#ifndef GF2_MUL_CUH
#define GF2_MUL_CUH

#include "gf2_mat.cuh"

namespace gf2
{
    __global__ void gpu_addmul_kernel(base_t *a, size_t a_rowstride, base_t *tb, size_t tb_rowstride, base_t *c, size_t c_rowstride, size_t ncols, size_t width, size_t nrows)
    {
        size_t w = blockIdx.x * blockDim.x + threadIdx.x;
        size_t r = blockIdx.y * blockDim.y + threadIdx.y;

        if (w >= width || r >= nrows)
        {
            return;
        }

        base_t val = *at_base(a, a_rowstride, r, 0);
        base_t temp = base_zero;
        for (size_t i = 0; i < gf2_table_num; i++)
        {
            temp ^= *at_base(tb, tb_rowstride, i * (1 << gf2_table_len) + (val & gf2_table_mask), w);
            val >>= gf2_table_len;
        }
        *at_base(c, c_rowstride, r, w) ^= temp;
    }
    __global__ void gpu_mktb_kernel(base_t *tb, size_t tb_rowstride, base_t *b, size_t b_rowstride, size_t tb_width)
    {
        size_t w = blockIdx.x * blockDim.x + threadIdx.x;
        size_t r = blockIdx.y * blockDim.y + threadIdx.y;

        if (w >= tb_width)
        {
            return;
        }

        base_t val = base_zero;
        base_t idx = r & gf2_table_mask;
        base_t st_row = (r >> gf2_table_len) * gf2_table_len;

        for (size_t i = 0; i < gf2_table_len; i++)
        {
            if (get(idx, i) != 0)
            {
                val ^= *at_base(b, b_rowstride, st_row + i, w);
            }
        }
        *at_base(tb, tb_rowstride, r, w) = val;
    }

    __host__ void MatGF2::gpu_addmul(const MatGF2 &a, const MatGF2 &b)
    {
        assert(a.ncols == b.nrows && a.nrows == nrows && b.ncols == ncols);
        MatGF2 tb(gf2_table_num * (1 << gf2_table_len), b.ncols);

        progress::ProgressBar pb("GPU MULTIPLY", a.width);
        for (size_t w = 0; w < a.width; w++, pb.tick_display())
        {
            size_t size = min(base_len, a.ncols - w * base_len);
            size_t tb_nrows = (size / gf2_table_len) * (1 << gf2_table_len) + (size % gf2_table_len == 0 ? 0 : 1 << (size % gf2_table_len));

            dim3 block_tb(THREAD_X, THREAD_Y);
            dim3 grid_tb((b.width - 1) / block_tb.x + 1, (tb_nrows - 1) / block_tb.y + 1);
            gpu_mktb_kernel<<<grid_tb, block_tb>>>(tb.data, tb.rowstride, b.at_base(w * base_len, 0), b.rowstride, tb.width);
            cudaDeviceSynchronize();

            dim3 block(THREAD_X, THREAD_Y);
            dim3 grid((b.width - 1) / block.x + 1, (nrows - 1) / block.y + 1);
            gpu_addmul_kernel<<<grid, block>>>(a.at_base(0, w), a.rowstride, tb.data, tb.rowstride, data, rowstride, size, width, nrows);
            cudaDeviceSynchronize();
        }
    }

    __host__ MatGF2 gpu_mul(const MatGF2 &a, const MatGF2 &b)
    {
        assert(a.ncols == b.nrows);
        MatGF2 c(a.nrows, b.ncols);
        c.gpu_addmul(a, b);
        return c;
    }
}

#endif