// SPDX-License-Identifier: Apache-2.0
// BRES compute (meteor-p150 OPT2 round 2, tt/bresize.py): per output row block b and channel tile ct
//   D = sum_e Wt[e] @ x[e, ct]       (FPU matmuls into one fp32 DEST tile; bf16 x bf16 products, BRES_FID fidelity)
//   out = bf16_rne(D)                (SFPU typecast, BRES_OUT_BF16) or the fp32 D (packed to an fp32 CB)
// CB 1 holds the block's KM weight tiles (in0), CB 0 its KM x CT input tiles (in1, entry-major).
// NX = 2: D also sums Wt[e] @ x_lo[e, ct] (an fp32 map as its two bf16 terms; tiles KM x CT on in CB 0).
// CT args: [0] NB, [1] NC, [2] GX, [3] KM, [4] CT, [5] NX. Define BRES_OUT_BF16 (1: bf16 output).
#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_unary/eltwise_unary.h"
#include "api/compute/eltwise_unary/typecast.h"
#include "api/compute/matmul.h"

#ifndef BRES_OUT_BF16
#define BRES_OUT_BF16 1
#endif

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 KM = get_compile_time_arg_val(3);
constexpr uint32_t CT = get_compile_time_arg_val(4);
constexpr uint32_t NX = get_compile_time_arg_val(5);
constexpr uint32_t cb_x = 0, cb_w = 1, 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<SrcOrder::Reverse>(cb_w, cb_x, cb_out);
    for (uint32_t b = b0; b < b0 + n; ++b) {
        cb_wait_front(cb_w, KM);
        cb_wait_front(cb_x, NX * KM * CT);
        for (uint32_t ct = 0; ct < CT; ++ct) {
            tile_regs_acquire();
            matmul_init(cb_w, cb_x);
            for (uint32_t e = 0; e < KM; ++e) {
                matmul_tiles(cb_w, cb_x, e, e * CT + ct, 0);
            }
            if constexpr (NX > 1) {
                for (uint32_t e = 0; e < KM; ++e) {
                    matmul_tiles(cb_w, cb_x, e, (KM + e) * CT + ct, 0);
                }
            }
#if BRES_OUT_BF16
            typecast_tile_init<(uint32_t)DataFormat::Float32, (uint32_t)DataFormat::Float16_b>();
            typecast_tile<(uint32_t)DataFormat::Float32, (uint32_t)DataFormat::Float16_b>(0);
#endif
            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_x, NX * KM * CT);
        cb_pop_front(cb_w, KM);
    }
}