Skip to content

Mace
Mace

慢慢CG / Mace
与 Fork 源项目一致

Fork自 Xiaomi / Mace

通知 1
Star 0
Fork 0
Mace
Mace
“d7e753546a813f7493c8834ca1a4b3f37a7ff139”上不存在“paddlespeech/s2t/exps/u2_kaldi/bin/test.py”
提交 2940ce5d 编写于 作者: 吴 吴承辉
浏览文件
操作

Merge branch 'mace-runtime-status' into 'master' 

Return mace status for allocate

See merge request !537
上级 fff1483d 33415ee9
展开全部 显示空白变更内容
内联 并排
Showing with 2763 addition and 3629 deletion
mace/core/mace.cc
浏览文件 @ 2940ce5d
......@@ -155,13 +155,13 @@ MaceStatus MaceEngine::Impl::Init(
}
} else {
#endif
MACE_FAILURE_RETURN(ws_->LoadModelTensor(
MACE_RETURN_IF_ERROR(ws_->LoadModelTensor(
*net_def, device_type_, model_data));
// Init model
auto net = CreateNet(op_registry_, *net_def, ws_.get(), device_type_,
NetMode::INIT);
MACE_FAILURE_RETURN(net->Run());
MACE_RETURN_IF_ERROR(net->Run());
net_ = CreateNet(op_registry_, *net_def, ws_.get(), device_type_);
#ifdef MACE_ENABLE_HEXAGON
}
......@@ -195,7 +195,7 @@ MaceStatus MaceEngine::Impl::Run(
" please use 1 to fill missing dimensions");
Tensor *input_tensor =
ws_->GetTensor(MakeString("mace_input_node_", input.first));
input_tensor->Resize(input.second.shape());
MACE_RETURN_IF_ERROR(input_tensor->Resize(input.second.shape()));
{
Tensor::MappingGuard input_guard(input_tensor);
float *input_data = input_tensor->mutable_data<float>();
......@@ -221,7 +221,7 @@ MaceStatus MaceEngine::Impl::Run(
hexagon_controller_->ExecuteGraph(*input_tensors[0], output_tensors[0]);
} else {
#endif
MACE_FAILURE_RETURN(net_->Run(run_metadata));
MACE_RETURN_IF_ERROR(net_->Run(run_metadata));
#ifdef MACE_ENABLE_HEXAGON
}
#endif
......
mace/core/net.cc
浏览文件 @ 2940ce5d
......@@ -71,7 +71,7 @@ MaceStatus SerialNet::Run(RunMetadata *run_metadata) {
CallStats call_stats;
if (future_wait) {
StatsFuture future;
MACE_FAILURE_RETURN(op->Run(&future));
MACE_RETURN_IF_ERROR(op->Run(&future));
if (run_metadata != nullptr) {
future.wait_fn(&call_stats);
} else {
......@@ -79,10 +79,10 @@ MaceStatus SerialNet::Run(RunMetadata *run_metadata) {
}
} else if (run_metadata != nullptr) {
call_stats.start_micros = NowMicros();
MACE_FAILURE_RETURN(op->Run(nullptr));
MACE_RETURN_IF_ERROR(op->Run(nullptr));
call_stats.end_micros = NowMicros();
} else {
MACE_FAILURE_RETURN(op->Run(nullptr));
MACE_RETURN_IF_ERROR(op->Run(nullptr));
}
if (run_metadata != nullptr) {
......
mace/core/workspace.cc
浏览文件 @ 2940ce5d
......@@ -83,10 +83,7 @@ MaceStatus Workspace::LoadModelTensor(const NetDef &net_def,
} else {
tensor_buffer_ = std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(type)));
MaceStatus status = tensor_buffer_->Allocate(model_data_size);
if (status != MaceStatus::MACE_SUCCESS) {
return status;
}
MACE_RETURN_IF_ERROR(tensor_buffer_->Allocate(model_data_size));
tensor_buffer_->Map(nullptr);
tensor_buffer_->Copy(const_cast<unsigned char*>(model_data),
0, model_data_size);
......@@ -156,11 +153,8 @@ MaceStatus Workspace::CreateOutputTensorBuffer(const NetDef &net_def,
if (mem_block.mem_id() >= 20000) {
std::unique_ptr<BufferBase> image_buf(
new Image());
MaceStatus status = image_buf->Allocate(
{mem_block.x(), mem_block.y()}, dtype);
if (status != MaceStatus::MACE_SUCCESS) {
return status;
}
MACE_RETURN_IF_ERROR(image_buf->Allocate(
{mem_block.x(), mem_block.y()}, dtype));
preallocated_allocator_.SetBuffer(mem_block.mem_id(),
std::move(image_buf));
}
......@@ -168,12 +162,9 @@ MaceStatus Workspace::CreateOutputTensorBuffer(const NetDef &net_def,
if (mem_block.mem_id() < 20000) {
std::unique_ptr<BufferBase> tensor_buf(
new Buffer(GetDeviceAllocator(device_type)));
MaceStatus status = tensor_buf->Allocate(
MACE_RETURN_IF_ERROR(tensor_buf->Allocate(
mem_block.x() * GetEnumTypeSize(dtype)
+ MACE_EXTRA_BUFFER_PAD_SIZE);
if (status != MaceStatus::MACE_SUCCESS) {
return status;
}
+ MACE_EXTRA_BUFFER_PAD_SIZE));
preallocated_allocator_.SetBuffer(mem_block.mem_id(),
std::move(tensor_buf));
}
......
mace/kernels/addn.h
浏览文件 @ 2940ce5d
......@@ -40,7 +40,7 @@ struct AddNFunctor {
Tensor *output_tensor,
StatsFuture *future) {
MACE_UNUSED(future);
MACE_FAILURE_RETURN(output_tensor->ResizeLike(input_tensors[0]));
MACE_RETURN_IF_ERROR(output_tensor->ResizeLike(input_tensors[0]));
index_t size = output_tensor->size();
Tensor::MappingGuard output_map(output_tensor);
float *output_data = output_tensor->mutable_data<float>();
......
mace/kernels/arm/conv_2d_neon.h
浏览文件 @ 2940ce5d
......@@ -103,9 +103,9 @@ inline void Conv2dCPUKHxKWCalc(const float *in_ptr,
for (index_t w = 0; w < out_width; ++w) {
for (int i = 0; i < filter_height; ++i) {
for (int j = 0; j < filter_width; ++j) {
out_ptr[h * out_width + w]
+= in_ptr[(h * stride + i) * in_width + (w * stride + j)]
* filter_ptr[i * filter_width + j];
out_ptr[h * out_width + w] +=
in_ptr[(h * stride + i) * in_width + (w * stride + j)] *
filter_ptr[i * filter_width + j];
}
}
}
......
mace/kernels/arm/conv_2d_neon_15x1.cc
浏览文件 @ 2940ce5d
......@@ -38,16 +38,15 @@ inline void Conv2dCPUK15x1Calc(const float *in_ptr,
for (index_t iw = 0; iw < tile_width && w + iw < out_width; ++iw) {
for (int i = 0; i < 15; ++i) {
for (int j = 0; j < 1; ++j) {
out_ptr[io * out_image_size + ih * out_width + w + iw]
+= in_ptr[(ih * stride + i) * in_width + ((w + iw) * stride + j)]
* filter_ptr[io * in_channels * 15 + i * 1 + j];
out_ptr[io * out_image_size + ih * out_width + w + iw] +=
in_ptr[(ih * stride + i) * in_width + ((w + iw) * stride + j)] *
filter_ptr[io * in_channels * 15 + i * 1 + j];
}
}
}
}
}
// Ho = 4, Wo = 1, Co = 1
void Conv2dNeonK15x1S1(const float *input,
const float *filter,
......@@ -69,8 +68,7 @@ void Conv2dNeonK15x1S1(const float *input,
const index_t out_width = out_shape[3];
const index_t in_channels = in_shape[1];
const index_t in_width = in_shape[3];
float *out_ptr_base =
output + b * out_batch_size + m * out_image_size;
float *out_ptr_base = output + b * out_batch_size + m * out_image_size;
for (index_t c = 0; c < in_channels; ++c) {
const float *in_ptr_base =
input + b * in_batch_size + c * in_image_size;
......
mace/kernels/arm/conv_2d_neon_1x1.cc
浏览文件 @ 2940ce5d
......@@ -31,12 +31,8 @@ void Conv2dNeonK1x1S1(const float *input,
const index_t out_channels,
float *output) {
for (index_t b = 0; b < batch; ++b) {
Gemm(filter,
input + b * in_channels * height * width,
1,
out_channels,
in_channels,
height * width,
Gemm(filter, input + b * in_channels * height * width, 1, out_channels,
in_channels, height * width,
output + b * out_channels * height * width);
}
}
......
mace/kernels/arm/conv_2d_neon_1x15.cc
浏览文件 @ 2940ce5d
......@@ -17,8 +17,8 @@
#endif
#include "mace/kernels/arm/conv_2d_neon.h"
#include "mace/utils/utils.h"
#include "mace/utils/logging.h"
#include "mace/utils/utils.h"
namespace mace {
namespace kernels {
......@@ -39,16 +39,15 @@ inline void Conv2dCPUK1x15Calc(const float *in_ptr,
for (index_t iw = 0; iw < out_width; ++iw) {
for (int i = 0; i < 1; ++i) {
for (int j = 0; j < 15; ++j) {
out_ptr[io * out_image_size + (h + ih) * out_width + iw]
+= in_ptr[((h + ih) * stride + i) * in_width + (iw * stride + j)]
* filter_ptr[io * in_channels * 15 + i * 15 + j];
out_ptr[io * out_image_size + (h + ih) * out_width + iw] +=
in_ptr[((h + ih) * stride + i) * in_width + (iw * stride + j)] *
filter_ptr[io * in_channels * 15 + i * 15 + j];
}
}
}
}
}
// Ho = 1, Wo = 4, Co = 1
void Conv2dNeonK1x15S1(const float *input,
const float *filter,
......@@ -70,8 +69,7 @@ void Conv2dNeonK1x15S1(const float *input,
const index_t out_width = out_shape[3];
const index_t in_channels = in_shape[1];
const index_t in_width = in_shape[3];
float *out_ptr_base =
output + b * out_batch_size + m * out_image_size;
float *out_ptr_base = output + b * out_batch_size + m * out_image_size;
for (index_t c = 0; c < in_channels; ++c) {
const float *in_ptr_base =
input + b * in_batch_size + c * in_image_size;
......
mace/kernels/arm/conv_2d_neon_1x7.cc
浏览文件 @ 2940ce5d
......@@ -41,8 +41,7 @@ void Conv2dNeonK1x7S1(const float *input,
const index_t in_channels = in_shape[1];
const index_t in_width = in_shape[3];
if (m + 3 < out_channels) {
float *out_ptr0_base =
output + b * out_batch_size + m * out_image_size;
float *out_ptr0_base = output + b * out_batch_size + m * out_image_size;
#if defined(MACE_ENABLE_NEON)
float *out_ptr1_base =
output + b * out_batch_size + (m + 1) * out_image_size;
......@@ -56,12 +55,9 @@ void Conv2dNeonK1x7S1(const float *input,
input + b * in_batch_size + c * in_image_size;
const float *filter_ptr0 = filter + m * in_channels * 7 + c * 7;
#if defined(MACE_ENABLE_NEON)
const float *filter_ptr1 =
filter + (m + 1) * in_channels * 7 + c * 7;
const float *filter_ptr2 =
filter + (m + 2) * in_channels * 7 + c * 7;
const float *filter_ptr3 =
filter + (m + 3) * in_channels * 7 + c * 7;
const float *filter_ptr1 = filter + (m + 1) * in_channels * 7 + c * 7;
const float *filter_ptr2 = filter + (m + 2) * in_channels * 7 + c * 7;
const float *filter_ptr3 = filter + (m + 3) * in_channels * 7 + c * 7;
/* load filter (4 outch x 1 height x 4 width) */
float32x4_t vf00, vf01;
float32x4_t vf10, vf11;
......@@ -241,9 +237,8 @@ void Conv2dNeonK1x7S1(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0,
in_width, 1, 7, out_height, out_width,
out_ptr0_base, 1);
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0, in_width, 1, 7,
out_height, out_width, out_ptr0_base, 1);
#endif
} // c
}
......
mace/kernels/arm/conv_2d_neon_3x3.cc
浏览文件 @ 2940ce5d
......@@ -75,7 +75,6 @@ void Conv2dNeonK3x3S1(const float *input,
vf11 = vld1q_f32(filter_ptr1 + 3);
vf12 = vld1q_f32(filter_ptr1 + 6);
for (index_t h = 0; h + 1 < out_height; h += 2) {
for (index_t w = 0; w + 3 < out_width; w += 4) {
// input (4 height x 3 slide): vi_height_slide
......@@ -198,7 +197,6 @@ void Conv2dNeonK3x3S1(const float *input,
vf167 = vld1_f32(filter_ptr1 + 6);
vf189 = vld1_f32(filter_ptr1 + 8);
for (index_t h = 0; h + 1 < out_height; h += 2) {
for (index_t w = 0; w + 3 < out_width; w += 4) {
// input (4 height x 3 slide): vi_height_slide
......@@ -313,11 +311,11 @@ void Conv2dNeonK3x3S1(const float *input,
} // c
} else {
for (index_t mm = m; mm < out_channels; ++mm) {
float
*out_ptr0_base = output + b * out_batch_size + mm * out_image_size;
float *out_ptr0_base =
output + b * out_batch_size + mm * out_image_size;
for (index_t c = 0; c < in_channels; ++c) {
const float
*in_ptr0 = input + b * in_batch_size + c * in_image_size;
const float *in_ptr0 =
input + b * in_batch_size + c * in_image_size;
#if defined(MACE_ENABLE_NEON)
const float *in_ptr1 =
input + b * in_batch_size + c * in_image_size + 1 * in_width;
......@@ -396,7 +394,6 @@ void Conv2dNeonK3x3S1(const float *input,
vst1q_f32(out_ptr0, vo00);
vst1q_f32(out_ptr0 + out_width, vo01);
in_ptr0 += 4;
in_ptr1 += 4;
in_ptr2 += 4;
......@@ -482,7 +479,6 @@ void Conv2dNeonK3x3S1(const float *input,
vst1q_f32(out_ptr0, vo00);
vst1q_f32(out_ptr0 + out_width, vo01);
in_ptr0 += 4;
in_ptr1 += 4;
in_ptr2 += 4;
......@@ -499,9 +495,8 @@ void Conv2dNeonK3x3S1(const float *input,
out_ptr0 += out_width;
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr0, filter_ptr0,
in_width, 3, 3, out_height, out_width,
out_ptr0_base, 1);
Conv2dCPUKHxKWCalc(in_ptr0, filter_ptr0, in_width, 3, 3, out_height,
out_width, out_ptr0_base, 1);
#endif
} // c
} // mm
......@@ -529,8 +524,7 @@ void Conv2dNeonK3x3S2(const float *input,
const index_t out_height = out_shape[2];
const index_t out_width = out_shape[3];
const float *in_base = input + b * in_batch_size + c * in_image_size;
const float
*filter_ptr = filter + m * in_channels * 9 + c * 9;
const float *filter_ptr = filter + m * in_channels * 9 + c * 9;
float *out_base = output + b * out_batch_size + m * out_image_size;
#if defined(MACE_ENABLE_NEON) && defined(__aarch64__)
......@@ -656,9 +650,8 @@ void Conv2dNeonK3x3S2(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_base, filter_ptr,
in_width, 3, 3, out_height, out_width,
out_base, 2);
Conv2dCPUKHxKWCalc(in_base, filter_ptr, in_width, 3, 3, out_height,
out_width, out_base, 2);
#endif
} // c
} // m
......
mace/kernels/arm/conv_2d_neon_5x5.cc
浏览文件 @ 2940ce5d
......@@ -205,9 +205,8 @@ void Conv2dNeonK5x5S1(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0,
in_width, 5, 5, out_height, out_width,
out_ptr0_base, 1);
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0, in_width, 5, 5,
out_height, out_width, out_ptr0_base, 1);
#endif
} // c
} // mm
......
mace/kernels/arm/conv_2d_neon_7x1.cc
浏览文件 @ 2940ce5d
......@@ -41,8 +41,7 @@ void Conv2dNeonK7x1S1(const float *input,
const index_t in_channels = in_shape[1];
const index_t in_width = in_shape[3];
if (m + 3 < out_channels) {
float *out_ptr0_base =
output + b * out_batch_size + m * out_image_size;
float *out_ptr0_base = output + b * out_batch_size + m * out_image_size;
#if defined(MACE_ENABLE_NEON)
float *out_ptr1_base =
output + b * out_batch_size + (m + 1) * out_image_size;
......@@ -56,12 +55,9 @@ void Conv2dNeonK7x1S1(const float *input,
input + b * in_batch_size + c * in_image_size;
const float *filter_ptr0 = filter + m * in_channels * 7 + c * 7;
#if defined(MACE_ENABLE_NEON)
const float *filter_ptr1 =
filter + (m + 1) * in_channels * 7 + c * 7;
const float *filter_ptr2 =
filter + (m + 2) * in_channels * 7 + c * 7;
const float *filter_ptr3 =
filter + (m + 3) * in_channels * 7 + c * 7;
const float *filter_ptr1 = filter + (m + 1) * in_channels * 7 + c * 7;
const float *filter_ptr2 = filter + (m + 2) * in_channels * 7 + c * 7;
const float *filter_ptr3 = filter + (m + 3) * in_channels * 7 + c * 7;
/* load filter (4 outch x 4 height x 1 width) */
float32x4_t vf00, vf01;
float32x4_t vf10, vf11;
......@@ -98,7 +94,6 @@ void Conv2dNeonK7x1S1(const float *input,
out_ptr3_base[out_offset + 2 * out_width],
out_ptr3_base[out_offset + 3 * out_width]};
// input offset
index_t in_offset = h * in_width + w;
// input (3 slide)
......@@ -282,9 +277,8 @@ void Conv2dNeonK7x1S1(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0,
in_width, 7, 1, out_height, out_width,
out_ptr0_base, 1);
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0, in_width, 7, 1,
out_height, out_width, out_ptr0_base, 1);
#endif
} // c
}
......
mace/kernels/arm/conv_2d_neon_7x7.cc
浏览文件 @ 2940ce5d
......@@ -298,9 +298,8 @@ void Conv2dNeonK7x7S1(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0,
in_width, 7, 7, out_height, out_width,
out_ptr0_base, 1);
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0, in_width, 7, 7,
out_height, out_width, out_ptr0_base, 1);
#endif
} // c
} // mm
......@@ -464,9 +463,8 @@ void Conv2dNeonK7x7S2(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0,
in_width, 7, 7, out_height, out_width,
out_ptr0_base, 2);
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0, in_width, 7, 7,
out_height, out_width, out_ptr0_base, 2);
#endif
} // c
} // mm
......@@ -630,9 +628,8 @@ void Conv2dNeonK7x7S3(const float *input,
} // w
} // h
#else
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0,
in_width, 7, 7, out_height, out_width,
out_ptr0_base, 3);
Conv2dCPUKHxKWCalc(in_ptr_base, filter_ptr0, in_width, 7, 7,
out_height, out_width, out_ptr0_base, 3);
#endif
} // c
} // mm
......
mace/kernels/arm/conv_winograd.cc
浏览文件 @ 2940ce5d
......@@ -17,8 +17,8 @@
#include "mace/kernels/arm/conv_winograd.h"
#include "mace/kernels/gemm.h"
#include "mace/utils/utils.h"
#include "mace/utils/logging.h"
#include "mace/utils/utils.h"
namespace mace {
namespace kernels {
......@@ -49,9 +49,8 @@ void TransformInput4x4(const float *input,
s15;
// load tile data
const float *input_ptr =
input + n * input_batch_size + c * in_height_width + h * in_width
+ w;
const float *input_ptr = input + n * input_batch_size +
c * in_height_width + h * in_width + w;
d0 = input_ptr[0];
d1 = input_ptr[1];
d2 = input_ptr[2];
......@@ -166,9 +165,8 @@ void TransformInput8x8(const float *input,
float s[8][8];
for (index_t h = 0; h < in_height - 2; h += 6) {
for (index_t w = 0; w < in_width - 2; w += 6) {
const float *input_ptr =
input + n * input_batch_size + c * in_height_width + h * in_width
+ w;
const float *input_ptr = input + n * input_batch_size +
c * in_height_width + h * in_width + w;
for (int i = 0; i < 8; ++i) {
float d0, d1, d2, d3, d4, d5, d6, d7;
......@@ -258,25 +256,16 @@ void BatchGemm(const float *input,
const index_t out_stride = out_channels * tile_count;
if (batch == 1) {
Gemm(filter,
input,
in_tile_area,
out_channels,
in_channels,
tile_count,
Gemm(filter, input, in_tile_area, out_channels, in_channels, tile_count,
output);
} else {
#pragma omp parallel for collapse(2)
for (int b = 0; b < batch; ++b) {
for (int i = 0; i < in_tile_area; ++i) {
const float
*in_ptr = input + b * in_batch_size + i * in_stride;
const float *in_ptr = input + b * in_batch_size + i * in_stride;
const float *filter_ptr = filter + i * filter_stride;
float *out_ptr = output + b * out_batch_size + i * out_stride;
Gemm(filter_ptr,
in_ptr,
1,
out_channels, /* rows */
Gemm(filter_ptr, in_ptr, 1, out_channels, /* rows */
in_channels, /* K */
tile_count, /* cols */
out_ptr);
......@@ -345,9 +334,8 @@ void TransformOutput4x4(const float *input,
v2 = s2 - s4 - s6;
v3 = s3 - s5 - s7;
float *output_ptr =
output + n * output_batch_size + m * out_image_size + h * out_width
+ w;
float *output_ptr = output + n * output_batch_size +
m * out_image_size + h * out_width + w;
output_ptr[0] = v0;
output_ptr[1] = v1;
output_ptr[out_width] = v2;
......@@ -433,9 +421,8 @@ void TransformOutput8x8(const float *input,
input_ptr += 8 * stride;
}
float *output_ptr =
output + n * output_batch_size + m * out_image_size + h * out_width
+ w;
float *output_ptr = output + n * output_batch_size +
m * out_image_size + h * out_width + w;
for (int i = 0; i < 6; ++i) {
float d0, d1, d2, d3, d4, d5, d6, d7;
......@@ -471,7 +458,6 @@ void TransformOutput8x8(const float *input,
}
} // namespace
// OCHW => TOC
// no need to optimize, it will exist in converter
void TransformFilter4x4(const float *filter,
......@@ -573,16 +559,14 @@ void TransformFilter8x8(const float *filter,
float *output) {
const index_t stride = out_channels * in_channels;
const float G[8][3] = {
{1.0f, 0.0f, 0.0f},
const float G[8][3] = {{1.0f, 0.0f, 0.0f},
{-2.0f / 9, -2.0f / 9, -2.0f / 9},
{-2.0f / 9, 2.0f / 9, -2.0f / 9},
{1.0f / 90, 1.0f / 45, 2.0f / 45},
{1.0f / 90, -1.0f / 45, 2.0f / 45},
{1.0f / 45, 1.0f / 90, 1.0f / 180},
{1.0f / 45, -1.0f / 90, 1.0f / 180},
{0.0f, 0.0f, 1.0f}
};
{0.0f, 0.0f, 1.0f}};
#pragma omp parallel for collapse(2)
for (index_t m = 0; m < out_channels; ++m) {
......@@ -640,55 +624,31 @@ void WinoGradConv3x3s1(const float *input,
switch (out_tile_size) {
case 2:
TransformInput4x4(input,
batch,
in_height,
in_width,
in_channels,
tile_count,
transformed_input);
TransformInput4x4(input, batch, in_height, in_width, in_channels,
tile_count, transformed_input);
break;
case 6:
TransformInput8x8(input,
batch,
in_height,
in_width,
in_channels,
tile_count,
transformed_input);
TransformInput8x8(input, batch, in_height, in_width, in_channels,
tile_count, transformed_input);
break;
default:MACE_NOT_IMPLEMENTED;
default:
MACE_NOT_IMPLEMENTED;
}
BatchGemm(transformed_input,
transformed_filter,
batch,
in_channels,
out_channels,
tile_count,
out_tile_size,
transformed_output);
BatchGemm(transformed_input, transformed_filter, batch, in_channels,
out_channels, tile_count, out_tile_size, transformed_output);
switch (out_tile_size) {
case 2:
TransformOutput4x4(transformed_output,
batch,
out_height,
out_width,
out_channels,
tile_count,
output);
TransformOutput4x4(transformed_output, batch, out_height, out_width,
out_channels, tile_count, output);
break;
case 6:
TransformOutput8x8(transformed_output,
batch,
out_height,
out_width,
out_channels,
tile_count,
output);
TransformOutput8x8(transformed_output, batch, out_height, out_width,
out_channels, tile_count, output);
break;
default:MACE_NOT_IMPLEMENTED;
default:
MACE_NOT_IMPLEMENTED;
}
}
......@@ -712,8 +672,8 @@ void WinoGradConv3x3s1(const float *input,
index_t transformed_input_size =
in_tile_area * batch * in_channels * tile_count;
index_t transformed_filter_size = in_tile_area * out_channels * in_channels;
index_t
transformed_output_size = in_tile_area * batch * out_channels * tile_count;
index_t transformed_output_size =
in_tile_area * batch * out_channels * tile_count;
float *transformed_input = new float[transformed_input_size]; // TNCB
float *transformed_filter = new float[transformed_filter_size]; // TOC
......@@ -721,35 +681,22 @@ void WinoGradConv3x3s1(const float *input,
switch (out_tile_size) {
case 2:
TransformFilter4x4(filter,
in_channels,
out_channels,
transformed_filter);
TransformFilter4x4(filter, in_channels, out_channels, transformed_filter);
break;
case 6:
TransformFilter8x8(filter,
in_channels,
out_channels,
transformed_filter);
TransformFilter8x8(filter, in_channels, out_channels, transformed_filter);
break;
default:MACE_NOT_IMPLEMENTED;
}
WinoGradConv3x3s1(input,
transformed_filter,
batch,
in_height,
in_width,
in_channels,
out_channels,
out_tile_size,
transformed_input,
transformed_output,
output);
default:
MACE_NOT_IMPLEMENTED;
}
WinoGradConv3x3s1(input, transformed_filter, batch, in_height, in_width,
in_channels, out_channels, out_tile_size, transformed_input,
transformed_output, output);
delete[]transformed_input;
delete[]transformed_filter;
delete[]transformed_output;
delete[] transformed_input;
delete[] transformed_filter;
delete[] transformed_output;
}
void ConvRef3x3s1(const float *input,
......@@ -778,10 +725,9 @@ void ConvRef3x3s1(const float *input,
index_t iw = w + kw;
index_t in_offset =
((b * in_channels + c) * in_height + ih) * in_width + iw;
index_t
filter_offset = (((m * in_channels) + c) * 3 + kh) * 3 + kw;
output[out_offset] +=
input[in_offset] * filter[filter_offset];
index_t filter_offset =
(((m * in_channels) + c) * 3 + kh) * 3 + kw;
output[out_offset] += input[in_offset] * filter[filter_offset];
}
}
}
......
mace/kernels/arm/conv_winograd_test.cc
浏览文件 @ 2940ce5d
......@@ -13,13 +13,13 @@
// limitations under the License.
#include <gtest/gtest.h>
#include <random>
#include <algorithm>
#include <memory>
#include <random>
#include "mace/kernels/arm/conv_winograd.h"
#include "mace/core/types.h"
#include "mace/core/tensor.h"
#include "mace/core/types.h"
#include "mace/kernels/arm/conv_winograd.h"
namespace mace {
namespace kernels {
......@@ -55,32 +55,18 @@ TEST(ConvWinogradTest, winograd) {
std::random_device rd;
std::mt19937 gen(rd());
std::normal_distribution<float> nd(0, 1);
std::generate(input_data, input_data + input_size,
[&gen, &nd] {
std::generate(input_data, input_data + input_size, [&gen, &nd] {
return std::max(-1.0f, std::min(1.0f, nd(gen)));
});
std::generate(filter_data, filter_data + filter_size,
[&gen, &nd] {
std::generate(filter_data, filter_data + filter_size, [&gen, &nd] {
return std::max(-1.0f, std::min(1.0f, nd(gen)));
});
kernels::ConvRef3x3s1(input_data,
filter_data,
batch,
in_height,
in_width,
in_channels,
out_channels,
output_data_ref);
kernels::ConvRef3x3s1(input_data, filter_data, batch, in_height, in_width,
in_channels, out_channels, output_data_ref);
kernels::WinoGradConv3x3s1(input_data,
filter_data,
batch,
in_height,
in_width,
in_channels,
out_channels,
6,
kernels::WinoGradConv3x3s1(input_data, filter_data, batch, in_height,
in_width, in_channels, out_channels, 6,
output_data);
// test
......
mace/kernels/arm/depthwise_conv2d_neon_3x3.cc
浏览文件 @ 2940ce5d
......@@ -16,8 +16,8 @@
#include <arm_neon.h>
#endif
#include "mace/kernels/arm/depthwise_conv2d_neon.h"
#include "mace/core/macros.h"
#include "mace/kernels/arm/depthwise_conv2d_neon.h"
namespace mace {
namespace kernels {
......@@ -52,9 +52,9 @@ void DepthwiseConv2dPixel(const float *in_base,
// Ho = 2, Wo = 4, Co = 1
void DepthwiseConv2dNeonK3x3S1(const float *input,
const float *filter,
const index_t* in_shape,
const index_t* out_shape,
const int* pad_hw,
const index_t *in_shape,
const index_t *out_shape,
const int *pad_hw,
const index_t valid_h_start,
const index_t valid_h_stop,
const index_t valid_w_start,
......@@ -88,18 +88,9 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
// top
for (h = 0; h < valid_h_start; ++h) {
for (w = 0; w < out_shape[3]; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h - pad_top,
w - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h - pad_top,
w - pad_left, out_width, in_height, in_width, 3,
3, out_base);
}
}
......@@ -113,30 +104,12 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
for (h = valid_h_start; h + 1 < valid_h_stop; h += 2) {
// left
for (w = 0; w < valid_w_start; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h - pad_top,
w - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base,
filter_ptr,
h + 1,
w,
h + 1 - pad_top,
w - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h - pad_top,
w - pad_left, out_width, in_height, in_width, 3,
3, out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h + 1, w, h + 1 - pad_top,
w - pad_left, out_width, in_height, in_width, 3,
3, out_base);
}
for (w = valid_w_start; w + 3 < valid_w_stop; w += 4) {
......@@ -227,47 +200,20 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
// right
for (; w < out_width; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h - pad_top,
w - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base,
filter_ptr,
h + 1,
w,
h + 1 - pad_top,
w - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h - pad_top,
w - pad_left, out_width, in_height, in_width, 3,
3, out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h + 1, w, h + 1 - pad_top,
w - pad_left, out_width, in_height, in_width, 3,
3, out_base);
}
} // h
#else
for (index_t ih = valid_h_start; ih < valid_h_stop; ++ih) {
for (index_t iw = 0; iw < out_shape[3]; ++iw) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
ih,
iw,
ih - pad_top,
iw - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, ih, iw, ih - pad_top,
iw - pad_left, out_width, in_height, in_width, 3,
3, out_base);
}
}
#endif
......@@ -275,18 +221,9 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
// bottom
for (; h < out_shape[2]; ++h) {
for (w = 0; w < out_shape[3]; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h - pad_top,
w - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h - pad_top,
w - pad_left, out_width, in_height, in_width, 3,
3, out_base);
}
}
} // m
......@@ -295,9 +232,9 @@ void DepthwiseConv2dNeonK3x3S1(const float *input,
void DepthwiseConv2dNeonK3x3S2(const float *input,
const float *filter,
const index_t* in_shape,
const index_t* out_shape,
const int* pad_hw,
const index_t *in_shape,
const index_t *out_shape,
const int *pad_hw,
const index_t valid_h_start,
const index_t valid_h_stop,
const index_t valid_w_start,
......@@ -330,18 +267,9 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
// top
for (h = 0; h < valid_h_start; ++h) {
for (w = 0; w < out_width; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h * 2 - pad_top,
w * 2 - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h * 2 - pad_top,
w * 2 - pad_left, out_width, in_height, in_width,
3, 3, out_base);
}
}
......@@ -355,18 +283,9 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
for (h = valid_h_start; h < valid_h_stop; ++h) {
// left
for (w = 0; w < valid_w_start; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h * 2 - pad_top,
w * 2 - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h * 2 - pad_top,
w * 2 - pad_left, out_width, in_height, in_width,
3, 3, out_base);
}
for (w = valid_w_start; w + 3 < valid_w_stop; w += 4) {
......@@ -435,35 +354,17 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
// right
for (; w < out_width; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h * 2 - pad_top,
w * 2 - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h * 2 - pad_top,
w * 2 - pad_left, out_width, in_height, in_width,
3, 3, out_base);
}
} // h
#else
for (index_t ih = valid_h_start; ih < valid_h_stop; ++ih) {
for (index_t iw = 0; iw < out_width; ++iw) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
ih,
iw,
ih * 2 - pad_top,
iw * 2 - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, ih, iw, ih * 2 - pad_top,
iw * 2 - pad_left, out_width, in_height,
in_width, 3, 3, out_base);
}
}
#endif
......@@ -471,18 +372,9 @@ void DepthwiseConv2dNeonK3x3S2(const float *input,
// bottom
for (; h < out_shape[2]; ++h) {
for (w = 0; w < out_shape[3]; ++w) {
DepthwiseConv2dPixel(in_base,
filter_ptr,
h,
w,
h * 2 - pad_top,
w * 2 - pad_left,
out_width,
in_height,
in_width,
3,
3,
out_base);
DepthwiseConv2dPixel(in_base, filter_ptr, h, w, h * 2 - pad_top,
w * 2 - pad_left, out_width, in_height, in_width,
3, 3, out_base);
}
}
} // m
......
mace/kernels/channel_shuffle.h
浏览文件 @ 2940ce5d
......@@ -32,7 +32,7 @@ struct ChannelShuffleFunctor {
Tensor *output,
StatsFuture *future) {
MACE_UNUSED(future);
MACE_FAILURE_RETURN(output->ResizeLike(input));
MACE_RETURN_IF_ERROR(output->ResizeLike(input));
Tensor::MappingGuard logits_guard(input);
Tensor::MappingGuard output_guard(output);
......
mace/kernels/concat.h
浏览文件 @ 2940ce5d
......@@ -68,7 +68,7 @@ struct ConcatFunctor : ConcatFunctorBase {
outer_sizes[i] = input->size() / inner_size;
output_shape[axis_] += input->dim(axis_);
}
MACE_FAILURE_RETURN(output->Resize(output_shape));
MACE_RETURN_IF_ERROR(output->Resize(output_shape));
T *output_ptr = output->mutable_data<T>();
......
mace/kernels/conv_2d.h
浏览文件 @ 2940ce5d
......@@ -296,7 +296,7 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
RoundType::FLOOR,
output_shape.data());
}
MACE_FAILURE_RETURN(output->Resize(output_shape));
MACE_RETURN_IF_ERROR(output->Resize(output_shape));
index_t batch = output->dim(0);
index_t channels = output->dim(1);
......@@ -497,7 +497,7 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
if (is_filter_transformed_) {
transformed_filter_ptr = filter_data;
} else {
MACE_FAILURE_RETURN(transformed_filter_.Resize(
MACE_RETURN_IF_ERROR(transformed_filter_.Resize(
transformed_filter_shape));
switch (winograd_out_tile_size) {
case 2:
......@@ -644,7 +644,7 @@ struct Conv2dFunctor<DeviceType::CPU, float> : Conv2dFunctorBase {
const Tensor *pad_input_ptr = input;
if (extra_input_height != input_height
|| extra_input_width != input_width) {
MACE_FAILURE_RETURN(ConstructNCHWInputWithSpecificPadding(input,
MACE_RETURN_IF_ERROR(ConstructNCHWInputWithSpecificPadding(input,
pad_top,
pad_bottom,
pad_left,
......
mace/kernels/conv_pool_2d_util.cc
浏览文件 @ 2940ce5d
......@@ -306,7 +306,7 @@ MaceStatus ConstructNCHWInputWithPadding(const Tensor *input_tensor,
const int padded_top = paddings[0] / 2;
const int padded_left = paddings[1] / 2;
MACE_FAILURE_RETURN(output_tensor->Resize(output_shape));
MACE_RETURN_IF_ERROR(output_tensor->Resize(output_shape));
Tensor::MappingGuard padded_output_mapper(output_tensor);
float *output_data = output_tensor->mutable_data<float>();
......@@ -378,7 +378,7 @@ MaceStatus ConstructNCHWInputWithSpecificPadding(const Tensor *input_tensor,
const int pad_width = pad_left + pad_right;
std::vector<index_t> output_shape(
{batch, channels, height + pad_height, width + pad_width});
MACE_FAILURE_RETURN(output_tensor->Resize(output_shape));
MACE_RETURN_IF_ERROR(output_tensor->Resize(output_shape));
output_tensor->Clear();
Tensor::MappingGuard padded_output_mapper(output_tensor);
float *output_data = output_tensor->mutable_data<float>();
......@@ -428,7 +428,7 @@ MaceStatus ConstructNHWCInputWithPadding(const Tensor *input_tensor,
const int padded_top = paddings[0] / 2;
const int padded_left = paddings[1] / 2;
MACE_FAILURE_RETURN(output_tensor->Resize(output_shape));
MACE_RETURN_IF_ERROR(output_tensor->Resize(output_shape));
Tensor::MappingGuard padded_output_mapper(output_tensor);
float *output_data = output_tensor->mutable_data<float>();
......
mace/kernels/deconv_2d.h
浏览文件 @ 2940ce5d
......@@ -250,7 +250,7 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
strides_, padding_type_,
output_shape.data(),
paddings_.data(), true);
MACE_FAILURE_RETURN(output->Resize(output_shape));
MACE_RETURN_IF_ERROR(output->Resize(output_shape));
} else {
output_shape_.clear();
output_shape_ = std::vector<index_t>(4, 0);
......@@ -259,7 +259,7 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
strides_,
output_shape_.data(),
paddings_.data(), true);
MACE_FAILURE_RETURN(output->Resize(output_shape_));
MACE_RETURN_IF_ERROR(output->Resize(output_shape_));
}
index_t kernel_h = filter->dim(2);
index_t kernel_w = filter->dim(3);
......
mace/kernels/depth_to_space.h
浏览文件 @ 2940ce5d
......@@ -55,7 +55,7 @@ struct DepthToSpaceOpFunctor {
std::vector<index_t> output_shape = {batch_size, output_depth,
output_height, output_width};
MACE_FAILURE_RETURN(output->Resize(output_shape));
MACE_RETURN_IF_ERROR(output->Resize(output_shape));
Tensor::MappingGuard logits_guard(input);
Tensor::MappingGuard output_guard(output);
......
mace/kernels/depthwise_conv2d.h
浏览文件 @ 2940ce5d
......@@ -161,7 +161,7 @@ struct DepthwiseConv2dFunctor<DeviceType::CPU, float>
RoundType::FLOOR,
output_shape.data());
}
MACE_FAILURE_RETURN(output->Resize(output_shape));
MACE_RETURN_IF_ERROR(output->Resize(output_shape));
output->Clear();
index_t batch = output->dim(0);
......
mace/kernels/eltwise.h
浏览文件 @ 2940ce5d
......@@ -494,7 +494,7 @@ struct EltwiseFunctor<DeviceType::CPU, float>: EltwiseFunctorBase {
}
}
}
MACE_FAILURE_RETURN(output->ResizeLike(input0));
MACE_RETURN_IF_ERROR(output->ResizeLike(input0));
Tensor::MappingGuard input0_guard(input0);
Tensor::MappingGuard output_guard(output);
......
mace/kernels/fully_connected.h
浏览文件 @ 2940ce5d
......@@ -57,7 +57,7 @@ struct FullyConnectedFunctor<DeviceType::CPU, float>: FullyConnectedBase {
StatsFuture *future) {
MACE_UNUSED(future);
std::vector<index_t> output_shape = {input->dim(0), weight->dim(0), 1, 1};
MACE_FAILURE_RETURN(output->Resize(output_shape));
MACE_RETURN_IF_ERROR(output->Resize(output_shape));
const index_t N = output->dim(0);
const index_t input_size = weight->dim(1) * weight->dim(2) * weight->dim(3);
const index_t output_size = weight->dim(0);
......
mace/kernels/matmul.h
浏览文件 @ 2940ce5d
......@@ -44,7 +44,7 @@ struct MatMulFunctor {
StatsFuture *future) {
MACE_UNUSED(future);
std::vector<index_t> c_shape = {A->dim(0), A->dim(1), B->dim(2), 1};
MACE_FAILURE_RETURN(C->Resize(c_shape));
MACE_RETURN_IF_ERROR(C->Resize(c_shape));
Tensor::MappingGuard guarda(A);
Tensor::MappingGuard guardb(B);
......
mace/kernels/opencl/activation.cc
浏览文件 @ 2940ce5d
......@@ -21,9 +21,9 @@
namespace mace {
namespace kernels {
template<typename T>
MaceStatus ActivationFunctor<DeviceType::GPU,
T>::operator()(const Tensor *input,
template <typename T>
MaceStatus ActivationFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input,
const Tensor *alpha,
Tensor *output,
StatsFuture *future) {
......@@ -47,7 +47,7 @@ MaceStatus ActivationFunctor<DeviceType::GPU,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -56,22 +56,28 @@ MaceStatus ActivationFunctor<DeviceType::GPU,
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
switch (activation_) {
case RELU:tuning_key_prefix_ = "relu_opencl_kernel";
case RELU:
tuning_key_prefix_ = "relu_opencl_kernel";
built_options.emplace("-DUSE_RELU");
break;
case RELUX:tuning_key_prefix_ = "relux_opencl_kernel";
case RELUX:
tuning_key_prefix_ = "relux_opencl_kernel";
built_options.emplace("-DUSE_RELUX");
break;
case PRELU:tuning_key_prefix_ = "prelu_opencl_kernel";
case PRELU:
tuning_key_prefix_ = "prelu_opencl_kernel";
built_options.emplace("-DUSE_PRELU");
break;
case TANH:tuning_key_prefix_ = "tanh_opencl_kernel";
case TANH:
tuning_key_prefix_ = "tanh_opencl_kernel";
built_options.emplace("-DUSE_TANH");
break;
case SIGMOID:tuning_key_prefix_ = "sigmoid_opencl_kernel";
case SIGMOID:
tuning_key_prefix_ = "sigmoid_opencl_kernel";
built_options.emplace("-DUSE_SIGMOID");
break;
default:LOG(FATAL) << "Unknown activation type: " << activation_;
default:
LOG(FATAL) << "Unknown activation type: " << activation_;
}
kernel_ = runtime->BuildKernel("activation", kernel_name, built_options);
......@@ -121,9 +127,7 @@ MaceStatus ActivationFunctor<DeviceType::GPU,
return MACE_SUCCESS;
}
template
struct ActivationFunctor<DeviceType::GPU, float>;
template
struct ActivationFunctor<DeviceType::GPU, half>;
template struct ActivationFunctor<DeviceType::GPU, float>;
template struct ActivationFunctor<DeviceType::GPU, half>;
} // namespace kernels
} // namespace mace
mace/kernels/opencl/addn.cc
浏览文件 @ 2940ce5d
......@@ -59,7 +59,7 @@ MaceStatus AddNFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -87,8 +87,8 @@ MaceStatus AddNFunctor<DeviceType::GPU, T>::operator()(
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output_tensor->ResizeImage(output_shape,
output_image_shape));
MACE_RETURN_IF_ERROR(
output_tensor->ResizeImage(output_shape, output_image_shape));
uint32_t idx = 0;
if (runtime->IsOutOfRangeCheckEnabled()) {
......
mace/kernels/opencl/batch_norm.cc
浏览文件 @ 2940ce5d
......@@ -23,7 +23,8 @@ namespace mace {
namespace kernels {
template <typename T>
MaceStatus BatchNormFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
MaceStatus BatchNormFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input,
const Tensor *scale,
const Tensor *offset,
const Tensor *mean,
......@@ -57,7 +58,7 @@ MaceStatus BatchNormFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......
mace/kernels/opencl/bias_add.cc
浏览文件 @ 2940ce5d
......@@ -50,7 +50,7 @@ MaceStatus BiasAddFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -91,8 +91,7 @@ MaceStatus BiasAddFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
} else {
std::vector<uint32_t> roundup_gws(lws.size());
for (size_t i = 0; i < lws.size(); ++i) {
if (lws[i] != 0)
roundup_gws[i] = RoundUp(gws[i], lws[i]);
if (lws[i] != 0) roundup_gws[i] = RoundUp(gws[i], lws[i]);
}
error = runtime->command_queue().enqueueNDRangeKernel(
......
mace/kernels/opencl/buffer_to_image.cc
浏览文件 @ 2940ce5d
......@@ -25,14 +25,13 @@ MaceStatus BufferToImageFunctor<DeviceType::GPU, T>::operator()(
const BufferType type,
Tensor *image,
StatsFuture *future) {
std::vector<size_t> image_shape;
CalImage2DShape(buffer->shape(), type, &image_shape);
if (type == WINOGRAD_FILTER) {
std::vector<index_t> new_shape = CalWinogradShape(buffer->shape(), type);
MACE_FAILURE_RETURN(image->ResizeImage(new_shape, image_shape));
MACE_RETURN_IF_ERROR(image->ResizeImage(new_shape, image_shape));
} else {
MACE_FAILURE_RETURN(image->ResizeImage(buffer->shape(), image_shape));
MACE_RETURN_IF_ERROR(image->ResizeImage(buffer->shape(), image_shape));
}
uint32_t gws[2] = {static_cast<uint32_t>(image_shape[0]),
......@@ -94,7 +93,7 @@ MaceStatus BufferToImageFunctor<DeviceType::GPU, T>::operator()(
if (!kernel_error_) {
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -120,8 +119,7 @@ MaceStatus BufferToImageFunctor<DeviceType::GPU, T>::operator()(
static_cast<uint32_t>(buffer->buffer_offset() /
GetEnumTypeSize(buffer->dtype())));
if (type == CONV2D_FILTER) {
const index_t inner_size =
buffer->dim(1) * buffer->dim(2) * buffer->dim(3);
const index_t inner_size = buffer->dim(1) * buffer->dim(2) * buffer->dim(3);
b2f_kernel.setArg(idx++, static_cast<uint32_t>(buffer->dim(0)));
b2f_kernel.setArg(idx++, static_cast<uint32_t>(buffer->dim(2)));
b2f_kernel.setArg(idx++, static_cast<uint32_t>(buffer->dim(3)));
......
mace/kernels/opencl/channel_shuffle.cc
浏览文件 @ 2940ce5d
......@@ -16,18 +16,16 @@
#include "mace/core/runtime/opencl/cl2_header.h"
#include "mace/core/runtime/opencl/opencl_runtime.h"
#include "mace/kernels/opencl/helper.h"
#include "mace/utils/utils.h"
#include "mace/utils/tuner.h"
#include "mace/utils/utils.h"
namespace mace {
namespace kernels {
template <typename T>
MaceStatus ChannelShuffleFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input,
Tensor *output,
StatsFuture *future) {
MACE_FAILURE_RETURN(output->ResizeLike(input));
const Tensor *input, Tensor *output, StatsFuture *future) {
MACE_RETURN_IF_ERROR(output->ResizeLike(input));
const index_t batch = input->dim(0);
const index_t height = input->dim(1);
......@@ -36,8 +34,7 @@ MaceStatus ChannelShuffleFunctor<DeviceType::GPU, T>::operator()(
const index_t channels_per_group = channels / groups_;
MACE_CHECK(channels_per_group % 4 == 0,
"channels per group must be multiple of 4");
MACE_CHECK(groups_ % 4 == 0,
"groups must be multiple of 4");
MACE_CHECK(groups_ % 4 == 0, "groups must be multiple of 4");
const index_t group_channel_blocks = RoundUpDiv4(channels_per_group);
const uint32_t gws[3] = {static_cast<uint32_t>(group_channel_blocks),
......@@ -57,7 +54,7 @@ MaceStatus ChannelShuffleFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -65,8 +62,8 @@ MaceStatus ChannelShuffleFunctor<DeviceType::GPU, T>::operator()(
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
kernel_ = runtime->BuildKernel("channel_shuffle", kernel_name,
built_options);
kernel_ =
runtime->BuildKernel("channel_shuffle", kernel_name, built_options);
kwg_size_ =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(kernel_));
......@@ -93,8 +90,8 @@ MaceStatus ChannelShuffleFunctor<DeviceType::GPU, T>::operator()(
const std::vector<uint32_t> lws = Default3DLocalWS(gws, kwg_size_);
std::string tuning_key =
Concat("channel_shuffle_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("channel_shuffle_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -107,9 +104,7 @@ MaceStatus ChannelShuffleFunctor<DeviceType::GPU, T>::operator()(
return MACE_SUCCESS;
}
template
struct ChannelShuffleFunctor<DeviceType::GPU, float>;
template
struct ChannelShuffleFunctor<DeviceType::GPU, half>;
template struct ChannelShuffleFunctor<DeviceType::GPU, float>;
template struct ChannelShuffleFunctor<DeviceType::GPU, half>;
} // namespace kernels
} // namespace mace
mace/kernels/opencl/concat.cc
浏览文件 @ 2940ce5d
......@@ -22,11 +22,9 @@ namespace mace {
namespace kernels {
namespace {
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
lws[0] = std::min<uint32_t>(base, kwg_size / lws[1]);
......@@ -37,8 +35,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
} // namespace
static void Concat2(cl::Kernel *kernel,
static MaceStatus Concat2(cl::Kernel *kernel,
const Tensor *input0,
const Tensor *input1,
const DataType dt,
......@@ -68,7 +65,7 @@ static void Concat2(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -115,8 +112,8 @@ static void Concat2(cl::Kernel *kernel,
const std::vector<uint32_t> lws = LocalWS(gws, *kwg_size);
std::string tuning_key =
Concat("concat_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("concat_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(*kernel, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -125,9 +122,11 @@ static void Concat2(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
static void ConcatN(cl::Kernel *kernel,
static MaceStatus ConcatN(cl::Kernel *kernel,
const std::vector<const Tensor *> &input_list,
const DataType dt,
Tensor *output,
......@@ -150,7 +149,7 @@ static void ConcatN(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -218,8 +217,8 @@ static void ConcatN(cl::Kernel *kernel,
if (runtime->is_profiling_enabled()) {
CallStats tmp_stats;
runtime->GetCallStats(event, &tmp_stats);
call_stats.start_micros = std::min<int64_t>(tmp_stats.start_micros,
call_stats.start_micros);
call_stats.start_micros =
std::min<int64_t>(tmp_stats.start_micros, call_stats.start_micros);
call_stats.end_micros += tmp_stats.end_micros - tmp_stats.start_micros;
}
}
......@@ -232,6 +231,8 @@ static void ConcatN(cl::Kernel *kernel,
}
};
}
return MACE_SUCCESS;
}
template <typename T>
......@@ -266,17 +267,17 @@ MaceStatus ConcatFunctor<DeviceType::GPU, T>::operator()(
"Dimensions of inputs should be divisible by 4 when inputs_count > 2.");
std::vector<size_t> image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL, &image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, image_shape));
switch (inputs_count) {
case 2:
Concat2(&kernel_, input_list[0], input_list[1], DataTypeToEnum<T>::value,
&input_shape_, output, future, &kwg_size_, &kernel_error_);
break;
return Concat2(&kernel_, input_list[0], input_list[1],
DataTypeToEnum<T>::value, &input_shape_, output, future,
&kwg_size_, &kernel_error_);
default:
if (divisible_four) {
ConcatN(&kernel_, input_list, DataTypeToEnum<T>::value, output, future,
&kwg_size_, &kernel_error_);
return ConcatN(&kernel_, input_list, DataTypeToEnum<T>::value, output,
future, &kwg_size_, &kernel_error_);
} else {
MACE_NOT_IMPLEMENTED;
}
......
mace/kernels/opencl/conv_2d.cc
浏览文件 @ 2940ce5d
......@@ -18,7 +18,7 @@
namespace mace {
namespace kernels {
extern void Conv2dOpenclK1x1(cl::Kernel *kernel,
extern MaceStatus Conv2dOpenclK1x1(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -34,7 +34,7 @@ extern void Conv2dOpenclK1x1(cl::Kernel *kernel,
uint32_t *kwg_size,
std::unique_ptr<BufferBase> *kernel_error);
extern void Conv2dOpenclK3x3(cl::Kernel *kernel,
extern MaceStatus Conv2dOpenclK3x3(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -50,7 +50,7 @@ extern void Conv2dOpenclK3x3(cl::Kernel *kernel,
uint32_t *kwg_size,
std::unique_ptr<BufferBase> *kernel_error);
extern void Conv2dOpencl(cl::Kernel *kernel,
extern MaceStatus Conv2dOpencl(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -72,7 +72,7 @@ MaceStatus Conv2dFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
const Tensor *bias,
Tensor *output,
StatsFuture *future) {
typedef void (*Conv2dOpenclFunction)(
typedef MaceStatus (*Conv2dOpenclFunction)(
cl::Kernel * kernel, const Tensor *input, const Tensor *filter,
const Tensor *bias, const int stride, const int *padding,
const int *dilations, const ActivationType activation,
......@@ -111,23 +111,21 @@ MaceStatus Conv2dFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, output_image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, output_image_shape));
if (kernel_h == kernel_w && kernel_h <= 5 &&
selector[kernel_h - 1] != nullptr) {
auto conv2d_func = selector[kernel_h - 1];
conv2d_func(&kernel_, input, filter, bias, strides_[0], paddings.data(),
dilations_, activation_, relux_max_limit_,
DataTypeToEnum<T>::value, &input_shape_, output, future,
&kwg_size_, &kernel_error_);
return conv2d_func(
&kernel_, input, filter, bias, strides_[0], paddings.data(), dilations_,
activation_, relux_max_limit_, DataTypeToEnum<T>::value, &input_shape_,
output, future, &kwg_size_, &kernel_error_);
} else {
Conv2dOpencl(&kernel_, input, filter, bias, strides_[0], paddings.data(),
dilations_, activation_, relux_max_limit_,
DataTypeToEnum<T>::value, &input_shape_, output, future,
&kwg_size_, &kernel_error_);
return Conv2dOpencl(
&kernel_, input, filter, bias, strides_[0], paddings.data(), dilations_,
activation_, relux_max_limit_, DataTypeToEnum<T>::value, &input_shape_,
output, future, &kwg_size_, &kernel_error_);
}
return MACE_SUCCESS;
}
template struct Conv2dFunctor<DeviceType::GPU, float>;
......
mace/kernels/opencl/conv_2d_1x1.cc
浏览文件 @ 2940ce5d
......@@ -12,8 +12,8 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "mace/kernels/conv_2d.h"
#include "mace/core/runtime/opencl/opencl_runtime.h"
#include "mace/kernels/conv_2d.h"
#include "mace/kernels/opencl/helper.h"
#include "mace/utils/tuner.h"
......@@ -25,11 +25,9 @@ namespace {
const uint32_t kernel_cache_size = (4 + 4 + 4) * 4 * 4;
// TODO(liuqi): Fix the specific value.
const uint32_t lws_limit = 128;
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t compute_units = OpenCLRuntime::Global()->device_compute_units();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
......@@ -46,8 +44,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
lws[0] = std::min<uint32_t>(lws[0], kwg_size / lws[1]);
const uint32_t lws_size = lws[0] * lws[1];
lws[2] = std::min<uint32_t>(
(cache_size / kernel_cache_size / lws_size / compute_units) * 8,
gws[2]);
(cache_size / kernel_cache_size / lws_size / compute_units) * 8, gws[2]);
if (lws[2] == 0) {
lws[2] = std::min<uint32_t>(gws[2], base);
}
......@@ -57,7 +54,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
} // namespace
extern void Conv2dOpenclK1x1(cl::Kernel *kernel,
extern MaceStatus Conv2dOpenclK1x1(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -101,7 +98,7 @@ extern void Conv2dOpenclK1x1(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -172,8 +169,8 @@ extern void Conv2dOpenclK1x1(cl::Kernel *kernel,
std::vector<uint32_t> lws = LocalWS(gws, *kwg_size);
std::string tuning_key =
Concat("conv2d_1x1_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("conv2d_1x1_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(*kernel, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -182,6 +179,8 @@ extern void Conv2dOpenclK1x1(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
} // namespace kernels
......
mace/kernels/opencl/conv_2d_3x3.cc
浏览文件 @ 2940ce5d
......@@ -12,9 +12,9 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "mace/kernels/conv_2d.h"
#include "mace/core/runtime/opencl/opencl_runtime.h"
#include "mace/kernels/activation.h"
#include "mace/kernels/conv_2d.h"
#include "mace/kernels/opencl/helper.h"
#include "mace/utils/tuner.h"
#include "mace/utils/utils.h"
......@@ -24,22 +24,20 @@ namespace kernels {
namespace {
// (inputs + weights + outputs) * array_size * sizeof(float)
const uint32_t kernel_cache_size = (5 + 4 + 5) * 4 * 4;
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t compute_units = std::max<uint32_t>(
OpenCLRuntime::Global()->device_compute_units() / 2, 1);
const uint32_t base = std::min<uint32_t>(cache_size / kBaseGPUMemCacheSize,
4);
const uint32_t base =
std::min<uint32_t>(cache_size / kBaseGPUMemCacheSize, 4);
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
lws[0] = std::min<uint32_t>(std::min<uint32_t>(gws[0], base),
kwg_size / lws[1]);
lws[0] =
std::min<uint32_t>(std::min<uint32_t>(gws[0], base), kwg_size / lws[1]);
const uint32_t lws_size = lws[0] * lws[1];
lws[2] = std::min<uint32_t>(
RoundUp<uint32_t>(cache_size / kernel_cache_size /
lws_size / compute_units, base),
RoundUp<uint32_t>(
cache_size / kernel_cache_size / lws_size / compute_units, base),
gws[2]);
if (lws[2] == 0) {
lws[2] = std::min<uint32_t>(gws[2], base);
......@@ -50,7 +48,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
} // namespace
extern void Conv2dOpenclK3x3(cl::Kernel *kernel,
extern MaceStatus Conv2dOpenclK3x3(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -87,7 +85,7 @@ extern void Conv2dOpenclK3x3(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -159,8 +157,8 @@ extern void Conv2dOpenclK3x3(cl::Kernel *kernel,
std::vector<uint32_t> lws = LocalWS(gws, *kwg_size);
std::string tuning_key =
Concat("conv2d_3x3_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("conv2d_3x3_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(*kernel, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -169,6 +167,8 @@ extern void Conv2dOpenclK3x3(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
} // namespace kernels
......
mace/kernels/opencl/conv_2d_general.cc
浏览文件 @ 2940ce5d
......@@ -12,9 +12,9 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "mace/kernels/conv_2d.h"
#include "mace/core/runtime/opencl/opencl_runtime.h"
#include "mace/kernels/activation.h"
#include "mace/kernels/conv_2d.h"
#include "mace/kernels/opencl/helper.h"
#include "mace/utils/tuner.h"
#include "mace/utils/utils.h"
......@@ -30,8 +30,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kernel_size,
const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t compute_units = OpenCLRuntime::Global()->device_compute_units();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
......@@ -41,9 +40,9 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
}
lws[0] = std::min<uint32_t>(lws[0], kwg_size / lws[1]);
const uint32_t lws_size = lws[0] * lws[1];
lws[2] = std::min<uint32_t>(
(cache_size / kernel_cache_size / kernel_size / lws_size / compute_units)
* 8,
lws[2] = std::min<uint32_t>((cache_size / kernel_cache_size / kernel_size /
lws_size / compute_units) *
8,
gws[2]);
if (lws[2] == 0) {
if (gws[2] < lws_limit) {
......@@ -58,7 +57,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
} // namespace
extern void Conv2dOpencl(cl::Kernel *kernel,
extern MaceStatus Conv2dOpencl(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -95,7 +94,7 @@ extern void Conv2dOpencl(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -168,9 +167,8 @@ extern void Conv2dOpencl(cl::Kernel *kernel,
}
std::string tuning_key =
Concat("conv2d_general_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3),
filter->dim(2), filter->dim(3));
Concat("conv2d_general_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3), filter->dim(2), filter->dim(3));
std::vector<uint32_t> lws =
LocalWS(gws, filter->dim(2) * filter->dim(3), *kwg_size);
TuningOrRun3DKernel(*kernel, tuning_key, gws, lws, future);
......@@ -181,6 +179,8 @@ extern void Conv2dOpencl(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
} // namespace kernels
......
mace/kernels/opencl/deconv_2d_opencl.cc
浏览文件 @ 2940ce5d
......@@ -20,7 +20,7 @@ namespace kernels {
namespace {
void Deconv2dOpencl(cl::Kernel *kernel,
MaceStatus Deconv2dOpencl(cl::Kernel *kernel,
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
......@@ -46,10 +46,10 @@ void Deconv2dOpencl(cl::Kernel *kernel,
#define MACE_WIDTH_BLK 5
const index_t n_strides = (width + stride - 1) / stride;
const index_t width_blocks =
((n_strides + MACE_WIDTH_BLK -1)/ MACE_WIDTH_BLK) * stride;
((n_strides + MACE_WIDTH_BLK - 1) / MACE_WIDTH_BLK) * stride;
const float stride_r = 1.f / static_cast<float>(stride);
const int padding_h = (paddings[0]+1) >> 1;
const int padding_w = (paddings[0]+1) >> 1;
const int padding_h = (paddings[0] + 1) >> 1;
const int padding_w = (paddings[0] + 1) >> 1;
const int align_h = stride - 1 - padding_h;
const int align_w = stride - 1 - padding_w;
......@@ -67,7 +67,7 @@ void Deconv2dOpencl(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -77,16 +77,22 @@ void Deconv2dOpencl(cl::Kernel *kernel,
}
built_options.emplace(bias != nullptr ? "-DBIAS" : "");
switch (activation) {
case NOOP:break;
case RELU:built_options.emplace("-DUSE_RELU");
case NOOP:
break;
case RELUX:built_options.emplace("-DUSE_RELUX");
case RELU:
built_options.emplace("-DUSE_RELU");
break;
case TANH:built_options.emplace("-DUSE_TANH");
case RELUX:
built_options.emplace("-DUSE_RELUX");
break;
case SIGMOID:built_options.emplace("-DUSE_SIGMOID");
case TANH:
built_options.emplace("-DUSE_TANH");
break;
default:LOG(FATAL) << "Unknown activation type: " << activation;
case SIGMOID:
built_options.emplace("-DUSE_SIGMOID");
break;
default:
LOG(FATAL) << "Unknown activation type: " << activation;
}
*kernel = runtime->BuildKernel("deconv_2d", kernel_name, built_options);
......@@ -150,12 +156,15 @@ void Deconv2dOpencl(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
} // namespace
template <typename T>
MaceStatus Deconv2dFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
MaceStatus Deconv2dFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input,
const Tensor *filter,
const Tensor *bias,
Tensor *output,
......@@ -167,34 +176,25 @@ MaceStatus Deconv2dFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
if (output_shape_.size() == 4) {
paddings_.clear();
paddings_ = std::vector<int>(2, 0);
CalcDeconvPaddingAndInputSize(
input->shape().data(),
filter->shape().data(),
strides_, padding_type_,
output_shape_.data(),
CalcDeconvPaddingAndInputSize(input->shape().data(), filter->shape().data(),
strides_, padding_type_, output_shape_.data(),
paddings_.data());
} else {
output_shape_.clear();
output_shape_ = std::vector<index_t>(4, 0);
CalcDeconvOutputSize(input->shape().data(),
filter->shape().data(),
strides_,
output_shape_.data(),
paddings_.data());
CalcDeconvOutputSize(input->shape().data(), filter->shape().data(),
strides_, output_shape_.data(), paddings_.data());
}
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape_, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape_, output_image_shape));
Deconv2dOpencl(&kernel_, input, filter, bias,
strides_[0], paddings_.data(),
activation_, relux_max_limit_,
DataTypeToEnum<T>::value, &input_shape_,
output, future, &kwg_size_, &kernel_error_);
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape_, output_image_shape));
return MACE_SUCCESS;
return Deconv2dOpencl(&kernel_, input, filter, bias, strides_[0],
paddings_.data(), activation_, relux_max_limit_,
DataTypeToEnum<T>::value, &input_shape_, output, future,
&kwg_size_, &kernel_error_);
}
template struct Deconv2dFunctor<DeviceType::GPU, float>;
......
mace/kernels/opencl/depth_to_space.cc
浏览文件 @ 2940ce5d
......@@ -70,7 +70,7 @@ MaceStatus DepthToSpaceOpFunctor<DeviceType::GPU, T>::operator()(
std::vector<size_t> image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL, &image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, image_shape));
auto runtime = OpenCLRuntime::Global();
......@@ -87,7 +87,7 @@ MaceStatus DepthToSpaceOpFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -95,9 +95,8 @@ MaceStatus DepthToSpaceOpFunctor<DeviceType::GPU, T>::operator()(
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
kernel_ =
runtime->BuildKernel("depth_to_space",
obfuscated_kernel_name, built_options);
kernel_ = runtime->BuildKernel("depth_to_space", obfuscated_kernel_name,
built_options);
kwg_size_ =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(kernel_));
......
mace/kernels/opencl/depthwise_conv.cc
浏览文件 @ 2940ce5d
......@@ -24,8 +24,7 @@ namespace kernels {
namespace {
// (inputs + weights + outputs) * array_size * sizeof(float)
const uint32_t kernel_cache_size = (4 + 4 + 1) * 4 * 4;
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t min_lws0 = cache_size / kBaseGPUMemCacheSize;
......@@ -40,8 +39,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
}
}
const uint32_t lws_size = lws[0] * lws[1];
lws[2] = std::min<uint32_t>(
(cache_size / kernel_cache_size / lws_size) * 4,
lws[2] = std::min<uint32_t>((cache_size / kernel_cache_size / lws_size) * 4,
gws[2]);
if (lws[2] == 0) {
lws[2] = gws[2];
......@@ -52,7 +50,7 @@ std::vector<uint32_t> LocalWS(const uint32_t *gws,
} // namespace
static void DepthwiseConv2d(cl::Kernel *kernel,
static MaceStatus DepthwiseConv2d(cl::Kernel *kernel,
const Tensor *input, // NHWC
const Tensor *filter, // HWIM
const Tensor *bias,
......@@ -98,7 +96,7 @@ static void DepthwiseConv2d(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -181,8 +179,8 @@ static void DepthwiseConv2d(cl::Kernel *kernel,
}
const std::vector<uint32_t> lws = LocalWS(gws, *kwg_size);
std::string tuning_key = Concat("depthwise_conv2d_ocl_kernel",
gws[0], gws[1], gws[2], multiplier);
std::string tuning_key =
Concat("depthwise_conv2d_ocl_kernel", gws[0], gws[1], gws[2], multiplier);
TuningOrRun3DKernel(*kernel, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -191,6 +189,8 @@ static void DepthwiseConv2d(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
template <typename T>
......@@ -200,7 +200,6 @@ MaceStatus DepthwiseConv2dFunctor<DeviceType::GPU, T>::operator()(
const Tensor *bias,
Tensor *output,
StatsFuture *future) {
index_t kernel_h = filter->dim(2);
index_t kernel_w = filter->dim(3);
if (strides_[0] != strides_[1]) {
......@@ -237,14 +236,12 @@ MaceStatus DepthwiseConv2dFunctor<DeviceType::GPU, T>::operator()(
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, output_image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, output_image_shape));
DepthwiseConv2d(&kernel_, input, filter, bias, strides_[0], paddings.data(),
dilations_, activation_, relux_max_limit_,
DataTypeToEnum<T>::value, &input_shape_, output, future,
&kwg_size_, &kernel_error_);
return MACE_SUCCESS;
return DepthwiseConv2d(
&kernel_, input, filter, bias, strides_[0], paddings.data(), dilations_,
activation_, relux_max_limit_, DataTypeToEnum<T>::value, &input_shape_,
output, future, &kwg_size_, &kernel_error_);
}
template struct DepthwiseConv2dFunctor<DeviceType::GPU, float>;
......
mace/kernels/opencl/eltwise.cc
浏览文件 @ 2940ce5d
......@@ -28,9 +28,8 @@ MaceStatus EltwiseFunctor<DeviceType::GPU, T>::operator()(const Tensor *input0,
MACE_UNUSED(future);
bool swapped = false;
if (input1 != nullptr) {
MACE_CHECK(input0->dim_size() == input1->dim_size()
|| input0->dim_size() == 1
|| input1->dim_size() == 1)
MACE_CHECK(input0->dim_size() == input1->dim_size() ||
input0->dim_size() == 1 || input1->dim_size() == 1)
<< "Inputs of Eltwise op must be same shape";
if (input0->size() != input1->size()) {
if (input0->size() < input1->size()) {
......@@ -42,25 +41,23 @@ MaceStatus EltwiseFunctor<DeviceType::GPU, T>::operator()(const Tensor *input0,
<< "Element-Wise op only support channel dimension broadcast";
} else {
MACE_CHECK((input0->dim(0) == input1->dim(0) || input1->dim(0) == 1) &&
input0->dim(3) == input1->dim(3) &&
input1->dim(1) == 1 &&
input0->dim(3) == input1->dim(3) && input1->dim(1) == 1 &&
input1->dim(2) == 1)
<< "Element-Wise op only support channel dimension broadcast";
}
}
}
std::vector<index_t > output_shape(4);
std::vector<index_t> output_shape(4);
output_shape[0] = input0->dim(0);
output_shape[1] = input0->dim(1);
output_shape[2] = input0->dim(2);
output_shape[3] = input0->dim(3);
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape,
BufferType::IN_OUT_CHANNEL,
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, output_image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, output_image_shape));
const index_t batch = output->dim(0);
const index_t height = output->dim(1);
......@@ -98,7 +95,7 @@ MaceStatus EltwiseFunctor<DeviceType::GPU, T>::operator()(const Tensor *input0,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -142,8 +139,8 @@ MaceStatus EltwiseFunctor<DeviceType::GPU, T>::operator()(const Tensor *input0,
const std::vector<uint32_t> lws = Default3DLocalWS(gws, kwg_size_);
std::string tuning_key =
Concat("eltwise_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("eltwise_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
kernel_error_->Map(nullptr);
......
mace/kernels/opencl/fully_connected.cc
浏览文件 @ 2940ce5d
......@@ -20,7 +20,7 @@ namespace kernels {
namespace {
template <typename T>
void FCWXKernel(cl::Kernel *kernel,
MaceStatus FCWXKernel(cl::Kernel *kernel,
const Tensor *input,
const Tensor *weight,
const Tensor *bias,
......@@ -75,7 +75,7 @@ void FCWXKernel(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -170,10 +170,12 @@ void FCWXKernel(cl::Kernel *kernel,
}
};
}
return MACE_SUCCESS;
}
template <typename T>
void FCWTXKernel(cl::Kernel *kernel,
MaceStatus FCWTXKernel(cl::Kernel *kernel,
const Tensor *input,
const Tensor *weight,
const Tensor *bias,
......@@ -202,7 +204,7 @@ void FCWTXKernel(cl::Kernel *kernel,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
*kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
(*kernel_error)->Allocate(1);
MACE_RETURN_IF_ERROR((*kernel_error)->Allocate(1));
(*kernel_error)->Map(nullptr);
*((*kernel_error)->mutable_data<char>()) = 0;
(*kernel_error)->UnMap();
......@@ -233,7 +235,7 @@ void FCWTXKernel(cl::Kernel *kernel,
uint32_t kwg_size =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(*kernel));
*lws = {16, kwg_size/16, 0};
*lws = {16, kwg_size / 16, 0};
}
if (!IsVecEqual(*prev_input_shape, input->shape())) {
const index_t batch = output->dim(0);
......@@ -268,8 +270,8 @@ void FCWTXKernel(cl::Kernel *kernel,
}
std::string tuning_key =
Concat("fc_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("fc_opencl_kernel", output->dim(0), output->dim(1), output->dim(2),
output->dim(3));
TuningOrRun2DKernel(*kernel, tuning_key, gws->data(), *lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -278,6 +280,8 @@ void FCWTXKernel(cl::Kernel *kernel,
MACE_CHECK(*kerror_code == 0) << "Kernel error code: " << *kerror_code;
(*kernel_error)->UnMap();
}
return MACE_SUCCESS;
}
} // namespace
......@@ -292,13 +296,11 @@ MaceStatus FullyConnectedFunctor<DeviceType::GPU, T>::operator()(
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, output_image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, output_image_shape));
FCWXKernel<T>(&kernel_, input, weight, bias, &input_shape_, output,
return FCWXKernel<T>(&kernel_, input, weight, bias, &input_shape_, output,
activation_, &gws_, &lws_, relux_max_limit_, future,
&kernel_error_);
return MACE_SUCCESS;
}
template struct FullyConnectedFunctor<DeviceType::GPU, float>;
......
mace/kernels/opencl/helper.cc
浏览文件 @ 2940ce5d
......@@ -209,12 +209,11 @@ std::string DtToUpstreamCLCMDDt(const DataType dt) {
std::vector<uint32_t> Default3DLocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
lws[2] = std::min<uint32_t>(std::min<uint32_t>(gws[2], base),
kwg_size / lws[1]);
lws[2] =
std::min<uint32_t>(std::min<uint32_t>(gws[2], base), kwg_size / lws[1]);
const uint32_t lws_size = lws[1] * lws[2];
lws[0] = std::min<uint32_t>(base, kwg_size / lws_size);
return lws;
......@@ -278,7 +277,7 @@ void TuningOrRun3DKernel(const cl::Kernel &kernel,
MACE_CHECK(params.size() == 4)
<< "Tuning parameters of 3D kernel must be 4D";
cl_int error = CL_SUCCESS;
std::vector<uint32_t> internal_gws(gws, gws+3);
std::vector<uint32_t> internal_gws(gws, gws + 3);
if (!runtime->IsNonUniformWorkgroupsSupported()) {
for (size_t i = 0; i < 3; ++i) {
internal_gws[i] = RoundUp(gws[i], params[i]);
......@@ -287,12 +286,12 @@ void TuningOrRun3DKernel(const cl::Kernel &kernel,
if (timer == nullptr) {
uint32_t block_size = params[3] == 0 ? internal_gws[2] : params[3];
const uint32_t num_blocks = RoundUpDiv<uint32_t>(internal_gws[2],
block_size);
const uint32_t num_blocks =
RoundUpDiv<uint32_t>(internal_gws[2], block_size);
for (uint32_t i = 0; i < num_blocks; ++i) {
uint32_t gws2 = block_size;
if (runtime->IsNonUniformWorkgroupsSupported()
&& (i == num_blocks - 1)) {
if (runtime->IsNonUniformWorkgroupsSupported() &&
(i == num_blocks - 1)) {
gws2 = (internal_gws[2] - (i * block_size));
}
error = runtime->command_queue().enqueueNDRangeKernel(
......@@ -324,8 +323,8 @@ void TuningOrRun3DKernel(const cl::Kernel &kernel,
num_blocks = RoundUpDiv<uint32_t>(internal_gws[2], block_size);
for (uint32_t i = 0; i < num_blocks; ++i) {
uint32_t gws2 = block_size;
if (runtime->IsNonUniformWorkgroupsSupported()
&& (i == num_blocks - 1)) {
if (runtime->IsNonUniformWorkgroupsSupported() &&
(i == num_blocks - 1)) {
gws2 = (internal_gws[2] - (i * block_size));
}
error = runtime->command_queue().enqueueNDRangeKernel(
......@@ -365,17 +364,11 @@ void TuningOrRun2DKernel(const cl::Kernel &kernel,
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(kernel));
std::vector<std::vector<uint32_t>> results;
std::vector<std::vector<uint32_t>> candidates = {
{kwg_size / 2, 2, 0},
{kwg_size / 4, 4, 0},
{kwg_size / 8, 8, 0},
{kwg_size / 16, 16, 0},
{kwg_size / 32, 32, 0},
{kwg_size / 64, 64, 0},
{kwg_size / 128, 128, 0},
{kwg_size / 256, 256, 0},
{kwg_size, 1, 0},
{1, kwg_size, 0}
};
{kwg_size / 2, 2, 0}, {kwg_size / 4, 4, 0},
{kwg_size / 8, 8, 0}, {kwg_size / 16, 16, 0},
{kwg_size / 32, 32, 0}, {kwg_size / 64, 64, 0},
{kwg_size / 128, 128, 0}, {kwg_size / 256, 256, 0},
{kwg_size, 1, 0}, {1, kwg_size, 0}};
for (auto &ele : candidates) {
const uint32_t tmp = ele[0] * ele[1] * ele[2];
if (0 < tmp && tmp <= kwg_size) {
......@@ -390,7 +383,7 @@ void TuningOrRun2DKernel(const cl::Kernel &kernel,
MACE_CHECK(params.size() == 3)
<< "Tuning parameters of 2D kernel must be 3d";
cl_int error = CL_SUCCESS;
std::vector<uint32_t> internal_gws(gws, gws+2);
std::vector<uint32_t> internal_gws(gws, gws + 2);
if (!runtime->IsNonUniformWorkgroupsSupported()) {
for (size_t i = 0; i < 2; ++i) {
internal_gws[i] = RoundUp(gws[i], params[i]);
......@@ -399,12 +392,12 @@ void TuningOrRun2DKernel(const cl::Kernel &kernel,
if (timer == nullptr) {
uint32_t block_size = params[2] == 0 ? internal_gws[1] : params[2];
const uint32_t num_blocks = RoundUpDiv<uint32_t>(internal_gws[1],
block_size);
const uint32_t num_blocks =
RoundUpDiv<uint32_t>(internal_gws[1], block_size);
for (uint32_t i = 0; i < num_blocks; ++i) {
uint32_t gws1 = block_size;
if (runtime->IsNonUniformWorkgroupsSupported()
&& (i == num_blocks - 1)) {
if (runtime->IsNonUniformWorkgroupsSupported() &&
(i == num_blocks - 1)) {
gws1 = (internal_gws[1] - (i * block_size));
}
error = runtime->command_queue().enqueueNDRangeKernel(
......@@ -435,8 +428,8 @@ void TuningOrRun2DKernel(const cl::Kernel &kernel,
num_blocks = RoundUpDiv<uint32_t>(internal_gws[1], block_size);
for (uint32_t i = 0; i < num_blocks; ++i) {
uint32_t gws1 = block_size;
if (runtime->IsNonUniformWorkgroupsSupported()
&& (i == num_blocks - 1)) {
if (runtime->IsNonUniformWorkgroupsSupported() &&
(i == num_blocks - 1)) {
gws1 = (internal_gws[1] - (i * block_size));
}
error = runtime->command_queue().enqueueNDRangeKernel(
......@@ -463,6 +456,5 @@ void TuningOrRun2DKernel(const cl::Kernel &kernel,
}
}
} // namespace kernels
} // namespace mace
mace/kernels/opencl/helper.h
浏览文件 @ 2940ce5d
......@@ -88,8 +88,7 @@ inline bool LimitKernelTime() {
}
template <typename T>
bool IsVecEqual(const std::vector<T> &input0,
const std::vector<T> &input1) {
bool IsVecEqual(const std::vector<T> &input0, const std::vector<T> &input1) {
return ((input0.size() == input1.size()) &&
(std::equal(input0.begin(), input0.end(), input1.begin())));
}
......
mace/kernels/opencl/image_to_buffer.cc
浏览文件 @ 2940ce5d
......@@ -25,10 +25,9 @@ MaceStatus ImageToBufferFunctor<DeviceType::GPU, T>::operator()(
const BufferType type,
Tensor *buffer,
StatsFuture *future) {
std::vector<size_t> image_shape;
CalImage2DShape(image->shape(), type, &image_shape);
MACE_FAILURE_RETURN(buffer->Resize(image->shape()));
MACE_RETURN_IF_ERROR(buffer->Resize(image->shape()));
uint32_t gws[2] = {static_cast<uint32_t>(image_shape[0]),
static_cast<uint32_t>(image_shape[1])};
......@@ -87,7 +86,7 @@ MaceStatus ImageToBufferFunctor<DeviceType::GPU, T>::operator()(
if (!kernel_error_) {
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -108,8 +107,7 @@ MaceStatus ImageToBufferFunctor<DeviceType::GPU, T>::operator()(
}
b2f_kernel.setArg(idx++, *(buffer->opencl_buffer()));
if (type == CONV2D_FILTER) {
const index_t inner_size =
buffer->dim(1) * buffer->dim(2) * buffer->dim(3);
const index_t inner_size = buffer->dim(1) * buffer->dim(2) * buffer->dim(3);
b2f_kernel.setArg(idx++, static_cast<uint32_t>(buffer->dim(0)));
b2f_kernel.setArg(idx++, static_cast<uint32_t>(buffer->dim(2)));
b2f_kernel.setArg(idx++, static_cast<uint32_t>(buffer->dim(3)));
......
mace/kernels/opencl/matmul.cc
浏览文件 @ 2940ce5d
......@@ -29,7 +29,7 @@ MaceStatus MatMulFunctor<DeviceType::GPU, T>::operator()(const Tensor *A,
std::vector<index_t> c_shape = {A->dim(0), A->dim(1), B->dim(2), 1};
std::vector<size_t> c_image_shape;
CalImage2DShape(c_shape, BufferType::IN_OUT_HEIGHT, &c_image_shape);
MACE_FAILURE_RETURN(C->ResizeImage(c_shape, c_image_shape));
MACE_RETURN_IF_ERROR(C->ResizeImage(c_shape, c_image_shape));
const index_t batch = C->dim(0);
const index_t height = C->dim(1);
......@@ -55,7 +55,7 @@ MaceStatus MatMulFunctor<DeviceType::GPU, T>::operator()(const Tensor *A,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -87,9 +87,8 @@ MaceStatus MatMulFunctor<DeviceType::GPU, T>::operator()(const Tensor *A,
kernel_.setArg(idx++, static_cast<int>(RoundUpDiv4(A->dim(2))));
const std::vector<uint32_t> lws = {kwg_size_ / 64, 64, 0};
std::string tuning_key =
Concat("matmul_opencl_kernel", C->dim(0),
C->dim(1), C->dim(2), C->dim(3));
std::string tuning_key = Concat("matmul_opencl_kernel", C->dim(0), C->dim(1),
C->dim(2), C->dim(3));
TuningOrRun2DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......
mace/kernels/opencl/out_of_range_check_test.cc
浏览文件 @ 2940ce5d
......@@ -58,7 +58,7 @@ bool BufferToImageOpImpl(Tensor *buffer,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error->Allocate(1));
kernel_error->Map(nullptr);
*(kernel_error->mutable_data<char>()) = 0;
kernel_error->UnMap();
......@@ -113,8 +113,7 @@ bool BufferToImageOpImpl(Tensor *buffer,
bool is_out_of_range = false;
if (runtime->IsOutOfRangeCheckEnabled()) {
kernel_error->Map(nullptr);
is_out_of_range =
*(kernel_error->mutable_data<char>()) == 1 ? true : false;
is_out_of_range = *(kernel_error->mutable_data<char>()) == 1 ? true : false;
kernel_error->UnMap();
}
return is_out_of_range;
......@@ -124,9 +123,7 @@ bool BufferToImageOpImpl(Tensor *buffer,
class OutOfRangeCheckTest : public ::testing::Test {
protected:
virtual void SetUp() {
setenv("OUT_OF_RANGE_CHECK", "1", 1);
}
virtual void SetUp() { setenv("OUT_OF_RANGE_CHECK", "1", 1); }
};
TEST(OutOfRangeCheckTest, RandomTest) {
......@@ -137,14 +134,13 @@ TEST(OutOfRangeCheckTest, RandomTest) {
std::vector<index_t> buffer_shape = {batch, height, width, channels};
Workspace ws;
Tensor *buffer = ws.CreateTensor("Buffer",
GetDeviceAllocator(DeviceType::GPU),
Tensor *buffer =
ws.CreateTensor("Buffer", GetDeviceAllocator(DeviceType::GPU),
DataTypeToEnum<float>::v());
buffer->Resize(buffer_shape);
std::vector<size_t> image_shape;
Tensor *image = ws.CreateTensor("Image",
GetDeviceAllocator(DeviceType::GPU),
Tensor *image = ws.CreateTensor("Image", GetDeviceAllocator(DeviceType::GPU),
DataTypeToEnum<float>::v());
CalImage2DShape(buffer->shape(), IN_OUT_CHANNEL, &image_shape);
image->ResizeImage(buffer->shape(), image_shape);
......
mace/kernels/opencl/pad.cc
浏览文件 @ 2940ce5d
......@@ -20,26 +20,25 @@
namespace mace {
namespace kernels {
template<typename T>
MaceStatus PadFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input,
template <typename T>
MaceStatus PadFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
Tensor *output,
StatsFuture *future) {
MACE_CHECK(
this->paddings_.size() == static_cast<size_t>((input->dim_size() * 2)));
MACE_CHECK((this->paddings_[0] == 0) && (this->paddings_[1] == 0)
&& (this->paddings_[6] == 0) && (this->paddings_[7] == 0))
MACE_CHECK(this->paddings_.size() ==
static_cast<size_t>((input->dim_size() * 2)));
MACE_CHECK((this->paddings_[0] == 0) && (this->paddings_[1] == 0) &&
(this->paddings_[6] == 0) && (this->paddings_[7] == 0))
<< "Mace only support height/width dimension now";
auto input_shape = input->shape();
std::vector<index_t>
output_shape = {input_shape[0] + this->paddings_[0] + this->paddings_[1],
std::vector<index_t> output_shape = {
input_shape[0] + this->paddings_[0] + this->paddings_[1],
input_shape[1] + this->paddings_[2] + this->paddings_[3],
input_shape[2] + this->paddings_[4] + this->paddings_[5],
input_shape[3] + this->paddings_[6] + this->paddings_[7]};
std::vector<size_t> image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL, &image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, image_shape));
const index_t batch = output->dim(0);
const index_t height = output->dim(1);
......@@ -61,7 +60,7 @@ MaceStatus PadFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -103,9 +102,8 @@ MaceStatus PadFunctor<DeviceType::GPU, T>::operator()(
}
const std::vector<uint32_t> lws = Default3DLocalWS(gws, kwg_size_);
std::string tuning_key =
Concat("pad", output->dim(0), output->dim(1), output->dim(2),
output->dim(3));
std::string tuning_key = Concat("pad", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -118,10 +116,8 @@ MaceStatus PadFunctor<DeviceType::GPU, T>::operator()(
return MACE_SUCCESS;
}
template
struct PadFunctor<DeviceType::GPU, float>;
template
struct PadFunctor<DeviceType::GPU, half>;
template struct PadFunctor<DeviceType::GPU, float>;
template struct PadFunctor<DeviceType::GPU, half>;
} // namespace kernels
} // namespace mace
mace/kernels/opencl/pooling.cc
浏览文件 @ 2940ce5d
......@@ -23,15 +23,13 @@ namespace kernels {
namespace {
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
lws[2] = std::min<uint32_t>(std::min<uint32_t>(gws[2], base),
kwg_size / lws[1]);
lws[2] =
std::min<uint32_t>(std::min<uint32_t>(gws[2], base), kwg_size / lws[1]);
const uint32_t lws_size = lws[1] * lws[2];
lws[0] = gws[0] / 4;
if (lws[0] == 0) {
......@@ -73,7 +71,7 @@ MaceStatus PoolingFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -108,7 +106,7 @@ MaceStatus PoolingFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, output_image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, output_image_shape));
index_t batch = output->dim(0);
index_t out_height = output->dim(1);
......@@ -159,8 +157,8 @@ MaceStatus PoolingFunctor<DeviceType::GPU, T>::operator()(const Tensor *input,
const std::vector<uint32_t> lws = LocalWS(gws.data(), kwg_size_);
std::string tuning_key =
Concat("pooling_opencl_kernel_", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("pooling_opencl_kernel_", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(kernel_, tuning_key, gws.data(), lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......
mace/kernels/opencl/resize_bilinear.cc
浏览文件 @ 2940ce5d
......@@ -23,11 +23,9 @@ namespace mace {
namespace kernels {
namespace {
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
std::vector<uint32_t> lws(4, 0);
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
if (lws[1] >= base) {
......@@ -79,7 +77,7 @@ MaceStatus ResizeBilinearFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -100,7 +98,7 @@ MaceStatus ResizeBilinearFunctor<DeviceType::GPU, T>::operator()(
std::vector<size_t> output_image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
MACE_FAILURE_RETURN(output->ResizeImage(output_shape, output_image_shape));
MACE_RETURN_IF_ERROR(output->ResizeImage(output_shape, output_image_shape));
float height_scale =
CalculateResizeScale(in_height, out_height, align_corners_);
......@@ -130,8 +128,8 @@ MaceStatus ResizeBilinearFunctor<DeviceType::GPU, T>::operator()(
const std::vector<uint32_t> lws = LocalWS(gws, kwg_size_);
std::string tuning_key =
Concat("resize_bilinear_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("resize_bilinear_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......
mace/kernels/opencl/slice.cc
浏览文件 @ 2940ce5d
......@@ -20,7 +20,7 @@
namespace mace {
namespace kernels {
template<typename T>
template <typename T>
MaceStatus SliceFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input,
const std::vector<Tensor *> &output_list,
......@@ -30,13 +30,14 @@ MaceStatus SliceFunctor<DeviceType::GPU, T>::operator()(
const index_t output_channels = input_channels / outputs_count;
MACE_CHECK(output_channels % 4 == 0)
<< "output channels of slice op must be divisible by 4";
std::vector<index_t> output_shape({input->dim(0), input->dim(1),
input->dim(2), output_channels});
std::vector<index_t> output_shape(
{input->dim(0), input->dim(1), input->dim(2), output_channels});
std::vector<size_t> image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL, &image_shape);
for (size_t i= 0; i < outputs_count; ++i) {
MACE_FAILURE_RETURN(output_list[i]->ResizeImage(output_shape, image_shape));
for (size_t i = 0; i < outputs_count; ++i) {
MACE_RETURN_IF_ERROR(
output_list[i]->ResizeImage(output_shape, image_shape));
}
auto runtime = OpenCLRuntime::Global();
......@@ -46,13 +47,13 @@ MaceStatus SliceFunctor<DeviceType::GPU, T>::operator()(
std::string kernel_name = MACE_OBFUSCATE_SYMBOL("slice");
built_options.emplace("-Dslice=" + kernel_name);
built_options.emplace("-DDATA_TYPE=" + DtToCLDt(DataTypeToEnum<T>::value));
built_options.emplace("-DCMD_DATA_TYPE="
+ DtToCLCMDDt(DataTypeToEnum<T>::value));
built_options.emplace("-DCMD_DATA_TYPE=" +
DtToCLCMDDt(DataTypeToEnum<T>::value));
if (runtime->IsOutOfRangeCheckEnabled()) {
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -68,8 +69,7 @@ MaceStatus SliceFunctor<DeviceType::GPU, T>::operator()(
const index_t channel_blk = RoundUpDiv4(output_channels);
const uint32_t gws[3] = {
static_cast<uint32_t>(channel_blk),
static_cast<uint32_t>(input->dim(2)),
static_cast<uint32_t>(channel_blk), static_cast<uint32_t>(input->dim(2)),
static_cast<uint32_t>(input->dim(0) * input->dim(1)),
};
......@@ -117,8 +117,8 @@ MaceStatus SliceFunctor<DeviceType::GPU, T>::operator()(
if (runtime->is_profiling_enabled()) {
CallStats tmp_stats;
runtime->GetCallStats(event, &tmp_stats);
call_stats.start_micros = std::min<int64_t>(tmp_stats.start_micros,
call_stats.start_micros);
call_stats.start_micros =
std::min<int64_t>(tmp_stats.start_micros, call_stats.start_micros);
call_stats.end_micros += tmp_stats.end_micros - tmp_stats.start_micros;
}
}
......@@ -135,10 +135,8 @@ MaceStatus SliceFunctor<DeviceType::GPU, T>::operator()(
return MACE_SUCCESS;
}
template
struct SliceFunctor<DeviceType::GPU, float>;
template
struct SliceFunctor<DeviceType::GPU, half>;
template struct SliceFunctor<DeviceType::GPU, float>;
template struct SliceFunctor<DeviceType::GPU, half>;
} // namespace kernels
} // namespace mace
mace/kernels/opencl/softmax.cc
浏览文件 @ 2940ce5d
......@@ -24,10 +24,8 @@ namespace kernels {
namespace {
std::vector<uint32_t> LocalWS(const uint32_t *gws,
const uint32_t kwg_size) {
uint64_t cache_size =
OpenCLRuntime::Global()->device_global_mem_cache_size();
std::vector<uint32_t> LocalWS(const uint32_t *gws, const uint32_t kwg_size) {
uint64_t cache_size = OpenCLRuntime::Global()->device_global_mem_cache_size();
uint32_t base = cache_size / kBaseGPUMemCacheSize;
std::vector<uint32_t> lws(4, 0);
lws[1] = std::min<uint32_t>(gws[1], kwg_size);
......@@ -71,7 +69,7 @@ MaceStatus SoftmaxFunctor<DeviceType::GPU, T>::operator()(const Tensor *logits,
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -105,8 +103,8 @@ MaceStatus SoftmaxFunctor<DeviceType::GPU, T>::operator()(const Tensor *logits,
std::vector<uint32_t> lws = LocalWS(gws, kwg_size_);
std::string tuning_key =
Concat("softmax_opencl_kernel", output->dim(0),
output->dim(1), output->dim(2), output->dim(3));
Concat("softmax_opencl_kernel", output->dim(0), output->dim(1),
output->dim(2), output->dim(3));
TuningOrRun3DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......
mace/kernels/opencl/space_to_batch.cc
浏览文件 @ 2940ce5d
......@@ -26,17 +26,13 @@ namespace kernels {
template <typename T>
MaceStatus SpaceToBatchFunctor<DeviceType::GPU, T>::operator()(
Tensor *space_tensor,
Tensor *batch_tensor,
StatsFuture *future) {
Tensor *space_tensor, Tensor *batch_tensor, StatsFuture *future) {
std::vector<index_t> output_shape(4, 0);
if (b2s_) {
CalculateBatchToSpaceOutputShape(batch_tensor,
DataFormat::NHWC,
CalculateBatchToSpaceOutputShape(batch_tensor, DataFormat::NHWC,
output_shape.data());
} else {
CalculateSpaceToBatchOutputShape(space_tensor,
DataFormat::NHWC,
CalculateSpaceToBatchOutputShape(space_tensor, DataFormat::NHWC,
output_shape.data());
}
......@@ -45,12 +41,12 @@ MaceStatus SpaceToBatchFunctor<DeviceType::GPU, T>::operator()(
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL,
&output_image_shape);
if (b2s_) {
MACE_FAILURE_RETURN(space_tensor->ResizeImage(output_shape,
output_image_shape));
MACE_RETURN_IF_ERROR(
space_tensor->ResizeImage(output_shape, output_image_shape));
kernel_name = "batch_to_space";
} else {
MACE_FAILURE_RETURN(batch_tensor->ResizeImage(output_shape,
output_image_shape));
MACE_RETURN_IF_ERROR(
batch_tensor->ResizeImage(output_shape, output_image_shape));
kernel_name = "space_to_batch";
}
const uint32_t chan_blk = RoundUpDiv4<uint32_t>(batch_tensor->dim(3));
......@@ -73,7 +69,7 @@ MaceStatus SpaceToBatchFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -81,9 +77,8 @@ MaceStatus SpaceToBatchFunctor<DeviceType::GPU, T>::operator()(
if (runtime->IsNonUniformWorkgroupsSupported()) {
built_options.emplace("-DNON_UNIFORM_WORK_GROUP");
}
kernel_ =
runtime->BuildKernel("space_to_batch",
obfuscated_kernel_name, built_options);
kernel_ = runtime->BuildKernel("space_to_batch", obfuscated_kernel_name,
built_options);
kwg_size_ =
static_cast<uint32_t>(runtime->GetKernelMaxWorkGroupSize(kernel_));
......
mace/kernels/opencl/winograd_transform.cc
浏览文件 @ 2940ce5d
......@@ -24,7 +24,6 @@ namespace kernels {
template <typename T>
MaceStatus WinogradTransformFunctor<DeviceType::GPU, T>::operator()(
const Tensor *input_tensor, Tensor *output_tensor, StatsFuture *future) {
auto runtime = OpenCLRuntime::Global();
if (kernel_.get() == nullptr) {
......@@ -40,7 +39,7 @@ MaceStatus WinogradTransformFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -78,7 +77,7 @@ MaceStatus WinogradTransformFunctor<DeviceType::GPU, T>::operator()(
output_shape = {16, input_tensor->dim(3), out_width, 1};
std::vector<size_t> image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_HEIGHT, &image_shape);
MACE_FAILURE_RETURN(output_tensor->ResizeImage(output_shape, image_shape));
MACE_RETURN_IF_ERROR(output_tensor->ResizeImage(output_shape, image_shape));
uint32_t idx = 0;
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -103,10 +102,9 @@ MaceStatus WinogradTransformFunctor<DeviceType::GPU, T>::operator()(
}
const std::vector<uint32_t> lws = {kwg_size_ / 8, 8, 0};
std::string tuning_key =
Concat("winograd_transform_kernel", output_tensor->dim(0),
output_tensor->dim(1), output_tensor->dim(2),
output_tensor->dim(3));
std::string tuning_key = Concat("winograd_transform_kernel",
output_tensor->dim(0), output_tensor->dim(1),
output_tensor->dim(2), output_tensor->dim(3));
TuningOrRun2DKernel(kernel_, tuning_key, gws, lws, future);
if (runtime->IsOutOfRangeCheckEnabled()) {
......@@ -125,7 +123,6 @@ MaceStatus WinogradInverseTransformFunctor<DeviceType::GPU, T>::operator()(
const Tensor *bias,
Tensor *output_tensor,
StatsFuture *future) {
auto runtime = OpenCLRuntime::Global();
if (kernel_.get() == nullptr) {
......@@ -142,7 +139,7 @@ MaceStatus WinogradInverseTransformFunctor<DeviceType::GPU, T>::operator()(
built_options.emplace("-DOUT_OF_RANGE_CHECK");
kernel_error_ = std::move(std::unique_ptr<Buffer>(
new Buffer(GetDeviceAllocator(DeviceType::GPU))));
kernel_error_->Allocate(1);
MACE_RETURN_IF_ERROR(kernel_error_->Allocate(1));
kernel_error_->Map(nullptr);
*(kernel_error_->mutable_data<char>()) = 0;
kernel_error_->UnMap();
......@@ -188,7 +185,7 @@ MaceStatus WinogradInverseTransformFunctor<DeviceType::GPU, T>::operator()(
input_tensor->dim(1)};
std::vector<size_t> image_shape;
CalImage2DShape(output_shape, BufferType::IN_OUT_CHANNEL, &image_shape);
MACE_FAILURE_RETURN(output_tensor->ResizeImage(output_shape, image_shape));
MACE_RETURN_IF_ERROR(output_tensor->ResizeImage(output_shape, image_shape));
const uint32_t round_h = (height_ + 1) / 2;
const uint32_t round_w = (width_ + 1) / 2;
......
mace/kernels/pad.h
浏览文件 @ 2940ce5d
......@@ -51,7 +51,7 @@ struct PadFunctor : public PadFunctorBase {
MACE_CHECK(
this->paddings_.size() == static_cast<size_t>(input->dim_size()) * 2);
auto input_shape = input->shape();
MACE_FAILURE_RETURN(output->Resize({input_shape[0] + this->paddings_[0]
MACE_RETURN_IF_ERROR(output->Resize({input_shape[0] + this->paddings_[0]
+ this->paddings_[1],
input_shape[1] + this->paddings_[2]
+ this->paddings_[3],
......
mace/kernels/pooling.h
浏览文件 @ 2940ce5d
......@@ -190,7 +190,7 @@ struct PoolingFunctor<DeviceType::CPU, float>: PoolingFunctorBase {
RoundType::CEIL,
output_shape.data());
}
MACE_FAILURE_RETURN(output_tensor->Resize(output_shape));
MACE_RETURN_IF_ERROR(output_tensor->Resize(output_shape));
Tensor::MappingGuard input_guard(input_tensor);
Tensor::MappingGuard output_guard(output_tensor);
......
mace/kernels/proposal.h
浏览文件 @ 2940ce5d
......@@ -267,7 +267,7 @@ struct ProposalFunctor {
// Our RPN implementation only supports a single input image, so all
// batch inds are 0
size = static_cast<int>(nms_result.size());
MACE_FAILURE_RETURN(output->Resize({size, 1, 1, 5}));
MACE_RETURN_IF_ERROR(output->Resize({size, 1, 1, 5}));
auto output_ptr = output->mutable_data<float>();
#pragma omp parallel for
for (int i = 0; i < size; ++i) {
......
mace/kernels/psroi_align.h
浏览文件 @ 2940ce5d
......@@ -50,7 +50,7 @@ struct PSROIAlignFunctor {
const index_t num_rois = rois->dim(0);
const index_t batch_size = input->dim(0);
MACE_FAILURE_RETURN(output->Resize({num_rois, pooled_height, pooled_width,
MACE_RETURN_IF_ERROR(output->Resize({num_rois, pooled_height, pooled_width,
output_dim_}));
T *output_ptr = output->mutable_data<T>();
......
mace/kernels/resize_bilinear.h
浏览文件 @ 2940ce5d
......@@ -150,7 +150,7 @@ struct ResizeBilinearFunctor<DeviceType::CPU, float>
index_t out_width = out_width_;
MACE_CHECK(out_height > 0 && out_width > 0);
std::vector<index_t> out_shape{batch, channels, out_height, out_width};
MACE_FAILURE_RETURN(output->Resize(out_shape));
MACE_RETURN_IF_ERROR(output->Resize(out_shape));
Tensor::MappingGuard input_mapper(input);
Tensor::MappingGuard output_mapper(output);
......
mace/kernels/slice.h
浏览文件 @ 2940ce5d
......@@ -61,7 +61,7 @@ struct SliceFunctor : SliceFunctorBase {
1,
std::multiplies<index_t>());
for (size_t i= 0; i < outputs_count; ++i) {
MACE_FAILURE_RETURN(output_list[i]->Resize(output_shape));
MACE_RETURN_IF_ERROR(output_list[i]->Resize(output_shape));
output_ptrs[i] = output_list[i]->mutable_data<T>();
}
const T *input_ptr = input->data<T>();
......
mace/kernels/space_to_batch.h
浏览文件 @ 2940ce5d
......@@ -150,12 +150,12 @@ struct SpaceToBatchFunctor<DeviceType::CPU, float> : SpaceToBatchFunctorBase {
CalculateBatchToSpaceOutputShape(batch_tensor,
DataFormat::NCHW,
output_shape.data());
MACE_FAILURE_RETURN(space_tensor->Resize(output_shape));
MACE_RETURN_IF_ERROR(space_tensor->Resize(output_shape));
} else {
CalculateSpaceToBatchOutputShape(space_tensor,
DataFormat::NCHW,
output_shape.data());
MACE_FAILURE_RETURN(batch_tensor->Resize(output_shape));
MACE_RETURN_IF_ERROR(batch_tensor->Resize(output_shape));
}
Tensor::MappingGuard input_guard(space_tensor);
......
mace/ops/BUILD
浏览文件 @ 2940ce5d
......@@ -15,7 +15,6 @@ cc_library(
hdrs = [
"ops_test_util.h",
],
copts = ["-Werror", "-Wextra", "-Wno-missing-field-initializers"],
deps = [
"//mace/core",
"@gtest//:gtest",
......@@ -36,18 +35,23 @@ cc_library(
[
"buffer_to_image.cc",
"image_to_buffer.cc",
]),
],
),
hdrs = glob(
["*.h"],
exclude = ["ops_test_util.h"],
),
copts = ["-Werror", "-Wextra", "-Wno-missing-field-initializers"] +
if_openmp_enabled(["-fopenmp"]) +
if_neon_enabled(["-DMACE_ENABLE_NEON"]) +
if_android_armv7(["-mfpu=neon"]) +
if_android_armv7(["-mfloat-abi=softfp"]) +
if_android(["-DMACE_ENABLE_OPENCL"]) +
if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled([
"-DMACE_ENABLE_NEON",
]) + if_android_armv7([
"-mfpu=neon",
]) + if_android_armv7([
"-mfloat-abi=softfp",
]) + if_android([
"-DMACE_ENABLE_OPENCL",
]) + if_hexagon_enabled([
"-DMACE_ENABLE_HEXAGON",
]),
deps = [
"//mace/kernels",
],
......@@ -60,13 +64,17 @@ cc_test(
srcs = glob(
["*_test.cc"],
),
copts = ["-Werror", "-Wextra", "-Wno-missing-field-initializers"] +
if_openmp_enabled(["-fopenmp"]) +
if_neon_enabled(["-DMACE_ENABLE_NEON"]) +
if_android_armv7(["-mfpu=neon"]) +
if_android_armv7(["-mfloat-abi=softfp"]) +
if_android(["-DMACE_ENABLE_OPENCL"]) +
if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled([
"-DMACE_ENABLE_NEON",
]) + if_android_armv7([
"-mfpu=neon",
]) + if_android_armv7([
"-mfloat-abi=softfp",
]) + if_android([
"-DMACE_ENABLE_OPENCL",
]) + if_hexagon_enabled([
"-DMACE_ENABLE_HEXAGON",
]),
linkopts = ["-fopenmp"],
linkstatic = 1,
deps = [
......@@ -80,13 +88,17 @@ cc_test(
name = "ops_benchmark",
testonly = 1,
srcs = glob(["*_benchmark.cc"]),
copts = ["-Werror", "-Wextra", "-Wno-missing-field-initializers"] +
if_openmp_enabled(["-fopenmp"]) +
if_neon_enabled(["-DMACE_ENABLE_NEON"]) +
if_android_armv7(["-mfpu=neon"]) +
if_android_armv7(["-mfloat-abi=softfp"]) +
if_android(["-DMACE_ENABLE_OPENCL"]) +
if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled([
"-DMACE_ENABLE_NEON",
]) + if_android_armv7([
"-mfpu=neon",
]) + if_android_armv7([
"-mfloat-abi=softfp",
]) + if_android([
"-DMACE_ENABLE_OPENCL",
]) + if_hexagon_enabled([
"-DMACE_ENABLE_HEXAGON",
]),
linkopts = ["-fopenmp"],
linkstatic = 1,
deps = [
......
mace/ops/activation.h
浏览文件 @ 2940ce5d
......@@ -31,15 +31,15 @@ class ActivationOp : public Operator<D, T> {
functor_(kernels::StringToActivationType(
OperatorBase::GetOptionalArg<std::string>("activation",
"NOOP")),
static_cast<T>(OperatorBase::GetOptionalArg<float>(
"max_limit", 0.0f))) {}
static_cast<T>(
OperatorBase::GetOptionalArg<float>("max_limit", 0.0f))) {}
MaceStatus Run(StatsFuture *future) override {
const Tensor *input_tensor = this->Input(0);
const Tensor *alpha_tensor =
this->InputSize() >= 2 ? this->Input(1) : nullptr;
Tensor *output_tensor = this->Output(0);
MACE_FAILURE_RETURN(output_tensor->ResizeLike(input_tensor));
MACE_RETURN_IF_ERROR(output_tensor->ResizeLike(input_tensor));
return functor_(input_tensor, alpha_tensor, output_tensor, future);
}
......
mace/ops/activation_test.cc
浏览文件 @ 2940ce5d
......@@ -120,7 +120,6 @@ TEST_F(ActivationOpTest, OPENCLUnalignedSimpleRelu) {
TestUnalignedSimpleRelu<DeviceType::GPU>();
}
namespace {
template <DeviceType D>
void TestSimpleRelux() {
......@@ -169,9 +168,7 @@ void TestSimpleRelux() {
TEST_F(ActivationOpTest, CPUSimple) { TestSimpleRelux<DeviceType::CPU>(); }
TEST_F(ActivationOpTest, OPENCLSimple) {
TestSimpleRelux<DeviceType::GPU>();
}
TEST_F(ActivationOpTest, OPENCLSimple) { TestSimpleRelux<DeviceType::GPU>(); }
namespace {
template <DeviceType D>
......@@ -278,9 +275,7 @@ void TestSimplePrelu() {
}
} // namespace
TEST_F(ActivationOpTest, CPUSimplePrelu) {
TestSimplePrelu<DeviceType::CPU>();
}
TEST_F(ActivationOpTest, CPUSimplePrelu) { TestSimplePrelu<DeviceType::CPU>(); }
TEST_F(ActivationOpTest, OPENCLSimplePrelu) {
TestSimplePrelu<DeviceType::GPU>();
......
mace/ops/addn_test.cc
浏览文件 @ 2940ce5d
......@@ -97,8 +97,8 @@ void SimpleAdd3() {
net.RunOp(D);
}
auto expected = CreateTensor<float>({1, 2, 3, 1},
{-0.000713, 8, 12, 16, 20, 24});
auto expected =
CreateTensor<float>({1, 2, 3, 1}, {-0.000713, 8, 12, 16, 20, 24});
ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-4, 1e-3);
}
......@@ -160,8 +160,8 @@ void RandomTest() {
ImageToBuffer<D, float>(&net, "OutputImage", "OPENCLOutput",
kernels::BufferType::IN_OUT_CHANNEL);
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
1e-2, 1e-2);
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2,
1e-2);
}
}
} // namespace
......
mace/ops/batch_norm.h
浏览文件 @ 2940ce5d
......@@ -51,7 +51,7 @@ class BatchNormOp : public Operator<D, T> {
var->dim_size());
Tensor *output = this->Output(OUTPUT);
MACE_FAILURE_RETURN(output->ResizeLike(input));
MACE_RETURN_IF_ERROR(output->ResizeLike(input));
return functor_(input, scale, offset, mean, var, epsilon_, output, future);
}
......
mace/ops/batch_norm_test.cc
浏览文件 @ 2940ce5d
......@@ -22,7 +22,7 @@ namespace test {
class BatchNormOpTest : public OpsTestBase {};
namespace {
template<DeviceType D>
template <DeviceType D>
void Simple() {
OpsTestNet net;
......@@ -79,10 +79,9 @@ void Simple() {
}
// Check
auto expected =
CreateTensor<float>({1, 6, 2, 1}, {-3.8543, -3.8543, -1.5125, -1.5125,
0.8291, 0.8291, 3.1708, 3.1708,
5.5125, 5.5125, 7.8543, 7.8543});
auto expected = CreateTensor<float>(
{1, 6, 2, 1}, {-3.8543, -3.8543, -1.5125, -1.5125, 0.8291, 0.8291, 3.1708,
3.1708, 5.5125, 5.5125, 7.8543, 7.8543});
ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-4);
}
......@@ -103,16 +102,14 @@ TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Mean", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Var", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
// Construct graph
......@@ -129,9 +126,7 @@ TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -186,16 +181,14 @@ TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Mean", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Var", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("BatchNorm", "BatchNormTest")
......@@ -211,9 +204,7 @@ TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -269,16 +260,14 @@ TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Mean", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Var", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("BatchNorm", "BatchNormTest")
......@@ -294,9 +283,7 @@ TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -351,16 +338,14 @@ TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Mean", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Var", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("BatchNorm", "BatchNormTest")
......@@ -376,9 +361,7 @@ TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......
mace/ops/batch_to_space.h
浏览文件 @ 2940ce5d
......@@ -36,8 +36,7 @@ class BatchToSpaceNDOp : public Operator<D, T> {
MaceStatus Run(StatsFuture *future) override {
const Tensor *batch_tensor = this->Input(INPUT);
Tensor *space_tensor = this->Output(OUTPUT);
return functor_(space_tensor, const_cast<Tensor *>(batch_tensor),
future);
return functor_(space_tensor, const_cast<Tensor *>(batch_tensor), future);
}
private:
......
mace/ops/bias_add.h
浏览文件 @ 2940ce5d
......@@ -37,7 +37,7 @@ class BiasAddOp : public Operator<D, T> {
bias->dim_size());
Tensor *output = this->Output(OUTPUT);
MACE_FAILURE_RETURN(output->ResizeLike(input));
MACE_RETURN_IF_ERROR(output->ResizeLike(input));
return functor_(input, bias, output, future);
}
......
mace/ops/bias_add_test.cc
浏览文件 @ 2940ce5d
......@@ -32,9 +32,7 @@ void BiasAddSimple() {
net.AddInputFromArray<D, float>("Bias", {1}, {0.5f});
if (D == DeviceType::CPU) {
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("BiasAdd", "BiasAddTest")
.Input("InputNCHW")
......@@ -43,10 +41,8 @@ void BiasAddSimple() {
.Finalize(net.NewOperatorDef());
// Run
net.RunOp(D);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
NHWC);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW,
"Output", NHWC);
} else if (D == DeviceType::GPU) {
BufferToImage<D, float>(&net, "Input", "InputImage",
kernels::BufferType::IN_OUT_CHANNEL);
......@@ -79,9 +75,7 @@ void BiasAddSimple() {
TEST_F(BiasAddOpTest, BiasAddSimpleCPU) { BiasAddSimple<DeviceType::CPU>(); }
TEST_F(BiasAddOpTest, BiasAddSimpleOPENCL) {
BiasAddSimple<DeviceType::GPU>();
}
TEST_F(BiasAddOpTest, BiasAddSimpleOPENCL) { BiasAddSimple<DeviceType::GPU>(); }
TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
// generate random input
......@@ -94,13 +88,11 @@ TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Bias", {channels}, true);
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
// Construct graph
......@@ -113,9 +105,7 @@ TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -154,13 +144,11 @@ TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Bias", {channels}, true);
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
// Construct graph
......@@ -173,9 +161,7 @@ TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
Tensor expected;
......
mace/ops/buffer_to_image_test.cc
浏览文件 @ 2940ce5d
......@@ -233,8 +233,8 @@ TEST(BufferToImageTest, ArgStringHalfToHalfSmall) {
const unsigned char input_data[] = {
0xCD, 0x3C, 0x33, 0x40,
};
TestStringHalfBidirectionTransform<DeviceType::GPU, half>(
kernels::ARGUMENT, {2}, input_data);
TestStringHalfBidirectionTransform<DeviceType::GPU, half>(kernels::ARGUMENT,
{2}, input_data);
}
} // namespace test
......
mace/ops/channel_shuffle_test.cc
浏览文件 @ 2940ce5d
......@@ -29,9 +29,7 @@ TEST_F(ChannelShuffleOpTest, C8G4_CPU) {
"Input", {1, 1, 2, 8},
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
// Construct graph
......@@ -43,9 +41,7 @@ TEST_F(ChannelShuffleOpTest, C8G4_CPU) {
// Run
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......
mace/ops/concat_test.cc
浏览文件 @ 2940ce5d
......@@ -12,12 +12,12 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <string>
#include <functional>
#include <string>
#include "gmock/gmock.h"
#include "mace/ops/ops_test_util.h"
#include "mace/ops/concat.h"
#include "mace/ops/ops_test_util.h"
namespace mace {
namespace ops {
......@@ -163,7 +163,7 @@ void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
int concat_axis_size = 0;
// Construct graph
std::vector<std::vector<float>> inputs(num_inputs, std::vector<float>());
std::vector<const float*> input_ptrs(num_inputs);
std::vector<const float *> input_ptrs(num_inputs);
OpsTestNet net;
for (int i = 0; i < num_inputs; ++i) {
const std::string input_name = MakeString("Input", i);
......@@ -171,8 +171,8 @@ void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
concat_axis_size += shapes[i][axis];
GenerateRandomRealTypeData(shapes[i], &inputs[i]);
input_ptrs[i] = inputs[i].data();
net.AddInputFromArray<DeviceType::GPU, float>(input_name,
shapes[i], inputs[i]);
net.AddInputFromArray<DeviceType::GPU, float>(input_name, shapes[i],
inputs[i]);
BufferToImage<DeviceType::GPU, T>(&net, input_name, image_name,
kernels::BufferType::IN_OUT_CHANNEL);
}
......
mace/ops/conv_2d_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/conv_pool_2d_base.h
浏览文件 @ 2940ce5d
......@@ -32,8 +32,7 @@ class ConvPool2dOpBase : public Operator<D, T> {
padding_type_(static_cast<Padding>(OperatorBase::GetOptionalArg<int>(
"padding", static_cast<int>(SAME)))),
paddings_(OperatorBase::GetRepeatedArgs<int>("padding_values")),
dilations_(
OperatorBase::GetRepeatedArgs<int>("dilations", {1, 1})) {}
dilations_(OperatorBase::GetRepeatedArgs<int>("dilations", {1, 1})) {}
protected:
std::vector<int> strides_;
......
mace/ops/core_test.cc
浏览文件 @ 2940ce5d
......@@ -31,8 +31,7 @@ TEST(CoreTest, INIT_MODE) {
.AddIntArg("mode", static_cast<int>(NetMode::INIT))
.Finalize(&op_defs[op_defs.size() - 1]);
Tensor *input =
ws.CreateTensor("Input", GetDeviceAllocator(DeviceType::GPU),
Tensor *input = ws.CreateTensor("Input", GetDeviceAllocator(DeviceType::GPU),
DataTypeToEnum<float>::v());
input->Resize({1, 3, 3, 3});
{
......
mace/ops/deconv_2d_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/depth_to_space_test.cc
浏览文件 @ 2940ce5d
......@@ -36,9 +36,7 @@ void RunDepthToSpace(const bool d2s,
const char *ops_test_name = (d2s) ? "DepthToSpaceTest" : "SpaceToDepthTest";
// Construct graph
if (D == DeviceType::CPU) {
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder(ops_name, ops_test_name)
.Input("InputNCHW")
......@@ -47,10 +45,8 @@ void RunDepthToSpace(const bool d2s,
.Finalize(net.NewOperatorDef());
// Run
net.RunOp(D);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
NHWC);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW,
"Output", NHWC);
} else {
BufferToImage<D, float>(&net, "Input", "InputImage",
......@@ -64,7 +60,6 @@ void RunDepthToSpace(const bool d2s,
net.RunOp(D);
}
if (D == DeviceType::GPU) {
ImageToBuffer<DeviceType::GPU, float>(&net, "OutputImage", "Output",
kernels::BufferType::IN_OUT_CHANNEL);
......@@ -74,108 +69,89 @@ void RunDepthToSpace(const bool d2s,
}
} // namespace
class SpaceToDepthOpTest : public OpsTestBase {};
TEST_F(SpaceToDepthOpTest, Input2x4x4_B2_CPU) {
RunDepthToSpace<DeviceType::CPU>(false, {1, 2, 4, 4},
RunDepthToSpace<DeviceType::CPU>(
false, {1, 2, 4, 4},
{0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23,
8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31},
2,
{1, 1, 2, 16},
2, {1, 1, 2, 16},
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31});
}
TEST_F(SpaceToDepthOpTest, Input2x4x4_B2_OPENCL) {
RunDepthToSpace<DeviceType::GPU>(false, {1, 2, 4, 4},
RunDepthToSpace<DeviceType::GPU>(
false, {1, 2, 4, 4},
{0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23,
8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31},
2,
{1, 1, 2, 16},
2, {1, 1, 2, 16},
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31});
}
TEST_F(SpaceToDepthOpTest, Input2x2x4_B2_CPU) {
RunDepthToSpace<DeviceType::CPU>(false, {1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16},
2,
{1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16});
RunDepthToSpace<DeviceType::CPU>(
false, {1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}, 2, {1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16});
}
TEST_F(SpaceToDepthOpTest, Input4x4x1_B2_OPENCL) {
RunDepthToSpace<DeviceType::GPU>(false, {1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16},
2,
{1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16});
RunDepthToSpace<DeviceType::GPU>(
false, {1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}, 2, {1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16});
}
class DepthToSpaceOpTest : public OpsTestBase {};
TEST_F(DepthToSpaceOpTest, Input1x2x16_B2_CPU) {
RunDepthToSpace<DeviceType::CPU>(true, {1, 1, 2, 16},
RunDepthToSpace<DeviceType::CPU>(
true, {1, 1, 2, 16},
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31},
2,
{1, 2, 4, 4},
2, {1, 2, 4, 4},
{0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23,
8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31});
}
TEST_F(DepthToSpaceOpTest, Input1x2x16_B2_OPENCL) {
RunDepthToSpace<DeviceType::GPU>(true, {1, 1, 2, 16},
RunDepthToSpace<DeviceType::GPU>(
true, {1, 1, 2, 16},
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31},
2,
{1, 2, 4, 4},
2, {1, 2, 4, 4},
{0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23,
8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31});
}
TEST_F(DepthToSpaceOpTest, Input1x1x16_B2_CPU) {
RunDepthToSpace<DeviceType::CPU>(true, {1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16},
2,
{1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16});
RunDepthToSpace<DeviceType::CPU>(
true, {1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}, 2, {1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16});
}
TEST_F(DepthToSpaceOpTest, Input1x1x16_B2_OPENCL) {
RunDepthToSpace<DeviceType::GPU>(true, {1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16},
2,
{1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16});
RunDepthToSpace<DeviceType::GPU>(
true, {1, 1, 1, 16},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}, 2, {1, 2, 2, 4},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16});
}
TEST_F(DepthToSpaceOpTest, InputLarger_B2_OPENCL) {
const std::vector<float > in = std::vector<float >(192 * 192 *128, 1.0);
const std::vector<float> in = std::vector<float>(192 * 192 * 128, 1.0);
RunDepthToSpace<DeviceType::GPU>(true, {1, 192, 192, 128},
in,
2,
{1, 384, 384, 32},
in);
RunDepthToSpace<DeviceType::GPU>(true, {1, 192, 192, 128}, in, 2,
{1, 384, 384, 32}, in);
}
namespace {
template <DeviceType D, typename T>
void RandomTest(const bool d2s, const int block_size,
void RandomTest(const bool d2s,
const int block_size,
const std::vector<index_t> &shape) {
testing::internal::LogToStderr();
srand(time(NULL));
......@@ -188,9 +164,7 @@ void RandomTest(const bool d2s, const int block_size,
// Add input data
net.AddRandomInput<D, float>("Input", shape);
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder(ops_name, ops_test_name)
.Input("InputNCHW")
......@@ -201,12 +175,9 @@ void RandomTest(const bool d2s, const int block_size,
// Run
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
BufferToImage<D, T>(&net, "Input", "InputImg",
kernels::BufferType::IN_OUT_CHANNEL);
......@@ -238,15 +209,15 @@ TEST_F(DepthToSpaceOpTest, OPENCLRandomFloat) {
}
TEST_F(DepthToSpaceOpTest, OPENCLRandomHalf) {
RandomTest<DeviceType::GPU, half>(true, 2, {1, 192, 192, 128});
RandomTest<DeviceType::GPU, half>(true, 2, {1, 192, 192, 128});
}
TEST_F(SpaceToDepthOpTest, OPENCLRandomFloat) {
RandomTest<DeviceType::GPU, float>(false, 2, {1, 384, 384, 32});
RandomTest<DeviceType::GPU, float>(false, 2, {1, 384, 384, 32});
}
TEST_F(SpaceToDepthOpTest, OPENCLRandomHalf) {
RandomTest<DeviceType::GPU, half>(false, 2, {1, 384, 384, 32});
RandomTest<DeviceType::GPU, half>(false, 2, {1, 384, 384, 32});
}
} // namespace test
......
mace/ops/depthwise_conv2d_test.cc
浏览文件 @ 2940ce5d
......@@ -22,7 +22,7 @@ namespace test {
class DepthwiseConv2dOpTest : public OpsTestBase {};
namespace {
template<DeviceType D, typename T>
template <DeviceType D, typename T>
void SimpleValidTest() {
testing::internal::LogToStderr();
// Construct graph
......@@ -36,9 +36,7 @@ void SimpleValidTest() {
"Filter", {1, 2, 2, 2}, {1.0f, 2.0f, 3.0f, 4.0f, 2.0f, 4.0f, 6.0f, 8.0f});
net.AddInputFromArray<D, float>("Bias", {2}, {.1f, .2f});
if (D == DeviceType::CPU) {
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("DepthwiseConv2d", "DepthwiseConv2DTest")
.Input("InputNCHW")
......@@ -51,10 +49,8 @@ void SimpleValidTest() {
.Finalize(net.NewOperatorDef());
// Run
net.RunOp(D);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
NHWC);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW,
"Output", NHWC);
} else if (D == DeviceType::GPU) {
BufferToImage<D, T>(&net, "Input", "InputImage",
kernels::BufferType::IN_OUT_CHANNEL);
......@@ -85,8 +81,8 @@ void SimpleValidTest() {
// Check
auto expected = CreateTensor<float>(
{1, 2, 2, 2}, {37.1f, 148.2f, 47.1f, 188.2f,
67.1f, 268.2f, 77.1f, 308.2f});
{1, 2, 2, 2},
{37.1f, 148.2f, 47.1f, 188.2f, 67.1f, 268.2f, 77.1f, 308.2f});
if (DataTypeToEnum<T>::value == DT_HALF) {
ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-3, 1e-3);
......@@ -109,9 +105,13 @@ TEST_F(DepthwiseConv2dOpTest, SimpleOpenCLHalf) {
}
namespace {
template<DeviceType D, typename T>
void ComplexValidTest(index_t batch, index_t channel, index_t height,
index_t width, index_t kernel, index_t multiplier,
template <DeviceType D, typename T>
void ComplexValidTest(index_t batch,
index_t channel,
index_t height,
index_t width,
index_t kernel,
index_t multiplier,
int stride) {
testing::internal::LogToStderr();
// Construct graph
......@@ -125,18 +125,14 @@ void ComplexValidTest(index_t batch, index_t channel, index_t height,
std::vector<float> filter_data(kernel * kernel * channel * multiplier);
GenerateRandomRealTypeData({multiplier, channel, kernel, kernel},
&filter_data);
net.AddInputFromArray<D, float>("Filter",
{multiplier, channel, kernel, kernel},
filter_data);
net.AddInputFromArray<D, float>(
"Filter", {multiplier, channel, kernel, kernel}, filter_data);
std::vector<float> bias_data(channel * multiplier);
GenerateRandomRealTypeData({channel * multiplier}, &bias_data);
net.AddInputFromArray<D, float>("Bias", {channel * multiplier},
bias_data);
net.AddInputFromArray<D, float>("Bias", {channel * multiplier}, bias_data);
if (D == DeviceType::CPU) {
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("DepthwiseConv2d", "DepthwiseConv2DTest")
.Input("InputNCHW")
......@@ -150,10 +146,8 @@ void ComplexValidTest(index_t batch, index_t channel, index_t height,
.Finalize(net.NewOperatorDef());
// Run
net.RunOp(D);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
NHWC);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW,
"Output", NHWC);
} else if (D == DeviceType::GPU) {
BufferToImage<D, T>(&net, "Input", "InputImage",
kernels::BufferType::IN_OUT_CHANNEL);
......@@ -217,8 +211,8 @@ void ComplexValidTest(index_t batch, index_t channel, index_t height,
}
}
auto expected = CreateTensor<T>(
{1, out_height, out_width, out_channels}, expect);
auto expected =
CreateTensor<T>({1, out_height, out_width, out_channels}, expect);
if (DataTypeToEnum<T>::value == DT_FLOAT) {
ExpectTensorNear<T>(*expected, *net.GetOutput("Output"), 1e-5);
......@@ -249,7 +243,7 @@ TEST_F(DepthwiseConv2dOpTest, ComplexOpenCLHalf) {
}
namespace {
template<typename T>
template <typename T>
void TestNxNS12(const index_t height, const index_t width) {
testing::internal::LogToStderr();
auto func = [&](int kernel_h, int kernel_w, int stride_h, int stride_w,
......@@ -263,18 +257,14 @@ void TestNxNS12(const index_t height, const index_t width) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width,
input_channels});
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, input_channels});
net.AddRandomInput<DeviceType::GPU, float>(
"Filter", {multiplier, input_channels, kernel_h, kernel_w});
net.AddRandomInput<DeviceType::GPU, float>("Bias",
{multiplier
* input_channels});
{multiplier * input_channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("DepthwiseConv2d", "DepthwiseConv2DTest")
.Input("InputNCHW")
......@@ -290,10 +280,8 @@ void TestNxNS12(const index_t height, const index_t width) {
// Run on cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
NHWC);
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW,
"Output", NHWC);
// Check
Tensor expected;
......@@ -319,18 +307,16 @@ void TestNxNS12(const index_t height, const index_t width) {
net.RunOp(DeviceType::GPU);
// Transfer output
ImageToBuffer<DeviceType::GPU, float>(&net,
"OutputImage",
"DeviceOutput",
ImageToBuffer<DeviceType::GPU, float>(&net, "OutputImage", "DeviceOutput",
kernels::BufferType::IN_OUT_CHANNEL);
// Check
if (DataTypeToEnum<T>::value == DT_FLOAT) {
ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"),
1e-5, 1e-4);
ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"), 1e-5,
1e-4);
} else {
ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"),
1e-2, 1e-2);
ExpectTensorNear<float>(expected, *net.GetOutput("DeviceOutput"), 1e-2,
1e-2);
}
};
......@@ -343,9 +329,7 @@ void TestNxNS12(const index_t height, const index_t width) {
}
} // namespace
TEST_F(DepthwiseConv2dOpTest, OpenCLSimpleNxNS12) {
TestNxNS12<float>(4, 4);
}
TEST_F(DepthwiseConv2dOpTest, OpenCLSimpleNxNS12) { TestNxNS12<float>(4, 4); }
TEST_F(DepthwiseConv2dOpTest, OpenCLSimpleNxNS12Half) {
TestNxNS12<half>(4, 4);
......
mace/ops/eltwise.h
浏览文件 @ 2940ce5d
......@@ -26,15 +26,15 @@ class EltwiseOp : public Operator<D, T> {
public:
EltwiseOp(const OperatorDef &op_def, Workspace *ws)
: Operator<D, T>(op_def, ws),
functor_(static_cast<kernels::EltwiseType>(
OperatorBase::GetOptionalArg<int>(
functor_(
static_cast<kernels::EltwiseType>(OperatorBase::GetOptionalArg<int>(
"type", static_cast<int>(kernels::EltwiseType::NONE))),
OperatorBase::GetRepeatedArgs<float>("coeff"),
OperatorBase::GetOptionalArg<float>("x", 1.0)) {}
MaceStatus Run(StatsFuture *future) override {
const Tensor* input0 = this->Input(0);
const Tensor* input1 = this->InputSize() == 2 ? this->Input(1) : nullptr;
const Tensor *input0 = this->Input(0);
const Tensor *input1 = this->InputSize() == 2 ? this->Input(1) : nullptr;
Tensor *output = this->Output(OUTPUT);
return functor_(input0, input1, output, future);
}
......
mace/ops/eltwise_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/folded_batch_norm.h
浏览文件 @ 2940ce5d
......@@ -47,7 +47,7 @@ class FoldedBatchNormOp : public Operator<D, T> {
offset->dim_size());
Tensor *output = this->Output(OUTPUT);
MACE_FAILURE_RETURN(output->ResizeLike(input));
MACE_RETURN_IF_ERROR(output->ResizeLike(input));
return functor_(input, scale, offset, nullptr, nullptr, 0, output, future);
}
......
mace/ops/folded_batch_norm_test.cc
浏览文件 @ 2940ce5d
......@@ -36,7 +36,7 @@ void CalculateScaleOffset(const std::vector<float> &gamma,
}
}
template<DeviceType D>
template <DeviceType D>
void Simple() {
OpsTestNet net;
......@@ -83,10 +83,9 @@ void Simple() {
}
// Check
auto expected =
CreateTensor<float>({1, 6, 2, 1}, {-3.8543, -3.8543, -1.5125, -1.5125,
0.8291, 0.8291, 3.1708, 3.1708,
5.5125, 5.5125, 7.8543, 7.8543});
auto expected = CreateTensor<float>(
{1, 6, 2, 1}, {-3.8543, -3.8543, -1.5125, -1.5125, 0.8291, 0.8291, 3.1708,
3.1708, 5.5125, 5.5125, 7.8543, 7.8543});
ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-4);
}
......@@ -108,14 +107,12 @@ TEST_F(FoldedBatchNormOpTest, SimpleRandomOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("FoldedBatchNorm", "FoldedBatchNormTest")
......@@ -128,9 +125,7 @@ TEST_F(FoldedBatchNormOpTest, SimpleRandomOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -173,14 +168,12 @@ TEST_F(FoldedBatchNormOpTest, SimpleRandomHalfOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("FoldedBatchNorm", "FoldedBatchNormTest")
......@@ -193,9 +186,7 @@ TEST_F(FoldedBatchNormOpTest, SimpleRandomHalfOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -239,14 +230,12 @@ TEST_F(FoldedBatchNormOpTest, ComplexRandomOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("FoldedBatchNorm", "FoldedBatchNormTest")
......@@ -259,9 +248,7 @@ TEST_F(FoldedBatchNormOpTest, ComplexRandomOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......@@ -303,14 +290,12 @@ TEST_F(FoldedBatchNormOpTest, ComplexRandomHalfOPENCL) {
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Scale", {channels});
net.AddRandomInput<DeviceType::GPU, float>("Offset", {channels});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("FoldedBatchNorm", "FoldedBatchNormTest")
......@@ -323,9 +308,7 @@ TEST_F(FoldedBatchNormOpTest, ComplexRandomHalfOPENCL) {
// run cpu
net.RunOp();
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW",
NCHW,
"Output",
net.TransformDataFormat<DeviceType::CPU, float>("OutputNCHW", NCHW, "Output",
NHWC);
// Check
......
mace/ops/fully_connected.h
浏览文件 @ 2940ce5d
......@@ -23,7 +23,7 @@
namespace mace {
namespace ops {
template<DeviceType D, class T>
template <DeviceType D, class T>
class FullyConnectedOp : public Operator<D, T> {
public:
FullyConnectedOp(const OperatorDef &operator_def, Workspace *ws)
......@@ -40,29 +40,19 @@ class FullyConnectedOp : public Operator<D, T> {
Tensor *output = this->Output(OUTPUT);
if (D == DeviceType::CPU) {
MACE_CHECK(input->dim(1) == weight->dim(1)
&& input->dim(2) == weight->dim(2)
&& input->dim(3) == weight->dim(3)
&& weight->dim(0) == bias->dim(0),
"The shape of Input: ",
MakeString(input->shape()),
"The shape of Weight: ",
MakeString(weight->shape()),
" and Bias ",
bias->dim(0),
" don't match.");
MACE_CHECK(
input->dim(1) == weight->dim(1) && input->dim(2) == weight->dim(2) &&
input->dim(3) == weight->dim(3) && weight->dim(0) == bias->dim(0),
"The shape of Input: ", MakeString(input->shape()),
"The shape of Weight: ", MakeString(weight->shape()), " and Bias ",
bias->dim(0), " don't match.");
} else {
MACE_CHECK(input->dim(1) == weight->dim(2)
&& input->dim(2) == weight->dim(3)
&& input->dim(3) == weight->dim(1)
&& weight->dim(0) == bias->dim(0),
"The shape of Input: ",
MakeString(input->shape()),
"The shape of Weight: ",
MakeString(weight->shape()),
" and Bias ",
bias->dim(0),
" don't match.");
MACE_CHECK(
input->dim(1) == weight->dim(2) && input->dim(2) == weight->dim(3) &&
input->dim(3) == weight->dim(1) && weight->dim(0) == bias->dim(0),
"The shape of Input: ", MakeString(input->shape()),
"The shape of Weight: ", MakeString(weight->shape()), " and Bias ",
bias->dim(0), " don't match.");
}
return functor_(input, weight, bias, output, future);
......
mace/ops/fully_connected_test.cc
浏览文件 @ 2940ce5d
......@@ -24,7 +24,7 @@ namespace test {
class FullyConnectedOpTest : public OpsTestBase {};
namespace {
template<DeviceType D>
template <DeviceType D>
void Simple(const std::vector<index_t> &input_shape,
const std::vector<float> &input_value,
const std::vector<index_t> &weight_shape,
......@@ -111,8 +111,8 @@ TEST_F(FullyConnectedOpTest, SimpleOPENCL) {
{2}, {2, 3}, {1, 1, 1, 2}, {387, 3853});
Simple<DeviceType::GPU>(
{1, 1, 2, 3}, {1, 2, 3, 4, 5, 6}, {5, 3, 1, 2},
{1, 4, 2, 5, 3, 6, 10, 40, 20, 50, 30, 60, 1, 4, 2, 5, 3, 6,
10, 40, 20, 50, 30, 60, 1, 4, 2, 5, 3, 6},
{1, 4, 2, 5, 3, 6, 10, 40, 20, 50, 30, 60, 1, 4, 2,
5, 3, 6, 10, 40, 20, 50, 30, 60, 1, 4, 2, 5, 3, 6},
{5}, {1, 2, 3, 4, 5}, {1, 1, 1, 5}, {92, 912, 94, 914, 96});
}
......@@ -122,7 +122,7 @@ TEST_F(FullyConnectedOpTest, SimpleGPUWithBatch) {
}
namespace {
template<typename T>
template <typename T>
void Random(const index_t batch,
const index_t height,
const index_t width,
......@@ -134,15 +134,13 @@ void Random(const index_t batch,
OpsTestNet net;
// Add input data
net.AddRandomInput<DeviceType::GPU, float>(
"Input", {batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>("Input",
{batch, height, width, channels});
net.AddRandomInput<DeviceType::GPU, float>(
"Weight", {out_channel, channels, height, width});
net.AddRandomInput<DeviceType::GPU, float>("Bias", {out_channel});
net.TransformDataFormat<DeviceType::CPU, float>("Input",
NHWC,
"InputNCHW",
net.TransformDataFormat<DeviceType::CPU, float>("Input", NHWC, "InputNCHW",
NCHW);
OpDefBuilder("FullyConnected", "FullyConnectedTest")
.Input("InputNCHW")
......@@ -182,11 +180,11 @@ void Random(const index_t batch,
ImageToBuffer<DeviceType::GPU, float>(&net, "OutputImage", "OPENCLOutput",
kernels::BufferType::IN_OUT_CHANNEL);
if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
1e-1, 1e-1);
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-1,
1e-1);
} else {
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
1e-2, 1e-3);
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2,
1e-3);
}
}
} // namespace
......
mace/ops/local_response_norm.h
浏览文件 @ 2940ce5d
......@@ -25,8 +25,7 @@ template <DeviceType D, class T>
class LocalResponseNormOp : public Operator<D, T> {
public:
LocalResponseNormOp(const OperatorDef &operator_def, Workspace *ws)
: Operator<D, T>(operator_def, ws),
functor_() {
: Operator<D, T>(operator_def, ws), functor_() {
depth_radius_ = OperatorBase::GetOptionalArg<int>("depth_radius", 5);
bias_ = OperatorBase::GetOptionalArg<float>("bias", 1.0f);
alpha_ = OperatorBase::GetOptionalArg<float>("alpha", 1.0f);
......@@ -40,7 +39,7 @@ class LocalResponseNormOp : public Operator<D, T> {
input->dim_size());
Tensor *output = this->Output(OUTPUT);
MACE_FAILURE_RETURN(output->ResizeLike(input));
MACE_RETURN_IF_ERROR(output->ResizeLike(input));
return functor_(input, depth_radius_, bias_, alpha_, beta_, output, future);
}
......
mace/ops/local_response_norm_test.cc
浏览文件 @ 2940ce5d
......@@ -21,7 +21,7 @@ namespace test {
class LocalResponseNormOpTest : public OpsTestBase {};
template<DeviceType D>
template <DeviceType D>
void Simple() {
OpsTestNet net;
......@@ -46,9 +46,9 @@ void Simple() {
}
// Check
auto expected =
CreateTensor<float>({1, 1, 2, 6}, {0.28, 0.28, 0.39, 0.39, 0.51, 0.51,
0.34, 0.34, 0.40, 0.40, 0.47, 0.47});
auto expected = CreateTensor<float>(
{1, 1, 2, 6},
{0.28, 0.28, 0.39, 0.39, 0.51, 0.51, 0.34, 0.34, 0.40, 0.40, 0.47, 0.47});
ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0, 1e-2);
}
......
mace/ops/matmul_test.cc
浏览文件 @ 2940ce5d
......@@ -92,8 +92,7 @@ TEST_F(MatMulOpTest, SimpleCPUWithBatch) {
TEST_F(MatMulOpTest, SimpleOPENCL) {
Simple<DeviceType::GPU>({1, 2, 3, 1}, {1, 2, 3, 4, 5, 6}, {1, 3, 2, 1},
{1, 2, 3, 4, 5, 6}, {1, 2, 2, 1},
{22, 28, 49, 64});
{1, 2, 3, 4, 5, 6}, {1, 2, 2, 1}, {22, 28, 49, 64});
Simple<DeviceType::GPU>(
{1, 5, 5, 1}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,
14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25},
......@@ -127,10 +126,9 @@ void Complex(const index_t batch,
.Finalize(net.NewOperatorDef());
// Add input data
net.AddRandomInput<DeviceType::GPU, float>("A",
{batch, height, channels, 1});
net.AddRandomInput<DeviceType::GPU, float>(
"B", {batch, channels, out_width, 1});
net.AddRandomInput<DeviceType::GPU, float>("A", {batch, height, channels, 1});
net.AddRandomInput<DeviceType::GPU, float>("B",
{batch, channels, out_width, 1});
// run cpu
net.RunOp();
......@@ -158,11 +156,11 @@ void Complex(const index_t batch,
ImageToBuffer<DeviceType::GPU, float>(&net, "OutputImage", "OPENCLOutput",
kernels::BufferType::IN_OUT_HEIGHT);
if (DataTypeToEnum<T>::value == DataType::DT_HALF) {
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
1e-2, 1e-1);
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2,
1e-1);
} else {
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"),
1e-5, 1e-5);
ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-5,
1e-5);
}
}
} // namespace
......
mace/ops/ops_test_util.h
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/pad.h
浏览文件 @ 2940ce5d
......@@ -29,8 +29,7 @@ class PadOp : public Operator<D, T> {
PadOp(const OperatorDef &operator_def, Workspace *ws)
: Operator<D, T>(operator_def, ws),
functor_(OperatorBase::GetRepeatedArgs<int>("paddings"),
OperatorBase::GetOptionalArg<float>("constant_value", 0.0))
{}
OperatorBase::GetOptionalArg<float>("constant_value", 0.0)) {}
MaceStatus Run(StatsFuture *future) override {
const Tensor *input_tensor = this->Input(0);
......
mace/ops/pad_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/pooling_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/proposal_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/quantize.h
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/quantize_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/resize_bilinear.h
浏览文件 @ 2940ce5d
......@@ -26,8 +26,7 @@ class ResizeBilinearOp : public Operator<D, T> {
public:
ResizeBilinearOp(const OperatorDef &operator_def, Workspace *ws)
: Operator<D, T>(operator_def, ws),
functor_(
OperatorBase::GetRepeatedArgs<index_t>("size", {-1, -1}),
functor_(OperatorBase::GetRepeatedArgs<index_t>("size", {-1, -1}),
OperatorBase::GetOptionalArg<bool>("align_corners", false)) {}
MaceStatus Run(StatsFuture *future) override {
......
mace/ops/resize_bilinear_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/slice_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/softmax.h
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/softmax_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/space_to_batch.h
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/space_to_batch_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/space_to_depth.h
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/transpose.h
浏览文件 @ 2940ce5d
此差异已折叠。
mace/ops/transpose_test.cc
浏览文件 @ 2940ce5d
此差异已折叠。
mace/public/mace.h
浏览文件 @ 2940ce5d
此差异已折叠。
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