| 1 | #include "conv-transpose-1d.cuh" |
| 2 | |
| 3 | static __global__ void conv_transpose_1d_kernel( |
| 4 | const int s0, const int p0, const int d0, const int output_size, |
| 5 | const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3, |
| 6 | const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3, |
| 7 | const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3, |
| 8 | const float * src0, const float * src1, float * dst) { |
| 9 | int global_index = threadIdx.x + blockIdx.x * blockDim.x; |
| 10 | if (global_index >= output_size) { |
| 11 | return; |
| 12 | } |
| 13 | |
| 14 | int out_index = global_index / dst_ne0; |
| 15 | |
| 16 | float accumulator = 0; |
| 17 | |
| 18 | for (int c = 0; c < src0_ne2; c++) { |
| 19 | int idx = global_index % dst_ne0; |
| 20 | |
| 21 | int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0); |
| 22 | int input_offset = src1_ne0 * c; |
| 23 | |
| 24 | for (int i = 0; i < src1_ne0; i++) { |
| 25 | if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) { |
| 26 | continue; |
| 27 | } |
| 28 | int weight_idx = idx - i*s0; |
| 29 | |
| 30 | float kernel_weight = src0[kernel_offset + weight_idx]; |
| 31 | float input_value = src1[input_offset+i]; |
| 32 | |
| 33 | accumulator += kernel_weight * input_value; |
| 34 | } |
| 35 | } |
| 36 | dst[global_index] = accumulator; |
| 37 | GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2); |
| 38 | } |
| 39 | |
| 40 | static void conv_transpose_1d_f32_f32_cuda( |
| 41 | const int s0, const int p0, const int d0, const int output_size, |
| 42 | const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3, |
| 43 | const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3, |
| 44 | const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3, |
| 45 | const float * src0, const float * src1, float * dst, |
| 46 | cudaStream_t stream) { |
| 47 | |
| 48 | const int num_blocks = (output_size + CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE; |
| 49 | conv_transpose_1d_kernel<<<gridDim: num_blocks,CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE, sharedMem: 0, stream>>>( |
| 50 | s0,p0,d0,output_size, |
| 51 | src0_ne0, src0_ne1, src0_ne2, src0_ne3, |
| 52 | src1_ne0, src1_ne1, src1_ne2, src1_ne3, |
| 53 | dst_ne0, dst_ne1, dst_ne2, dst_ne3, |
| 54 | src0,src1, dst); |
| 55 | } |
| 56 | |
| 57 | void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { |
| 58 | const ggml_tensor * src0 = dst->src[0]; |
| 59 | const float * src0_d = (const float *)src0->data; |
| 60 | |
| 61 | const ggml_tensor * src1 = dst->src[1]; |
| 62 | const float * src1_d = (const float *)src1->data; |
| 63 | |
| 64 | float * dst_d = (float *)dst->data; |
| 65 | cudaStream_t stream = ctx.stream(); |
| 66 | |
| 67 | GGML_ASSERT(src0->type == GGML_TYPE_F32); |
| 68 | GGML_ASSERT( dst->type == GGML_TYPE_F32); |
| 69 | |
| 70 | GGML_ASSERT(ggml_is_contiguous(src0)); |
| 71 | GGML_ASSERT(ggml_is_contiguous(src1)); |
| 72 | |
| 73 | const int32_t * opts = (const int32_t *)dst->op_params; |
| 74 | |
| 75 | const int s0 = opts[0]; |
| 76 | const int p0 = 0;//opts[3]; |
| 77 | const int d0 = 1;//opts[4]; |
| 78 | |
| 79 | const int64_t output_size = ggml_nelements(dst); |
| 80 | |
| 81 | conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size, |
| 82 | src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], |
| 83 | src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], |
| 84 | dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], |
| 85 | src0_d, src1_d, dst_d, stream); |
| 86 | } |
| 87 | |