fix bug of operator quantize

mindspore/lite/src/ops/primitive_c.h

@@ -39,7 +39,8 @@ constexpr uint32_t kDoubleNum = 2;
 constexpr uint32_t kMultiNum = 3;
 constexpr uint32_t kDimension_4d = 4;
 
-const std::set<int> kSupportDataType = {kNumberTypeUInt8, kNumberTypeInt32, kNumberTypeFloat32, kNumberTypeFloat16};
+const std::set<int> kSupportDataType = {kNumberTypeUInt8, kNumberTypeInt8, kNumberTypeInt32,
+                                        kNumberTypeFloat32, kNumberTypeFloat16};
 
 #ifdef PRIMITIVE_WRITEABLE
 using TensorPtr = std::shared_ptr<mindspore::tensor::Tensor>;

mindspore/lite/src/runtime/kernel/arm/fp32/cast.cc

@@ -147,6 +147,7 @@ kernel::LiteKernel *CpuCastFp32KernelCreator(const std::vector<lite::Tensor *> &
 
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Cast, CpuCastFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeUInt8, PrimitiveType_Cast, CpuCastFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Cast, CpuCastFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_Cast, CpuCastFp32KernelCreator)
 #ifndef ENABLE_ARM64
 REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Cast, CpuCastFp32KernelCreator)

mindspore/lite/tools/converter/parser/tflite/tflite_quantize_parser.cc

@@ -36,27 +36,37 @@ STATUS TfliteQuantizeParser::Parse(const std::unique_ptr<tflite::OperatorT> &tfl
     return RET_NULL_PTR;
   }
 
-  std::unique_ptr<schema::QuantDTypeCastT> attr = std::make_unique<schema::QuantDTypeCastT>();
-  if (attr == nullptr) {
-    MS_LOG(ERROR) << "new op failed";
-    return RET_NULL_PTR;
-  }
-
   const auto &in_tensor = tflite_tensors[tflite_op->inputs[0]];
   if (in_tensor == nullptr) {
     MS_LOG(ERROR) << "input tensor is null";
     return RET_NULL_PTR;
   }
-  attr->srcT = GetTfliteDataType(in_tensor->type);
   const auto &out_tensor = tflite_tensors[tflite_op->outputs[0]];
   if (out_tensor == nullptr) {
     MS_LOG(ERROR) << "output tensor is null";
     return RET_NULL_PTR;
   }
-  attr->dstT = GetTfliteDataType(out_tensor->type);
-
-  op->primitive->value.type = schema::PrimitiveType_QuantDTypeCast;
-  op->primitive->value.value = attr.release();
+  if (GetTfliteDataType(in_tensor->type) != kNumberTypeInt8) {
+    std::unique_ptr<schema::QuantDTypeCastT> attr = std::make_unique<schema::QuantDTypeCastT>();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new op failed";
+      return RET_NULL_PTR;
+    }
+    attr->srcT = GetTfliteDataType(in_tensor->type);
+    attr->dstT = GetTfliteDataType(out_tensor->type);
+    op->primitive->value.type = schema::PrimitiveType_QuantDTypeCast;
+    op->primitive->value.value = attr.release();
+  } else {
+    std::unique_ptr<schema::CastT> attr = std::make_unique<schema::CastT>();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new op failed";
+      return RET_NULL_PTR;
+    }
+    attr->srcT = GetTfliteDataType(in_tensor->type);
+    attr->dstT = GetTfliteDataType(out_tensor->type);
+    op->primitive->value.type = schema::PrimitiveType_Cast;
+    op->primitive->value.value = attr.release();
+  }
 
   AddOpInput(op, tensors_id, tensors_format, tensors_id_map, tflite_op->inputs[0], tensors_id->size(),
              tflite_tensors.size(), schema::Format_NHWC);