cublasXgerc(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, args.m, args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc, diff --git a/test/routines/level2/xgeru.hpp b/test/routines/level2/xgeru.hpp index 1ba83107..96dab22e 100644 --- a/test/routines/level2/xgeru.hpp +++ b/test/routines/level2/xgeru.hpp @@ -120,7 +120,7 @@ class TestXgeru { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXgeru(args.layout, + auto status = cublasXgeru(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, args.m, args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, buffers.y_vec, args.y_offset, args.y_inc, diff --git a/test/routines/level2/xhbmv.hpp b/test/routines/level2/xhbmv.hpp index 2faf86d9..b6844744 100644 --- a/test/routines/level2/xhbmv.hpp +++ b/test/routines/level2/xhbmv.hpp @@ -120,7 +120,7 @@ class TestXhbmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXhbmv(args.layout, + auto status = cublasXhbmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.kl, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, diff --git a/test/routines/level2/xhemv.hpp b/test/routines/level2/xhemv.hpp index b2b6b337..e1f23592 100644 --- a/test/routines/level2/xhemv.hpp +++ b/test/routines/level2/xhemv.hpp @@ -120,7 +120,7 @@ class TestXhemv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXhemv(args.layout, + auto status = cublasXhemv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, diff --git a/test/routines/level2/xher.hpp b/test/routines/level2/xher.hpp index c313d0f5..1ac1247b 100644 --- a/test/routines/level2/xher.hpp +++ b/test/routines/level2/xher.hpp @@ -113,7 +113,7 @@ class TestXher { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<U> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXher(args.layout, + auto status = cublasXher(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xher2.hpp b/test/routines/level2/xher2.hpp index e60486a8..18ccc1ac 100644 --- a/test/routines/level2/xher2.hpp +++ b/test/routines/level2/xher2.hpp @@ -120,7 +120,7 @@ class TestXher2 { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXher2(args.layout, + auto status = cublasXher2(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xhpmv.hpp b/test/routines/level2/xhpmv.hpp index 40ec5475..ad91fe15 100644 --- a/test/routines/level2/xhpmv.hpp +++ b/test/routines/level2/xhpmv.hpp @@ -120,7 +120,7 @@ class TestXhpmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXhpmv(args.layout, + auto status = cublasXhpmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.ap_mat, args.ap_offset, diff --git a/test/routines/level2/xhpr.hpp b/test/routines/level2/xhpr.hpp index 986059bd..f9d580cd 100644 --- a/test/routines/level2/xhpr.hpp +++ b/test/routines/level2/xhpr.hpp @@ -113,7 +113,7 @@ class TestXhpr { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<U> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXhpr(args.layout, + auto status = cublasXhpr(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xhpr2.hpp b/test/routines/level2/xhpr2.hpp index 651989a4..f946ba5c 100644 --- a/test/routines/level2/xhpr2.hpp +++ b/test/routines/level2/xhpr2.hpp @@ -120,7 +120,7 @@ class TestXhpr2 { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXhpr2(args.layout, + auto status = cublasXhpr2(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xsbmv.hpp b/test/routines/level2/xsbmv.hpp index efcdbe34..6481d19b 100644 --- a/test/routines/level2/xsbmv.hpp +++ b/test/routines/level2/xsbmv.hpp @@ -120,7 +120,7 @@ class TestXsbmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsbmv(args.layout, + auto status = cublasXsbmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.kl, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, diff --git a/test/routines/level2/xspmv.hpp b/test/routines/level2/xspmv.hpp index c7d3d348..9815dbee 100644 --- a/test/routines/level2/xspmv.hpp +++ b/test/routines/level2/xspmv.hpp @@ -120,7 +120,7 @@ class TestXspmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXspmv(args.layout, + auto status = cublasXspmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.ap_mat, args.ap_offset, diff --git a/test/routines/level2/xspr.hpp b/test/routines/level2/xspr.hpp index 8d50074c..01a50c38 100644 --- a/test/routines/level2/xspr.hpp +++ b/test/routines/level2/xspr.hpp @@ -113,7 +113,7 @@ class TestXspr { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXspr(args.layout, + auto status = cublasXspr(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xspr2.hpp b/test/routines/level2/xspr2.hpp index 2ee9538a..55f8a141 100644 --- a/test/routines/level2/xspr2.hpp +++ b/test/routines/level2/xspr2.hpp @@ -120,7 +120,7 @@ class TestXspr2 { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXspr2(args.layout, + auto status = cublasXspr2(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xsymv.hpp b/test/routines/level2/xsymv.hpp index 9411fa8d..aec0dfb0 100644 --- a/test/routines/level2/xsymv.hpp +++ b/test/routines/level2/xsymv.hpp @@ -120,7 +120,7 @@ class TestXsymv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsymv(args.layout, + auto status = cublasXsymv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.a_mat, args.a_offset, args.a_ld, diff --git a/test/routines/level2/xsyr.hpp b/test/routines/level2/xsyr.hpp index 8c62f586..78b686d8 100644 --- a/test/routines/level2/xsyr.hpp +++ b/test/routines/level2/xsyr.hpp @@ -113,7 +113,7 @@ class TestXsyr { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsyr(args.layout, + auto status = cublasXsyr(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xsyr2.hpp b/test/routines/level2/xsyr2.hpp index 80838174..38aa4f43 100644 --- a/test/routines/level2/xsyr2.hpp +++ b/test/routines/level2/xsyr2.hpp @@ -120,7 +120,7 @@ class TestXsyr2 { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsyr2(args.layout, + auto status = cublasXsyr2(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), args.n, args.alpha, buffers.x_vec, args.x_offset, args.x_inc, diff --git a/test/routines/level2/xtbmv.hpp b/test/routines/level2/xtbmv.hpp index 9aff2cea..8c7aa381 100644 --- a/test/routines/level2/xtbmv.hpp +++ b/test/routines/level2/xtbmv.hpp @@ -116,7 +116,7 @@ class TestXtbmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXtbmv(args.layout, + auto status = cublasXtbmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.diagonal), diff --git a/test/routines/level2/xtpmv.hpp b/test/routines/level2/xtpmv.hpp index e950b892..3afab978 100644 --- a/test/routines/level2/xtpmv.hpp +++ b/test/routines/level2/xtpmv.hpp @@ -116,7 +116,7 @@ class TestXtpmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXtpmv(args.layout, + auto status = cublasXtpmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.diagonal), diff --git a/test/routines/level2/xtrmv.hpp b/test/routines/level2/xtrmv.hpp index a773b1ca..2b71f151 100644 --- a/test/routines/level2/xtrmv.hpp +++ b/test/routines/level2/xtrmv.hpp @@ -116,7 +116,7 @@ class TestXtrmv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXtrmv(args.layout, + auto status = cublasXtrmv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.diagonal), diff --git a/test/routines/level2/xtrsv.hpp b/test/routines/level2/xtrsv.hpp index 4428271a..85b50e85 100644 --- a/test/routines/level2/xtrsv.hpp +++ b/test/routines/level2/xtrsv.hpp @@ -131,7 +131,7 @@ class TestXtrsv { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXtrsv(args.layout, + auto status = cublasXtrsv(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.diagonal), diff --git a/test/routines/level3/xgemm.hpp b/test/routines/level3/xgemm.hpp index 36fa2f43..7e0ead6d 100644 --- a/test/routines/level3/xgemm.hpp +++ b/test/routines/level3/xgemm.hpp @@ -130,7 +130,7 @@ class TestXgemm { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXgemm(args.layout, + auto status = cublasXgemm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.b_transpose), args.m, args.n, args.k, args.alpha, diff --git a/test/routines/level3/xhemm.hpp b/test/routines/level3/xhemm.hpp index 9400a1fc..a89617b5 100644 --- a/test/routines/level3/xhemm.hpp +++ b/test/routines/level3/xhemm.hpp @@ -130,7 +130,7 @@ class TestXhemm { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXhemm(args.layout, + auto status = cublasXhemm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.side), convertToCUBLAS(args.triangle), args.m, args.n, args.alpha, diff --git a/test/routines/level3/xher2k.hpp b/test/routines/level3/xher2k.hpp index 2b0fff64..55e6d894 100644 --- a/test/routines/level3/xher2k.hpp +++ b/test/routines/level3/xher2k.hpp @@ -132,7 +132,7 @@ class TestXher2k { #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<U> &args, BuffersCUDA<T> &buffers, Queue &) { auto alpha2 = T{args.alpha, args.alpha}; - auto status = cublasXher2k(args.layout, + auto status = cublasXher2k(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), args.n, args.k, alpha2, diff --git a/test/routines/level3/xherk.hpp b/test/routines/level3/xherk.hpp index 3a676f59..3e1e7e02 100644 --- a/test/routines/level3/xherk.hpp +++ b/test/routines/level3/xherk.hpp @@ -118,7 +118,7 @@ class TestXherk { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<U> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXherk(args.layout, + auto status = cublasXherk(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), args.n, args.k, args.alpha, diff --git a/test/routines/level3/xsymm.hpp b/test/routines/level3/xsymm.hpp index 4888091b..5d840d40 100644 --- a/test/routines/level3/xsymm.hpp +++ b/test/routines/level3/xsymm.hpp @@ -130,7 +130,7 @@ class TestXsymm { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsymm(args.layout, + auto status = cublasXsymm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.side), convertToCUBLAS(args.triangle), args.m, args.n, args.alpha, diff --git a/test/routines/level3/xsyr2k.hpp b/test/routines/level3/xsyr2k.hpp index 2fc4090c..4a4a2f10 100644 --- a/test/routines/level3/xsyr2k.hpp +++ b/test/routines/level3/xsyr2k.hpp @@ -128,7 +128,7 @@ class TestXsyr2k { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsyr2k(args.layout, + auto status = cublasXsyr2k(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), args.n, args.k, args.alpha, diff --git a/test/routines/level3/xsyrk.hpp b/test/routines/level3/xsyrk.hpp index 34f8f41a..90e46727 100644 --- a/test/routines/level3/xsyrk.hpp +++ b/test/routines/level3/xsyrk.hpp @@ -118,7 +118,7 @@ class TestXsyrk { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXsyrk(args.layout, + auto status = cublasXsyrk(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), args.n, args.k, args.alpha, diff --git a/test/routines/level3/xtrmm.hpp b/test/routines/level3/xtrmm.hpp index abf77db9..acc00e01 100644 --- a/test/routines/level3/xtrmm.hpp +++ b/test/routines/level3/xtrmm.hpp @@ -122,7 +122,7 @@ class TestXtrmm { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXtrmm(args.layout, + auto status = cublasXtrmm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.side), convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), diff --git a/test/routines/level3/xtrsm.hpp b/test/routines/level3/xtrsm.hpp index 10b216cc..d63c9d79 100644 --- a/test/routines/level3/xtrsm.hpp +++ b/test/routines/level3/xtrsm.hpp @@ -133,7 +133,7 @@ class TestXtrsm { // Describes how to run the cuBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { - auto status = cublasXtrsm(args.layout, + auto status = cublasXtrsm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.side), convertToCUBLAS(args.triangle), convertToCUBLAS(args.a_transpose), diff --git a/test/routines/levelx/xaxpybatched.hpp b/test/routines/levelx/xaxpybatched.hpp index add6c1e1..5385e86e 100644 --- a/test/routines/levelx/xaxpybatched.hpp +++ b/test/routines/levelx/xaxpybatched.hpp @@ -129,7 +129,7 @@ class TestXaxpyBatched { #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { - auto status = cublasXaxpy(args.n, args.alphas[batch], + auto status = cublasXaxpy(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.n, args.alphas[batch], buffers.x_vec, args.x_offsets[batch], args.x_inc, buffers.y_vec, args.y_offsets[batch], args.y_inc); if (status != CUBLAS_STATUS_SUCCESS) { return StatusCode::kUnknownError; } diff --git a/test/routines/levelx/xgemmbatched.hpp b/test/routines/levelx/xgemmbatched.hpp index ae8630c0..ebfd8b19 100644 --- a/test/routines/levelx/xgemmbatched.hpp +++ b/test/routines/levelx/xgemmbatched.hpp @@ -164,7 +164,7 @@ class TestXgemmBatched { #ifdef CLBLAST_REF_CUBLAS static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { - auto status = cublasXgemm(args.layout, + auto status = cublasXgemm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, convertToCUBLAS(args.a_transpose), convertToCUBLAS(args.b_transpose), args.m, args.n, args.k, args.alphas[batch], diff --git a/test/wrapper_cublas.hpp b/test/wrapper_cublas.hpp index 4de8451a..35b1b9c6 100644 --- a/test/wrapper_cublas.hpp +++ b/test/wrapper_cublas.hpp @@ -34,258 +34,226 @@ cublasSideMode_t convertToCUBLAS(const Side v) { return (v == Side::kLeft) ? CUB // Forwards the cuBLAS calls for SROTG/DROTG template <typename T> -cublasStatus_t cublasXrotg(T* sa_buffer, const size_t sa_offset, +cublasStatus_t cublasXrotg(cublasHandle_t handle, T* sa_buffer, const size_t sa_offset, T* sb_buffer, const size_t sb_offset, T* sc_buffer, const size_t sc_offset, T* ss_buffer, const size_t ss_offset); template <> -cublasStatus_t cublasXrotg<float>(float* sa_buffer, const size_t sa_offset, +cublasStatus_t cublasXrotg<float>(cublasHandle_t handle, float* sa_buffer, const size_t sa_offset, float* sb_buffer, const size_t sb_offset, float* sc_buffer, const size_t sc_offset, float* ss_buffer, const size_t ss_offset) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSrotg(handle, &sa_buffer[sa_offset], &sb_buffer[sb_offset], &sc_buffer[sc_offset], &ss_buffer[ss_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXrotg<double>(double* sa_buffer, const size_t sa_offset, +cublasStatus_t cublasXrotg<double>(cublasHandle_t handle, double* sa_buffer, const size_t sa_offset, double* sb_buffer, const size_t sb_offset, double* sc_buffer, const size_t sc_offset, double* ss_buffer, const size_t ss_offset) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDrotg(handle, &sa_buffer[sa_offset], &sb_buffer[sb_offset], &sc_buffer[sc_offset], &ss_buffer[ss_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SROTMG/DROTMG template <typename T> -cublasStatus_t cublasXrotmg(T* sd1_buffer, const size_t sd1_offset, +cublasStatus_t cublasXrotmg(cublasHandle_t handle, T* sd1_buffer, const size_t sd1_offset, T* sd2_buffer, const size_t sd2_offset, T* sx1_buffer, const size_t sx1_offset, const T* sy1_buffer, const size_t sy1_offset, T* sparam_buffer, const size_t sparam_offset); template <> -cublasStatus_t cublasXrotmg<float>(float* sd1_buffer, const size_t sd1_offset, +cublasStatus_t cublasXrotmg<float>(cublasHandle_t handle, float* sd1_buffer, const size_t sd1_offset, float* sd2_buffer, const size_t sd2_offset, float* sx1_buffer, const size_t sx1_offset, const float* sy1_buffer, const size_t sy1_offset, float* sparam_buffer, const size_t sparam_offset) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSrotmg(handle, &sd1_buffer[sd1_offset], &sd2_buffer[sd2_offset], &sx1_buffer[sx1_offset], &sy1_buffer[sy1_offset], &sparam_buffer[sparam_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXrotmg<double>(double* sd1_buffer, const size_t sd1_offset, +cublasStatus_t cublasXrotmg<double>(cublasHandle_t handle, double* sd1_buffer, const size_t sd1_offset, double* sd2_buffer, const size_t sd2_offset, double* sx1_buffer, const size_t sx1_offset, const double* sy1_buffer, const size_t sy1_offset, double* sparam_buffer, const size_t sparam_offset) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDrotmg(handle, &sd1_buffer[sd1_offset], &sd2_buffer[sd2_offset], &sx1_buffer[sx1_offset], &sy1_buffer[sy1_offset], &sparam_buffer[sparam_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SROT/DROT -cublasStatus_t cublasXrot(const size_t n, +cublasStatus_t cublasXrot(cublasHandle_t handle, const size_t n, float* x_buffer, const size_t x_offset, const size_t x_inc, float* y_buffer, const size_t y_offset, const size_t y_inc, const float cos, const float sin) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSrot(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &cos, &sin); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXrot(const size_t n, +cublasStatus_t cublasXrot(cublasHandle_t handle, const size_t n, double* x_buffer, const size_t x_offset, const size_t x_inc, double* y_buffer, const size_t y_offset, const size_t y_inc, const double cos, const double sin) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDrot(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &cos, &sin); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SROTM/DROTM template <typename T> -cublasStatus_t cublasXrotm(const size_t n, +cublasStatus_t cublasXrotm(cublasHandle_t handle, const size_t n, T* x_buffer, const size_t x_offset, const size_t x_inc, T* y_buffer, const size_t y_offset, const size_t y_inc, T* sparam_buffer, const size_t sparam_offset); template <> -cublasStatus_t cublasXrotm<float>(const size_t n, +cublasStatus_t cublasXrotm<float>(cublasHandle_t handle, const size_t n, float* x_buffer, const size_t x_offset, const size_t x_inc, float* y_buffer, const size_t y_offset, const size_t y_inc, float* sparam_buffer, const size_t sparam_offset) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSrotm(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &sparam_buffer[sparam_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXrotm<double>(const size_t n, +cublasStatus_t cublasXrotm<double>(cublasHandle_t handle, const size_t n, double* x_buffer, const size_t x_offset, const size_t x_inc, double* y_buffer, const size_t y_offset, const size_t y_inc, double* sparam_buffer, const size_t sparam_offset) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDrotm(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &sparam_buffer[sparam_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SSWAP/DSWAP/CSWAP/ZSWAP template <typename T> -cublasStatus_t cublasXswap(const size_t n, +cublasStatus_t cublasXswap(cublasHandle_t handle, const size_t n, T* x_buffer, const size_t x_offset, const size_t x_inc, T* y_buffer, const size_t y_offset, const size_t y_inc); template <> -cublasStatus_t cublasXswap<float>(const size_t n, +cublasStatus_t cublasXswap<float>(cublasHandle_t handle, const size_t n, float* x_buffer, const size_t x_offset, const size_t x_inc, float* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSswap(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXswap<double>(const size_t n, +cublasStatus_t cublasXswap<double>(cublasHandle_t handle, const size_t n, double* x_buffer, const size_t x_offset, const size_t x_inc, double* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDswap(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXswap<float2>(const size_t n, +cublasStatus_t cublasXswap<float2>(cublasHandle_t handle, const size_t n, float2* x_buffer, const size_t x_offset, const size_t x_inc, float2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCswap(handle, static_cast<int>(n), reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXswap<double2>(const size_t n, +cublasStatus_t cublasXswap<double2>(cublasHandle_t handle, const size_t n, double2* x_buffer, const size_t x_offset, const size_t x_inc, double2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZswap(handle, static_cast<int>(n), reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXswap<half>(const size_t n, +cublasStatus_t cublasXswap<half>(cublasHandle_t handle, const size_t n, half* x_buffer, const size_t x_offset, const size_t x_inc, half* y_buffer, const size_t y_offset, const size_t y_inc) { return CUBLAS_STATUS_NOT_SUPPORTED; } // Forwards the cuBLAS calls for SSCAL/DSCAL/CSCAL/ZSCAL -cublasStatus_t cublasXscal(const size_t n, +cublasStatus_t cublasXscal(cublasHandle_t handle, const size_t n, const float alpha, float* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSscal(handle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXscal(const size_t n, +cublasStatus_t cublasXscal(cublasHandle_t handle, const size_t n, const double alpha, double* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDscal(handle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXscal(const size_t n, +cublasStatus_t cublasXscal(cublasHandle_t handle, const size_t n, const float2 alpha, float2* x_buffer, const size_t x_offset, const size_t x_inc) { cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCscal(handle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXscal(const size_t n, +cublasStatus_t cublasXscal(cublasHandle_t handle, const size_t n, const double2 alpha, double2* x_buffer, const size_t x_offset, const size_t x_inc) { cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZscal(handle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXscal(const size_t n, +cublasStatus_t cublasXscal(cublasHandle_t handle, const size_t n, const half alpha, half* x_buffer, const size_t x_offset, const size_t x_inc) { return CUBLAS_STATUS_NOT_SUPPORTED; @@ -293,124 +261,108 @@ cublasStatus_t cublasXscal(const size_t n, // Forwards the cuBLAS calls for SCOPY/DCOPY/CCOPY/ZCOPY template <typename T> -cublasStatus_t cublasXcopy(const size_t n, +cublasStatus_t cublasXcopy(cublasHandle_t handle, const size_t n, const T* x_buffer, const size_t x_offset, const size_t x_inc, T* y_buffer, const size_t y_offset, const size_t y_inc); template <> -cublasStatus_t cublasXcopy<float>(const size_t n, +cublasStatus_t cublasXcopy<float>(cublasHandle_t handle, const size_t n, const float* x_buffer, const size_t x_offset, const size_t x_inc, float* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasScopy(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXcopy<double>(const size_t n, +cublasStatus_t cublasXcopy<double>(cublasHandle_t handle, const size_t n, const double* x_buffer, const size_t x_offset, const size_t x_inc, double* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDcopy(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXcopy<float2>(const size_t n, +cublasStatus_t cublasXcopy<float2>(cublasHandle_t handle, const size_t n, const float2* x_buffer, const size_t x_offset, const size_t x_inc, float2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCcopy(handle, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXcopy<double2>(const size_t n, +cublasStatus_t cublasXcopy<double2>(cublasHandle_t handle, const size_t n, const double2* x_buffer, const size_t x_offset, const size_t x_inc, double2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZcopy(handle, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXcopy<half>(const size_t n, +cublasStatus_t cublasXcopy<half>(cublasHandle_t handle, const size_t n, const half* x_buffer, const size_t x_offset, const size_t x_inc, half* y_buffer, const size_t y_offset, const size_t y_inc) { return CUBLAS_STATUS_NOT_SUPPORTED; } // Forwards the cuBLAS calls for SAXPY/DAXPY/CAXPY/ZAXPY -cublasStatus_t cublasXaxpy(const size_t n, +cublasStatus_t cublasXaxpy(cublasHandle_t handle, const size_t n, const float alpha, const float* x_buffer, const size_t x_offset, const size_t x_inc, float* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSaxpy(handle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXaxpy(const size_t n, +cublasStatus_t cublasXaxpy(cublasHandle_t handle, const size_t n, const double alpha, const double* x_buffer, const size_t x_offset, const size_t x_inc, double* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDaxpy(handle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXaxpy(const size_t n, +cublasStatus_t cublasXaxpy(cublasHandle_t handle, const size_t n, const float2 alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, float2* y_buffer, const size_t y_offset, const size_t y_inc) { cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCaxpy(handle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXaxpy(const size_t n, +cublasStatus_t cublasXaxpy(cublasHandle_t handle, const size_t n, const double2 alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, double2* y_buffer, const size_t y_offset, const size_t y_inc) { cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZaxpy(handle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXaxpy(const size_t n, +cublasStatus_t cublasXaxpy(cublasHandle_t handle, const size_t n, const half alpha, const half* x_buffer, const size_t x_offset, const size_t x_inc, half* y_buffer, const size_t y_offset, const size_t y_inc) { @@ -419,40 +371,36 @@ cublasStatus_t cublasXaxpy(const size_t n, // Forwards the cuBLAS calls for SDOT/DDOT template <typename T> -cublasStatus_t cublasXdot(const size_t n, +cublasStatus_t cublasXdot(cublasHandle_t handle, const size_t n, T* dot_buffer, const size_t dot_offset, const T* x_buffer, const size_t x_offset, const size_t x_inc, const T* y_buffer, const size_t y_offset, const size_t y_inc); template <> -cublasStatus_t cublasXdot<float>(const size_t n, +cublasStatus_t cublasXdot<float>(cublasHandle_t handle, const size_t n, float* dot_buffer, const size_t dot_offset, const float* x_buffer, const size_t x_offset, const size_t x_inc, const float* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSdot(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &dot_buffer[dot_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXdot<double>(const size_t n, +cublasStatus_t cublasXdot<double>(cublasHandle_t handle, const size_t n, double* dot_buffer, const size_t dot_offset, const double* x_buffer, const size_t x_offset, const size_t x_inc, const double* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDdot(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &dot_buffer[dot_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXdot<half>(const size_t n, +cublasStatus_t cublasXdot<half>(cublasHandle_t handle, const size_t n, half* dot_buffer, const size_t dot_offset, const half* x_buffer, const size_t x_offset, const size_t x_inc, const half* y_buffer, const size_t y_offset, const size_t y_inc) { @@ -461,129 +409,113 @@ cublasStatus_t cublasXdot<half>(const size_t n, // Forwards the cuBLAS calls for CDOTU/ZDOTU template <typename T> -cublasStatus_t cublasXdotu(const size_t n, +cublasStatus_t cublasXdotu(cublasHandle_t handle, const size_t n, T* dot_buffer, const size_t dot_offset, const T* x_buffer, const size_t x_offset, const size_t x_inc, const T* y_buffer, const size_t y_offset, const size_t y_inc); template <> -cublasStatus_t cublasXdotu<float2>(const size_t n, +cublasStatus_t cublasXdotu<float2>(cublasHandle_t handle, const size_t n, float2* dot_buffer, const size_t dot_offset, const float2* x_buffer, const size_t x_offset, const size_t x_inc, const float2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCdotu(handle, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuComplex*>(&dot_buffer[dot_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXdotu<double2>(const size_t n, +cublasStatus_t cublasXdotu<double2>(cublasHandle_t handle, const size_t n, double2* dot_buffer, const size_t dot_offset, const double2* x_buffer, const size_t x_offset, const size_t x_inc, const double2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZdotu(handle, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuDoubleComplex*>(&dot_buffer[dot_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CDOTC/ZDOTC template <typename T> -cublasStatus_t cublasXdotc(const size_t n, +cublasStatus_t cublasXdotc(cublasHandle_t handle, const size_t n, T* dot_buffer, const size_t dot_offset, const T* x_buffer, const size_t x_offset, const size_t x_inc, const T* y_buffer, const size_t y_offset, const size_t y_inc); template <> -cublasStatus_t cublasXdotc<float2>(const size_t n, +cublasStatus_t cublasXdotc<float2>(cublasHandle_t handle, const size_t n, float2* dot_buffer, const size_t dot_offset, const float2* x_buffer, const size_t x_offset, const size_t x_inc, const float2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCdotc(handle, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuComplex*>(&dot_buffer[dot_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXdotc<double2>(const size_t n, +cublasStatus_t cublasXdotc<double2>(cublasHandle_t handle, const size_t n, double2* dot_buffer, const size_t dot_offset, const double2* x_buffer, const size_t x_offset, const size_t x_inc, const double2* y_buffer, const size_t y_offset, const size_t y_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZdotc(handle, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuDoubleComplex*>(&dot_buffer[dot_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SNRM2/DNRM2/ScNRM2/DzNRM2 template <typename T> -cublasStatus_t cublasXnrm2(const size_t n, +cublasStatus_t cublasXnrm2(cublasHandle_t handle, const size_t n, T* nrm2_buffer, const size_t nrm2_offset, const T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXnrm2<float>(const size_t n, +cublasStatus_t cublasXnrm2<float>(cublasHandle_t handle, const size_t n, float* nrm2_buffer, const size_t nrm2_offset, const float* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSnrm2(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &nrm2_buffer[nrm2_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXnrm2<double>(const size_t n, +cublasStatus_t cublasXnrm2<double>(cublasHandle_t handle, const size_t n, double* nrm2_buffer, const size_t nrm2_offset, const double* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDnrm2(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &nrm2_buffer[nrm2_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXnrm2<float2>(const size_t n, +cublasStatus_t cublasXnrm2<float2>(cublasHandle_t handle, const size_t n, float2* nrm2_buffer, const size_t nrm2_offset, const float2* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasScnrm2(handle, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<float*>(&nrm2_buffer[nrm2_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXnrm2<double2>(const size_t n, +cublasStatus_t cublasXnrm2<double2>(cublasHandle_t handle, const size_t n, double2* nrm2_buffer, const size_t nrm2_offset, const double2* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDznrm2(handle, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<double*>(&nrm2_buffer[nrm2_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXnrm2<half>(const size_t n, +cublasStatus_t cublasXnrm2<half>(cublasHandle_t handle, const size_t n, half* nrm2_buffer, const size_t nrm2_offset, const half* x_buffer, const size_t x_offset, const size_t x_inc) { return CUBLAS_STATUS_NOT_SUPPORTED; @@ -591,59 +523,51 @@ cublasStatus_t cublasXnrm2<half>(const size_t n, // Forwards the cuBLAS calls for SASUM/DASUM/ScASUM/DzASUM template <typename T> -cublasStatus_t cublasXasum(const size_t n, +cublasStatus_t cublasXasum(cublasHandle_t handle, const size_t n, T* asum_buffer, const size_t asum_offset, const T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXasum<float>(const size_t n, +cublasStatus_t cublasXasum<float>(cublasHandle_t handle, const size_t n, float* asum_buffer, const size_t asum_offset, const float* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSasum(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &asum_buffer[asum_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXasum<double>(const size_t n, +cublasStatus_t cublasXasum<double>(cublasHandle_t handle, const size_t n, double* asum_buffer, const size_t asum_offset, const double* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDasum(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), &asum_buffer[asum_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXasum<float2>(const size_t n, +cublasStatus_t cublasXasum<float2>(cublasHandle_t handle, const size_t n, float2* asum_buffer, const size_t asum_offset, const float2* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasScasum(handle, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<float*>(&asum_buffer[asum_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXasum<double2>(const size_t n, +cublasStatus_t cublasXasum<double2>(cublasHandle_t handle, const size_t n, double2* asum_buffer, const size_t asum_offset, const double2* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDzasum(handle, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<double*>(&asum_buffer[asum_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXasum<half>(const size_t n, +cublasStatus_t cublasXasum<half>(cublasHandle_t handle, const size_t n, half* asum_buffer, const size_t asum_offset, const half* x_buffer, const size_t x_offset, const size_t x_inc) { return CUBLAS_STATUS_NOT_SUPPORTED; @@ -651,59 +575,51 @@ cublasStatus_t cublasXasum<half>(const size_t n, // Forwards the cuBLAS calls for iSAMAX/iDAMAX/iCAMAX/iZAMAX/iHAMAX template <typename T> -cublasStatus_t cublasXamax(const size_t n, +cublasStatus_t cublasXamax(cublasHandle_t handle, const size_t n, T* imax_buffer, const size_t imax_offset, const T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXamax<float>(const size_t n, +cublasStatus_t cublasXamax<float>(cublasHandle_t handle, const size_t n, float* imax_buffer, const size_t imax_offset, const float* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasIsamax(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), reinterpret_cast<int*>(&imax_buffer[imax_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXamax<double>(const size_t n, +cublasStatus_t cublasXamax<double>(cublasHandle_t handle, const size_t n, double* imax_buffer, const size_t imax_offset, const double* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasIdamax(handle, static_cast<int>(n), &x_buffer[x_offset], static_cast<int>(x_inc), reinterpret_cast<int*>(&imax_buffer[imax_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXamax<float2>(const size_t n, +cublasStatus_t cublasXamax<float2>(cublasHandle_t handle, const size_t n, float2* imax_buffer, const size_t imax_offset, const float2* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasIcamax(handle, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<int*>(&imax_buffer[imax_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXamax<double2>(const size_t n, +cublasStatus_t cublasXamax<double2>(cublasHandle_t handle, const size_t n, double2* imax_buffer, const size_t imax_offset, const double2* x_buffer, const size_t x_offset, const size_t x_inc) { - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasIzamax(handle, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<int*>(&imax_buffer[imax_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXamax<half>(const size_t n, +cublasStatus_t cublasXamax<half>(cublasHandle_t handle, const size_t n, half* imax_buffer, const size_t imax_offset, const half* x_buffer, const size_t x_offset, const size_t x_inc) { return CUBLAS_STATUS_NOT_SUPPORTED; @@ -714,7 +630,7 @@ cublasStatus_t cublasXamax<half>(const size_t n, // ================================================================================================= // Forwards the cuBLAS calls for SGEMV/DGEMV/CGEMV/ZGEMV -cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgemv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -722,8 +638,6 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp const float beta, float* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSgemv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), &alpha, @@ -731,10 +645,10 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgemv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -742,8 +656,6 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp const double beta, double* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDgemv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), &alpha, @@ -751,10 +663,10 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgemv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -768,8 +680,6 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCgemv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, @@ -777,10 +687,10 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgemv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -794,8 +704,6 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZgemv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, @@ -803,10 +711,10 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgemv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -817,7 +725,7 @@ cublasStatus_t cublasXgemv(const Layout layout, const cublasOperation_t a_transp } // Forwards the cuBLAS calls for SGBMV/DGBMV/CGBMV/ZGBMV -cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgbmv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const size_t kl, const size_t ku, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -825,8 +733,6 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp const float beta, float* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSgbmv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(kl), static_cast<int>(ku), &alpha, @@ -834,10 +740,10 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgbmv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const size_t kl, const size_t ku, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -845,8 +751,6 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp const double beta, double* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDgbmv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(kl), static_cast<int>(ku), &alpha, @@ -854,10 +758,10 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgbmv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const size_t kl, const size_t ku, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -871,8 +775,6 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCgbmv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(kl), static_cast<int>(ku), &alpha_cuda, @@ -880,10 +782,10 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgbmv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const size_t kl, const size_t ku, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -897,8 +799,6 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZgbmv(handle, a_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(kl), static_cast<int>(ku), &alpha_cuda, @@ -906,10 +806,10 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transpose, +cublasStatus_t cublasXgbmv(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const size_t m, const size_t n, const size_t kl, const size_t ku, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -920,7 +820,7 @@ cublasStatus_t cublasXgbmv(const Layout layout, const cublasOperation_t a_transp } // Forwards the cuBLAS calls for CHEMV/ZHEMV -cublasStatus_t cublasXhemv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhemv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -934,8 +834,6 @@ cublasStatus_t cublasXhemv(const Layout layout, const cublasFillMode_t triangle, cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasChemv(handle, triangle, static_cast<int>(n), &alpha_cuda, @@ -943,10 +841,10 @@ cublasStatus_t cublasXhemv(const Layout layout, const cublasFillMode_t triangle, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXhemv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhemv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -960,8 +858,6 @@ cublasStatus_t cublasXhemv(const Layout layout, const cublasFillMode_t triangle, cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZhemv(handle, triangle, static_cast<int>(n), &alpha_cuda, @@ -969,12 +865,12 @@ cublasStatus_t cublasXhemv(const Layout layout, const cublasFillMode_t triangle, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CHBMV/ZHBMV -cublasStatus_t cublasXhbmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhbmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const size_t k, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -988,8 +884,6 @@ cublasStatus_t cublasXhbmv(const Layout layout, const cublasFillMode_t triangle, cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasChbmv(handle, triangle, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -997,10 +891,10 @@ cublasStatus_t cublasXhbmv(const Layout layout, const cublasFillMode_t triangle, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXhbmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhbmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const size_t k, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1014,8 +908,6 @@ cublasStatus_t cublasXhbmv(const Layout layout, const cublasFillMode_t triangle, cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZhbmv(handle, triangle, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -1023,12 +915,12 @@ cublasStatus_t cublasXhbmv(const Layout layout, const cublasFillMode_t triangle, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CHPMV/ZHPMV -cublasStatus_t cublasXhpmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhpmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float2 alpha, const float2* ap_buffer, const size_t ap_offset, @@ -1042,8 +934,6 @@ cublasStatus_t cublasXhpmv(const Layout layout, const cublasFillMode_t triangle, cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasChpmv(handle, triangle, static_cast<int>(n), &alpha_cuda, @@ -1051,10 +941,10 @@ cublasStatus_t cublasXhpmv(const Layout layout, const cublasFillMode_t triangle, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXhpmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhpmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double2 alpha, const double2* ap_buffer, const size_t ap_offset, @@ -1068,8 +958,6 @@ cublasStatus_t cublasXhpmv(const Layout layout, const cublasFillMode_t triangle, cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZhpmv(handle, triangle, static_cast<int>(n), &alpha_cuda, @@ -1077,12 +965,12 @@ cublasStatus_t cublasXhpmv(const Layout layout, const cublasFillMode_t triangle, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SSYMV/DSYMV -cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1090,8 +978,6 @@ cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, const float beta, float* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsymv(handle, triangle, static_cast<int>(n), &alpha, @@ -1099,10 +985,10 @@ cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1110,8 +996,6 @@ cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, const double beta, double* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsymv(handle, triangle, static_cast<int>(n), &alpha, @@ -1119,10 +1003,10 @@ cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1133,7 +1017,7 @@ cublasStatus_t cublasXsymv(const Layout layout, const cublasFillMode_t triangle, } // Forwards the cuBLAS calls for SSBMV/DSBMV -cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsbmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const size_t k, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1141,8 +1025,6 @@ cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, const float beta, float* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsbmv(handle, triangle, static_cast<int>(n), static_cast<int>(k), &alpha, @@ -1150,10 +1032,10 @@ cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsbmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const size_t k, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1161,8 +1043,6 @@ cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, const double beta, double* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsbmv(handle, triangle, static_cast<int>(n), static_cast<int>(k), &alpha, @@ -1170,10 +1050,10 @@ cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsbmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const size_t k, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -1184,7 +1064,7 @@ cublasStatus_t cublasXsbmv(const Layout layout, const cublasFillMode_t triangle, } // Forwards the cuBLAS calls for SSPMV/DSPMV -cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float* ap_buffer, const size_t ap_offset, @@ -1192,8 +1072,6 @@ cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, const float beta, float* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSspmv(handle, triangle, static_cast<int>(n), &alpha, @@ -1201,10 +1079,10 @@ cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double* ap_buffer, const size_t ap_offset, @@ -1212,8 +1090,6 @@ cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, const double beta, double* y_buffer, const size_t y_offset, const size_t y_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDspmv(handle, triangle, static_cast<int>(n), &alpha, @@ -1221,10 +1097,10 @@ cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, &x_buffer[x_offset], static_cast<int>(x_inc), &beta, &y_buffer[y_offset], static_cast<int>(y_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const half alpha, const half* ap_buffer, const size_t ap_offset, @@ -1236,72 +1112,64 @@ cublasStatus_t cublasXspmv(const Layout layout, const cublasFillMode_t triangle, // Forwards the cuBLAS calls for STRMV/DTRMV/CTRMV/ZTRMV template <typename T> -cublasStatus_t cublasXtrmv(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const T* a_buffer, const size_t a_offset, const size_t a_ld, T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXtrmv<float>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmv<float>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float* a_buffer, const size_t a_offset, const size_t a_ld, float* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStrmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrmv<double>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmv<double>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double* a_buffer, const size_t a_offset, const size_t a_ld, double* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtrmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrmv<float2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmv<float2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float2* a_buffer, const size_t a_offset, const size_t a_ld, float2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtrmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrmv<double2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmv<double2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double2* a_buffer, const size_t a_offset, const size_t a_ld, double2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtrmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrmv<half>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmv<half>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const half* a_buffer, const size_t a_offset, const size_t a_ld, half* x_buffer, const size_t x_offset, const size_t x_inc) { @@ -1310,72 +1178,64 @@ cublasStatus_t cublasXtrmv<half>(const Layout layout, const cublasFillMode_t tri // Forwards the cuBLAS calls for STBMV/DTBMV/CTBMV/ZTBMV template <typename T> -cublasStatus_t cublasXtbmv(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const T* a_buffer, const size_t a_offset, const size_t a_ld, T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXtbmv<float>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbmv<float>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const float* a_buffer, const size_t a_offset, const size_t a_ld, float* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStbmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbmv<double>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbmv<double>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const double* a_buffer, const size_t a_offset, const size_t a_ld, double* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtbmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbmv<float2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbmv<float2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const float2* a_buffer, const size_t a_offset, const size_t a_ld, float2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtbmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbmv<double2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbmv<double2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const double2* a_buffer, const size_t a_offset, const size_t a_ld, double2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtbmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbmv<half>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbmv<half>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const half* a_buffer, const size_t a_offset, const size_t a_ld, half* x_buffer, const size_t x_offset, const size_t x_inc) { @@ -1384,72 +1244,64 @@ cublasStatus_t cublasXtbmv<half>(const Layout layout, const cublasFillMode_t tri // Forwards the cuBLAS calls for STPMV/DTPMV/CTPMV/ZTPMV template <typename T> -cublasStatus_t cublasXtpmv(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpmv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const T* ap_buffer, const size_t ap_offset, T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXtpmv<float>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpmv<float>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float* ap_buffer, const size_t ap_offset, float* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStpmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &ap_buffer[ap_offset], &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpmv<double>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpmv<double>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double* ap_buffer, const size_t ap_offset, double* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtpmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &ap_buffer[ap_offset], &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpmv<float2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpmv<float2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float2* ap_buffer, const size_t ap_offset, float2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtpmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&ap_buffer[ap_offset]), reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpmv<double2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpmv<double2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double2* ap_buffer, const size_t ap_offset, double2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtpmv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&ap_buffer[ap_offset]), reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpmv<half>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpmv<half>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const half* ap_buffer, const size_t ap_offset, half* x_buffer, const size_t x_offset, const size_t x_inc) { @@ -1458,241 +1310,213 @@ cublasStatus_t cublasXtpmv<half>(const Layout layout, const cublasFillMode_t tri // Forwards the cuBLAS calls for STRSV/DTRSV/CTRSV/ZTRSV template <typename T> -cublasStatus_t cublasXtrsv(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const T* a_buffer, const size_t a_offset, const size_t a_ld, T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXtrsv<float>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsv<float>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float* a_buffer, const size_t a_offset, const size_t a_ld, float* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStrsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrsv<double>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsv<double>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double* a_buffer, const size_t a_offset, const size_t a_ld, double* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtrsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrsv<float2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsv<float2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float2* a_buffer, const size_t a_offset, const size_t a_ld, float2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtrsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtrsv<double2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsv<double2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double2* a_buffer, const size_t a_offset, const size_t a_ld, double2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtrsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for STBSV/DTBSV/CTBSV/ZTBSV template <typename T> -cublasStatus_t cublasXtbsv(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbsv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const T* a_buffer, const size_t a_offset, const size_t a_ld, T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXtbsv<float>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbsv<float>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const float* a_buffer, const size_t a_offset, const size_t a_ld, float* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStbsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbsv<double>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbsv<double>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const double* a_buffer, const size_t a_offset, const size_t a_ld, double* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtbsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), &a_buffer[a_offset], a_ld, &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbsv<float2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbsv<float2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const float2* a_buffer, const size_t a_offset, const size_t a_ld, float2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtbsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtbsv<double2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtbsv<double2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const size_t k, const double2* a_buffer, const size_t a_offset, const size_t a_ld, double2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtbsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), static_cast<int>(k), reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for STPSV/DTPSV/CTPSV/ZTPSV template <typename T> -cublasStatus_t cublasXtpsv(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpsv(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const T* ap_buffer, const size_t ap_offset, T* x_buffer, const size_t x_offset, const size_t x_inc); template <> -cublasStatus_t cublasXtpsv<float>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpsv<float>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float* ap_buffer, const size_t ap_offset, float* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStpsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &ap_buffer[ap_offset], &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpsv<double>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpsv<double>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double* ap_buffer, const size_t ap_offset, double* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtpsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), &ap_buffer[ap_offset], &x_buffer[x_offset], static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpsv<float2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpsv<float2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const float2* ap_buffer, const size_t ap_offset, float2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtpsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuComplex*>(&ap_buffer[ap_offset]), reinterpret_cast<cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } template <> -cublasStatus_t cublasXtpsv<double2>(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtpsv<double2>(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t n, const double2* ap_buffer, const size_t ap_offset, double2* x_buffer, const size_t x_offset, const size_t x_inc) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtpsv(handle, triangle, a_transpose, diagonal, static_cast<int>(n), reinterpret_cast<const cuDoubleComplex*>(&ap_buffer[ap_offset]), reinterpret_cast<cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc)); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SGER/DGER -cublasStatus_t cublasXger(const Layout layout, +cublasStatus_t cublasXger(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const float alpha, const float* x_buffer, const size_t x_offset, const size_t x_inc, const float* y_buffer, const size_t y_offset, const size_t y_inc, float* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSger(handle, static_cast<int>(m), static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &a_buffer[a_offset], a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXger(const Layout layout, +cublasStatus_t cublasXger(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const double alpha, const double* x_buffer, const size_t x_offset, const size_t x_inc, const double* y_buffer, const size_t y_offset, const size_t y_inc, double* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDger(handle, static_cast<int>(m), static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &a_buffer[a_offset], a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXger(const Layout layout, +cublasStatus_t cublasXger(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const half alpha, const half* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1702,7 +1526,7 @@ cublasStatus_t cublasXger(const Layout layout, } // Forwards the cuBLAS calls for CGERU/ZGERU -cublasStatus_t cublasXgeru(const Layout layout, +cublasStatus_t cublasXgeru(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const float2 alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1712,17 +1536,15 @@ cublasStatus_t cublasXgeru(const Layout layout, cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCgeru(handle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgeru(const Layout layout, +cublasStatus_t cublasXgeru(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const double2 alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1732,19 +1554,17 @@ cublasStatus_t cublasXgeru(const Layout layout, cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZgeru(handle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuDoubleComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CGERC/ZGERC -cublasStatus_t cublasXgerc(const Layout layout, +cublasStatus_t cublasXgerc(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const float2 alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1754,17 +1574,15 @@ cublasStatus_t cublasXgerc(const Layout layout, cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCgerc(handle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgerc(const Layout layout, +cublasStatus_t cublasXgerc(cublasHandle_t handle, const Layout layout, const size_t m, const size_t n, const double2 alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1774,87 +1592,77 @@ cublasStatus_t cublasXgerc(const Layout layout, cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZgerc(handle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuDoubleComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CHER/ZHER -cublasStatus_t cublasXher(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXher(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, float2* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCher(handle, triangle, static_cast<int>(n), &alpha, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXher(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXher(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, double2* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZher(handle, triangle, static_cast<int>(n), &alpha, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuDoubleComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CHPR/ZHPR -cublasStatus_t cublasXhpr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhpr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, float2* ap_buffer, const size_t ap_offset) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasChpr(handle, triangle, static_cast<int>(n), &alpha, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuComplex*>(&ap_buffer[ap_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXhpr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhpr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, double2* ap_buffer, const size_t ap_offset) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZhpr(handle, triangle, static_cast<int>(n), &alpha, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<cuDoubleComplex*>(&ap_buffer[ap_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CHER2/ZHER2 -cublasStatus_t cublasXher2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXher2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float2 alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1864,18 +1672,16 @@ cublasStatus_t cublasXher2(const Layout layout, const cublasFillMode_t triangle, cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCher2(handle, triangle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXher2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXher2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double2 alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1885,20 +1691,18 @@ cublasStatus_t cublasXher2(const Layout layout, const cublasFillMode_t triangle, cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZher2(handle, triangle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuDoubleComplex*>(&a_buffer[a_offset]), a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for CHPR2/ZHPR2 -cublasStatus_t cublasXhpr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhpr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float2 alpha, const float2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1908,18 +1712,16 @@ cublasStatus_t cublasXhpr2(const Layout layout, const cublasFillMode_t triangle, cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasChpr2(handle, triangle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuComplex*>(&ap_buffer[ap_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXhpr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXhpr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double2 alpha, const double2* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1929,52 +1731,46 @@ cublasStatus_t cublasXhpr2(const Layout layout, const cublasFillMode_t triangle, cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZhpr2(handle, triangle, static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&x_buffer[x_offset]), static_cast<int>(x_inc), reinterpret_cast<const cuDoubleComplex*>(&y_buffer[y_offset]), static_cast<int>(y_inc), reinterpret_cast<cuDoubleComplex*>(&ap_buffer[ap_offset])); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SSYR/DSYR -cublasStatus_t cublasXsyr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsyr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float* x_buffer, const size_t x_offset, const size_t x_inc, float* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsyr(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &a_buffer[a_offset], a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsyr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double* x_buffer, const size_t x_offset, const size_t x_inc, double* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsyr(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &a_buffer[a_offset], a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsyr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const half alpha, const half* x_buffer, const size_t x_offset, const size_t x_inc, @@ -1983,39 +1779,35 @@ cublasStatus_t cublasXsyr(const Layout layout, const cublasFillMode_t triangle, } // Forwards the cuBLAS calls for SSPR/DSPR -cublasStatus_t cublasXspr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float* x_buffer, const size_t x_offset, const size_t x_inc, float* ap_buffer, const size_t ap_offset) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSspr(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &ap_buffer[ap_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXspr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double* x_buffer, const size_t x_offset, const size_t x_inc, double* ap_buffer, const size_t ap_offset) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDspr(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &ap_buffer[ap_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXspr(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspr(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const half alpha, const half* x_buffer, const size_t x_offset, const size_t x_inc, @@ -2024,43 +1816,39 @@ cublasStatus_t cublasXspr(const Layout layout, const cublasFillMode_t triangle, } // Forwards the cuBLAS calls for SSYR2/DSYR2 -cublasStatus_t cublasXsyr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsyr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float* x_buffer, const size_t x_offset, const size_t x_inc, const float* y_buffer, const size_t y_offset, const size_t y_inc, float* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsyr2(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &a_buffer[a_offset], a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsyr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double* x_buffer, const size_t x_offset, const size_t x_inc, const double* y_buffer, const size_t y_offset, const size_t y_inc, double* a_buffer, const size_t a_offset, const size_t a_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsyr2(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &a_buffer[a_offset], a_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXsyr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const half alpha, const half* x_buffer, const size_t x_offset, const size_t x_inc, @@ -2070,43 +1858,39 @@ cublasStatus_t cublasXsyr2(const Layout layout, const cublasFillMode_t triangle, } // Forwards the cuBLAS calls for SSPR2/DSPR2 -cublasStatus_t cublasXspr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const float alpha, const float* x_buffer, const size_t x_offset, const size_t x_inc, const float* y_buffer, const size_t y_offset, const size_t y_inc, float* ap_buffer, const size_t ap_offset) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSspr2(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &ap_buffer[ap_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXspr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const double alpha, const double* x_buffer, const size_t x_offset, const size_t x_inc, const double* y_buffer, const size_t y_offset, const size_t y_inc, double* ap_buffer, const size_t ap_offset) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDspr2(handle, triangle, static_cast<int>(n), &alpha, &x_buffer[x_offset], static_cast<int>(x_inc), &y_buffer[y_offset], static_cast<int>(y_inc), &ap_buffer[ap_offset]); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXspr2(const Layout layout, const cublasFillMode_t triangle, +cublasStatus_t cublasXspr2(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const size_t n, const half alpha, const half* x_buffer, const size_t x_offset, const size_t x_inc, @@ -2120,7 +1904,7 @@ cublasStatus_t cublasXspr2(const Layout layout, const cublasFillMode_t triangle, // ================================================================================================= // Forwards the cuBLAS calls for SGEMM/DGEMM/CGEMM/ZGEMM -cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, +cublasStatus_t cublasXgemm(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, const size_t m, const size_t n, const size_t k, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2128,8 +1912,6 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp const float beta, float* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSgemm(handle, a_transpose, b_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(k), &alpha, @@ -2137,10 +1919,10 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp &b_buffer[b_offset], b_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, +cublasStatus_t cublasXgemm(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, const size_t m, const size_t n, const size_t k, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2148,8 +1930,6 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp const double beta, double* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDgemm(handle, a_transpose, b_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(k), &alpha, @@ -2157,10 +1937,10 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp &b_buffer[b_offset], b_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, +cublasStatus_t cublasXgemm(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, const size_t m, const size_t n, const size_t k, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2174,8 +1954,6 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCgemm(handle, a_transpose, b_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -2183,10 +1961,10 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp reinterpret_cast<const cuComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, +cublasStatus_t cublasXgemm(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, const size_t m, const size_t n, const size_t k, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2200,8 +1978,6 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZgemm(handle, a_transpose, b_transpose, static_cast<int>(m), static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -2209,10 +1985,10 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp reinterpret_cast<const cuDoubleComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, +cublasStatus_t cublasXgemm(cublasHandle_t handle, const Layout layout, const cublasOperation_t a_transpose, const cublasOperation_t b_transpose, const size_t m, const size_t n, const size_t k, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2223,7 +1999,7 @@ cublasStatus_t cublasXgemm(const Layout layout, const cublasOperation_t a_transp } // Forwards the cuBLAS calls for SSYMM/DSYMM/CSYMM/ZSYMM -cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2231,8 +2007,6 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con const float beta, float* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsymm(handle, side, triangle, static_cast<int>(m), static_cast<int>(n), &alpha, @@ -2240,10 +2014,10 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con &b_buffer[b_offset], b_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2251,8 +2025,6 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con const double beta, double* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsymm(handle, side, triangle, static_cast<int>(m), static_cast<int>(n), &alpha, @@ -2260,10 +2032,10 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con &b_buffer[b_offset], b_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2277,8 +2049,6 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCsymm(handle, side, triangle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, @@ -2286,10 +2056,10 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con reinterpret_cast<const cuComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2303,8 +2073,6 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZsymm(handle, side, triangle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, @@ -2312,10 +2080,10 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con reinterpret_cast<const cuDoubleComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXsymm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2326,7 +2094,7 @@ cublasStatus_t cublasXsymm(const Layout layout, const cublasSideMode_t side, con } // Forwards the cuBLAS calls for CHEMM/ZHEMM -cublasStatus_t cublasXhemm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXhemm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2340,8 +2108,6 @@ cublasStatus_t cublasXhemm(const Layout layout, const cublasSideMode_t side, con cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasChemm(handle, side, triangle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, @@ -2349,10 +2115,10 @@ cublasStatus_t cublasXhemm(const Layout layout, const cublasSideMode_t side, con reinterpret_cast<const cuComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXhemm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, +cublasStatus_t cublasXhemm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const size_t m, const size_t n, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2366,8 +2132,6 @@ cublasStatus_t cublasXhemm(const Layout layout, const cublasSideMode_t side, con cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZhemm(handle, side, triangle, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, @@ -2375,48 +2139,44 @@ cublasStatus_t cublasXhemm(const Layout layout, const cublasSideMode_t side, con reinterpret_cast<const cuDoubleComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SSYRK/DSYRK/CSYRK/ZSYRK -cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXsyrk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, const float beta, float* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsyrk(handle, triangle, a_transpose, static_cast<int>(n), static_cast<int>(k), &alpha, &a_buffer[a_offset], a_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXsyrk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, const double beta, double* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsyrk(handle, triangle, a_transpose, static_cast<int>(n), static_cast<int>(k), &alpha, &a_buffer[a_offset], a_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXsyrk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2429,18 +2189,16 @@ cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCsyrk(handle, triangle, a_transpose, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, &beta_cuda, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXsyrk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2453,18 +2211,16 @@ cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZsyrk(handle, triangle, a_transpose, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXsyrk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2474,45 +2230,41 @@ cublasStatus_t cublasXsyrk(const Layout layout, const cublasFillMode_t triangle, } // Forwards the cuBLAS calls for CHERK/ZHERK -cublasStatus_t cublasXherk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXherk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const float alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, const float beta, float2* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCherk(handle, triangle, a_transpose, static_cast<int>(n), static_cast<int>(k), &alpha, reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, &beta, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXherk(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, +cublasStatus_t cublasXherk(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const size_t n, const size_t k, const double alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, const double beta, double2* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZherk(handle, triangle, a_transpose, static_cast<int>(n), static_cast<int>(k), &alpha, reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, &beta, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for SSYR2K/DSYR2K/CSYR2K/ZSYR2K -cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXsyr2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2520,8 +2272,6 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle const float beta, float* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasSsyr2k(handle, triangle, ab_transpose, static_cast<int>(n), static_cast<int>(k), &alpha, @@ -2529,10 +2279,10 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle &b_buffer[b_offset], b_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXsyr2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2540,8 +2290,6 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle const double beta, double* c_buffer, const size_t c_offset, const size_t c_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDsyr2k(handle, triangle, ab_transpose, static_cast<int>(n), static_cast<int>(k), &alpha, @@ -2549,10 +2297,10 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle &b_buffer[b_offset], b_ld, &beta, &c_buffer[c_offset], c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXsyr2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2566,8 +2314,6 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle cuComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCsyr2k(handle, triangle, ab_transpose, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -2575,10 +2321,10 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle reinterpret_cast<const cuComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXsyr2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2592,8 +2338,6 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle cuDoubleComplex beta_cuda; beta_cuda.x = beta.real(); beta_cuda.y = beta.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZsyr2k(handle, triangle, ab_transpose, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -2601,10 +2345,10 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle reinterpret_cast<const cuDoubleComplex*>(&b_buffer[b_offset]), b_ld, &beta_cuda, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXsyr2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2615,7 +2359,7 @@ cublasStatus_t cublasXsyr2k(const Layout layout, const cublasFillMode_t triangle } // Forwards the cuBLAS calls for CHER2K/ZHER2K -cublasStatus_t cublasXher2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXher2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2626,8 +2370,6 @@ cublasStatus_t cublasXher2k(const Layout layout, const cublasFillMode_t triangle cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCher2k(handle, triangle, ab_transpose, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -2635,10 +2377,10 @@ cublasStatus_t cublasXher2k(const Layout layout, const cublasFillMode_t triangle reinterpret_cast<const cuComplex*>(&b_buffer[b_offset]), b_ld, &beta, reinterpret_cast<cuComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXher2k(const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, +cublasStatus_t cublasXher2k(cublasHandle_t handle, const Layout layout, const cublasFillMode_t triangle, const cublasOperation_t ab_transpose, const size_t n, const size_t k, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2649,8 +2391,6 @@ cublasStatus_t cublasXher2k(const Layout layout, const cublasFillMode_t triangle cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZher2k(handle, triangle, ab_transpose, static_cast<int>(n), static_cast<int>(k), &alpha_cuda, @@ -2658,46 +2398,42 @@ cublasStatus_t cublasXher2k(const Layout layout, const cublasFillMode_t triangle reinterpret_cast<const cuDoubleComplex*>(&b_buffer[b_offset]), b_ld, &beta, reinterpret_cast<cuDoubleComplex*>(&c_buffer[c_offset]), c_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } // Forwards the cuBLAS calls for STRMM/DTRMM/CTRMM/ZTRMM -cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, float* b_buffer, const size_t b_offset, const size_t b_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStrmm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha, &a_buffer[a_offset], a_ld, &a_buffer[a_offset], a_ld, &b_buffer[b_offset], b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, double* b_buffer, const size_t b_offset, const size_t b_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtrmm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha, &a_buffer[a_offset], a_ld, &a_buffer[a_offset], a_ld, &b_buffer[b_offset], b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2706,18 +2442,16 @@ cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, con cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtrmm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuComplex*>(&b_buffer[b_offset]), b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2726,18 +2460,16 @@ cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, con cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtrmm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuDoubleComplex*>(&b_buffer[b_offset]), b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrmm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const half alpha, const half* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2746,39 +2478,35 @@ cublasStatus_t cublasXtrmm(const Layout layout, const cublasSideMode_t side, con } // Forwards the cuBLAS calls for STRSM/DTRSM/CTRSM/ZTRSM -cublasStatus_t cublasXtrsm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const float alpha, const float* a_buffer, const size_t a_offset, const size_t a_ld, float* b_buffer, const size_t b_offset, const size_t b_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasStrsm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha, &a_buffer[a_offset], a_ld, &b_buffer[b_offset], b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrsm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const double alpha, const double* a_buffer, const size_t a_offset, const size_t a_ld, double* b_buffer, const size_t b_offset, const size_t b_ld) { if (layout == Layout::kRowMajor) { return CUBLAS_STATUS_NOT_SUPPORTED; } - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasDtrsm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha, &a_buffer[a_offset], a_ld, &b_buffer[b_offset], b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrsm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const float2 alpha, const float2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2787,17 +2515,15 @@ cublasStatus_t cublasXtrsm(const Layout layout, const cublasSideMode_t side, con cuComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasCtrsm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuComplex*>(&b_buffer[b_offset]), b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } -cublasStatus_t cublasXtrsm(const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, +cublasStatus_t cublasXtrsm(cublasHandle_t handle, const Layout layout, const cublasSideMode_t side, const cublasFillMode_t triangle, const cublasOperation_t a_transpose, const cublasDiagType_t diagonal, const size_t m, const size_t n, const double2 alpha, const double2* a_buffer, const size_t a_offset, const size_t a_ld, @@ -2806,14 +2532,12 @@ cublasStatus_t cublasXtrsm(const Layout layout, const cublasSideMode_t side, con cuDoubleComplex alpha_cuda; alpha_cuda.x = alpha.real(); alpha_cuda.y = alpha.imag(); - cublasHandle_t handle; - if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS) { return CUBLAS_STATUS_NOT_INITIALIZED; } auto status = cublasZtrsm(handle, side, triangle, a_transpose, diagonal, static_cast<int>(m), static_cast<int>(n), &alpha_cuda, reinterpret_cast<const cuDoubleComplex*>(&a_buffer[a_offset]), a_ld, reinterpret_cast<cuDoubleComplex*>(&b_buffer[b_offset]), b_ld); - cublasDestroy(handle); + cudaDeviceSynchronize(); return status; } diff --git a/test/wrapper_cuda.hpp b/test/wrapper_cuda.hpp index 509de9d1..51f897c4 100644 --- a/test/wrapper_cuda.hpp +++ b/test/wrapper_cuda.hpp @@ -29,17 +29,47 @@ namespace clblast { // ================================================================================================= +#ifdef CLBLAST_REF_CUBLAS + template <typename T> + void cublasSetup(Arguments<T> &args) { + cudaSetDevice(static_cast<int>(args.device_id)); + auto status = cublasCreate(reinterpret_cast<cublasHandle_t*>(&args.cublas_handle)); + if (status != CUBLAS_STATUS_SUCCESS) { + throw std::runtime_error("CUDA cublasCreate error"); + } + } +#endif + +#ifdef CLBLAST_REF_CUBLAS + template <typename T> + void cublasTeardown(Arguments<T> &args) { + auto status = cublasDestroy(reinterpret_cast<cublasHandle_t>(args.cublas_handle)); + if (status != CUBLAS_STATUS_SUCCESS) { + throw std::runtime_error("CUDA cublasDestroy error"); + } + } +#endif + +// ================================================================================================= + // Copies data from the CUDA device to the host and frees-up the CUDA memory afterwards #ifdef CLBLAST_REF_CUBLAS template <typename T> - void CUDAToHost(T* buffer_cuda, std::vector<T> &buffer_host, const size_t size) { - cudaMemcpy( + void CUDAToHost(T** buffer_cuda, std::vector<T> &buffer_host, const size_t size) { + auto status1 = cudaMemcpy( reinterpret_cast<void*>(buffer_host.data()), - reinterpret_cast<void*>(buffer_cuda), + reinterpret_cast<void*>(*buffer_cuda), size*sizeof(T), cudaMemcpyDeviceToHost ); - cudaFree(buffer_cuda); + if (status1 != cudaSuccess) { + throw std::runtime_error("CUDA cudaMemcpy error with status: "+ToString(static_cast<int>(status1))); + } + auto status2 = cudaFree(*buffer_cuda); + if (status2 != cudaSuccess) { + throw std::runtime_error("CUDA cudaFree error with status: "+ToString(static_cast<int>(status2))); + } + *buffer_cuda = nullptr; } #else template <typename T> void CUDAToHost(T*, const std::vector<T>&, const size_t) { } @@ -48,14 +78,22 @@ namespace clblast { // Allocates space on the CUDA device and copies in data from the host #ifdef CLBLAST_REF_CUBLAS template <typename T> - void HostToCUDA(T* buffer_cuda, std::vector<T> &buffer_host, const size_t size) { - cudaMalloc(reinterpret_cast<void**>(&buffer_cuda), size*sizeof(T)); - cudaMemcpy( - reinterpret_cast<void*>(buffer_cuda), + void HostToCUDA(T** buffer_cuda, std::vector<T> &buffer_host, const size_t size) { + if (*buffer_cuda == nullptr) { + auto status1 = cudaMalloc(reinterpret_cast<void**>(buffer_cuda), size*sizeof(T)); + if (status1 != cudaSuccess) { + throw std::runtime_error("CUDA cudaMalloc error with status: "+ToString(static_cast<int>(status1))); + } + } + auto status2 = cudaMemcpy( + reinterpret_cast<void*>(*buffer_cuda), reinterpret_cast<void*>(buffer_host.data()), size*sizeof(T), cudaMemcpyHostToDevice ); + if (status2 != cudaSuccess) { + throw std::runtime_error("CUDA cudaMemcpy error with status: "+ToString(static_cast<int>(status2))); + } } #else template <typename T> void HostToCUDA(T*, const std::vector<T>&, const size_t) { } @@ -65,26 +103,26 @@ namespace clblast { template <typename T> struct BuffersCUDA { - T* x_vec; - T* y_vec; - T* a_mat; - T* b_mat; - T* c_mat; - T* ap_mat; - T* scalar; + T* x_vec = nullptr; + T* y_vec = nullptr; + T* a_mat = nullptr; + T* b_mat = nullptr; + T* c_mat = nullptr; + T* ap_mat = nullptr; + T* scalar = nullptr; }; template <typename T, typename U> void CUDAToHost(const Arguments<U> &args, BuffersCUDA<T> &buffers, BuffersHost<T> &buffers_host, const std::vector<std::string> &names) { for (auto &name: names) { - if (name == kBufVecX) { buffers_host.x_vec = std::vector<T>(args.x_size, static_cast<T>(0)); CUDAToHost(buffers.x_vec, buffers_host.x_vec, args.x_size); } - else if (name == kBufVecY) { buffers_host.y_vec = std::vector<T>(args.y_size, static_cast<T>(0)); CUDAToHost(buffers.y_vec, buffers_host.y_vec, args.y_size); } - else if (name == kBufMatA) { buffers_host.a_mat = std::vector<T>(args.a_size, static_cast<T>(0)); CUDAToHost(buffers.a_mat, buffers_host.a_mat, args.a_size); } - else if (name == kBufMatB) { buffers_host.b_mat = std::vector<T>(args.b_size, static_cast<T>(0)); CUDAToHost(buffers.b_mat, buffers_host.b_mat, args.b_size); } - else if (name == kBufMatC) { buffers_host.c_mat = std::vector<T>(args.c_size, static_cast<T>(0)); CUDAToHost(buffers.c_mat, buffers_host.c_mat, args.c_size); } - else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector<T>(args.ap_size, static_cast<T>(0)); CUDAToHost(buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } - else if (name == kBufScalar) { buffers_host.scalar = std::vector<T>(args.scalar_size, static_cast<T>(0)); CUDAToHost(buffers.scalar, buffers_host.scalar, args.scalar_size); } + if (name == kBufVecX) { buffers_host.x_vec = std::vector<T>(args.x_size, static_cast<T>(0)); CUDAToHost(&buffers.x_vec, buffers_host.x_vec, args.x_size); } + else if (name == kBufVecY) { buffers_host.y_vec = std::vector<T>(args.y_size, static_cast<T>(0)); CUDAToHost(&buffers.y_vec, buffers_host.y_vec, args.y_size); } + else if (name == kBufMatA) { buffers_host.a_mat = std::vector<T>(args.a_size, static_cast<T>(0)); CUDAToHost(&buffers.a_mat, buffers_host.a_mat, args.a_size); } + else if (name == kBufMatB) { buffers_host.b_mat = std::vector<T>(args.b_size, static_cast<T>(0)); CUDAToHost(&buffers.b_mat, buffers_host.b_mat, args.b_size); } + else if (name == kBufMatC) { buffers_host.c_mat = std::vector<T>(args.c_size, static_cast<T>(0)); CUDAToHost(&buffers.c_mat, buffers_host.c_mat, args.c_size); } + else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector<T>(args.ap_size, static_cast<T>(0)); CUDAToHost(&buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } + else if (name == kBufScalar) { buffers_host.scalar = std::vector<T>(args.scalar_size, static_cast<T>(0)); CUDAToHost(&buffers.scalar, buffers_host.scalar, args.scalar_size); } else { throw std::runtime_error("Invalid buffer name"); } } } @@ -93,13 +131,13 @@ template <typename T, typename U> void HostToCUDA(const Arguments<U> &args, BuffersCUDA<T> &buffers, BuffersHost<T> &buffers_host, const std::vector<std::string> &names) { for (auto &name: names) { - if (name == kBufVecX) { HostToCUDA(buffers.x_vec, buffers_host.x_vec, args.x_size); } - else if (name == kBufVecY) { HostToCUDA(buffers.y_vec, buffers_host.y_vec, args.y_size); } - else if (name == kBufMatA) { HostToCUDA(buffers.a_mat, buffers_host.a_mat, args.a_size); } - else if (name == kBufMatB) { HostToCUDA(buffers.b_mat, buffers_host.b_mat, args.b_size); } - else if (name == kBufMatC) { HostToCUDA(buffers.c_mat, buffers_host.c_mat, args.c_size); } - else if (name == kBufMatAP) { HostToCUDA(buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } - else if (name == kBufScalar) { HostToCUDA(buffers.scalar, buffers_host.scalar, args.scalar_size); } + if (name == kBufVecX) { HostToCUDA(&buffers.x_vec, buffers_host.x_vec, args.x_size); } + else if (name == kBufVecY) { HostToCUDA(&buffers.y_vec, buffers_host.y_vec, args.y_size); } + else if (name == kBufMatA) { HostToCUDA(&buffers.a_mat, buffers_host.a_mat, args.a_size); } + else if (name == kBufMatB) { HostToCUDA(&buffers.b_mat, buffers_host.b_mat, args.b_size); } + else if (name == kBufMatC) { HostToCUDA(&buffers.c_mat, buffers_host.c_mat, args.c_size); } + else if (name == kBufMatAP) { HostToCUDA(&buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } + else if (name == kBufScalar) { HostToCUDA(&buffers.scalar, buffers_host.scalar, args.scalar_size); } else { throw std::runtime_error("Invalid buffer name"); } } } |