summaryrefslogtreecommitdiff
path: root/test
diff options
context:
space:
mode:
Diffstat (limited to 'test')
-rw-r--r--test/correctness/testblas.cc68
-rw-r--r--test/correctness/testblas.h4
-rw-r--r--test/correctness/tester.cc8
-rw-r--r--test/correctness/tester.h5
-rw-r--r--test/performance/client.cc28
-rw-r--r--test/performance/client.h9
-rw-r--r--test/routines/level1/xaxpy.h11
-rw-r--r--test/routines/level2/xgemv.h11
-rw-r--r--test/routines/level3/xgemm.h11
-rw-r--r--test/routines/level3/xhemm.h11
-rw-r--r--test/routines/level3/xher2k.h11
-rw-r--r--test/routines/level3/xherk.h11
-rw-r--r--test/routines/level3/xsymm.h11
-rw-r--r--test/routines/level3/xsyr2k.h11
-rw-r--r--test/routines/level3/xsyrk.h11
-rw-r--r--test/routines/level3/xtrmm.h11
16 files changed, 98 insertions, 134 deletions
diff --git a/test/correctness/testblas.cc b/test/correctness/testblas.cc
index 5951b177..ff81f4c3 100644
--- a/test/correctness/testblas.cc
+++ b/test/correctness/testblas.cc
@@ -76,31 +76,31 @@ void TestBlas<T,U>::TestRegular(std::vector<Arguments<U>> &test_vector, const st
for (auto &args: test_vector) {
// Runs the reference clBLAS code
- auto x_vec1 = Buffer(context_, CL_MEM_READ_WRITE, args.x_size*sizeof(T));
- auto y_vec1 = Buffer(context_, CL_MEM_READ_WRITE, args.y_size*sizeof(T));
- auto a_mat1 = Buffer(context_, CL_MEM_READ_WRITE, args.a_size*sizeof(T));
- auto b_mat1 = Buffer(context_, CL_MEM_READ_WRITE, args.b_size*sizeof(T));
- auto c_mat1 = Buffer(context_, CL_MEM_READ_WRITE, args.c_size*sizeof(T));
- x_vec1.WriteBuffer(queue_, args.x_size*sizeof(T), x_source_);
- y_vec1.WriteBuffer(queue_, args.y_size*sizeof(T), y_source_);
- a_mat1.WriteBuffer(queue_, args.a_size*sizeof(T), a_source_);
- b_mat1.WriteBuffer(queue_, args.b_size*sizeof(T), b_source_);
- c_mat1.WriteBuffer(queue_, args.c_size*sizeof(T), c_source_);
- auto buffers1 = Buffers{x_vec1, y_vec1, a_mat1, b_mat1, c_mat1};
+ auto x_vec1 = Buffer<T>(context_, args.x_size);
+ auto y_vec1 = Buffer<T>(context_, args.y_size);
+ auto a_mat1 = Buffer<T>(context_, args.a_size);
+ auto b_mat1 = Buffer<T>(context_, args.b_size);
+ auto c_mat1 = Buffer<T>(context_, args.c_size);
+ x_vec1.Write(queue_, args.x_size, x_source_);
+ y_vec1.Write(queue_, args.y_size, y_source_);
+ a_mat1.Write(queue_, args.a_size, a_source_);
+ b_mat1.Write(queue_, args.b_size, b_source_);
+ c_mat1.Write(queue_, args.c_size, c_source_);
+ auto buffers1 = Buffers<T>{x_vec1, y_vec1, a_mat1, b_mat1, c_mat1};
auto status1 = run_reference_(args, buffers1, queue_);
// Runs the CLBlast code
- auto x_vec2 = Buffer(context_, CL_MEM_READ_WRITE, args.x_size*sizeof(T));
- auto y_vec2 = Buffer(context_, CL_MEM_READ_WRITE, args.y_size*sizeof(T));
- auto a_mat2 = Buffer(context_, CL_MEM_READ_WRITE, args.a_size*sizeof(T));
- auto b_mat2 = Buffer(context_, CL_MEM_READ_WRITE, args.b_size*sizeof(T));
- auto c_mat2 = Buffer(context_, CL_MEM_READ_WRITE, args.c_size*sizeof(T));
- x_vec2.WriteBuffer(queue_, args.x_size*sizeof(T), x_source_);
- y_vec2.WriteBuffer(queue_, args.y_size*sizeof(T), y_source_);
- a_mat2.WriteBuffer(queue_, args.a_size*sizeof(T), a_source_);
- b_mat2.WriteBuffer(queue_, args.b_size*sizeof(T), b_source_);
- c_mat2.WriteBuffer(queue_, args.c_size*sizeof(T), c_source_);
- auto buffers2 = Buffers{x_vec2, y_vec2, a_mat2, b_mat2, c_mat2};
+ auto x_vec2 = Buffer<T>(context_, args.x_size);
+ auto y_vec2 = Buffer<T>(context_, args.y_size);
+ auto a_mat2 = Buffer<T>(context_, args.a_size);
+ auto b_mat2 = Buffer<T>(context_, args.b_size);
+ auto c_mat2 = Buffer<T>(context_, args.c_size);
+ x_vec2.Write(queue_, args.x_size, x_source_);
+ y_vec2.Write(queue_, args.y_size, y_source_);
+ a_mat2.Write(queue_, args.a_size, a_source_);
+ b_mat2.Write(queue_, args.b_size, b_source_);
+ c_mat2.Write(queue_, args.c_size, c_source_);
+ auto buffers2 = Buffers<T>{x_vec2, y_vec2, a_mat2, b_mat2, c_mat2};
auto status2 = run_routine_(args, buffers2, queue_);
// Tests for equality of the two status codes
@@ -149,25 +149,25 @@ void TestBlas<T,U>::TestInvalid(std::vector<Arguments<U>> &test_vector, const st
auto a1 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.a_size*sizeof(T), nullptr,nullptr);
auto b1 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.b_size*sizeof(T), nullptr,nullptr);
auto c1 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.c_size*sizeof(T), nullptr,nullptr);
- auto x_vec1 = Buffer(x1);
- auto y_vec1 = Buffer(y1);
- auto a_mat1 = Buffer(a1);
- auto b_mat1 = Buffer(b1);
- auto c_mat1 = Buffer(c1);
+ auto x_vec1 = Buffer<T>(x1);
+ auto y_vec1 = Buffer<T>(y1);
+ auto a_mat1 = Buffer<T>(a1);
+ auto b_mat1 = Buffer<T>(b1);
+ auto c_mat1 = Buffer<T>(c1);
auto x2 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.x_size*sizeof(T), nullptr,nullptr);
auto y2 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.y_size*sizeof(T), nullptr,nullptr);
auto a2 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.a_size*sizeof(T), nullptr,nullptr);
auto b2 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.b_size*sizeof(T), nullptr,nullptr);
auto c2 = clCreateBuffer(context_(), CL_MEM_READ_WRITE, args.c_size*sizeof(T), nullptr,nullptr);
- auto x_vec2 = Buffer(x2);
- auto y_vec2 = Buffer(y2);
- auto a_mat2 = Buffer(a2);
- auto b_mat2 = Buffer(b2);
- auto c_mat2 = Buffer(c2);
+ auto x_vec2 = Buffer<T>(x2);
+ auto y_vec2 = Buffer<T>(y2);
+ auto a_mat2 = Buffer<T>(a2);
+ auto b_mat2 = Buffer<T>(b2);
+ auto c_mat2 = Buffer<T>(c2);
// Runs the two routines
- auto status1 = run_reference_(args, Buffers{x_vec1, y_vec1, a_mat1, b_mat1, c_mat1}, queue_);
- auto status2 = run_routine_(args, Buffers{x_vec2, y_vec2, a_mat2, b_mat2, c_mat2}, queue_);
+ auto status1 = run_reference_(args, Buffers<T>{x_vec1, y_vec1, a_mat1, b_mat1, c_mat1}, queue_);
+ auto status2 = run_routine_(args, Buffers<T>{x_vec2, y_vec2, a_mat2, b_mat2, c_mat2}, queue_);
// Tests for equality of the two status codes
TestErrorCodes(status1, status2, args);
diff --git a/test/correctness/testblas.h b/test/correctness/testblas.h
index 96c140c1..af8a4d0e 100644
--- a/test/correctness/testblas.h
+++ b/test/correctness/testblas.h
@@ -66,8 +66,8 @@ class TestBlas: public Tester<T,U> {
static const std::vector<Transpose> kTransposes; // Data-type dependent, see .cc-file
// Shorthand for the routine-specific functions passed to the tester
- using Routine = std::function<StatusCode(const Arguments<U>&, const Buffers&, CommandQueue&)>;
- using ResultGet = std::function<std::vector<T>(const Arguments<U>&, Buffers&, CommandQueue&)>;
+ using Routine = std::function<StatusCode(const Arguments<U>&, const Buffers<T>&, Queue&)>;
+ using ResultGet = std::function<std::vector<T>(const Arguments<U>&, Buffers<T>&, Queue&)>;
using ResultIndex = std::function<size_t(const Arguments<U>&, const size_t, const size_t)>;
using ResultIterator = std::function<size_t(const Arguments<U>&)>;
diff --git a/test/correctness/tester.cc b/test/correctness/tester.cc
index 378968ed..002cb1a6 100644
--- a/test/correctness/tester.cc
+++ b/test/correctness/tester.cc
@@ -28,9 +28,9 @@ Tester<T,U>::Tester(int argc, char *argv[], const bool silent,
const std::string &name, const std::vector<std::string> &options):
help_("Options given/available:\n"),
platform_(Platform(GetArgument(argc, argv, help_, kArgPlatform, size_t{0}))),
- device_(Device(platform_, kDeviceType, GetArgument(argc, argv, help_, kArgDevice, size_t{0}))),
+ device_(Device(platform_, GetArgument(argc, argv, help_, kArgDevice, size_t{0}))),
context_(Context(device_)),
- queue_(CommandQueue(context_, device_)),
+ queue_(Queue(context_, device_)),
full_test_(CheckArgument(argc, argv, help_, kArgFullTest)),
error_log_{},
num_passed_{0},
@@ -339,11 +339,11 @@ template <> const std::vector<double2> GetExampleScalars(const bool full_test) {
template <> bool PrecisionSupported<float>(const Device &) { return true; }
template <> bool PrecisionSupported<float2>(const Device &) { return true; }
template <> bool PrecisionSupported<double>(const Device &device) {
- auto extensions = device.Extensions();
+ auto extensions = device.Capabilities();
return (extensions.find(kKhronosDoublePrecision) == std::string::npos) ? false : true;
}
template <> bool PrecisionSupported<double2>(const Device &device) {
- auto extensions = device.Extensions();
+ auto extensions = device.Capabilities();
return (extensions.find(kKhronosDoublePrecision) == std::string::npos) ? false : true;
}
diff --git a/test/correctness/tester.h b/test/correctness/tester.h
index 93515138..06f4afbe 100644
--- a/test/correctness/tester.h
+++ b/test/correctness/tester.h
@@ -36,9 +36,6 @@ template <typename T, typename U>
class Tester {
public:
- // Types of devices to consider
- const cl_device_type kDeviceType = CL_DEVICE_TYPE_ALL;
-
// Maximum number of test results printed on a single line
static constexpr auto kResultsPerLine = size_t{64};
@@ -92,7 +89,7 @@ class Tester {
Platform platform_;
Device device_;
Context context_;
- CommandQueue queue_;
+ Queue queue_;
// Whether or not to run the full test-suite or just a smoke test
bool full_test_;
diff --git a/test/performance/client.cc b/test/performance/client.cc
index 676e88e4..893bb55d 100644
--- a/test/performance/client.cc
+++ b/test/performance/client.cc
@@ -110,9 +110,9 @@ void Client<T,U>::PerformanceTest(Arguments<U> &args, const SetMetric set_sizes)
// Initializes OpenCL and the libraries
auto platform = Platform(args.platform_id);
- auto device = Device(platform, kDeviceType, args.device_id);
+ auto device = Device(platform, args.device_id);
auto context = Context(device);
- auto queue = CommandQueue(context, device);
+ auto queue = Queue(context, device);
if (args.compare_clblas) { clblasSetup(); }
// Iterates over all "num_step" values jumping by "step" each time
@@ -135,17 +135,17 @@ void Client<T,U>::PerformanceTest(Arguments<U> &args, const SetMetric set_sizes)
PopulateVector(c_source);
// Creates the matrices on the device
- auto x_vec = Buffer(context, CL_MEM_READ_WRITE, args.x_size*sizeof(T));
- auto y_vec = Buffer(context, CL_MEM_READ_WRITE, args.y_size*sizeof(T));
- auto a_mat = Buffer(context, CL_MEM_READ_WRITE, args.a_size*sizeof(T));
- auto b_mat = Buffer(context, CL_MEM_READ_WRITE, args.b_size*sizeof(T));
- auto c_mat = Buffer(context, CL_MEM_READ_WRITE, args.c_size*sizeof(T));
- x_vec.WriteBuffer(queue, args.x_size*sizeof(T), x_source);
- y_vec.WriteBuffer(queue, args.y_size*sizeof(T), y_source);
- a_mat.WriteBuffer(queue, args.a_size*sizeof(T), a_source);
- b_mat.WriteBuffer(queue, args.b_size*sizeof(T), b_source);
- c_mat.WriteBuffer(queue, args.c_size*sizeof(T), c_source);
- auto buffers = Buffers{x_vec, y_vec, a_mat, b_mat, c_mat};
+ auto x_vec = Buffer<T>(context, args.x_size);
+ auto y_vec = Buffer<T>(context, args.y_size);
+ auto a_mat = Buffer<T>(context, args.a_size);
+ auto b_mat = Buffer<T>(context, args.b_size);
+ auto c_mat = Buffer<T>(context, args.c_size);
+ x_vec.Write(queue, args.x_size, x_source);
+ y_vec.Write(queue, args.y_size, y_source);
+ a_mat.Write(queue, args.a_size, a_source);
+ b_mat.Write(queue, args.b_size, b_source);
+ c_mat.Write(queue, args.c_size, c_source);
+ auto buffers = Buffers<T>{x_vec, y_vec, a_mat, b_mat, c_mat};
// Runs the routines and collects the timings
auto ms_clblast = TimedExecution(args.num_runs, args, buffers, queue, run_routine_, "CLBlast");
@@ -176,7 +176,7 @@ void Client<T,U>::PerformanceTest(Arguments<U> &args, const SetMetric set_sizes)
// value found in the vector of timing results. The return value is in milliseconds.
template <typename T, typename U>
double Client<T,U>::TimedExecution(const size_t num_runs, const Arguments<U> &args,
- const Buffers &buffers, CommandQueue &queue,
+ const Buffers<T> &buffers, Queue &queue,
Routine run_blas, const std::string &library_name) {
auto timings = std::vector<double>(num_runs);
for (auto &timing: timings) {
diff --git a/test/performance/client.h b/test/performance/client.h
index c9095967..9f6852d0 100644
--- a/test/performance/client.h
+++ b/test/performance/client.h
@@ -38,11 +38,8 @@ template <typename T, typename U>
class Client {
public:
- // Types of devices to consider
- const cl_device_type kDeviceType = CL_DEVICE_TYPE_ALL;
-
// Shorthand for the routine-specific functions passed to the tester
- using Routine = std::function<StatusCode(const Arguments<U>&, const Buffers&, CommandQueue&)>;
+ using Routine = std::function<StatusCode(const Arguments<U>&, const Buffers<T>&, Queue&)>;
using SetMetric = std::function<void(Arguments<U>&)>;
using GetMetric = std::function<size_t(const Arguments<U>&)>;
@@ -63,8 +60,8 @@ class Client {
private:
// Runs a function a given number of times and returns the execution time of the shortest instance
- double TimedExecution(const size_t num_runs, const Arguments<U> &args, const Buffers &buffers,
- CommandQueue &queue, Routine run_blas, const std::string &library_name);
+ double TimedExecution(const size_t num_runs, const Arguments<U> &args, const Buffers<T> &buffers,
+ Queue &queue, Routine run_blas, const std::string &library_name);
// Prints the header of a performance-data table
void PrintTableHeader(const bool silent, const std::vector<std::string> &args);
diff --git a/test/routines/level1/xaxpy.h b/test/routines/level1/xaxpy.h
index 6ce5d7e2..866fb620 100644
--- a/test/routines/level1/xaxpy.h
+++ b/test/routines/level1/xaxpy.h
@@ -57,8 +57,7 @@ class TestXaxpy {
static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Axpy(args.n, args.alpha,
@@ -70,8 +69,7 @@ class TestXaxpy {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXaxpy(args.n, args.alpha,
@@ -83,10 +81,9 @@ class TestXaxpy {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.y_size, static_cast<T>(0));
- buffers.y_vec.ReadBuffer(queue, args.y_size*sizeof(T), result);
+ buffers.y_vec.Read(queue, args.y_size, result);
return result;
}
diff --git a/test/routines/level2/xgemv.h b/test/routines/level2/xgemv.h
index 73f7d76e..056dec30 100644
--- a/test/routines/level2/xgemv.h
+++ b/test/routines/level2/xgemv.h
@@ -68,8 +68,7 @@ class TestXgemv {
static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Gemv(args.layout, args.a_transpose,
@@ -83,8 +82,7 @@ class TestXgemv {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXgemv(static_cast<clblasOrder>(args.layout),
@@ -99,10 +97,9 @@ class TestXgemv {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
std::vector<T> result(args.y_size, static_cast<T>(0));
- buffers.y_vec.ReadBuffer(queue, args.y_size*sizeof(T), result);
+ buffers.y_vec.Read(queue, args.y_size, result);
return result;
}
diff --git a/test/routines/level3/xgemm.h b/test/routines/level3/xgemm.h
index 86a304d1..f06719d6 100644
--- a/test/routines/level3/xgemm.h
+++ b/test/routines/level3/xgemm.h
@@ -70,8 +70,7 @@ class TestXgemm {
static size_t DefaultLDC(const Arguments<T> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Gemm(args.layout, args.a_transpose, args.b_transpose,
@@ -85,8 +84,7 @@ class TestXgemm {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXgemm(static_cast<clblasOrder>(args.layout),
@@ -102,10 +100,9 @@ class TestXgemm {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xhemm.h b/test/routines/level3/xhemm.h
index 75878b06..0c3b9c31 100644
--- a/test/routines/level3/xhemm.h
+++ b/test/routines/level3/xhemm.h
@@ -70,8 +70,7 @@ class TestXhemm {
static size_t DefaultLDC(const Arguments<T> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Hemm(args.layout, args.side, args.triangle,
@@ -85,8 +84,7 @@ class TestXhemm {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXhemm(static_cast<clblasOrder>(args.layout),
@@ -102,10 +100,9 @@ class TestXhemm {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xher2k.h b/test/routines/level3/xher2k.h
index f13e8a62..b20ec973 100644
--- a/test/routines/level3/xher2k.h
+++ b/test/routines/level3/xher2k.h
@@ -68,8 +68,7 @@ class TestXher2k {
static size_t DefaultLDC(const Arguments<U> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<U> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<U> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto alpha2 = T{args.alpha, args.alpha};
@@ -84,8 +83,7 @@ class TestXher2k {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<U> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<U> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto alpha2 = T{args.alpha, args.alpha};
@@ -102,10 +100,9 @@ class TestXher2k {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<U> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xherk.h b/test/routines/level3/xherk.h
index 780b9b52..20c2b4b8 100644
--- a/test/routines/level3/xherk.h
+++ b/test/routines/level3/xherk.h
@@ -61,8 +61,7 @@ class TestXherk {
static size_t DefaultLDC(const Arguments<U> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<U> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<U> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Herk(args.layout, args.triangle, args.a_transpose,
@@ -75,8 +74,7 @@ class TestXherk {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<U> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<U> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXherk(static_cast<clblasOrder>(args.layout),
@@ -91,10 +89,9 @@ class TestXherk {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<U> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xsymm.h b/test/routines/level3/xsymm.h
index 10476349..5b5ad351 100644
--- a/test/routines/level3/xsymm.h
+++ b/test/routines/level3/xsymm.h
@@ -70,8 +70,7 @@ class TestXsymm {
static size_t DefaultLDC(const Arguments<T> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Symm(args.layout, args.side, args.triangle,
@@ -85,8 +84,7 @@ class TestXsymm {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXsymm(static_cast<clblasOrder>(args.layout),
@@ -102,10 +100,9 @@ class TestXsymm {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xsyr2k.h b/test/routines/level3/xsyr2k.h
index f3b1b542..21fcee2a 100644
--- a/test/routines/level3/xsyr2k.h
+++ b/test/routines/level3/xsyr2k.h
@@ -68,8 +68,7 @@ class TestXsyr2k {
static size_t DefaultLDC(const Arguments<T> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Syr2k(args.layout, args.triangle, args.a_transpose,
@@ -83,8 +82,7 @@ class TestXsyr2k {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXsyr2k(static_cast<clblasOrder>(args.layout),
@@ -100,10 +98,9 @@ class TestXsyr2k {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xsyrk.h b/test/routines/level3/xsyrk.h
index 2ec9fb65..c92693c2 100644
--- a/test/routines/level3/xsyrk.h
+++ b/test/routines/level3/xsyrk.h
@@ -61,8 +61,7 @@ class TestXsyrk {
static size_t DefaultLDC(const Arguments<T> &args) { return args.n; }
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Syrk(args.layout, args.triangle, args.a_transpose,
@@ -75,8 +74,7 @@ class TestXsyrk {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXsyrk(static_cast<clblasOrder>(args.layout),
@@ -91,10 +89,9 @@ class TestXsyrk {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+ buffers.c_mat.Read(queue, args.c_size, result);
return result;
}
diff --git a/test/routines/level3/xtrmm.h b/test/routines/level3/xtrmm.h
index 7b7e7af1..d5a52903 100644
--- a/test/routines/level3/xtrmm.h
+++ b/test/routines/level3/xtrmm.h
@@ -61,8 +61,7 @@ class TestXtrmm {
static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine
// Describes how to run the CLBlast routine
- static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Trmm(args.layout, args.side, args.triangle, args.a_transpose, args.diagonal,
@@ -75,8 +74,7 @@ class TestXtrmm {
}
// Describes how to run the clBLAS routine (for correctness/performance comparison)
- static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
- CommandQueue &queue) {
+ static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = clblasXtrmm(static_cast<clblasOrder>(args.layout),
@@ -93,10 +91,9 @@ class TestXtrmm {
}
// Describes how to download the results of the computation (more importantly: which buffer)
- static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
- CommandQueue &queue) {
+ 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));
- buffers.b_mat.ReadBuffer(queue, args.b_size*sizeof(T), result);
+ buffers.b_mat.Read(queue, args.b_size, result);
return result;
}
generated by cgit v1.2.3 (git 2.39.1) at 2024-08-10 01:10:10 +0000