summaryrefslogtreecommitdiff
path: root/test/routines/level3
diff options
context:
space:
mode:
Diffstat (limited to 'test/routines/level3')
-rw-r--r--test/routines/level3/xgemm.hpp14
-rw-r--r--test/routines/level3/xhemm.hpp14
-rw-r--r--test/routines/level3/xher2k.hpp14
-rw-r--r--test/routines/level3/xherk.hpp13
-rw-r--r--test/routines/level3/xsymm.hpp14
-rw-r--r--test/routines/level3/xsyr2k.hpp14
-rw-r--r--test/routines/level3/xsyrk.hpp13
-rw-r--r--test/routines/level3/xtrmm.hpp15
-rw-r--r--test/routines/level3/xtrsm.hpp15
9 files changed, 126 insertions, 0 deletions
diff --git a/test/routines/level3/xgemm.hpp b/test/routines/level3/xgemm.hpp
index d6ad98f9..36fa2f43 100644
--- a/test/routines/level3/xgemm.hpp
+++ b/test/routines/level3/xgemm.hpp
@@ -127,6 +127,20 @@ class TestXgemm {
}
#endif
+ // 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,
+ convertToCUBLAS(args.a_transpose),
+ convertToCUBLAS(args.b_transpose),
+ args.m, args.n, args.k, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld, args.beta,
+ buffers.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xhemm.hpp b/test/routines/level3/xhemm.hpp
index beadf62d..9400a1fc 100644
--- a/test/routines/level3/xhemm.hpp
+++ b/test/routines/level3/xhemm.hpp
@@ -127,6 +127,20 @@ class TestXhemm {
}
#endif
+ // 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,
+ convertToCUBLAS(args.side),
+ convertToCUBLAS(args.triangle),
+ args.m, args.n, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld, args.beta,
+ buffers.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xher2k.hpp b/test/routines/level3/xher2k.hpp
index b5d22579..b341c4d7 100644
--- a/test/routines/level3/xher2k.hpp
+++ b/test/routines/level3/xher2k.hpp
@@ -128,6 +128,20 @@ class TestXher2k {
}
#endif
+ // 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 = cublasXher2k(args.layout,
+ convertToCUBLAS(args.triangle),
+ convertToCUBLAS(args.a_transpose),
+ args.n, args.k, alpha2,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld, args.beta,
+ buffers.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<U> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xherk.hpp b/test/routines/level3/xherk.hpp
index 558f4e76..676d5286 100644
--- a/test/routines/level3/xherk.hpp
+++ b/test/routines/level3/xherk.hpp
@@ -115,6 +115,19 @@ class TestXherk {
}
#endif
+ // 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 = cublasXherk(args.layout,
+ convertToCUBLAS(args.triangle),
+ convertToCUBLAS(args.a_transpose),
+ args.n, args.k, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld, args.beta,
+ buffers.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<U> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xsymm.hpp b/test/routines/level3/xsymm.hpp
index 704a8f9e..4888091b 100644
--- a/test/routines/level3/xsymm.hpp
+++ b/test/routines/level3/xsymm.hpp
@@ -127,6 +127,20 @@ class TestXsymm {
}
#endif
+ // 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,
+ convertToCUBLAS(args.side),
+ convertToCUBLAS(args.triangle),
+ args.m, args.n, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld, args.beta,
+ buffers.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xsyr2k.hpp b/test/routines/level3/xsyr2k.hpp
index c321b9cf..2fc4090c 100644
--- a/test/routines/level3/xsyr2k.hpp
+++ b/test/routines/level3/xsyr2k.hpp
@@ -125,6 +125,20 @@ class TestXsyr2k {
}
#endif
+ // 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,
+ convertToCUBLAS(args.triangle),
+ convertToCUBLAS(args.a_transpose),
+ args.n, args.k, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld, args.beta,
+ buffers.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xsyrk.hpp b/test/routines/level3/xsyrk.hpp
index 00a3013d..2ee24a2e 100644
--- a/test/routines/level3/xsyrk.hpp
+++ b/test/routines/level3/xsyrk.hpp
@@ -115,6 +115,19 @@ class TestXsyrk {
}
#endif
+ // 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,
+ convertToCUBLAS(args.triangle),
+ convertToCUBLAS(args.a_transpose),
+ args.n, args.k, args.alpha,
+ buffer.a_mat, args.a_offset, args.a_ld, args.beta,
+ buffer.c_mat, args.c_offset, args.c_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.c_size, static_cast<T>(0));
diff --git a/test/routines/level3/xtrmm.hpp b/test/routines/level3/xtrmm.hpp
index 660001df..abf77db9 100644
--- a/test/routines/level3/xtrmm.hpp
+++ b/test/routines/level3/xtrmm.hpp
@@ -119,6 +119,21 @@ class TestXtrmm {
}
#endif
+ // 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,
+ convertToCUBLAS(args.side),
+ convertToCUBLAS(args.triangle),
+ convertToCUBLAS(args.a_transpose),
+ convertToCUBLAS(args.diagonal),
+ args.m, args.n, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.b_size, static_cast<T>(0));
diff --git a/test/routines/level3/xtrsm.hpp b/test/routines/level3/xtrsm.hpp
index 9e8b9565..10b216cc 100644
--- a/test/routines/level3/xtrsm.hpp
+++ b/test/routines/level3/xtrsm.hpp
@@ -130,6 +130,21 @@ class TestXtrsm {
}
#endif
+ // 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,
+ convertToCUBLAS(args.side),
+ convertToCUBLAS(args.triangle),
+ convertToCUBLAS(args.a_transpose),
+ convertToCUBLAS(args.diagonal),
+ args.m, args.n, args.alpha,
+ buffers.a_mat, args.a_offset, args.a_ld,
+ buffers.b_mat, args.b_offset, args.b_ld);
+ if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+ }
+ #endif
+
// Describes how to download the results of the computation (more importantly: which buffer)
static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.b_size, static_cast<T>(0));
generated by cgit v1.2.3 (git 2.39.1) at 2024-07-12 00:09:04 +0000