CUDA: support F32 kernel type for CONV_TRANSPOSE_2D
#17094
Conversation
…types and improve parameter handling - Introduced a `conv2d_transpose_params` struct for better parameter management. - Updated `conv2d_transpose_kernel` to be templated for different kernel types (float and half). - Modified `ggml_cuda_conv_2d_transpose_p0` to handle both F16 and F32 kernel types. - Enhanced test cases to validate functionality for both kernel types.
…ernel types - Updated `test_conv_transpose_2d` structure to improve parameter handling by reordering constructor arguments. - Enhanced test case generation to iterate over kernel types, allowing for flexible testing of different configurations. - Removed hardcoded kernel type instances in favor of a loop for better maintainability and scalability.
github-actions
bot
added
testing
There was a problem hiding this comment.
Does this PR make a difference to something? From what I understand, the kernel value is upcast into float before doing any accumulation (and accumulation is anyway in f32). So unless there are kernels around which don't fit into f16 I don't see a benefit to supporting this, especially when we don't support the f16 inputs yet (which incidentally might be more relevant than kernels being f32 as we could potentially do half2 multiplications)
| input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride, | ||
| channels_in, channels_out, batches); | ||
| if (kernel->type == GGML_TYPE_F16) { | ||
| conv2d_transpose_cuda_f16(input_data, (const half *) kernel_data, output_data, params, st); |
There was a problem hiding this comment.
you don't need separate cuda_f16 and cuda_f32 functions here, you can straight away dispatch here to conv2d_transpose_cuda<type> and remove those two functions
| const int total; | ||
| }; | ||
|
|
||
| template <typename T> |
There was a problem hiding this comment.
probably a better name for this should be kernel_t
2 participants
also updated test case in
test-backend-ops.But since F32 kernel type is not supported on CPU, only
GGML_TYPE_F16is kept andGGML_TYPE_F32can be uncommented back in the future.