From 266965e2007f86817b7fd4ca4e3bbe141b2d47b8 Mon Sep 17 00:00:00 2001
From: zhangwen31 <edimetia3d@gmail.com>
Date: Mon, 14 Sep 2020 14:51:54 +0000
Subject: [PATCH] [arm][kernel][math] feat: add i32 support for elementwise div

---
 lite/backends/arm/math/elementwise.cc   | 27 +++++++++++++++++++++++++
 lite/kernels/arm/elementwise_compute.cc | 10 +++++++++
 2 files changed, 37 insertions(+)

diff --git a/lite/backends/arm/math/elementwise.cc b/lite/backends/arm/math/elementwise.cc
index eb3874d74e..658a235d62 100644
--- a/lite/backends/arm/math/elementwise.cc
+++ b/lite/backends/arm/math/elementwise.cc
@@ -44,6 +44,12 @@ static T naive_sub(T l, T r) {
   return l - r;
 }
 
+// todo: remove this function when all elementwise div works
+template <typename T>
+static T naive_div(T l, T r) {
+  return l / r;
+}
+
 // todo: use arm intrinsics
 template <>
 void elementwise_add<int32_t>(const int32_t* dinx,
@@ -1511,6 +1517,15 @@ void elementwise_max_relu_broadcast<float>(const float* dinx,
   }
 }
 
+// todo: use arm intrinsics
+template <>
+void elementwise_div<int32_t>(const int32_t* dinx,
+                              const int32_t* diny,
+                              int32_t* dout,
+                              int num) {
+  naive_elementwise_op<int32_t>(dinx, diny, dout, num, naive_div<int32_t>);
+}
+
 template <>
 void elementwise_div<int64_t>(const int64_t* dinx,
                               const int64_t* diny,
@@ -1576,6 +1591,18 @@ void elementwise_div<float>(const float* dinx,
   }
 }
 
+// todo: use arm intrinsics
+template <>
+void elementwise_div_broadcast<int32_t>(const int32_t* dinx,
+                                        const int32_t* diny,
+                                        int32_t* dout,
+                                        int batch,
+                                        int channels,
+                                        int num) {
+  naive_elementwise_op_broadcast<int32_t>(
+      dinx, diny, dout, batch, channels, num, naive_div<int32_t>);
+}
+
 template <>
 void elementwise_div_broadcast<int64_t>(const int64_t* dinx,
                                         const int64_t* diny,
diff --git a/lite/kernels/arm/elementwise_compute.cc b/lite/kernels/arm/elementwise_compute.cc
index cf38f76989..42cfa24715 100644
--- a/lite/kernels/arm/elementwise_compute.cc
+++ b/lite/kernels/arm/elementwise_compute.cc
@@ -548,6 +548,16 @@ REGISTER_LITE_KERNEL(
     .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM))})
     .Finalize();
 
+using elementwise_div_int32_t =
+    paddle::lite::kernels::arm::ElementwiseDivCompute<int32_t,
+                                                      PRECISION(kInt32)>;
+REGISTER_LITE_KERNEL(
+    elementwise_div, kARM, kInt32, kNCHW, elementwise_div_int32_t, def)
+    .BindInput("X", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt32))})
+    .BindInput("Y", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt32))})
+    .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt32))})
+    .Finalize();
+
 using elementwise_div_int64_t =
     paddle::lite::kernels::arm::ElementwiseDivCompute<int64_t,
                                                       PRECISION(kInt64)>;
-- 
GitLab