Skip to content

Mace
Mace

Xiaomi / Mace

通知 106
Star 40
Fork 27
Mace
Mace

体验新版 GitCode，发现更多精彩内容 >>

提交 fdb477e2 编写于 作者: 刘 刘托
浏览文件
操作

Merge branch 'refactor-eltwise' into 'master' 

Refactor eltwise op.

See merge request !428
上级 d5cbd799 25d2ad2d
展开全部 隐藏空白更改
内联 并排
Showing with 969 addition and 677 deletion
mace/kernels/eltwise.h
浏览文件 @ fdb477e2
......@@ -18,6 +18,7 @@
#include <algorithm>
#include <memory>
#include <vector>
#include <utility>
#include "mace/core/future.h"
#include "mace/core/tensor.h"
......@@ -30,216 +31,331 @@ namespace mace {
namespace kernels {
enum EltwiseType {
PROD = 0,
SUM = 1,
MAX = 2,
MIN = 3,
SUB = 4,
DIV = 5,
SUM = 0,
SUB = 1,
PROD = 2,
DIV = 3,
MIN = 4,
MAX = 5,
NEG = 6,
ABS = 7,
SQR_DIFF = 8,
NONE = 9,
};
struct EltwiseFunctorBase {
EltwiseFunctorBase(const EltwiseType type,
const std::vector<float> &coeff)
: type_(type), coeff_(coeff) {}
EltwiseType type_;
std::vector<float> coeff_;
};
template <DeviceType D, typename T>
struct EltwiseFunctor : EltwiseFunctorBase {
EltwiseFunctor(const EltwiseType type,
const std::vector<float> &coeff)
: EltwiseFunctorBase(type, coeff) {}
void operator()(const Tensor *input0,
const Tensor *input1,
const index_t start_axis,
const bool is_scaler,
const float value,
const bool swap,
Tensor *output,
StatsFuture *future) {
if (is_scaler) {
Tensor::MappingGuard input0_guard(input0);
Tensor::MappingGuard output_guard(output);
const T *input0_ptr = input0->data<T>();
T *output_ptr = output->mutable_data<T>();
const index_t num = input0->size();
switch (type_) {
case PROD:
inline void TensorScalar(const EltwiseType type,
const float *input0,
const float value,
const index_t size,
float *output) {
switch (type) {
case SUM:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = input0_ptr[i] * value;
}
break;
case SUM:
if (coeff_.empty()) {
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] + value;
}
break;
case SUB:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = input0_ptr[i] + value;
}
} else {
const float coeff_0 = swap ? coeff_[1] : coeff_[0];
const float coeff_1 = swap ? coeff_[0] : coeff_[1];
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] - value;
}
break;
case PROD:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = coeff_0 * input0_ptr[i] +
coeff_1 * value;
}
}
break;
case MAX:
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] * value;
}
break;
case DIV:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = std::max<T>(input0_ptr[i], value);
}
break;
case MIN:
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] / value;
}
break;
case MIN:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = std::min<T>(input0_ptr[i], value);
}
break;
case SUB:
for (index_t i = 0; i < size; ++i) {
output[i] = std::min<float>(input0[i], value);
}
break;
case MAX:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = swap ? value - input0_ptr[i] :
input0_ptr[i] - value;
}
break;
case DIV:
if (!swap) {
MACE_CHECK(fabs(value) > 1e-6, "cannot divided by 0.");
for (index_t i = 0; i < size; ++i) {
output[i] = std::max<float>(input0[i], value);
}
break;
case NEG:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
output_ptr[i] = input0_ptr[i] / value;
}
} else {
for (index_t i = 0; i < size; ++i) {
output[i] = -input0[i];
}
break;
case ABS:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
MACE_CHECK(fabs(input0_ptr[i]) > 1e-6, "cannot divided by 0.");
output_ptr[i] = value / input0_ptr[i];
}
}
break;
case SQR_DIFF:
for (index_t i = 0; i < size; ++i) {
output[i] = std::abs(input0[i]);
}
break;
case SQR_DIFF:
#pragma omp parallel for
for (index_t i = 0; i < num; ++i) {
const float tmp = input0_ptr[i] - value;
output_ptr[i] = tmp * tmp;
}
break;
default:
LOG(FATAL) << "Eltwise op not support type " << type_;
for (index_t i = 0; i < size; ++i) {
output[i] = std::pow(input0[i] - value, 2.f);
}
} else {
MACE_CHECK_NOTNULL(input0);
MACE_CHECK_NOTNULL(input1);
Tensor::MappingGuard input0_guard(input0);
Tensor::MappingGuard input1_guard(input1);
Tensor::MappingGuard output_guard(output);
const T *input0_ptr = input0->data<T>();
const T *input1_ptr = input1->data<T>();
T *output_ptr = output->mutable_data<T>();
const index_t size0 = input0->size();
const index_t size1 = input1->size();
break;
default:
LOG(FATAL) << "Eltwise op not support type " << type;
}
}
const index_t num = size0 / size1;
switch (type_) {
case PROD:
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j= 0; j < size1; ++j) {
output_ptr[i * size1 + j] =
input0_ptr[i * size1 + j] * input1_ptr[j];
}
inline void TensorVector(const EltwiseType type,
const float *input0,
const float *input1,
const index_t batch,
const index_t channel,
const index_t hw,
const bool swapped,
float *output) {
switch (type) {
case SUM:
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = input0[idx0] + input1[idx1];
}
break;
case SUM:
if (coeff_.empty()) {
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
output_ptr[i * size1 + j] =
input0_ptr[i * size1 + j] + input1_ptr[j];
}
}
} else {
const float coeff_0 = swap ? coeff_[1] : coeff_[0];
const float coeff_1 = swap ? coeff_[0] : coeff_[1];
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
output_ptr[i * size1 + j] =
coeff_0 * input0_ptr[i * size1 + j] +
coeff_1 * input1_ptr[j];
}
}
}
break;
case SUB:
if (swapped) {
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = input1[idx1] - input0[idx0];
}
}
break;
case MAX:
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
output_ptr[i * size1 + j] =
std::max<T>(input0_ptr[i * size1 + j], input1_ptr[j]);
}
} else {
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = input0[idx0] - input1[idx1];
}
}
break;
case MIN:
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
output_ptr[i * size1 + j] =
std::min<T>(input0_ptr[i * size1 + j], input1_ptr[j]);
}
}
}
break;
case PROD:
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = input0[idx0] * input1[idx1];
}
break;
case SUB:
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
output_ptr[i * size1 + j] = swap ?
input0_ptr[i * size1 + j] - input1_ptr[j] :
input1_ptr[j] - input0_ptr[i * size1 + j];
}
}
break;
case DIV:
if (swapped) {
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = input1[idx1] / input0[idx0];
}
}
break;
case DIV:
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
if (!swap) {
MACE_CHECK(fabs(input1_ptr[j]) > 1e-6, "cannot divided by 0.");
output_ptr[i * size1 + j] =
input0_ptr[i * size1 + j] / input1_ptr[j];
} else {
MACE_CHECK(fabs(input0_ptr[i * size1 + j]) > 1e-6,
"cannot divided by 0.");
output_ptr[i * size1 + j] =
input1_ptr[j] / input0_ptr[i * size1 + j];
}
}
} else {
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = input0[idx0] / input1[idx1];
}
}
break;
case SQR_DIFF:
#pragma omp parallel for collapse(2)
for (index_t i = 0; i < num; ++i) {
for (index_t j = 0; j < size1; ++j) {
const T tmp = input0_ptr[i * size1 + j] - input1_ptr[j];
output_ptr[i * size1 + j] = tmp * tmp;
}
}
}
break;
case MIN:
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = std::min<float>(input0[idx0], input1[idx1]);
}
}
}
break;
case MAX:
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = std::max<float>(input0[idx0], input1[idx1]);
}
}
}
break;
case SQR_DIFF:
#pragma omp parallel for collapse(3)
for (index_t b = 0; b < batch; ++b) {
for (index_t c = 0; c < channel; ++c) {
for (index_t i = 0; i < hw; ++i) {
const index_t idx0 = (b * channel + c) * hw + i;
const index_t idx1 = b * channel + c;
output[idx0] = std::pow(input0[idx0] - input1[idx1], 2.f);
}
}
}
break;
default:
LOG(FATAL) << "Eltwise op not support type " << type;
}
}
inline void TensorEltwise(const EltwiseType type,
const float *input0,
const float *input1,
const index_t size,
float *output) {
switch (type) {
case SUM:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] + input1[i];
}
break;
case SUB:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] - input1[i];
}
break;
case PROD:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] * input1[i];
}
break;
case DIV:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = input0[i] / input1[i];
}
break;
case MIN:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = std::min<float>(input0[i], input1[i]);
}
break;
case MAX:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = std::max<float>(input0[i], input1[i]);
}
break;
case SQR_DIFF:
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output[i] = std::pow(input0[i] - input1[i], 2.f);
}
break;
default:
LOG(FATAL) << "Eltwise op not support type " << type;
}
}
struct EltwiseFunctorBase {
EltwiseFunctorBase(const EltwiseType type,
const std::vector<float> &coeff,
const float value)
: type_(type), coeff_(coeff), value_(value) {}
EltwiseType type_;
std::vector<float> coeff_;
float value_;
};
template <DeviceType D, typename T>
struct EltwiseFunctor;
template <>
struct EltwiseFunctor<DeviceType::CPU, float>: EltwiseFunctorBase {
EltwiseFunctor(const EltwiseType type,
const std::vector<float> &coeff,
const float value)
: EltwiseFunctorBase(type, coeff, value) {}
void operator()(const Tensor *input0,
const Tensor *input1,
Tensor *output,
StatsFuture *future) {
bool swapped = false;
if (input1 != nullptr) {
MACE_CHECK(input0->dim_size() == input1->dim_size())
<< "Inputs of Eltwise op must be same shape";
if (input0->size() != input1->size()) {
if (input0->size() < input1->size()) {
std::swap(input0, input1);
swapped = true;
}
MACE_CHECK(input0->dim(0) == input1->dim(0) &&
input0->dim(1) == input1->dim(1) &&
input1->dim(2) == 1 &&
input1->dim(3) == 1)
<< "Element-Wise op only support channel dimension broadcast";
}
}
output->ResizeLike(input0);
Tensor::MappingGuard input0_guard(input0);
Tensor::MappingGuard output_guard(output);
const float *input0_ptr = input0->data<float>();
float *output_ptr = output->mutable_data<float>();
const index_t size = input0->size();
if (input1 == nullptr) {
TensorScalar(type_, input0_ptr, value_, size, output_ptr);
} else {
Tensor::MappingGuard input1_guard(input1);
const float *input1_ptr = input1->data<float>();
if (input1->size() != input0->size()) {
const index_t batch = input0->dim(0);
const index_t channel = input0->dim(1);
const index_t hw = input0->dim(2) * input0->dim(3);
TensorVector(type_, input0_ptr, input1_ptr,
batch, channel, hw, swapped, output_ptr);
} else {
if (!coeff_.empty() && type_ == SUM) {
#pragma omp parallel for
for (index_t i = 0; i < size; ++i) {
output_ptr[i] = coeff_[0] * input0_ptr[i] +
coeff_[1] * input1_ptr[i];
}
break;
default:
LOG(FATAL) << "Eltwise op not support type " << type_;
} else {
TensorEltwise(type_, input0_ptr, input1_ptr, size, output_ptr);
}
}
}
}
......@@ -249,15 +365,12 @@ struct EltwiseFunctor : EltwiseFunctorBase {
template <typename T>
struct EltwiseFunctor<DeviceType::OPENCL, T> : EltwiseFunctorBase {
EltwiseFunctor(const EltwiseType type,
const std::vector<float> &coeff)
: EltwiseFunctorBase(type, coeff) {}
const std::vector<float> &coeff,
const float value)
: EltwiseFunctorBase(type, coeff, value) {}
void operator()(const Tensor *input0,
const Tensor *input1,
const index_t start_axis,
const bool is_scaler,
const float value,
const bool swap,
Tensor *output,
StatsFuture *future);
......
mace/kernels/opencl/cl/eltwise.cl
浏览文件 @ fdb477e2
......@@ -3,8 +3,11 @@
__kernel void eltwise(KERNEL_ERROR_PARAMS
GLOBAL_WORK_GROUP_SIZE_DIM3
__read_only image2d_t input0,
__read_only image2d_t input1,
#if INPUT_TYPE == 1
__private const float value,
#else
__read_only image2d_t input1,
#endif
__private const int height,
__private const int width,
__private const int channel,
......@@ -13,101 +16,76 @@ __kernel void eltwise(KERNEL_ERROR_PARAMS
__private const float coeff1,
#endif
__write_only image2d_t output) {
const int c = get_global_id(0);
const int w = get_global_id(1);
const int chan_idx = get_global_id(0);
const int width_idx = get_global_id(1);
const int hb = get_global_id(2);
#ifndef NON_UNIFORM_WORK_GROUP
if (c >= global_size_dim0 || w >= global_size_dim1 || hb >= global_size_dim2)
if (chan_idx >= global_size_dim0 ||
width_idx >= global_size_dim1 || hb >= global_size_dim2)
return;
#endif
int pos_w;
int pos_h;
#if START_AXIS == 0
pos_w = mad24(c, width, w);
pos_h = hb;
#elif START_AXIS == 1
pos_w = mad24(c, width, w);
pos_h = hb % height;
#elif START_AXIS == 2
pos_w = mad24(c, width, w);
pos_h = 0;
#elif START_AXIS == 3
pos_w = c;
pos_h = 0;
#endif
const int pos = mad24(c, width, w);
const int remain_channel = channel - 4 * c;
const int pos = mad24(chan_idx, width, width_idx);
DATA_TYPE4 in0 = READ_IMAGET(input0, SAMPLER, (int2)(pos, hb));
DATA_TYPE4 in1 ;
#if IS_SCALER == 1
in1 = (DATA_TYPE4){value, value, value, value};
#if INPUT_TYPE == 1
DATA_TYPE4 in1 = (DATA_TYPE4)(value, value, value, value);
#elif INPUT_TYPE == 2
const int batch_idx = hb / height;
DATA_TYPE4 in1 = READ_IMAGET(input1, SAMPLER, (int2)(chan_idx, batch_idx));
#else
in1 = READ_IMAGET(input1, SAMPLER, (int2)(pos_w, pos_h));
DATA_TYPE4 in1 = READ_IMAGET(input1, SAMPLER, (int2)(pos, hb));
#endif
DATA_TYPE4 out;
#if ELTWISE_TYPE == 0
out = in0 * in1;
#ifdef COEFF_SUM
out = mad(coeff1, in0, mad(coeff0, in1, 0));
#else
out = in0 + in1;
#endif
#elif ELTWISE_TYPE == 1
#ifdef COEFF_SUM
#if NEEDSWAP == 0
out = mad(coeff0, in0, mad(coeff1, in1, 0));
#ifdef SWAPPED
out = in1 - in0;
#else
out = mad(coeff1, in0, mad(coeff0, in1, 0));
out = in0 - in1;
#endif
#else
out = in0 + in1;
#endif
#elif ELTWISE_TYPE == 2
out = fmax(in0, in1);
out = in0 * in1;
#elif ELTWISE_TYPE == 3
out = fmin(in0, in1);
#elif ELTWISE_TYPE == 4
#if NEED_SWAP == 0
out = in0 - in1;
#ifdef SWAPPED
out = in1 / in0;
#else
out = in1 - in0;
out = in0 / in1;
#endif
#elif ELTWISE_TYPE == 4
out = fmin(in0, in1);
#elif ELTWISE_TYPE == 5
#if NEED_SWAP == 0
if (fabs(in1.x) > 0.000001f)
out.x = in0.x / in1.x;
if (fabs(in1.y) > 0.000001f)
out.y = in0.y / in1.y;
if (fabs(in1.z) > 0.000001f)
out.z = in0.z / in1.z;
if (fabs(in1.w) > 0.000001f)
out.w = in0.w / in1.w;
#else
if (fabs(in1.x) > 0.000001f)
out.x = in1.x / in0.x;
if (fabs(in1.y) > 0.000001f)
out.y = in1.y / in0.y;
if (fabs(in1.z) > 0.000001f)
out.z = in1.z / in0.z;
if (fabs(in1.w) > 0.000001f)
out.w = in1.w / in0.w;
#endif
out = fmax(in0, in1);
#elif ELTWISE_TYPE == 6
in1 = (DATA_TYPE4)(0, 0, 0, 0);
out = in1 - in0;
#elif ELTWISE_TYPE == 7
out = fabs(in0);
#elif ELTWISE_TYPE == 8
DATA_TYPE4 diff = in0 - in1;
out = diff * diff;
#endif
#if ELTWISE_TYPE == 1 || ELTWISE_TYPE == 2 || ELTWISE_TYPE == 3 \
|| ELTWISE_TYPE == 4 || ELTWISE_TYPE == 8
if (remain_channel < 4) {
switch (remain_channel) {
case 1:
out.y = 0;
case 2:
out.z = 0;
case 3:
out.w = 0;
#if INPUT_TYPE == 1
#if ELTWISE_TYPE == 0 || ELTWISE_TYPE == 1 || ELTWISE_TYPE == 4 || ELTWISE_TYPE == 5 || ELTWISE_TYPE == 8
const int remain_channel = channel - 4 * chan_idx;
if (remain_channel < 4) {
switch (remain_channel) {
case 1:
out.y = 0;
case 2:
out.z = 0;
case 3:
out.w = 0;
}
}
}
#endif
#endif
WRITE_IMAGET(output, (int2)(pos, hb), out);
......
mace/kernels/opencl/eltwise_opencl.cc
浏览文件 @ fdb477e2
......@@ -23,16 +23,29 @@ namespace kernels {
template <typename T>
void EltwiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input0,
const Tensor *input1,
const index_t start_axis,
const bool is_scaler,
const float value,
const bool swap,
Tensor *output,
StatsFuture *future) {
const index_t batch = input0->dim(0);
const index_t height = input0->dim(1);
const index_t width = input0->dim(2);
const index_t channels = input0->dim(3);
bool swapped = false;
if (input1 != nullptr) {
MACE_CHECK(input0->dim_size() == input1->dim_size())
<< "Inputs of Eltwise op must be same shape";
if (input0->size() != input1->size()) {
if (input0->size() < input1->size()) {
std::swap(input0, input1);
swapped = true;
}
MACE_CHECK(input0->dim(0) == input1->dim(0) &&
input1->dim(1) == 1 &&
input1->dim(2) == 1 &&
input0->dim(3) == input1->dim(3))
<< "Element-Wise op only support channel dimension broadcast";
}
}
output->ResizeLike(input0);
const index_t batch = output->dim(0);
const index_t height = output->dim(1);
const index_t width = output->dim(2);
const index_t channels = output->dim(3);
const index_t channel_blocks = RoundUpDiv4(channels);
const index_t batch_height_pixels = batch * height;
......@@ -41,8 +54,6 @@ void EltwiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input0,
static_cast<uint32_t>(width),
static_cast<uint32_t>(batch_height_pixels)};
const int scaler = is_scaler ? 1 : 0;
const int need_swap = swap ? 1 : 0;
auto runtime = OpenCLRuntime::Global();
if (kernel_.get() == nullptr) {
std::set<std::string> built_options;
......@@ -52,9 +63,14 @@ void EltwiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input0,
built_options.emplace("-DDATA_TYPE=" + DtToUpstreamCLDt(dt));
built_options.emplace("-DCMD_DATA_TYPE=" + DtToUpstreamCLCMDDt(dt));
built_options.emplace(MakeString("-DELTWISE_TYPE=", type_));
built_options.emplace(MakeString("-DSTART_AXIS=", start_axis));
built_options.emplace(MakeString("-DIS_SCALER=", scaler));
built_options.emplace(MakeString("-DNEEDSWAP=", need_swap));
if (input1 == nullptr) {
built_options.emplace("-DINPUT_TYPE=1");
} else if (input0->size() != input1->size()) {
built_options.emplace("-DINPUT_TYPE=2");
if (swapped) built_options.emplace("-DSWAPPED");
}
if (!coeff_.empty()) built_options.emplace("-DCOEFF_SUM");
if (runtime->IsOutOfRangeCheckEnabled()) {
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
......@@ -66,7 +82,6 @@ void EltwiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input0,
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
if (!coeff_.empty()) built_options.emplace("-DCOEFF_SUM");
kernel_ = runtime->BuildKernel("eltwise", kernel_name, built_options);
kwg_size_ =
......@@ -84,8 +99,11 @@ void EltwiseFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input0,
kernel_.setArg(idx++, gws[2]);
}
kernel_.setArg(idx++, *(input0->opencl_image()));
kernel_.setArg(idx++, *(input1->opencl_image()));
kernel_.setArg(idx++, value);
if (input1 == nullptr) {
kernel_.setArg(idx++, value_);
} else {
kernel_.setArg(idx++, *(input1->opencl_image()));
}
kernel_.setArg(idx++, static_cast<int32_t>(height));
kernel_.setArg(idx++, static_cast<int32_t>(width));
kernel_.setArg(idx++, static_cast<int32_t>(channels));
......
mace/ops/eltwise.h
浏览文件 @ fdb477e2
......@@ -28,57 +28,20 @@ class EltwiseOp : public Operator<D, T> {
: Operator<D, T>(op_def, ws),
functor_(static_cast<kernels::EltwiseType>(
OperatorBase::GetSingleArgument<int>(
"type", static_cast<int>(kernels::EltwiseType::SUM))),
OperatorBase::GetRepeatedArgument<float>("coeff")) {}
"type", static_cast<int>(kernels::EltwiseType::NONE))),
OperatorBase::GetRepeatedArgument<float>("coeff"),
OperatorBase::GetSingleArgument<float>("x", 1.0)) {}
bool Run(StatsFuture *future) override {
if (this->InputSize() == 1) {
const Tensor* input = this->Input(0);
Tensor *output = this->Output(OUTPUT);
start_axis_ = input->dim_size() - 1;
is_scaler_ = true;
output->ResizeLike(input);
const float x = OperatorBase::GetSingleArgument<float>("x", 1.0);
functor_(input, nullptr, start_axis_,
is_scaler_, x, false, output, future);
} else {
const index_t size0 = this->Input(0)->size();
const index_t size1 = this->Input(1)->size();
const bool swap = (size0 < size1);
const Tensor *input0 = swap ? this->Input(1) : this->Input(0);
const Tensor *input1 = swap ? this->Input(0) : this->Input(1);
Tensor *output = this->Output(OUTPUT);
MACE_CHECK(input0->dim_size() == input1->dim_size())
<< "Inputs of Eltwise op must be same shape";
start_axis_ = input0->dim_size() - 1;
is_scaler_ = (input1->size() == 1);
uint32_t compared_size = 1;
if (!is_scaler_) {
while (start_axis_ >= 0) {
MACE_CHECK(input0->dim(start_axis_) == input1->dim(start_axis_),
"Invalid inputs dimension at axis: ") << start_axis_
<< "input 0: " << input0->dim(start_axis_)
<< "input 1: " << input1->dim(start_axis_);
compared_size *= input1->dim(start_axis_);
if (compared_size == input1->size()) {
break;
}
start_axis_--;
}
}
output->ResizeLike(input0);
const float x = OperatorBase::GetSingleArgument<float>("x", 1.0);
functor_(input0, input1, start_axis_,
is_scaler_, x, swap, output, future);
}
const Tensor* input0 = this->Input(0);
const Tensor* input1 = this->InputSize() == 2 ? this->Input(1) : nullptr;
Tensor *output = this->Output(OUTPUT);
functor_(input0, input1, output, future);
return true;
}
private:
kernels::EltwiseFunctor<D, T> functor_;
index_t start_axis_;
bool is_scaler_;
private:
OP_OUTPUT_TAGS(OUTPUT);
......
mace/ops/eltwise_benchmark.cc
浏览文件 @ fdb477e2
......@@ -35,10 +35,10 @@ void EltwiseBenchmark(
net.AddRandomInput<D, T>("Input1", {n, h, w, c});
if (D == DeviceType::OPENCL) {
BufferToImage<D, half>(&net, "Input0", "InputImg0",
kernels::BufferType::IN_OUT_CHANNEL);
BufferToImage<D, half>(&net, "Input1", "InputImg1",
kernels::BufferType::IN_OUT_CHANNEL);
BufferToImage<D, T>(&net, "Input0", "InputImg0",
kernels::BufferType::IN_OUT_CHANNEL);
BufferToImage<D, T>(&net, "Input1", "InputImg1",
kernels::BufferType::IN_OUT_CHANNEL);
OpDefBuilder("Eltwise", "EltwiseTest")
.Input("InputImg0")
.Input("InputImg1")
......@@ -48,9 +48,13 @@ void EltwiseBenchmark(
.Output("OutputImg")
.Finalize(net.NewOperatorDef());
} else {
net.TransformDataFormat<D, float>("Input0", NHWC,
"TInput0", NCHW);
net.TransformDataFormat<D, float>("Input1", NHWC,
"TInput1", NCHW);
OpDefBuilder("Eltwise", "EltwiseTest")
.Input("Input0")
.Input("Input1")
.Input("TInput0")
.Input("TInput1")
.AddIntArg("type", static_cast<int>(type))
.AddFloatsArg("coeff", {1.2, 2.1})
.AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
......@@ -89,13 +93,13 @@ void EltwiseBenchmark(
BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, float, OPENCL); \
BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, half, OPENCL);
BM_ELTWISE(0, 1, 256, 256, 32);
BM_ELTWISE(0, 1, 128, 128, 32);
BM_ELTWISE(1, 1, 128, 128, 32);
BM_ELTWISE(2, 1, 128, 128, 32);
BM_ELTWISE(0, 1, 240, 240, 256);
BM_ELTWISE(1, 1, 240, 240, 256);
BM_ELTWISE(2, 1, 240, 240, 256);
BM_ELTWISE(2, 1, 256, 256, 32);
BM_ELTWISE(0, 1, 128, 128, 32);
BM_ELTWISE(0, 1, 240, 240, 256);
BM_ELTWISE(5, 1, 128, 128, 32);
BM_ELTWISE(5, 1, 240, 240, 256);
} // namespace test
} // namespace ops
......
mace/ops/eltwise_test.cc
浏览文件 @ fdb477e2
此差异已折叠。
mace/python/tools/caffe_converter_lib.py
浏览文件 @ fdb477e2
......@@ -41,6 +41,12 @@ activation_name_map = {
'TanH': 'TANH',
}
math_type_mode = {
0: 2, # PROD
1: 0, # SUM
2: 5, # MAX
}
MACE_INPUT_NODE_NAME = "mace_input_node"
MACE_OUTPUT_NODE_NAME = "mace_output_node"
......@@ -921,11 +927,11 @@ class CaffeConverter(object):
param = op.layer.eltwise_param
type_arg = op_def.arg.add()
type_arg.name = 'type'
type_arg.i = param.operation
type_arg.i = math_type_mode[param.operation]
if len(param.coeff) > 0:
coeff_arg = op_def.arg.add()
coeff_arg.name = 'coeff'
coeff_arg.ints.extend(list(param.coeff))
coeff_arg.floats.extend(list(param.coeff))
output_shape = op.parents[0].output_shape_map[op.layer.bottom[0]]
op.output_shape_map[op.layer.top[0]] = output_shape
......
mace/python/tools/tf_converter_lib.py
浏览文件 @ fdb477e2
......@@ -30,14 +30,14 @@ pooling_type_mode = {'AvgPool': 1, 'MaxPool': 2}
# and also cwise type's in mace/kernels/cwise.h
# cuz these math ops should have compatible with "EltWise" and "CWise"
math_type_mode = {
'MUL': 0,
'ADD': 1,
'MAX': 2,
'MIN': 3,
'SUB': 4,
'DIV': 5,
'ADD': 0,
'SUB': 1,
'MUL': 2,
'DIV': 3,
'MIN': 4,
'MAX': 5,
'NEG': 6,
'ABS': 7
'ABS': 7,
}
buffer_type_map = {
......@@ -859,18 +859,26 @@ class TFConverter(object):
arg.i = self.dt
op_def.name = op.name
op_def.type = "Eltwise"
op_def.input.extend([input.name for input in op.inputs])
x_value = op.get_attr('x')
if len(op.inputs) >= 2:
if len(op.inputs) == 2:
input_tensor0 = get_input_tensor(op, 0)
input_tensor1 = get_input_tensor(op, 1)
if len(input_tensor0) == 1:
x_value = input_tensor0.eval().astype(np.float32)
elif len(input_tensor1) == 1:
x_value = input_tensor1.eval().astype(np.float32)
x_arg = op_def.arg.add()
x_arg.name = 'x'
x_arg.f = x_value
x_value = None
if np.asarray(input_tensor1.shape).size == 0:
x_value = input_tensor1.eval()
op_def.input.extend([op.inputs[0].name])
self.unused_tensor.add(input_tensor1.name)
elif np.asarray(input_tensor0.shape).size == 0:
x_value = input_tensor0.eval()
op_def.input.extend([op.inputs[1].name])
self.unused_tensor.add(input_tensor0.name)
else:
op_def.input.extend([input.name for input in op.inputs])
if x_value is not None:
x_arg = op_def.arg.add()
x_arg.name = 'x'
x_arg.f = x_value
else:
op_def.input.extend([input.name for input in op.inputs])
type_arg = op_def.arg.add()
type_arg.name = 'type'
type_arg.i = math_type_mode[math_type]
......
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