haraldwolff
/
CLBlast
mirror of https://github.com/CNugteren/CLBlast.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Made buffers of batched routines read/write (was: read-only)

pull/204/head
Cedric Nugteren 2017-10-17 19:56:47 +02:00
parent f349731d54
commit b1270f04b8
3 changed files with 17 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/clpp11.hpp
View File

@ -668,6 +668,9 @@ class Buffer {
// Copies from host to device: writing the device buffer a-synchronously
void WriteAsync(const Queue &queue, const size_t size, const T* host, const size_t offset = 0) {
if (access_ == BufferAccess::kReadOnly) {
throw LogicError("Buffer: writing to a read-only buffer");
}
if (GetSize() < (offset+size)*sizeof(T)) {
throw LogicError("Buffer: target device buffer is too small");
}

6
src/routines/levelx/xaxpybatched.cpp
View File

@ -59,9 +59,9 @@ void XaxpyBatched<T>::DoAxpyBatched(const size_t n, const std::vector<T> &alphas
x_offsets_int[batch] = static_cast<int>(x_offsets[batch]);
y_offsets_int[batch] = static_cast<int>(y_offsets[batch]);
}
auto x_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto y_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto alphas_device = Buffer<T>(context_, BufferAccess::kReadOnly, batch_count);
auto x_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto y_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto alphas_device = Buffer<T>(context_, BufferAccess::kReadWrite, batch_count);
x_offsets_device.Write(queue_, batch_count, x_offsets_int);
y_offsets_device.Write(queue_, batch_count, y_offsets_int);
alphas_device.Write(queue_, batch_count, alphas);

22
src/routines/levelx/xgemmbatched.cpp
View File

@ -100,8 +100,8 @@ void XgemmBatched<T>::DoGemmBatched(const Layout layout, const Transpose a_trans
}
// Upload the scalar arguments to the device
auto alphas_device = Buffer<T>(context_, BufferAccess::kReadOnly, batch_count);
auto betas_device = Buffer<T>(context_, BufferAccess::kReadOnly, batch_count);
auto alphas_device = Buffer<T>(context_, BufferAccess::kReadWrite, batch_count);
auto betas_device = Buffer<T>(context_, BufferAccess::kReadWrite, batch_count);
alphas_device.Write(queue_, batch_count, alphas);
betas_device.Write(queue_, batch_count, betas);
@ -200,8 +200,8 @@ void XgemmBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const
// to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In
// case nothing has to be done, these kernels can be skipped.
if (!a_no_temp) {
auto a_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto a_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto a_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto a_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
a_offsets_device.Write(queue_, batch_count, a_offsets);
a_offsets_i_device.Write(queue_, batch_count, a_offsets_i);
auto eventProcessA = Event();
@ -214,8 +214,8 @@ void XgemmBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const
// As above, but now for matrix B
if (!b_no_temp) {
auto b_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto b_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto b_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto b_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
b_offsets_device.Write(queue_, batch_count, b_offsets);
b_offsets_i_device.Write(queue_, batch_count, b_offsets_i);
auto eventProcessB = Event();
@ -227,8 +227,8 @@ void XgemmBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const
}
// As above, but now for matrix C
auto c_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto c_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto c_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto c_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
if (!c_no_temp) {
c_offsets_device.Write(queue_, batch_count, c_offsets);
c_offsets_i_device.Write(queue_, batch_count, c_offsets_i);
@ -297,9 +297,9 @@ void XgemmBatched<T>::BatchedGemmDirect(const size_t m, const size_t n, const si
const size_t batch_count) {
// Uploads the offsets to the device
auto a_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto b_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto c_offsets_device = Buffer<int>(context_, BufferAccess::kReadOnly, batch_count);
auto a_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto b_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
auto c_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
a_offsets_device.Write(queue_, batch_count, a_offsets);
b_offsets_device.Write(queue_, batch_count, b_offsets);
c_offsets_device.Write(queue_, batch_count, c_offsets);