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

ggml: CUDA unary op EXP

Browse source 
Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
This commit is contained in:
Molly Sophia 2024-09-05 18:18:51 +08:00
parent 78203641fe
commit 0b174abc3d
3 changed files with 35 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml/src/ggml-cuda.cu
View file

@ -2225,6 +2225,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_UNARY_OP_HARDSWISH:
ggml_cuda_op_hardswish(ctx, dst);
break;
case GGML_UNARY_OP_EXP:
ggml_cuda_op_exp(ctx, dst);
break;
default:
return false;
}
@ -2769,6 +2772,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_TANH:
case GGML_UNARY_OP_EXP:
return ggml_is_contiguous(op->src[0]);
default:
return false;

28
ggml/src/ggml-cuda/unary.cu
View file

@ -85,6 +85,15 @@ static __global__ void hardswish_f32(const float * x, float * dst, const int k)
dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
}
static __global__ void exp_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = expf(x[i]);
}
static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
@ -174,6 +183,11 @@ static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaSt
hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void exp_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_EXP_BLOCK_SIZE - 1) / CUDA_EXP_BLOCK_SIZE;
exp_f32<<<num_blocks, CUDA_EXP_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
@ -325,6 +339,20 @@ void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_exp(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);
exp_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;

3
ggml/src/ggml-cuda/unary.cuh
View file

@ -7,6 +7,7 @@
#define CUDA_RELU_BLOCK_SIZE 256
#define CUDA_SIGMOID_BLOCK_SIZE 256
#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
#define CUDA_EXP_BLOCK_SIZE 256
#define CUDA_HARDSWISH_BLOCK_SIZE 256
#define CUDA_SQR_BLOCK_SIZE 256
#define CUDA_SQRT_BLOCK_SIZE 256
@ -29,6 +30,8 @@ void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
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);