Fixed CUDA malloc and cuBLAS handles: cuBLAS as a performance-reference now works

44 files changed, 44 insertions, 44 deletions

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],