From 08254c538d686b1855245a0fb1203661f904a0de Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E6=9D=8E=E5=AF=85?= <liyin@xiaomi.com>
Date: Fri, 1 Sep 2017 13:51:13 +0800
Subject: [PATCH] Add relu neon

---
 mace/kernels/BUILD             | 22 ++++++++++++++++++++
 mace/kernels/neon/relu_neon.cc | 30 +++++++++++++++++++++++++++
 mace/kernels/neon/relu_neon.h  | 19 +++++++++++++++++
 mace/kernels/relu.h            | 28 +++++++++++++++++++++++++
 mace/ops/BUILD                 |  1 +
 mace/ops/relu.cc               | 38 +++++++++++++---------------------
 6 files changed, 114 insertions(+), 24 deletions(-)
 create mode 100644 mace/kernels/BUILD
 create mode 100644 mace/kernels/neon/relu_neon.cc
 create mode 100644 mace/kernels/neon/relu_neon.h
 create mode 100644 mace/kernels/relu.h

diff --git a/mace/kernels/BUILD b/mace/kernels/BUILD
new file mode 100644
index 00000000..a842356a
--- /dev/null
+++ b/mace/kernels/BUILD
@@ -0,0 +1,22 @@
+# Description:
+# Mace neon kernels.
+#
+package(
+    default_visibility = ["//visibility:public"],
+)
+
+
+licenses(["notice"])  # Apache 2.0
+
+load("//mace:mace.bzl", "if_android")
+
+cc_library(
+    name = "kernels",
+    srcs = glob(["*.cc"]) + if_android(glob(["neon/*.cc"])),
+    hdrs = glob(["*.h"]) + if_android(glob(["neon/*.h"])),
+    deps = [
+        "//mace/core:core",
+    ],
+    copts = ['-std=c++11'],
+)
+
diff --git a/mace/kernels/neon/relu_neon.cc b/mace/kernels/neon/relu_neon.cc
new file mode 100644
index 00000000..84542217
--- /dev/null
+++ b/mace/kernels/neon/relu_neon.cc
@@ -0,0 +1,30 @@
+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include <arm_neon.h>
+#include "mace/kernels/neon/relu_neon.h"
+
+namespace mace {
+namespace kernels{
+
+void NeonReluFuntion_float(const Tensor *input_tensor,
+                           Tensor *output_tensor) {
+  int64 size = input_tensor->size();
+  output_tensor->ResizeLike(input_tensor);
+  const float* input = input_tensor->data<float>();
+  float* output = output_tensor->mutable_data<float>();
+
+  float32x4_t _zero = vdupq_n_f32(0.f);
+  for (; size > 0; size--) {
+    float32x4_t _inp = vld1q_f32(input);
+    float32x4_t _outp = vmaxq_f32(_inp, _zero);
+    vst1q_f32(output, _outp);
+
+    input += 4;
+    output += 4;
+  }
+}
+
+} // namespace kernels
+} // namespace mace
\ No newline at end of file
diff --git a/mace/kernels/neon/relu_neon.h b/mace/kernels/neon/relu_neon.h
new file mode 100644
index 00000000..0be3be6f
--- /dev/null
+++ b/mace/kernels/neon/relu_neon.h
@@ -0,0 +1,19 @@
+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_KERNELS_RELU_NEON_H_
+#define MACE_KERNELS_RELU_NEON_H_
+
+#include "mace/core/tensor.h"
+
+namespace mace {
+namespace kernels {
+
+void NeonReluFuntion_float(const Tensor *input_tensor,
+                           Tensor *output_tensor);
+
+} // namespace kernels
+} // namespace mace
+
+#endif // MACE_KERNELS_RELU_NEON_H_
diff --git a/mace/kernels/relu.h b/mace/kernels/relu.h
new file mode 100644
index 00000000..cc613f1d
--- /dev/null
+++ b/mace/kernels/relu.h
@@ -0,0 +1,28 @@
+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_KERNELS_RELU_H_
+#define MACE_KERNELS_RELU_H_
+
+#include "mace/core/tensor.h"
+
+namespace mace {
+namespace kernels {
+
+template<typename T>
+void ReluFuntion(const Tensor *input_tensor, Tensor *output_tensor) {
+  int64 size = input_tensor->size();
+  output_tensor->ResizeLike(input_tensor);
+  const float* input = input_tensor->data<float>();
+  float* output = output_tensor->mutable_data<float>();
+
+  for (int64 i = 0; i < size; ++i) {
+    output[i] = std::max(input[i], static_cast<T>(0));
+  }
+}
+
+} //  namespace kernels
+} //  namespace mace
+
+#endif // MACE_KERNELS_RELU_H_
\ No newline at end of file
diff --git a/mace/ops/BUILD b/mace/ops/BUILD
index 7cde44fa..1acbc1fd 100644
--- a/mace/ops/BUILD
+++ b/mace/ops/BUILD
@@ -17,6 +17,7 @@ cc_library(
     deps = [
         "//mace/proto:cc_proto",
         "//mace/core:core",
+        "//mace/kernels:kernels",
     ],
     copts = ['-std=c++11'],
     alwayslink = 1,
diff --git a/mace/ops/relu.cc b/mace/ops/relu.cc
index 66662f38..59d4e3b7 100644
--- a/mace/ops/relu.cc
+++ b/mace/ops/relu.cc
@@ -4,42 +4,32 @@
 
 #include "mace/ops/relu.h"
 #include "mace/proto/mace.pb.h"
+#include "mace/kernels/relu.h"
+#if __ARM_NEON
+#include "mace/kernels/neon/relu_neon.h"
+#endif // __ARM_NEON
 
 namespace mace {
 
 template <>
 bool ReluOp<DeviceType::CPU, float>::Run() {
-  const Tensor* X = Input(0);
-  Tensor* Y = Output(0);
-  Y->ResizeLike(X);
-
-  const float* Xdata = X-> data<float>();
-  float* Ydata = Y->mutable_data<float>();
-  for (int i = 0; i < X->size(); ++i) {
-    Ydata[i] = std::max(Xdata[i], 0.f);
-    VLOG(0) << i << ": " << Xdata[i] << " " << Ydata[i];
-  }
-
+  const Tensor* input_tensor = Input(0);
+  Tensor* output_tensor = Output(0);
+  kernels::ReluFuntion<float>(input_tensor, output_tensor);
   return true;
 }
+REGISTER_CPU_OPERATOR(Relu, ReluOp<DeviceType::CPU, float>);
+
 
+#if __ARM_NEON
 template <>
 bool ReluOp<DeviceType::NEON, float>::Run() {
-  const Tensor* X = Input(0);
-  Tensor* Y = Output(0);
-  Y->ResizeLike(X);
-
-  const float* Xdata = X-> data<float>();
-  float* Ydata = Y->mutable_data<float>();
-  for (int i = 0; i < X->size(); ++i) {
-    Ydata[i] = std::max(Xdata[i], 0.f);
-    VLOG(0) << i << ": " << Xdata[i] << " " << Ydata[i];
-  }
-
+  const Tensor* input_tensor = Input(0);
+  Tensor* output_tensor = Output(0);
+  kernels::NeonReluFuntion_float(input_tensor, output_tensor);
   return true;
 }
-
-REGISTER_CPU_OPERATOR(Relu, ReluOp<DeviceType::CPU, float>);
 REGISTER_NEON_OPERATOR(Relu, ReluOp<DeviceType::NEON, float>);
+#endif // __ARM_NEON
 
 } //  namespace mace
-- 
GitLab