mirror/trigdx
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update GPU backend (#29)

Browse Source 
* Update GPU memory management
* Add allocate_memory and free_memory
This commit is contained in:
Bram Veenboer
2025-09-03 09:16:28 +02:00
committed by GitHub
parent 9d3af8c202
commit 8fe8314905
4 changed files with 87 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

84
src/gpu.cpp
View File

@@ -10,79 +10,63 @@
struct GPUBackend::Impl {
~Impl() {
if (h_x) {
cudaFreeHost(h_x);
}
if (h_s) {
cudaFreeHost(h_s);
}
if (h_c) {
cudaFreeHost(h_c);
}
if (d_x) {
cudaFree(d_x);
}
if (d_s) {
cudaFree(d_s);
}
if (d_c) {
cudaFree(d_c);
}
void *allocate_memory(size_t bytes) const {
void *ptr;
cudaMallocHost(&ptr, bytes);
return ptr;
}
void init(size_t n) {
const size_t bytes = n * sizeof(float);
cudaMallocHost(&h_x, bytes);
cudaMallocHost(&h_s, bytes);
cudaMallocHost(&h_c, bytes);
cudaMalloc(&d_x, bytes);
cudaMalloc(&d_s, bytes);
cudaMalloc(&d_c, bytes);
}
void free_memory(void *ptr) const { cudaFreeHost(ptr); }
void compute_sinf(size_t n, const float *x, float *s) const {
const size_t bytes = n * sizeof(float);
std::memcpy(h_x, x, bytes);
cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
float *d_x, *d_s;
cudaMalloc(&d_x, bytes);
cudaMalloc(&d_s, bytes);
cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
launch_sinf_kernel(d_x, d_s, n);
cudaMemcpy(h_s, d_s, bytes, cudaMemcpyDeviceToHost);
std::memcpy(s, h_s, bytes);
cudaMemcpy(s, d_s, bytes, cudaMemcpyDeviceToHost);
cudaFree(d_x);
cudaFree(d_s);
}
void compute_cosf(size_t n, const float *x, float *c) const {
const size_t bytes = n * sizeof(float);
std::memcpy(h_x, x, bytes);
cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
float *d_x, *d_c;
cudaMalloc(&d_x, bytes);
cudaMalloc(&d_c, bytes);
cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
launch_cosf_kernel(d_x, d_c, n);
cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
std::memcpy(c, h_c, bytes);
cudaMemcpy(c, d_c, bytes, cudaMemcpyDeviceToHost);
cudaFree(d_x);
cudaFree(d_c);
}
void compute_sincosf(size_t n, const float *x, float *s, float *c) const {
const size_t bytes = n * sizeof(float);
std::memcpy(h_x, x, bytes);
cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
float *d_x, *d_s, *d_c;
cudaMalloc(&d_x, bytes);
cudaMalloc(&d_s, bytes);
cudaMalloc(&d_c, bytes);
cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
launch_sincosf_kernel(d_x, d_s, d_c, n);
cudaMemcpy(h_s, d_s, bytes, cudaMemcpyDeviceToHost);
cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
std::memcpy(s, h_s, bytes);
std::memcpy(c, h_c, bytes);
cudaMemcpy(s, d_s, bytes, cudaMemcpyDeviceToHost);
cudaMemcpy(c, d_c, bytes, cudaMemcpyDeviceToHost);
cudaFree(d_x);
cudaFree(d_s);
cudaFree(d_c);
}
float *h_x = nullptr;
float *h_s = nullptr;
float *h_c = nullptr;
float *d_x = nullptr;
float *d_s = nullptr;
float *d_c = nullptr;
};
GPUBackend::GPUBackend() : impl(std::make_unique<Impl>()) {}
GPUBackend::~GPUBackend() = default;
void GPUBackend::init(size_t n) { impl->init(n); }
void *GPUBackend::allocate_memory(size_t bytes) const {
return impl->allocate_memory(bytes);
}
void GPUBackend::free_memory(void *ptr) const { impl->free_memory(ptr); }
void GPUBackend::compute_sinf(size_t n, const float *x, float *s) const {
impl->compute_sinf(n, x, s);