Skip to content

GitLab

Commit 1596cf1f authored by Shucai Xiao's avatar Shucai Xiao
Browse files

clang format

parent 6d0742b6
Hide whitespace changes
Inline Side-by-side
Showing with 74 additions and 62 deletions
src/onnx/onnx.cpp
View file @ 1596cf1f
......@@ -741,12 +741,12 @@ struct onnx_parser
act_funcs[1] = attributes.at("activations").strings(1);
}
if (act_funcs.size() != 2)
if(act_funcs.size() != 2)
{
MIGRAPHX_THROW("GRU: wrong activation function attribute");
}
for (std::size_t i = 0; i < act_funcs.size(); ++i)
for(std::size_t i = 0; i < act_funcs.size(); ++i)
{
if(actv_funcs.count(act_funcs.at(i)) == 0)
{
......@@ -762,15 +762,18 @@ struct onnx_parser
}
int linear_before_reset = 0;
if (contains(attributes, "linear_before_reset"))
if(contains(attributes, "linear_before_reset"))
{
linear_before_reset = parse_value(attributes.at("linear_before_reset")).at<int>();
}
return prog.add_instruction(op::gru{hidden_size,
{actv_funcs[act_funcs.at(0)], actv_funcs[act_funcs.at(1)]},
dirct, clip, linear_before_reset},
std::move(args));
return prog.add_instruction(
op::gru{hidden_size,
{actv_funcs[act_funcs.at(0)], actv_funcs[act_funcs.at(1)]},
dirct,
clip,
linear_before_reset},
std::move(args));
}
void parse_from(std::istream& is)
......
src/rewrite_gru.cpp
View file @ 1596cf1f
......@@ -24,7 +24,7 @@ void rewrite_gru::apply(program& prog) const
shape seq_shape = args[0]->get_shape();
std::size_t hidden_size = args[2]->get_shape().lens()[2];
std::size_t batchs = seq_shape.lens()[1];
std::size_t batchs = seq_shape.lens()[1];
shape::type_t type = seq_shape.type();
migraphx::shape ih_shape{type, {batchs, hidden_size}};
std::vector<char> data(ih_shape.bytes(), 0);
......@@ -36,17 +36,17 @@ void rewrite_gru::apply(program& prog) const
long hs = static_cast<long>(hidden_size);
// forward weight
auto uw_forward = prog.insert_instruction(ins, op::slice{{0}, {0}, {1}}, args[1]);
auto w_forward = prog.insert_instruction(ins, op::squeeze{{0}, uw_forward});
auto w_forward = prog.insert_instruction(ins, op::squeeze{{0}, uw_forward});
auto ur_forward = prog.insert_instruction(ins, op::slice{{0}, {0}, {1}}, args[2]);
auto r_forward = prog.insert_instruction(ins, op::squeeze{{0}}, ur_forward);
auto r_forward = prog.insert_instruction(ins, op::squeeze{{0}}, ur_forward);
// reverse weight
auto uw_reverse = prog.insert_instruction(ins, op::slice{{0}, {1}, {2}}, args[1]);
auto w_reverse = prog.insert_instruction(ins, op::squeeze{{0}, uw_reverse});
auto w_reverse = prog.insert_instruction(ins, op::squeeze{{0}, uw_reverse});
auto ur_reverse = prog.insert_instruction(ins, op::slice{{0}, {1}, {2}}, args[2]);
auto r_reverse = prog.insert_instruction(ins, op::squeeze{{0}}, ur_reverse);
auto r_reverse = prog.insert_instruction(ins, op::squeeze{{0}}, ur_reverse);
// process bias
instruction_ref bias_forward, bias_reverse;
......@@ -55,12 +55,12 @@ void rewrite_gru::apply(program& prog) const
{
// forward bias
auto uwb_forward = prog.insert_instruction(ins, op::slice{{0}, {0}, {1}}, args[3]);
bias_forward = prog.insert_instruction(ins, op::squeeze{{0}}, uwb_forward);
bias_forward = prog.insert_instruction(ins, op::squeeze{{0}}, uwb_forward);
// backward bias
auto uwb_reverse = prog.insert_instruction(ins, op::slice{{0}, {1}, {2}}, args[3]);
bias_reverse = prog.insert_instruction(ins, op::squeeze{{0}}, uwb_reverse);
}
bias_reverse = prog.insert_instruction(ins, op::squeeze{{0}}, uwb_reverse);
}
// intial hidden state
instruction_ref ih_forward, ih_reverse;
......@@ -104,7 +104,8 @@ void rewrite_gru::apply(program& prog) const
gru_op.actv_funcs.at(2),
gru_op.actv_funcs.at(3));
// auto final_output = prog.insert_instruction(ins, op::concat{0}, ret_forward[1], ret_reverse[1]);
// auto final_output = prog.insert_instruction(ins, op::concat{0}, ret_forward[1],
// ret_reverse[1]);
// add the dimension of num_direction
ret_forward[0] = prog.insert_instruction(ins, op::unsqueeze{{1}}, ret_forward[0]);
......@@ -117,14 +118,14 @@ void rewrite_gru::apply(program& prog) const
{
bool is_forward = (dicrt == op::gru::forward) ? true : false;
// weight matrix
auto w = prog.insert_instruction(ins, op::squeeze{{0}}, args[1]);
auto r = prog.insert_instruction(ins, op::squeeze{{0}}, args[2]);
auto w = prog.insert_instruction(ins, op::squeeze{{0}}, args[1]);
auto r = prog.insert_instruction(ins, op::squeeze{{0}}, args[2]);
// bias
instruction_ref bias = prog.end();
if(args.size() >= 4)
{
bias = prog.insert_instruction(ins, op::squeeze{{0}}, args[3]);
bias = prog.insert_instruction(ins, op::squeeze{{0}}, args[3]);
}
// intial hidden state
......@@ -138,8 +139,17 @@ void rewrite_gru::apply(program& prog) const
ih = prog.add_literal(migraphx::literal{s, data});
}
auto ret = gru_oper(
is_forward, prog, ins, args[0], w, r, ih, bias, gru_op.linear_before_reset, gru_op.actv_funcs.at(0), gru_op.actv_funcs.at(1));
auto ret = gru_oper(is_forward,
prog,
ins,
args[0],
w,
r,
ih,
bias,
gru_op.linear_before_reset,
gru_op.actv_funcs.at(0),
gru_op.actv_funcs.at(1));
// add the dimension of num_direction
prog.replace_instruction(ins, op::unsqueeze{{1}}, ret[0]);
......@@ -160,48 +170,48 @@ std::vector<instruction_ref> rewrite_gru::gru_oper(bool is_forward,
operation& actv_func2) const
{
instruction_ref hidden_out, final_out;
long seq_len = static_cast<long>(input->get_shape().lens()[0]);
long hs = static_cast<long>(r->get_shape().lens()[1]);
long seq_index = is_forward ? 0 : seq_len - 1;
long seq_len = static_cast<long>(input->get_shape().lens()[0]);
long hs = static_cast<long>(r->get_shape().lens()[1]);
long seq_index = is_forward ? 0 : seq_len - 1;
migraphx::shape s(input->get_shape().type(), {1});
auto l1 = prog.add_literal(migraphx::leteral{s, {1}});
// weight matrix
std::vector<int64_t> perm{1, 0};
auto wz = prog.insert_instruction(ins, op::slice{{0}, {0}, {hs}}, w);
auto wz = prog.insert_instruction(ins, op::slice{{0}, {0}, {hs}}, w);
auto twz = prog.insert_instruction(ins, op::transpose{perm}, wz);
auto wr = prog.insert_instruction(ins, op::slice{{0}, {hs}, {2 * hs}}, w);
auto wr = prog.insert_instruction(ins, op::slice{{0}, {hs}, {2 * hs}}, w);
auto twr = prog.insert_instruction(ins, op::transpose{perm}, wr);
auto wh = prog.insert_instruction(ins, op::slice{{0}, {2 * hs}, {3 * hs}}, w);
auto wh = prog.insert_instruction(ins, op::slice{{0}, {2 * hs}, {3 * hs}}, w);
auto twh = prog.insert_instruction(ins, op::transpose{perm}, wh);
auto rz = prog.insert_instruction(ins, op::slice{{0}, {0}, {hs}}, r);
auto rz = prog.insert_instruction(ins, op::slice{{0}, {0}, {hs}}, r);
auto trz = prog.insert_instruction(ins, op::transpose{perm}, rz);
auto rr = prog.insert_instruction(ins, op::slice{{0}, {hs}, {2 * hs}}, r);
auto rr = prog.insert_instruction(ins, op::slice{{0}, {hs}, {2 * hs}}, r);
auto trr = prog.insert_instruction(ins, op::transpose{perm}, rr);
auto rh = prog.insert_instruction(ins, op::slice{{0}, {2 * hs}, {3 * hs}}, r);
auto rh = prog.insert_instruction(ins, op::slice{{0}, {2 * hs}, {3 * hs}}, r);
auto trh = prog.insert_instruction(ins, op::transpose{perm}, rh);
// bias
instruction_ref br_bz, br_br, br_wbh, br_rbh, br_bh;
if (bias != prog.end())
if(bias != prog.end())
{
auto wbz = prog.insert_instruction(ins, op::slice{{0}, {0}, {hs}}, bias);
auto wbr = prog.insert_instruction(ins, op::slice{{0}, {hs}, {2*hs}}, bias);
wbh = prog.insert_instruction(ins, op::slice{{0}, {2*hs}, {3*hs}}, bias);
br_wbh = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, wbh);
auto wbr = prog.insert_instruction(ins, op::slice{{0}, {hs}, {2 * hs}}, bias);
wbh = prog.insert_instruction(ins, op::slice{{0}, {2 * hs}, {3 * hs}}, bias);
br_wbh = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, wbh);
auto rbz = prog.insert_instruction(ins, op::slice{{0}, {3*hs}, {4*hs}}, bias);
auto rbr = prog.insert_instruction(ins, op::slice{{0}, {4*hs}, {5*hs}}, bias);
rbh = prog.insert_instruction(ins, op::slice{{0}, {5*hs}, {6*hs}}, bias);
br_rbh = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, rbh);
auto rbz = prog.insert_instruction(ins, op::slice{{0}, {3 * hs}, {4 * hs}}, bias);
auto rbr = prog.insert_instruction(ins, op::slice{{0}, {4 * hs}, {5 * hs}}, bias);
rbh = prog.insert_instruction(ins, op::slice{{0}, {5 * hs}, {6 * hs}}, bias);
br_rbh = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, rbh);
auto bz = prog.insert_instruction(ins, op::add{}, wbz, rbz);
br_bz = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, bz);
br_bz = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, bz);
auto br = prog.insert_instruction(ins, op::add{}, wbr, rbr);
br_br = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, br);
br_bh = prog.insert_instruction(ins, op::add{}, br_wbh, br_rbh);
br_br = prog.insert_instruction(ins, op::broadcast{1, ih->get_shape()}, br);
br_bh = prog.insert_instruction(ins, op::add{}, br_wbh, br_rbh);
}
for(long i = 0; i < seq_len; i++)
......@@ -209,50 +219,50 @@ std::vector<instruction_ref> rewrite_gru::gru_oper(bool is_forward,
auto xt = prog.insert_instruction(ins, op::slice{{0}, {seq_index}, {seq_index + 1}}, input);
xt = prog.insert_instruction(ins, op::squeeze{{0}}, xt);
// equation f(xt*(Wz^T) + Ht-1 * (Rz^T) + Wbz + Rbz)
auto xwzt = prog.insert_instruction(ins, op::dot{}, xt, twz);
auto hrzt = prog.insert_instruction(ins, op::dot{}, ih, trz);
auto xwzt = prog.insert_instruction(ins, op::dot{}, xt, twz);
auto hrzt = prog.insert_instruction(ins, op::dot{}, ih, trz);
auto xwhr_zt = prog.insert_instruction(ins, op::add{}, xwzt, hrzt);
if (bias != prog.end())
if(bias != prog.end())
{
xwhr_zt = prog.insert_instruction(ins, op::add{}, xwhr_zt, br_bz);
}
auto zt = prog.insert_instruction(ins, actv_func1, xwhr_zt);
// equation f(Xt*(Wr^T) + Ht-1*(Rr^T) + Wbr + Rbr)
auto xwrt = prog.insert_instruction(ins, op::dot{}, xt, twr);
auto hrrt = prog.insert_instruction(ins, op::dot{}, xt, trr);
auto xwrt = prog.insert_instruction(ins, op::dot{}, xt, twr);
auto hrrt = prog.insert_instruction(ins, op::dot{}, xt, trr);
auto xwhr_rt = prog.insert_instruction(ins, op::add{}, xwrt, hrrt);
if (bias != prog.end())
if(bias != prog.end())
{
xwhr_rt = prog.insert_instruction(ins, op::add{}, xwhr_rt, br_br);
}
auto rt = prog.insert_instruction(ins, actv_func1, xwhr_rt);
instruction_ref xwhh_rt;
if (linear_before_reset == 0)
if(linear_before_reset == 0)
{
// equation g(Xt*(Wh^T) + (rt (.) Ht-1)*(Rh^T) + Rbh + Wbh)
auto xwht = prog.insert_instruction(ins, op::dot{}, xt, twh);
auto xwht = prog.insert_instruction(ins, op::dot{}, xt, twh);
auto rt_ht = prog.insert_instruction(ins, op::mul{}, rt, ih);
auto rt_rh = prog.insert_instruction(ins, op::dot{}, rt_ht, trh);
xwhh_rt = prog.insert_instruction(ins, op::add{}, xwht, rt_rt);
if (bias != prog.end())
xwhh_rt = prog.insert_instruction(ins, op::add{}, xwht, rt_rt);
if(bias != prog.end())
{
xwhh_rt = prog.insert_instruction(ins, op::add{}, xwhh_rt, br_bh);
}
}
else
else
{
// equation ht = g(Xt*(Wh^T) + (rt (.) (Ht-1*(Rh^T) + Rbh)) + Wbh)
auto xwht = prog.insert_instruction(ins, op::dot{}, xt, twh);
auto xwht = prog.insert_instruction(ins, op::dot{}, xt, twh);
auto ih_rht = prog.insert_instruction(ins, op::dot{}, ih, twh);
if (bias != prog.end())
if(bias != prog.end())
{
ih_rht = prog.insert_instruction(ins, op::add{}, ih_rht, br_rbh);
}
auto rt_rh = prog.insert_instruction(ins, op::mul{}, rt, ih_rht);
xwhh_rt = prog.insert_instruction(ins, op::add{}, xwht, rt_rh);
if (bias != prog.end())
xwhh_rt = prog.insert_instruction(ins, op::add{}, xwht, rt_rh);
if(bias != prog.end())
{
xwhh_rt = prog.insert_instruction(ins, op::add{}, xwhh_rt, br_wbh);
}
......@@ -260,17 +270,16 @@ std::vector<instruction_ref> rewrite_gru::gru_oper(bool is_forward,
ht = prog.insert_instruction(ins, actv_func2, xwhh_rt);
// equation Ht = (1 - zt) (.) ht + zt (.) Ht-1
auto 1zt = prog.insert_instruction(ins, op::sub{}, l1, zt);
auto 1zt = prog.insert_instruction(ins, op::sub{}, l1, zt);
auto 1ztht = prog.insert_instruction(ins, op::mul{}, 1zt, ht);
auto ztht1 = prog.insert_instruction(ins, op::mul{}, zt, ih);
ih = prog.insert_instruction(ins, op::add{}, 1ztht ztht1);
final_out = ih;
ih = prog.insert_instruction(ins, op::add{}, 1ztht ztht1);
final_out = ih;
if(is_forward)
{
hidden_out = (seq_index == 0)
? ih
: prog.insert_instruction(ins, op::concat{0}, hidden_out, ih);
hidden_out =
(seq_index == 0) ? ih : prog.insert_instruction(ins, op::concat{0}, hidden_out, ih);
}
else
{
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment