1#include "clamp.cuh"
 2
 3static __device__ __forceinline__ float op_clamp(float x, float min, float max) {
 4    return fminf(fmaxf(x, min), max);
 5}
 6
 7template <class T>
 8static __global__ void op_clamp_kernel(const T * x, T * dst, const T min, const T max, const int k) {
 9    const int i = blockDim.x*blockIdx.x + threadIdx.x;
10
11    if (i >= k) {
12        return;
13    }
14
15    dst[i] = (T)op_clamp((float)x[i], (float)min, (float)max);
16}
17
18template <class T>
19static void clamp_cuda(const T * x, T * dst, const T min, const T max, const int k, cudaStream_t stream) {
20    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
21    op_clamp_kernel<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
22}
23
24
25void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
26    const ggml_tensor * src0 = dst->src[0];
27    const void * src0_d = src0->data;
28    void * dst_d = dst->data;
29    cudaStream_t stream = ctx.stream();
30
31    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
32    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
33    GGML_ASSERT(src0->type == dst->type);
34
35    float min;
36    float max;
37    memcpy(&min, dst->op_params, sizeof(float));
38    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
39
40    if (src0->type == GGML_TYPE_F16) {
41        clamp_cuda((const half *)src0_d, (half *)dst_d, (half)min, (half)max, ggml_nelements(src0), stream);
42    } else {
43        clamp_cuda((const float *)src0_d, (float *)dst_d, (float)min, (float)max, ggml_nelements(src0), stream);
44    }
45}