LCPQ/qp2
9
1
Fork 0
mirror of https://github.com/QuantumPackage/qp2.git synced 2024-12-22 03:23:29 +01:00
Code Releases Activity

Merge branch 'dev-stable' of github.com:QuantumPackage/qp2 into dev-stable

Browse Source
This commit is contained in:
Anthony Scemama 2024-07-09 22:28:08 +02:00
commit e4ad73a87e
2 changed files with 26 additions and 115 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
plugins/local/gpu_intel/README.rst
View File

@ -4,5 +4,5 @@ gpu_intel
Intel implementation of GPU routines. Uses MKL and SYCL.
```bash
dpcpp -O3 -c gpu.o gpu.sycl
icpx -fsycl gpu.cxx -c -qmkl=sequential
```

139
plugins/local/gpu_intel/gpu.sycl
View File

@ -18,7 +18,7 @@ void gpu_set_device(int32_t igpu) {
/* Allocation functions */
void gpu_allocate(void** ptr, int64_t size) {
auto queue = sycl::queue(sycl::default_selector{});
auto queue = sycl::queue(sycl::default_selector_v);
try {
*ptr = sycl::malloc_shared(size, queue);
@ -31,25 +31,25 @@ void gpu_allocate(void** ptr, int64_t size) {
void gpu_deallocate(void** ptr) {
assert(*ptr != nullptr);
sycl::free(*ptr, sycl::queue(sycl::default_selector{}));
sycl::free(*ptr, sycl::queue(sycl::default_selector_v));
*ptr = nullptr;
}
/* Upload data from host to device */
void gpu_upload(const void* cpu_ptr, void* gpu_ptr, const int64_t n) {
sycl::queue queue(sycl::default_selector{});
sycl::queue queue(sycl::default_selector_v);
queue.memcpy(gpu_ptr, cpu_ptr, n).wait();
}
/* Download data from device to host */
void gpu_download(const void* gpu_ptr, void* cpu_ptr, const int64_t n) {
sycl::queue queue(sycl::default_selector{});
sycl::queue queue(sycl::default_selector_v);
queue.memcpy(cpu_ptr, gpu_ptr, n).wait();
}
/* Copy data from one GPU memory location to another */
void gpu_copy(const void* gpu_ptr_src, void* gpu_ptr_dest, const int64_t n) {
sycl::queue queue(sycl::default_selector{});
sycl::queue queue(sycl::default_selector_v);
queue.memcpy(gpu_ptr_dest, gpu_ptr_src, n).wait();
}
@ -57,7 +57,7 @@ void gpu_copy(const void* gpu_ptr_src, void* gpu_ptr_dest, const int64_t n) {
/* SYCL queue as a replacement for CUDA stream */
void gpu_stream_create(sycl::queue** ptr) {
*ptr = new sycl::queue(sycl::default_selector{});
*ptr = new sycl::queue(sycl::default_selector_v);
}
void gpu_stream_destroy(sycl::queue** ptr) {
@ -66,59 +66,8 @@ void gpu_stream_destroy(sycl::queue** ptr) {
*ptr = nullptr;
}
To translate the CUDA functions related to stream management to SYCL, you will need to adapt to SYCL's approach to command groups and queues. SYCL uses queues to manage execution order and parallelism, similar to CUDA streams but integrated within the SYCL ecosystem.
### Original CUDA Code
```c
/* Create a CUDA stream */
void gpu_stream_create(cudaStream_t* ptr) {
cudaError_t rc = cudaStreamCreate(ptr);
assert(rc == cudaSuccess);
}
/* Destroy a CUDA stream */
void gpu_stream_destroy(cudaStream_t* ptr) {
assert(ptr != NULL);
cudaError_t rc = cudaStreamDestroy(*ptr);
assert(rc == cudaSuccess);
*ptr = NULL;
}
/* Set a specific stream for cuBLAS operations */
void gpu_set_stream(cublasHandle_t handle, cudaStream_t stream) {
cublasSetStream(handle, stream);
}
/* Synchronize all streams */
void gpu_synchronize() {
cudaDeviceSynchronize();
}
```
### Translated SYCL Code
```cpp
#include <CL/sycl.hpp>
#include <cassert>
/* SYCL queue as a replacement for CUDA stream */
void gpu_stream_create(sycl::queue** ptr) {
*ptr = new sycl::queue(sycl::default_selector{});
}
void gpu_stream_destroy(sycl::queue** ptr) {
*ptr->wait_and_throw();
assert(*ptr != nullptr);
delete *ptr;
*ptr = nullptr;
}
/* SYCL does not need an equivalent for setting a stream on a cuBLAS handle,
because each SYCL queue acts independently and can be used directly. */
void gpu_synchronize() {
sycl::queue queue(sycl::default_selector{});
sycl::queue queue(sycl::default_selector_v);
queue.wait_and_throw();
}
@ -132,17 +81,17 @@ void gpu_set_stream(blasHandle_t* handle, sycl::queue* ptr) {
handle->queue = ptr;
}
void gpu_blas_create(blasHandle_t* ptr) {
*ptr = new blasHandle_t;
void gpu_blas_create(blasHandle_t** ptr) {
*ptr = (blasHandle_t*) malloc(sizeof(blasHandle_t));
assert(*ptr != nullptr);
ptr->queue = new sycl::queue(sycl::default_selector{});
assert(ptr->queue != nullptr);
(*ptr)->queue = new sycl::queue(sycl::default_selector_v);
assert((*ptr)->queue != nullptr);
}
void gpu_blas_destroy(blasHandle_t* ptr) {
void gpu_blas_destroy(blasHandle_t** ptr) {
assert(*ptr != nullptr);
delete ptr->queue;
delete *ptr;
delete (*ptr)->queue;
free(*ptr);
*ptr = nullptr;
}
@ -159,21 +108,7 @@ void gpu_ddot(blasHandle_t* handle, const int64_t n, const double* x, const int6
assert(y != nullptr);
assert(result != nullptr);
// SYCL buffer for the result
sycl::buffer<double, 1> result_buf(result, sycl::range<1>(1));
sycl::queue& queue = handle->queue;
// Perform the dot product operation
queue.submit([&](sycl::handler& cgh) {
// Accessors for the buffers
auto result_acc = result_buf.get_access<sycl::access::mode::write>(cgh);
// This is an asynchronous call to compute dot product
cgh.single_task([=]() {
result_acc[0] = oneapi::mkl::blas::dot(cgh, n, x, incx, y, incy);
});
});
oneapi::mkl::blas::dot(*handle->queue, n, x, incx, y, incy, result);
}
@ -194,16 +129,7 @@ void gpu_dgemv(blasHandle_t* handle, const char* transa, const int64_t m, const
}
// Perform DGEMV operation using oneMKL
handle->queue->submit([&](sycl::handler& cgh) {
// Use accessors to ensure data consistency and dependency resolution
auto a_acc = sycl::accessor(a, sycl::range(m * lda), sycl::read_only, cgh);
auto x_acc = sycl::accessor(x, sycl::range(n * incx), sycl::read_only, cgh);
auto y_acc = sycl::accessor(y, sycl::range(m * incy), sycl::read_write, cgh);
cgh.parallel_for(sycl::range(1), [=](sycl::id<1>) {
oneapi::mkl::blas::gemv(*handle->queue, transa_, m, n, *alpha, a_acc.get_pointer(), lda, x_acc.get_pointer(), incx, *beta, y_acc.get_pointer(), incy);
});
});
oneapi::mkl::blas::column_major::gemv(*handle->queue, transa_, m, n, *alpha, a, lda, x, incx, *beta, y, incy);
}
@ -218,23 +144,12 @@ void gpu_dgemm(blasHandle_t* handle, const char* transa, const char* transb, con
auto transa_ = (*transa == 'T' || *transa == 't') ? oneapi::mkl::transpose::trans : oneapi::mkl::transpose::nontrans;
auto transb_ = (*transb == 'T' || *transb == 't') ? oneapi::mkl::transpose::trans : oneapi::mkl::transpose::nontrans;
// Ensure queue is ready
handle->queue->submit([&](sycl::handler& cgh) {
// Accessors for matrices
auto a_acc = sycl::accessor(a, sycl::range<1>(m * lda), sycl::read_only, cgh);
auto b_acc = sycl::accessor(b, sycl::range<1>(k * ldb), sycl::read_only, cgh);
auto c_acc = sycl::accessor(c, sycl::range<1>(m * ldc), sycl::read_write, cgh);
cgh.parallel_for(sycl::range(1), [=](sycl::id<1>) {
oneapi::mkl::blas::gemm(*handle->queue, transa_, transb_, m, n, k,
*alpha, a_acc.get_pointer(), lda,
b_acc.get_pointer(), ldb,
*beta, c_acc.get_pointer(), ldc);
});
});
oneapi::mkl::blas::column_major::gemm(*handle->queue, transa_, transb_, m, n, k,
*alpha, a, lda, b, ldb, *beta, c, ldc);
}
void gpu_dgeam(blasHandle_t* handle, const char* transa, const char* transb, const int64_t m, const int64_t n, const double* alpha,
const double* a, const int64_t lda, const double* beta, const double* b, const int64_t ldb, double* c, const int64_t ldc) {
assert(handle != nullptr && handle->queue != nullptr);
@ -246,18 +161,14 @@ void gpu_dgeam(blasHandle_t* handle, const char* transa, const char* transb, con
bool transB = (*transb == 'T' || *transb == 't');
handle->queue->submit([&](sycl::handler& cgh) {
auto a_acc = sycl::accessor(a, sycl::range(m * lda), sycl::read_only, cgh);
auto b_acc = sycl::accessor(b, sycl::range(n * ldb), sycl::read_only, cgh);
auto c_acc = sycl::accessor(c, sycl::range(m * ldc), sycl::read_write, cgh);
cgh.parallel_for(sycl::range<2>(m, n), [=](sycl::id<2> idx) {
int i = idx[0];
int j = idx[1];
int ai = transA ? j * lda + i : i * lda + j;
int bi = transB ? j * ldb + i : i * ldb + j;
int ci = i * ldc + j;
const int i = idx[0];
const int j = idx[1];
const int ai = transA ? j * lda + i : i * lda + j;
const int bi = transB ? j * ldb + i : i * ldb + j;
const int ci = i * ldc + j;
c_acc[ci] = (*alpha) * a_acc[ai] + (*beta) * b_acc[bi];
c[ci] = (*alpha) * a[ai] + (*beta) * b[bi];
});
});