Skip to content

cuda : support contiguous_rows unary kernels #28821

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
BensonChen2000 wants to merge 2 commits into from
+48 −20
Closed

cuda : support contiguous_rows unary kernels #28821

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
95d725d
cuda : support contiguous_rows unary kernels
BensonChen2000 Sep 12, 2026
c589083
Merge branch 'ggml-org:master' into non_contiguous_unary_cuda
BensonChen2000 Sep 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 1 addition & 3 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5091,9 +5091,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_UNARY_OP_CEIL:
case GGML_UNARY_OP_ROUND:
case GGML_UNARY_OP_TRUNC:
// TODO: should become:
//return ggml_is_contiguous_rows(op->src[0]);
return ggml_is_contiguous(op->src[0]);
return ggml_is_contiguous_rows(op->src[0]);
default:
return false;
}
Expand Down
64 changes: 47 additions & 17 deletions ggml/src/ggml-cuda/unary.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,23 +115,46 @@ static __device__ __forceinline__ float op_trunc(float x) {
}

template <float (*op)(float), typename T>
static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
static __global__ void unary_op_kernel(const T * x, T * dst,
const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
const size_t nb0, const size_t nb1, const size_t nb2, const size_t nb3) {
ggml_cuda_pdl_lc();
const int i = blockDim.x*blockIdx.x + threadIdx.x;
const int64_t k = ne00 * ne01 * ne02 * ne03;
const int64_t i = blockDim.x*(int64_t)blockIdx.x + threadIdx.x;

if (i >= k) {
return;
}

const int64_t ne_mat = ne00 * ne01;
const int64_t ne_batch = ne_mat * ne02;

// logical index
const int64_t i03 = i / ne_batch;
const int64_t i02 = (i - i03 * ne_batch) / ne_mat;
const int64_t i01 = (i - i03 * ne_batch - i02 * ne_mat) / ne00;
const int64_t i00 = i - i03 * ne_batch - i02 * ne_mat - i01 * ne00;

// skip padding
const T * src_ptr = (const T *)((const char *)x + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
T * dst_ptr = (T *) ((char *)dst + i03*nb3 + i02*nb2 + i01*nb1 + i00*nb0);

ggml_cuda_pdl_sync();
dst[i] = (T)op((float)x[i]);
*dst_ptr = (T)op((float)(*src_ptr));
}

template <float (*op)(float), typename T>
static void unary_cuda(const T * x, T * dst, const int k, cudaStream_t stream) {
static void unary_cuda(const T * x, T * dst,
const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
const size_t nb0, const size_t nb1, const size_t nb2, const size_t nb3,
cudaStream_t stream) {
const int k = ne00 * ne01 * ne02 * ne03;
const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream);
ggml_cuda_kernel_launch(unary_op_kernel<op, T>, launch_params, x, dst, k);
ggml_cuda_kernel_launch(unary_op_kernel<op, T>, launch_params, x, dst,
ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, nb0, nb1, nb2, nb3);
}

template <float (*op)(float)>
Expand All @@ -141,16 +164,20 @@ void ggml_cuda_op_unary(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
void * dst_d = dst->data;
cudaStream_t stream = ctx.stream();

GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous_rows(src0));

GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == dst->type);

GGML_TENSOR_UNARY_OP_LOCALS;

if (src0->type == GGML_TYPE_F16) {
unary_cuda<op>((const half *)src0_d, (half *)dst_d, ggml_nelements(src0), stream);
unary_cuda<op>((const half *)src0_d, (half *)dst_d,
ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, nb0, nb1, nb2, nb3, stream);
} else {
unary_cuda<op>((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), stream);
unary_cuda<op>((const float *)src0_d, (float *)dst_d,
ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, nb0, nb1, nb2, nb3, stream);
}
}

Expand Down Expand Up @@ -704,18 +731,21 @@ void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary

/* fused relu + sqr */

void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node) {
const ggml_tensor * src = relu_node->src[0];
void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * dst) {
const ggml_tensor * src0 = relu_node->src[0];
cudaStream_t stream = ctx.stream();

GGML_ASSERT(ggml_is_contiguous(src));
GGML_ASSERT(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
GGML_ASSERT(src->type == sqr_node->type);
GGML_ASSERT(ggml_is_contiguous_rows(src0));
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == dst->type);

const int k = ggml_nelements(src);
if (src->type == GGML_TYPE_F16) {
unary_cuda<op_relu_sqr>((const half *)src->data, (half *)sqr_node->data, k, stream);
GGML_TENSOR_UNARY_OP_LOCALS;

if (src0->type == GGML_TYPE_F16) {
unary_cuda<op_relu_sqr>((const half *)src0->data, (half *)dst->data,
ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, nb0, nb1, nb2, nb3, stream);
} else {
unary_cuda<op_relu_sqr>((const float *)src->data, (float *)sqr_node->data, k, stream);
unary_cuda<op_relu_sqr>((const float *)src0->data, (float *)dst->data,
ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, nb0, nb1, nb2, nb3, stream);
}
}