Function ggml_compute_forward_conv_transpose_2d

ggml.c:11843–11941

// ggml_compute_forward_conv_transpose_2d

static void ggml_compute_forward_conv_transpose_2d(
        const struct ggml_compute_params * params,
        const struct ggml_tensor * src0,
        const struct ggml_tensor * src1,
              struct ggml_tensor * dst) {
    GGML_ASSERT(src0->type == GGML_TYPE_F16);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);
    GGML_ASSERT( dst->type == GGML_TYPE_F32);

    int64_t t0 = ggml_perf_time_us();
    UNUSED(t0);

    GGML_TENSOR_BINARY_OP_LOCALS

    const int ith = params->ith;
    const int nth = params->nth;

    const int nk = ne00*ne01*ne02*ne03;

    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
    GGML_ASSERT(nb10 == sizeof(float));

    if (params->type == GGML_TASK_INIT) {
        memset(params->wdata, 0, params->wsize);

        // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout)
        {
            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;

            for (int64_t i03 = 0; i03 < ne03; i03++) {
                for (int64_t i02 = 0; i02 < ne02; i02++) {
                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02);
                    ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03;
                    for (int64_t i01 = 0; i01 < ne01; i01++) {
                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                            dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00];
                        }
                    }
                }
            }
        }

        // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh)
        {
            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk;
            for (int i12 = 0; i12 < ne12; i12++) {
                for (int i11 = 0; i11 < ne11; i11++) {
                    const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
                    ggml_fp16_t * dst_data = wdata + i11*ne10*ne12;
                    for (int i10 = 0; i10 < ne10; i10++) {
                        dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]);
                    }
                }
            }
        }

        memset(dst->data, 0, ggml_nbytes(dst));