// SPDX-License-Identifier: Apache-2.0
// RSZ compute (frnet-p150 OPT round 2, tt/resize.py): per unit (ct, y)
//   V_k = x0_k * A0 + x1_k * A1      (the H-direction lerp of input tile k, fp32 SFPU, k < NWI) -> CB 5 (fp32)
//   out_wt = g0_wt * B0_wt + g1_wt * B1_wt   (the W-direction lerp; g0 / g1: the V rows of each output pixel's
//                                             two input columns, gathered by the writer) -> bf16 (typecast) -> CB 16
// fp32 DEST, every operand unpacked to DEST as fp32 (no TF32 rounding). CT args: [0] U, [1] NC, [2] GX, [3] NWI,
// [4] NWT (work split as the reader).
#include <cstdint>

#include "api/compute/common.h"
#include "api/compute/compute_kernel_api.h"
#include "api/compute/eltwise_binary_sfpu.h"
#include "api/compute/eltwise_unary/eltwise_unary.h"
#include "api/compute/eltwise_unary/typecast.h"
#include "api/compute/reconfig_data_format.h"
#include "api/compute/tile_move_copy.h"

using namespace ckernel;

constexpr uint32_t U = 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 NWI = get_compile_time_arg_val(3);
constexpr uint32_t NWT = get_compile_time_arg_val(4);
constexpr uint32_t cb_x0 = 0, cb_x1 = 1, cb_a = 2, cb_b = 3, cb_v = 5, cb_g0 = 6, cb_g1 = 7, cb_out = 16;

// D[d] = cb[i] * w[j]; then D[0] += D[d]  (d = 1 for the second product)
FORCE_INLINE void lerp(uint32_t ca, uint32_t ia, uint32_t cw0, uint32_t iw0, uint32_t cb, uint32_t ib, uint32_t cw1,
                       uint32_t iw1) {
    copy_init(ca);
    copy_tile(ca, ia, 0);
    copy_init(cw0);
    copy_tile(cw0, iw0, 1);
    mul_binary_tile_init();
    mul_binary_tile(0, 1, 0);
    copy_init(cb);
    copy_tile(cb, ib, 1);
    copy_init(cw1);
    copy_tile(cw1, iw1, 2);
    mul_binary_tile_init();
    mul_binary_tile(1, 2, 1);
    add_binary_tile_init();
    add_binary_tile<ckernel::DstRoundingMode::NearestEven>(0, 1, 0);
}

void kernel_main() {
    const uint32_t core = get_absolute_logical_y() * GX + get_absolute_logical_x();
    if (core >= NC) {
        return;
    }
    constexpr uint32_t B = U / NC, E = U % NC;
    const uint32_t n = B + (core < E ? 1 : 0);
    if (n == 0) {
        return;
    }
    compute_kernel_hw_startup(cb_x0, cb_v);
    cb_wait_front(cb_b, 2 * NWT);
    for (uint32_t u = 0; u < n; ++u) {
        cb_wait_front(cb_a, 2);
        pack_reconfig_data_format(cb_v);
        for (uint32_t k = 0; k < NWI; ++k) {
            cb_wait_front(cb_x0, 1);
            cb_wait_front(cb_x1, 1);
            tile_regs_acquire();
            lerp(cb_x0, 0, cb_a, 0, cb_x1, 0, cb_a, 1);
            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(cb_x0, 1);
            cb_pop_front(cb_x1, 1);
        }
        cb_pop_front(cb_a, 2);
        pack_reconfig_data_format(cb_out);
        for (uint32_t wt = 0; wt < NWT; ++wt) {
            cb_wait_front(cb_g0, 1);
            cb_wait_front(cb_g1, 1);
            tile_regs_acquire();
#ifndef RSZ_SKIP_MATH
            lerp(cb_g0, 0, cb_b, 2 * wt, cb_g1, 0, cb_b, 2 * wt + 1);
#else
            copy_init(cb_g0);
            copy_tile(cb_g0, 0, 0);
#endif
            typecast_tile_init<(uint32_t)DataFormat::Float32, (uint32_t)DataFormat::Float16_b>();
            typecast_tile<(uint32_t)DataFormat::Float32, (uint32_t)DataFormat::Float16_b>(0);
            tile_regs_commit();
            cb_reserve_back(cb_out, 1);
            tile_regs_wait();
            pack_tile(0, cb_out);
            tile_regs_release();
            cb_push_back(cb_out, 1);
            cb_pop_front(cb_g0, 1);
            cb_pop_front(cb_g1, 1);
        }
    }
    cb_pop_front(cb_b, 2 * NWT);
}