Datasets:
shibing624
/
source_code

Dataset card Data Studio Files
xet
 Community
2
text
stringlengths
0
2.2M
tdpbusd(x1, x2, x3);
} else if (jcp.src_dt == data_type::s8 && jcp.wei_dt == data_type::u8) {
tdpbsud(x1, x2, x3);
} else if (jcp.src_dt == data_type::s8 && jcp.wei_dt == data_type::s8) {
tdpbssd(x1, x2, x3);
} else {
assert(!"unsupported combination");
}
};
auto tileloadd_nt = [=](const Tmm &t1, int offset) {
int ab_size = jcp.nb_os2_blocking * jcp.nb_os_blocking * jcp.tile_width
* (jcp.nb_ic_int * jcp.ic_block_int_np
+ jcp.nb_oc_blocking * jcp.oc_block);
int c_size = (jcp.nb_ic_int * jcp.ic_block_int_np * jcp.nb_oc_blocking
* jcp.oc_block);
// If the size of src + wei used in the kernel cannot fit into L1 cache,
// use non-temporal load of weights to help keep src in L1 cache
if (static_cast<size_t>(jcp.typesize_in * (ab_size + c_size))
>= platform::get_per_core_cache_size(1))
tileloaddt1(t1, ptr[wei_ptr + offset + stride_seq]);
else
tileloadd(t1, ptr[wei_ptr + offset + stride_seq]);
};
auto compute_block = [=](int icb, int os_b) {
for (int osb = 0; osb < os_b; osb++) {
int ih = ((osb * jcp.tile_width) / jcp.ow) * jcp.stride_h;
int iw = ((osb * jcp.tile_width) % jcp.ow) * jcp.stride_w;
tileloadd(Tmm(get_inp_tensor(osb)),
ptr[inp_ptr + stride_nhwc + inp_offset(ih, iw, icb)]);
}
for (int ocb = 0; ocb < jcp.nb_oc_blocking; ocb++) {
const int wei_offset = jcp.typesize_in
* (ocb
* utils::rnd_up(jcp.ic_without_padding,
jcp.ic_block_int)
* jcp.oc_block
+ icb * jcp.ic_block_int_np * jcp.oc_block);
tileloadd_nt(Tmm(get_wei_tensor(ocb)), wei_offset);
for (int osb = 0; osb < os_b; osb++) {
tdpbxxd(Tmm(get_out_tensor(osb, ocb)), Tmm(get_inp_tensor(osb)),
Tmm(get_wei_tensor(ocb)));
interleave_store();
}
}
};
auto reconfig_tiles = [=](tiles_cfg_t cfg) {
tilerelease();
if (cfg == cfg_tiles) {
mov(reg_scratch, ptr[param1 + GET_OFF(tile_cfg)]);
} else if (cfg == cfg_tiles_tail) {
mov(reg_scratch, ptr[param1 + GET_OFF(tile_cfg_tail)]);
}
ldtilecfg(ptr[reg_scratch]);
};
auto restore_output_tiles = [=](int os_b, restore_tiles_t restore) {
mov(reg_tilebuff, ptr[param1 + GET_OFF(src_prf)]);
for (int ocb = 0; ocb < jcp.nb_oc_blocking; ocb++)
for (int osb = 0; osb < os_b; osb++) {
const int wsp_offset = jcp.typesize_acc
* (osb * jcp.nb_oc_blocking * jcp.max_width
* jcp.oc_block
+ ocb * jcp.max_width * jcp.oc_block);
if (restore == write_tiles)
tilestored(ptr[reg_tilebuff + stride_seq + wsp_offset],
Tmm(get_out_tensor(osb, ocb)));
else if (restore == read_tiles)
tileloadd(Tmm(get_out_tensor(osb, ocb)),
ptr[reg_tilebuff + stride_seq + wsp_offset]);
}
};
auto reset_tiles = [=](int os_b, bool tail) {
if (jcp.nb_ic_int != 1) {
restore_output_tiles(os_b, write_tiles);
reconfig_tiles((tail) ? cfg_tiles_tail : cfg_tiles);
restore_output_tiles(os_b, read_tiles);
}
};
auto compute_icb_loop = [=](int os_b = 1) {
int shift = (get_ic_tail() && os_b == 1) ? 1 : 0;
int nb_ic_int = jcp.nb_ic_int - shift;
if (jcp.src_zero_point) {
mov(reg_zp_compensation, ptr[param1 + GET_OFF(zp_compensation)]);
mov(reg_src_zero_point, ptr[param1 + GET_OFF(src_zero_point)]);
vpbroadcastd(zmm_src_zp, EVEX_compress_addr(reg_src_zero_point, 0));
}
if (jcp.dst_zero_point) {
mov(reg_dst_zero_point, ptr[param1 + GET_OFF(dst_zero_point)]);
vcvtdq2ps(zmm_dst_zp,
EVEX_compress_addr(reg_dst_zero_point, 0, true));
}
for (int icb = 0; icb < nb_ic_int; icb++)
compute_block(icb, os_b);