Skip to content

N
ngraph
Unverified Commit 97a9dbb3 authored by Scott Cyphers's avatar Scott Cyphers Committed by GitHub
Browse files
Options

Merge branch 'master' into tomdol/plaidml

parents b534a674 05535a35
Hide whitespace changes
Inline Side-by-side
Showing with 672 additions and 71 deletions
.ci/onnx/onnxruntime/Jenkinsfile 0 → 100644
View file @ 97a9dbb3
// ******************************************************************************
// Copyright 2018-2019 Intel Corporation
//
// 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.
// ******************************************************************************
try{ if(LABEL.trim() == "") {throw new Exception();} }catch(Exception e){LABEL="onnx && ci"}; echo "${LABEL}"
NGRPAH_REPOSITORY = "https://github.com/NervanaSystems/ngraph.git"
NGRAPH_COMMIT_HASH = "${ghprbActualCommit}" // particular nGraph PR commit hash
ONNX_REPOSITORY = "https://github.com/NervanaSystems/onnxruntime.git"
ONNX_RUNTIME_BRANCH = "release"
def main(){
timeout(activity: true, time: 15) {
try{
stage("CloneRepos"){
CloneRepos()
}
stage("Apply Patch"){
ApplyPatch()
}
stage("Onnx Models"){
BuildAndTest()
}
}
catch(e) {
// Set result to ABORTED if exception contains exit code of a process interrupted by SIGTERM
if ("$e".contains("143")) {
currentBuild.result = "ABORTED"
} else {
currentBuild.result = "FAILURE"
}
}
stage("Clean"){
Clean()
}
}
}
def CloneRepos() {
dir("ngraph"){
checkout([
$class: 'GitSCM',
branches: [[name: "${NGRAPH_COMMIT_HASH}"]],
doGenerateSubmoduleConfigurations: false,
extensions: [[
$class: 'SubmoduleOption',
disableSubmodules: false,
parentCredentials: true,
recursiveSubmodules: true,
reference: '',
trackingSubmodules: false,
timeout: 15
]],
submoduleCfg: [],
userRemoteConfigs: [[
refspec: '+refs/pull/*:refs/remotes/origin/pr/*',
url: "${NGRPAH_REPOSITORY}"
]]
])
}
dir("onnxruntime") {
checkout([
$class: 'GitSCM',
branches: [[name: "${ONNX_RUNTIME_BRANCH}"]],
doGenerateSubmoduleConfigurations: false,
extensions: [[
$class: 'SubmoduleOption',
disableSubmodules: false,
parentCredentials: true,
recursiveSubmodules: true,
reference: '',
trackingSubmodules: false,
timeout: 15
]],
submoduleCfg: [],
userRemoteConfigs: [[
url: "${ONNX_REPOSITORY}"
]]
])
}
}
def ApplyPatch(){
dir("onnxruntime"){
echo "Update cmake/external/ngraph.cmake with ${NGRAPH_COMMIT_HASH}"
sh """
sed -i 's/set(ngraph_TAG ".*")/set(ngraph_TAG "${NGRAPH_COMMIT_HASH}")/g' cmake/external/ngraph.cmake
grep -q "${NGRAPH_COMMIT_HASH}" cmake/external/ngraph.cmake
"""
echo "Add proxy to tools/ci_build/github/linux/docker/Dockerfile.ubuntu"
sh """
sed -i 's|{HTTP_PROXY}|${env.http_proxy}|g' ../ngraph/.ci/onnx/onnxruntime/proxy.patch
sed -i 's|{SOCKS_PROXY}|${env.socks_proxy}|g' ../ngraph/.ci/onnx/onnxruntime/proxy.patch
grep -q "${env.http_proxy}" ../ngraph/.ci/onnx/onnxruntime/proxy.patch
git apply ../ngraph/.ci/onnx/onnxruntime/proxy.patch
"""
}
}
def BuildAndTest(){
dir("onnxruntime"){
sh "mkdir -p `pwd`/build/models && chmod 777 `pwd`/build/models"
sh """
//!/bin/bash
./tools/ci_build/github/linux/run_dockerbuild.sh \
-o ubuntu16.04 \
-d ngraph \
-r `pwd`/build -x '--use_ngraph --use_full_protobuf --test_data_url https://onnxruntimetestdata.blob.core.windows.net/models/20190327.zip --test_data_checksum 45166d81c021c8aae212b53c92101792'
"""
}
}
def Clean(){
deleteDir()
}
node(LABEL) {
main()
}
.ci/onnx/onnxruntime/proxy.patch 0 → 100644
View file @ 97a9dbb3
diff --git a/tools/ci_build/github/linux/docker/Dockerfile.ubuntu b/tools/ci_build/github/linux/docker/Dockerfile.ubuntu
index bdff95e1..cd9c0008 100644
--- a/tools/ci_build/github/linux/docker/Dockerfile.ubuntu
+++ b/tools/ci_build/github/linux/docker/Dockerfile.ubuntu
@@ -3,6 +3,18 @@ FROM ubuntu:${OS_VERSION}
ARG PYTHON_VERSION=3.5
+ENV http_proxy={HTTP_PROXY}
+ENV socks_proxy={SOCKS_PROXY}
+ENV https_proxy={HTTP_PROXY}
+ENV ftp_proxy={HTTP_PROXY}
+ENV rsync_proxy={HTTP_PROXY}
+ENV no_proxy=intel.com,.intel.com,localhost
+ENV HTTP_PROXY={HTTP_PROXY}
+ENV HTTPS_PROXY={HTTP_PROXY}
+ENV FTP_PROXY={HTTP_PROXY}
+ENV SOCKS_PROXY={SOCKS_PROXY}
+ENV NO_PROXY=intel.com,.intel.com,localhost
+
ADD scripts /tmp/scripts
RUN /tmp/scripts/install_ubuntu.sh -p ${PYTHON_VERSION} && /tmp/scripts/install_deps.sh && rm -rf /tmp/scripts
CONTRIB.md
View file @ 97a9dbb3
Contributor Guidelines
======================
https://ngraph.nervanasys.com/docs/latest/project/code-contributor-README.html
The latest version of this file can be found at:
https://ngraph.nervanasys.com/docs/latest/project/contribution-guide.html
License
......
src/ngraph/runtime/cpu/cpu_call_frame.cpp
View file @ 97a9dbb3
......@@ -216,8 +216,6 @@ void runtime::cpu::CPU_CallFrame::setup_runtime_context()
{
// single thread for codegen
NGRAPH_CHECK(m_num_ctx == 1);
ctx->mkldnn_primitives.swap(mkldnn_emitter->get_mkldnn_primitives());
ctx->mkldnn_workspaces = mkldnn_emitter->get_mkldnn_workspaces();
}
ctx->states = m_external_function->m_states.data();
......
src/ngraph/runtime/cpu/kernel/sigmoid_multiply.hpp
View file @ 97a9dbb3
......@@ -57,28 +57,28 @@ namespace ngraph
{
case 0 /*Logistic|Logistic*/:
{
auto c = (in0.exp() * in1.exp()) / ((in0.exp() + 1.f) * (in1.exp() + 1.f));
auto c = 1.f / (((-in0).exp() + 1.f) * ((-in1).exp() + 1.f));
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
break;
case 1 /*Logistic|Tanh*/:
{
auto c = (in0.exp() * in1.tanh()) / (in0.exp() + 1.f);
auto c = in1.tanh() / ((-in0).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
break;
case 2 /*Logistic|Identity*/:
{
auto c = (in0.exp() * in1) / (in0.exp() + 1.f);
auto c = in1 / ((-in0).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
break;
case 3 /*Tanh|Logistic*/:
{
auto c = (in0.tanh() * in1.exp()) / (in1.exp() + 1.f);
auto c = in0.tanh() / ((-in1).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
......@@ -99,7 +99,7 @@ namespace ngraph
break;
case 6 /*Identity|Logistic*/:
{
auto c = (in0 * in1.exp()) / (in1.exp() + 1.f);
auto c = in0 / ((-in1).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
......@@ -141,10 +141,15 @@ namespace ngraph
{
case 0 /*Logistic|Logistic*/:
{
auto i0 = delta * (in1.exp() * in0.exp()) /
((in1.exp() + 1.f) * ((in0.exp() + 1.f) * (in0.exp() + 1.f)));
auto i1 = delta * (in0.exp() * in1.exp()) /
((in0.exp() + 1.f) * ((in1.exp() + 1.f) * (in1.exp() + 1.f)));
auto in0_neg_exp = (-in0).exp();
auto in0_log_denominator = in0_neg_exp + 1.f;
auto in1_neg_exp = (-in1).exp();
auto in1_log_denominator = in1_neg_exp + 1.f;
auto i0 = delta * in0_neg_exp /
(in1_log_denominator * in0_log_denominator * in0_log_denominator);
auto i1 = delta * in1_neg_exp /
(in0_log_denominator * in1_log_denominator * in1_log_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -153,12 +158,17 @@ namespace ngraph
break;
case 1 /*Logistic|Tanh*/:
{
auto in0_neg_exp = (-in0).exp();
auto in0_log_denominator = in0_neg_exp + 1.f;
auto in1_2exp = (in1 * 2.f).exp();
auto in1_tanh_denominator = in1_2exp + 1.f;
auto i0 =
delta * (((in1 * 2.f).exp() - 1.f) * in0.exp()) /
(((in1 * 2.f).exp() + 1.f) * ((in0.exp() + 1.f) * (in0.exp() + 1.f)));
auto i1 = delta * (in0.exp() * (4.f * (in1 * 2.f).exp())) /
((in0.exp() + 1.f) *
(((in1 * 2.f).exp() + 1.f) * ((in1 * 2.f).exp() + 1.f)));
delta * ((in1_2exp - 1.f) * in0_neg_exp) /
(in1_tanh_denominator * in0_log_denominator * in0_log_denominator);
auto i1 =
delta * (4.f * in1_2exp) /
(in0_log_denominator * in1_tanh_denominator * in1_tanh_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -167,9 +177,12 @@ namespace ngraph
break;
case 2 /*Logistic|Identity*/:
{
auto i0 =
delta * (in1 * in0.exp()) / ((in0.exp() + 1.f) * (in0.exp() + 1.f));
auto i1 = delta * in0.exp() / ((in0.exp() + 1.f));
auto in0_neg_exp = (-in0).exp();
auto in0_log_denominator = in0_neg_exp + 1.f;
auto i0 = delta * (in1 * in0_neg_exp) /
(in0_log_denominator * in0_log_denominator);
auto i1 = delta / in0_log_denominator;
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -178,12 +191,17 @@ namespace ngraph
break;
case 3 /*Tanh|Logistic*/:
{
auto i0 = delta * (in1.exp() * (4.f * (in0 * 2.f).exp())) /
((in1.exp() + 1.f) * ((in0 * 2.f).exp() + 1.f) *
((in0 * 2.f).exp() + 1.f));
auto in0_2exp = (in0 * 2.f).exp();
auto in0_tanh_denominator = in0_2exp + 1.f;
auto in1_neg_exp = (-in1).exp();
auto in1_log_denominator = in1_neg_exp + 1.f;
auto i0 =
delta * (4.f * in0_2exp) /
(in1_log_denominator * in0_tanh_denominator * in0_tanh_denominator);
auto i1 =
delta * (((in0 * 2.f).exp() - 1.f) * in1.exp()) /
(((in0 * 2.f).exp() + 1.f) * ((in1.exp() + 1.f) * (in1.exp() + 1.f)));
delta * ((in0_2exp - 1.f) * in1_neg_exp) /
(in0_tanh_denominator * in1_log_denominator * in1_log_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -192,12 +210,17 @@ namespace ngraph
break;
case 4 /*Tanh|Tanh*/:
{
auto i0 = delta * (((in1 * 2.f).exp() - 1.f) * (4.f * (in0 * 2.f).exp())) /
(((in1 * 2.f).exp() + 1.f) *
(((in0 * 2.f).exp() + 1.f) * ((in0 * 2.f).exp() + 1.f)));
auto i1 = delta * (((in0 * 2.f).exp() - 1.f) * (4.f * (in1 * 2.f).exp())) /
(((in0 * 2.f).exp() + 1.f) *
(((in1 * 2.f).exp() + 1.f) * ((in1 * 2.f).exp() + 1.f)));
auto in0_2exp = (in0 * 2.f).exp();
auto in0_tanh_denominator = in0_2exp + 1.f;
auto in1_2exp = (in1 * 2.f).exp();
auto in1_tanh_denominator = in1_2exp + 1.f;
auto i0 =
delta * (in1_2exp - 1.f) * 4.f * in0_2exp /
(in1_tanh_denominator * in0_tanh_denominator * in0_tanh_denominator);
auto i1 =
delta * (in0_2exp - 1.f) * 4.f * in1_2exp /
(in0_tanh_denominator * in1_tanh_denominator * in1_tanh_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -206,9 +229,12 @@ namespace ngraph
break;
case 5 /*Tanh|Identity*/:
{
auto i0 = delta * (in1 * (4.f * (in0 * 2.f).exp())) /
(((in0 * 2.f).exp() + 1.f) * ((in0 * 2.f).exp() + 1.f));
auto i1 = delta * ((in0 * 2.f).exp() - 1.f) / ((in0 * 2.f).exp() + 1.f);
auto in0_2exp = (in0 * 2.f).exp();
auto in0_tanh_denominator = in0_2exp + 1.f;
auto i0 = delta * in1 * 4.f * in0_2exp /
(in0_tanh_denominator * in0_tanh_denominator);
auto i1 = delta * (in0_2exp - 1.f) / in0_tanh_denominator;
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -217,9 +243,12 @@ namespace ngraph
break;
case 6 /*Identity|Logistic*/:
{
auto i0 = delta * (in1.exp()) / (in1.exp() + 1.f);
auto in1_neg_exp = (-in1).exp();
auto in1_log_denominator = in1_neg_exp + 1.f;
auto i0 = delta * 1.f / in1_log_denominator;
auto i1 =
delta * (in0 * in1.exp()) / ((in1.exp() + 1.f) * (in1.exp() + 1.f));
delta * in0 * in1_neg_exp / (in1_log_denominator * in1_log_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -228,9 +257,12 @@ namespace ngraph
break;
case 7 /*Identity|Tanh*/:
{
auto i0 = delta * ((in1 * 2.f).exp() - 1.f) / ((in1 * 2.f).exp() + 1.f);
auto i1 = delta * (in0 * (4.f * (in1 * 2.f).exp())) /
(((in1 * 2.f).exp() + 1.f) * ((in1 * 2.f).exp() + 1.f));
auto in1_2exp = (in1 * 2.f).exp();
auto in1_tanh_denominator = in1_2exp + 1.f;
auto i0 = delta * (in1_2exp - 1.f) / in1_tanh_denominator;
auto i1 = delta * (in0 * (4.f * in1_2exp)) /
(in1_tanh_denominator * in1_tanh_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......
src/ngraph/runtime/plaidml/CMakeLists.txt
View file @ 97a9dbb3
......@@ -41,6 +41,7 @@ set(SRC
plaidml_ops_one_hot.cpp
plaidml_ops_passthrough.cpp
plaidml_ops_pool.cpp
plaidml_ops_quantize.cpp
plaidml_ops_reduce.cpp
plaidml_ops_replace_slice.cpp
plaidml_ops_replicate.cpp
......
src/ngraph/runtime/plaidml/plaidml_builder.hpp
View file @ 97a9dbb3
......@@ -188,7 +188,8 @@ class ngraph::runtime::plaidml::builder::Elementwise final : public Statement
{
public:
Elementwise(std::string lhs, std::string rhs);
void set_lhs(const std::string& lhs) { m_lhs = lhs; }
void set_rhs(const std::string& rhs) { m_rhs = rhs; }
private:
friend class Function;
......
src/ngraph/runtime/plaidml/plaidml_ops_quantize.cpp 0 → 100644
View file @ 97a9dbb3
//*****************************************************************************
// Copyright 2017-2019 Intel Corporation
//
// 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 "ngraph/op/dequantize.hpp"
#include "ngraph/op/quantize.hpp"
#include "ngraph/runtime/plaidml/plaidml_impl.hpp"
namespace ngraph
{
namespace runtime
{
namespace plaidml
{
NGRAPH_PLAIDML_OP_CLASS(ImplDequantize, OpImpl<op::Dequantize>);
NGRAPH_PLAIDML_OP_CLASS(ImplQuantize, OpImpl<op::Quantize>);
}
}
}
void ngraph::runtime::plaidml::ImplDequantize::Apply()
{
check_inputs(3);
check_outputs(1);
const auto& axes = op().get_axes();
const auto& input_shape = op().get_input_shape(0);
const auto& scale_shape = op().get_input_shape(1);
const auto& zp_shape = op().get_input_shape(2);
const auto& input_type = op().get_input_element_type(0);
if (!input_type.is_signed() && input_type.size() >= 8)
{
throw std::runtime_error("PlaidML does not yet support dequantizing from uint64+");
}
if (scale_shape != zp_shape)
{
throw std::runtime_error("Dequantize given mismatched scale & zero point shapes.");
}
if (scale_shape.size() != axes.size())
{
std::ostringstream msg;
msg << "Dequantize received " << axes.size()
<< " axes to use for scale & zero point, but those tensors have " << scale_shape.size()
<< " dimensions instead.";
throw std::runtime_error(msg.str());
}
std::vector<std::string> short_idxs;
for (size_t i = 0; i < input_shape.size(); ++i)
{
if (axes.count(i))
{
std::ostringstream name;
name << "i" << i;
short_idxs.push_back(name.str());
}
}
builder::ContractionInput scale_input{"S"};
builder::ContractionInput neg_zp_input{"NegZ"};
for (const auto& idx : short_idxs)
{
scale_input.add_indices({idx});
neg_zp_input.add_indices({idx});
}
std::function<std::string(std::string)> cast_uint_to_wider_int =
[input_type](std::string tensor_name) {
std::ostringstream cast_str;
if (!input_type.is_signed())
{
cast_str << "as_int(" << tensor_name << ", " << 2 * 8 * input_type.size() << ")";
}
else
{
cast_str << tensor_name;
}
return cast_str.str();
};
builder::Elementwise CastI{"CastI", cast_uint_to_wider_int("I")};
builder::Elementwise CastZ{"CastZ", cast_uint_to_wider_int("Z")};
auto f = start_tile_function();
f.add(builder::Input{op_input(0), "I"}.add_dims("I", 0, input_shape.size()))
.add(builder::Input{op_input(1), "S"}.add_dims("S", 0, scale_shape.size()))
.add(builder::Input{op_input(2), "Z"}.add_dims("Z", 0, zp_shape.size()))
.add(builder::Output{"O"})
.add(CastI)
.add(CastZ)
.add(builder::Elementwise{"NegZ", "-CastZ"})
.add(
builder::BinaryContraction{"=", "+"}
.set(builder::ContractionOutput{"Offset"}
.add_indices("i", 0, input_shape.size())
.add_dims("I", 0, input_shape.size()))
.set_lhs(builder::ContractionInput{"CastI"}.add_indices("i", 0, input_shape.size()))
.set_rhs(neg_zp_input))
.add(builder::BinaryContraction{"=", "*"}
.set(builder::ContractionOutput{"O"}
.add_indices("i", 0, input_shape.size())
.add_dims("I", 0, input_shape.size()))
.set_lhs(
builder::ContractionInput{"Offset"}.add_indices("i", 0, input_shape.size()))
.set_rhs(scale_input));
set_output(f.finalize());
}
void ngraph::runtime::plaidml::ImplQuantize::Apply()
{
check_inputs(3);
check_outputs(1);
const auto& type = op().get_output_element_type(0);
const auto& axes = op().get_axes();
const auto& round_mode = op().get_round_mode();
const auto& input_shape = op().get_input_shape(0);
const auto& scale_shape = op().get_input_shape(1);
const auto& zp_shape = op().get_input_shape(2);
std::function<std::string(std::string)> cast_to_output_type = [type](std::string tensor_name) {
std::ostringstream cast_str;
if (type.is_signed())
{
cast_str << "as_int";
}
else
{
cast_str << "as_uint";
}
cast_str << "(" << tensor_name << ", " << 8 * type.size() << ")";
return cast_str.str();
};
if (scale_shape != zp_shape)
{
throw std::runtime_error("Quantize given mismatched scale & zero point shapes.");
}
if (scale_shape.size() != axes.size())
{
std::ostringstream msg;
msg << "Quantize received " << axes.size()
<< " axes to use for scale & zero point, but those tensors have " << scale_shape.size()
<< " dimensions instead.";
throw std::runtime_error(msg.str());
}
std::vector<std::string> short_idxs;
for (size_t i = 0; i < input_shape.size(); ++i)
{
if (axes.count(i))
{
std::ostringstream name;
name << "i" << i;
short_idxs.push_back(name.str());
}
}
if (!type.is_integral())
{
throw std::runtime_error("Quantize output type must be integral");
}
builder::Elementwise Rounded{"Rounded", ""};
builder::Elementwise Clamped{"Clamped", ""};
builder::Elementwise O{"O", ""};
int64_t q_min;
int64_t q_max;
std::ostringstream clamp_formula;
if (type.size() > 4)
{
// PlaidML doesn't support quantization clamping for types wider than 32 bits
if (!type.is_signed())
{
clamp_formula << "Uncast < 0 ? 0 : Uncast";
}
else
{
clamp_formula << "Uncast";
}
}
else
{
if (type.is_signed())
{
q_max = (1 << (8 * type.size() - 1)) - 1;
q_min = -q_max - 1;
}
else
{
q_max = (1 << (8 * type.size())) - 1;
q_min = 0;
}
clamp_formula << "Uncast < " << q_min << " ? " << q_min << " : "
<< "(Uncast > " << q_max << " ? " << q_max << " : Uncast)";
}
Clamped.set_rhs(clamp_formula.str());
std::ostringstream round_formula;
std::string lower_rounded_int;
switch (round_mode)
{
case ngraph::op::Quantize::RoundMode::ROUND_DOWN: Rounded.set_rhs("floor(Frac)"); break;
case ngraph::op::Quantize::RoundMode::ROUND_UP: Rounded.set_rhs("ceil(Frac)"); break;
case ngraph::op::Quantize::RoundMode::ROUND_NEAREST_DOWNWARD:
Rounded.set_rhs("ceil(Frac - 0.5)");
break;
case ngraph::op::Quantize::RoundMode::ROUND_NEAREST_UPWARD:
Rounded.set_rhs("floor(Frac + 0.5)");
break;
case ngraph::op::Quantize::RoundMode::ROUND_TOWARD_ZERO:
Rounded.set_rhs("Frac > 0 ? floor(Frac) : ceil(Frac)");
break;
case ngraph::op::Quantize::RoundMode::ROUND_TOWARD_INFINITY:
Rounded.set_rhs("Frac < 0 ? floor(Frac) : ceil(Frac)");
break;
case ngraph::op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_ZERO:
Rounded.set_rhs("Frac > 0 ? ceil(Frac - 0.5) : floor(Frac + 0.5)");
break;
case ngraph::op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_INFINITY:
Rounded.set_rhs("Frac < 0 ? ceil(Frac - 0.5) : floor(Frac + 0.5)");
break;
case ngraph::op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN:
// This is ugly, but it produces correct output
lower_rounded_int = cast_to_output_type("ceil(Frac - 0.5)");
round_formula << "2 * (" << lower_rounded_int << " / 2) == " << lower_rounded_int
<< " ? ceil(Frac - 0.5) : floor(Frac + 0.5)";
Rounded.set_rhs(round_formula.str());
break;
default:
throw std::runtime_error("Requested quantize round mode not yet implemented in PlaidML");
}
O.set_rhs(cast_to_output_type("Clamped"));
builder::ContractionInput scale_recip_input{"SRecip"};
builder::ContractionInput zp_input{"Z"};
for (const auto& idx : short_idxs)
{
scale_recip_input.add_indices({idx});
zp_input.add_indices({idx});
}
auto f = start_tile_function();
f.add(builder::Input{op_input(0), "I"}.add_dims("I", 0, input_shape.size()))
.add(builder::Input{op_input(1), "S"}.add_dims("S", 0, scale_shape.size()))
.add(builder::Input{op_input(2), "Z"}.add_dims("Z", 0, zp_shape.size()))
.add(builder::Output{"O"})
.add(builder::Elementwise{"SRecip", "1 / S"})
.add(builder::BinaryContraction{"=", "*"}
.set(builder::ContractionOutput{"Frac"}
.add_indices("i", 0, input_shape.size())
.add_dims("I", 0, input_shape.size()))
.set_lhs(builder::ContractionInput{"I"}.add_indices("i", 0, input_shape.size()))
.set_rhs(scale_recip_input))
.add(Rounded)
.add(builder::BinaryContraction{"=", "+"}
.set(builder::ContractionOutput{"Uncast"}
.add_indices("i", 0, input_shape.size())
.add_dims("I", 0, input_shape.size()))
.set_lhs(
builder::ContractionInput{"Rounded"}.add_indices("i", 0, input_shape.size()))
.set_rhs(zp_input))
.add(Clamped)
.add(O);
set_output(f.finalize());
}
src/ngraph/runtime/plaidml/unit_test.manifest
View file @ 97a9dbb3
......@@ -60,37 +60,7 @@ generate_mask
generate_mask2
avg_pool_3d
avg_pool_3d_uneven_strided_padded_include_in_computation
quantize_dynamic_offset # Quantization/Dequantization is unimplemented
dequantize_dynamic_offset # Quantization/Dequantization is unimplemented
dequantize_int8_zero_offset # Quantization/Dequantization is unimplemented
dequantize_int32 # Quantization/Dequantization is unimplemented
dequantize_int32_zero_offset # Quantization/Dequantization is unimplemented
dequantize_zero_offset # Quantization/Dequantization is unimplemented
quantize_ROUND_NEAREST_TOWARD_ZERO # Quantization/Dequantization is unimplemented
quantize_ROUND_NEAREST_UPWARD # Quantization/Dequantization is unimplemented
quantize_ROUND_NEAREST_DOWNWARD # Quantization/Dequantization is unimplemented
quantize_ROUND_NEAREST_TOWARD_EVEN # Quantization/Dequantization is unimplemented
quantize_ROUND_NEAREST_TOWARD_INFINITY # Quantization/Dequantization is unimplemented
quantize_ROUND_TOWARD_INFINITY # Quantization/Dequantization is unimplemented
quantize_ROUND_TOWARD_ZERO # Quantization/Dequantization is unimplemented
quantize_ROUND_UP # Quantization/Dequantization is unimplemented
quantize_ROUND_DOWN # Quantization/Dequantization is unimplemented
quantize # Quantization/Dequantization is unimplemented
quantize_zero_offset # Quantization/Dequantization is unimplemented
quantize_axes # Quantization/Dequantization is unimplemented
quantize_dynamic_offset # Quantization/Dequantization is unimplemented
quantize_int8 # Quantization/Dequantization is unimplemented
quantize_int8_zero_offset # Quantization/Dequantization is unimplemented
quantize_int32 # Quantization/Dequantization is unimplemented
quantize_int32_zero_offset # Quantization/Dequantization is unimplemented
quantize_clamp # Quantization/Dequantization is unimplemented
quantize_clamp_int8 # Quantization/Dequantization is unimplemented
quantize_clamp_int32 # Quantization/Dequantization is unimplemented
quantize_clamp_int32_zero_offset # Quantization/Dequantization is unimplemented
quantize_clamp_uint8 # Quantization/Dequantization is unimplemented
dequantize # Quantization/Dequantization is unimplemented
dequantize_axes # Quantization/Dequantization is unimplemented
dequantize_int8 # Quantization/Dequantization is unimplemented
quantize_clamp_int32 # Requires fp64 inputs, which won't work on GPUs
numeric_float_nan
numeric_double_nan
shape_of_scalar
......
test/models/onnx/lstm_fwd_large_batch_no_clip.prototxt 0 → 100644
View file @ 97a9dbb3
ir_version: 4
producer_name: "nGraph ONNX Importer"
graph {
node {
input: "X"
input: "W"
input: "R"
output: ""
output: "Y_h"
op_type: "LSTM"
attribute {
name: "clip"
f: 9999.0
type: FLOAT
}
attribute {
name: "direction"
s: "forward"
type: STRING
}
attribute {
name: "hidden_size"
i: 3
type: INT
}
}
name: "compute_graph"
input {
name: "X"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 2
}
dim {
dim_value: 32
}
dim {
dim_value: 1
}
}
}
}
}
input {
name: "W"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 1
}
dim {
dim_value: 12
}
dim {
dim_value: 1
}
}
}
}
}
input {
name: "R"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 1
}
dim {
dim_value: 12
}
dim {
dim_value: 3
}
}
}
}
}
output {
name: "Y_h"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 1
}
dim {
dim_value: 32
}
dim {
dim_value: 3
}
}
}
}
}
}
opset_import {
version: 7
}
test/onnx/onnx_import_rnn.in.cpp
View file @ 97a9dbb3
......@@ -20,6 +20,7 @@
#include <fstream>
#include <iterator>
#include <limits>
#include <numeric>
#include <sstream>
#include <stdexcept>
#include <vector>
......@@ -203,3 +204,48 @@ NGRAPH_TEST(onnx_${BACKEND_NAME}, model_lstm_fwd_hardsigmoid_activation)
test_case.set_tolerance(6);
test_case.run();
}
NGRAPH_TEST(onnx_${BACKEND_NAME}, model_lstm_fwd_large_batch_no_clip)
{
auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/lstm_fwd_large_batch_no_clip.prototxt"));
auto test_case = ngraph::test::NgraphTestCase(function, "${BACKEND_NAME}");
std::size_t seq_length = 2;
std::size_t batch_size = 32;
std::size_t input_size = 1;
std::size_t hidden_size = 3;
std::vector<float> in_X(seq_length * batch_size * input_size);
std::iota(std::begin(in_X), std::end(in_X), 1.f);
std::vector<float> in_R(4 * hidden_size * hidden_size, 0.1f);
// X
test_case.add_input<float>(in_X);
// W
test_case.add_input<float>(
{0.1f, 0.2f, 0.3f, 0.4f, 1.f, 2.f, 3.f, 4.f, 10.f, 11.f, 12.f, 13.f});
// R
test_case.add_input<float>(in_R);
// Y_h_data
test_case.add_expected_output<float>(
Shape{1, batch_size, hidden_size},
{0.90387899f, 0.9135572f, 0.91772245f, 0.90897038f, 0.92132433f, 0.92825467f, 0.91365823f,
0.92815113f, 0.93676105f, 0.91799162f, 0.93406357f, 0.94344562f, 0.92199681f, 0.93912057f,
0.94859476f, 0.92569357f, 0.94340185f, 0.95250664f, 0.92909964f, 0.94699686f, 0.95545127f,
0.93223207f, 0.94999634f, 0.95765468f, 0.93510761f, 0.9524867f, 0.95929726f, 0.93774272f,
0.9545467f, 0.96051891f, 0.9401536f, 0.95624603f, 0.96142619f, 0.94235605f, 0.95764499f,
0.96209939f, 0.94436539f, 0.95879495f, 0.96259862f, 0.94619635f, 0.95973921f, 0.96296872f,
0.94786299f, 0.96051397f, 0.96324302f, 0.94937864f, 0.96114929f, 0.96344629f, 0.95075587f,
0.96167006f, 0.96359692f, 0.95200645f, 0.96209679f, 0.96370852f, 0.95314133f, 0.9624464f,
0.9637912f, 0.95417069f, 0.96273278f, 0.96385246f, 0.95510395f, 0.96296733f, 0.96389785f,
0.95594975f, 0.96315942f, 0.96393147f, 0.95671607f, 0.96331673f, 0.96395638f, 0.9574102f,
0.96344554f, 0.96397483f, 0.9580388f, 0.96355102f, 0.9639885f, 0.95860795f, 0.96363739f,
0.96399863f, 0.95912322f, 0.96370811f, 0.96400613f, 0.95958963f, 0.96376601f, 0.96401169f,
0.96001179f, 0.96381342f, 0.96401581f, 0.96039386f, 0.96385224f, 0.96401886f, 0.96073964f,
0.96388402f, 0.96402112f, 0.96105254f, 0.96391004f, 0.96402279f});
test_case.run();
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment