Ggml/cuda col2im 1d (#24417)

* cuda: add GGML_OP_COL2IM_1D, follow-up to the CPU op

* cuda: col2im_1d use fast_div_modulo for the index decomposition

* cuda: col2im_1d tighten supports_op, type match and contiguous dst
This commit is contained in:
Pascal
2026-06-18 22:23:01 +02:00
committed by GitHub
parent 40f3aafc45
commit 3a3edc9ac6
3 changed files with 96 additions and 0 deletions
+12
View File
@@ -11,6 +11,7 @@
#include "ggml-cuda/argsort.cuh"
#include "ggml-cuda/binbcast.cuh"
#include "ggml-cuda/clamp.cuh"
#include "ggml-cuda/col2im-1d.cuh"
#include "ggml-cuda/concat.cuh"
#include "ggml-cuda/conv-transpose-1d.cuh"
#include "ggml-cuda/conv2d.cuh"
@@ -3051,6 +3052,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_CONV_TRANSPOSE_1D:
ggml_cuda_op_conv_transpose_1d(ctx,dst);
break;
case GGML_OP_COL2IM_1D:
ggml_cuda_op_col2im_1d(ctx, dst);
break;
case GGML_OP_POOL_2D:
ggml_cuda_op_pool2d(ctx, dst);
break;
@@ -5316,6 +5320,14 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
}
return false;
} break;
case GGML_OP_COL2IM_1D:
{
ggml_type src0_type = op->src[0]->type;
return (src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_F16 || src0_type == GGML_TYPE_BF16) &&
op->type == src0_type &&
ggml_is_contiguous(op->src[0]) &&
ggml_is_contiguous(op);
} break;
case GGML_OP_SILU_BACK:
return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
break;