fix compile error and isolate opencl for deconv

mace/kernels/deconv_2d.h

@@ -23,12 +23,15 @@
 #include <vector>
 
 #include "mace/core/future.h"
-#include "mace/core/runtime/opencl/cl2_header.h"
 #include "mace/core/tensor.h"
 #include "mace/kernels/activation.h"
 #include "mace/kernels/conv_pool_2d_util.h"
 #include "mace/utils/utils.h"
 
+#ifdef MACE_ENABLE_OPENCL
+#include "mace/core/runtime/opencl/cl2_header.h"
+#endif  // MACE_ENABLE_OPENCL
+
 namespace mace {
 namespace kernels {
 
@@ -126,7 +129,6 @@ struct Deconv2dFunctorBase {
 
     const index_t in_height = isNCHW ? input_shape[2] : input_shape[1];
     const index_t in_width = isNCHW ? input_shape[3] : input_shape[2];
-    const index_t in_channels = isNCHW ? input_shape[1] : input_shape[3];
 
     const index_t extended_input_height =
         (in_height - 1) * strides[0] + 1 + padding_size[0];
@@ -168,11 +170,9 @@ struct Deconv2dFunctorBase {
 
     const index_t in_height = isNCHW ? input_shape[2] : input_shape[1];
     const index_t in_width = isNCHW ? input_shape[3] : input_shape[2];
-    const index_t in_channels = isNCHW ? input_shape[1] : input_shape[3];
 
     const index_t out_height = isNCHW ? output_shape[2] : output_shape[1];
     const index_t out_width = isNCHW ? output_shape[3] : output_shape[2];
-    const index_t out_channels = isNCHW ? output_shape[1] : output_shape[3];
 
     const index_t extended_input_height = (in_height - 1) * strides[0] + 1;
     const index_t extended_input_width = (in_width - 1) * strides[1] + 1;
@@ -216,9 +216,9 @@ struct Deconv2dFunctorBase {
   const int *strides_;  // [stride_h, stride_w]
   const Padding padding_type_;
   std::vector<int> paddings_;
+  std::vector<index_t> output_shape_;
   const ActivationType activation_;
   const float relux_max_limit_;
-  std::vector<index_t> output_shape_;
 };
 
 template <DeviceType D, typename T>
@@ -228,9 +228,7 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
                   const std::vector<int> &paddings,
                   const std::vector<index_t> &output_shape,
                   const ActivationType activation,
-                  const float relux_max_limit,
-                  const bool is_filter_transformed,
-                  ScratchBuffer *scratch)
+                  const float relux_max_limit)
       : Deconv2dFunctorBase(strides,
                             padding_type,
                             paddings,
@@ -243,6 +241,7 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
                   const Tensor *bias,
                   Tensor *output,
                   StatsFuture *future) {
+    MACE_UNUSED(future);
     MACE_CHECK_NOTNULL(input);
     MACE_CHECK_NOTNULL(filter);
     MACE_CHECK_NOTNULL(output);
@@ -315,6 +314,7 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
   }
 };
 
+#ifdef MACE_ENABLE_OPENCL
 template <typename T>
 struct Deconv2dFunctor<DeviceType::GPU, T> : Deconv2dFunctorBase {
   Deconv2dFunctor(const int *strides,
@@ -322,9 +322,7 @@ struct Deconv2dFunctor<DeviceType::GPU, T> : Deconv2dFunctorBase {
                   const std::vector<int> &paddings,
                   const std::vector<index_t> &output_shape,
                   const ActivationType activation,
-                  const float relux_max_limit,
-                  const bool is_filter_transformed,
-                  ScratchBuffer *scratch)
+                  const float relux_max_limit)
       : Deconv2dFunctorBase(strides,
                             padding_type,
                             paddings,
@@ -343,6 +341,7 @@ struct Deconv2dFunctor<DeviceType::GPU, T> : Deconv2dFunctorBase {
   std::unique_ptr<BufferBase> kernel_error_;
   std::vector<index_t> input_shape_;
 };
+#endif  // MACE_ENABLE_OPENCL
 
 }  // namespace kernels
 }  // namespace mace

mace/ops/deconv_2d.cc

@@ -24,6 +24,7 @@ void Register_Deconv2D(OperatorRegistry *op_registry) {
                                      .Build(),
                     Deconv2dOp<DeviceType::CPU, float>);
 
+#ifdef MACE_ENABLE_OPENCL
   REGISTER_OPERATOR(op_registry, OpKeyBuilder("Deconv2D")
                                      .Device(DeviceType::GPU)
                                      .TypeConstraint<float>("T")
@@ -35,6 +36,7 @@ void Register_Deconv2D(OperatorRegistry *op_registry) {
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     Deconv2dOp<DeviceType::GPU, half>);
+#endif  // MACE_ENABLE_OPENCL
 }
 
 }  // namespace ops

mace/ops/deconv_2d.h

@@ -34,10 +34,7 @@ class Deconv2dOp : public ConvPool2dOpBase<D, T> {
                  this->paddings_,
                  OperatorBase::GetRepeatedArgument<index_t>("output_shape"),
                  kernels::ActivationType::NOOP,
-                 0.0f,
-                 static_cast<bool>(OperatorBase::GetSingleArgument<int>(
-                     "is_filter_transformed", false)),
-                 ws->GetScratchBuffer(D)) {}
+                 0.0f) {}
 
   bool Run(StatsFuture *future) override {
     const Tensor *input = this->Input(INPUT);

mace/python/tools/tensor_data.jinja2

@@ -17,6 +17,7 @@
 #include <vector>
 #include <string>
 
+#include "mace/core/macros.h"
 #include "mace/public/mace.h"
 #include "mace/utils/env_time.h"
 #include "mace/utils/logging.h"
@@ -42,6 +43,7 @@ alignas(4) const unsigned char model_data[{{ model_data_size }}] = {
 
 const unsigned char *LoadModelData(const char *model_data_file) {
 {% if embed_model_data %}
+  MACE_UNUSED(model_data_file);
   return model_data;
 {% else %}
   int fd = open(model_data_file, O_RDONLY);
@@ -67,6 +69,8 @@ void UnloadModelData(const unsigned char *model_data) {
   int ret = munmap(const_cast<unsigned char *>(model_data),
                    {{ model_data_size }});
   MACE_CHECK(ret == 0, "Failed to unmap model data file, error code: ", errno);
+{% else %}
+  MACE_UNUSED(model_data);
 {% endif %}
 }