// SPDX-FileCopyrightText: © 2026 Tenstorrent Inc.
// SPDX-License-Identifier: Apache-2.0
// Round 11: cores a cmm program leaves idle copy the first K blocks of a LATER xmm's packed weight from DRAM into the
// persistent L1 prefetch buffers of that xmm's weight senders (tt_kernels.XmmPrefetch): their own buffer directly, other
// cores' through a local scratch CB 22 (DRAM -> scratch -> NoC write, two scratch halves so a write overlaps the next
// read).  No semaphores: the consumer is a later program.
// runtime: w_addr, n, then n x (bank, bank offset, dst x, dst y, dst addr, bytes, local)  (fixed length).
#include <stdint.h>
#include "api/dataflow/dataflow_api.h"

void kernel_main() {
    const uint32_t w_addr = get_arg_val<uint32_t>(0);
    const uint32_t n = get_arg_val<uint32_t>(1);
    if (n == 0) {
        return;
    }
    const uint32_t scr = get_write_ptr(22);
    const uint32_t half = get_local_cb_interface(22).fifo_size / 2;
    // local entries first: plain reads into this core's own buffer
    for (uint32_t i = 0; i < n; ++i) {
        const uint32_t a = 2 + 7 * i;
        if (get_arg_val<uint32_t>(a + 6)) {
            noc_async_read(get_noc_addr_from_bank_id<true>(get_arg_val<uint32_t>(a), w_addr + get_arg_val<uint32_t>(a + 1)),
                           get_arg_val<uint32_t>(a + 4), get_arg_val<uint32_t>(a + 5));
        }
    }
    uint32_t k = 0;
    for (uint32_t i = 0; i < n; ++i) {
        const uint32_t a = 2 + 7 * i;
        if (get_arg_val<uint32_t>(a + 6)) {
            continue;
        }
        const uint32_t buf = scr + (k & 1) * half;
        const uint32_t bytes = get_arg_val<uint32_t>(a + 5);
        // this half was last the source of the write two remote entries ago: every issued write has left L1 (flushed)
        noc_async_writes_flushed();
        noc_async_read(get_noc_addr_from_bank_id<true>(get_arg_val<uint32_t>(a), w_addr + get_arg_val<uint32_t>(a + 1)),
                       buf, bytes);
        noc_async_read_barrier();
        noc_async_write(buf, get_noc_addr(get_arg_val<uint32_t>(a + 2), get_arg_val<uint32_t>(a + 3), get_arg_val<uint32_t>(a + 4)),
                        bytes);
        ++k;
    }
    noc_async_read_barrier();
    noc_async_write_barrier();
}