vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

cuda sqrt support

Browse source
This commit is contained in:
Calvin Laurenson 2024-06-15 17:35:57 -07:00
parent 0c7b3595b9
commit 80321f4b49
No known key found for this signature in database
GPG key ID: 2265749925202DAD
3 changed files with 34 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml-cuda.cu
View file

@ -2267,6 +2267,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_SQR:
ggml_cuda_op_sqr(ctx, dst);
break;
case GGML_OP_SQRT:
ggml_cuda_op_sqrt(ctx, dst);
break;
case GGML_OP_CLAMP:
ggml_cuda_op_clamp(ctx, dst);
break;

28
ggml-cuda/unary.cu
View file

@ -92,6 +92,15 @@ static __global__ void sqr_f32(const float * x, float * dst, const int k) {
dst[i] = x[i] * x[i];
}
static __global__ void sqrt_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = sqrt(x[i]);
}
static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@ -142,6 +151,11 @@ static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t
sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_SQRT_BLOCK_SIZE - 1) / CUDA_SQRT_BLOCK_SIZE;
sqrt_f32<<<num_blocks, CUDA_SQRT_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
@ -284,3 +298,17 @@ void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}

3
ggml-cuda/unary.cuh
View file

@ -8,6 +8,7 @@
#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
#define CUDA_HARDSWISH_BLOCK_SIZE 256
#define CUDA_SQR_BLOCK_SIZE 256
#define CUDA_SQRT_BLOCK_SIZE 256
void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@ -28,3 +29,5 @@ void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst);