Skip to content

Commit

Permalink
style: consistent use of trailing underscore in jit_1x1_conv_conf_t
Browse files Browse the repository at this point in the history
  • Loading branch information
@ankalinin
ankalinin committed Dec 21, 2017
1 parent 2137049 commit 73a4f28
Show file tree
Hide file tree
Showing 3 changed files with 47 additions and 47 deletions.
30 changes: 15 additions & 15 deletions src/cpu/jit_avx512_common_1x1_conv_kernel.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1132,16 +1132,16 @@ void jit_avx512_common_1x1_conv_kernel::balance(jit_1x1_conv_conf_t &jcp,
{
if (nthreads < jcp.ngroups) {
/* simplification... fortunately it doesn't hurt much */
jcp.nthr_ = jcp.nthr_mb_ = jcp.nthr_g_ =
jcp.nthr_oc_b_ = jcp.nthr_ic_b_ = 1;
jcp.nthr = jcp.nthr_mb = jcp.nthr_g =
jcp.nthr_oc_b = jcp.nthr_ic_b = 1;
return;
}
const int nb_bcast = div_up(jcp.bcast_dim, jcp.bcast_block);
const int nb_load = div_up(jcp.load_dim, jcp.load_block);
const int nb_reduce = div_up(jcp.reduce_dim, jcp.reduce_block);

jcp.nthr_g_ = jcp.ngroups;
const int nthr = nthreads / jcp.nthr_g_;
jcp.nthr_g = jcp.ngroups;
const int nthr = nthreads / jcp.nthr_g;

auto calc_mem_cost = [=](int nthr_mb, int nthr_oc_b, int nthr_ic_b) {
/* calculate per thread memory cost (read/write). high level
Expand All @@ -1162,14 +1162,14 @@ void jit_avx512_common_1x1_conv_kernel::balance(jit_1x1_conv_conf_t &jcp,
}
return 0
+ bcast_koeff * div_up(jcp.mb * nb_reduce, nthr_mb)
* div_up(jcp.ngroups, jcp.nthr_g_)
* div_up(jcp.ngroups, jcp.nthr_g)
* div_up(nb_bcast, nthr_ic_b) * jcp.ic_block * jcp.reduce_block
/ jcp.stride_h / jcp.stride_w /* (n1) */
+ load_koeff * div_up(jcp.mb * nb_reduce, nthr_mb)
* div_up(jcp.ngroups, jcp.nthr_g_)
* div_up(jcp.ngroups, jcp.nthr_g)
* div_up(nb_load, nthr_oc_b) * jcp.oc_block * jcp.reduce_block
+ output_koeff /* (n2) */
* div_up(jcp.ngroups, jcp.nthr_g_) * div_up(nb_load, nthr_oc_b)
* div_up(jcp.ngroups, jcp.nthr_g) * div_up(nb_load, nthr_oc_b)
* div_up(nb_bcast, nthr_ic_b) * jcp.ic_block
* jcp.oc_block;
};
Expand All @@ -1184,22 +1184,22 @@ void jit_avx512_common_1x1_conv_kernel::balance(jit_1x1_conv_conf_t &jcp,
const int nthr_oc_b_max = nstl::min(nthr_par, nb_load);
for (nthr_oc_b = 1; nthr_oc_b <= nthr_oc_b_max; ++nthr_oc_b) {
nthr_ic_b = nstl::min(nthr_par / nthr_oc_b, nb_bcast);
if (nthr_mb * jcp.nthr_g_ * nthr_oc_b * nthr_ic_b < nthreads)
if (nthr_mb * jcp.nthr_g * nthr_oc_b * nthr_ic_b < nthreads)
continue;
int mem_cost = calc_mem_cost(nthr_mb, nthr_oc_b, nthr_ic_b);
if (mem_cost <= best_mem_cost) {
best_mem_cost = mem_cost;
jcp.nthr_mb_ = nthr_mb;
jcp.nthr_oc_b_ = nthr_oc_b;
jcp.nthr_ic_b_ = nthr_ic_b;
jcp.nthr_mb = nthr_mb;
jcp.nthr_oc_b = nthr_oc_b;
jcp.nthr_ic_b = nthr_ic_b;
}
}
}
if (jcp.nthr_mb_ > nthreads / 2 && jcp.nthr_mb_ < nthreads)
jcp.nthr_mb_ = nstl::min(jcp.mb, nthreads);
if (jcp.nthr_mb > nthreads / 2 && jcp.nthr_mb < nthreads)
jcp.nthr_mb = nstl::min(jcp.mb, nthreads);

jcp.nthr_ = jcp.nthr_mb_ * jcp.nthr_g_ * jcp.nthr_oc_b_ * jcp.nthr_ic_b_;
assert(jcp.nthr_ <= nthreads);
jcp.nthr = jcp.nthr_mb * jcp.nthr_g * jcp.nthr_oc_b * jcp.nthr_ic_b;
assert(jcp.nthr <= nthreads);
}

}
Expand Down
62 changes: 31 additions & 31 deletions src/cpu/jit_avx512_common_1x1_convolution.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -417,24 +417,24 @@ jit_avx512_common_1x1_convolution_bwd_weights_t ::
const auto &jcp = kernel_->jcp;

bctx_ = (simple_barrier::ctx_t *)malloc(
jcp.nthr_ * sizeof(simple_barrier::ctx_t), 64);
for (int i = 0; i < jcp.nthr_; ++i)
jcp.nthr * sizeof(simple_barrier::ctx_t), 64);
for (int i = 0; i < jcp.nthr; ++i)
simple_barrier::ctx_init(&bctx_[i]);

const int wei_size = jcp.ngroups * jcp.oc * jcp.ic;
ws_reduction_ =
(data_t *)malloc((jcp.nthr_mb_ - 1) * wei_size * sizeof(data_t), 64);
(data_t *)malloc((jcp.nthr_mb - 1) * wei_size * sizeof(data_t), 64);
acc_ker_ = new cpu_accumulator_1d_t<data_type::f32>();

if (conf_.with_bias()) {
const size_t max_buffer_size = jcp.nthr_ * 3 * 5 * 5 * 16 * 16;
const size_t max_buffer_size = jcp.nthr * 3 * 5 * 5 * 16 * 16;
reducer_bias_ = new cpu_reducer_t<data_type::f32>(
reduce_balancer_t(jcp.nthr_, jcp.oc_block,
reduce_balancer_t(jcp.nthr, jcp.oc_block,
jcp.ngroups * jcp.nb_load, jcp.mb, max_buffer_size));
}
if (jcp.transpose_src) {
const size_t tr_src_size =
jcp.nthr_mb_ * jcp.ngroups * jcp.ic * jcp.tr_is;
jcp.nthr_mb * jcp.ngroups * jcp.ic * jcp.tr_is;
tr_src_ = (data_t *)malloc(tr_src_size * sizeof(data_t), 64);
# pragma omp parallel for
for (size_t i = 0; i < tr_src_size; i++)
Expand Down Expand Up @@ -544,37 +544,37 @@ void jit_avx512_common_1x1_convolution_bwd_weights_t::execute_backward_weights()
};

auto ker = [&](const int ithr, const int nthr) {
const int ithr_ic_b = ithr % jcp.nthr_ic_b_;
const int ithr_oc_b = ithr / jcp.nthr_ic_b_ % jcp.nthr_oc_b_;
const int ithr_g = ithr / jcp.nthr_ic_b_ / jcp.nthr_oc_b_ % jcp.nthr_g_;
const int ithr_mb = ithr / jcp.nthr_ic_b_ / jcp.nthr_oc_b_ /
jcp.nthr_g_;
const int ithr_ic_b = ithr % jcp.nthr_ic_b;
const int ithr_oc_b = ithr / jcp.nthr_ic_b % jcp.nthr_oc_b;
const int ithr_g = ithr / jcp.nthr_ic_b / jcp.nthr_oc_b % jcp.nthr_g;
const int ithr_mb = ithr / jcp.nthr_ic_b / jcp.nthr_oc_b /
jcp.nthr_g;

const int ithr_but_oc
= (ithr_mb * jcp.nthr_g_ + ithr_g) * jcp.nthr_ic_b_ + ithr_ic_b;
= (ithr_mb * jcp.nthr_g + ithr_g) * jcp.nthr_ic_b + ithr_ic_b;

/* reduction dimension */
int mb_sp_b_start{ 0 }, mb_sp_b_end{ 0 };
if (jcp.transpose_src && jcp.nthr_mb_ < jcp.mb / 2) {
if (jcp.transpose_src && jcp.nthr_mb < jcp.mb / 2) {
// it's preferable to parallelize by mb if possible
int img_start{ 0 }, img_end{ 0 };
balance211(jcp.mb, jcp.nthr_mb_, ithr_mb, img_start, img_end);
balance211(jcp.mb, jcp.nthr_mb, ithr_mb, img_start, img_end);
mb_sp_b_start = img_start * sp_nb;
mb_sp_b_end = img_end * sp_nb;
}
else {
balance211(mb_sp_work, jcp.nthr_mb_, ithr_mb, mb_sp_b_start,
balance211(mb_sp_work, jcp.nthr_mb, ithr_mb, mb_sp_b_start,
mb_sp_b_end);
}

/* independent dimensions */
int g_start{ 0 }, oc_b_start{ 0 }, ic_b_start{ 0 };
int g_end{ 0 }, oc_b_end{ 0 }, ic_b_end{ 0 };

balance211(jcp.ngroups, jcp.nthr_g_, ithr_g, g_start, g_end);
balance211(jcp.nb_load, jcp.nthr_oc_b_, ithr_oc_b, oc_b_start,
balance211(jcp.ngroups, jcp.nthr_g, ithr_g, g_start, g_end);
balance211(jcp.nb_load, jcp.nthr_oc_b, ithr_oc_b, oc_b_start,
oc_b_end);
balance211(jcp.nb_bcast, jcp.nthr_ic_b_, ithr_ic_b, ic_b_start,
balance211(jcp.nb_bcast, jcp.nthr_ic_b, ithr_ic_b, ic_b_start,
ic_b_end);

const int g_work = g_end - g_start;
Expand Down Expand Up @@ -602,17 +602,17 @@ void jit_avx512_common_1x1_convolution_bwd_weights_t::execute_backward_weights()
bcast_step = step(nb_ic_blocking, ic_b_end - ic_b,
jcp.nb_bcast_blocking_max);
if (jcp.transpose_src) {
if (jcp.nthr_oc_b_ > 1)
if (jcp.nthr_oc_b > 1)
simple_barrier::barrier(
&bctx_[ithr_but_oc], jcp.nthr_oc_b_);
&bctx_[ithr_but_oc], jcp.nthr_oc_b);
const int sp_size
= nstl::min(sp_b_step * jcp.reduce_block,
jcp.is - sp_b * jcp.reduce_block);
uker_trans(ithr_mb, img, sp_b, sp_size, g, 1, ic_b,
bcast_step, ithr_oc_b, jcp.nthr_oc_b_, ic_b_start);
if (jcp.nthr_oc_b_ > 1)
bcast_step, ithr_oc_b, jcp.nthr_oc_b, ic_b_start);
if (jcp.nthr_oc_b > 1)
simple_barrier::barrier(
&bctx_[ithr_but_oc], jcp.nthr_oc_b_);
&bctx_[ithr_but_oc], jcp.nthr_oc_b);
}

for (int oc_b = oc_b_start; oc_b < oc_b_end;
Expand Down Expand Up @@ -687,15 +687,15 @@ void jit_avx512_common_1x1_convolution_bwd_weights_t::execute_backward_weights()
}

/* diff_weights[:] += sum(ws_reduction_[thr_mb][:]) */
if (jcp.nthr_mb_ > 1) {
simple_barrier::barrier(&reduction_barrier, jcp.nthr_);
if (jcp.nthr_mb > 1) {
simple_barrier::barrier(&reduction_barrier, jcp.nthr);
const int work = g_work * oc_b_work * ic_b_work;
int start{ 0 }, end{ 0 };
balance211(work, jcp.nthr_mb_, ithr_mb, start, end);
balance211(work, jcp.nthr_mb, ithr_mb, start, end);
if (start == end)
return;

for (int thr_mb = 1; thr_mb < jcp.nthr_mb_; ++thr_mb) {
for (int thr_mb = 1; thr_mb < jcp.nthr_mb; ++thr_mb) {
int w = start;
int sub_g_start{ 0 }, sub_oc_b_start{ 0 },
sub_ic_b_start{ 0 };
Expand Down Expand Up @@ -771,13 +771,13 @@ void jit_avx512_common_1x1_convolution_bwd_weights_t::execute_backward_weights()
rb->reduce(ithr, diff_bias);
};

#pragma omp parallel num_threads(jcp.nthr_)
#pragma omp parallel num_threads(jcp.nthr)
{
int ithr = omp_get_thread_num();
assert(jcp.nthr_ == omp_get_num_threads());
ker(ithr, jcp.nthr_);
assert(jcp.nthr == omp_get_num_threads());
ker(ithr, jcp.nthr);
if (conf_.with_bias())
ker_bias(ithr, jcp.nthr_);
ker_bias(ithr, jcp.nthr);
}
}

Expand Down
2 changes: 1 addition & 1 deletion src/cpu/jit_primitive_conf.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -227,7 +227,7 @@ struct jit_1x1_conv_conf_t {
/* 4fma */
bool transpose_src;
int tr_is;
int nthr_, nthr_mb_, nthr_g_, nthr_oc_b_, nthr_ic_b_;
int nthr, nthr_mb, nthr_g, nthr_oc_b, nthr_ic_b;
};

struct jit_gemm_conv_conf_t {
Expand Down

0 comments on commit 73a4f28

Please sign in to comment.