提交 25f8d24c 编写于 作者: J jingqinghe

add support for three-dimentional input in reduce max test=develop

上级 d132b8f3
...@@ -47,62 +47,62 @@ void reduce_n<float>(const float* src, ...@@ -47,62 +47,62 @@ void reduce_n<float>(const float* src,
} }
template <> template <>
void reduce_first_of_three<float>(const float* src, void reduce_first_of_three<float>(
float* dst, const float* src, float* dst, int first_in, int second_in, int third_in) {
int first_in, for (int i = 0; i < second_in; i++) {
int second_in, for (int j = 0; j < third_in; j++) {
int third_in){ dst[i * third_in + j] = src[i * third_in + j];
for (int i = 0; i < second_in; i++){ for (int k = 1; k < first_in; k++) {
for (int j = 0; j < third_in; j++){ dst[i * third_in + j] =
dst[i*third_in+j] = src[i*third_in+j]; src[k * second_in * third_in + i * third_in + j] >
for (int k = 1; k < first_in; k++){ dst[i * third_in + j]
dst[i*third_in+j] = src[k*second_in*third_in+i*third_in+j] > dst[i*third_in+j] ? src[k*second_in*third_in+i*third_in+j] : dst[i*third_in+j]; ? src[k * second_in * third_in + i * third_in + j]
: dst[i * third_in + j];
} }
} }
} }
} }
template <> template <>
void reduce_second_of_three<float>(const float* src, void reduce_second_of_three<float>(
float* dst, const float* src, float* dst, int first_in, int second_in, int third_in) {
int first_in, for (int i = 0; i < first_in; i++) {
int second_in, for (int j = 0; j < third_in; j++) {
int third_in){ dst[i * third_in + j] = src[i * second_in * third_in + j];
for (int i = 0; i < first_in; i++){ for (int k = 1; k < second_in; k++) {
for (int j = 0; j < third_in; j++){ dst[i * third_in + j] =
dst[i*third_in+j] = src[i*second_in*third_in+j]; src[i * second_in * third_in + third_in * k + j] >
for (int k = 1; k < second_in; k++){ dst[i * third_in + j]
dst[i*third_in+j] = src[i*second_in*third_in+third_in*k+j] > dst[i*third_in+j] ? src[i*second_in*third_in+third_in*k+j] : dst[i*third_in+j]; ? src[i * second_in * third_in + third_in * k + j]
: dst[i * third_in + j];
} }
} }
} }
} }
template <> template <>
void reduce_third_of_three<float>(const float* src, void reduce_third_of_three<float>(
float* dst, const float* src, float* dst, int first_in, int second_in, int third_in) {
int first_in, for (int i = 0; i < first_in; i++) {
int second_in, for (int j = 0; j < second_in; j++) {
int third_in){ dst[i * second_in + j] = src[i * second_in * third_in + j * second_in];
for (int i = 0; i < first_in; i++){ for (int k = 0; k < third_in; k++) {
for (int j = 0; j < second_in; j++){ dst[i * second_in + j] =
dst[i*second_in+j] = src[i*second_in*third_in+j*second_in]; src[i * second_in * third_in + j * second_in + k] >
for (int k = 0; k< third_in; k++){ dst[i * second_in + j]
dst[i*second_in+j] = src[i*second_in*third_in+j*second_in+k] > dst[i*second_in+j] ? src[i*second_in*third_in+j*second_in+k] : dst[i*second_in+j]; ? src[i * second_in * third_in + j * second_in + k]
: dst[i * second_in + j];
} }
} }
} }
} }
template <> template <>
void reduce_all_of_three<float>(const float* src, void reduce_all_of_three<float>(
float* dst, const float* src, float* dst, int first_in, int second_in, int third_in) {
int first_in,
int second_in,
int third_in){
float max = src[0]; float max = src[0];
int total_element = first_in * second_in * third_in; int total_element = first_in * second_in * third_in;
for (int i = 0; i <total_element; i++){ for (int i = 0; i < total_element; i++) {
max = src[i] > max ? src[i] : max; max = src[i] > max ? src[i] : max;
} }
dst[0] = max; dst[0] = max;
......
...@@ -36,32 +36,20 @@ void reduce_c(const T* src, ...@@ -36,32 +36,20 @@ void reduce_c(const T* src,
int width_in); int width_in);
template <typename T> template <typename T>
void reduce_all_of_three(const T* src, void reduce_all_of_three(
T* dst, const T* src, T* dst, int first_in, int second_in, int third_in);
int first_in,
int second_in,
int third_in);
template <typename T> template <typename T>
void reduce_first_of_three(const T* src, void reduce_first_of_three(
T* dst, const T* src, T* dst, int first_in, int second_in, int third_in);
int first_in,
int second_in,
int third_in);
template <typename T> template <typename T>
void reduce_second_of_three(const T* src, void reduce_second_of_three(
T* dst, const T* src, T* dst, int first_in, int second_in, int third_in);
int first_in,
int second_in,
int third_in);
template <typename T> template <typename T>
void reduce_third_of_three(const T* src, void reduce_third_of_three(
T* dst, const T* src, T* dst, int first_in, int second_in, int third_in);
int first_in,
int second_in,
int third_in);
template <typename T> template <typename T>
void reduce_h(const T* src, void reduce_h(const T* src,
......
...@@ -25,7 +25,7 @@ void ReduceMaxCompute::Run() { ...@@ -25,7 +25,7 @@ void ReduceMaxCompute::Run() {
auto& param = Param<operators::ReduceMaxParam>(); auto& param = Param<operators::ReduceMaxParam>();
const float* input = param.X->data<float>(); const float* input = param.X->data<float>();
auto x_dims = param.X->dims(); auto x_dims = param.X->dims();
int x_rank = x_dims.size(); int x_rank = x_dims.size();
float* output = param.Out->mutable_data<float>(); float* output = param.Out->mutable_data<float>();
bool keep_dim = param.keep_dim; bool keep_dim = param.keep_dim;
...@@ -39,37 +39,33 @@ void ReduceMaxCompute::Run() { ...@@ -39,37 +39,33 @@ void ReduceMaxCompute::Run() {
} }
} }
if (x_dims.size()==3){ if (x_dims.size() == 3) {
if (dim.size() == 0 || dim.size() == 3){ if (dim.size() == 0 || dim.size() == 3) {
lite::arm::math::reduce_all_of_three(input, output, x_dims[0], x_dims[1], x_dims[2]); lite::arm::math::reduce_all_of_three(
} input, output, x_dims[0], x_dims[1], x_dims[2]);
else if (dim.size() == 1){ } else if (dim.size() == 1) {
switch (dim[0]) switch (dim[0]) {
{ case 0:
case 0: lite::arm::math::reduce_first_of_three(
lite::arm::math::reduce_first_of_three(input, output, x_dims[0], x_dims[1], x_dims[2]); input, output, x_dims[0], x_dims[1], x_dims[2]);
break; break;
case 1: case 1:
lite::arm::math::reduce_second_of_three(input, output, x_dims[0], x_dims[1], x_dims[2]); lite::arm::math::reduce_second_of_three(
break; input, output, x_dims[0], x_dims[1], x_dims[2]);
break;
case 2: case 2:
lite::arm::math::reduce_third_of_three(input, output, x_dims[0], x_dims[1], x_dims[2]); lite::arm::math::reduce_third_of_three(
break; input, output, x_dims[0], x_dims[1], x_dims[2]);
default: break;
LOG(FATAL) << "error!!!"; default:
LOG(FATAL) << "error!!!";
} }
} } else if (dim.size() == 2) {
else if (dim.size() == 2){ } else {
}
else {
LOG(FATAL) << "dim size should not larger than 3!!!"; LOG(FATAL) << "dim size should not larger than 3!!!";
} }
} else if (x_dims.size() == 4) {
}
else if (x_dims.size()==4){
int n_in = x_dims[0]; int n_in = x_dims[0];
int c_in = x_dims[1]; int c_in = x_dims[1];
int h_in = x_dims[2]; int h_in = x_dims[2];
...@@ -78,37 +74,38 @@ void ReduceMaxCompute::Run() { ...@@ -78,37 +74,38 @@ void ReduceMaxCompute::Run() {
if (dim.size() == 0) { if (dim.size() == 0) {
lite::arm::math::reduce_all(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_all(input, output, n_in, c_in, h_in, w_in);
} else if (dim.size() == 1) { } else if (dim.size() == 1) {
switch (dim[0]) { switch (dim[0]) {
case 0: case 0:
lite::arm::math::reduce_n(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_n(input, output, n_in, c_in, h_in, w_in);
break; break;
case 1: case 1:
lite::arm::math::reduce_c(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_c(input, output, n_in, c_in, h_in, w_in);
break; break;
case 2: case 2:
lite::arm::math::reduce_h(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_h(input, output, n_in, c_in, h_in, w_in);
break; break;
case 3: case 3:
lite::arm::math::reduce_w(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_w(input, output, n_in, c_in, h_in, w_in);
break; break;
default: default:
LOG(FATAL) << "error!!!"; LOG(FATAL) << "error!!!";
} }
} else if (dim.size() == 2) { } else if (dim.size() == 2) {
if (dim[0] == 0 && dim[1] == 1) { if (dim[0] == 0 && dim[1] == 1) {
lite::arm::math::reduce_nc(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_nc(input, output, n_in, c_in, h_in, w_in);
} else if (dim[0] == 1 && dim[1] == 2) { } else if (dim[0] == 1 && dim[1] == 2) {
lite::arm::math::reduce_ch(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_ch(input, output, n_in, c_in, h_in, w_in);
} else if (dim[0] == 2 && dim[1] == 3) { } else if (dim[0] == 2 && dim[1] == 3) {
lite::arm::math::reduce_hw(input, output, n_in, c_in, h_in, w_in); lite::arm::math::reduce_hw(input, output, n_in, c_in, h_in, w_in);
} else { } else {
LOG(FATAL) << "invalid dim!!"; LOG(FATAL) << "invalid dim!!";
} }
} else { } else {
LOG(FATAL) << "dim's size over than 2, which is not supported now!!"; LOG(FATAL) << "dim's size over than 2, which is not supported now!!";
} }
} else {
LOG(FATAL) << "only support input with 3&4 dimensions now!!";
} }
} }
} // namespace arm } // namespace arm
......
...@@ -190,71 +190,64 @@ void reduce_hw(const float* src, ...@@ -190,71 +190,64 @@ void reduce_hw(const float* src,
reduce_w(tmp_out, dst, num_in, channel_in, 1, width_in); reduce_w(tmp_out, dst, num_in, channel_in, 1, width_in);
} }
void reduce_first_of_three(const float* src, void reduce_first_of_three(
float* dst, const float* src, float* dst, int first_in, int second_in, int third_in) {
int first_in, for (int i = 0; i < second_in; i++) {
int second_in, for (int j = 0; j < third_in; j++) {
int third_in){ dst[i * third_in + j] = src[i * third_in + j];
for (int k = 1; k < first_in; k++) {
for (int i = 0; i < second_in; i++){ dst[i * third_in + j] =
for (int j = 0; j < third_in; j++){ src[k * second_in * third_in + i * third_in + j] >
dst[i*third_in+j] = src[i*third_in+j]; dst[i * third_in + j]
for (int k = 1; k < first_in; k++){ ? src[k * second_in * third_in + i * third_in + j]
dst[i*third_in+j] = src[k*second_in*third_in+i*third_in+j] > dst[i*third_in+j] ? src[k*second_in*third_in+i*third_in+j] : dst[i*third_in+j]; : dst[i * third_in + j];
} }
} }
} }
} }
void reduce_second_of_three(
void reduce_second_of_three(const float* src, const float* src, float* dst, int first_in, int second_in, int third_in) {
float* dst, for (int i = 0; i < first_in; i++) {
int first_in, for (int j = 0; j < third_in; j++) {
int second_in, dst[i * third_in + j] = src[i * second_in * third_in + j];
int third_in){ for (int k = 1; k < second_in; k++) {
dst[i * third_in + j] =
for (int i = 0; i < first_in; i++){ src[i * second_in * third_in + third_in * k + j] >
for (int j = 0; j < third_in; j++){ dst[i * third_in + j]
dst[i*third_in+j] = src[i*second_in*third_in+j]; ? src[i * second_in * third_in + third_in * k + j]
for (int k = 1; k < second_in; k++){ : dst[i * third_in + j];
dst[i*third_in+j] = src[i*second_in*third_in+third_in*k+j] > dst[i*third_in+j] ? src[i*second_in*third_in+third_in*k+j] : dst[i*third_in+j];
} }
} }
} }
} }
void reduce_third_of_three(
void reduce_third_of_three(const float* src, const float* src, float* dst, int first_in, int second_in, int third_in) {
float* dst, for (int i = 0; i < first_in; i++) {
int first_in, for (int j = 0; j < second_in; j++) {
int second_in, dst[i * second_in + j] = src[i * second_in * third_in + j * second_in];
int third_in){ for (int k = 0; k < third_in; k++) {
dst[i * second_in + j] =
for (int i = 0; i < first_in; i++){ src[i * second_in * third_in + j * second_in + k] >
for (int j = 0; j < second_in; j++){ dst[i * second_in + j]
dst[i*second_in+j] = src[i*second_in*third_in+j*second_in]; ? src[i * second_in * third_in + j * second_in + k]
for (int k = 0; k< third_in; k++){ : dst[i * second_in + j];
dst[i*second_in+j] = src[i*second_in*third_in+j*second_in+k] > dst[i*second_in+j] ? src[i*second_in*third_in+j*second_in+k] : dst[i*second_in+j];
} }
} }
} }
} }
void reduce_all_of_three(
void reduce_all_of_three(const float* src, const float* src, float* dst, int first_in, int second_in, int third_in) {
float* dst,
int first_in,
int second_in,
int third_in){
float max = src[0]; float max = src[0];
int total_element = first_in * second_in * third_in; int total_element = first_in * second_in * third_in;
for (int i = 0; i <total_element; i++){ for (int i = 0; i < total_element; i++) {
max = src[i] > max ? src[i] : max; max = src[i] > max ? src[i] : max;
} }
dst[0] = max; dst[0] = max;
} }
class ReduceMaxComputeTester : public arena::TestCase { class ReduceMaxComputeTester : public arena::TestCase {
protected: protected:
// common attributes for this op. // common attributes for this op.
...@@ -321,37 +314,36 @@ class ReduceMaxComputeTester : public arena::TestCase { ...@@ -321,37 +314,36 @@ class ReduceMaxComputeTester : public arena::TestCase {
} }
auto* out_data = out->mutable_data<float>(); auto* out_data = out->mutable_data<float>();
if (x_dims_.size()==3){
if (dim_.size() == 0 || dim_.size() == 3){
reduce_all_of_three(x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
}
else if (dim_.size() == 1){
switch (dim_[0])
{
case 0:
reduce_first_of_three(x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
break;
case 1:
reduce_second_of_three(x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
break;
case 2: if (x_dims_.size() == 3) {
reduce_third_of_three(x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]); if (dim_.size() == 0 || dim_.size() == 3) {
break; reduce_all_of_three(
default: x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
LOG(FATAL) << "error!!!"; } else if (dim_.size() == 1) {
switch (dim_[0]) {
case 0:
reduce_first_of_three(
x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
break;
case 1:
reduce_second_of_three(
x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
break;
case 2:
reduce_third_of_three(
x_data, out_data, x_dims_[0], x_dims_[1], x_dims_[2]);
break;
default:
LOG(FATAL) << "error!!!";
} }
} } else if (dim_.size() == 2) {
else if (dim_.size() == 2){ LOG(FATAL) << "invalid dims_!!";
LOG(FATAL) << "invalid dims_!!"; } else {
}
else {
LOG(FATAL) << "dim size should not larger than 3!!!"; LOG(FATAL) << "dim size should not larger than 3!!!";
} }
} } else if (x_dims_.size() == 4) {
else if (x_dims_.size()==4){
int in_n = x_dims_[0]; int in_n = x_dims_[0];
int in_c = x_dims_[1]; int in_c = x_dims_[1];
int in_h = x_dims_[2]; int in_h = x_dims_[2];
...@@ -384,13 +376,9 @@ class ReduceMaxComputeTester : public arena::TestCase { ...@@ -384,13 +376,9 @@ class ReduceMaxComputeTester : public arena::TestCase {
reduce_hw(x_data, out_data, in_n, in_c, in_h, in_w); reduce_hw(x_data, out_data, in_n, in_c, in_h, in_w);
} else { } else {
LOG(FATAL) << "invalid dims_!!"; LOG(FATAL) << "invalid dims_!!";
} }
} }
} }
} }
void PrepareOpDesc(cpp::OpDesc* op_desc) { void PrepareOpDesc(cpp::OpDesc* op_desc) {
...@@ -434,26 +422,23 @@ void test_reduce_max(Place place) { ...@@ -434,26 +422,23 @@ void test_reduce_max(Place place) {
} }
void test_reduce_max_for_three(Place place) { void test_reduce_max_for_three(Place place) {
std::vector<std::vector<int>> reduce_dim{ std::vector<std::vector<int>> reduce_dim{{0}, {1}, {2}};
{0}, {1}, {2}};
for (auto f : {1, 3}) { for (auto f : {1, 3}) {
for (auto s : {1, 2}) { for (auto s : {1, 2}) {
for (auto t : {1, 3}) { for (auto t : {1, 3}) {
for (bool keep_dim : {false, true}) { for (bool keep_dim : {false, true}) {
for (auto dim : reduce_dim) { for (auto dim : reduce_dim) {
auto x_dims = DDim(std::vector<int64_t>({f, s, t})); auto x_dims = DDim(std::vector<int64_t>({f, s, t}));
std::unique_ptr<arena::TestCase> tester( std::unique_ptr<arena::TestCase> tester(new ReduceMaxComputeTester(
new ReduceMaxComputeTester( place, "def", dim, keep_dim, x_dims));
place, "def", dim, keep_dim, x_dims));
arena::Arena arena(std::move(tester), place, 2e-5); arena::Arena arena(std::move(tester), place, 2e-5);
arena.TestPrecision(); arena.TestPrecision();
}
} }
} }
} }
} }
} }
}
TEST(ReduceMax, precision) { TEST(ReduceMax, precision) {
// #ifdef LITE_WITH_X86 // #ifdef LITE_WITH_X86
......
Markdown is supported
0% .
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