CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-17 11:37:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

cuda: unary ops as float + de-duplicate (ggml/1130)

Browse Source
This commit is contained in:
cmdr2
2025-03-03 20:51:31 +05:30
committed by Georgi Gerganov
parent 3d1cf3cf33
commit b64d7cc272
2 changed files with 210 additions and 645 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-cuda/clamp.cu
View File

@@ -1,20 +1,24 @@
#include "clamp.cuh"
static __device__ __forceinline__ float op_clamp(float x, float min, float max) {
return fminf(fmaxf(x, min), max);
}
template <class T>
static __global__ void op_clamp(const T * x, T * dst, const T min, const T max, const int k) {
static __global__ void op_clamp_kernel(const T * x, T * dst, const T min, const T max, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
dst[i] = (T)op_clamp((float)x[i], (float)min, (float)max);
}
template <class T>
static void clamp_cuda(const T * x, T * dst, const T min, const T max, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
op_clamp<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
op_clamp_kernel<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
}