// SPDX-License-Identifier: Apache-2.0
// TGATE reader (meteor-p150 OPT2 round 3 R3.2, tt/tgate.py): the merged temporal-fuse / gate conv output g
// [R, GT * 32] fp32 TILE (columns 0..95 = tfuse.0's W x, columns 96..98 = the gate logit differences l_j - l_0) per
// 32-row block b: tiles b * GT + 0..2 into CB 0, tile b * GT + 3 into CB 1; once: the ones tile (CB 2) and the 3
// tfuse.0 bias tiles (CB 3, every row the bias).
// Work split (no per-core RT args, probe P3): core i = y * GX + x of the first NC cores owns the contiguous block
// range [i * B + min(i, E), + B + (i < E)) of the NB blocks, B = NB / NC, E = NB % NC.
// CT args: [0] NB, [1] NC, [2] GX, [3] GT, then the TensorAccessorArgs of g, ones, bias.
// Common RT args: [g_addr, ones_addr, bias_addr].
#include <cstdint>

#include "api/dataflow/dataflow_api.h"

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 GT = get_compile_time_arg_val(3);
constexpr auto g_args = TensorAccessorArgs<4>();
constexpr auto o_args = TensorAccessorArgs<g_args.next_compile_time_args_offset()>();
constexpr auto b_args = TensorAccessorArgs<o_args.next_compile_time_args_offset()>();
constexpr uint32_t TB = 4096;
constexpr uint32_t cb_t = 0, cb_l = 1, cb_one = 2, cb_bias = 3;

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;
    }
    const auto g = TensorAccessor(g_args, get_common_arg_val<uint32_t>(0), TB);
    const auto ones = TensorAccessor(o_args, get_common_arg_val<uint32_t>(1), TB);
    const auto bias = TensorAccessor(b_args, get_common_arg_val<uint32_t>(2), TB);
    cb_reserve_back(cb_one, 1);
    noc_async_read(ones.get_noc_addr(0), get_write_ptr(cb_one), TB);
    cb_reserve_back(cb_bias, 3);
    for (uint32_t k = 0; k < 3; ++k) {
        noc_async_read(bias.get_noc_addr(k), get_write_ptr(cb_bias) + k * TB, TB);
    }
    noc_async_read_barrier();
    cb_push_back(cb_one, 1);
    cb_push_back(cb_bias, 3);
    for (uint32_t b = b0; b < b0 + n; ++b) {
        cb_reserve_back(cb_l, 1);
        cb_reserve_back(cb_t, 3);
        noc_async_read(g.get_noc_addr(b * GT + 3), get_write_ptr(cb_l), TB);
        const uint32_t wt = get_write_ptr(cb_t);
        for (uint32_t k = 0; k < 3; ++k) {
            noc_async_read(g.get_noc_addr(b * GT + k), wt + k * TB, TB);
        }
        noc_async_read_barrier();
        cb_push_back(cb_l, 1);
        cb_push_back(cb_t, 3);
    }
}