| 1 | #include "pad_reflect_1d.cuh" |
| 2 | |
| 3 | static __global__ __launch_bounds__(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1) void |
| 4 | pad_reflect_1d_kernel_f32( |
| 5 | const void * __restrict__ src0, |
| 6 | void * __restrict__ dst, |
| 7 | const int64_t ne0, |
| 8 | const int64_t ne00, |
| 9 | const uint3 ne01, |
| 10 | const int64_t ne02, |
| 11 | const int64_t ne03, |
| 12 | const int64_t nb00, |
| 13 | const int64_t nb01, |
| 14 | const int64_t nb02, |
| 15 | const int64_t nb03, |
| 16 | const int64_t nb0, |
| 17 | const int64_t nb1, |
| 18 | const int64_t nb2, |
| 19 | const int64_t nb3, |
| 20 | const int p0, |
| 21 | const int p1) { |
| 22 | const int64_t i3 = blockIdx.z; |
| 23 | const int64_t i2 = blockIdx.y; |
| 24 | |
| 25 | const uint2 div_mod_packed = fast_div_modulo(n: blockIdx.x, fastdiv_values: ne01); |
| 26 | const int64_t tile1 = div_mod_packed.y; // i1 |
| 27 | const int64_t tile0 = div_mod_packed.x; // nth i0 tile |
| 28 | const int64_t i1 = tile1; |
| 29 | const int64_t i0 = threadIdx.x + tile0 * blockDim.x; |
| 30 | |
| 31 | // ne01.z is original value of unpacked ne01 (see init_fastdiv_values in common.cuh) |
| 32 | if (i0 >= ne0 || i1 >= ne01.z || i2 >= ne02 || i3 >= ne03) { |
| 33 | return; |
| 34 | } |
| 35 | |
| 36 | const char * src0_ptr = (const char *) src0 + i3 * nb03 + i2 * nb02 + i1 * nb01; |
| 37 | char * dst_ptr = (char *) dst + i3 * nb3 + i2 * nb2 + i1 * nb1; |
| 38 | |
| 39 | const int64_t rel_i0 = i0 - p0; // relative i0 in src0 |
| 40 | int64_t src_idx; |
| 41 | |
| 42 | if (rel_i0 < 0) { |
| 43 | // Left padding - reflect |
| 44 | src_idx = -rel_i0; |
| 45 | } else if (rel_i0 < ne00) { |
| 46 | // Middle - copy |
| 47 | src_idx = rel_i0; |
| 48 | } else { |
| 49 | // Right padding - reflect |
| 50 | src_idx = 2 * ne00 - 2 - rel_i0; |
| 51 | } |
| 52 | const float value = *(const float *) (src0_ptr + src_idx * nb00); |
| 53 | *(float *) (dst_ptr + i0 * nb0) = value; |
| 54 | |
| 55 | GGML_UNUSED(p1); |
| 56 | } |
| 57 | |
| 58 | void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { |
| 59 | const ggml_tensor * src0 = dst->src[0]; |
| 60 | cudaStream_t stream = ctx.stream(); |
| 61 | |
| 62 | GGML_ASSERT(src0->type == GGML_TYPE_F32); |
| 63 | GGML_ASSERT(dst->type == GGML_TYPE_F32); |
| 64 | |
| 65 | const int32_t * opts = (const int32_t *) dst->op_params; |
| 66 | const int p0 = opts[0]; |
| 67 | const int p1 = opts[1]; |
| 68 | |
| 69 | const int64_t ne00 = src0->ne[0]; |
| 70 | const int64_t ne01 = src0->ne[1]; |
| 71 | const uint3 ne01_packed = init_fastdiv_values(d_64: ne01); |
| 72 | const int64_t ne02 = src0->ne[2]; |
| 73 | const int64_t ne03 = src0->ne[3]; |
| 74 | |
| 75 | const int64_t ne0 = dst->ne[0]; |
| 76 | |
| 77 | // sanity: padded length matches |
| 78 | GGML_ASSERT(ne0 == ne00 + p0 + p1); |
| 79 | |
| 80 | constexpr int64_t bx = CUDA_PAD_REFLECT_1D_BLOCK_SIZE; // threads per block (x) |
| 81 | const int64_t tiles0 = (ne0 + bx - 1) / bx; // number of tiles along i0 |
| 82 | // grid.x covers i1 and all tiles of i0: [ne01 * tiles0] |
| 83 | // grid.y covers i2: [ne02] |
| 84 | // grid.z covers i3: [ne03] |
| 85 | const dim3 grid_dims((unsigned) (ne01 * tiles0), (unsigned) ne02, (unsigned) ne03); |
| 86 | const dim3 block_dims((unsigned) bx, 1, 1); |
| 87 | |
| 88 | pad_reflect_1d_kernel_f32<<<gridDim: grid_dims, blockDim: block_dims, sharedMem: 0, stream>>>( |
| 89 | src0: src0->data, dst: dst->data, ne0, ne00, ne01: ne01_packed, ne02, ne03, nb00: src0->nb[0], nb01: src0->nb[1], nb02: src0->nb[2], nb03: src0->nb[3], |
| 90 | nb0: dst->nb[0], nb1: dst->nb[1], nb2: dst->nb[2], nb3: dst->nb[3], p0, p1); |
| 91 | } |
| 92 | |