mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-04-29 02:41:47 +00:00
0ceeb11721
* Merging mainline - WIP * Merging mainline - WIP AVX2 and CUDA appear to work. CUDA performance seems slightly (~1-2%) lower as it is so often the case with llama.cpp/ggml after some "improvements" have been made. * Merging mainline - fix Metal * Remove check --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
35 lines
1.2 KiB
Plaintext
35 lines
1.2 KiB
Plaintext
#include "arange.cuh"
|
|
|
|
static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
|
|
// blockIDx.x: idx of ne0 / BLOCK_SIZE
|
|
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
|
|
if (nidx >= ne0) {
|
|
return;
|
|
}
|
|
dst[nidx] = start + step * nidx;
|
|
}
|
|
|
|
static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
|
|
int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
|
|
arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start, step);
|
|
}
|
|
|
|
void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
|
|
float * dst_d = (float *)dst->data;
|
|
cudaStream_t stream = ctx.stream();
|
|
|
|
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
|
|
|
float start;
|
|
float stop;
|
|
float step;
|
|
memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
|
|
memcpy(&stop, (float *)dst->op_params + 1, sizeof(float));
|
|
memcpy(&step, (float *)dst->op_params + 2, sizeof(float));
|
|
|
|
int64_t steps = (int64_t)ceil((stop - start) / step);
|
|
GGML_ASSERT(ggml_nelements(dst) == steps);
|
|
|
|
arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
|
|
}
|