// SPDX-License-Identifier: Apache-2.0
// STEM_POOL compute (meteor-p150 OPT round 3 R3.2, tt/pool.py): the vertical half of the 3x3 / s2 max pool, per tile
// of a unit's three input rows (CBs 0 / 1 / 2, WT tiles each): v = max(max(r0, r1), r2) with binary_max_tile, the
// fp32 SFPU max of ttnn.maximum (binary_ng), operands unpacked to DEST as fp32 (UnpackToDestFp32 on CBs 0..2), packed
// fp32 to CB 16. The writer does the horizontal half. Max is exact; the order of the operands does not matter.
// CT args: [0] NU, [1] NC, [2] GX, [3] WT.
#include <cstdint>

#include "api/compute/binary_max_min.h"
#include "api/compute/common.h"
#include "api/compute/compute_kernel_api.h"
#include "api/compute/tile_move_copy.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 WT = get_compile_time_arg_val(3);
constexpr uint32_t cb_v = 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 nu = B + (core < E ? 1 : 0);
    if (nu == 0) {
        return;
    }
    compute_kernel_hw_startup(0, cb_v);
    for (uint32_t u = 0; u < nu; ++u) {
        cb_wait_front(0, WT);
        cb_wait_front(1, WT);
        cb_wait_front(2, WT);
        for (uint32_t tx = 0; tx < WT; ++tx) {
            tile_regs_acquire();
            copy_init(0);
            copy_tile(0, tx, 0);
            copy_init(1);
            copy_tile(1, tx, 1);
            binary_max_tile_init();
            binary_max_tile(0, 1, 0);
            copy_init(2);
            copy_tile(2, tx, 1);
            binary_max_tile_init();
            binary_max_tile(0, 1, 0);
            tile_regs_commit();
            cb_reserve_back(cb_v, 1);
            tile_regs_wait();
            pack_tile(0, cb_v);
            tile_regs_release();
            cb_push_back(cb_v, 1);
        }
        cb_pop_front(0, WT);
        cb_pop_front(1, WT);
        cb_pop_front(2, WT);
    }
}