p89165320
/
InfiniTensor
forked from jiuyuan/InfiniTensor
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

add softmax/element_wise kernel

This commit is contained in:
OdinaryWord 2024-01-26 15:40:21 +08:00
parent c970c93ba1
commit f6176124ec
19 changed files with 736 additions and 328 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
examples/NNmodel

@ -1 +1 @@
Subproject commit b896cec2dba5b8522b141ac4f89eb43074ee1b98
Subproject commit 51d3105277f3774ed31c02ed4cd11fa92925af77

32
include/ascend/ascend_runtime.h
View File

@ -18,28 +18,34 @@ namespace infini {
class ASCENDRuntimeObj : public RuntimeObj {
private:
aclrtContext aclnn;
aclrtContext context;
aclrtStream stream;
ASCENDPtr workspace;
ASCENDPtr workspace = nullptr;
size_t workspaceSize;
public:
ASCENDRuntimeObj(int deviceId = 0) : RuntimeObj(Device::ASCEND, deviceId) {
// #ifndef _ACL_INIT
// #define _ACL_INIT
// aclInit(nullptr);
// // auto ret_init =
// // CHECK_RET(ret == ACL_SUCCESS,
// // LOG_PRINT("aclInit failed. ERROR: %d\n",
// ret));
// #endif
auto ret = aclrtSetDevice(deviceId);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret));
ret = aclrtCreateContext(&aclnn, deviceId);
ret = aclrtCreateContext(&context, deviceId);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret));
ret = aclrtSetCurrentContext(aclnn);
ret = aclrtSetCurrentContext(context);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret));
ret = aclrtCreateStream(&stream);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret));
ret = aclInit(nullptr);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclInit failed. ERROR: %d\n", ret));
// 10GB for Longformer
// size_t longformerNum = 3lu * (1 << 30);
workspaceSize = 3ll << 30; // 3 GB
@ -50,9 +56,9 @@ class ASCENDRuntimeObj : public RuntimeObj {
virtual ~ASCENDRuntimeObj() {
dealloc(workspace);
aclrtDestroyStream(stream);
aclrtDestroyContext(aclnn);
aclrtDestroyContext(context);
aclrtResetDevice(deviceId);
aclFinalize();
// aclFinalize();
}
string toString() const override;
@ -68,7 +74,7 @@ class ASCENDRuntimeObj : public RuntimeObj {
return ptr;
}
void dealloc(void *ptr) override { aclrtFree(ptr); }
aclrtContext *ASCENDHandle() const { return nullptr; }
aclrtStream ASCENDHandle() const { return stream; }
ASCENDPtr getWorkspace(size_t size) const {
IT_ASSERT(size <= workspaceSize);
return workspace;
@ -76,19 +82,19 @@ class ASCENDRuntimeObj : public RuntimeObj {
void copyBlobFromCPU(void *dst, const void *src,
size_t bytes) const override {
aclrtMemcpy(dst, 1024 * 1024 * 1024, const_cast<void *>(src), bytes,
aclrtMemcpy(dst, bytes, const_cast<void *>(src), bytes,
ACL_MEMCPY_HOST_TO_DEVICE);
}
void copyBlobToCPU(void *dst, const void *src,
size_t bytes) const override {
aclrtMemcpy(dst, 1024 * 1024 * 1024, const_cast<void *>(src), bytes,
aclrtMemcpy(dst, bytes, const_cast<void *>(src), bytes,
ACL_MEMCPY_DEVICE_TO_HOST);
}
void copyBlobInsideRuntime(void *dst, const void *src,
size_t bytes) const override {
aclrtMemcpy(dst, 1024 * 1024 * 1024, const_cast<void *>(src), bytes,
aclrtMemcpy(dst, bytes, const_cast<void *>(src), bytes,
ACL_MEMCPY_DEVICE_TO_DEVICE);
}

3
src/ascend/ascend_runtime.cc
View File

@ -13,8 +13,7 @@ void ASCENDRuntimeObj::runWithoutSync(const Graph &graph, bool tune = false,
std::map<OpType, int> opCnt;
for (auto &op : graph->getOperators()) {
// HACK: set correct data type
auto kernelAttrs =
KernelAttrs{device, op->getOpType().underlying(), op->getDType()};
auto kernelAttrs = KernelAttrs{device, op->getOpType().underlying()};
Kernel *kernel = kernelRegistry.getKernel(kernelAttrs);
auto perfKey = PerfEngine::Key{kernelAttrs, op->getOpPerfKey()};
auto perfData = perfEngine.getPerfData(perfKey);

72
src/kernels/ascend/batch_norm.cc
View File

@ -5,10 +5,8 @@
namespace infini {
class BatchNormAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<BatchNormObj>(_op);
@ -35,36 +33,31 @@ class BatchNormAclnn : public ASCENDKernelWithoutConfig {
std::vector<int64_t> outputDim = MycastTo64(outD);
std::vector<int64_t> outputStride = MycastTo64(outS);
//std::vector<int64_t> inputDim(in.size(), 1);
//for (size_t i = 0; i < a.size(); ++i) {
// inputDim[i] = int64_t(in[i]);
//}
//std::vector<int64_t> inputStride(inS.size(), 1);
//for (size_t i = 0; i < inS.size(); ++i) {
// inputStride[i] = int64_t(inS[i]);
//}
auto inputTensor = aclCreateTensor(
inputDim.data(), inputDim.size(), ACL_FLOAT, inputStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, inputDim.data(), inputDim.size(), inData);
auto outputTensor = aclCreateTensor(
outputDim.data(), outputDim.size(), ACL_FLOAT, outputStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, outputDim.data(), outputDim.size(), outData);
auto inputTensor =
aclCreateTensor(inputDim.data(), inputDim.size(), ACL_FLOAT,
inputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
inputDim.data(), inputDim.size(), inData);
auto outputTensor =
aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
outputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
outputDim.data(), outputDim.size(), outData);
auto meanTensor = aclCreateTensor(
paraDim.data(), paraDim.size(), ACL_FLOAT, paraStride.data(), 0,
aclFormat::ACL_FORMAT_ND, paraDim.data(), paraDim.size(), meanData);
auto varTensor = aclCreateTensor(
paraDim.data(), paraDim.size(), ACL_FLOAT, paraStride.data(), 0,
aclFormat::ACL_FORMAT_ND, paraDim.data(), paraDim.size(), varData);
auto scaleTensor = aclCreateTensor(
paraDim.data(), paraDim.size(), ACL_FLOAT, paraStride.data(), 0,
aclFormat::ACL_FORMAT_ND, paraDim.data(), paraDim.size(), scaleData);
auto scaleTensor =
aclCreateTensor(paraDim.data(), paraDim.size(), ACL_FLOAT,
paraStride.data(), 0, aclFormat::ACL_FORMAT_ND,
paraDim.data(), paraDim.size(), scaleData);
auto biasTensor = aclCreateTensor(
paraDim.data(), paraDim.size(), ACL_FLOAT, paraStride.data(), 0,
aclFormat::ACL_FORMAT_ND, paraDim.data(), paraDim.size(), biasData);
auto savemeanTensor = aclCreateTensor(
paraDim.data(), paraDim.size(), ACL_FLOAT, paraStride.data(), 0,
aclFormat::ACL_FORMAT_ND, paraDim.data(), paraDim.size(), scaleData);
auto savemeanTensor =
aclCreateTensor(paraDim.data(), paraDim.size(), ACL_FLOAT,
paraStride.data(), 0, aclFormat::ACL_FORMAT_ND,
paraDim.data(), paraDim.size(), scaleData);
auto saveinvstdTensor = aclCreateTensor(
paraDim.data(), paraDim.size(), ACL_FLOAT, paraStride.data(), 0,
aclFormat::ACL_FORMAT_ND, paraDim.data(), paraDim.size(), biasData);
@ -72,38 +65,35 @@ class BatchNormAclnn : public ASCENDKernelWithoutConfig {
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret =
aclnnBatchNormGetWorkspaceSize(inputTensor, scaleTensor, biasTensor, meanTensor, varTensor, false, op->getMomentum(), op->getEps(), outputTensor, savemeanTensor, saveinvstdTensor, &workspaceSize, &executor);
auto ret = aclnnBatchNormGetWorkspaceSize(
inputTensor, scaleTensor, biasTensor, meanTensor, varTensor, false,
op->getMomentum(), op->getEps(), outputTensor, savemeanTensor,
saveinvstdTensor, &workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize,
ACL_MEM_MALLOC_HUGE_FIRST);
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnBatchNorm(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
aclDestroyTensor(inputTensor);
aclDestroyTensor(outputTensor);
aclDestroyTensor(meanTensor);
aclDestroyTensor(varTensor);
aclDestroyTensor(scaleTensor);
aclDestroyTensor(biasTensor);
aclDestroyTensor(savemeanTensor);
aclDestroyTensor(saveinvstdTensor);
// aclDestroyTensor(inputTensor);
// aclDestroyTensor(outputTensor);
// aclDestroyTensor(meanTensor);
// aclDestroyTensor(varTensor);
// aclDestroyTensor(scaleTensor);
// aclDestroyTensor(biasTensor);
// aclDestroyTensor(savemeanTensor);
// aclDestroyTensor(saveinvstdTensor);
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::BatchNormalization, DataType::Float32, BatchNormAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::BatchNormalization, BatchNormAclnn,
"batchnorm_ASCEND_float");
}; // namespace infini

83
src/kernels/ascend/concat.cc
View File

@ -5,56 +5,33 @@
namespace infini {
class ConcatAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<ConcatObj>(_op);
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
int dim = op->getDim();
//int num = op->numInputs();
int num = op->numInputs();
std::vector<aclTensor *> inputsData{};
std::vector<aclTensor*> inputsData{};
auto inD0 = op->getInputs(0)->getDims();
auto inS0 = op->getInputs(0)->getStride();
std::vector<int64_t> inputDim0 = MycastTo64(inD0);
std::vector<int64_t> inputStride0 = MycastTo64(inS0);
for (int i = 0; i < num; ++i) {
auto inD = op->getInputs(i)->getDims();
auto inS = op->getInputs(i)->getStride();
std::vector<int64_t> inputDim = MycastTo64(inD);
std::vector<int64_t> inputStride = MycastTo64(inS);
void *const inData0 = (op->getInputs(0)->getRawDataPtr<void *>());
auto tmpTensor0 = aclCreateTensor(
inputDim0.data(), inputDim0.size(), ACL_FLOAT, inputStride0.data(), 0,
aclFormat::ACL_FORMAT_ND, inputDim0.data(), inputDim0.size(), inData0);
void *const inData = (op->getInputs(i)->getRawDataPtr<void *>());
auto tmpTensor =
aclCreateTensor(inputDim.data(), inputDim.size(), ACL_FLOAT,
inputStride.data(), 0, aclFormat::ACL_FORMAT_ND,
inputDim.data(), inputDim.size(), inData);
inputsData.push_back(tmpTensor0);
auto inD = op->getInputs(1)->getDims();
auto inS = op->getInputs(1)->getStride();
std::vector<int64_t> inputDim = MycastTo64(inD);
std::vector<int64_t> inputStride = MycastTo64(inS);
void *const inData = (op->getInputs(1)->getRawDataPtr<void *>());
auto tmpTensor = aclCreateTensor(
inputDim.data(), inputDim.size(), ACL_FLOAT, inputStride.data(), 0,
aclFormat::ACL_FORMAT_ND, inputDim.data(), inputDim.size(), inData);
inputsData.push_back(tmpTensor);
//for (int i = 0; i < num; ++i) {
// auto inD = op->getInputs(i)->getDims();
// auto inS = op->getInputs(i)->getStride();
// std::vector<int64_t> inputDim = MycastTo64(inD);
// std::vector<int64_t> inputStride = MycastTo64(inS);
// void *const inData = (op->getInputs(i)->getRawDataPtr<void *>());
// auto tmpTensor = aclCreateTensor(
// inputDim.data(), inputDim.size(), ACL_FLOAT, inputStride.data(), 0,
// aclFormat::ACL_FORMAT_ND, inputDim.data(), inputDim.size(), inData);
// inputsData.push_back(tmpTensor);
//}
aclTensorList* tensorList = aclCreateTensorList(inputsData.data(), inputsData.size());
inputsData.push_back(tmpTensor);
}
aclTensorList *tensorList =
aclCreateTensorList(inputsData.data(), inputsData.size());
void *const outData = (op->getOutput()->getRawDataPtr<void *>());
auto outD = op->getOutput()->getDims();
@ -62,39 +39,35 @@ class ConcatAclnn : public ASCENDKernelWithoutConfig {
std::vector<int64_t> outputDim = MycastTo64(outD);
std::vector<int64_t> outputStride = MycastTo64(outS);
auto outputTensor = aclCreateTensor(
outputDim.data(), outputDim.size(), ACL_FLOAT, outputStride.data(), 0,
aclFormat::ACL_FORMAT_ND, outputDim.data(), outputDim.size(), outData);
auto outputTensor =
aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
outputStride.data(), 0, aclFormat::ACL_FORMAT_ND,
outputDim.data(), outputDim.size(), outData);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret =
aclnnCatGetWorkspaceSize(tensorList, int64_t(dim), outputTensor, &workspaceSize, &executor);
auto ret = aclnnCatGetWorkspaceSize(
tensorList, int64_t(dim), outputTensor, &workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize,
ACL_MEM_MALLOC_HUGE_FIRST);
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnCat(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
aclDestroyTensorList(tensorList);
aclDestroyTensor(outputTensor);
// aclDestroyTensorList(tensorList);
// aclDestroyTensor(outputTensor);
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::Concat, DataType::Float32, ConcatAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::Concat, ConcatAclnn,
"concat_ASCEND_float");
}; // namespace infini

67
src/kernels/ascend/conv.cc
View File

@ -5,7 +5,6 @@
namespace infini {
class ConvAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
@ -14,20 +13,23 @@ class ConvAclnn : public ASCENDKernelWithoutConfig {
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
const auto [ph, pw, sh, sw, dh, dw] = op->getPadStrideDilation();
//const auto [n, c, h, w, f, r, s] = op->getNCHWFRS();
//const int cpg = op->getChannelPerGroup();
//const int g = c / cpg;
// const auto [n, c, h, w, f, r, s] = op->getNCHWFRS();
// const int cpg = op->getChannelPerGroup();
// const int g = c / cpg;
std::vector<int64_t> pads = {ph, pw};
//std::vector<int64_t> ksize = {r, s};
// std::vector<int64_t> ksize = {r, s};
std::vector<int64_t> stride = {sh, sw};
std::vector<int64_t> dilation = {dh, dw};
std::vector<int64_t> outputPadding = {sh-1, sw-1};
std::vector<int64_t> outputPadding = {sh - 1, sw - 1};
aclIntArray *convpads = aclCreateIntArray(pads.data(), pads.size());
aclIntArray *convstride = aclCreateIntArray(stride.data(), stride.size());
aclIntArray *convdilation = aclCreateIntArray(dilation.data(), dilation.size());
aclIntArray *convOutputpadding = aclCreateIntArray(outputPadding.data(), outputPadding.size());
aclIntArray *convpads = aclCreateIntArray(pads.data(), pads.size());
aclIntArray *convstride =
aclCreateIntArray(stride.data(), stride.size());
aclIntArray *convdilation =
aclCreateIntArray(dilation.data(), dilation.size());
aclIntArray *convOutputpadding =
aclCreateIntArray(outputPadding.data(), outputPadding.size());
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
@ -47,46 +49,45 @@ class ConvAclnn : public ASCENDKernelWithoutConfig {
std::vector<int64_t> outputDim = MycastTo64(outD);
std::vector<int64_t> outputStride = MycastTo64(outS);
auto inputTensor = aclCreateTensor(
inputDim.data(), inputDim.size(), ACL_FLOAT, inputStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, inputDim.data(), inputDim.size(), aData);
auto weightTensor = aclCreateTensor(
weightDim.data(), weightDim.size(), ACL_FLOAT, weightStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, weightDim.data(), weightDim.size(), bData);
auto outputTensor = aclCreateTensor(
outputDim.data(), outputDim.size(), ACL_FLOAT, outputStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, outputDim.data(), outputDim.size(), cData);
auto inputTensor =
aclCreateTensor(inputDim.data(), inputDim.size(), ACL_FLOAT,
inputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
inputDim.data(), inputDim.size(), aData);
auto weightTensor =
aclCreateTensor(weightDim.data(), weightDim.size(), ACL_FLOAT,
weightStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
weightDim.data(), weightDim.size(), bData);
auto outputTensor =
aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
outputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
outputDim.data(), outputDim.size(), cData);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret =
aclnnConvolutionGetWorkspaceSize(inputTensor, weightTensor, nullptr, convstride, convpads, convdilation, false, convOutputpadding, 1, outputTensor, 1, &workspaceSize, &executor);
auto ret = aclnnConvolutionGetWorkspaceSize(
inputTensor, weightTensor, nullptr, convstride, convpads,
convdilation, false, convOutputpadding, 1, outputTensor, 1,
&workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize,
ACL_MEM_MALLOC_HUGE_FIRST);
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnConvolution(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
aclDestroyTensor(inputTensor);
aclDestroyTensor(weightTensor);
aclDestroyTensor(outputTensor);
// aclDestroyTensor(inputTensor);
// aclDestroyTensor(weightTensor);
// aclDestroyTensor(outputTensor);
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::Conv, DataType::Float32, ConvAclnn,
"conv_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Conv, ConvAclnn, "conv_ASCEND_float");
}; // namespace infini

278
src/kernels/ascend/element_wise.cc Normal file
View File

@ -0,0 +1,278 @@
#include "operators/element_wise.h"
#include "aclnnop/level2/aclnn_add.h"
#include "aclnnop/level2/aclnn_div.h"
#include "aclnnop/level2/aclnn_mul.h"
#include "aclnnop/level2/aclnn_pow_tensor_tensor.h"
#include "aclnnop/level2/aclnn_sub.h"
#include "ascend/ascend_kernel_without_config.h"
#include "ascend/ascend_runtime.h"
namespace infini {
/*
class PowAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<ElementWiseObj>(_op);
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
void *const cData = (op->getOutput()->getRawDataPtr<void *>());
auto a = op->getInputs(0)->getDims();
auto aS = op->getInputs(0)->getStride();
auto b = op->getInputs(1)->getDims();
auto bS = op->getInputs(1)->getStride();
auto c = op->getInputs(0)->getDims();
auto cS = op->getInputs(0)->getStride();
std::vector<int64_t> aDim = MycastTo64(a);
std::vector<int64_t> aStride = MycastTo64(aS);
std::vector<int64_t> bDim = MycastTo64(b);
std::vector<int64_t> bStride = MycastTo64(bS);
std::vector<int64_t> cDim = MycastTo64(c);
std::vector<int64_t> cStride = MycastTo64(cS);
auto inputA = aclCreateTensor(
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0,
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData);
auto inputB = aclCreateTensor(
bDim.data(), bDim.size(), ACL_FLOAT, bStride.data(), 0,
aclFormat::ACL_FORMAT_ND, bDim.data(), bDim.size(), bData);
auto output = aclCreateTensor(
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0,
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret = aclnnPowTensorTensorGetWorkspaceSize(
inputA, inputB, output, &workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnPowTensorTensor(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclDestroyTensor(inputA);
ret = aclDestroyTensor(inputB);
ret = aclDestroyTensor(output);
return;
}
};
*/
#define DEFINE_ELEMENT_WISE_Aclnn(prefix) \
class prefix##Aclnn : public ASCENDKernelWithoutConfig { \
void compute(const Operator &_op, \
const RuntimeObj *_context) const override { \
auto op = as<ElementWiseObj>(_op); \
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context); \
\
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>()); \
void *const bData = (op->getInputs(1)->getRawDataPtr<void *>()); \
void *const cData = (op->getOutput()->getRawDataPtr<void *>()); \
\
auto a = op->getInputs(0) -> getDims(); \
auto aS = op->getInputs(0) -> getStride(); \
auto b = op->getInputs(1) -> getDims(); \
auto bS = op->getInputs(1) -> getStride(); \
auto c = op->getInputs(0) -> getDims(); \
auto cS = op->getInputs(0) -> getStride(); \
\
std::vector<int64_t> aDim = MycastTo64(a); \
std::vector<int64_t> aStride = MycastTo64(aS); \
std::vector<int64_t> bDim = MycastTo64(b); \
std::vector<int64_t> bStride = MycastTo64(bS); \
std::vector<int64_t> cDim = MycastTo64(c); \
std::vector<int64_t> cStride = MycastTo64(cS); \
\
auto inputA = aclCreateTensor( \
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData); \
auto inputB = aclCreateTensor( \
bDim.data(), bDim.size(), ACL_FLOAT, bStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, bDim.data(), bDim.size(), bData); \
auto output = aclCreateTensor( \
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData); \
\
uint64_t workspaceSize = 0; \
aclOpExecutor *executor; \
\
auto ret = aclnn##prefix##GetWorkspaceSize( \
inputA, inputB, output, &workspaceSize, &executor); \
void *workspaceAddr = nullptr; \
if (workspaceSize > 0) { \
workspaceAddr = context->getWorkspace(workspaceSize); \
} \
assert(ret == ACL_SUCCESS); \
ret = aclnn##prefix(workspaceAddr, workspaceSize, executor, \
context->ASCENDHandle()); \
assert(ret == ACL_SUCCESS); \
\
ret = aclrtSynchronizeStream(context->ASCENDHandle()); \
assert(ret == ACL_SUCCESS); \
\
ret = aclDestroyTensor(inputA); \
ret = aclDestroyTensor(inputB); \
ret = aclDestroyTensor(output); \
\
return; \
} \
};
class AddAclnn : public ASCENDKernelWithoutConfig {
virtual tuple<float, float, float> getAlphBeta() const {
return {1.f, 1.f, 0.f};
}
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<ElementWiseObj>(_op);
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
void *const cData = (op->getOutput()->getRawDataPtr<void *>());
auto a = op->getInputs(0)->getDims();
auto aS = op->getInputs(0)->getStride();
auto b = op->getInputs(1)->getDims();
auto bS = op->getInputs(1)->getStride();
auto c = op->getInputs(0)->getDims();
auto cS = op->getInputs(0)->getStride();
std::vector<int64_t> aDim = MycastTo64(a);
std::vector<int64_t> aStride = MycastTo64(aS);
std::vector<int64_t> bDim = MycastTo64(b);
std::vector<int64_t> bStride = MycastTo64(bS);
std::vector<int64_t> cDim = MycastTo64(c);
std::vector<int64_t> cStride = MycastTo64(cS);
auto inputA = aclCreateTensor(
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0,
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData);
auto inputB = aclCreateTensor(
bDim.data(), bDim.size(), ACL_FLOAT, bStride.data(), 0,
aclFormat::ACL_FORMAT_ND, bDim.data(), bDim.size(), bData);
auto output = aclCreateTensor(
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0,
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData);
auto [aAlpha, bAlpha, beta] = getAlphBeta();
auto alpha = aclCreateScalar(&bAlpha, ACL_FLOAT);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret = aclnnAddGetWorkspaceSize(inputA, inputB, alpha, output,
&workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnAdd(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
// ret = aclDestroyTensor(inputA);
// ret = aclDestroyTensor(inputB);
// ret = aclDestroyScalar(alpha);
// ret = aclDestroyTensor(output);
return;
}
};
class SubAclnn : public ASCENDKernelWithoutConfig {
virtual tuple<float, float, float> getAlphBeta() const {
return {1.f, 1.f, 0.f};
}
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<ElementWiseObj>(_op);
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
void *const cData = (op->getOutput()->getRawDataPtr<void *>());
auto a = op->getInputs(0)->getDims();
auto aS = op->getInputs(0)->getStride();
auto b = op->getInputs(1)->getDims();
auto bS = op->getInputs(1)->getStride();
auto c = op->getInputs(0)->getDims();
auto cS = op->getInputs(0)->getStride();
std::vector<int64_t> aDim = MycastTo64(a);
std::vector<int64_t> aStride = MycastTo64(aS);
std::vector<int64_t> bDim = MycastTo64(b);
std::vector<int64_t> bStride = MycastTo64(bS);
std::vector<int64_t> cDim = MycastTo64(c);
std::vector<int64_t> cStride = MycastTo64(cS);
auto inputA = aclCreateTensor(
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0,
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData);
auto inputB = aclCreateTensor(
bDim.data(), bDim.size(), ACL_FLOAT, bStride.data(), 0,
aclFormat::ACL_FORMAT_ND, bDim.data(), bDim.size(), bData);
auto output = aclCreateTensor(
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0,
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData);
auto [aAlpha, bAlpha, beta] = getAlphBeta();
auto alpha = aclCreateScalar(&bAlpha, ACL_FLOAT);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret = aclnnSubGetWorkspaceSize(inputA, inputB, alpha, output,
&workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnSub(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclDestroyTensor(inputA);
ret = aclDestroyTensor(inputB);
ret = aclDestroyScalar(alpha);
ret = aclDestroyTensor(output);
return;
}
};
DEFINE_ELEMENT_WISE_Aclnn(PowTensorTensor);
DEFINE_ELEMENT_WISE_Aclnn(Div);
DEFINE_ELEMENT_WISE_Aclnn(Mul);
REGISTER_KERNEL(Device::ASCEND, OpType::Pow, PowTensorTensorAclnn,
"pow_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Div, DivAclnn, "div_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Mul, MulAclnn, "mul_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Add, AddAclnn, "add_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Sub, SubAclnn, "sub_ASCEND_float");
// REGISTER_KERNEL(Device::ASCEND, OpType::Abs, AbsAclnn, "abs_ASCEND_float");
}; // namespace infini

30
src/kernels/ascend/matmul.cc
View File

@ -5,10 +5,8 @@
namespace infini {
class MatmulAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<MatmulObj>(_op);
@ -38,40 +36,36 @@ class MatmulAclnn : public ASCENDKernelWithoutConfig {
auto matTensor = aclCreateTensor(
matDim.data(), matDim.size(), ACL_FLOAT, matStride.data(), 0,
aclFormat::ACL_FORMAT_ND, matDim.data(), matDim.size(), bData);
auto outputTensor = aclCreateTensor(
outputDim.data(), outputDim.size(), ACL_FLOAT, outputStride.data(), 0,
aclFormat::ACL_FORMAT_ND, outputDim.data(), outputDim.size(), cData);
auto outputTensor =
aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
outputStride.data(), 0, aclFormat::ACL_FORMAT_ND,
outputDim.data(), outputDim.size(), cData);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret =
aclnnMatmulGetWorkspaceSize(selfTensor, matTensor, outputTensor, 1, &workspaceSize, &executor);
auto ret = aclnnMatmulGetWorkspaceSize(
selfTensor, matTensor, outputTensor, 1, &workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize,
ACL_MEM_MALLOC_HUGE_FIRST);
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnMatmul(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
aclDestroyTensor(selfTensor);
aclDestroyTensor(matTensor);
aclDestroyTensor(outputTensor);
// aclDestroyTensor(selfTensor);
// aclDestroyTensor(matTensor);
// aclDestroyTensor(outputTensor);
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::MatMul, DataType::Float32, MatmulAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::MatMul, MatmulAclnn,
"matmul_ASCEND_float");
}; // namespace infini

38
src/kernels/ascend/pooling.cc
View File

@ -5,10 +5,8 @@
namespace infini {
class AvgPooling : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<PoolingObj>(_op);
@ -24,8 +22,7 @@ class AvgPooling : public ASCENDKernelWithoutConfig {
std::vector<int64_t> stride = {sh, sw};
std::vector<int64_t> pad = {ph, pw};
int64_t divisorOverride = kh * kw;
int64_t divisorOverride = kh * kw;
auto selfD = op->getInputs(0)->getDims();
auto selfS = op->getInputs(0)->getStride();
@ -37,46 +34,43 @@ class AvgPooling : public ASCENDKernelWithoutConfig {
std::vector<int64_t> outputDim = MycastTo64(outD);
std::vector<int64_t> outputStride = MycastTo64(outS);
aclIntArray *kernelSize = aclCreateIntArray(ksize.data(), ksize.size());
aclIntArray *strides = aclCreateIntArray(stride.data(), stride.size());
aclIntArray *paddings = aclCreateIntArray(pad.data(), pad.size());
aclIntArray *kernelSize = aclCreateIntArray(ksize.data(), ksize.size());
aclIntArray *strides = aclCreateIntArray(stride.data(), stride.size());
aclIntArray *paddings = aclCreateIntArray(pad.data(), pad.size());
auto selfTensor = aclCreateTensor(
selfDim.data(), selfDim.size(), ACL_FLOAT, selfStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, selfDim.data(), selfDim.size(), aData);
auto outputTensor = aclCreateTensor(
outputDim.data(), outputDim.size(), ACL_FLOAT, outputStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, outputDim.data(), outputDim.size(), cData);
auto outputTensor =
aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
outputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
outputDim.data(), outputDim.size(), cData);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret =
aclnnAvgPool2dGetWorkspaceSize(selfTensor, kernelSize, strides, paddings, false, true, divisorOverride, 1, outputTensor, &workspaceSize, &executor);
auto ret = aclnnAvgPool2dGetWorkspaceSize(
selfTensor, kernelSize, strides, paddings, false, true,
divisorOverride, 1, outputTensor, &workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize,
ACL_MEM_MALLOC_HUGE_FIRST);
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnAvgPool2d(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
aclDestroyTensor(selfTensor);
aclDestroyTensor(outputTensor);
// aclDestroyTensor(selfTensor);
// aclDestroyTensor(outputTensor);
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::AveragePool, DataType::Float32, AvgPooling,
REGISTER_KERNEL(Device::ASCEND, OpType::AveragePool, AvgPooling,
"avgpooling_ASCEND_float");
}; // namespace infini

62
src/kernels/ascend/softmax.cc Normal file
View File

@ -0,0 +1,62 @@
#include "operators/softmax.h"
#include "aclnnop/level2/aclnn_softmax.h"
#include "ascend/ascend_kernel_without_config.h"
#include "ascend/ascend_runtime.h"
namespace infini {
class SoftmaxAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<SoftmaxObj>(_op);
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
void *const cData = (op->getOutput()->getRawDataPtr<void *>());
int64_t axis = int64_t(op->getAxis());
auto a = op->getInputs(0)->getDims();
auto aS = op->getInputs(0)->getStride();
auto c = op->getInputs(0)->getDims();
auto cS = op->getInputs(0)->getStride();
std::vector<int64_t> aDim = MycastTo64(a);
std::vector<int64_t> aStride = MycastTo64(aS);
std::vector<int64_t> cDim = MycastTo64(c);
std::vector<int64_t> cStride = MycastTo64(cS);
auto input = aclCreateTensor(
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0,
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData);
auto output = aclCreateTensor(
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0,
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret = aclnnSoftmaxGetWorkspaceSize(input, axis, output,
&workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnSoftmax(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
// aclDestroyTensor(input);
// aclDestroyTensor(output);
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::Softmax, SoftmaxAclnn,
"softmax_ASCEND_float");
}; // namespace infini

222
src/kernels/ascend/unary.cc
View File

@ -1,21 +1,21 @@
#include "operators/unary.h"
#include "aclnnop/level2/aclnn_relu.h"
#include "aclnnop/level2/aclnn_abs.h"
#include "aclnnop/level2/aclnn_sigmoid.h"
#include "aclnnop/level2/aclnn_hardswish.h"
#include "aclnnop/level2/aclnn_tanh.h"
#include "aclnnop/level2/aclnn_gelu.h"
#include "aclnnop/level2/aclnn_sin.h"
#include "aclnnop/level2/aclnn_cos.h"
#include "aclnnop/level2/aclnn_acos.h"
#include "aclnnop/level2/aclnn_atan.h"
#include "aclnnop/level2/aclnn_ceil.h"
#include "aclnnop/level2/aclnn_floor.h"
#include "aclnnop/level2/aclnn_cos.h"
#include "aclnnop/level2/aclnn_exp.h"
#include "aclnnop/level2/aclnn_floor.h"
#include "aclnnop/level2/aclnn_gelu.h"
#include "aclnnop/level2/aclnn_hardswish.h"
#include "aclnnop/level2/aclnn_neg.h"
#include "aclnnop/level2/aclnn_reciprocal.h"
#include "aclnnop/level2/aclnn_sqrt.h"
#include "aclnnop/level2/aclnn_relu.h"
#include "aclnnop/level2/aclnn_round.h"
#include "aclnnop/level2/aclnn_sigmoid.h"
#include "aclnnop/level2/aclnn_sin.h"
#include "aclnnop/level2/aclnn_sqrt.h"
#include "aclnnop/level2/aclnn_tanh.h"
#include "ascend/ascend_kernel_without_config.h"
#include "ascend/ascend_runtime.h"
@ -64,138 +64,120 @@ class ReluAclnn : public ASCENDKernelWithoutConfig {
aclnnReluGetWorkspaceSize(input, output, &workspaceSize, &executor);
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize,
ACL_MEM_MALLOC_HUGE_FIRST);
workspaceAddr = context->getWorkspace(workspaceSize);
}
assert(ret == ACL_SUCCESS);
ret = aclnnRelu(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
//ret = aclDestroyTensor(input);
//assert(ret == ACL_SUCCESS);
//ret = aclDestroyTensor(output);
//assert(ret == ACL_SUCCESS);
// aclDestroyTensor(input);
// aclDestroyTensor(output);
ret = aclrtSynchronizeStream(context->ASCENDHandle());
assert(ret == ACL_SUCCESS);
return;
}
};
#define DEFINE_UNARY_Aclnn(prefix) \
class prefix##Aclnn : public ASCENDKernelWithoutConfig { \
void compute(const Operator &_op, \
const RuntimeObj *_context) const override { \
auto op = as<UnaryObj>(_op); \
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context); \
\
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>()); \
void *const cData = (op->getOutput()->getRawDataPtr<void *>()); \
\
auto a = op->getInputs(0) -> getDims(); \
std::vector<int64_t> aDim(a.size(), 1); \
for (size_t i = 0; i < a.size(); ++i) { \
aDim[i] = int64_t(a[i]); \
} \
auto aS = op->getInputs(0) -> getStride(); \
std::vector<int64_t> aStride(aS.size(), 1); \
for (size_t i = 0; i < aS.size(); ++i) { \
aStride[i] = int64_t(aS[i]); \
} \
auto c = op->getInputs(0) -> getDims(); \
std::vector<int64_t> cDim(c.size(), 1); \
for (size_t i = 0; i < c.size(); ++i) { \
cDim[i] = int64_t(c[i]); \
} \
auto cS = op->getInputs(0) -> getStride(); \
std::vector<int64_t> cStride(cS.size(), 1); \
for (size_t i = 0; i < cS.size(); ++i) { \
cStride[i] = int64_t(cS[i]); \
} \
\
auto input = aclCreateTensor( \
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData); \
auto output = aclCreateTensor( \
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData); \
\
uint64_t workspaceSize = 0; \
aclOpExecutor *executor; \
\
auto ret = aclnn##prefix##GetWorkspaceSize( \
input, output, &workspaceSize, &executor); \
void *workspaceAddr = nullptr; \
if (workspaceSize > 0) { \
workspaceAddr = context->getWorkspace(workspaceSize); \
} \
assert(ret == ACL_SUCCESS); \
ret = aclnn##prefix(workspaceAddr, workspaceSize, executor, \
context->ASCENDHandle()); \
assert(ret == ACL_SUCCESS); \
ret = aclrtSynchronizeStream(context->ASCENDHandle()); \
assert(ret == ACL_SUCCESS); \
\
return; \
} \
};
#define DEFINE_UNARY_Aclnn(prefix) \
class prefix##Aclnn : public ASCENDKernelWithoutConfig { \
void compute(const Operator &_op, \
const RuntimeObj *_context) const override { \
auto op = as<UnaryObj>(_op); \
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context); \
\
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>()); \
void *const cData = (op->getOutput()->getRawDataPtr<void *>()); \
\
auto a = op->getInputs(0)->getDims(); \
std::vector<int64_t> aDim(a.size(), 1); \
for (size_t i = 0; i < a.size(); ++i) { \
aDim[i] = int64_t(a[i]); \
} \
auto aS = op->getInputs(0)->getStride(); \
std::vector<int64_t> aStride(aS.size(), 1); \
for (size_t i = 0; i < aS.size(); ++i) { \
aStride[i] = int64_t(aS[i]); \
} \
auto c = op->getInputs(0)->getDims(); \
std::vector<int64_t> cDim(c.size(), 1); \
for (size_t i = 0; i < c.size(); ++i) { \
cDim[i] = int64_t(c[i]); \
} \
auto cS = op->getInputs(0)->getStride(); \
std::vector<int64_t> cStride(cS.size(), 1); \
for (size_t i = 0; i < cS.size(); ++i) { \
cStride[i] = int64_t(cS[i]); \
} \
\
auto input = aclCreateTensor( \
aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData); \
auto output = aclCreateTensor( \
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0, \
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData); \
\
uint64_t workspaceSize = 0; \
aclOpExecutor *executor; \
\
auto ret = aclnn##prefix##GetWorkspaceSize(input, output, &workspaceSize, &executor); \
void *workspaceAddr = nullptr; \
if (workspaceSize > 0) { \
ret = aclrtMalloc(&workspaceAddr, workspaceSize, \
ACL_MEM_MALLOC_HUGE_FIRST); \
} \
assert(ret == ACL_SUCCESS); \
ret = aclnn##prefix(workspaceAddr, workspaceSize, executor, \
context->ASCENDHandle()); \
assert(ret == ACL_SUCCESS); \
ret = aclrtSynchronizeStream(context->ASCENDHandle()); \
assert(ret == ACL_SUCCESS); \
\
return; \
} \
};
DEFINE_UNARY_Aclnn(Abs);
DEFINE_UNARY_Aclnn(Sigmoid);
DEFINE_UNARY_Aclnn(Hardswish);
DEFINE_UNARY_Aclnn(Gelu);
DEFINE_UNARY_Aclnn(Abs)
DEFINE_UNARY_Aclnn(Sigmoid)
DEFINE_UNARY_Aclnn(Hardswish)
DEFINE_UNARY_Aclnn(Gelu)
DEFINE_UNARY_Aclnn(Tanh);
DEFINE_UNARY_Aclnn(Sin);
DEFINE_UNARY_Aclnn(Cos);
DEFINE_UNARY_Aclnn(Acos);
DEFINE_UNARY_Aclnn(Atan);
DEFINE_UNARY_Aclnn(Tanh)
DEFINE_UNARY_Aclnn(Sin)
DEFINE_UNARY_Aclnn(Cos)
DEFINE_UNARY_Aclnn(Acos)
DEFINE_UNARY_Aclnn(Atan)
DEFINE_UNARY_Aclnn(Ceil);
DEFINE_UNARY_Aclnn(Floor);
DEFINE_UNARY_Aclnn(Exp);
DEFINE_UNARY_Aclnn(Neg);
DEFINE_UNARY_Aclnn(Reciprocal);
DEFINE_UNARY_Aclnn(Sqrt);
DEFINE_UNARY_Aclnn(Round);
DEFINE_UNARY_Aclnn(Ceil)
DEFINE_UNARY_Aclnn(Floor)
DEFINE_UNARY_Aclnn(Exp)
DEFINE_UNARY_Aclnn(Neg)
DEFINE_UNARY_Aclnn(Reciprocal)
DEFINE_UNARY_Aclnn(Sqrt)
DEFINE_UNARY_Aclnn(Round)
REGISTER_KERNEL(Device::ASCEND, OpType::Relu, DataType::Float32, ReluAclnn,
"relu_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Abs, DataType::Float32, AbsAclnn,
"abs_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Sigmoid, DataType::Float32, SigmoidAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::Relu, ReluAclnn, "relu_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Abs, AbsAclnn, "abs_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Sigmoid, SigmoidAclnn,
"sigmoid_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::HardSwish, DataType::Float32, HardswishAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::HardSwish, HardswishAclnn,
"hardswish_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Tanh, DataType::Float32, TanhAclnn,
"tanh_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Gelu, DataType::Float32, GeluAclnn,
"gelu_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Sin, DataType::Float32, SinAclnn,
"sin_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Cos, DataType::Float32, CosAclnn,
"cos_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Acos, DataType::Float32, AcosAclnn,
"acos_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Atan, DataType::Float32, AtanAclnn,
"atan_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Neg, DataType::Float32, NegAclnn,
"neg_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Ceil, DataType::Float32, CeilAclnn,
"ceil_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Floor, DataType::Float32, FloorAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::Tanh, TanhAclnn, "tanh_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Gelu, GeluAclnn, "gelu_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Sin, SinAclnn, "sin_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Cos, CosAclnn, "cos_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Acos, AcosAclnn, "acos_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Atan, AtanAclnn, "atan_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Neg, NegAclnn, "neg_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Ceil, CeilAclnn, "ceil_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Floor, FloorAclnn,
"floor_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Exp, DataType::Float32, ExpAclnn,
"exp_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Reciprocal, DataType::Float32, ReciprocalAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::Exp, ExpAclnn, "exp_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Reciprocal, ReciprocalAclnn,
"reciprocal_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Sqrt, DataType::Float32, SqrtAclnn,
"sqrt_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Round, DataType::Float32, RoundAclnn,
REGISTER_KERNEL(Device::ASCEND, OpType::Sqrt, SqrtAclnn, "sqrt_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Round, RoundAclnn,
"round_ASCEND_float");
}; // namespace infini

3
test/kernels/ascend/test_ascend_batch_norm.cc
View File

@ -9,6 +9,7 @@
namespace infini {
TEST(ascend_BatchNorm, run) {
aclInit(nullptr);
Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
auto npuRuntime = make_ref<ASCENDRuntimeObj>();
@ -51,5 +52,7 @@ TEST(ascend_BatchNorm, run) {
// check results on CPU
EXPECT_TRUE(ocpu->equalData(vector<float>{
-0.5, 0, 0.5, 1, -2, -1, 0, 1, -0.333333, 0, 0.333333, 0.666667}));
aclFinalize();
}
} // namespace infini

22
test/kernels/ascend/test_ascend_concat.cc
View File

@ -24,40 +24,42 @@ void testConcat(const std::function<void(void *, size_t, DataType)> &generator,
make_ref<TensorObj>(shape, DataType::Float32, cpuRuntime);
inputCpu2->dataMalloc();
inputCpu2->setData(generator);
Tensor inputCpu3 =
make_ref<TensorObj>(shape, DataType::Float32, cpuRuntime);
inputCpu3->dataMalloc();
inputCpu3->setData(generator);
// NPU
Graph npuGraph = make_ref<GraphObj>(npuRuntime);
auto inputNpu1 = npuGraph->cloneTensor(inputCpu1);
auto inputNpu2 = npuGraph->cloneTensor(inputCpu2);
auto npuOp =
npuGraph->addOp<T>(TensorVec{inputNpu1, inputNpu2}, nullptr, 2);
auto inputNpu3 = npuGraph->cloneTensor(inputCpu3);
auto npuOp = npuGraph->addOp<T>(TensorVec{inputNpu1, inputNpu2, inputNpu3},
nullptr, 2);
npuGraph->dataMalloc();
inputNpu1->setData(generator);
inputNpu2->setData(generator);
inputNpu3->setData(generator);
npuRuntime->run(npuGraph);
auto outputNpu = npuOp->getOutput();
auto outputNpu2Cpu = outputNpu->clone(cpuRuntime);
/********************************************************/
auto inputTest1 = inputNpu1->clone(cpuRuntime);
auto inputTest2 = inputNpu2->clone(cpuRuntime);
inputTest1->printData();
inputTest2->printData();
/********************************************************/
// Check
inputCpu1->print();
inputCpu1->printData();
inputCpu2->print();
inputCpu2->printData();
inputCpu3->print();
inputCpu3->printData();
outputNpu2Cpu->print();
outputNpu2Cpu->printData();
EXPECT_TRUE(1);
}
TEST(ascend_Concat, run) {
aclInit(nullptr);
testConcat<ConcatObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
aclFinalize();
}
} // namespace infini

5
test/kernels/ascend/test_ascend_conv.cc
View File

@ -1,12 +1,11 @@
#include "ascend/ascend_runtime.h"
#include "core/graph.h"
#include "core/kernel.h"
#include "core/runtime.h"
#include "ascend/ascend_runtime.h"
#include "operators/conv.h"
#include "test.h"
namespace infini {
template <class T>
@ -50,8 +49,10 @@ void testConv(const std::function<void(void *, size_t, DataType)> &generatorA,
}
TEST(ascend_Conv, run) {
aclInit(nullptr);
testConv<ConvObj>(IncrementalGenerator(), IncrementalGenerator(),
Shape{1, 3, 32, 32}, Shape{2, 3, 3, 3});
aclFinalize();
}
} // namespace infini

61
test/kernels/ascend/test_ascend_element_wise.cc Normal file
View File

@ -0,0 +1,61 @@
#include "ascend/ascend_runtime.h"
#include "core/graph.h"
#include "core/kernel.h"
#include "core/runtime.h"
#include "operators/element_wise.h"
#include "test.h"
namespace infini {
template <class T>
void testElementWise(
const std::function<void(void *, size_t, DataType)> &generator,
const Shape &shape) {
// Runtime
Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
auto npuRuntime = make_ref<ASCENDRuntimeObj>();
// Build input data on CPU
Tensor inputCpu1 =
make_ref<TensorObj>(shape, DataType::Float32, cpuRuntime);
Tensor inputCpu2 =
make_ref<TensorObj>(shape, DataType::Float32, cpuRuntime);
inputCpu1->dataMalloc();
inputCpu2->dataMalloc();
inputCpu1->setData(generator);
inputCpu2->setData(generator);
// NPU
Graph npuGraph = make_ref<GraphObj>(npuRuntime);
auto inputNpu1 = npuGraph->cloneTensor(inputCpu1);
auto inputNpu2 = npuGraph->cloneTensor(inputCpu2);
auto npuOp = npuGraph->addOp<T>(inputNpu1, inputNpu2, nullptr);
npuGraph->dataMalloc();
inputNpu1->setData(generator);
inputNpu2->setData(generator);
npuRuntime->run(npuGraph);
auto outputNpu = npuOp->getOutput();
auto outputNpu2Cpu = outputNpu->clone(cpuRuntime);
// Check
inputCpu1->print();
inputCpu1->printData();
inputCpu2->print();
inputCpu2->printData();
outputNpu2Cpu->print();
outputNpu2Cpu->printData();
EXPECT_TRUE(1);
}
TEST(ascend_ElementWise, run) {
aclInit(nullptr);
testElementWise<PowObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testElementWise<AddObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testElementWise<SubObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testElementWise<DivObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testElementWise<MulObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
aclFinalize();
}
} // namespace infini

2
test/kernels/ascend/test_ascend_matmul.cc
View File

@ -50,8 +50,10 @@ void testMatmul(const std::function<void(void *, size_t, DataType)> &generatorA,
}
TEST(ascend_Matmul, run) {
aclInit(nullptr);
testMatmul<MatmulObj>(IncrementalGenerator(), IncrementalGenerator(), false,
false, Shape{1, 2, 3}, Shape{1, 3, 4});
aclFinalize();
}
} // namespace infini

3
test/kernels/ascend/test_ascend_pooling.cc
View File

@ -29,6 +29,7 @@ void testPooling(const std::function<void(void *, size_t, DataType)> &generator,
npuGraph->dataMalloc();
inputNpu->setData(generator);
npuRuntime->run(npuGraph);
auto outputNpu = npuOp->getOutput();
auto outputNpu2Cpu = outputNpu->clone(cpuRuntime);
inputCpu->printData();
@ -37,8 +38,10 @@ void testPooling(const std::function<void(void *, size_t, DataType)> &generator,
}
TEST(cnnl_Pooling, run) {
aclInit(nullptr);
// testPooling<MaxPoolObj>(IncrementalGenerator(), Shape{1, 1, 5, 5});
testPooling<AvgPoolObj>(IncrementalGenerator(), Shape{1, 1, 5, 5});
aclFinalize();
}
} // namespace infini

55
test/kernels/ascend/test_ascend_softmax.cc Normal file
View File

@ -0,0 +1,55 @@
#include "ascend/ascend_runtime.h"
#include "core/graph.h"
#include "core/kernel.h"
#include "core/runtime.h"
#include "operators/softmax.h"
#include "test.h"
namespace infini {
template <class T>
void testSoftmax(const std::function<void(void *, size_t, DataType)> &generator,
const Shape &shape, int axis, vector<float> Out) {
// Runtime
Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
auto npuRuntime = make_ref<ASCENDRuntimeObj>();
// Build input data on CPU
Tensor inputCpu1 =
make_ref<TensorObj>(shape, DataType::Float32, cpuRuntime);
inputCpu1->dataMalloc();
// inputCpu1->setData(generator);
// NPU
Graph npuGraph = make_ref<GraphObj>(npuRuntime);
auto inputNpu1 = npuGraph->cloneTensor(inputCpu1);
auto npuOp = npuGraph->addOp<T>(inputNpu1, nullptr, axis);
npuGraph->dataMalloc();
inputNpu1->setData(generator);
npuRuntime->run(npuGraph);
auto outputNpu = npuOp->getOutput();
auto outputNpu2Cpu = outputNpu->clone(cpuRuntime);
// Check
EXPECT_TRUE(outputNpu2Cpu->equalData(Out));
}
TEST(ascend_ElementWise, run) {
aclInit(nullptr);
testSoftmax<SoftmaxObj>(
IncrementalGenerator(), Shape{2, 2, 2, 2}, 1,
vector<float>{0.0179862, 0.0179862, 0.0179862, 0.0179862, 0.9820138,
0.9820138, 0.9820138, 0.9820138, 0.0179862, 0.0179862,
0.0179862, 0.0179862, 0.9820138, 0.9820138, 0.9820138,
0.9820138});
testSoftmax<SoftmaxObj>(
IncrementalGenerator(), Shape{2, 2, 2, 2}, 3,
vector<float>{0.2689414, 0.7310586, 0.2689414, 0.7310586, 0.2689414,
0.7310586, 0.2689414, 0.7310586, 0.2689414, 0.7310586,
0.2689414, 0.7310586, 0.2689414, 0.7310586, 0.2689414,
0.7310586});
aclFinalize();
}
} // namespace infini

24
test/kernels/ascend/test_ascend_unary.cc
View File

@ -13,20 +13,20 @@ void testUnary(const std::function<void(void *, size_t, DataType)> &generator,
const Shape &shape) {
// Runtime
Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
auto xpuRuntime = make_ref<ASCENDRuntimeObj>();
auto npuRuntime = make_ref<ASCENDRuntimeObj>();
// Build input data on CPU
Tensor inputCpu = make_ref<TensorObj>(shape, DataType::Float32, cpuRuntime);
// GPU
Graph xpuGraph = make_ref<GraphObj>(xpuRuntime);
auto inputGpu = xpuGraph->cloneTensor(inputCpu);
auto gpuOp = xpuGraph->addOp<T>(inputGpu, nullptr);
xpuGraph->dataMalloc();
inputGpu->setData(generator);
xpuRuntime->run(xpuGraph);
auto outputGpu = gpuOp->getOutput();
auto outputGpu2Cpu = outputGpu->clone(cpuRuntime);
Graph npuGraph = make_ref<GraphObj>(npuRuntime);
auto inputNpu = npuGraph->cloneTensor(inputCpu);
auto npuOp = npuGraph->addOp<T>(inputNpu, nullptr);
npuGraph->dataMalloc();
inputNpu->setData(generator);
npuRuntime->run(npuGraph);
auto outputNpu = npuOp->getOutput();
auto outputNpu2Cpu = outputNpu->clone(cpuRuntime);
// CPU
Graph cpuGraph = make_ref<GraphObj>(cpuRuntime);
auto cpuOp = cpuGraph->addOp<T>(inputCpu, nullptr);
@ -36,10 +36,11 @@ void testUnary(const std::function<void(void *, size_t, DataType)> &generator,
cpuRuntime->run(cpuGraph);
auto outputCpu = cpuOp->getOutput();
// Check
EXPECT_TRUE(outputCpu->equalData(outputGpu2Cpu, 1e-3));
EXPECT_TRUE(outputCpu->equalData(outputNpu2Cpu, 1e-3));
}
TEST(ascend_Unary, run) {
aclInit(nullptr);
testUnary<ReluObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testUnary<AbsObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testUnary<SigmoidObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
@ -52,11 +53,12 @@ TEST(ascend_Unary, run) {
testUnary<ATanObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
// testUnary<CeilObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
// testUnary<FloorObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
// testUnary<ExpObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
// testUnary<ExpObj>(IncrementalGenerators(), Shape{1, 2, 2, 3});
testUnary<NegObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
// testUnary<ReciprocalObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
testUnary<SqrtObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
// testUnary<RoundObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
aclFinalize();
}
} // namespace infini