From d65fecc2f3a50f324152006284be2530c2443995 Mon Sep 17 00:00:00 2001
From: zengxianglong <zengxianglong1@huawei.com>
Date: Fri, 25 Dec 2020 16:00:52 +0800
Subject: [PATCH] fix some fp16 subgraph bugs

---
 mindspore/lite/src/ops/arithmetic_compare.cc           | 10 +++-------
 mindspore/lite/src/runtime/kernel/arm/fp32/pad_fp32.cc |  6 ++++--
 .../lite/src/runtime/kernel/arm/fp32/shape_fp32.cc     |  1 +
 mindspore/lite/src/sub_graph_kernel.cc                 |  3 +++
 mindspore/lite/test/models_onnx.cfg                    |  1 +
 mindspore/lite/test/models_tflite_fp16.cfg             |  6 +++++-
 6 files changed, 17 insertions(+), 10 deletions(-)

diff --git a/mindspore/lite/src/ops/arithmetic_compare.cc b/mindspore/lite/src/ops/arithmetic_compare.cc
index 661883b98a..57c3db7f74 100644
--- a/mindspore/lite/src/ops/arithmetic_compare.cc
+++ b/mindspore/lite/src/ops/arithmetic_compare.cc
@@ -21,13 +21,9 @@ namespace lite {
 
 int ArithmeticCompare::InferShape(std::vector<Tensor *> inputs_, std::vector<Tensor *> outputs_) {
   auto res = Arithmetic::InferShape(inputs_, outputs_);
-  if (res == RET_OK) {
-    auto output = outputs_.front();
-    output->set_data_type(TypeId::kNumberTypeBool);
-    return RET_OK;
-  } else {
-    return res;
-  }
+  auto output = outputs_.front();
+  output->set_data_type(TypeId::kNumberTypeBool);
+  return res;
 }
 }  // namespace lite
 }  // namespace mindspore
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/pad_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/pad_fp32.cc
index f0ca63ff5e..6515c2a79e 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/pad_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/pad_fp32.cc
@@ -304,10 +304,12 @@ int PadCPUKernel::CheckPaddings(int *paddings, int length, int *input_shape, int
   for (auto i = 0; i < length; ++i) {
     int max_valid = input_shape[i] - offset;
     if (paddings[i * 2] > max_valid) {
-      MS_LOG(WARNING) << prefix << "paddings " << paddings[i * 2] << "should be less than " << max_valid + 1;
+      MS_LOG(ERROR) << prefix << "paddings " << paddings[i * 2] << "should be less than " << max_valid + 1;
+      return RET_ERROR;
     }
     if (paddings[i * 2 + 1] > max_valid) {
-      MS_LOG(WARNING) << prefix << "paddings " << paddings[i * 2 + 1] << "should be less than " << max_valid + 1;
+      MS_LOG(ERROR) << prefix << "paddings " << paddings[i * 2 + 1] << "should be less than " << max_valid + 1;
+      return RET_ERROR;
     }
   }
   return RET_OK;
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/shape_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/shape_fp32.cc
index 452c590454..5cd4b8c438 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/shape_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/shape_fp32.cc
@@ -77,5 +77,6 @@ kernel::LiteKernel *CpuShapeFp32KernelCreator(const std::vector<lite::Tensor *>
 }
 
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Shape, CpuShapeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Shape, CpuShapeFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Shape, CpuShapeFp32KernelCreator)
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/sub_graph_kernel.cc b/mindspore/lite/src/sub_graph_kernel.cc
index 76c77c0429..b4cc23e3f2 100644
--- a/mindspore/lite/src/sub_graph_kernel.cc
+++ b/mindspore/lite/src/sub_graph_kernel.cc
@@ -238,6 +238,9 @@ int CpuFp16SubGraph::PreProcess() {
   }
   for (auto kernel : this->nodes_) {
     for (auto tensor : kernel->out_tensors()) {
+      if (kernel->Type() == schema::PrimitiveType_Cast) {
+        continue;
+      }
       if (tensor->data_type() == kNumberTypeFloat32) {
         tensor->set_data_type(kNumberTypeFloat16);
       }
diff --git a/mindspore/lite/test/models_onnx.cfg b/mindspore/lite/test/models_onnx.cfg
index 039bcf498c..ba8c614ccf 100644
--- a/mindspore/lite/test/models_onnx.cfg
+++ b/mindspore/lite/test/models_onnx.cfg
@@ -36,3 +36,4 @@ udnie-9.onnx
 adversarial_pruning.onnx
 residual_distill_res34_cifar10_bs_1_update.onnx
 residual_distill_res50_cifar10_bs_1_update.onnx
+encoder.onnx;1,32,83
diff --git a/mindspore/lite/test/models_tflite_fp16.cfg b/mindspore/lite/test/models_tflite_fp16.cfg
index c337f9041b..5309061db0 100644
--- a/mindspore/lite/test/models_tflite_fp16.cfg
+++ b/mindspore/lite/test/models_tflite_fp16.cfg
@@ -40,7 +40,11 @@ deeplabv3_257_mv_gpu.tflite 3
 multi_person_mobilenet_v1_075_float.tflite 9
 #hiai_vad.tflite 20
 ide_label_base.tflite 22
-#ide_label_retrained.tflite 80
+# ide_label_retrained.tflite involves a softmax-like structure whose output channel is 12.
+# The values in the first few channels are extremely small and casted into 0 in the fp16 subgraph.
+# The next slice operator takes out data from the first 6 channels, which are all zero, causing the error of
+# dividing 0 in the following operator.
+#ide_label_retrained.tflite
 ml_ei_headpose.tflite 3
 ml_ei_landmark.tflite 3
 mnist.tflite 4