timekiller
/
mindspore2022
forked from huawei/mindspore2022
0
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
mindspore2022/mindspore/ccsrc/utils/ms_device_shape_transfer.cc

1954 lines
73 KiB
C++
Raw Blame History

/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "utils/ms_device_shape_transfer.h"
#include <functional>
#include <numeric>
#include <utility>
#include <algorithm>
namespace mindspore {
namespace trans {
const int b1 = 1;
const int b2 = 2;
const int b4 = 4;
const int b8 = 8;
const int64_t kCubeSize = 16;
const int64_t kCube16 = kCubeSize;
const int64_t kCube32 = 32;
const int64_t kCube64 = 64;
const int64_t kCubeSize_C04 = 4;
const int64_t kNiSize = 16;
constexpr int kDims2 = 2;
constexpr int64_t k4 = 4;
static const std::set<TypeId> C0_64 = {kNumberTypeInt4};
static const std::set<TypeId> C0_32 = {kNumberTypeUInt8, kNumberTypeInt8};
namespace {
bool HasShapeDynamic(const ShapeVector &shape_list) {
return std::any_of(shape_list.begin(), shape_list.end(), [](int64_t v) { return v == abstract::Shape::SHP_ANY; });
}
template <typename T>
T Gcd(T a, T b) {
if (b == 0) {
return 0;
}
T c = b;
while (a % b != 0) {
c = a % b;
a = b;
b = c;
}
return c;
}
template <typename T>
T Lcm(T a, T b) {
if (b == 0) {
return 0;
}
T ret = (a * b) / (Gcd(a, b));
return ret;
}
template <typename T>
T DivCeil(T n1, T n2) {
if (n2 != 0) {
return (n1 + n2 - 1) / n2;
}
return 0;
}
template <typename T>
bool CheckDims(const std::vector<T> &shape) {
if (shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Host shape dims should be 4";
return false;
}
return true;
}
int64_t GetCubeSizeByType(const TypeId &data_type) {
if (C0_32.find(data_type) != C0_32.end()) {
return kCube32;
}
if (C0_64.find(data_type) != C0_64.end()) {
return kCube64;
}
return kCube16;
}
RangePair PaddingRangeTo5D(const RangePair &ori_range) {
RangePair dst_range(kNcdhw, std::pair<int64_t, int64_t>(1, 1));
switch (ori_range.size()) {
case N_ncdhw:
return ori_range;
case C_ncdhw:
dst_range[C_ncdhw] = ori_range[N_ncdhw];
break;
case D_ncdhw:
dst_range[C_ncdhw] = ori_range[N_ncdhw];
dst_range[D_ncdhw] = ori_range[C_ncdhw];
break;
case H_ncdhw:
dst_range[C_ncdhw] = ori_range[N_ncdhw];
dst_range[D_ncdhw] = ori_range[C_ncdhw];
dst_range[H_ncdhw] = ori_range[D_ncdhw];
break;
case W_ncdhw:
dst_range[C_ncdhw] = ori_range[N_ncdhw];
dst_range[D_ncdhw] = ori_range[C_ncdhw];
dst_range[H_ncdhw] = ori_range[D_ncdhw];
dst_range[W_ncdhw] = ori_range[H_ncdhw];
break;
default:
MS_LOG(EXCEPTION) << "Unexpected shape size = " << ori_range.size();
}
return dst_range;
}
RangePair PaddingRangeTo4D(const RangePair &ori_range) {
RangePair dst_range(kNchwDims, std::pair<int64_t, int64_t>(1, 1));
switch (ori_range.size()) {
case kN:
return dst_range;
case kC:
dst_range[kC] = ori_range[kN];
break;
case kH:
dst_range[kC] = ori_range[kN];
dst_range[kH] = ori_range[kC];
break;
case kW:
dst_range[kC] = ori_range[kN];
dst_range[kH] = ori_range[kC];
dst_range[kW] = ori_range[kH];
break;
case kNchwDims:
(void)std::copy(ori_range.begin(), ori_range.end(), dst_range.begin());
break;
default:
MS_LOG(EXCEPTION) << "Unexpected range size: " << ori_range.size();
}
return dst_range;
}
} // namespace
void StringToAxisVector4D(const std::string &reshape_type_str, std::vector<Axis> *reshape_type_vec) {
MS_EXCEPTION_IF_NULL(reshape_type_vec);
if (reshape_type_str.empty()) {
MS_LOG(DEBUG) << "Reshape type str is empty, no need padding.";
return;
}
for (const auto &c : reshape_type_str) {
switch (c) {
case 'N':
reshape_type_vec->push_back(N);
break;
case 'C':
reshape_type_vec->push_back(C);
break;
case 'H':
reshape_type_vec->push_back(H);
break;
case 'W':
reshape_type_vec->push_back(W);
break;
default:
MS_LOG(EXCEPTION) << "Unknown axis " << c << "in reshape type.";
}
}
}
void StringToAxisVector5D(const std::string &reshape_type_str, std::vector<Axis5D> *reshape_type_vec) {
MS_EXCEPTION_IF_NULL(reshape_type_vec);
if (reshape_type_str.empty()) {
MS_LOG(DEBUG) << "Reshape type str is empty, no need padding.";
return;
}
for (const auto &c : reshape_type_str) {
switch (c) {
case 'N':
reshape_type_vec->push_back(N_ncdhw);
break;
case 'C':
reshape_type_vec->push_back(C_ncdhw);
break;
case 'D':
reshape_type_vec->push_back(D_ncdhw);
break;
case 'H':
reshape_type_vec->push_back(H_ncdhw);
break;
case 'W':
reshape_type_vec->push_back(W_ncdhw);
break;
default:
MS_LOG(EXCEPTION) << "Unknown axis " << c << "in reshape type.";
}
}
}
bool IsNeedPadding(const std::string &format, size_t shape_size) {
if (shape_size == 0) {
return false;
}
if (format == kOpFormat_DEFAULT || format == kOpFormat_NCHW ||
kNoPaddingFormatSet.find(format) != kNoPaddingFormatSet.end()) {
return false;
} else if (shape_size < kNchwDims) {
return true;
}
return false;
}
ShapeVector GetRuntimePaddingShape(const AnfNodePtr &node, size_t index) {
MS_EXCEPTION_IF_NULL(node);
ShapeVector shape;
std::vector<size_t> host_shape;
if (node->isa<ValueNode>()) {
auto value_node = node->cast<ValueNodePtr>();
MS_EXCEPTION_IF_NULL(value_node);
auto node_value = value_node->value();
MS_EXCEPTION_IF_NULL(node_value);
auto tensor = node_value->cast<tensor::TensorPtr>();
if (tensor == nullptr) {
MS_LOG(EXCEPTION) << " The node[ " << node->DebugString() << "]'s cannot convert ";
}
auto shape_temp = tensor->shape();
(void)std::transform(shape_temp.begin(), shape_temp.end(), std::back_inserter(host_shape), LongToSize);
if (host_shape.empty()) {
host_shape.push_back(1);
}
} else {
host_shape = AnfAlgo::GetOutputInferShape(node, index);
}
auto format = AnfAlgo::GetOutputFormat(node, index);
if (IsNeedPadding(format, host_shape.size())) {
host_shape = PaddingShape(host_shape, format, AnfAlgo::GetOutputReshapeType(node, index));
}
std::transform(host_shape.begin(), host_shape.end(), std::back_inserter(shape), SizeToLong);
return shape;
}
bool TransDataType(const TypeIdArgs &args, void *result) {
DataTypeTransfer dataTypeTransfer;
return dataTypeTransfer.TransDataType(args, result);
}
bool TransFormat(const FormatArgs &args, void *result, const AnfNodePtr &node, size_t index) {
FormatTransfer formatTransfer;
return formatTransfer.TransDataByFormat(args, result, node, index, true);
}
bool TransFormatFromDeviceToHost(const FormatArgs &args, void *result, int64_t groups) {
FormatTransfer formatTransfer;
return formatTransfer.TransDataBackwordCore(args, result, groups);
}
bool TransFormatFromDeviceToHost(const FormatArgs &args, void *result, const AnfNodePtr &node, size_t index) {
FormatTransfer formatTransfer;
return formatTransfer.TransDataByFormat(args, result, node, index, false);
}
/**###################### DATA TYPE TRANS ################################*/
void CheckMemSize(const TypeIdArgs &args) {
auto src_type_size = abstract::TypeIdSize(args.src_data_type);
auto dst_type_size = abstract::TypeIdSize(args.dst_data_type);
if (src_type_size < 1 || dst_type_size < 1) {
MS_LOG(EXCEPTION) << "Invalid src or dst data type. Src type: " << TypeIdLabel(args.src_data_type)
<< ", dst type: " << TypeIdLabel(args.dst_data_type);
}
if (SizeToLong(args.data_size / src_type_size) != args.src_shape_size) {
MS_LOG(EXCEPTION) << "Invalid src or dst data shape size. Src shape size: " << args.src_shape_size
<< ", dst shape size: " << args.data_size / src_type_size;
}
}
template <typename SrcT, typename DstT>
void TransDataSrc2Dst(const TypeIdArgs &args, void *dst, const int64_t data_size) {
CheckMemSize(args);
for (int64_t idx = 0; idx != data_size; idx++) {
SrcT src_data = static_cast<const SrcT *>(args.data)[idx];
static_cast<DstT *>(dst)[idx] = static_cast<DstT>(src_data);
}
}
template <typename SrcT>
void TransDataSrc2Fp16(const TypeIdArgs &args, void *dst, const int64_t data_size) {
CheckMemSize(args);
auto src_data = static_cast<const SrcT *>(args.data);
auto half_data = static_cast<float16 *>(dst);
for (int64_t i = 0; i < data_size; i++) {
half_data[i] = float16(src_data[i]);
}
}
bool DataTypeTransfer::CastKernel(const TypeIdArgs &args, void *dst, int64_t data_size, DataTypeTransMode mode) {
using DtypeKernel = std::function<void(const TypeIdArgs &, void *, const int64_t)>;
const std::map<DataTypeTransMode, DtypeKernel> cast_kernel_map{
{DataTypeTransMode::FROM_BOOL_TO_UINT8, TransDataSrc2Dst<int8_t, uint8_t>},
{DataTypeTransMode::FROM_BOOL_TO_INT32, TransDataSrc2Dst<int8_t, int32_t>},
{DataTypeTransMode::FROM_BOOL_TO_FLOAT16, TransDataSrc2Fp16<int8_t>},
{DataTypeTransMode::FROM_BOOL_TO_FLOAT, TransDataSrc2Dst<int8_t, float>},
{DataTypeTransMode::FROM_INT8_TO_INT32, TransDataSrc2Dst<int8_t, int32_t>},
{DataTypeTransMode::FROM_INT8_TO_FLOAT16, TransDataSrc2Fp16<int8_t>},
{DataTypeTransMode::FROM_INT8_TO_FLOAT, TransDataSrc2Dst<int8_t, float>},
{DataTypeTransMode::FROM_UINT8_TO_INT32, TransDataSrc2Dst<uint8_t, int32_t>},
{DataTypeTransMode::FROM_UINT8_TO_FLOAT16, TransDataSrc2Fp16<uint8_t>},
{DataTypeTransMode::FROM_UINT8_TO_FLOAT, TransDataSrc2Dst<uint8_t, float>},
{DataTypeTransMode::FROM_UINT16_TO_INT32, TransDataSrc2Dst<uint16_t, int32_t>},
{DataTypeTransMode::FROM_INT32_TO_BOOL, TransDataSrc2Dst<int32_t, int8_t>},
{DataTypeTransMode::FROM_INT32_TO_INT8, TransDataSrc2Dst<int32_t, int8_t>},
{DataTypeTransMode::FROM_INT32_TO_UINT8, TransDataSrc2Dst<int32_t, uint8_t>},
{DataTypeTransMode::FROM_INT32_TO_FLOAT16, TransDataSrc2Fp16<int32_t>},
{DataTypeTransMode::FROM_INT32_TO_FLOAT, TransDataSrc2Dst<int32_t, float>},
{DataTypeTransMode::FROM_INT32_TO_INT64, TransDataSrc2Dst<int32_t, int64_t>},
{DataTypeTransMode::FROM_INT64_TO_INT32, TransDataSrc2Dst<int64_t, int32_t>},
{DataTypeTransMode::FROM_FLOAT16_TO_UINT8, TransDataSrc2Dst<float16, uint8_t>},
{DataTypeTransMode::FROM_FLOAT16_TO_INT32, TransDataSrc2Dst<float16, int32_t>},
{DataTypeTransMode::FROM_FLOAT_TO_INT32, TransDataSrc2Dst<float, int32_t>},
{DataTypeTransMode::FROM_FLOAT32_TO_FLOAT64, TransDataSrc2Dst<float, double>},
{DataTypeTransMode::FROM_FLOAT64_TO_FLOAT32, TransDataSrc2Dst<double, float>}};
if (mode == DataTypeTransMode::FROM_FLOAT_TO_FLOAT16) {
device::FloatToHalf(dst, args.data, data_size);
return true;
} else if (mode == DataTypeTransMode::FROM_FLOAT16_TO_FLOAT) {
device::HalfToFloat(dst, args.data, data_size);
return true;
}
auto iter = cast_kernel_map.find(mode);
if (iter != cast_kernel_map.end()) {
iter->second(args, dst, data_size);
return true;
} else {
MS_LOG(ERROR) << "Can not find a datatype trans function. Src type :" << TypeIdLabel(args.src_data_type)
<< ", dst_type:" << TypeIdLabel(args.dst_data_type);
return false;
}
}
bool DataTypeTransfer::TransDataType(const TypeIdArgs &args, void *result) {
MS_LOG(DEBUG) << "Begin trans datatype from " << TypeIdLabel(args.src_data_type) << " to "
<< TypeIdLabel(args.dst_data_type);
MS_EXCEPTION_IF_NULL(result);
std::pair<TypeId, TypeId> type_info(args.src_data_type, args.dst_data_type);
auto iter = mode_map.find(type_info);
if (iter == mode_map.end()) {
MS_LOG(ERROR) << "Can not find a datatype trans type. src_type :" << TypeIdLabel(args.src_data_type)
<< ", dst_type:" << TypeIdLabel(args.dst_data_type);
return false;
}
auto trans_mode = iter->second;
if (!CastKernel(args, result, args.src_shape_size, trans_mode)) {
MS_LOG(ERROR) << "Failed to trans datatype. Src: " << TypeIdLabel(args.src_data_type)
<< ", dst: " << TypeIdLabel(args.dst_data_type);
return false;
}
return true;
}
/**###################### DATA SHAPE TRANS ################################*/
ShapeVector DeviceShapeTransfer::GetDeviceShapeByFormat(const ShapeVector &shape, const std::string &format,
const AnfNodePtr &node, size_t index, const TypeId &type,
bool is_output) {
auto dev_shape = GetFixedDeviceShape(shape, node, index, is_output);
if (dev_shape.has_value()) {
return dev_shape.value();
}
int64_t groups = 1;
if (format == kOpFormat_FRAC_Z) {
groups = AnfAlgo::GetAttrGroups(node, index);
}
ShapeVector input_hidden_size = {kAlign16, kAlign16};
if (format == kOpFormat_FRACTAL_ZN_RNN || format == kOpFormat_ND_RNN_BIAS) {
input_hidden_size = GetAttrInputAndHiddenSize(node);
}
if (node != nullptr) {
MS_LOG(DEBUG) << "Start trans infer shape to device shape for node: " << node->DebugString()
<< ", format: " << format;
}
return TransCore(shape, format, type, groups, input_hidden_size);
}
ShapeVector DeviceShapeTransfer::GetDeviceShapeByFormat(const ShapeVector &shape, const std::string &format,
const TypeId &type, int64_t groups,
const ShapeVector &input_hidden_size) {
return TransCore(shape, format, type, groups, input_hidden_size);
}
std::optional<ShapeVector> DeviceShapeTransfer::GetFixedDeviceShape(const ShapeVector &, const AnfNodePtr &node,
size_t index, bool is_output) {
if (node == nullptr || !node->isa<CNode>()) {
return {};
}
auto attr_name = is_output ? kAttrFixedOutputDeviceShape : kAttrFixedInputDeviceShape;
auto cnode = node->cast<CNodePtr>();
if (!AnfAlgo::HasNodeAttr(attr_name, cnode)) {
return {};
}
auto shapes = AnfAlgo::GetNodeAttr<std::vector<ShapeVector>>(cnode, attr_name);
if (index >= shapes.size()) {
MS_LOG(INFO) << "Index is out of range, got index: " << index << ", shape size: " << shapes.size();
return {};
}
return std::optional<ShapeVector>(std::move(shapes[index]));
}
ShapeVector DeviceShapeTransfer::TransCore(const ShapeVector &shape, const std::string &format, const TypeId &type,
int64_t groups, const ShapeVector &input_hidden_size) {
using DeviceShapeTransfer = std::function<ShapeVector(const ShapeVector &, const TypeId &)>;
const std::map<std::string, DeviceShapeTransfer> device_shape_map = {
{kOpFormat_NCHW, NCHWDeviceShape},
{kOpFormat_NHWC, NHWCDeviceShape},
{kOpFormat_HWCN, HWCNDeviceShape},
{kOpFormat_NCDHW, NCDHWDeviceShape},
{kOpFormat_FRAC_Z, FRAC_ZDeviceShape},
{kOpFormat_FRAC_NZ, FRAC_NZDeviceShape},
{kOpFormat_NC1HWC0, NC1HWC0DeviceShape},
{kOpFormat_NDC1HWC0, NDC1HWC0DeviceShape},
{kOpFormat_C1HWNCoC0, C1HWNCOC0DeviceShape},
{kOpFormat_NC1HWC0_C04, NC1HWC04DeviceShape},
{kOpFormat_FRACTAL_Z_3D, FRAC_Z3DDeviceShape},
{kOpFormat_FRACTAL_Z_C04, FRAC_ZC04DeviceShape},
{kOpFormat_ChannelLast, ChannelLastDeviceShape},
{kOpFormat_FRACTAL_ZN_LSTM, FRAC_ZN_LSTMDeviceShape}};
if (format == kOpFormat_ND || format == kOpFormat_DEFAULT || format == kOpFormat_NCHW) {
return shape;
}
if (groups > 1 && format == kOpFormat_FRAC_Z) {
return FRAC_ZDeviceShapeWithGroups(shape, type, groups);
}
if (format == kOpFormat_FRACTAL_ZN_RNN) {
return FRAC_ZN_RNNDeviceShape(shape, type, input_hidden_size);
}
if (format == kOpFormat_ND_RNN_BIAS) {
return NDRNNBiasDeviceShape(shape, type, input_hidden_size[1]);
}
auto temp_shape = shape;
if (kNoPaddingFormatSet.find(format) == kNoPaddingFormatSet.end() && format != kOpFormat_FRACTAL_ZN_LSTM &&
shape.size() < kNchwDims && k3DFormatSet.find(format) == k3DFormatSet.end()) {
MS_LOG(WARNING) << "Origin shape size is less than 4, should be Padding shape by Default firstly";
temp_shape = PaddingShapeTo4dDefault(shape);
}
if (shape.size() != kNcdhw && k3DFormatSet.find(format) != k3DFormatSet.end()) {
temp_shape = PaddingShapeTo5dDefault(shape);
}
auto iter = device_shape_map.find(format);
if (iter == device_shape_map.end()) {
MS_LOG(EXCEPTION) << "Unexpected format[" << format << "]";
}
return iter->second(temp_shape, type);
}
ShapeVector DeviceShapeTransfer::NCHWDeviceShape(const ShapeVector &shape, const TypeId &) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
return shape;
}
ShapeVector DeviceShapeTransfer::NHWCDeviceShape(const ShapeVector &shape, const TypeId &) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
device_shape.push_back(shape[kN]);
device_shape.push_back(shape[kH]);
device_shape.push_back(shape[kW]);
device_shape.push_back(shape[kC]);
return device_shape;
}
ShapeVector DeviceShapeTransfer::HWCNDeviceShape(const ShapeVector &shape, const TypeId &) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
device_shape.push_back(shape[kH]);
device_shape.push_back(shape[kW]);
device_shape.push_back(shape[kC]);
device_shape.push_back(shape[kN]);
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_ZDeviceShape(const ShapeVector &shape, const TypeId &type) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
auto c0 = GetCubeSizeByType(type);
if (HasShapeDynamic({shape[kC], shape[kH], shape[kW]})) {
device_shape.push_back(abstract::Shape::SHP_ANY);
} else {
auto c1 = (shape[kC] + c0 - 1) / c0;
device_shape.push_back(shape[kH] * shape[kW] * c1);
}
if (shape[kN] == abstract::Shape::SHP_ANY) {
device_shape.push_back(abstract::Shape::SHP_ANY);
} else {
auto no = (shape[kN] + kNiSize - 1) / kNiSize;
device_shape.push_back(no);
}
device_shape.push_back(kNiSize);
device_shape.push_back(c0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::NC1HWC0DeviceShape(const ShapeVector &shape, const TypeId &type) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
auto c0 = GetCubeSizeByType(type);
auto c1 = (shape[kC] == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : (shape[kC] + c0 - 1) / c0;
device_shape.push_back(shape[kN]);
device_shape.push_back(c1);
device_shape.push_back(shape[kH]);
device_shape.push_back(shape[kW]);
device_shape.push_back(c0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::NDC1HWC0DeviceShape(const ShapeVector &shape, const TypeId &type) {
if (shape.size() != kNcdhw) {
MS_LOG(EXCEPTION) << "Check dims failed, expect shape dim 5, but got shape dim : " << shape.size();
}
ShapeVector device_shape;
auto c0 = GetCubeSizeByType(type);
auto c1 = (shape[1] == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : (shape[1] + c0 - 1) / c0;
device_shape.push_back(shape[N_ncdhw]);
device_shape.push_back(shape[D_ncdhw]);
device_shape.push_back(c1);
device_shape.push_back(shape[H_ncdhw]);
device_shape.push_back(shape[W_ncdhw]);
device_shape.push_back(c0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_Z3DDeviceShape(const ShapeVector &shape, const TypeId &type) {
if (shape.size() != kNcdhw) {
MS_LOG(EXCEPTION) << "Check dims failed, expect shape dim 5, but got shape dim : " << shape.size();
}
ShapeVector device_shape;
auto c0 = GetCubeSizeByType(type);
if (HasShapeDynamic({shape[C_ncdhw], shape[D_ncdhw], shape[H_ncdhw], shape[W_ncdhw]})) {
device_shape.push_back(abstract::Shape::SHP_ANY);
} else {
auto c1 = (shape[1] + c0 - 1) / c0;
device_shape.push_back(shape[D_ncdhw] * c1 * shape[H_ncdhw] * shape[W_ncdhw]);
}
auto no = (shape[0] == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : (shape[0] + kNiSize - 1) / kNiSize;
device_shape.push_back(no);
device_shape.push_back(kNiSize);
device_shape.push_back(c0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::C1HWNCOC0DeviceShape(const ShapeVector &shape, const TypeId &type) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
auto c0 = GetCubeSizeByType(type);
if (shape[kC] == abstract::Shape::SHP_ANY) {
device_shape.push_back(abstract::Shape::SHP_ANY);
} else {
device_shape.push_back((shape[kC] - 1) / c0 + 1);
}
device_shape.push_back(shape[kH]);
device_shape.push_back(shape[kW]);
device_shape.push_back(shape[kN]);
device_shape.push_back(c0);
device_shape.push_back(c0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_ZC04DeviceShape(const ShapeVector &shape, const TypeId &type) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
const int64_t C04 = 4;
int64_t first_dim;
if (HasShapeDynamic({shape[kH], shape[kW]})) {
first_dim = abstract::Shape::SHP_ANY;
} else {
first_dim = DivCeil(C04 * shape[kH] * shape[kW], kCubeSize);
}
auto no = (shape[kN] == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : DivCeil(shape.at(kN), kCubeSize);
device_shape.push_back(first_dim);
device_shape.push_back(no);
device_shape.push_back(kCubeSize);
device_shape.push_back(kCubeSize);
return device_shape;
}
ShapeVector DeviceShapeTransfer::NC1HWC04DeviceShape(const ShapeVector &shape, const TypeId &) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
ShapeVector device_shape;
const int64_t C04 = 4;
const int64_t C1 = (shape[kC] == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : DivCeil(shape.at(kC), C04);
device_shape.push_back(shape[kN]);
device_shape.push_back(C1);
device_shape.push_back(shape[kH]);
device_shape.push_back(shape[kW]);
device_shape.push_back(C04);
return device_shape;
}
ShapeVector DeviceShapeTransfer::NCDHWDeviceShape(const ShapeVector &shape, const TypeId &) {
if (shape.size() < kNcdhw) {
MS_LOG(EXCEPTION) << "Shape dims must be 5 when format is ndhwc.";
}
return shape;
}
ShapeVector DeviceShapeTransfer::ChannelLastDeviceShape(const ShapeVector &shape, const TypeId &) {
auto dim = shape.size();
ShapeVector axis;
axis.resize(dim);
const int step_value = 2;
std::iota(axis.begin() + 1, axis.end(), step_value);
axis[dim - 1] = 1;
ShapeVector device_shape;
(void)std::transform(axis.begin(), axis.end(), std::back_inserter(device_shape),
[&shape](int64_t n) { return shape[n]; });
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_NZDeviceShape(const ShapeVector &shape, const TypeId &type) {
ShapeVector device_shape;
auto c0 = GetCubeSizeByType(type);
if (shape.size() == 1 && (shape[0] == 1 || shape[0] % c0 == 0)) {
// For [1] and [1024] shape we can trait it as NZ shape
return shape;
}
if (shape.size() < kShape2dDims) {
MS_LOG(EXCEPTION) << "Format FRACTAL_NZ don't support shape with " << shape.size() << " dims";
} else {
const auto remove_dim = 2;
(void)std::copy(shape.begin(), shape.end() - remove_dim, std::back_inserter(device_shape));
}
int64_t h_shape = shape[shape.size() - kH];
int64_t w_shape = shape[shape.size() - 1];
int64_t w1 = (w_shape == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : (w_shape - 1) / c0 + 1;
int64_t h1 = (h_shape == abstract::Shape::SHP_ANY) ? abstract::Shape::SHP_ANY : (h_shape - 1) / kCubeSize + 1;
device_shape.push_back(w1);
device_shape.push_back(h1);
device_shape.push_back(kCubeSize);
device_shape.push_back(c0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_ZN_LSTMDeviceShape(const ShapeVector &shape, const TypeId &type) {
ShapeVector device_shape;
const int64_t lstm_ni = 4;
const int64_t ni = 16;
int64_t first = abstract::Shape::SHP_ANY;
int64_t second = abstract::Shape::SHP_ANY;
if (!HasShapeDynamic({shape[kN], shape[kC]})) {
const int64_t h = shape.at(kN) / lstm_ni;
const int64_t i = shape.at(kC) - h;
first = DivCeil(i, ni) + DivCeil(h, ni);
second = lstm_ni * DivCeil(h, ni);
}
device_shape.push_back(first);
device_shape.push_back(second);
device_shape.push_back(ni);
device_shape.push_back(ni);
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_ZDeviceShapeWithGroups(const ShapeVector &shape, const TypeId &type,
int64_t groups) {
if (!CheckDims(shape)) {
MS_LOG(EXCEPTION) << "Check dims failed.";
}
if (groups <= 0) {
MS_LOG(EXCEPTION) << "The value of groups should be greater than 0, but got " << groups;
}
auto cube_size = GetCubeSizeByType(type);
auto c1_dim = abstract::Shape::SHP_ANY;
auto g_dim = abstract::Shape::SHP_ANY;
auto n1 = abstract::Shape::SHP_ANY;
if (!HasShapeDynamic({shape[kC], shape[kN]})) {
auto group_size = groups;
auto cin_ori_tmp = static_cast<int64_t>(shape[kC]);
auto cout_ori_tmp = static_cast<int64_t>(shape[kN]) / group_size;
auto e_mult =
std::min(Lcm(Lcm(cin_ori_tmp, cube_size) / cin_ori_tmp, Lcm(cout_ori_tmp, cube_size) / cout_ori_tmp), group_size);
auto cin_opt = DivCeil(e_mult * cin_ori_tmp, cube_size) * cube_size;
c1_dim = cin_opt / cube_size;
g_dim = DivCeil(group_size, e_mult);
n1 = DivCeil(cout_ori_tmp * e_mult, cube_size);
}
ShapeVector device_shape;
if (!HasShapeDynamic({shape[kC], shape[kN], shape[kH], shape[kW]})) {
device_shape.push_back(g_dim * c1_dim * shape[kH] * shape[kW]);
} else {
device_shape.push_back(abstract::Shape::SHP_ANY);
}
device_shape.push_back(n1);
device_shape.push_back(kNiSize);
device_shape.push_back(cube_size);
return device_shape;
}
ShapeVector DeviceShapeTransfer::FRAC_ZN_RNNDeviceShape(const ShapeVector &shape, const TypeId &type,
const ShapeVector &input_hidden_size) {
if (shape.size() < kShape2dDims) {
MS_LOG(EXCEPTION) << "Format FRACTAL_NZ_RNN don't support shape with " << shape.size() << " dims";
}
auto C0 = GetCubeSizeByType(type);
auto input_size = input_hidden_size[0];
auto hidden_size = input_hidden_size[1];
auto dim_last1 = shape[shape.size() - 1];
auto dim_last2 = shape[shape.size() - kDim2];
const int64_t NUM16 = 16;
ShapeVector device_shape = shape;
if (dim_last2 == abstract::Shape::SHP_ANY) {
device_shape[shape.size() - kDim2] = abstract::Shape::SHP_ANY;
} else if (dim_last2 == input_size || dim_last2 == hidden_size) {
device_shape[shape.size() - kDim2] = DivCeil(dim_last2, NUM16);
} else if (dim_last2 == input_size + hidden_size) {
device_shape[shape.size() - kDim2] = DivCeil(input_size, NUM16) + DivCeil(hidden_size, NUM16);
} else {
MS_LOG(EXCEPTION) << "The second-last dim value of shape is invalid.";
}
if (dim_last1 == abstract::Shape::SHP_ANY) {
device_shape[shape.size() - kDim1] = abstract::Shape::SHP_ANY;
} else {
if (dim_last1 % hidden_size != 0) {
MS_LOG(EXCEPTION) << "Last dim of shape " << shape << " should be multiple of hidden_size " << hidden_size;
}
int64_t n_num = shape[shape.size() - 1] / hidden_size;
device_shape[shape.size() - kDim1] = n_num * DivCeil(hidden_size, C0);
}
device_shape.push_back(NUM16);
device_shape.push_back(C0);
return device_shape;
}
ShapeVector DeviceShapeTransfer::NDRNNBiasDeviceShape(const ShapeVector &shape, const TypeId &type,
int64_t hidden_size) {
if (shape.empty()) {
MS_LOG(EXCEPTION) << "Format ND_RNN_BIAS don't support empty shape.";
}
auto C0 = GetCubeSizeByType(type);
ShapeVector device_shape = shape;
// cppcheck-suppress *
auto dim_last1 = shape[shape.size() - 1];
if (dim_last1 == abstract::Shape::SHP_ANY) {
device_shape[shape.size() - 1] = abstract::Shape::SHP_ANY;
} else {
if (hidden_size <= 0 || dim_last1 % hidden_size != 0) {
MS_LOG(EXCEPTION) << "Last dim of shape " << shape << " should be multiple of hidden_size " << hidden_size;
}
int64_t n_num = shape[shape.size() - 1] / hidden_size;
device_shape[shape.size() - 1] = n_num * DivCeil(hidden_size, C0) * C0;
}
return device_shape;
}
ShapeVector DeviceShapeTransfer::GetAttrInputAndHiddenSize(const AnfNodePtr &node) {
MS_EXCEPTION_IF_NULL(node);
std::vector<int64_t> input_hidden_size = {kAlign16, kAlign16};
if (!node->isa<CNode>() && !node->isa<Parameter>()) {
return input_hidden_size;
}
if (node->isa<Parameter>()) {
auto param = node->cast<ParameterPtr>();
input_hidden_size[0] = param->input_size();
input_hidden_size[1] = param->hidden_size();
} else {
CNodePtr cnode = node->cast<CNodePtr>();
if (cnode == nullptr || !AnfAlgo::HasNodeAttr(kAttrHiddenSize, cnode) ||
!AnfAlgo::HasNodeAttr(kAttrInputSize, cnode)) {
MS_LOG(EXCEPTION)
<< "Node with format FRACTAL_ZN_RNN or ND_RNN_BIAS should have hidden_size or input_size attr. Node info:"
<< node->DebugString();
}
input_hidden_size[0] = AnfAlgo::GetNodeAttr<int64_t>(cnode, kAttrInputSize);
input_hidden_size[1] = AnfAlgo::GetNodeAttr<int64_t>(cnode, kAttrHiddenSize);
}
return input_hidden_size;
}
/**###################### DATA FORMAT TRANS ################################*/
inline void SetData(int64_t size, bool pad_zero, int64_t src_idx, int64_t dst_idx, const FormatArgs &args,
void *result) {
switch (size) {
case b1:
static_cast<uint8_t *>(result)[dst_idx] = pad_zero ? 0 : static_cast<const uint8_t *>(args.data)[src_idx];
break;
case b2:
static_cast<uint16_t *>(result)[dst_idx] = pad_zero ? 0 : static_cast<const uint16_t *>(args.data)[src_idx];
break;
case b4:
static_cast<uint32_t *>(result)[dst_idx] = pad_zero ? 0 : static_cast<const uint32_t *>(args.data)[src_idx];
break;
case b8:
static_cast<uint64_t *>(result)[dst_idx] = pad_zero ? 0 : static_cast<const uint64_t *>(args.data)[src_idx];
break;
default:
MS_LOG(EXCEPTION) << "Trans data not support size " << size;
}
}
bool FormatTransfer::TransDataByFormat(const FormatArgs &args, void *result, const AnfNodePtr &node, size_t index,
bool is_forward) {
int64_t groups = 1;
if (args.device_format == kOpFormat_FRAC_Z && node != nullptr) {
groups = AnfAlgo::GetAttrGroups(node, index);
}
if (is_forward) {
return TransDataForwardCore(args, result, groups);
}
return TransDataBackwordCore(args, result, groups);
}
bool FormatTransfer::TransDataForwardCore(const FormatArgs &args, void *result, int64_t groups) {
MS_LOG(DEBUG) << "Start trans format.";
if (abstract::TypeIdSize(args.src_data_type) < 1) {
MS_LOG(ERROR) << "Invalid datatype: " << args.src_data_type;
return false;
}
if (groups > 1 && args.device_format == kOpFormat_FRAC_Z) {
return NCHW_TO_FRAC_Z_WITH_GROPUS(args, result, true, groups);
}
auto iter = format_trans_fp_map.find(args.device_format);
if (iter == format_trans_fp_map.end()) {
MS_LOG(EXCEPTION) << "Unexpected format[" << args.device_format << "]";
}
return iter->second(args, result);
}
bool FormatTransfer::TransDataBackwordCore(const FormatArgs &args, void *result, int64_t groups) {
MS_LOG(DEBUG) << "Start trans format.";
if (abstract::TypeIdSize(args.src_data_type) < 1) {
MS_LOG(ERROR) << "Invalid datatype, type: " << args.src_data_type;
return false;
}
if (groups > 1 && args.device_format == kOpFormat_FRAC_Z) {
return FRAC_Z_TO_NCHW_WITH_GROUPS(args, result, groups);
}
auto iter = format_trans_bp_map.find(args.device_format);
if (iter == format_trans_bp_map.end()) {
MS_LOG(EXCEPTION) << "Unexpected format[" << args.device_format << "]";
}
return iter->second(args, result);
}
bool FormatTransfer::CheckArgs(const FormatArgs &args, int64_t *size) {
if (args.host_shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Invalid host shape, host shape dims:" << args.host_shape.size() << ", expect dims:" << kNchwDims;
return false;
}
MS_EXCEPTION_IF_NULL(size);
*size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (*size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * (*size);
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
return true;
}
bool FormatTransfer::TransShapeToHW_NZ(const ShapeVector &host_shape, ShapeVector *hw_shape) {
MS_EXCEPTION_IF_NULL(hw_shape);
if (host_shape.empty()) {
MS_LOG(ERROR) << "Size of vector is 0.";
return false;
}
switch (host_shape.size()) {
case 1:
hw_shape->push_back(1);
hw_shape->push_back(1);
hw_shape->push_back(host_shape[0]);
return true;
default:
auto size = host_shape.size();
if (size < kDim2) {
MS_LOG(ERROR) << "Illegal size: " << size;
return false;
}
int64_t times = 1;
for (size_t i = 0; i != size - kDim2; i++) {
times *= host_shape[i];
}
hw_shape->push_back(times);
hw_shape->push_back(host_shape[size - kDim2]);
hw_shape->push_back(host_shape[size - kDim1]);
return true;
}
}
bool FormatTransfer::NCHW_TO_4D(const FormatArgs &args, void *result) {
// trans nchw to NHWC or HWCN
MS_LOG(DEBUG) << "Trans format from nchw to " << args.device_format;
MS_EXCEPTION_IF_NULL(result);
int64_t size = 0;
if (!CheckArgs(args, &size)) {
MS_LOG(ERROR) << "Check args failed.";
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
for (int64_t ni = 0; ni < n; ni++) {
for (int64_t ci = 0; ci < c; ci++) {
for (int64_t hi = 0; hi < h; hi++) {
for (int64_t wi = 0; wi < w; wi++) {
auto src_idx = ni * c * h * w + ci * h * w + hi * w + wi;
int64_t dst_idx = 0;
if (args.device_format == kOpFormat_NHWC) {
dst_idx = ni * h * w * c + hi * w * c + wi * c + ci;
} else if (args.device_format == kOpFormat_HWCN) {
dst_idx = hi * w * c * n + wi * c * n + ci * n + ni;
}
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
}
return true;
}
bool FormatTransfer::TO_NCHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format to nchw from " << args.device_format;
MS_EXCEPTION_IF_NULL(result);
int64_t size = 0;
if (!CheckArgs(args, &size)) {
MS_LOG(ERROR) << "Check args failed.";
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
for (int64_t ni = 0; ni < n; ni++) {
for (int64_t ci = 0; ci < c; ci++) {
for (int64_t hi = 0; hi < h; hi++) {
for (int64_t wi = 0; wi < w; wi++) {
auto dst_idx = ni * c * h * w + ci * h * w + hi * w + wi;
int64_t src_idx = 0;
if (args.device_format == kOpFormat_NHWC) {
src_idx = ni * h * w * c + hi * w * c + wi * c + ci;
} else if (args.device_format == kOpFormat_HWCN) {
src_idx = hi * w * c * n + wi * c * n + ci * n + ni;
}
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
}
return true;
}
bool FormatTransfer::NCHW_TO_FRAC_Z(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from nchw to frac_z";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Invalid host shape, host shape dims:" << args.host_shape.size() << ", expect dims:" << kNchwDims;
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
auto c0 = GetCubeSizeByType(args.src_data_type);
auto c1 = DivCeil(c, c0);
auto hw = h * w;
auto chw = c * hw;
auto hwc0 = hw * c0;
auto nchw = n * chw;
auto hf_cnt = DivCeil(n, kNiSize);
auto vf_cnt = c1 * hw;
auto fractal_ele_cnt = c0 * kNiSize;
auto total_ele_cnt = hf_cnt * vf_cnt * fractal_ele_cnt;
auto dst_size = total_ele_cnt * size;
if (dst_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size."
<< "dst size is :" << dst_size << "device size is :" << args.device_size;
return false;
}
for (int64_t vfi = 0; vfi < vf_cnt; vfi++) {
auto vf_base_i = vfi * hf_cnt; // vertical fractal matrix base index
for (int64_t hfi = 0; hfi < hf_cnt; hfi++) {
auto gfi = vf_base_i + hfi; // global fractal matrix index
auto src_n_offset = hfi * chw * kNiSize;
auto src_f_offset = src_n_offset + vfi % hw + vfi / hw * hwc0;
for (int64_t row = 0; row < c0; row++) {
auto src_ci = vfi / hw * c0 + row;
auto src_row_offset = src_f_offset + row * hw;
for (int64_t col = 0; col < kNiSize; col++) {
auto src_ni = hfi * kNiSize + col;
auto src_idx = src_row_offset + chw * col;
auto dst_idx = gfi * fractal_ele_cnt + col * c0 + row;
auto pad_zero = src_ni >= n || src_idx >= nchw || src_ci >= c;
SetData(size, pad_zero, src_idx, dst_idx, args, result);
}
}
}
}
return true;
}
bool FormatTransfer::NCHW_TO_FRAC_NZ(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from nchw to frac_nz.";
MS_EXCEPTION_IF_NULL(result);
ShapeVector hw_shape;
if (!TransShapeToHW_NZ(args.host_shape, &hw_shape)) {
MS_LOG(ERROR) << "Trans shape failed..";
return false;
}
if (hw_shape.size() < kDim3 || args.device_shape.size() < kDim4) {
MS_LOG(ERROR) << "Invalid shape size.";
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto dst_size = abstract::ShapeSize(args.device_shape) * size;
if (dst_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << dst_size << ", device_size:" << args.device_size;
return false;
}
auto times = hw_shape.at(0);
auto h = hw_shape.at(1);
auto w = hw_shape.at(2);
auto hw = h * w;
auto shape_size = args.device_shape.size();
auto w1 = args.device_shape[shape_size - 4];
auto h1 = args.device_shape[shape_size - 3];
auto h0 = args.device_shape[shape_size - 2];
auto w0 = args.device_shape[shape_size - 1];
auto h1h0w0 = h1 * h0 * w0;
auto w1h1h0w0 = w1 * h1h0w0;
auto num_w1 = w / w0;
for (int64_t times_idx = 0; times_idx < times; times_idx++) {
auto times_head = times_idx * w1h1h0w0;
auto src_times_head = times_idx * hw;
for (int64_t h1h0_idx = 0; h1h0_idx < h; h1h0_idx++) {
auto h1h0_head = times_head + h1h0_idx * w0;
auto src_h_head = src_times_head + h1h0_idx * w;
for (int64_t w1_idx = 0; w1_idx < num_w1; w1_idx++) {
for (int64_t i = 0; i < w0; ++i) {
int64_t src_idx = src_h_head + w1_idx * w0 + i;
int64_t dst_idx = h1h0_head + w1_idx * h1h0w0 + i;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
auto w1_head = num_w1 * w0;
for (int64_t w0_idx = 0; w1_head + w0_idx < w; w0_idx++) {
auto src_w_idx = w1_head + w0_idx;
int64_t dst_idx = h1h0_head + num_w1 * h1h0w0 + w0_idx;
int64_t src_idx = src_h_head + src_w_idx;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
return true;
}
bool FormatTransfer::NCHW_TO_FRAC_ZC04(const FormatArgs &args, void *result) {
// trans nchw to FracZc04
MS_LOG(DEBUG) << "Trans format from nchw to FracZc04.";
MS_EXCEPTION_IF_NULL(result);
int64_t size = 0;
if (!CheckArgs(args, &size)) {
MS_LOG(ERROR) << "Check args failed.";
return false;
}
auto cube = GetCubeSizeByType(args.src_data_type);
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
const int64_t c0 = 4;
auto c1 = DivCeil(c, c0);
auto hwc0 = h * w * c0;
auto hwc = h * w * c;
auto nhwc = n * h * w * c;
auto n_cnt = DivCeil(n, kNiSize);
auto v_cnt = DivCeil(h * w * c0 * c1, cube);
int64_t dst_idx = 0;
for (int64_t vi = 0; vi < v_cnt; vi++) {
for (int64_t ni = 0; ni < n_cnt; ni++) {
for (int64_t col = 0; col < kNiSize; col++) {
for (int64_t row = 0; row < kNiSize; row++) {
int64_t cur_cube_n = kNiSize * ni + col;
int64_t cur_cube_c1hwc0 = kNiSize * vi + row;
auto desc_g = cur_cube_n / n;
auto desc_n = cur_cube_n % n;
auto desc_c1 = cur_cube_c1hwc0 / hwc0;
auto desc_c0 = cur_cube_c1hwc0 % c0;
auto desc_h = (cur_cube_c1hwc0 - hwc0 * desc_c1) / (w * c0);
auto desc_w = (cur_cube_c1hwc0 - hwc0 * desc_c1 - w * c0 * desc_h) / c0;
auto c_idx = desc_c1 * c0 + desc_c0;
auto src_idx = desc_g * nhwc + desc_n * hwc + c_idx * h * w + desc_h * w + desc_w;
auto pad_zero = desc_g >= 1 || desc_n >= n || c_idx >= c;
SetData(size, pad_zero, src_idx, dst_idx, args, result);
dst_idx++;
}
}
}
}
return true;
}
bool FormatTransfer::NCHW_TO_NC1HWC0(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from nchw to Nc1h1wc0";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Invalid host shape, host shape dims:" << args.host_shape.size() << ", expect dims:" << kNchwDims;
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
auto c0 = GetCubeSizeByType(args.src_data_type);
if (args.device_format == kOpFormat_NC1HWC0_C04) {
c0 = kCubeSize_C04;
}
auto c1 = DivCeil(c, c0);
auto hw = h * w;
auto chw = c * hw;
auto c1hwc0 = c1 * hw * c0;
auto wc0 = w * c0;
for (int64_t n_idx = 0; n_idx < n; n_idx++) {
int64_t n_head_addr = n_idx * c1hwc0;
for (int64_t c1_idx = 0; c1_idx < c1; c1_idx++) {
int64_t c1_head_addr = n_head_addr + c1_idx * hw * c0;
for (int64_t h_idx = 0; h_idx < h; h_idx++) {
int64_t h_head_addr = c1_head_addr + h_idx * wc0;
for (int64_t w_idx = 0; w_idx < w; w_idx++) {
int64_t w_head_addr = h_head_addr + w_idx * c0;
for (int64_t c0_idx = 0; c0_idx < c0; c0_idx++) {
int64_t dst_idx = c0_idx + w_head_addr;
int64_t c_idx = c0_idx + c1_idx * c0;
int64_t src_idx = n_idx * chw + c_idx * hw + h_idx * w + w_idx;
auto pad_zero = c_idx >= c;
SetData(size, pad_zero, src_idx, dst_idx, args, result);
}
}
}
}
}
return true;
}
bool FormatTransfer::NCHW_TO_NC1HWC04(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from nchw to Nc1hwc04.";
return NCHW_TO_NC1HWC0(args, result);
}
bool FormatTransfer::NCHW_TO_C1HWNCOC0(const FormatArgs &args, void *result) {
// trans nchw to c1hwncoc0
MS_LOG(DEBUG) << "Trans format from nchw to c1hwncoc0.";
MS_EXCEPTION_IF_NULL(result);
int64_t size = 0;
if (!CheckArgs(args, &size)) {
MS_LOG(ERROR) << "Check args failed.";
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
const int co_idx = 4;
const int c0_idx = 5;
auto c1 = args.device_shape[0];
auto co = args.device_shape[co_idx];
auto c0 = args.device_shape[c0_idx];
for (int64_t c1_i = 0; c1_i < c1; c1_i++) {
for (int64_t h_i = 0; h_i < h; h_i++) {
for (int64_t w_i = 0; w_i < w; w_i++) {
for (int64_t n_i = 0; n_i < n; n_i++) {
for (int64_t co_i = 0; co_i < co; co_i++) {
for (int64_t c0_i = 0; c0_i < c0; c0_i++) {
int64_t dst_idx = c1_i * h * w * n * co * c0 + h_i * w * n * co * c0 + w_i * n * co * c0 + n_i * co * c0 +
co_i * c0 + c0_i;
int64_t c_i = c0_i + c1_i * c0;
int64_t src_idx = n_i * c * h * w + c_i * h * w + h_i * w + w_i;
auto pad_zero = !(c_i < c && c0_i == co_i);
SetData(size, pad_zero, src_idx, dst_idx, args, result);
}
}
}
}
}
}
return true;
}
bool FormatTransfer::NCDHW_TO_NDC1HWC0(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans from ncdhw to ndc1hwc0";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNcdhw) {
MS_LOG(ERROR) << "Illegal host shape dim, expect dim: 5, but got " << args.host_shape.size();
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n = args.host_shape[N_ncdhw];
auto c = args.host_shape[C_ncdhw];
auto d = args.host_shape[D_ncdhw];
auto h = args.host_shape[H_ncdhw];
auto w = args.host_shape[W_ncdhw];
auto c0 = GetCubeSizeByType(args.src_data_type);
auto c1 = DivCeil(c, c0);
const int64_t cdhw = c * d * h * w;
const int64_t dhw = d * h * w;
const int64_t hw = h * w;
const int64_t dc1hwc0 = d * c1 * h * w * c0;
const int64_t c1hwc0 = c1 * h * w * c0;
const int64_t hwc0 = h * w * c0;
const int64_t wc0 = w * c0;
for (int64_t n_i = 0; n_i < n; n_i++) {
int64_t n_head = n_i * dc1hwc0;
for (int64_t d_i = 0; d_i < d; d_i++) {
int64_t d_head = n_head + d_i * c1hwc0;
for (int64_t c1_i = 0; c1_i < c1; c1_i++) {
int64_t c1_head = d_head + c1_i * hwc0;
for (int64_t h_i = 0; h_i < h; h_i++) {
int64_t h_head = c1_head + h_i * wc0;
for (int64_t w_i = 0; w_i < w; w_i++) {
int64_t w_head = h_head + w_i * c0;
for (int64_t c0_i = 0; c0_i < c0; c0_i++) {
int64_t dst_i = c0_i + w_head;
int64_t c_i = c0_i + c1_i * c0;
int64_t src_i = n_i * cdhw + c_i * dhw + d_i * hw + h_i * w + w_i;
auto pad_zero = c_i >= c;
SetData(size, pad_zero, src_i, dst_i, args, result);
}
}
}
}
}
}
return true;
}
bool FormatTransfer::NCDHW_TO_FRAC_Z3D(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans from ncdhw to frac_z_3d";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNcdhw) {
MS_LOG(ERROR) << "Illegal host shape dim, expect dim: 5, but got " << args.host_shape.size();
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n = args.host_shape[N_ncdhw];
auto c = args.host_shape[C_ncdhw];
auto d = args.host_shape[D_ncdhw];
auto h = args.host_shape[H_ncdhw];
auto w = args.host_shape[W_ncdhw];
auto n1n0 = DivCeil(n, kNiSize) * kNiSize;
auto c0 = GetCubeSizeByType(args.src_data_type);
auto c1 = DivCeil(c, c0);
auto hw = h * w;
auto dhw = d * hw;
auto cdhw = c * dhw;
auto n1n0c0 = n1n0 * c0;
auto wn1n0c0 = w * n1n0c0;
auto hwn1n0c0 = h * wn1n0c0;
auto c1hwn1n0c0 = c1 * hwn1n0c0;
for (int64_t d_i = 0; d_i < d; d_i++) {
for (int64_t c1_i = 0; c1_i < c1; c1_i++) {
for (int64_t h_i = 0; h_i < h; h_i++) {
for (int64_t w_i = 0; w_i < w; w_i++) {
for (int64_t n1n0_i = 0; n1n0_i < n1n0; n1n0_i++) {
for (int64_t c0_i = 0; c0_i < c0; c0_i++) {
auto dst_i = d_i * c1hwn1n0c0 + c1_i * hwn1n0c0 + h_i * wn1n0c0 + w_i * n1n0c0 + n1n0_i * c0 + c0_i;
// ncdhw
int64_t src_i = n1n0_i * cdhw + (c1_i * c0 + c0_i) * dhw + d_i * hw + h_i * w + w_i;
auto pad_zero = ((c1_i * c0 + c0_i) >= c) || (n1n0_i >= n);
SetData(size, pad_zero, src_i, dst_i, args, result);
}
}
}
}
}
}
return true;
}
bool FormatTransfer::NCHW_TO_FRAC_Z_WITH_GROPUS(const FormatArgs &args, void *result, bool to_device, int64_t groups) {
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Invalid host shape, host shape dims:" << args.host_shape.size() << ", expect dims:" << kNchwDims;
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto n_dim = args.host_shape[kN];
auto c_dim = args.host_shape[kC];
auto h_dim = args.host_shape[kH];
auto w_dim = args.host_shape[kW];
auto d_dim = 1;
auto cin_ori = c_dim;
if (groups <= 0) {
MS_LOG(EXCEPTION) << "The value of groups should be greater than 0, but got " << groups;
}
// cppcheck-suppress *
auto cout_ori = n_dim / groups;
if (cin_ori == 0 || cout_ori == 0) {
MS_LOG(ERROR) << "cin_ori, cout_ori must not equal to 0";
return false;
}
auto cube_k = GetCubeSizeByType(args.src_data_type);
auto e_mult = std::min(Lcm(Lcm(cin_ori, cube_k) / cin_ori, Lcm(cout_ori, kCubeSize) / cout_ori), groups);
if (e_mult == 0) {
MS_LOG(EXCEPTION) << "The value of e_mult should be greater than 0, but got " << e_mult;
}
auto cin_opt = DivCeil(e_mult * cin_ori, cube_k) * cube_k;
auto cout_opt = DivCeil(e_mult * cout_ori, kCubeSize) * kCubeSize;
// cppcheck-suppress *
auto c1_dim = cin_opt / cube_k;
auto dst_size =
to_device ? abstract::ShapeSize(args.device_shape) * size : abstract::ShapeSize(args.host_shape) * size;
if (dst_size == 0) {
return true;
}
auto ret = memset_s(result, dst_size, 0, dst_size);
if (ret != EOK) {
MS_LOG(ERROR) << "memset failed";
return false;
}
for (int64_t g = 0; g < groups; ++g) {
for (int64_t d = 0; d < d_dim; ++d) {
for (int64_t c = 0; c < c_dim; ++c) {
for (int64_t h = 0; h < h_dim; ++h) {
for (int64_t w = 0; w < w_dim; ++w) {
for (int64_t n = 0; n < cout_ori; ++n) {
int64_t e_val = g % e_mult;
int64_t dst_ci = e_val * cin_ori + c;
int64_t dst_co = e_val * cout_ori + n;
int64_t src_co = g * cout_ori + n;
int64_t temporary = dst_ci % cube_k;
int64_t dev_idx = (g / e_mult) * d_dim * c1_dim * h_dim * w_dim * cout_opt * cube_k +
d * c1_dim * h_dim * w_dim * cout_opt * cube_k +
(dst_ci / cube_k) * h_dim * w_dim * cout_opt * cube_k + h * w_dim * cout_opt * cube_k +
w * cout_opt * cube_k + dst_co * cube_k + temporary;
int64_t hst_idx =
src_co * c_dim * d_dim * h_dim * w_dim + c * d_dim * h_dim * w_dim + d * h_dim * w_dim + h * w_dim + w;
if (to_device) {
SetData(size, false, hst_idx, dev_idx, args, result);
} else {
SetData(size, false, dev_idx, hst_idx, args, result);
}
}
}
}
}
}
}
return true;
}
bool FormatTransfer::NC1HWC0_TO_NCHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from nc1h1wc0 to nchw";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Invalid host shape, host shape dims:" << args.host_shape.size() << ", expect dims:" << kNchwDims;
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
auto c1 = args.device_shape[1];
auto c0 = args.device_shape[4];
auto hw = h * w;
auto chw = c * hw;
auto wc0 = w * c0;
auto hwc0 = h * wc0;
auto c1hwc0 = c1 * hwc0;
for (int64_t n_idx = 0; n_idx < n; n_idx++) {
int64_t n_head_addr = n_idx * chw;
for (int64_t c_idx = 0; c_idx < c; c_idx++) {
int64_t c_head_addr = n_head_addr + c_idx * hw;
for (int64_t h_idx = 0; h_idx < h; h_idx++) {
int64_t h_head_addr = c_head_addr + h_idx * w;
for (int64_t w_idx = 0; w_idx < w; w_idx++) {
int64_t dst_idx = h_head_addr + w_idx;
int64_t c1_idx = c_idx / c0;
int64_t c0_idx = c_idx % c0;
int64_t src_idx = n_idx * c1hwc0 + c1_idx * hwc0 + h_idx * wc0 + w_idx * c0 + c0_idx;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
}
return true;
}
bool FormatTransfer::NC1HWC04_TO_NCHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from Nc1hwc04 to nchw.";
return NC1HWC0_TO_NCHW(args, result);
}
bool FormatTransfer::C1HWNCOC0_TO_NCHW(const FormatArgs &args, void *result) {
// trans c1hwncoc0 to nchw
MS_LOG(DEBUG) << "Trans format from c1hwncoc0 to nchw";
MS_EXCEPTION_IF_NULL(result);
int64_t size = 0;
if (!CheckArgs(args, &size)) {
MS_LOG(ERROR) << "Check args failed.";
return false;
}
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
const int co_idx = 4;
const int c0_idx = 5;
auto co = args.device_shape[co_idx];
auto c0 = args.device_shape[c0_idx];
auto cube_k = GetCubeSizeByType(args.src_data_type);
for (int64_t n_i = 0; n_i < n; n_i++) {
for (int64_t c_i = 0; c_i < c; c_i++) {
for (int64_t h_i = 0; h_i < h; h_i++) {
for (int64_t w_i = 0; w_i < w; w_i++) {
int64_t dst_idx = n_i * c * h * w + c_i * h * w + h_i * w + w_i;
int64_t c1_i = c_i / cube_k;
int64_t c0_i = c_i % cube_k;
int64_t co_i = c0_i;
int64_t src_idx =
c1_i * h * w * n * co * c0 + h_i * w * n * co * c0 + w_i * n * co * c0 + n_i * co * c0 + co_i * c0 + c0_i;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
}
return true;
}
bool FormatTransfer::FRAC_Z_TO_NCHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from frac_z to nchw";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNchwDims) {
MS_LOG(ERROR) << "Invalid host shape, host shape dims:" << args.host_shape.size() << ", expect dims:" << kNchwDims;
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n0 = args.device_shape.at(1);
auto ni = args.device_shape.at(2);
auto c0 = args.device_shape.at(3);
auto n = args.host_shape[kN];
auto c = args.host_shape[kC];
auto h = args.host_shape[kH];
auto w = args.host_shape[kW];
auto nc = ni * n0;
auto ncc0 = nc * c0;
auto wncc0 = w * ncc0;
auto hwncc0 = h * wncc0;
auto hw = h * w;
auto chw = c * hw;
for (int64_t n_idx = 0; n_idx < n; n_idx++) {
int64_t n_head_addr = n_idx * chw;
for (int64_t c_idx = 0; c_idx < c; c_idx++) {
int64_t c_head_addr = n_head_addr + c_idx * hw;
for (int64_t h_idx = 0; h_idx < h; h_idx++) {
int64_t h_head_addr = c_head_addr + h_idx * w;
for (int64_t w_idx = 0; w_idx < w; w_idx++) {
auto dst_idx = h_head_addr + w_idx;
auto c1_idx = c_idx / c0;
auto c0_idx = c_idx % c0;
auto nc_idx = n_idx;
auto src_idx = c1_idx * hwncc0 + h_idx * wncc0 + w_idx * ncc0 + nc_idx * c0 + c0_idx;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
}
return true;
}
bool FormatTransfer::FRAC_NZ_TO_NCHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans format from frac_nz to nchw";
MS_EXCEPTION_IF_NULL(result);
ShapeVector hw_shape;
if (!TransShapeToHW_NZ(args.host_shape, &hw_shape)) {
MS_LOG(ERROR) << "Trans shape failed..";
return false;
}
if (hw_shape.size() < kDim3 || args.device_shape.size() < kDim4) {
MS_LOG(ERROR) << "Invalid shape size.";
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto dst_size = abstract::ShapeSize(args.device_shape) * size;
if (dst_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << dst_size << ", device_size:" << args.device_size;
return false;
}
auto times = hw_shape.at(0);
auto h = hw_shape.at(1);
auto w = hw_shape.at(2);
auto hw = h * w;
auto shape_size = args.device_shape.size();
auto w1 = args.device_shape[shape_size - 4];
auto h1 = args.device_shape[shape_size - 3];
auto h0 = args.device_shape[shape_size - 2];
auto w0 = args.device_shape[shape_size - 1];
auto h1h0w0 = h1 * h0 * w0;
auto w1h1h0w0 = w1 * h1h0w0;
auto num_w1 = w / w0;
for (int64_t times_idx = 0; times_idx < times; times_idx++) {
auto times_head = times_idx * w1h1h0w0;
auto src_times_head = times_idx * hw;
for (int64_t h1h0_idx = 0; h1h0_idx < h; h1h0_idx++) {
auto h1h0_head = times_head + h1h0_idx * w0;
auto src_h_head = src_times_head + h1h0_idx * w;
for (int64_t w1_idx = 0; w1_idx < num_w1; w1_idx++) {
for (int64_t i = 0; i < w0; ++i) {
int64_t src_idx = h1h0_head + w1_idx * h1h0w0 + i;
int64_t dst_idx = src_h_head + w1_idx * w0 + i;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
auto w1_head = num_w1 * w0;
for (int64_t w0_idx = 0; w1_head + w0_idx < w; w0_idx++) {
auto src_w_idx = w1_head + w0_idx;
int64_t src_idx = h1h0_head + num_w1 * h1h0w0 + w0_idx;
int64_t dst_idx = src_h_head + src_w_idx;
SetData(size, false, src_idx, dst_idx, args, result);
}
}
}
return true;
}
bool FormatTransfer::FRAC_Z3D_TO_NCDHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans from frac_z_3d to ncdhw";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNcdhw) {
MS_LOG(ERROR) << "Illegal host shape dim, expect dim: 5, but got " << args.host_shape.size();
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n = args.host_shape[N_ncdhw];
auto c = args.host_shape[C_ncdhw];
auto d = args.host_shape[D_ncdhw];
auto h = args.host_shape[H_ncdhw];
auto w = args.host_shape[W_ncdhw];
const int kFZ3D_C0 = 3;
auto c0 = args.device_shape[kFZ3D_C0];
auto cube_k = GetCubeSizeByType(args.src_data_type);
auto c1 = DivCeil(c, cube_k);
auto n1n0 = DivCeil(n, kNiSize) * kNiSize;
auto n1n0c0 = n1n0 * c0;
auto wn1n0c0 = w * n1n0c0;
auto hwn1n0c0 = h * wn1n0c0;
auto c1hwn1n0c0 = c1 * hwn1n0c0;
auto hw = h * w;
auto dhw = d * hw;
auto cdhw = c * dhw;
for (int64_t n_i = 0; n_i < n; n_i++) {
int64_t n_head = n_i * cdhw;
for (int64_t c_i = 0; c_i < c; c_i++) {
int64_t c_head = n_head + c_i * dhw;
for (int64_t d_i = 0; d_i < d; d_i++) {
int64_t d_head = c_head + d_i * hw;
for (int64_t h_i = 0; h_i < h; h_i++) {
int64_t h_head = d_head + h_i * w;
for (int64_t w_i = 0; w_i < w; w_i++) {
int64_t dst_i = h_head + w_i;
int64_t c1_i = c_i / c0;
int64_t c0_i = c_i % c0;
int64_t nc_i = n_i;
int64_t src_i = d_i * c1hwn1n0c0 + c1_i * hwn1n0c0 + h_i * wn1n0c0 + w_i * n1n0c0 + nc_i * c0 + c0_i;
SetData(size, false, src_i, dst_i, args, result);
}
}
}
}
}
return true;
}
bool FormatTransfer::NDC1HWC0_TO_NCDHW(const FormatArgs &args, void *result) {
MS_LOG(DEBUG) << "Trans from ndc1hwc0 to ncdhw";
MS_EXCEPTION_IF_NULL(result);
if (args.host_shape.size() != kNcdhw) {
MS_LOG(ERROR) << "Illegal host shape dim, expect dim: 5, but got " << args.host_shape.size();
return false;
}
auto size = SizeToLong(abstract::TypeIdSize(args.src_data_type));
if (size < 1) {
MS_LOG(ERROR) << "Illegal dtype: " << args.src_data_type;
return false;
}
auto total_size = abstract::ShapeSize(args.device_shape) * size;
if (total_size != SizeToLong(args.device_size)) {
MS_LOG(ERROR) << "Illegal total data size, total_size:" << total_size << ", device_size:" << args.device_size;
return false;
}
auto n = args.host_shape[N_ncdhw];
auto c = args.host_shape[C_ncdhw];
auto d = args.host_shape[D_ncdhw];
auto h = args.host_shape[H_ncdhw];
auto w = args.host_shape[W_ncdhw];
auto c1 = args.device_shape[C1_ndc1hwc0];
auto c0 = args.device_shape[C0_ndc1hwc0];
const int64_t cdhw = c * d * h * w;
const int64_t dhw = d * h * w;
const int64_t hw = h * w;
const int64_t dc1hwc0 = d * c1 * h * w * c0;
const int64_t c1hwc0 = c1 * h * w * c0;
const int64_t hwc0 = h * w * c0;
const int64_t wc0 = w * c0;
for (int64_t n_i = 0; n_i < n; n_i++) {
int64_t n_head = n_i * cdhw;
for (int64_t c_i = 0; c_i < c; c_i++) {
int64_t c_head = n_head + c_i * dhw;
for (int64_t d_i = 0; d_i < d; d_i++) {
int64_t d_head = c_head + d_i * hw;
for (int64_t h_i = 0; h_i < h; h_i++) {
int64_t h_head = d_head + h_i * w;
for (int64_t w_i = 0; w_i < w; w_i++) {
int64_t dst_i = h_head + w_i;
int64_t c1_i = c_i / c0;
int64_t c0_i = c_i % c0;
auto src_idx = n_i * dc1hwc0 + d_i * c1hwc0 + c1_i * hwc0 + h_i * wc0 + w_i * c0 + c0_i;
SetData(size, false, src_idx, dst_i, args, result);
}
}
}
}
}
return true;
}
bool FormatTransfer::FRAC_Z_TO_NCHW_WITH_GROUPS(const FormatArgs &args, void *result, int64_t groups) {
MS_LOG(DEBUG) << "Trans format from frac_z to nchw with groups=" << groups;
return NCHW_TO_FRAC_Z_WITH_GROPUS(args, result, false, groups);
}
// ######################## RANGE TRANS ########################
RangePair ShapeRangeTransfer::GetRealRange(const RangePair &ori_range, const std::string &format, const TypeId &type) {
const std::set<std::string> no_need_change = {kOpFormat_ND, kOpFormat_DEFAULT, kOpFormat_NCHW, kOpFormat_NCDHW};
using RangeTransfer = std::function<RangePair(const RangePair &, const TypeId &)>;
const std::map<std::string, RangeTransfer> format_range_map = {{kOpFormat_NHWC, NHWCRange},
{kOpFormat_HWCN, HWCNRange},
{kOpFormat_FRAC_Z, FRAC_ZRange},
{kOpFormat_NC1HWC0, NC1HWC0Range},
{kOpFormat_NDC1HWC0, NDC1HWC0Range},
{kOpFormat_C1HWNCoC0, C1HWNCOC0Range},
{kOpFormat_NC1HWC0_C04, NC1HWC04Range},
{kOpFormat_FRACTAL_Z_3D, FRAC_Z_3DRange},
{kOpFormat_FRACTAL_Z_C04, FRAC_ZC04Range}};
if (no_need_change.find(format) != no_need_change.end()) {
return ori_range;
}
// kOpFormat_FRACTAL_ZN_LSTM, kOpFormat_FRAC_NZ no need pad range
if (format == kOpFormat_FRACTAL_ZN_LSTM) {
return FRAC_ZN_LSTMRange(ori_range, type);
}
if (format == kOpFormat_FRAC_NZ) {
return FRAC_NZRange(ori_range, type);
}
auto temp_range = ori_range;
if (ori_range.size() < kNchwDims && k3DFormatSet.find(format) == k3DFormatSet.end()) {
MS_LOG(DEBUG) << "A special format:" << format << " with a range size less than 4, so padding the range firstly";
temp_range = PaddingRangeTo4D(ori_range);
}
if (ori_range.size() < kNcdhw && k3DFormatSet.find(format) != k3DFormatSet.end()) {
MS_LOG(DEBUG) << "A special format:" << format << " with a range size less than 5, so padding the range firstly";
temp_range = PaddingRangeTo5D(ori_range);
}
auto iter = format_range_map.find(format);
if (iter == format_range_map.end()) {
MS_LOG(INFO) << "Can not find a supported format: " << format << ", using default range";
return ori_range;
}
return iter->second(temp_range, type);
}
RangePair ShapeRangeTransfer::NHWCRange(const RangePair &ori_range, const TypeId &) {
RangePair dst_range;
dst_range.push_back(ori_range[kN]);
dst_range.push_back(ori_range[kH]);
dst_range.push_back(ori_range[kW]);
dst_range.push_back(ori_range[kC]);
return dst_range;
}
RangePair ShapeRangeTransfer::HWCNRange(const RangePair &ori_range, const TypeId &) {
RangePair dst_range;
dst_range.push_back(ori_range[kH]);
dst_range.push_back(ori_range[kW]);
dst_range.push_back(ori_range[kC]);
dst_range.push_back(ori_range[kN]);
return dst_range;
}
RangePair ShapeRangeTransfer::NC1HWC04Range(const RangePair &ori_range, const TypeId &) {
RangePair dst_range;
const std::pair<int64_t, int64_t> c0 = {k4, k4};
const std::pair<int64_t, int64_t> c1 = {(ori_range[kC].first + k4 - 1) / k4, (ori_range[kC].second + k4 - 1) / k4};
dst_range.push_back(ori_range[kN]);
dst_range.push_back(c1);
dst_range.push_back(ori_range[kH]);
dst_range.push_back(ori_range[kW]);
dst_range.push_back(c0);
return dst_range;
}
RangePair ShapeRangeTransfer::FRAC_ZC04Range(const RangePair &ori_range, const TypeId &) {
RangePair dst_range;
const std::pair<int64_t, int64_t> c0 = {k4, k4};
const std::pair<int64_t, int64_t> c16 = {kNiSize, kNiSize};
const std::pair<int64_t, int64_t> first_dim = {
(c0.first * ori_range[kH].first * ori_range[kW].first + kNiSize - 1) / kNiSize,
(c0.second * ori_range[kH].second * ori_range[kW].second + kNiSize - 1) / kNiSize};
const std::pair<int64_t, int64_t> no = {(ori_range[kN].first + kNiSize - 1) / kNiSize,
(ori_range[kN].second + kNiSize - 1) / kNiSize};
dst_range.push_back(first_dim);
dst_range.push_back(no);
dst_range.push_back(c16);
dst_range.push_back(c16);
return dst_range;
}
RangePair ShapeRangeTransfer::FRAC_ZRange(const RangePair &ori_range, const TypeId &type) {
RangePair dst_range;
auto cube = GetCubeSizeByType(type);
const std::pair<int64_t, int64_t> c0 = {cube, cube};
const std::pair<int64_t, int64_t> cout16 = {((ori_range[kN].first + kNiSize - 1) / kNiSize) * kNiSize,
((ori_range[kN].second + kNiSize - 1) / kNiSize) * kNiSize};
const std::pair<int64_t, int64_t> cin16 = {((ori_range[kC].first + cube - 1) / cube) * cube,
((ori_range[kC].second + cube - 1) / cube) * cube};
const std::pair<int64_t, int64_t> r0 = {ori_range[kH].first * ori_range[kW].first * cin16.first / cube,
ori_range[kH].second * ori_range[kW].second * cin16.second / cube};
const std::pair<int64_t, int64_t> r1 = {cout16.first / kNiSize, cout16.second / kNiSize};
dst_range.push_back(r0);
dst_range.push_back(r1);
dst_range.push_back({kNiSize, kNiSize});
dst_range.push_back(c0);
return dst_range;
}
RangePair ShapeRangeTransfer::FRAC_NZRange(const RangePair &ori_range, const TypeId &type) {
RangePair dst_range;
auto cube = GetCubeSizeByType(type);
auto ori_size = ori_range.size();
if (ori_size < kDims2) {
MS_LOG(EXCEPTION) << "Format FracNZ can not support range size: " << ori_size;
} else {
(void)std::copy(ori_range.begin(), ori_range.end() - kDims2, std::back_inserter(dst_range));
}
const std::pair<int64_t, int64_t> c0 = {cube, cube};
const std::pair<int64_t, int64_t> w1 = {(ori_range[ori_size - 1].first - 1) / cube + 1,
(ori_range[ori_size - 1].second - 1) / cube + 1};
const std::pair<int64_t, int64_t> h1 = {(ori_range[ori_size - kDims2].first - 1) / kNiSize + 1,
(ori_range[ori_size - kDims2].second - 1) / kNiSize + 1};
dst_range.push_back(w1);
dst_range.push_back(h1);
dst_range.push_back({kNiSize, kNiSize});
dst_range.push_back(c0);
return dst_range;
}
RangePair ShapeRangeTransfer::NC1HWC0Range(const RangePair &ori_range, const TypeId &type) {
RangePair dst_range;
auto cube = GetCubeSizeByType(type);
const std::pair<int64_t, int64_t> c0 = {cube, cube};
const std::pair<int64_t, int64_t> c1 = {(ori_range[kC].first + cube - 1) / cube,
(ori_range[kC].second + cube - 1) / cube};
dst_range.push_back(ori_range[kN]);
dst_range.push_back(c1);
dst_range.push_back(ori_range[kH]);
dst_range.push_back(ori_range[kW]);
dst_range.push_back(c0);
return dst_range;
}
RangePair ShapeRangeTransfer::FRAC_ZN_LSTMRange(const RangePair &ori_range, const TypeId &) {
RangePair dst_range;
const std::pair<int64_t, int64_t> c0 = {k4, k4};
const std::pair<int64_t, int64_t> c16 = {k4, k4};
const std::pair<int64_t, int64_t> h = {ori_range[kN].first / c0.first, ori_range[kN].second / c0.second};
const std::pair<int64_t, int64_t> i = {ori_range[kC].first - h.first, ori_range[kC].second - h.second};
const std::pair<int64_t, int64_t> first_dim = {
(i.first + kCube16 - 1) / kCube16 + (h.first + kCube16 - 1) / kCube16,
(i.second + kCube16 - 1) / kCube16 + (h.second + kCube16 - 1) / kCube16};
const std::pair<int64_t, int64_t> second = {c0.first * ((h.first + kCube16 - 1) / kCube16),
c0.second * ((h.second + kCube16 - 1) / kCube16)};
dst_range.push_back(first_dim);
dst_range.push_back(second);
dst_range.push_back(c16);
dst_range.push_back(c16);
return dst_range;
}
RangePair ShapeRangeTransfer::NDC1HWC0Range(const RangePair &ori_range, const TypeId &type) {
RangePair dst_range;
auto cube = GetCubeSizeByType(type);
const std::pair<int64_t, int64_t> c0 = {cube, cube};
const std::pair<int64_t, int64_t> c1 = {(ori_range[C_ncdhw].first + cube - 1) / cube,
(ori_range[C_ncdhw].second + cube - 1) / cube};
dst_range.push_back(ori_range[N_ncdhw]);
dst_range.push_back(ori_range[D_ncdhw]);
dst_range.push_back(c1);
dst_range.push_back(ori_range[H_ncdhw]);
dst_range.push_back(ori_range[W_ncdhw]);
dst_range.push_back(c0);
return dst_range;
}
RangePair ShapeRangeTransfer::C1HWNCOC0Range(const RangePair &ori_range, const TypeId &type) {
RangePair dst_range;
auto cube = GetCubeSizeByType(type);
const std::pair<int64_t, int64_t> c0 = {cube, cube};
const std::pair<int64_t, int64_t> r1 = {(ori_range[kC].first - 1) / cube + 1, (ori_range[kC].second - 1) / cube + 1};
dst_range.push_back(r1);
dst_range.push_back(ori_range[kH]);
dst_range.push_back(ori_range[kW]);
dst_range.push_back(ori_range[kN]);
dst_range.push_back(c0);
dst_range.push_back(c0);
return dst_range;
}
RangePair ShapeRangeTransfer::FRAC_Z_3DRange(const RangePair &ori_range, const TypeId &type) {
RangePair dst_range;
auto cube = GetCubeSizeByType(type);
const std::pair<int64_t, int64_t> c0 = {cube, cube};
const std::pair<int64_t, int64_t> c1 = {(ori_range[C_ncdhw].first + cube - 1) / cube,
(ori_range[C_ncdhw].second + cube - 1) / cube};
const std::pair<int64_t, int64_t> n1 = {(ori_range[N_ncdhw].first + kNiSize - 1) / kNiSize,
(ori_range[N_ncdhw].second + kNiSize - 1) / kNiSize};
const int64_t r1_0 = ori_range[D_ncdhw].first * c1.first * ori_range[H_ncdhw].first * ori_range[W_ncdhw].first;
const int64_t r1_1 = ori_range[D_ncdhw].second * c1.second * ori_range[H_ncdhw].second * ori_range[W_ncdhw].second;
const std::pair<int64_t, int64_t> r1 = {r1_0, r1_1};
dst_range.push_back(r1);
dst_range.push_back(n1);
dst_range.push_back(c1);
dst_range.push_back(c0);
return dst_range;
}
} // namespace trans
} // namespace mindspore
Reference in New Issue View Git Blame Copy Permalink