Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 95b356a0 编写于 作者: W Wilber 提交者： GitHub
浏览文件
操作

update embedding_eltwise_layernorm fuse and kernel. test=develop (#23114) 

update embedding_eltwise_layernorm fuse pass and fused kernel, to support multi input
上级 a31d7328
展开全部 隐藏空白更改
内联 并排
Showing with 585 addition and 320 deletion
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 95b356a0
......@@ -130,6 +130,7 @@ cc_test(test_skip_layernorm_fuse_pass SRCS skip_layernorm_fuse_pass_tester.cc DE
cc_test(test_multihead_matmul_fuse_pass SRCS multihead_matmul_fuse_pass_tester.cc DEPS multihead_matmul_fuse_pass)
cc_test(test_conv_bn_fuse_pass SRCS conv_bn_fuse_pass_tester.cc DEPS conv_bn_fuse_pass)
if(WITH_GPU)
cc_test(test_embedding_eltwise_layernorm_fuse_pass SRCS embedding_eltwise_layernorm_fuse_pass_tester.cc DEPS embedding_eltwise_layernorm_fuse_pass)
cc_test(test_cudnn_placement_pass SRCS cudnn_placement_pass_tester.cc DEPS cudnn_placement_pass)
endif()
if(NOT WIN32)
......
paddle/fluid/framework/ir/embedding_eltwise_layernorm_fuse_pass.h
浏览文件 @ 95b356a0
......@@ -16,6 +16,7 @@
#include <memory>
#include <string>
#include <utility>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
......@@ -25,35 +26,76 @@ namespace framework {
namespace ir {
namespace patterns {
struct EmbeddingEltwiseLayerNormPattern : public PatternBase {
EmbeddingEltwiseLayerNormPattern(PDPattern* pattern,
const std::string& name_scope)
: PatternBase(pattern, name_scope, "embedding_eltwise_layernorm") {}
// detect start pattern.
//
// in_var emb in_var emb
// | | | |
// lookup_table lookup_table
// | |
// lkt_var lkt_var
// \ /
// elementwise_add
// |
// elt_out_var
//
struct Embedding2Eltwise1Pattern : public PatternBase {
Embedding2Eltwise1Pattern(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "embedding2_eltwise1") {}
PDNode* operator()();
void operator()();
PATTERN_DECL_NODE(lookup_table1_x);
PATTERN_DECL_NODE(lookup_table2_x);
PATTERN_DECL_NODE(lookup_table3_x);
PATTERN_DECL_NODE(lookup_table1_w);
PATTERN_DECL_NODE(lookup_table2_w);
PATTERN_DECL_NODE(lookup_table3_w);
PATTERN_DECL_NODE(lookup_table1);
PATTERN_DECL_NODE(lookup_table2);
PATTERN_DECL_NODE(lookup_table3);
PATTERN_DECL_NODE(lookup_table1_out);
PATTERN_DECL_NODE(lookup_table2_out);
PATTERN_DECL_NODE(lookup_table3_out);
PATTERN_DECL_NODE(eltwise_add_12);
PATTERN_DECL_NODE(eltwise_add_12_out);
PATTERN_DECL_NODE(eltwise_add);
PATTERN_DECL_NODE(eltwise_add_out);
};
// detect repeats inner pattern
//
// elt_out_var in_var emb
// \ | |
// \ lookup_table
// \ |
// \ lkt_var
// \ /
// elementwise_add
// |
// elt_out_var
//
struct Embedding1Eltwise1Pattern : public PatternBase {
Embedding1Eltwise1Pattern(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "embedding1_eltwise1") {}
void operator()();
PATTERN_DECL_NODE(lookup_table1_x);
PATTERN_DECL_NODE(lookup_table1_w);
PATTERN_DECL_NODE(lookup_table1);
PATTERN_DECL_NODE(lookup_table1_out);
PATTERN_DECL_NODE(eltwise_add_in);
PATTERN_DECL_NODE(eltwise_add);
PATTERN_DECL_NODE(eltwise_add_out);
};
// detect end pattern
//
// elementwise_add
// |
// elt_out_var
// scale | bias
// \ | /
// layer_norm
//
struct SkipLayerNorm : public PatternBase {
SkipLayerNorm(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "skip_layernorm") {}
void operator()();
PATTERN_DECL_NODE(eltwise_add);
PATTERN_DECL_NODE(eltwise_add_out);
PATTERN_DECL_NODE(layer_norm);
PATTERN_DECL_NODE(layer_norm_bias);
PATTERN_DECL_NODE(layer_norm_scale);
......@@ -79,6 +121,23 @@ struct EmbeddingEltwiseLayerNormPattern : public PatternBase {
//
// (word, pos, sent, weights_0, weights_1, weights_2,
// scale, baias) embedding_eltwise_layernorm -> layer_norm_out
//
//
// in_var emb_var in_var emb_var in_var emb_var in_var emb_var
// | | | | | | | |
// lookup_table lookup_table lookup_table ... lookup_table
// | | | |
// lkt_var lkt_var lkt_var lkt_var
// \ / | ... |
// elementwise_add | |
// \ / |
// elementwise_add |
// | |
// elt_var /
// \ /
// elementwise_add
// |
// layer_norm
class EmbeddingEltwiseLayerNormFusePass : public FusePassBase {
public:
......
paddle/fluid/framework/ir/embedding_eltwise_layernorm_fuse_pass_tester.cc 0 → 100644
浏览文件 @ 95b356a0
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
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 "paddle/fluid/framework/ir/embedding_eltwise_layernorm_fuse_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/ir/pass_tester_helper.h"
namespace paddle {
namespace framework {
namespace ir {
TEST(SkipLayerNormFusePass, basic) {
// inputs operator output
// --------------------------------------------------------------------
// (x, y) elementwise_add -> elementwise_out
// (elementwise_out, scale, bias) layer_norm -> layer_norm_out...
Layers layers;
auto* x0 = layers.data("x0", {1, 256, 1});
auto* x1 = layers.data("x1", {1, 256, 1});
auto* x2 = layers.data("x2", {1, 256, 1});
auto* x3 = layers.data("x3", {1, 256, 1});
auto* emb0 = layers.data("emb0", {18000, 768}, true);
auto* emb1 = layers.data("emb1", {4, 768}, true);
auto* emb2 = layers.data("emb2", {513, 768}, true);
auto* emb3 = layers.data("emb3", {3, 768}, true);
auto* lkt0 = layers.embedding(x0, emb0);
auto* lkt1 = layers.embedding(x1, emb1);
auto* lkt2 = layers.embedding(x2, emb2);
auto* lkt3 = layers.embedding(x3, emb3);
auto* elementwise_out1 = layers.elementwise_add(lkt0, lkt2);
auto* elementwise_out2 = layers.elementwise_add(elementwise_out1, lkt1);
auto* elementwise_out3 = layers.elementwise_add(elementwise_out2, lkt3);
auto* scale = layers.data("scale", {768}, true);
auto* bias = layers.data("bias", {768}, true);
layers.layer_norm(elementwise_out3, scale, bias);
auto* y0 = layers.data("y0", {1, 256, 1});
auto* y1 = layers.data("y1", {1, 256, 1});
auto* y2 = layers.data("y2", {1, 256, 1});
auto* emb0y = layers.data("emb0y", {18000, 768}, true);
auto* emb1y = layers.data("emb1y", {4, 768}, true);
auto* emb2y = layers.data("emb2y", {513, 768}, true);
auto* lkt0y = layers.embedding(y0, emb0y);
auto* lkt1y = layers.embedding(y1, emb1y);
auto* lkt2y = layers.embedding(y2, emb2y);
auto* elementwise_out1y = layers.elementwise_add(lkt0y, lkt2y);
auto* elementwise_out2y = layers.elementwise_add(elementwise_out1y, lkt1y);
auto* scaley = layers.data("scaley", {768}, true);
auto* biasy = layers.data("biasy", {768}, true);
layers.layer_norm(elementwise_out2y, scaley, biasy);
std::unique_ptr<ir::Graph> graph(new ir::Graph(layers.main_program()));
auto pass =
PassRegistry::Instance().Get("embedding_eltwise_layernorm_fuse_pass");
int num_nodes_before = graph->Nodes().size();
VLOG(3) << DebugString(graph);
graph.reset(pass->Apply(graph.release()));
int num_nodes_after = graph->Nodes().size();
int num_fused_nodes_after =
GetNumOpNodes(graph, "fused_embedding_eltwise_layernorm");
VLOG(3) << DebugString(graph);
PADDLE_ENFORCE_EQ(num_nodes_before, num_nodes_after + 28,
platform::errors::PreconditionNotMet(
"The number of nodes before and after the fuse does "
"not meet expectations"));
PADDLE_ENFORCE_EQ(
num_fused_nodes_after, 2,
platform::errors::PreconditionNotMet(
"The number of fusion nodes does not meet expectations after fuse"));
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(embedding_eltwise_layernorm_fuse_pass);
paddle/fluid/operators/fused/fused_embedding_eltwise_layernorm_op.cc
浏览文件 @ 95b356a0
......@@ -26,33 +26,19 @@ class EmbeddingEltWiseLayerNormOp : public framework::OperatorWithKernel {
protected:
void InferShape(framework::InferShapeContext* context) const override {
PADDLE_ENFORCE_EQ(context->HasInput("WordId"), true,
PADDLE_ENFORCE_EQ(context->Inputs("Ids").size(),
context->Inputs("Embs").size(),
platform::errors::InvalidArgument(
"Input(WordId) of EmbeddingEltWiseLayerNormOp should "
"not be null."));
PADDLE_ENFORCE_EQ(
context->HasInput("PosId"), true,
platform::errors::InvalidArgument(
"Input(PosId) of EmbeddingEltWiseLayerNormOp should not be null."));
PADDLE_ENFORCE_EQ(context->HasInput("SentId"), true,
"Two inputs of EmbeddingEltWiseLayerNormOp shoube be "
"the same size"));
PADDLE_ENFORCE_GE(context->Inputs("Embs").size(), 2UL,
platform::errors::InvalidArgument(
"Input(SentId) of EmbeddingEltWiseLayerNormOp should "
"not be null."));
PADDLE_ENFORCE_EQ(context->HasInput("WordEmb"), true,
platform::errors::InvalidArgument(
"Input(WordEmb) of EmbeddingEltWiseLayerNormOp "
"should not be null."));
PADDLE_ENFORCE_EQ(context->HasInput("PosEmb"), true,
"Input Embs of EmbeddingEltWiseLayerNormOp should "
"have at least 2 tensors"));
PADDLE_ENFORCE_GE(context->Inputs("Ids").size(), 2UL,
platform::errors::InvalidArgument(
"Input(PosEmb) of EmbeddingEltWiseLayerNormOp should "
"not be null."));
PADDLE_ENFORCE_EQ(context->HasInput("SentEmb"), true,
platform::errors::InvalidArgument(
"Input(SentEmb) of EmbeddingEltWiseLayerNormOp "
"should not be null."));
"Input Ids of EmbeddingEltWiseLayerNormOp should "
"have at least 2 tensors"));
PADDLE_ENFORCE_EQ(
context->HasInput("Bias"), true,
......@@ -70,55 +56,55 @@ class EmbeddingEltWiseLayerNormOp : public framework::OperatorWithKernel {
"Output(Out) of EmbeddingEltWiseLayerNormOp should not be null."));
// batch * seq_len * 1
auto dims_word_id = context->GetInputDim("WordId");
auto ids_dims = context->GetInputsDim("Ids");
// word_num * hidden
auto dims_word_emb = context->GetInputDim("WordEmb");
auto dims_pos_emb = context->GetInputDim("PosEmb");
auto dims_sent_emb = context->GetInputDim("SentEmb");
auto embs_dims = context->GetInputsDim("Embs");
// hidden
auto dims_bias = context->GetInputDim("Bias");
PADDLE_ENFORCE_EQ(
dims_word_emb[1], dims_bias[0],
platform::errors::InvalidArgument(
"The second dims (%d) of the Word Embedding should be equal "
"to the Bias's size(%d).",
dims_word_emb[1], dims_bias[0]));
PADDLE_ENFORCE_EQ(dims_word_emb.size(), 2,
platform::errors::InvalidArgument(
"The WordEmb dim's size shoule be 2, but found %d.",
dims_word_emb.size()));
PADDLE_ENFORCE_EQ(dims_pos_emb.size(), 2,
platform::errors::InvalidArgument(
"The PosEmb dim's size shoule be 2, but found %d.",
dims_pos_emb.size()));
PADDLE_ENFORCE_EQ(dims_sent_emb.size(), 2,
platform::errors::InvalidArgument(
"The SentEmb dim's size shoule be 2, but found %d.",
dims_sent_emb.size()));
PADDLE_ENFORCE_EQ(
dims_word_emb[1], dims_pos_emb[1],
platform::errors::InvalidArgument(
"The WordEmb first dim size(%d) shoule equal to PosEmb ones(%d).",
dims_word_emb[1], dims_pos_emb[1]));
PADDLE_ENFORCE_EQ(
dims_word_emb[1], dims_sent_emb[1],
platform::errors::InvalidArgument(
"The WordEmb first dim size(%d) shoule equal to SentEmb ones(%d).",
dims_word_emb[1], dims_sent_emb[1]));
int batch = ids_dims[0][0];
int seq_len = ids_dims[0][1];
int hidden = embs_dims[0][1];
for (size_t i = 0; i < embs_dims.size(); ++i) {
PADDLE_ENFORCE_EQ(embs_dims[i].size(), 2,
platform::errors::InvalidArgument(
"The Emb dim's size shoule be 2, but found %d.",
embs_dims[i].size()));
PADDLE_ENFORCE_EQ(
embs_dims[i][1], dims_bias[0],
platform::errors::InvalidArgument(
"The second dims (%d) of the Embedding should be equal "
"to the Bias's size(%d).",
embs_dims[i][1], dims_bias[0]));
PADDLE_ENFORCE_EQ(
embs_dims[i][1], hidden,
platform::errors::InvalidArgument(
"The Emb first dim size(%d) shoule equal to hidden (%d).",
embs_dims[i][1], hidden));
}
int batch = dims_word_id[0];
int seq_len = dims_word_id[1];
int hidden = dims_word_emb[1];
auto dim_output = framework::make_ddim({batch, seq_len, hidden});
context->SetOutputDim("Out", dim_output);
context->ShareLoD("WordId", /*->*/ "Out");
context->ShareLoD("Ids", /*->*/ "Out");
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto data_type = OperatorWithKernel::IndicateVarDataType(ctx, "WordEmb");
return framework::OpKernelType(data_type, ctx.device_context());
auto inputs = ctx.MultiInput<framework::Tensor>("Embs");
auto input_data_type = framework::proto::VarType::Type(0);
bool flag = 0;
for (auto* input : inputs) {
if (input->IsInitialized() && input->numel() > 0) {
input_data_type = input->type();
flag = 1;
break;
}
}
if (flag == 0) {
PADDLE_THROW(platform::errors::PreconditionNotMet(
"All Inputs of fused_embedding_eltwise_layernorm OP are Empty!"));
}
return framework::OpKernelType(input_data_type, ctx.GetPlace());
}
};
......@@ -126,15 +112,10 @@ class EmbeddingEltWiseLayerNormOpMaker
: public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("WordId", "The word id input of EmbeddingEltWiseLayerNorm op");
AddInput("PosId", "The position id input of EmbeddingEltWiseLayerNorm op");
AddInput("SentId", "The sentence id input of EmbeddingEltWiseLayerNorm op");
AddInput("WordEmb",
"The Word embedding input of EmbeddingEltWiseLayerNorm op");
AddInput("PosEmb",
"The Position embedding input of EmbeddingEltWiseLayerNorm op");
AddInput("SentEmb",
"The Sent embedding input of EmbeddingEltWiseLayerNorm op");
AddInput("Ids", "Input id tensors of EmbeddingEltWiseLayerNorm op")
.AsDuplicable();
AddInput("Embs", "Input emb tensors of EmbeddingEltWiseLayerNorm op")
.AsDuplicable();
AddInput("Bias", "The LayerNorm Bias of EmbeddingEltWiseLayerNorm op");
AddInput("Scale", "The LayerNorm Scale of EmbeddingEltWiseLayerNorm op");
AddOutput("Out", "The output of EmbeddingEltWiseLayerNorm op");
......@@ -157,10 +138,11 @@ class EmbeddingEltWiseLayerNormOpMaker
EmbeddingEltWiseLayerNorm Operator.
This op is used for optimize the following structure in ernie model.
wordid -> lookup_table_op -> word
posid -> lookup_table_op -> pos
sentdid -> lookup_table_op -> sent
word + pos + sent -> Y
id1 -> lookup_table_op -> data1
id2 -> lookup_table_op -> data2
...
idn -> lookup_table_op -> data_n
data1 + data2 + ... + data_n -> Y
Y -> layer_norm -> Out
Not suggest to use in other case except has same structure as ernie.
......
paddle/fluid/operators/fused/fused_embedding_eltwise_layernorm_op.cu
浏览文件 @ 95b356a0
......@@ -16,6 +16,7 @@
#include <paddle/fluid/platform/device_context.h>
#include <algorithm>
#include <cub/cub.cuh> // NOLINT
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/malloc.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
......@@ -57,32 +58,28 @@ __device__ inline void LayerNorm(const cv2<T> &thread_data, const int ld,
}
template <typename T, unsigned TPB>
__global__ void EmbEltwiseLayernormKernel(
int hidden, const int64_t *word_id_d, const int64_t *pos_id_d,
const int64_t *sent_id_d, const T *scale, const T *bias, const T *word_emb,
const T *pos_emb, const T *sent_emb, T *output, float eps) {
__global__ void EmbEltwiseLayernormKernel(int hidden, const int64_t *ids,
const T *scale, const T *bias,
const int64_t *embs, T *output,
float eps, int input_num) {
cub::Sum pair_sum;
// blockIdx.x: position in the sequence
// blockIdx.y: batch
// gridDim.x: Seq
// gridDim.y: Batch
__shared__ int64_t word_id;
__shared__ int64_t pos_id;
__shared__ int64_t sent_id;
extern __shared__ int64_t array_id[];
const T rhidden = T(1.f) / T(hidden);
const int64_t seq_pos = blockIdx.y + blockIdx.x * gridDim.y;
if (threadIdx.x == 0) {
word_id = word_id_d[seq_pos];
pos_id = pos_id_d[seq_pos];
sent_id = sent_id_d[seq_pos];
for (int i = 0; i < input_num; ++i) {
const int64_t *ids_p = reinterpret_cast<const int64_t *>(ids[i]);
array_id[i] = ids_p[seq_pos];
}
}
__syncthreads();
// load word, pos, sentence embeddings and add them toghether
const int64_t woffset = word_id * hidden;
const int64_t poffset = pos_id * hidden;
const int64_t soffset = sent_id * hidden;
const int64_t out_offset = seq_pos * hidden;
cv2<T> thread_data;
......@@ -91,10 +88,10 @@ __global__ void EmbEltwiseLayernormKernel(
#pragma unroll
for (int it = threadIdx.x; it < hidden; it += TPB) {
const T w(word_emb[woffset + it]);
const T p(pos_emb[poffset + it]);
const T s(sent_emb[soffset + it]);
const T val = w + s + p;
T val = 0;
for (int i = 0; i < input_num; ++i) {
val += reinterpret_cast<const T *>(embs[i])[array_id[i] * hidden + it];
}
output[out_offset + it] = val;
const T rhiddenval = rhidden * val;
......@@ -112,47 +109,58 @@ class EmbeddingEltWiseLayerNormKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {
using Tensor = framework::Tensor;
auto *word_id = context.Input<framework::Tensor>("WordId");
auto *pos_id = context.Input<framework::Tensor>("PosId");
auto *sent_id = context.Input<framework::Tensor>("SentId");
auto *word_emb = context.Input<framework::Tensor>("WordEmb");
auto *pos_emb = context.Input<framework::Tensor>("PosEmb");
auto *sent_emb = context.Input<framework::Tensor>("SentEmb");
auto &device_ctx = context.template device_context<DeviceContext>();
auto ids = context.MultiInput<framework::Tensor>("Ids");
auto embs = context.MultiInput<framework::Tensor>("Embs");
int input_num = static_cast<int>(ids.size());
framework::Tensor in_ids_(framework::proto::VarType::INT64),
in_embs_(framework::proto::VarType::INT64);
framework::DDim in_dim{input_num};
int device_id;
cudaGetDevice(&device_id);
in_ids_.Resize(in_dim);
in_embs_.Resize(in_dim);
int64_t *in_ids_d =
in_ids_.mutable_data<int64_t>(platform::CUDAPlace(device_id));
int64_t *in_embs_d =
in_embs_.mutable_data<int64_t>(platform::CUDAPlace(device_id));
std::vector<int64_t> in1s, in2s;
for (int i = 0; i < input_num; ++i) {
in1s.push_back(reinterpret_cast<uintptr_t>(ids[i]->data<int64_t>()));
in2s.push_back(reinterpret_cast<uintptr_t>(embs[i]->data<T>()));
}
cudaMemcpyAsync(in_ids_d, in1s.data(), sizeof(int64_t) * input_num,
cudaMemcpyHostToDevice, device_ctx.stream());
cudaMemcpyAsync(in_embs_d, in2s.data(), sizeof(int64_t) * input_num,
cudaMemcpyHostToDevice, device_ctx.stream());
auto *bias = context.Input<framework::Tensor>("Bias");
auto *scale = context.Input<framework::Tensor>("Scale");
auto *out = context.Output<framework::Tensor>("Out");
auto *word_id_d = word_id->data<int64_t>();
auto *pos_id_d = pos_id->data<int64_t>();
auto *sent_id_d = sent_id->data<int64_t>();
// should be (B * S * hidden)
auto id0_dims = ids[0]->dims();
auto emb0_dims = embs[0]->dims();
auto *word_emb_d = word_emb->data<T>();
auto *pos_emb_d = pos_emb->data<T>();
auto *sent_emb_d = sent_emb->data<T>();
int batch = id0_dims[0];
int seq_len = id0_dims[1];
int hidden = emb0_dims[1];
auto *bias_d = bias->data<T>();
auto *scale_d = scale->data<T>();
auto *output_d = out->mutable_data<T>(context.GetPlace());
// compute q*k with eltadd
auto &device_ctx = context.template device_context<DeviceContext>();
float eps = context.Attr<float>("epsilon");
// should be (B * S * hidden)
auto word_id_dims = word_id->dims();
auto word_emb_dims = word_emb->dims();
int batch = word_id_dims[0];
int seq_len = word_id_dims[1];
int hidden = word_emb_dims[1];
const unsigned tpb = 256;
const dim3 grid(seq_len, batch, 1);
const dim3 block(tpb, 1, 1);
EmbEltwiseLayernormKernel<T, tpb><<<grid, block, 0, device_ctx.stream()>>>(
hidden, word_id_d, pos_id_d, sent_id_d, scale_d, bias_d, word_emb_d,
pos_emb_d, sent_emb_d, output_d, eps);
int shared_bytes = input_num * sizeof(int64_t);
EmbEltwiseLayernormKernel<
T, tpb><<<grid, block, shared_bytes, device_ctx.stream()>>>(
hidden, in_ids_d, scale_d, bias_d, in_embs_d, output_d, eps, input_num);
}
};
......
python/paddle/fluid/tests/unittests/ir/test_ir_embedding_eltwise_layernorm_fuse_pass.py
浏览文件 @ 95b356a0
......@@ -48,6 +48,39 @@ class EmbEltwiseLayerNormFusePassTest(PassTest):
add2 = fluid.layers.elementwise_add(add1, sent_emb)
hidden1 = fluid.layers.layer_norm(input=add2, begin_norm_axis=2)
id1 = fluid.layers.data(
name="id1",
shape=[1, 128, 1],
dtype="int64",
append_batch_size=False)
id2 = fluid.layers.data(
name="id2",
shape=[1, 128, 1],
dtype="int64",
append_batch_size=False)
id3 = fluid.layers.data(
name="id3",
shape=[1, 128, 1],
dtype="int64",
append_batch_size=False)
id4 = fluid.layers.data(
name="id4",
shape=[1, 128, 1],
dtype="int64",
append_batch_size=False)
emb1 = fluid.layers.embedding(
input=id1, size=(128, 768), dtype='float32')
emb2 = fluid.layers.embedding(
input=id2, size=(128, 768), dtype='float32')
emb3 = fluid.layers.embedding(
input=id3, size=(128, 768), dtype='float32')
emb4 = fluid.layers.embedding(
input=id4, size=(128, 768), dtype='float32')
add_1 = fluid.layers.elementwise_add(emb1, emb2)
add_2 = fluid.layers.elementwise_add(add_1, emb3)
add_3 = fluid.layers.elementwise_add(add_2, emb4)
hidden_1 = fluid.layers.layer_norm(input=add_3, begin_norm_axis=2)
self.feeds = {
"word_id": np.random.randint(
low=0, high=128, size=(1, 128, 1)).astype("int64"),
......@@ -55,11 +88,19 @@ class EmbEltwiseLayerNormFusePassTest(PassTest):
low=0, high=128, size=(1, 128, 1)).astype("int64"),
"sent_id": np.random.randint(
low=0, high=128, size=(1, 128, 1)).astype("int64"),
"id1": np.random.randint(
low=0, high=128, size=(1, 128, 1)).astype("int64"),
"id2": np.random.randint(
low=0, high=128, size=(1, 128, 1)).astype("int64"),
"id3": np.random.randint(
low=0, high=128, size=(1, 128, 1)).astype("int64"),
"id4": np.random.randint(
low=0, high=128, size=(1, 128, 1)).astype("int64"),
}
self.fetch_list = [hidden1]
self.fetch_list = [hidden1, hidden_1]
self.pass_names = "embedding_eltwise_layernorm_fuse_pass"
self.fused_op_type = "fused_embedding_eltwise_layernorm"
self.num_fused_ops = 1
self.num_fused_ops = 2
def test_check_output(self):
use_gpu_set = [True]
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或