// SPDX-License-Identifier: Apache-2.0
// TGATE compute (meteor-p150 OPT2 round 3 R3.2, tt/tgate.py): the temporal fuse's gate glue in one pass per 32-row
// block. The merged 1x1 conv writes t = W_t x (96 columns, CB 0) and d_j = l_j - l_0 (j = 1..3, columns 0..2 of the
// fourth tile, CB 1; its columns 3..31 are the zero padding of the conv output):
//   E = exp(d)                          SFPU, fp32 DEST (exp(0) = 1 in the padding columns)
//   S = E @ ONES                        FPU matmul with the ones tile (rows 0..2 = 1): every column = sum_j exp(d_j)
//   G = SCALE / (1 + S)                 SFPU (= SCALE * softmax(l)[0], identical in every column of a row)
//   out_k = bf16_rne(relu(t_k * G + b_k))   fp32 SFPU mul / add, relu, RNE typecast (k = 0..2)
// Every CB reaches DEST through SrcA / SrcB (TF32; tt/tgate.py: fp32 unpack-to-DEST before the matmul broke it).
// CBs: 0 t (3 tiles), 1 d, 2 ones, 3 bias (3 tiles, every row the bias), 4 E (loop-back), 5 G (loop-back), 16 out.
// CT args: [0] NB, [1] NC, [2] GX, [3] SCALE (fp32 bits).
#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_sfpu.h"
#include "api/compute/eltwise_unary/binop_with_scalar.h"
#include "api/compute/eltwise_unary/eltwise_unary.h"
#include "api/compute/eltwise_unary/exp.h"
#include "api/compute/eltwise_unary/recip.h"
#include "api/compute/eltwise_unary/relu.h"
#include "api/compute/eltwise_unary/typecast.h"
#include "api/compute/matmul.h"
#include "api/compute/tile_move_copy.h"

using namespace ckernel;

constexpr uint32_t NB = 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 SCALE = get_compile_time_arg_val(3);
constexpr uint32_t ONE = 0x3F800000u;
constexpr uint32_t cb_t = 0, cb_l = 1, cb_one = 2, cb_bias = 3, cb_e = 4, cb_g = 5, 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 = NB / NC, E = NB % NC;
    const uint32_t b0 = core * B + (core < E ? core : E);
    const uint32_t n = B + (core < E ? 1 : 0);
    if (n == 0) {
        return;
    }
    compute_kernel_hw_startup(cb_l, cb_e);
    cb_wait_front(cb_one, 1);
    cb_wait_front(cb_bias, 3);
    for (uint32_t b = b0; b < b0 + n; ++b) {
        // E = exp(d)
        cb_wait_front(cb_l, 1);
        tile_regs_acquire();
        copy_init(cb_l);
        copy_tile(cb_l, 0, 0);
        exp_tile_init();
        exp_tile(0);
        tile_regs_commit();
        cb_reserve_back(cb_e, 1);
        tile_regs_wait();
        pack_reconfig_data_format(cb_e);
        pack_tile(0, cb_e);
        tile_regs_release();
        cb_push_back(cb_e, 1);
        cb_pop_front(cb_l, 1);
        // G = SCALE / (1 + E @ ONES)
        cb_wait_front(cb_e, 1);
        tile_regs_acquire();
        matmul_init(cb_e, cb_one);
        matmul_tiles(cb_e, cb_one, 0, 0, 0);
        binop_with_scalar_tile_init();
        add_unary_tile(0, ONE);
        recip_tile_init();
        recip_tile(0);
        binop_with_scalar_tile_init();
        mul_unary_tile(0, SCALE);
        tile_regs_commit();
        cb_reserve_back(cb_g, 1);
        tile_regs_wait();
        pack_reconfig_data_format(cb_g);
        pack_tile(0, cb_g);
        tile_regs_release();
        cb_push_back(cb_g, 1);
        cb_pop_front(cb_e, 1);
        // out_k = bf16(relu(t_k * G + b_k))
        cb_wait_front(cb_g, 1);
        cb_wait_front(cb_t, 3);
        cb_reserve_back(cb_out, 3);
        for (uint32_t k = 0; k < 3; ++k) {
            tile_regs_acquire();
            copy_init(cb_t);
            copy_tile(cb_t, k, 0);
            copy_init(cb_g);
            copy_tile(cb_g, 0, 1);
            mul_binary_tile_init();
            mul_binary_tile(0, 1, 0);
            copy_init(cb_bias);
            copy_tile(cb_bias, k, 1);
            add_binary_tile_init();
            add_binary_tile<ckernel::DstRoundingMode::NearestEven>(0, 1, 0);
            relu_tile_init();
            relu_tile(0);
            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();
            tile_regs_wait();
            pack_reconfig_data_format(cb_out);
            pack_tile(0, cb_out);
            tile_regs_release();
        }
        cb_push_back(cb_out, 3);
        cb_pop_front(cb_t, 3);
        cb_pop_front(cb_g, 1);
    }
}