// SPDX-License-Identifier: Apache-2.0
// FOUT compute (meteor-p150 OPT2 MK-FO, tt/fout.py): per unit fused_k = other_k + t_k for the 3 tile columns, one FPU
// eltwise add per tile (SrcA = the other operand (raw / r4), SrcB = the tilized t), bf16 in / out, as ttnn.add's
// binary_ng FPU kernel (no fp32 DEST accumulation); work split as the reader.
// CT args: [0] NU, [1] NC, [2] GX.
#include <cstdint>

#include "api/compute/common.h"
#include "api/compute/compute_kernel_api.h"
#include "api/compute/compute_kernel_hw_startup.h"
#include "api/compute/eltwise_binary.h"

using namespace ckernel;

constexpr uint32_t NU = get_compile_time_arg_val(0);
constexpr uint32_t NC = get_compile_time_arg_val(1);
constexpr uint32_t GX = get_compile_time_arg_val(2);
constexpr uint32_t cb_a = 0, cb_b = 2, cb_out = 16;

void kernel_main() {
    const uint32_t core = get_absolute_logical_y() * GX + get_absolute_logical_x();
    if (core >= NC) {
        return;
    }
    constexpr uint32_t B = NU / NC, E = NU % NC;
    const uint32_t n_units = B + (core < E ? 1 : 0);
    if (n_units == 0) {
        return;
    }
    compute_kernel_hw_startup(cb_b, cb_a, cb_out);
    add_init(cb_b, cb_a);
    for (uint32_t u = 0; u < n_units; ++u) {
        cb_wait_front(cb_b, 3);
        cb_wait_front(cb_a, 3);
        cb_reserve_back(cb_out, 3);
        tile_regs_acquire();
        for (uint32_t k = 0; k < 3; ++k) {
            add_tiles(cb_b, cb_a, k, k, k);
        }
        tile_regs_commit();
        tile_regs_wait();
        for (uint32_t k = 0; k < 3; ++k) {
            pack_tile(k, cb_out);
        }
        tile_regs_release();
        cb_push_back(cb_out, 3);
        cb_pop_front(cb_a, 3);
        cb_pop_front(cb_b, 3);
    }
}