Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 5ac1e63c 编写于 作者: H hjchen2
浏览文件
操作

Support 1x1 and 7x7 conv, fix quant scale, we can run the googlenet with int8 now

上级 a69fc85c
展开全部 隐藏空白更改
内联 并排
Showing with 1608 addition and 661 deletion
src/framework/operator.cpp
浏览文件 @ 5ac1e63c
...@@ -32,7 +32,7 @@ template <typename Dtype> ...@@ -32,7 +32,7 @@ template <typename Dtype>
vector<string> OperatorBase<Dtype>::GetInputKeys() const { vector<string> OperatorBase<Dtype>::GetInputKeys() const {
auto it = op_input_output_key.find(type_); auto it = op_input_output_key.find(type_);
if (it == op_input_output_key.end()) { if (it == op_input_output_key.end()) {
DLOG << type_ << " has no outputs"; DLOG << type_ << " has no inputs";
return {}; return {};
} }
return it->second.first; return it->second.first;
......
src/operators/kernel/arm/dequantize_kernel.cpp
浏览文件 @ 5ac1e63c
...@@ -38,7 +38,8 @@ void DequantizeKernel<CPU, float>::Compute( ...@@ -38,7 +38,8 @@ void DequantizeKernel<CPU, float>::Compute(
const int32_t *x = input->data<const int32_t>(); const int32_t *x = input->data<const int32_t>();
float *y = output->mutable_data<float>(); float *y = output->mutable_data<float>();
size_t size = output->numel(); size_t size = output->numel();
float scale = 1.f / (activation_scale * weight_scale); // float scale = 1.f / (activation_scale * weight_scale);
float scale = activation_scale / weight_scale;
#if defined(__ARM_NEON__) || defined(__ARM_NEON) #if defined(__ARM_NEON__) || defined(__ARM_NEON)
size_t loop = size >> 4; size_t loop = size >> 4;
size_t remain = size & 0xF; size_t remain = size & 0xF;
......
src/operators/kernel/arm/quantize_kernel.cpp
浏览文件 @ 5ac1e63c
...@@ -280,17 +280,18 @@ void QuantizeKernel<CPU, float>::Compute( ...@@ -280,17 +280,18 @@ void QuantizeKernel<CPU, float>::Compute(
} }
max_abs = std::max(max_abs, 1e-6f); max_abs = std::max(max_abs, 1e-6f);
// only support int8 currently // only support int8 currently
float online_scale = 127 / max_abs; float scale = 127 / max_abs;
param.online_scale_->mutable_data<float>()[0] = online_scale; param.online_scale_->mutable_data<float>()[0] = max_abs;
switch (param.round_type_) { switch (param.round_type_) {
case ROUND_NEAREST_TO_EVEN: case ROUND_NEAREST_TO_EVEN:
quantize_round_to_even(input, online_scale, output); quantize_round_to_even(input, scale, output);
break; break;
case ROUND_NEAREST_TOWARDS_ZERO: case ROUND_NEAREST_TOWARDS_ZERO:
quantize_round_to_zero(input, online_scale, output); quantize_round_to_zero(input, scale, output);
break; break;
case ROUND_NEAREST_AWAY_ZERO: case ROUND_NEAREST_AWAY_ZERO:
quantize_round_to_nearest(input, online_scale, output); quantize_round_to_nearest(input, scale, output);
break;
default: default:
LOG(kLOG_ERROR) << "round type is not supported."; LOG(kLOG_ERROR) << "round type is not supported.";
break; break;
......
src/operators/kernel/central-arm-func/conv_arm_func.h
浏览文件 @ 5ac1e63c
...@@ -28,15 +28,15 @@ limitations under the License. */ ...@@ -28,15 +28,15 @@ limitations under the License. */
namespace paddle_mobile { namespace paddle_mobile {
namespace operators { namespace operators {
template <typename Dtype>
inline void ConvBasic(const ConvParam<CPU> &param) { inline void ConvBasic(const ConvParam<CPU> &param) {
const Tensor *input = param.Input(); const Tensor *input = param.Input();
Tensor filter = *param.Filter(); Tensor filter = *param.Filter();
Tensor *output = param.Output(); Tensor *output = param.Output();
output->mutable_data<float>();
int groups = param.Groups(); int groups = param.Groups();
std::vector<int> strides = param.Strides(); const std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings(); const std::vector<int> paddings = param.Paddings();
std::vector<int> dilations = param.Dilations(); const std::vector<int> dilations = param.Dilations();
const int batch_size = static_cast<int>(input->dims()[0]); const int batch_size = static_cast<int>(input->dims()[0]);
...@@ -60,7 +60,7 @@ inline void ConvBasic(const ConvParam<CPU> &param) { ...@@ -60,7 +60,7 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
Tensor col; Tensor col;
Tensor col_matrix; Tensor col_matrix;
if (is_expand) { if (is_expand) {
col.mutable_data<float>(col_shape); col.mutable_data<Dtype>(col_shape);
col_matrix.ShareDataWith(col); col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape); col_matrix.Resize(col_matrix_shape);
} }
...@@ -79,8 +79,8 @@ inline void ConvBasic(const ConvParam<CPU> &param) { ...@@ -79,8 +79,8 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
int in_step = static_cast<int>(input->dims()[1]) / groups; int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups; int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<CPU, float> vol2col; math::Vol2ColFunctor<CPU, Dtype> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col; math::Im2ColFunctor<math::ColFormat::kCFO, CPU, Dtype> im2col;
for (int i = 0; i < batch_size; i++) { for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
...@@ -99,6 +99,7 @@ inline void ConvBasic(const ConvParam<CPU> &param) { ...@@ -99,6 +99,7 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
std::vector<int>{paddings[0], paddings[1], paddings[0], std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]}, paddings[1]},
&col); &col);
} else if (data_dim == 3U) { } else if (data_dim == 3U) {
// vol2col // vol2col
vol2col(in_slice, dilations, strides, paddings, &col); vol2col(in_slice, dilations, strides, paddings, &col);
...@@ -107,7 +108,8 @@ inline void ConvBasic(const ConvParam<CPU> &param) { ...@@ -107,7 +108,8 @@ inline void ConvBasic(const ConvParam<CPU> &param) {
// gemm // gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step); Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step); Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<float>(filter_slice, false, col_matrix, false,
math::matmul<Dtype>(filter_slice, false, col_matrix, false,
static_cast<float>(1), &out_slice, static_cast<float>(1), &out_slice,
static_cast<float>(0)); static_cast<float>(0));
} }
...@@ -126,42 +128,41 @@ inline void ConvCompute_int8(const ConvParam<CPU> &param) { ...@@ -126,42 +128,41 @@ inline void ConvCompute_int8(const ConvParam<CPU> &param) {
const Tensor *input = param.Input(); const Tensor *input = param.Input();
Tensor *filter = param.Filter(); Tensor *filter = param.Filter();
Tensor *output = param.Output(); Tensor *output = param.Output();
output->mutable_data<int32_t>();
int groups = param.Groups(); int groups = param.Groups();
std::vector<int> strides = param.Strides(); const std::vector<int> &strides = param.Strides();
std::vector<int> paddings = param.Paddings(); const std::vector<int> &paddings = param.Paddings();
std::vector<int> dilations = param.Dilations(); const std::vector<int> &dilations = param.Dilations();
int kernel_h = filter->dims()[2]; int kernel_h = filter->dims()[2];
int kernel_w = filter->dims()[3]; int kernel_w = filter->dims()[3];
const int batch_size = static_cast<int>(input->dims()[0]); output->mutable_data<int32_t>();
math::PadFunctor<CPU, int8_t> pad;
Tensor input_pad;
for (int i = 0; i < batch_size; ++i) {
Tensor in_batch = input->Slice(i, i + 1);
Tensor out_batch = output->Slice(i, i + 1);
if (paddings[0] == 0 && paddings[1] == 0) {
input_pad = in_batch;
} else {
framework::DDim pad_shape = in_batch.dims();
pad_shape[2] += 2 * paddings[0];
pad_shape[3] += 2 * paddings[1];
input_pad.mutable_data<int8_t>(pad_shape);
pad(in_batch, paddings[0], paddings[1], &input_pad);
}
if (strides[1] == strides[0] && strides[1] < 6 && kernel_h == kernel_w && ConvFunc conv_func = 0;
kernel_h < 8 && groups == 1 && dilations[0] == dilations[1] && if (strides[1] == strides[0] && strides[1] < 6 && kernel_h == kernel_w &&
dilations[1] == 1) { kernel_h < 8 && groups == 1 && dilations[0] == dilations[1] &&
ConvFunc conv_func = conv_funcs_table[kernel_h - 1][strides[0] - 1]; dilations[1] == 1) {
if (conv_func) { conv_func = conv_funcs_table[kernel_h - 1][strides[0] - 1];
conv_func(input_pad, *filter, &out_batch); }
if (conv_func) {
int batch_size = input->dims()[0];
math::PadFunctor<CPU, int8_t> pad;
Tensor input_pad;
for (int i = 0; i < batch_size; ++i) {
Tensor in_batch = input->Slice(i, i + 1);
Tensor out_batch = output->Slice(i, i + 1);
if (paddings[0] == 0 && paddings[1] == 0) {
input_pad = in_batch;
} else { } else {
// TODO(hjchen2) framework::DDim pad_shape = in_batch.dims();
pad_shape[2] += 2 * paddings[0];
pad_shape[3] += 2 * paddings[1];
input_pad.mutable_data<int8_t>(pad_shape);
pad(in_batch, paddings[0], paddings[1], &input_pad);
} }
} else { conv_func(input_pad, *filter, &out_batch);
// TODO(hjchen2)
} }
} else {
ConvBasic<int8_t>(param);
} }
} }
...@@ -170,6 +171,7 @@ void ConvCompute(const ConvParam<CPU> &param) { ...@@ -170,6 +171,7 @@ void ConvCompute(const ConvParam<CPU> &param) {
if (param.Input()->type() == typeid(int8_t)) { if (param.Input()->type() == typeid(int8_t)) {
ConvCompute_int8(param); ConvCompute_int8(param);
} else { } else {
param.Output()->mutable_data<float>();
if (param.Groups() == param.Input()->dims()[1] && if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] && param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] && param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
...@@ -183,7 +185,7 @@ void ConvCompute(const ConvParam<CPU> &param) { ...@@ -183,7 +185,7 @@ void ConvCompute(const ConvParam<CPU> &param) {
math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(), math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
param.Filter(), nullptr, param.Output(), false); param.Filter(), nullptr, param.Output(), false);
} else { } else {
ConvBasic(param); ConvBasic<float>(param);
} }
} }
} }
......
src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h
浏览文件 @ 5ac1e63c
...@@ -44,7 +44,7 @@ void DepthwiseConvCompute(const ConvParam<CPU> &param) { ...@@ -44,7 +44,7 @@ void DepthwiseConvCompute(const ConvParam<CPU> &param) {
Bias, false); Bias, false);
} else { } else {
ConvBasic(param); ConvBasic<float>(param);
} }
} }
......
src/operators/kernel/central-arm-func/mul_arm_func.h
浏览文件 @ 5ac1e63c
...@@ -58,7 +58,7 @@ void MulCompute(const MulParam<CPU> &param) { ...@@ -58,7 +58,7 @@ void MulCompute(const MulParam<CPU> &param) {
const Tensor *input_x = param.InputX(); const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY(); const Tensor *input_y = param.InputY();
Tensor *out = param.Out(); Tensor *out = param.Out();
out->mutable_data<float>();
const Tensor x_matrix = const Tensor x_matrix =
input_x->dims().size() > 2 input_x->dims().size() > 2
? framework::ReshapeToMatrix(*input_x, param.XNumColDims()) ? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
...@@ -71,15 +71,21 @@ void MulCompute(const MulParam<CPU> &param) { ...@@ -71,15 +71,21 @@ void MulCompute(const MulParam<CPU> &param) {
if (out_dim.size() != 2) { if (out_dim.size() != 2) {
out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]}); out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
} }
math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1), if (param.InputX()->type() == typeid(int8_t)) {
out, static_cast<float>(0)); out->mutable_data<int32_t>();
math::matmul<int8_t>(x_matrix, false, y_matrix, false,
static_cast<int8_t>(1), out, static_cast<int8_t>(0));
} else {
out->mutable_data<float>();
math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
out, static_cast<float>(0));
}
if (out_dim.size() != 2) { if (out_dim.size() != 2) {
out->Resize(out_dim); out->Resize(out_dim);
} }
} }
template class MulKernel<CPU, float>;
} // namespace operators } // namespace operators
} // namespace paddle_mobile } // namespace paddle_mobile
......
src/operators/math/conv3x3_arm_int8.cpp
浏览文件 @ 5ac1e63c
...@@ -112,15 +112,15 @@ void conv3x3s1_int8(const framework::Tensor& input, ...@@ -112,15 +112,15 @@ void conv3x3s1_int8(const framework::Tensor& input,
"vmull.s8 q7, d4, d7 \n" "vmull.s8 q7, d4, d7 \n"
"vmlal.s8 q6, d5, d8 \n" "vmlal.s8 q6, d5, d8 \n"
"vaddw.s16 q12, q12, d12 \n" "vaddw.s16 q12, q12, d12 \n"
"vaddw.s16 q12, q12, d14 \n"
"vaddw.s16 q13, q13, d13 \n" "vaddw.s16 q13, q13, d13 \n"
"vaddw.s16 q12, q12, d14 \n"
"vaddw.s16 q13, q13, d15 \n" "vaddw.s16 q13, q13, d15 \n"
"vmull.s8 q6, d2, d9 \n" "vmull.s8 q6, d2, d9 \n"
"vmull.s8 q7, d4, d10 \n" "vmull.s8 q7, d4, d10 \n"
"vmlal.s8 q6, d5, d11 \n" "vmlal.s8 q6, d5, d11 \n"
"vaddw.s16 q14, q14, d12 \n" "vaddw.s16 q14, q14, d12 \n"
"vaddw.s16 q14, q14, d14 \n"
"vaddw.s16 q15, q15, d13 \n" "vaddw.s16 q15, q15, d13 \n"
"vaddw.s16 q14, q14, d14 \n"
"vaddw.s16 q15, q15, d15 \n" "vaddw.s16 q15, q15, d15 \n"
"vld1.8 {d2-d3}, [%[r2]] \n" // r2 "vld1.8 {d2-d3}, [%[r2]] \n" // r2
...@@ -139,8 +139,8 @@ void conv3x3s1_int8(const framework::Tensor& input, ...@@ -139,8 +139,8 @@ void conv3x3s1_int8(const framework::Tensor& input,
"vmull.s8 q7, d4, d10 \n" "vmull.s8 q7, d4, d10 \n"
"vmlal.s8 q6, d5, d11 \n" "vmlal.s8 q6, d5, d11 \n"
"vaddw.s16 q10, q10, d12 \n" "vaddw.s16 q10, q10, d12 \n"
"vaddw.s16 q10, q10, d14 \n"
"vaddw.s16 q11, q11, d13 \n" "vaddw.s16 q11, q11, d13 \n"
"vaddw.s16 q10, q10, d14 \n"
"vaddw.s16 q11, q11, d15 \n" "vaddw.s16 q11, q11, d15 \n"
"vdup.s8 d6, d0[6] \n" "vdup.s8 d6, d0[6] \n"
...@@ -153,21 +153,23 @@ void conv3x3s1_int8(const framework::Tensor& input, ...@@ -153,21 +153,23 @@ void conv3x3s1_int8(const framework::Tensor& input,
"vmull.s8 q7, d4, d7 \n" "vmull.s8 q7, d4, d7 \n"
"vmlal.s8 q6, d5, d8 \n" "vmlal.s8 q6, d5, d8 \n"
"vaddw.s16 q12, q12, d12 \n" "vaddw.s16 q12, q12, d12 \n"
"vaddw.s16 q12, q12, d14 \n"
"vaddw.s16 q13, q13, d13 \n" "vaddw.s16 q13, q13, d13 \n"
"vaddw.s16 q12, q12, d14 \n"
"vaddw.s16 q13, q13, d15 \n" "vaddw.s16 q13, q13, d15 \n"
"vld1.32 {d12-d15}, [%[output0]] \n"
"vadd.s32 q6, q6, q12 \n"
"vadd.s32 q7, q7, q13 \n"
"vst1.32 {d12-d15}, [%[output0]]! \n"
"vmull.s8 q6, d2, d9 \n" "vmull.s8 q6, d2, d9 \n"
"vmull.s8 q7, d4, d10 \n" "vmull.s8 q7, d4, d10 \n"
"vmlal.s8 q6, d5, d11 \n" "vmlal.s8 q6, d5, d11 \n"
"vaddw.s16 q14, q14, d12 \n" "vaddw.s16 q14, q14, d12 \n"
"vaddw.s16 q14, q14, d14 \n"
"vaddw.s16 q15, q15, d13 \n" "vaddw.s16 q15, q15, d13 \n"
"vaddw.s16 q14, q14, d14 \n"
"vaddw.s16 q15, q15, d15 \n" "vaddw.s16 q15, q15, d15 \n"
"vld1.32 {d12-d15}, [%[output0]] \n"
"vadd.s32 q6, q6, q12 \n"
"vadd.s32 q7, q7, q13 \n"
"vst1.32 {d12-d15}, [%[output0]]! \n"
"vld1.32 {d12-d15}, [%[output1]] \n" "vld1.32 {d12-d15}, [%[output1]] \n"
"vadd.s32 q6, q6, q14 \n" "vadd.s32 q6, q6, q14 \n"
"vadd.s32 q7, q7, q15 \n" "vadd.s32 q7, q7, q15 \n"
...@@ -182,21 +184,23 @@ void conv3x3s1_int8(const framework::Tensor& input, ...@@ -182,21 +184,23 @@ void conv3x3s1_int8(const framework::Tensor& input,
"vmull.s8 q7, d4, d7 \n" "vmull.s8 q7, d4, d7 \n"
"vmlal.s8 q6, d5, d8 \n" "vmlal.s8 q6, d5, d8 \n"
"vaddw.s16 q8, q8, d12 \n" "vaddw.s16 q8, q8, d12 \n"
"vaddw.s16 q9, q9, d15 \n"
"vaddw.s16 q8, q8, d14 \n" "vaddw.s16 q8, q8, d14 \n"
"vaddw.s16 q9, q9, d13 \n" "vaddw.s16 q9, q9, d13 \n"
"vaddw.s16 q9, q9, d15 \n"
"vld1.32 {d12-d15}, [%[output0n]] \n"
"vadd.s32 q6, q6, q8 \n"
"vadd.s32 q7, q7, q9 \n"
"vst1.32 {d12-d15}, [%[output0n]]! \n"
"vmull.s8 q6, d2, d9 \n" "vmull.s8 q6, d2, d9 \n"
"vmull.s8 q7, d4, d10 \n" "vmull.s8 q7, d4, d10 \n"
"vmlal.s8 q6, d5, d11 \n" "vmlal.s8 q6, d5, d11 \n"
"vaddw.s16 q10, q10, d12 \n" "vaddw.s16 q10, q10, d12 \n"
"vaddw.s16 q11, q11, d15 \n"
"vaddw.s16 q10, q10, d14 \n" "vaddw.s16 q10, q10, d14 \n"
"vaddw.s16 q11, q11, d13 \n" "vaddw.s16 q11, q11, d13 \n"
"vaddw.s16 q11, q11, d15 \n"
"vld1.32 {d12-d15}, [%[output0n]] \n"
"vadd.s32 q6, q6, q8 \n"
"vadd.s32 q7, q7, q9 \n"
"vst1.32 {d12-d15}, [%[output0n]]! \n"
"vld1.32 {d12-d15}, [%[output1n]] \n" "vld1.32 {d12-d15}, [%[output1n]] \n"
"vadd.s32 q6, q6, q10 \n" "vadd.s32 q6, q6, q10 \n"
"vadd.s32 q7, q7, q11 \n" "vadd.s32 q7, q7, q11 \n"
......
src/operators/math/conv_arm_int8.h
浏览文件 @ 5ac1e63c
...@@ -24,6 +24,10 @@ namespace operators { ...@@ -24,6 +24,10 @@ namespace operators {
void conv3x3s1_int8(const framework::Tensor& input, void conv3x3s1_int8(const framework::Tensor& input,
const framework::Tensor& weight, framework::Tensor* output); const framework::Tensor& weight, framework::Tensor* output);
void conv3x3s1_int8_4c(const framework::Tensor& input,
const framework::Tensor& weight,
framework::Tensor* output);
void conv5x5s1_int8(const framework::Tensor& input, void conv5x5s1_int8(const framework::Tensor& input,
const framework::Tensor& weight, framework::Tensor* output); const framework::Tensor& weight, framework::Tensor* output);
......
src/operators/math/gemm.cpp
浏览文件 @ 5ac1e63c
...@@ -3662,7 +3662,7 @@ void Gemm::AddDot6x8(int k, const float *a, const float *b, float *c, int ldc) { ...@@ -3662,7 +3662,7 @@ void Gemm::AddDot6x8(int k, const float *a, const float *b, float *c, int ldc) {
b_ptr = b; b_ptr = b;
int kc1 = k / 8; int kc1 = k / 8;
int kc2 = k % 8; int kc2 = k % 8;
int step = 4 * ldc; int step = sizeof(float) * ldc;
asm volatile( asm volatile(
"pld [%[a_ptr]] \n\t" "pld [%[a_ptr]] \n\t"
"pld [%[a_ptr], #64] \n\t" "pld [%[a_ptr], #64] \n\t"
...@@ -3866,11 +3866,10 @@ void Gemm::AddDot6x8(int k, const float *a, const float *b, float *c, int ldc) { ...@@ -3866,11 +3866,10 @@ void Gemm::AddDot6x8(int k, const float *a, const float *b, float *c, int ldc) {
: :
: [a_ptr] "r"(a_ptr), [b_ptr] "r"(b_ptr), [c] "r"(c), [kc1] "r"(kc1), : [a_ptr] "r"(a_ptr), [b_ptr] "r"(b_ptr), [c] "r"(c), [kc1] "r"(kc1),
[kc2] "r"(kc2), [step] "r"(step) [kc2] "r"(kc2), [step] "r"(step)
: "memory", "r5", "r6", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", : "cc", "memory", "r5", "r6", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
"q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"); "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
#endif // __aarch64__ #endif // __aarch64__
#else
#endif // __ARM_NEON #endif // __ARM_NEON
} }
......
src/operators/math/gemm.h
浏览文件 @ 5ac1e63c
...@@ -96,6 +96,7 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb, ...@@ -96,6 +96,7 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
void InnerKernelWithPRelu(int mc, int nc, const float *a, const float *b, void InnerKernelWithPRelu(int mc, int nc, const float *a, const float *b,
float *c, float *C, int ldc, float *p, float *c, float *C, int ldc, float *p,
std::string mode, float *bias, float *bias1); std::string mode, float *bias, float *bias1);
/* /*
// 向量矩阵乘法 (M = 1) // 向量矩阵乘法 (M = 1)
void VectorKernel(int m, int n, int k, float alpha, const float *A, int lda, void VectorKernel(int m, int n, int k, float alpha, const float *A, int lda,
...@@ -139,6 +140,7 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb, ...@@ -139,6 +140,7 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
float *new_scale, float *new_bias); float *new_scale, float *new_bias);
void WriteWithBnAddRelu(int mc, int nc, float *c, float *C, int ldc, void WriteWithBnAddRelu(int mc, int nc, float *c, float *C, int ldc,
float *new_scale, float *new_bias, float *bias1); float *new_scale, float *new_bias, float *bias1);
/* /*
// 向量矩阵乘法结果回写 // 向量矩阵乘法结果回写
// C = A * B // C = A * B
...@@ -185,15 +187,63 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb, ...@@ -185,15 +187,63 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
const float *B, int ldb, float *C, int ldc, float *p, const float *B, int ldb, float *C, int ldc, float *p,
std::string mode, float *bias, float *bias1); std::string mode, float *bias, float *bias1);
/************************ 8 bit function cluster ************************/
// 8 bit int small block inner product
void AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
int32_t ldc);
// 8 bit int inner product
void InnerKernelWithBias(int32_t mc, int32_t nc, int8_t alpha,
const int8_t *a, const int8_t *b, int8_t beta,
int32_t *c, int32_t *C, int32_t ldc, bool relu,
int8_t *bias);
// 8 bit int pack function
void PackMatrixA_6r(int32_t m, int32_t k, int32_t m_tail, const int8_t *A,
int32_t lda, int8_t *buffer);
void PackMatrixB_8c(int32_t k, int32_t n, int32_t n_tail, const int8_t *B,
int32_t ldb, int8_t *buffer);
// 8 bit int matrix product
void Sgemm(int32_t m, int32_t n, int32_t k, int8_t alpha, const int8_t *A,
int32_t lda, const int8_t *B, int32_t ldb, int8_t beta, int32_t *C,
int32_t ldc, bool relu, int8_t *bias);
// 8 bit int write back
// C = alpha * A * B + beta * C
void WriteWithAlphaBeta(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
int32_t ldc);
// C = A * B
void WriteBasic(int32_t mc, int32_t nc, int32_t *c, int32_t *C, int32_t ldc);
// C = A * B + C
void WriteWithAdd(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
int32_t ldc);
// C = A * B + bias
void WriteWithAddV1(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
int32_t ldc, int8_t *bias);
// C = A * B + C, relu(C)
void WriteWithAddRelu(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
int32_t ldc);
// C = A * B + bias, relu(C)
void WriteWithAddReluV1(int32_t mc, int32_t nc, int32_t *c, int32_t *C,
int32_t ldc, int8_t *bias);
private: private:
int MC = 0; int MC = 0;
int KC = 0; int KC = 0;
int NC = 0; int NC = 0;
// 32位 float
float *packedA; float *packedA;
float *packedB; float *packedB;
float *packedC; float *packedC;
float *zero; float *zero;
// 8 bit int
int8_t *packedA_int8;
int8_t *packedB_int8;
int32_t *packedC_int8;
int8_t *zero_int8;
}; };
} // namespace math } // namespace math
......
src/operators/math/gemm_int8.cpp 0 → 100644
浏览文件 @ 5ac1e63c
此差异已折叠。
src/operators/math/im2col.cpp
浏览文件 @ 5ac1e63c
此差异已折叠。
src/operators/math/math_function.cpp
浏览文件 @ 5ac1e63c
...@@ -135,7 +135,7 @@ template <typename T> ...@@ -135,7 +135,7 @@ template <typename T>
struct ClearTensor<CPU, T> { struct ClearTensor<CPU, T> {
void operator()(framework::Tensor *tensor) { void operator()(framework::Tensor *tensor) {
auto size = tensor->numel(); auto size = tensor->numel();
auto *tensor_data = tensor->data<float>(); auto *tensor_data = tensor->data<T>();
memset((void *)tensor_data, 0, sizeof(T) * size); // NOLINT memset((void *)tensor_data, 0, sizeof(T) * size); // NOLINT
} }
}; };
...@@ -151,9 +151,9 @@ struct RowwiseAdd<CPU, T> { ...@@ -151,9 +151,9 @@ struct RowwiseAdd<CPU, T> {
PADDLE_MOBILE_ENFORCE((output->dims() == in_dims), PADDLE_MOBILE_ENFORCE((output->dims() == in_dims),
"output->dims() must be equal to in_dims."); "output->dims() must be equal to in_dims.");
auto *input_data = input.data<float>(); auto *input_data = input.data<T>();
auto *out_data = output->data<float>(); auto *out_data = output->data<T>();
auto *vec_data = vector.data<float>(); auto *vec_data = vector.data<T>();
for (int64_t i = 0; i < in_dims[0]; ++i) { for (int64_t i = 0; i < in_dims[0]; ++i) {
for (int64_t j = 0; j < size; ++j) { for (int64_t j = 0; j < size; ++j) {
out_data[i * size + j] = input_data[i * size + j] + vec_data[j]; out_data[i * size + j] = input_data[i * size + j] + vec_data[j];
......
src/operators/math/math_function.h
浏览文件 @ 5ac1e63c
...@@ -15,6 +15,7 @@ limitations under the License. */ ...@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once #pragma once
#include <cmath> #include <cmath>
#include <string>
#include "framework/tensor.h" #include "framework/tensor.h"
namespace paddle_mobile { namespace paddle_mobile {
...@@ -25,7 +26,7 @@ template <typename T> ...@@ -25,7 +26,7 @@ template <typename T>
void matmul(const framework::Tensor &matrix_a, bool trans_a, void matmul(const framework::Tensor &matrix_a, bool trans_a,
const framework::Tensor &matrix_b, bool trans_b, T alpha, const framework::Tensor &matrix_b, bool trans_b, T alpha,
framework::Tensor *matrix_out, T beta, bool relu = false, framework::Tensor *matrix_out, T beta, bool relu = false,
float *bias = nullptr); T *bias = nullptr);
template <typename T> template <typename T>
void matmulWithBn(const framework::Tensor &matrix_a, bool trans_a, void matmulWithBn(const framework::Tensor &matrix_a, bool trans_a,
......
src/operators/math/math_function_int8.cpp 0 → 100644
浏览文件 @ 5ac1e63c
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <cstring>
#include <string>
#include "operators/math/gemm.h"
#include "operators/math/math_function.h"
namespace paddle_mobile {
namespace operators {
namespace math {
template <>
void matmul<int8_t>(const framework::Tensor &matrix_a, bool trans_a,
const framework::Tensor &matrix_b, bool trans_b,
int8_t alpha, framework::Tensor *matrix_out, int8_t beta,
bool relu, int8_t *bias) {
auto dim_a = matrix_a.dims();
auto dim_b = matrix_b.dims();
auto dim_out = matrix_out->dims();
PADDLE_MOBILE_ENFORCE(
dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
"The input and output of matmul be matrix");
int32_t M = dim_out[0];
int32_t N = dim_out[1];
int32_t K = (!trans_a) ? dim_a[1] : dim_a[0];
Gemm gemm;
if (trans_a) {
int32_t numel = matrix_a.numel();
int32_t m = matrix_a.dims()[0];
int32_t n = matrix_a.dims()[1];
int8_t *tmp = (int8_t *)(matrix_a.data<int8_t>()); // NOLINT
int8_t *a = static_cast<int8_t *>(
paddle_mobile::memory::Alloc(sizeof(int8_t) * numel));
int32_t index = 0;
for (int32_t j = 0; j < n; j++) {
for (int32_t i = 0; i < m; i++) {
a[index++] = tmp[i * n + j];
}
}
gemm.Sgemm(M, N, K, alpha, a, K, matrix_b.data<int8_t>(), N, beta,
matrix_out->data<int32_t>(), N, relu, bias);
} else {
gemm.Sgemm(M, N, K, alpha, matrix_a.data<int8_t>(), K,
matrix_b.data<int8_t>(), N, beta, matrix_out->data<int32_t>(), N,
relu, bias);
}
}
} // namespace math
} // namespace operators
} // namespace paddle_mobile
src/operators/math/vol2col.cpp
浏览文件 @ 5ac1e63c
...@@ -32,9 +32,6 @@ class Vol2ColFunctor<CPU, T> { ...@@ -32,9 +32,6 @@ class Vol2ColFunctor<CPU, T> {
void operator()(const Tensor &vol, const std::vector<int> &dilations, void operator()(const Tensor &vol, const std::vector<int> &dilations,
const std::vector<int> &strides, const std::vector<int> &strides,
const std::vector<int> &paddings, Tensor *col) const { const std::vector<int> &paddings, Tensor *col) const {
// PADDLE_ENFORCE(vol.dims().size() == 4);
// PADDLE_ENFORCE(col->dims().size() == 7);
int input_channels = vol.dims()[0]; int input_channels = vol.dims()[0];
int input_depth = vol.dims()[1]; int input_depth = vol.dims()[1];
int input_height = vol.dims()[2]; int input_height = vol.dims()[2];
...@@ -48,32 +45,6 @@ class Vol2ColFunctor<CPU, T> { ...@@ -48,32 +45,6 @@ class Vol2ColFunctor<CPU, T> {
int channels_col = int channels_col =
input_channels * filter_depth * filter_height * filter_width; input_channels * filter_depth * filter_height * filter_width;
// PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] -
// ((dilations[0] * (filter_depth - 1)
// + 1))) /
// strides[0] +
// 1,
// output_depth,
// "input_depth and output_depth are "
// "mismatching.");
// PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] -
// ((dilations[1] * (filter_height -
// 1) + 1))) /
// strides[1] +
// 1,
// output_height,
// "input_height and output_height are
// "
// "mismatching.");
// PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] -
// ((dilations[2] * (filter_width - 1)
// + 1))) /
// strides[2] +
// 1,
// output_width,
// "input_width and output_width are "
// "mismatching.");
const T *vol_data = vol.data<T>(); const T *vol_data = vol.data<T>();
T *col_data = col->data<T>(); T *col_data = col->data<T>();
...@@ -119,9 +90,6 @@ class Col2VolFunctor<CPU, T> { ...@@ -119,9 +90,6 @@ class Col2VolFunctor<CPU, T> {
void operator()(const Tensor &col, const std::vector<int> &dilations, void operator()(const Tensor &col, const std::vector<int> &dilations,
const std::vector<int> &strides, const std::vector<int> &strides,
const std::vector<int> &paddings, Tensor *vol) const { const std::vector<int> &paddings, Tensor *vol) const {
// PADDLE_ENFORCE(vol->dims().size() == 4);
// PADDLE_ENFORCE(col.dims().size() == 7);
int input_channels = vol->dims()[0]; int input_channels = vol->dims()[0];
int input_depth = vol->dims()[1]; int input_depth = vol->dims()[1];
int input_height = vol->dims()[2]; int input_height = vol->dims()[2];
...@@ -135,31 +103,6 @@ class Col2VolFunctor<CPU, T> { ...@@ -135,31 +103,6 @@ class Col2VolFunctor<CPU, T> {
int channels_col = int channels_col =
input_channels * filter_depth * filter_height * filter_width; input_channels * filter_depth * filter_height * filter_width;
// PADDLE_ENFORCE_EQ((input_depth + 2 * paddings[0] -
// ((dilations[0] * (filter_depth - 1)
// + 1))) /
// strides[0] +
// 1,
// output_depth,
// "input_depth and output_depth are "
// "mismatching.");
// PADDLE_ENFORCE_EQ((input_height + 2 * paddings[1] -
// ((dilations[1] * (filter_height -
// 1) + 1))) /
// strides[1] +
// 1,
// output_height,
// "input_height and output_height are
// "
// "mismatching.");
// PADDLE_ENFORCE_EQ((input_width + 2 * paddings[2] -
// ((dilations[2] * (filter_width - 1)
// + 1))) /
// strides[2] +
// 1,
// output_width,
// "input_width and output_width are "
// "mismatching.");
T *vol_data = vol->data<T>(); T *vol_data = vol->data<T>();
const T *col_data = col.data<T>(); const T *col_data = col.data<T>();
...@@ -195,9 +138,9 @@ class Col2VolFunctor<CPU, T> { ...@@ -195,9 +138,9 @@ class Col2VolFunctor<CPU, T> {
}; };
template class Vol2ColFunctor<CPU, float>; template class Vol2ColFunctor<CPU, float>;
template class Vol2ColFunctor<CPU, double>; template class Vol2ColFunctor<CPU, int8_t>;
template class Col2VolFunctor<CPU, float>; template class Col2VolFunctor<CPU, float>;
template class Col2VolFunctor<CPU, double>; template class Col2VolFunctor<CPU, int8_t>;
} // namespace math } // namespace math
} // namespace operators } // namespace operators
......
src/operators/op_param.h
浏览文件 @ 5ac1e63c
...@@ -2150,14 +2150,12 @@ class QuantizeParam : public OpParam { ...@@ -2150,14 +2150,12 @@ class QuantizeParam : public OpParam {
const AttributeMap &attrs, const Scope &scope) { const AttributeMap &attrs, const Scope &scope) {
input_ = InputXFrom<GType>(inputs, scope); input_ = InputXFrom<GType>(inputs, scope);
out_ = OutFrom<GType>(outputs, scope); out_ = OutFrom<GType>(outputs, scope);
if (HasAttr("is_static", attrs)) {
is_static_ = GetAttr<bool>("is_static", attrs);
}
// online // online
// scale = max(abs(x)) // scale = max(abs(x))
online_scale_ = GetVarValue<GType>("OutScale", outputs, scope); online_scale_ = GetVarValue<GType>("OutScale", outputs, scope);
// offline // offline
if (HasAttr("static_scale", attrs)) { if (HasAttr("static_scale", attrs)) {
is_static_ = true;
static_scale_ = GetAttr<float>("static_scale", attrs); static_scale_ = GetAttr<float>("static_scale", attrs);
} }
// x = round(scale * x) // x = round(scale * x)
...@@ -2179,7 +2177,7 @@ class QuantizeParam : public OpParam { ...@@ -2179,7 +2177,7 @@ class QuantizeParam : public OpParam {
float static_scale_ = 1.0f; float static_scale_ = 1.0f;
// round method type // round method type
// nearest_zero and nearest_even is valid currently // nearest_zero and nearest_even is valid currently
RoundType round_type_ = ROUND_NEAREST_TO_EVEN; RoundType round_type_ = ROUND_NEAREST_AWAY_ZERO;
}; };
template <typename Dtype> template <typename Dtype>
......
test/CMakeLists.txt
浏览文件 @ 5ac1e63c
...@@ -258,6 +258,10 @@ if (NOT FOUND_MATCH) ...@@ -258,6 +258,10 @@ if (NOT FOUND_MATCH)
ADD_EXECUTABLE(test-gemm-accuracy common/test_gemm_accuracy.cpp) ADD_EXECUTABLE(test-gemm-accuracy common/test_gemm_accuracy.cpp)
target_link_libraries(test-gemm-accuracy paddle-mobile) target_link_libraries(test-gemm-accuracy paddle-mobile)
# gen test
ADD_EXECUTABLE(test-gemm-int8-accuracy common/test_gemm_int8_accuracy.cpp)
target_link_libraries(test-gemm-int8-accuracy paddle-mobile)
# gen test # gen test
ADD_EXECUTABLE(test-gemm-perf common/test_gemm_perf.cpp) ADD_EXECUTABLE(test-gemm-perf common/test_gemm_perf.cpp)
target_link_libraries(test-gemm-perf paddle-mobile) target_link_libraries(test-gemm-perf paddle-mobile)
......
test/common/test_gemm_accuracy.cpp
浏览文件 @ 5ac1e63c
...@@ -84,7 +84,7 @@ int do_sgemm(int m, int n, int k, bool relu, int t1, int t2, int pr) { ...@@ -84,7 +84,7 @@ int do_sgemm(int m, int n, int k, bool relu, int t1, int t2, int pr) {
} }
paddle_mobile::operators::math::Gemm gemm; paddle_mobile::operators::math::Gemm gemm;
gemm.SgemmWithBn(m, n, k, 0.9, a, lda, b, ldb, 0.3, c, ldc, relu, scale, bias, gemm.SgemmWithBn(m, n, k, 1, a, lda, b, ldb, 0.3, c, ldc, relu, scale, bias,
nullptr); nullptr);
int eq = 0; int eq = 0;
int neq = 0; int neq = 0;
......
test/common/test_gemm_int8_accuracy.cpp 0 → 100644
浏览文件 @ 5ac1e63c
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <cstdlib>
#include <ctime>
#include <iostream>
#include <random>
#include "../test_helper.h"
#include "common/log.h"
#include "memory/t_malloc.h"
#include "operators/math/gemm.h"
#define a(i, j) a[(i)*lda + (j)]
#define b(i, j) b[(i)*ldb + (j)]
#define c(i, j) c[(i)*ldc + (j)]
#define c1(i, j) c1[(i)*ldc + (j)]
using std::default_random_engine;
using std::uniform_int_distribution;
void print_matirx(int m, int n, int ldc, int32_t *c) {
for (int i = 0; i < m; ++i) {
std::cout << c(i, 0);
for (int j = 1; j < n; ++j) {
std::cout << " | " << c(i, j);
}
std::cout << std::endl;
}
std::cout << std::endl;
}
void print_matirx(int m, int n, int ldc, int8_t *c) {
for (int i = 0; i < m; ++i) {
std::cout << static_cast<int32_t>(c(i, 0));
for (int j = 1; j < n; ++j) {
std::cout << " | " << static_cast<int32_t>(c(i, j));
}
std::cout << std::endl;
}
std::cout << std::endl;
}
int do_sgemm(int m, int n, int k, bool relu, int pr) {
int lda = k;
int ldb = n;
int ldc = n;
default_random_engine e;
uniform_int_distribution<int8_t> pixel(-127, 127);
int8_t *a = static_cast<int8_t *>(
paddle_mobile::memory::Alloc(sizeof(int8_t) * m * k));
int8_t *b = static_cast<int8_t *>(
paddle_mobile::memory::Alloc(sizeof(int8_t) * k * n));
int32_t *c = static_cast<int32_t *>(
paddle_mobile::memory::Alloc(sizeof(int32_t) * m * n));
int32_t *c1 = static_cast<int32_t *>(
paddle_mobile::memory::Alloc(sizeof(int32_t) * m * n));
for (int i = 0; i < m * k; ++i) {
a[i] = pixel(e);
}
for (int i = 0; i < k * n; ++i) {
b[i] = pixel(e);
}
for (int i = 0; i < m; ++i) {
for (int j = 0; j < n; ++j) {
int32_t r = 0;
for (int p = 0; p < k; p++) {
r += static_cast<int32_t>(a(i, p)) * static_cast<int32_t>(b(p, j));
}
c1(i, j) = r;
}
}
paddle_mobile::operators::math::Gemm gemm;
gemm.Sgemm(m, n, k, static_cast<int8_t>(1), a, lda, b, ldb,
static_cast<int8_t>(0), c, ldc, relu, nullptr);
int eq = 0;
int neq = 0;
for (int i = 0; i < m * n; ++i) {
if (c[i] == c1[i]) {
++eq;
} else {
++neq;
}
}
if (pr > 0) {
std::cout << "A:" << std::endl;
print_matirx(m, k, lda, a);
std::cout << "B:" << std::endl;
print_matirx(k, n, ldb, b);
std::cout << "C:" << std::endl;
print_matirx(m, n, ldc, c);
std::cout << "C1:" << std::endl;
print_matirx(m, n, ldc, c1);
}
std::cout << "mnk=" << m << " " << n << " " << k << " relu=" << relu
<< " eq=" << eq << " neq=" << neq << std::endl;
paddle_mobile::memory::Free(a);
paddle_mobile::memory::Free(b);
paddle_mobile::memory::Free(c);
paddle_mobile::memory::Free(c1);
return 0;
}
int main() {
do_sgemm(9, 9, 9, false, 10);
do_sgemm(10, 6, 12, false, 0);
do_sgemm(512, 256, 384, false, 0);
do_sgemm(1366, 768, 256, false, 0);
do_sgemm(1255, 755, 333, false, 0);
do_sgemm(555, 777, 999, false, 0);
do_sgemm(1024, 1024, 1024, false, 0);
return 0;
}
test/common/test_gemm_perf.cpp
浏览文件 @ 5ac1e63c
...@@ -28,13 +28,11 @@ limitations under the License. */ ...@@ -28,13 +28,11 @@ limitations under the License. */
int main() { int main() {
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile; paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
paddle_mobile.SetThreadNum(4); paddle_mobile.SetThreadNum(1);
Tensor aa, bb, cc, scale, bias; Tensor aa, bb, cc;
auto aaptr = aa.mutable_data<float>({m, k}); auto aaptr = aa.mutable_data<float>({m, k});
auto bbptr = bb.mutable_data<float>({k, n}); auto bbptr = bb.mutable_data<float>({k, n});
auto ccptr = cc.mutable_data<float>({m, n}); auto ccptr = cc.mutable_data<float>({m, n});
auto scaleptr = scale.mutable_data<float>({m});
auto biasptr = bias.mutable_data<float>({m});
for (int i = 0; i < m * k; ++i) { for (int i = 0; i < m * k; ++i) {
aaptr[i] = 2; aaptr[i] = 2;
...@@ -45,23 +43,55 @@ int main() { ...@@ -45,23 +43,55 @@ int main() {
for (int i = 0; i < m * n; ++i) { for (int i = 0; i < m * n; ++i) {
ccptr[i] = 2; ccptr[i] = 2;
} }
for (int i = 0; i < m; ++i) {
scaleptr[i] = 1; Tensor aa_int8, bb_int8, cc_int8;
biasptr[i] = 0; auto aaptr_int8 = aa_int8.mutable_data<int8_t>({m, k});
auto bbptr_int8 = bb_int8.mutable_data<int8_t>({k, n});
auto ccptr_int8 = cc_int8.mutable_data<int32_t>({m, n});
for (int i = 0; i < m * k; ++i) {
aaptr_int8[i] = static_cast<int8_t>(2);
}
for (int i = 0; i < k * n; ++i) {
bbptr_int8[i] = static_cast<int8_t>(2);
}
for (int i = 0; i < m * n; ++i) {
ccptr_int8[i] = static_cast<int32_t>(2);
} }
auto time1 = time(); // float
// warm-up 10 times
for (int j = 0; j < 10; ++j) { for (int j = 0; j < 10; ++j) {
paddle_mobile::operators::math::matmul<float>( paddle_mobile::operators::math::matmul<float>(
aa, false, bb, false, static_cast<float>(1), &cc, static_cast<float>(0), aa, false, bb, false, static_cast<float>(1), &cc, static_cast<float>(0),
false, biasptr); false, nullptr);
}
// paddle_mobile::operators::math::matmulWithBn<float>( auto time1 = time();
// aa, false, bb, false, static_cast<float>(1), &cc, for (int j = 0; j < 10; ++j) {
// static_cast<float>(0), true, &scale, &bias, 0); paddle_mobile::operators::math::matmul<float>(
aa, false, bb, false, static_cast<float>(1), &cc, static_cast<float>(0),
false, nullptr);
} }
auto time2 = time(); auto time2 = time();
std::cout << "gemm cost :" << time_diff(time1, time2) / 10 << "ms\n"; std::cout << "float gemm cost :" << time_diff(time1, time2) / 10 << "ms\n";
// int8_t
// warm-up 10 times
for (int j = 0; j < 10; ++j) {
paddle_mobile::operators::math::matmul<int8_t>(
aa_int8, false, bb_int8, false, static_cast<int8_t>(1), &cc_int8,
static_cast<int8_t>(0), false, nullptr);
}
auto time3 = time();
for (int j = 0; j < 10; ++j) {
paddle_mobile::operators::math::matmul<int8_t>(
aa_int8, false, bb_int8, false, static_cast<int8_t>(1), &cc_int8,
static_cast<int8_t>(0), false, nullptr);
}
auto time4 = time();
std::cout << "int8_t gemm cost :" << time_diff(time3, time4) / 10 << "ms\n";
return 0; return 0;
} }
test/net/test_googlenet.cpp
浏览文件 @ 5ac1e63c
...@@ -25,27 +25,31 @@ int main() { ...@@ -25,27 +25,31 @@ int main() {
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile; paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
#endif #endif
paddle_mobile.SetThreadNum(4); paddle_mobile.SetThreadNum(1);
bool optimize = true; bool optimize = false;
auto time1 = time(); auto time1 = time();
if (paddle_mobile.Load(g_googlenet, optimize)) { if (paddle_mobile.Load(g_googlenet, optimize)) {
auto time2 = time(); auto time2 = time();
std::cout << "load cost :" << time_diff(time1, time2) << "ms" << std::endl; std::cout << "load cost :" << time_diff(time1, time2) << "ms" << std::endl;
std::vector<float> input; std::vector<float> input;
std::vector<float> output;
std::vector<int64_t> dims{1, 3, 224, 224}; std::vector<int64_t> dims{1, 3, 224, 224};
GetInput<float>(g_test_image_1x3x224x224, &input, dims); GetInput<float>(g_test_image_1x3x224x224, &input, dims);
// 预热十次 // // 预热十次
for (int i = 0; i < 10; ++i) { // for (int i = 0; i < 10; ++i) {
auto vec_result = paddle_mobile.Predict(input, dims); // output = paddle_mobile.Predict(input, dims);
} // }
auto time3 = time(); auto time3 = time();
for (int i = 0; i < 10; ++i) { for (int i = 0; i < 10; ++i) {
auto vec_result = paddle_mobile.Predict(input, dims); output = paddle_mobile.Predict(input, dims);
} }
auto time4 = time(); auto time4 = time();
std::cout << "predict cost :" << time_diff(time3, time4) / 10 << "ms" std::cout << "predict cost :" << time_diff(time3, time4) / 10 << "ms"
<< std::endl; << std::endl;
for (int i = 0; i < output.size(); ++i) {
DLOG << "result[" << i << "] = " << output[i];
}
} }
return 0; return 0;
} }
test/operators/test_dequantize_op.cpp
浏览文件 @ 5ac1e63c
...@@ -59,7 +59,7 @@ int TestDequqntizeOp() { ...@@ -59,7 +59,7 @@ int TestDequqntizeOp() {
framework::Tensor output_cmp; framework::Tensor output_cmp;
output_cmp.Resize(dim); output_cmp.Resize(dim);
float dequant_scale = 1.f / (1.27 * 1.74); float dequant_scale = 1.27 / 1.74;
dequantize(input, dequant_scale, &output_cmp); dequantize(input, dequant_scale, &output_cmp);
const float* output_cmp_data = output_cmp.data<float>(); const float* output_cmp_data = output_cmp.data<float>();
for (int i = 0; i < output->numel(); ++i) { for (int i = 0; i < output->numel(); ++i) {
......
test/operators/test_int8_conv_op.cpp
浏览文件 @ 5ac1e63c
...@@ -140,10 +140,10 @@ int TestConvOp() { ...@@ -140,10 +140,10 @@ int TestConvOp() {
int dilation_w = 1; int dilation_w = 1;
int batch_size = 1; int batch_size = 1;
int input_c = 3; int input_c = 63;
int input_h = 25; int input_h = 51;
int input_w = 25; int input_w = 51;
int output_c = 3; int output_c = 125;
framework::DDim input_shape = framework::DDim input_shape =
framework::make_ddim({batch_size, input_c, input_h, input_w}); framework::make_ddim({batch_size, input_c, input_h, input_w});
framework::DDim filter_shape = framework::DDim filter_shape =
...@@ -158,11 +158,11 @@ int TestConvOp() { ...@@ -158,11 +158,11 @@ int TestConvOp() {
auto input_var = scope.get()->Var("input"); auto input_var = scope.get()->Var("input");
auto input = input_var->template GetMutable<framework::LoDTensor>(); auto input = input_var->template GetMutable<framework::LoDTensor>();
SetupTensor<Itype>(input, input_shape, -127, 127); SetupTensor<Itype>(input, input_shape, -20, 20);
auto filter_var = scope.get()->Var("filter"); auto filter_var = scope.get()->Var("filter");
auto filter = filter_var->template GetMutable<framework::LoDTensor>(); auto filter = filter_var->template GetMutable<framework::LoDTensor>();
SetupTensor<Itype>(filter, filter_shape, -127, 127); SetupTensor<Itype>(filter, filter_shape, -20, 20);
auto output_var = scope.get()->Var("output"); auto output_var = scope.get()->Var("output");
framework::AttributeMap attrs; framework::AttributeMap attrs;
...@@ -174,28 +174,40 @@ int TestConvOp() { ...@@ -174,28 +174,40 @@ int TestConvOp() {
auto *op = new operators::ConvOp<CPU, float>("conv2d", inputs, outputs, attrs, auto *op = new operators::ConvOp<CPU, float>("conv2d", inputs, outputs, attrs,
scope); scope);
struct timespec ts_begin, ts_end;
op->InferShape(); op->InferShape();
// warmup
op->Run(); op->Run();
clock_gettime(CLOCK_MONOTONIC, &ts_begin);
int kernel_extent_h = dilation_h * (kernel_h - 1) + 1; for (int i = 0; i < 10; ++i) {
int kernel_extent_w = dilation_w * (kernel_w - 1) + 1; op->Run();
int output_h = (input_h + 2 * pad_h - kernel_extent_h) / stride_h + 1;
int output_w = (input_w + 2 * pad_w - kernel_extent_w) / stride_w + 1;
auto output_shape = framework::make_ddim(
std::vector<int>({batch_size, output_c, output_h, output_w}));
framework::Tensor output_cmp;
output_cmp.mutable_data<Otype>(output_shape);
conv2d<Itype, Otype>(input, filter, attrs, &output_cmp);
// compare results
auto output = output_var->template Get<framework::LoDTensor>();
const Otype *output_data = output->data<Otype>();
Otype *output_cmp_data = output_cmp.data<Otype>();
for (int i = 0; i < output->numel(); ++i) {
PADDLE_MOBILE_ENFORCE(output_data[i] == output_cmp_data[i],
"output[%d] = %d, output_cmp[%d] = %d", i,
output_data[i], i, output_cmp_data[i]);
} }
clock_gettime(CLOCK_MONOTONIC, &ts_end);
uint64_t elapsed = (ts_end.tv_sec - ts_begin.tv_sec) * 1e3 +
(ts_end.tv_nsec - ts_begin.tv_nsec) / 1e6;
LOG(kLOG_INFO) << "elapsed: " << elapsed / 10.0 << " ms";
/*
int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
int output_h = (input_h + 2 * pad_h - kernel_extent_h) / stride_h + 1;
int output_w = (input_w + 2 * pad_w - kernel_extent_w) / stride_w + 1;
auto output_shape = framework::make_ddim(
std::vector<int>({batch_size, output_c, output_h, output_w}));
framework::Tensor output_cmp;
output_cmp.mutable_data<Otype>(output_shape);
conv2d<Itype, Otype>(input, filter, attrs, &output_cmp);
// compare results
auto output = output_var->template Get<framework::LoDTensor>();
const Otype *output_data = output->data<Otype>();
Otype *output_cmp_data = output_cmp.data<Otype>();
for (int i = 0; i < output->numel(); ++i) {
PADDLE_MOBILE_ENFORCE(output_data[i] == output_cmp_data[i],
"output[%d] = %d, output_cmp[%d] = %d", i,
output_data[i], i, output_cmp_data[i]);
}
*/
delete op; delete op;
return 0; return 0;
} }
...@@ -203,12 +215,42 @@ int TestConvOp() { ...@@ -203,12 +215,42 @@ int TestConvOp() {
} // namespace paddle_mobile } // namespace paddle_mobile
int main() { int main() {
// kernel = 7, pad = 0, stride = 2
LOG(paddle_mobile::kLOG_INFO) << "int8, kernel=7, pad=0, stride=2";
paddle_mobile::TestConvOp<int8_t, int32_t, 7, 0, 2>();
// kernel = 7, pad = 3, stride = 2
LOG(paddle_mobile::kLOG_INFO) << "int8, kernel=7, pad=3, stride=2";
paddle_mobile::TestConvOp<int8_t, int32_t, 7, 3, 2>();
// kernel = 3, pad = 0, stride = 1 // kernel = 3, pad = 0, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "int8, kernel=3, pad=0, stride=1";
paddle_mobile::TestConvOp<int8_t, int32_t, 3, 0, 1>(); paddle_mobile::TestConvOp<int8_t, int32_t, 3, 0, 1>();
// kernel = 3, pad = 0, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "float, kernel=3, pad=0, stride=1";
paddle_mobile::TestConvOp<float, float, 3, 0, 1>();
LOG(paddle_mobile::kLOG_INFO) << "\n";
// kernel = 3, pad = 1, stride = 1 // kernel = 3, pad = 1, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "int8, kernel=3, pad=1, stride=1";
paddle_mobile::TestConvOp<int8_t, int32_t, 3, 1, 1>(); paddle_mobile::TestConvOp<int8_t, int32_t, 3, 1, 1>();
// kernel = 3, pad = 1, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "float, kernel=3, pad=1, stride=1";
paddle_mobile::TestConvOp<float, float, 3, 1, 1>();
LOG(paddle_mobile::kLOG_INFO) << "\n";
// kernel = 5, pad = 0, stride = 1 // kernel = 5, pad = 0, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "int8, kernel=5, pad=0, stride=1";
paddle_mobile::TestConvOp<int8_t, int32_t, 5, 0, 1>(); paddle_mobile::TestConvOp<int8_t, int32_t, 5, 0, 1>();
// kernel = 5, pad = 0, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "float, kernel=5, pad=0, stride=1";
paddle_mobile::TestConvOp<float, float, 5, 0, 1>();
LOG(paddle_mobile::kLOG_INFO) << "\n";
// kernel = 5, pad = 2, stride = 1 // kernel = 5, pad = 2, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "int8, kernel=5, pad=2, stride=1";
paddle_mobile::TestConvOp<int8_t, int32_t, 5, 2, 1>(); paddle_mobile::TestConvOp<int8_t, int32_t, 5, 2, 1>();
// kernel = 5, pad = 2, stride = 1
LOG(paddle_mobile::kLOG_INFO) << "float, kernel=5, pad=2, stride=1";
paddle_mobile::TestConvOp<float, float, 5, 2, 1>();
} }
test/operators/test_mul_op.cpp
浏览文件 @ 5ac1e63c
...@@ -12,80 +12,80 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,80 +12,80 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "../test_helper.h"
#include "../test_include.h" #include "../test_include.h"
#include "operators/mul_op.h" #include "operators/mul_op.h"
int main() { #define a(i, j) a[(i)*lda + (j)]
paddle_mobile::Loader<paddle_mobile::CPU> loader; #define b(i, j) b[(i)*ldb + (j)]
auto program = loader.Load(g_resnet); #define c(i, j) c[(i)*ldc + (j)]
PADDLE_MOBILE_ENFORCE(program.originProgram != nullptr,
"program file read fail"); namespace paddle_mobile {
using framework::AttributeMap;
Executor4Test<paddle_mobile::CPU, using framework::DDim;
paddle_mobile::operators::MulOp<paddle_mobile::CPU, float>> using framework::Scope;
executor(program, "mul"); using framework::make_ddim;
template <typename I, typename O>
// 1. input_tensors; int TestMulOP() {
vector<Tensor> input_tensors; int32_t m = 1024;
int32_t n = 1024;
Tensor input1; int32_t k = 1024;
auto input1_data = CreateInput<float>(&input1, {3, 2, 1, 1}, 0, 1); int32_t lda = k;
input_tensors.push_back(input1); int32_t ldb = n;
Tensor input2; int32_t ldc = n;
auto input2_data = CreateInput<float>(&input2, {2, 3}, 0, 1); DDim inputA_shape = make_ddim({m, k});
input_tensors.push_back(input2); DDim inputB_shape = make_ddim({k, n});
VariableNameMap inputs;
// 2. input_names VariableNameMap outputs;
vector<string> input_names({ auto scope = std::make_shared<Scope>();
"pool2d_0.tmp_0", inputs["X"] = std::vector<std::string>({"inputA"});
"fc_0.w_0", inputs["Y"] = std::vector<std::string>({"inputB"});
}); outputs["Out"] = std::vector<std::string>({"output"});
// 3. output_names auto inputA_var = scope.get()->Var("inputA");
vector<string> output_names({"fc_0.tmp_0"}); auto inputA = inputA_var->template GetMutable<framework::LoDTensor>();
SetupTensor<I>(inputA, inputA_shape, -127, 127);
// 4. out_dims; auto inputB_var = scope.get()->Var("inputB");
vector<DDim> out_ddims; auto inputB = inputB_var->template GetMutable<framework::LoDTensor>();
auto out_ddim = paddle_mobile::framework::make_ddim({3, 3}); SetupTensor<I>(inputB, inputB_shape, -127, 127);
out_ddims.push_back(out_ddim);
auto output_var = scope.get()->Var("output");
auto output = executor.Predict<LoDTensor>(input_tensors, input_names, AttributeMap attrs;
output_names, out_ddims); attrs["x_num_col_dims"].Set<int>(1);
attrs["y_num_col_dims"].Set<int>(1);
auto output0_data = output[0]->data<float>(); auto *op =
new operators::MulOp<CPU, float>("mul", inputs, outputs, attrs, scope);
auto dim_1 = input1.numel() / input1.dims()[0]; op->InferShape();
DLOG << " input1 : "; op->Run();
for (int i = 0; i < input1.dims()[0]; ++i) { auto output = output_var->template Get<framework::LoDTensor>();
for (int j = 0; j < dim_1; ++j) { const O *output_data = output->data<O>();
DLOGF("%f ", input1_data[i * dim_1 + j]); // compare
O *c = static_cast<O *>(memory::Alloc(sizeof(O) * m * n));
I *a = inputA->data<I>();
I *b = inputB->data<I>();
for (int32_t i = 0; i < m; ++i) {
for (int32_t j = 0; j < n; ++j) {
O r = 0;
for (int32_t p = 0; p < k; p++) {
r += static_cast<O>(a(i, p)) * static_cast<O>(b(p, j));
}
c(i, j) = r;
} }
DLOGF("\n");
} }
auto dim_2 = input2.numel() / input2.dims()[0]; for (int32_t i = 0; i < m * n; ++i) {
DLOG << " input2 : "; PADDLE_MOBILE_ENFORCE(
for (int i = 0; i < input2.dims()[0]; ++i) { output_data[i] == c[i], "output[%d] = %d, output_cmp[%d] = %d", i,
for (int j = 0; j < dim_2; ++j) { static_cast<int32_t>(output_data[i]), i, static_cast<int32_t>(c[i]));
DLOGF("%f ", input2_data[i * dim_2 + j]);
}
DLOGF("\n");
}
auto dim_output0 = output[0]->numel() / output[0]->dims()[0];
DLOG << " output : ";
for (int i = 0; i < output[0]->dims()[0]; ++i) {
for (int j = 0; j < dim_output0; ++j) {
DLOGF("%f ", output0_data[i * dim_2 + j]);
}
DLOGF("\n");
} }
DLOG << "Run MulOp successfully!";
delete op;
return 0;
}
} // namespace paddle_mobile
/// output (3,3) int main() {
DLOG << "output memory size : " << output[0]->memory_size(); paddle_mobile::TestMulOP<int8_t, int32_t>();
DLOG << "output numel : " << output[0]->numel(); paddle_mobile::TestMulOP<float, float>();
DLOG << input1_data[0] << " x " << input2_data[0] << " + " << input1_data[1]
<< " x " << input2_data[0 + 3] << " = " << output0_data[0];
return 0; return 0;
} }
test/operators/test_quantize_op.cpp
浏览文件 @ 5ac1e63c
...@@ -18,14 +18,6 @@ limitations under the License. */ ...@@ -18,14 +18,6 @@ limitations under the License. */
namespace paddle_mobile { namespace paddle_mobile {
// static float g_test_data[50] = {
// -5.55, -5.5, -5.45, -5.0, -4.55, -4.5, -4.45, -4.0, -3.55, -3.5,
// -3.45, -3.01, -2.75, -2.5, -2.501, -2.49, -2.01, -1.75, -1.5, -1.25,
// -1.0, -0.75, -0.5, -0.25, 0.0, 0.25, 0.5, 0.75, 1.0, 1.25,
// 1.5, 1.75, 2.01, 2.49, 2.501, 2.5, 2.75, 3.01, 3.45, 3.5,
// 3.55, 4.0, 4.45, 4.5, 4.55, 5.0, 5.45, 5.5, 5.55, 6.0,
// };
static float find_abs_max(const Tensor *input) { static float find_abs_max(const Tensor *input) {
float max_abs = 0.f; float max_abs = 0.f;
const float *x = input->data<const float>(); const float *x = input->data<const float>();
...@@ -60,6 +52,16 @@ static void quantize_round_to_even(const Tensor *input, const float scale, ...@@ -60,6 +52,16 @@ static void quantize_round_to_even(const Tensor *input, const float scale,
} }
} }
static void quantize_round_to_nearest(const Tensor *input, const float scale,
Tensor *output) {
const float *x = input->data<const float>();
int8_t *y = output->mutable_data<int8_t>();
size_t size = input->numel();
for (size_t i = 0; i < size; ++i) {
y[i] = round(x[i] * scale);
}
}
int TestQuqntizeOp() { int TestQuqntizeOp() {
framework::DDim dim = framework::make_ddim({1, 3, 224, 224}); framework::DDim dim = framework::make_ddim({1, 3, 224, 224});
...@@ -88,15 +90,16 @@ int TestQuqntizeOp() { ...@@ -88,15 +90,16 @@ int TestQuqntizeOp() {
auto output_scale = output_scale_var->template Get<framework::LoDTensor>(); auto output_scale = output_scale_var->template Get<framework::LoDTensor>();
const float *output_scale_data = output_scale->data<float>(); const float *output_scale_data = output_scale->data<float>();
float max_abs = find_abs_max(input); float output_scale_cmp = find_abs_max(input);
float output_scale_cmp = 127 / max_abs;
PADDLE_MOBILE_ENFORCE(output_scale_cmp == output_scale_data[0], PADDLE_MOBILE_ENFORCE(output_scale_cmp == output_scale_data[0],
"output_scale = %.6f, output_scale_cmp = %.6f", "output_scale = %.6f, output_scale_cmp = %.6f",
output_scale_cmp, output_scale_data[0]); output_scale_cmp, output_scale_data[0]);
framework::Tensor output_cmp; framework::Tensor output_cmp;
output_cmp.Resize(dim); output_cmp.Resize(dim);
quantize_round_to_even(input, output_scale_cmp, &output_cmp); float scale = 127 / output_scale_cmp;
// quantize_round_to_even(input, scale, &output_cmp);
quantize_round_to_nearest(input, scale, &output_cmp);
int8_t *output_cmp_data = output_cmp.data<int8_t>(); int8_t *output_cmp_data = output_cmp.data<int8_t>();
for (int i = 0; i < output->numel(); ++i) { for (int i = 0; i < output->numel(); ++i) {
PADDLE_MOBILE_ENFORCE(output_data[i] == output_cmp_data[i], PADDLE_MOBILE_ENFORCE(output_data[i] == output_cmp_data[i],
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或