p95768021
/
InfiniTensor
forked from jiuyuan/InfiniTensor
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

add layernorm

This commit is contained in:
OdinaryWord 2024-04-26 15:25:41 +08:00
parent a765cd2a3d
commit 6ba1a0648a
3 changed files with 306 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

105
src/kernels/ascend/layer_norm.cc Normal file
View File

@ -0,0 +1,105 @@
#include "operators/layer_norm.h"
#include "aclnnop/level2/aclnn_layer_norm.h"
#include "ascend/ascend_kernel_without_config.h"
#include "ascend/ascend_runtime.h"
#include "operators/gather.h"
namespace infini {
class LayerNormAclnn : public ASCENDKernelWithoutConfig {
void compute(const Operator &_op,
const RuntimeObj *_context) const override {
auto op = as<LayerNormObj>(_op);
auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
void *const inputData = (op->getInputs(0)->getRawDataPtr<void *>());
void *const weightData = (op->getInputs(1)->getRawDataPtr<void *>());
void *const outputData = (op->getOutput()->getRawDataPtr<void *>());
auto inputD = op->getInputs(0)->getDims();
auto inputS = op->getInputs(0)->getStride();
auto weightD = op->getInputs(1)->getDims();
auto weightS = op->getInputs(1)->getStride();
auto outD = op->getOutput()->getDims();
auto outS = op->getOutput()->getStride();
double eps = static_cast<double>(op->getEps());
std::vector<int64_t> inputDim = castTo64(inputD);
std::vector<int64_t> inputStride = castTo64(inputS);
std::vector<int64_t> weightDim = castTo64(weightD);
std::vector<int64_t> weightStride = castTo64(weightS);
std::vector<int64_t> outputDim = castTo64(outD);
std::vector<int64_t> outputStride = castTo64(outS);
auto axis = op->getAxis();
auto rank = static_cast<int>(inputDim.size());
std::vector<int64_t> normalizedShape(rank - axis, 0);
for (auto i = rank; i > axis; --i) {
normalizedShape[i - 1 - axis] = inputDim[i - 1];
}
auto inputTensor =
aclCreateTensor(inputDim.data(), inputDim.size(), ACL_FLOAT,
inputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
inputDim.data(), inputDim.size(), inputData);
auto weightTensor =
aclCreateTensor(weightDim.data(), weightDim.size(), ACL_FLOAT,
weightStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
weightDim.data(), weightDim.size(), weightData);
auto outputTensor =
aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
outputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
outputDim.data(), outputDim.size(), outputData);
auto *normArray =
aclCreateIntArray(normalizedShape.data(), normalizedShape.size());
aclTensor *biasTensor = NULL;
if (op->numInputs() == 3) {
void *const biasData = (op->getInputs(2)->getRawDataPtr<void *>());
auto biasD = op->getInputs(2)->getDims();
auto biasS = op->getInputs(2)->getStride();
std::vector<int64_t> biasDim = castTo64(biasD);
std::vector<int64_t> biasStride = castTo64(biasS);
biasTensor = aclCreateTensor(
biasDim.data(), biasDim.size(), ACL_FLOAT, biasStride.data(), 0,
aclFormat::ACL_FORMAT_NCHW, biasDim.data(), biasDim.size(),
biasData);
}
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
auto ret = aclnnLayerNormGetWorkspaceSize(
inputTensor, normArray, weightTensor, biasTensor, eps, outputTensor,
NULL, NULL, &workspaceSize, &executor);
CHECK_RET(
ret == ACL_SUCCESS,
LOG_PRINT("aclnnLayerNormGetWorkspaceSize failed. ERROR: %d\n",
ret));
void *workspaceAddr = nullptr;
if (workspaceSize > 0) {
workspaceAddr = context->getWorkspace(workspaceSize);
}
ret = aclnnLayerNorm(workspaceAddr, workspaceSize, executor,
context->ASCENDHandle());
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclnnLayerNorm failed. ERROR: %d\n", ret));
ret = aclrtSynchronizeStream(context->ASCENDHandle());
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret));
return;
}
};
REGISTER_KERNEL(Device::ASCEND, OpType::LayerNormalization, LayerNormAclnn,
"LayerNorm_ASCEND");
}; // namespace infini

96
src/kernels/ascend/unary.cc
View File

@ -92,52 +92,53 @@ class ReluAclnn : public ASCENDKernelWithoutConfig {
void *const aData = (op->getInputs(0)->getRawDataPtr<void *>()); \ void *const aData = (op->getInputs(0)->getRawDataPtr<void *>()); \
void *const cData = (op->getOutput()->getRawDataPtr<void *>()); \ void *const cData = (op->getOutput()->getRawDataPtr<void *>()); \
\ \
auto a = op->getInputs(0)->getDims(); \ auto a = op->getInputs(0) -> getDims();
std::vector<int64_t> aDim(a.size(), 1); \
for (size_t i = 0; i < a.size(); ++i) { \ std::vector<int64_t> aDim(a.size(), 1);
aDim[i] = int64_t(a[i]); \ 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); \ auto aS = op->getInputs(0)->getStride();
for (size_t i = 0; i < aS.size(); ++i) { \ std::vector<int64_t> aStride(aS.size(), 1);
aStride[i] = int64_t(aS[i]); \ 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); \ auto c = op->getInputs(0)->getDims();
for (size_t i = 0; i < c.size(); ++i) { \ std::vector<int64_t> cDim(c.size(), 1);
cDim[i] = int64_t(c[i]); \ 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); \ auto cS = op->getInputs(0)->getStride();
for (size_t i = 0; i < cS.size(); ++i) { \ std::vector<int64_t> cStride(cS.size(), 1);
cStride[i] = int64_t(cS[i]); \ 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, \ auto input =
aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData); \ aclCreateTensor(aDim.data(), aDim.size(), ACL_FLOAT, aStride.data(), 0,
auto output = aclCreateTensor( \ aclFormat::ACL_FORMAT_ND, aDim.data(), aDim.size(), aData);
cDim.data(), cDim.size(), ACL_FLOAT, cStride.data(), 0, \ auto output =
aclFormat::ACL_FORMAT_ND, cDim.data(), cDim.size(), cData); \ 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; \ uint64_t workspaceSize = 0;
\ aclOpExecutor *executor;
auto ret = aclnn##prefix##GetWorkspaceSize( \
input, output, &workspaceSize, &executor); \ auto ret =
void *workspaceAddr = nullptr; \ aclnn##prefix##GetWorkspaceSize(input, output, &workspaceSize, &executor);
if (workspaceSize > 0) { \ void *workspaceAddr = nullptr;
workspaceAddr = context->getWorkspace(workspaceSize); \ if (workspaceSize > 0) {
} \ workspaceAddr = context->getWorkspace(workspaceSize);
assert(ret == ACL_SUCCESS); \ }
ret = aclnn##prefix(workspaceAddr, workspaceSize, executor, \ assert(ret == ACL_SUCCESS);
context->ASCENDHandle()); \ ret = aclnn##prefix(workspaceAddr, workspaceSize, executor,
assert(ret == ACL_SUCCESS); \ context->ASCENDHandle());
ret = aclrtSynchronizeStream(context->ASCENDHandle()); \ assert(ret == ACL_SUCCESS);
assert(ret == ACL_SUCCESS); \ ret = aclrtSynchronizeStream(context->ASCENDHandle());
\ assert(ret == ACL_SUCCESS);
return; \
} \ return;
} // namespace infini \
}; };
DEFINE_UNARY_Aclnn(Abs); DEFINE_UNARY_Aclnn(Abs);
@ -184,4 +185,5 @@ REGISTER_KERNEL(Device::ASCEND, OpType::Sqrt, SqrtAclnn, "sqrt_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Round, RoundAclnn, REGISTER_KERNEL(Device::ASCEND, OpType::Round, RoundAclnn,
"round_ASCEND_float"); "round_ASCEND_float");
REGISTER_KERNEL(Device::ASCEND, OpType::Erf, ErfAclnn, "erf_ASCEND_float"); REGISTER_KERNEL(Device::ASCEND, OpType::Erf, ErfAclnn, "erf_ASCEND_float");
}; // namespace infini }
; // namespace infini

152
test/kernels/ascend/test_ascend_layernorm.cc Normal file
View File

@ -0,0 +1,152 @@
#include "ascend/ascend_runtime.h"
#include "core/graph.h"
#include "core/kernel.h"
#include "core/runtime.h"
#include "operators/layer_norm.h"
#include "test.h"
namespace infini {
void test_layernormFp32(
const Shape &inputShape, const vector<float> &inputData,
const Shape &scaleShape, const vector<float> &scaleData, float eps,
int axis, int stash_type, const vector<float> &ExpectData,
const std::optional<Shape> &bShape = std::nullopt,
const std::optional<std::vector<float>> &biasData = std::nullopt) {
Runtime runtime = NativeCpuRuntimeObj::getInstance();
Graph gCpu = make_ref<GraphObj>(runtime);
if (bShape.has_value() && biasData.has_value()) {
Shape biasShape = *bShape;
auto bias = gCpu->addTensor(biasShape, DataType::Float32);
auto input = gCpu->addTensor(inputShape, DataType::Float32);
auto scale = gCpu->addTensor(scaleShape, DataType::Float32);
gCpu->dataMalloc();
bias->copyin(*biasData); //
// bias->printData();
input->copyin(inputData);
scale->copyin(scaleData); //
auto ascendRuntime = make_ref<ASCENDRuntimeObj>();
Graph gAscend = make_ref<GraphObj>(ascendRuntime);
auto biasNpu = gAscend->cloneTensor(bias);
auto inputNpu = gAscend->cloneTensor(input);
auto scaleNpu = gAscend->cloneTensor(scale);
// gCpu->cloneTensor(biasNpu)->printData();
auto op =
gAscend->addOp<LayerNormObj>(inputNpu, scaleNpu, nullptr, biasNpu,
eps, axis, stash_type); // LayernormObj
gAscend->dataMalloc();
biasNpu->copyin(*biasData);
// gCpu->cloneTensor(biasNpu)->printData();
inputNpu->copyin(inputData);
scaleNpu->copyin(scaleData);
ascendRuntime->run(gAscend);
auto oCpu =
gCpu->cloneTensor(op->getOutput()); // move Data from npu to cpu
oCpu->printData(); //->printData
EXPECT_TRUE(oCpu->equalData(ExpectData));
} else {
auto input = gCpu->addTensor(inputShape, DataType::Float32);
auto scale = gCpu->addTensor(scaleShape, DataType::Float32);
gCpu->dataMalloc();
input->copyin(inputData);
scale->copyin(scaleData); //
auto ascendRuntime = make_ref<ASCENDRuntimeObj>();
Graph gAscend = make_ref<GraphObj>(ascendRuntime);
auto inputNpu = gAscend->cloneTensor(input);
auto scaleNpu = gAscend->cloneTensor(scale);
auto op =
gAscend->addOp<LayerNormObj>(inputNpu, scaleNpu, nullptr, nullptr,
eps, axis, stash_type); // LayernormObj
gAscend->dataMalloc();
inputNpu->copyin(inputData);
scaleNpu->copyin(scaleData);
ascendRuntime->run(gAscend);
auto oCpu =
gCpu->cloneTensor(op->getOutput()); // move Data from npu to cpu
oCpu->printData(); //->printData
EXPECT_TRUE(oCpu->equalData(ExpectData));
}
}
TEST(CUDA_LayernormFp32, run) {
aclInit(nullptr);
test_layernormFp32(
Shape{2, 3, 2, 3},
vector<float>{0., 1., 2., 3., 4., 5., 6., 7., 8.,
9., 10., 11., 12., 13., 14., 15., 16., 17.,
18., 19., 20., 21., 22., 23., 24., 25., 26.,
27., 28., 29., 30., 31., 32., 33., 34., 35.},
Shape{3}, vector<float>{0.3, 0.2, 0.5}, 1e-5, 3, 1,
vector<float>{
-0.3674207, 0.0000000, 0.6123678, -0.3674207, 0.0000000, 0.6123678,
-0.3674207, 0.0000000, 0.6123678, -0.3674207, 0.0000000, 0.6123678,
-0.3674207, 0.0000000, 0.6123678, -0.3674207, 0.0000000, 0.6123678,
-0.3674207, 0.0000000, 0.6123678, -0.3674207, 0.0000000, 0.6123678,
-0.3674207, 0.0000000, 0.6123678, -0.3674207, 0.0000000, 0.6123678,
-0.3674207, 0.0000000, 0.6123678, -0.3674207, 0.0000000, 0.6123678},
Shape{3}, vector<float>{0, 0, 0});
// test_layernormFp32(
// Shape{2, 3, 2, 3},
// vector<float>{0., 1., 2., 3., 4., 5., 6., 7., 8.,
// 9., 10., 11., 12., 13., 14., 15., 16., 17.,
// 18., 19., 20., 21., 22., 23., 24., 25., 26.,
// 27., 28., 29., 30., 31., 32., 33., 34., 35.},
// Shape{3}, vector<float>{0.3, 0.2, 0.5}, 1e-5, 3, 1,
// vector<float>{
// -0.0674207, 0.2000000, 1.1123679, -0.0674207,
// 0.2000000, 1.1123679, -0.0674207, 0.2000000, 1.1123679,
// -0.0674207, 0.2000000, 1.1123679, -0.0674207,
// 0.2000000, 1.1123679, -0.0674207, 0.2000000, 1.1123679,
// -0.0674207, 0.2000000, 1.1123679, -0.0674207,
// 0.2000000, 1.1123679, -0.0674207, 0.2000000, 1.1123679,
// -0.0674207, 0.2000000, 1.1123679, -0.0674207,
// 0.2000000, 1.1123679, -0.0674207, 0.2000000, 1.1123679},
// Shape{3}, vector<float>{0.3, 0.2, 0.5});
// test_layernormFp32(
// Shape{2, 3, 2, 3},
// vector<float>{0., 1., 2., 3., 4., 5., 6., 7., 8.,
// 9., 10., 11., 12., 13., 14., 15., 16., 17.,
// 18., 19., 20., 21., 22., 23., 24., 25., 26.,
// 27., 28., 29., 30., 31., 32., 33., 34., 35.},
// Shape{1}, vector<float>{0.3}, 1e-5, 3, 1,
// vector<float>{
// -0.0674207, 0.2000000, 0.8674207, -0.0674207, 0.2000000,
// 0.8674207, -0.0674207, 0.2000000, 0.8674207, -0.0674207,
// 0.2000000, 0.8674207, -0.0674207, 0.2000000, 0.8674207,
// -0.0674207, 0.2000000, 0.8674207, -0.0674207, 0.2000000,
// 0.8674207, -0.0674207, 0.2000000, 0.8674207, -0.0674207,
// 0.2000000, 0.8674207, -0.0674207, 0.2000000, 0.8674207,
// -0.0674207, 0.2000000, 0.8674207, -0.0674207, 0.2000000,
// 0.8674207},
// Shape{3}, vector<float>{0.3, 0.2, 0.5});
// test_layernormFp32(
// Shape{2, 3, 2, 3},
// vector<float>{0., 1., 2., 3., 4., 5., 6., 7., 8.,
// 9., 10., 11., 12., 13., 14., 15., 16., 17.,
// 18., 19., 20., 21., 22., 23., 24., 25., 26.,
// 27., 28., 29., 30., 31., 32., 33., 34., 35.},
// Shape{3}, vector<float>{0.3, 0.2, 0.5}, 1e-5, 3, 1,
// vector<float>{-0.3674207, 0.0000000, 0.6123678, -0.3674207,
// 0.0000000, 0.6123678, -0.3674207, 0.0000000,
// 0.6123678, -0.3674207, 0.0000000, 0.6123678,
// -0.3674207, 0.0000000, 0.6123678, -0.3674207,
// 0.0000000, 0.6123678, -0.3674207, 0.0000000,
// 0.6123678, -0.3674207, 0.0000000, 0.6123678,
// -0.3674207, 0.0000000, 0.6123678, -0.3674207,
// 0.0000000, 0.6123678, -0.3674207, 0.0000000,
// 0.6123678, -0.3674207, 0.0000000, 0.6123678});
aclFinalize();
} // python output
} // namespace infini