Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Polish codes #35

Merged
merged 1 commit into from
Jun 14, 2019
Merged

Polish codes #35

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 0 additions & 22 deletions common/common_bitpack.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,28 +9,6 @@

#include <common/helper.h>

inline void pack_128_fallback(const float *float_ptr, void *binary_ptr,
size_t size) {
uint64_t *ui64_ptr = static_cast<uint64_t *>(binary_ptr);
const size_t UNIT_LEN = 64;
std::bitset<UNIT_LEN> bits1;
std::bitset<UNIT_LEN> bits2;
static_assert(sizeof(decltype(bits1.to_ullong())) * CHAR_BIT == 64,
"bits.to_ullong() must return a 64-bit element");

FORZS(j, size, 128) {
FORZS(i, 128, 4) {
const auto t = i / 4;
bits1[t] = (*(float_ptr + j + i) > 0);
bits1[t + 32] = (*(float_ptr + j + i + 1) > 0);
bits2[t] = (*(float_ptr + j + i + 2) > 0);
bits2[t + 32] = (*(float_ptr + j + i + 3) > 0);
}
*ui64_ptr++ = bits1.to_ullong();
*ui64_ptr++ = bits2.to_ullong();
}
}

inline void pack_64_bitset(const float *fptr, uint64_t *buf) {
const size_t UNIT_LEN = 64;
std::bitset<UNIT_LEN> bits;
Expand Down
23 changes: 16 additions & 7 deletions dabnn/bitpack.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@
#include "mat.h"

#ifdef __aarch64__
inline void pack_128_2(const float *float_ptr, void *binary_ptr, size_t size) {
inline void pack_128_opt(const float *float_ptr, void *binary_ptr, size_t size) {
/**
* This is the optimized bit-packing.
*
Expand Down Expand Up @@ -122,7 +122,7 @@ inline void pack_128_2(const float *float_ptr, void *binary_ptr, size_t size) {
"v19", "v20", "v21", "v22", "v23", "x0");
}

inline void pack_128(const float *float_ptr, void *binary_ptr, size_t size) {
inline void pack_128_baseline(const float *float_ptr, void *binary_ptr, size_t size) {
size_t nn_size = size >> 7;

asm volatile(
Expand Down Expand Up @@ -210,19 +210,28 @@ inline void pack_128(const float *float_ptr, void *binary_ptr, size_t size) {
"x0");
}

inline void pack_mat_128_2(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
inline void pack_mat_128_opt(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
assert(!binary_mat.empty());

pack_128_2(static_cast<float *>(float_mat.data), binary_mat.data,
pack_128_opt(static_cast<float *>(float_mat.data), binary_mat.data,
float_mat.total());
}

inline void pack_mat_128(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
inline void pack_mat_128_baseline(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
assert(!binary_mat.empty());

pack_128(static_cast<float *>(float_mat.data), binary_mat.data,
pack_128_baseline(static_cast<float *>(float_mat.data), binary_mat.data,
float_mat.total());
}

inline void pack_mat_128(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
/**
* Delegate it to optimized implementation.
* The cost of function calling will be eliminated by compiler,
* don't bother.
*/
pack_mat_128_opt(float_mat, binary_mat);
}
#endif // __aarch64__

inline void pack_mat_64(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
Expand Down Expand Up @@ -251,7 +260,7 @@ inline void pack_mat(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {
BNN_ASSERT(float_mat.c % 64 == 0, float_mat.c);
#ifdef __aarch64__
if (float_mat.c % 128 == 0) {
pack_mat_128_2(float_mat, binary_mat);
pack_mat_128_opt(float_mat, binary_mat);
} else {
pack_mat_64(float_mat, binary_mat);
}
Expand Down
2 changes: 1 addition & 1 deletion dabnn/net.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ void Net::prepare() {
add_mat(name, std::make_shared<Mat>(shape[0], shape[1],
shape[2], shape[3],
bnn::DataType::Bit, false));
pack_mat_128_2(*tmp, *mat_map_[name]);
pack_mat_128(*tmp, *mat_map_[name]);
} else {
#endif // __aarch64__
add_mat(name, std::make_shared<Mat>(
Expand Down
22 changes: 1 addition & 21 deletions tests/bitpack_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@ TEST(bitpack, pack_mat_128) {
const bnn::Mat a(AHEIGHT, AWIDTH, CHANNEL, a_data, bnn::DataType::Float);
bnn::Mat a_binary(AHEIGHT, AWIDTH, CHANNEL, bnn::DataType::Bit);
bnn::Mat expected(AHEIGHT, AWIDTH, CHANNEL, bnn::DataType::Bit);
pack_mat_128(a, a_binary);
pack_mat_128_opt(a, a_binary);

baseline_pack_mat(a, expected);

Expand Down Expand Up @@ -59,26 +59,6 @@ TEST(bitpack, pack_mat_64) {
}
}

#ifdef __aarch64__
TEST(bitpack, pack_mat_fallback) {
const size_t AHEIGHT = 64;
const size_t AWIDTH = 64;
const size_t CHANNEL = 256;
const size_t ALEN = AHEIGHT * AWIDTH * CHANNEL;
float a_data[ALEN];
fill_rand_float(a_data, ALEN);

const bnn::Mat a(AHEIGHT, AWIDTH, CHANNEL, a_data, bnn::DataType::Float);
bnn::Mat a_binary(AHEIGHT, AWIDTH, CHANNEL, bnn::DataType::Bit);
bnn::Mat expected(AHEIGHT, AWIDTH, CHANNEL, bnn::DataType::Bit);
pack_mat_128(a, a_binary);

pack_128_fallback(a_data, expected.data, ALEN);

ASSERT_EQ(a_binary, expected);
}
#endif // __aarch64__

TEST(bitpack, addv_v7) {
uint64_t data[2];
fill_rand_uint64(data, 2);
Expand Down
57 changes: 15 additions & 42 deletions tools/onnx2bnn/OnnxConverter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,8 +59,8 @@ void OnnxConverter::AddBinConv(const std::string &input_name,
const auto param = flatbnn::CreateBinConv2DDirect(
builder_, bin_name.c_str(), weight_name.c_str(), nullptr, &pads,
&strides, &dilations, output_name.c_str());
const auto layer = flatbnn::CreateLayer(
builder_, flatbnn::LayerType::BinConv2D, 0, param);
const auto layer =
flatbnn::CreateLayer(builder_, flatbnn::LayerType::BinConv2D, 0, param);
const auto flat_tensor = flatbnn::CreateTensorDirect(
builder_, flatbnn::DataType::Bit, &bin_weight.data, nullptr,
&bin_weight.shape, weight_name.c_str());
Expand Down Expand Up @@ -145,18 +145,11 @@ OnnxConverter::BTensor OnnxConverter::bitpack(OnnxConverter::FTensor ftensor) {
BNN_ASSERT(c % 64 == 0, ftensor.shape);

vector<bin_t> packed_data;
// if (c % 128 == 0) {
if (false) {
const auto size = Shaper::total(ftensor.shape);
packed_data.resize(size / 64);
pack_128_fallback(&ftensor.data[0], &packed_data[0], size);
} else {
bin_t tmp;
bin_t tmp;

FORZS(i, Shaper::total(ftensor.shape), 64) {
pack_64_bitset(&ftensor.data[i], &tmp);
packed_data.push_back(tmp);
}
FORZS(i, Shaper::total(ftensor.shape), 64) {
pack_64_bitset(&ftensor.data[i], &tmp);
packed_data.push_back(tmp);
}

Shape shape = {ftensor.shape[0], ftensor.shape[1], ftensor.shape[2],
Expand Down Expand Up @@ -190,27 +183,6 @@ std::vector<OnnxConverter::BTensor> OnnxConverter::split(
return outputs;
}

vector<bin_t> bitpack(const float *data, Shape shape) {
static_assert(std::is_same<bin_t, uint64_t>::value,
"bitpack requires bin_t is 64 bit");

auto c = Shaper::onnx_kc(shape);

BNN_ASSERT(c % 64 == 0, shape);

vector<bin_t> packed;

bin_t tmp;

FORZS(i, Shaper::total(shape), 64) {
pack_64_bitset(&data[i], &tmp);
packed.push_back(tmp);
}
BNN_ASSERT(false, "");

return packed;
}

void OnnxConverter::Convert(const ONNX_NAMESPACE::ModelProto &model_proto,
const std::string &filepath,
const OnnxConverter::Level level) {
Expand All @@ -220,18 +192,19 @@ void OnnxConverter::Convert(const ONNX_NAMESPACE::ModelProto &model_proto,
// Please check out "dabnn_*" pases in
// https://github.com/daquexian/onnx/blob/optimizer_for_bnn/onnx/optimizer/passes
// for details.
vector<string> optimizers{"eliminate_nop_pad", "extract_constant_to_initializer",
"dabnn_bconv_strict"};
vector<string> optimizers{"eliminate_nop_pad",
"extract_constant_to_initializer",
"dabnn_bconv_strict"};
if (level == Level::kModerate || level == Level::kAggressive) {
optimizers.push_back("dabnn_bconv_moderate");
}
if (level == Level::kAggressive) {
optimizers.push_back("dabnn_bconv_aggressive");
}
// model_proto is only used here. Please use the member variable model_proto_
// in the following code
model_proto_ = ONNX_NAMESPACE::optimization::Optimize(
model_proto, optimizers);
// model_proto is only used here. Please use the member variable
// model_proto_ in the following code
model_proto_ =
ONNX_NAMESPACE::optimization::Optimize(model_proto, optimizers);

for (const auto &tensor : model_proto_.graph().initializer()) {
if (tensor.data_type() == ONNX_NAMESPACE::TensorProto_DataType_FLOAT) {
Expand Down Expand Up @@ -548,8 +521,8 @@ void OnnxConverter::CalculateCoeff(const ONNX_NAMESPACE::NodeProto &node,
coeff_b_data.push_back(b.data[i] - scale.data[i] * mean.data[i] / tmp);
}
for (const auto &node2 : model_proto_.graph().node()) {
if (node2.domain() == "dabnn" && node2.op_type() == "Conv"
&& node2.output(0) == node.input(0)) {
if (node2.domain() == "dabnn" && node2.op_type() == "Conv" &&
node2.output(0) == node.input(0)) {
const auto &weight = onnx_float_tensors_[node2.input(1)];
{
int channels = Shaper::onnx_kc(weight.shape);
Expand Down