// Copyright (2019-2023) Paul Scherrer Institute
// SPDX-License-Identifier: GPL-3.0-or-later

#include "hls_jfjoch.h"

void transfer_hbm(ap_uint<256> *d_hbm_p0,
                  ap_uint<256> *uram,
                  uint32_t hbm_offset,
                  bool hbm_to_uram,
                  uint8_t stride,
                  uint8_t uram_offset) {
#pragma HLS INTERFACE m_axi port=d_hbm_p0 bundle=d_hbm_p0 depth=512 offset=off \
        max_read_burst_length=16  max_write_burst_length=16 latency=120 num_write_outstanding=8 num_read_outstanding=9
#pragma HLS INTERFACE m_axi port=uram bundle=uram depth=512 offset=off \
        max_read_burst_length=16  max_write_burst_length=16 latency=8 num_write_outstanding=2 num_read_outstanding=2
#pragma HLS interface s_axilite port=hbm_offset
#pragma HLS interface s_axilite port=hbm_to_uram
#pragma HLS interface s_axilite port=stride
#pragma HLS interface s_axilite port=uram_offset
#pragma HLS INTERFACE s_axilite port=return

    if ((stride != 1) && (stride != 2) && (stride != 4))
        return;

    if (uram_offset >= stride)
        return;

    if (stride == 2) {
        if (hbm_to_uram) {
            for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 32 / 2; i++) {
#pragma HLS PIPELINE II=1
                uram[2 * i + uram_offset] = d_hbm_p0[hbm_offset + i];
            }
        } else {
            for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 32 / 2; i++) {
#pragma HLS PIPELINE II=1
                d_hbm_p0[hbm_offset + i] = uram[2 * i + uram_offset];
            }
        }
    } else if (stride == 4) {
        if (hbm_to_uram) {
            for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 32 / 4; i++) {
#pragma HLS PIPELINE II=1
                uram[4 * i + uram_offset] = d_hbm_p0[hbm_offset + i];
            }
        } else {
            for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 32 / 4; i++) {
#pragma HLS PIPELINE II=1
                d_hbm_p0[hbm_offset + i] = uram[4 * i + uram_offset];
            }
        }
    } else {
        if (hbm_to_uram) {
            for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 32; i++) {
#pragma HLS PIPELINE II=1
                uram[i + uram_offset] = d_hbm_p0[hbm_offset + i];
            }
        } else {
            for (int i = 0; i < RAW_MODULE_SIZE * sizeof(uint16_t) / 32; i++) {
#pragma HLS PIPELINE II=1
                d_hbm_p0[hbm_offset + i] = uram[i + uram_offset];
            }
        }
    }
}