| 1 | #include "diagmask.cuh" |
| 2 | |
| 3 | static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) { |
| 4 | const int col = blockDim.y*blockIdx.y + threadIdx.y; |
| 5 | const int row = blockDim.x*blockIdx.x + threadIdx.x; |
| 6 | |
| 7 | if (col >= ncols) { |
| 8 | return; |
| 9 | } |
| 10 | |
| 11 | const int i = row*ncols + col; |
| 12 | //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i]; |
| 13 | //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU |
| 14 | dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; |
| 15 | } |
| 16 | |
| 17 | static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) { |
| 18 | const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1); |
| 19 | const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE; |
| 20 | const dim3 block_nums(nrows_x, block_num_x, 1); |
| 21 | diag_mask_inf_f32<<<gridDim: block_nums, blockDim: block_dims, sharedMem: 0, stream>>>(x, dst, ncols: ncols_x, rows_per_channel, n_past); |
| 22 | } |
| 23 | |
| 24 | void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { |
| 25 | const ggml_tensor * src0 = dst->src[0]; |
| 26 | const float * src0_d = (const float *)src0->data; |
| 27 | float * dst_d = (float *)dst->data; |
| 28 | cudaStream_t stream = ctx.stream(); |
| 29 | |
| 30 | GGML_ASSERT(src0->type == GGML_TYPE_F32); |
| 31 | GGML_ASSERT( dst->type == GGML_TYPE_F32); |
| 32 | |
| 33 | const int64_t ne00 = src0->ne[0]; |
| 34 | const int64_t ne01 = src0->ne[1]; |
| 35 | const int nrows0 = ggml_nrows(src0); |
| 36 | |
| 37 | const int n_past = ((int32_t *) dst->op_params)[0]; |
| 38 | |
| 39 | diag_mask_inf_f32_cuda(x: src0_d, dst: dst_d, ncols_x: ne00, nrows_x: nrows0, rows_per_channel: ne01, n_past, stream); |
| 40 | } |
| 41 | |