Add group parameter for deconv API

Limitation: only support depthwise deconvolution

Signed-off-by: xiang.zhang <xiang.zhang@verisilicon.com>

include/tim/vx/ops/deconv.h

@@ -34,23 +34,25 @@ namespace ops {
 
 class DeConv2d : public Operation {
   public:
-    DeConv2d(Graph* graph, int32_t weights, PadType pad_type,
+    DeConv2d(Graph* graph, int32_t oc_count_, PadType pad_type,
         const std::array<uint32_t, 2>& ksize,
         const std::array<uint32_t, 2>& stride,
         const std::array<uint32_t, 2>& output_padding);
-    DeConv2d(Graph* graph, int32_t weights, PadType pad_type,
+    DeConv2d(Graph* graph, int32_t oc_count_, PadType pad_type,
         const std::array<uint32_t, 2>& ksize,
         const std::array<uint32_t, 2>& stride,
         const std::array<uint32_t, 2>& output_padding,
-        const std::array<uint32_t, 4>& pad);
+        const std::array<uint32_t, 4>& pad,
+        const uint32_t group = 1);
 
   protected:
-    const uint32_t weights_;
+    const uint32_t oc_count_; // output channel count
     const PadType pad_type_;
     const std::array<uint32_t, 2> ksize_;
     const std::array<uint32_t, 2> stride_;
     const std::array<uint32_t, 2> output_padding_;
     const std::array<uint32_t, 4> pad_;
+    const uint32_t group_;
 };
 
 } // namespace ops

src/tim/CMakeLists.txt

@@ -30,9 +30,6 @@ if(TIM_VX_ENABLE_LAYOUT_INFER)
     )
 endif()
 
-set(UT_SRC)
-aux_source_directory(./vx/ut VX_UT_SRC)
-list(APPEND UT_SRC ${VX_UT_SRC})
 foreach(src_file ${SRC})
     if(${src_file} MATCHES ".*_test\.cc")
         list(REMOVE_ITEM SRC ${src_file})

src/tim/vx/ops/deconv.cc

@@ -21,9 +21,10 @@
 *    DEALINGS IN THE SOFTWARE.
 *
 *****************************************************************************/
-
 #include "tim/vx/ops/deconv.h"
 
+#include <cassert>
+
 #include "operation_private.h"
 #include "type_utils.h"
 #include "vsi_nn_pub.h"
@@ -32,33 +33,38 @@ namespace tim {
 namespace vx {
 namespace ops {
 
-DeConv2d::DeConv2d(Graph* graph, int32_t weights, PadType pad_type,
+DeConv2d::DeConv2d(Graph* graph, int32_t oc_count, PadType pad_type,
     const std::array<uint32_t, 2>& ksize,
     const std::array<uint32_t, 2>& stride,
     const std::array<uint32_t, 2>& output_padding)
-  : DeConv2d(graph, weights, pad_type, ksize, stride, output_padding,
+  : DeConv2d(graph, oc_count, pad_type, ksize, stride, output_padding,
       {0, 0, 0, 0}) {
 }
 
-DeConv2d::DeConv2d(Graph* graph, int32_t weights, PadType pad_type,
+DeConv2d::DeConv2d(Graph* graph, int32_t oc_count, PadType pad_type,
     const std::array<uint32_t, 2>& ksize,
     const std::array<uint32_t, 2>& stride,
     const std::array<uint32_t, 2>& output_padding,
-    const std::array<uint32_t, 4>& pad)
+    const std::array<uint32_t, 4>& pad,
+    const uint32_t group)
   : Operation(graph, VSI_NN_OP_DECONVOLUTION),
-    weights_(weights),
+    oc_count_(oc_count),
     pad_type_(pad_type),
     ksize_(ksize),
     stride_(stride),
     output_padding_(output_padding),
-    pad_(pad) {
+    pad_(pad),
+    group_(group) {
+
+  // TODO(Sven): only support depthwise usage
+  assert(group != 1 && group == oc_count);
   this->impl()->node()->nn_param.deconv.ksize[0] = ksize_[0];
   this->impl()->node()->nn_param.deconv.ksize[1] = ksize_[1];
   this->impl()->node()->nn_param.deconv.stride[0] = stride_[0];
   this->impl()->node()->nn_param.deconv.stride[1] = stride_[1];
   this->impl()->node()->nn_param.deconv.pad_type = TranslatePadType(pad_type_);
-  this->impl()->node()->nn_param.deconv.weights = weights_;
+  this->impl()->node()->nn_param.deconv.weights = oc_count_;
-  this->impl()->node()->nn_param.deconv.group = 1;
+  this->impl()->node()->nn_param.deconv.group = group_;
   this->impl()->node()->nn_param.deconv.output_padding[0] = output_padding_[0];
   this->impl()->node()->nn_param.deconv.output_padding[1] = output_padding_[1];
   this->impl()->node()->nn_param.deconv.pad[0] = pad_[0];

src/tim/vx/ops/deconv2d_test.cc

@@ -0,0 +1,86 @@
+#include "tim/vx/context.h"
+#include "tim/vx/graph.h"
+#include "tim/vx/ops/deconv.h"
+
+#include "gtest/gtest.h"
+
+namespace {
+
+size_t element_count(const tim::vx::ShapeType& shape) {
+  size_t sz = 1;
+  for (auto d : shape) {
+    sz *= d;
+  }
+
+  return sz;
+}
+
+}  // namespace
+
+TEST(OP, deconv_group) {
+    auto ctx = tim::vx::Context::Create();
+    auto graph = ctx->CreateGraph();
+
+    tim::vx::ShapeType input_shape ({3, 3, 2, 1});  //whcn
+    tim::vx::ShapeType kernel_shape({3, 3, 2, 1});  //whc1 same as depthwise convolution
+    tim::vx::ShapeType output_shape({5, 5, 2, 1});  //whcn
+
+    tim::vx::TensorSpec input_spec  (tim::vx::DataType::FLOAT32, input_shape, tim::vx::TensorAttribute::INPUT);
+    tim::vx::TensorSpec kernel_spec (tim::vx::DataType::FLOAT32, kernel_shape, tim::vx::TensorAttribute::CONSTANT);
+    tim::vx::TensorSpec output_spec (tim::vx::DataType::FLOAT32, output_shape, tim::vx::TensorAttribute::OUTPUT);
+
+    auto input_tensor = graph->CreateTensor(input_spec);
+    auto output_tensor = graph->CreateTensor(output_spec);
+    auto kernel_tensor = graph->CreateTensor(kernel_spec);
+
+    std::vector<float> input_data = {3.0f, 8.0f, 1.0f,
+                                     9.0f, 5.0f, 7.0f,
+                                     3.0f, 2.0f, 3.0f,
+
+                                     7.0f, 9.0f, 1.0f,
+                                     5.0f, 2.0f, 3.0f,
+                                     9.0f, 0.0f, 2.0f};
+    std::vector<float> kernel_data =
+                                    {9.0f, 0.0f, 3.0f,
+                                     0.0f, 0.0f, 0.0f,
+                                     1.0f, 0.0f, 2.0f,
+
+                                    3.0f, 0.0f, 7.0f,
+                                    0.0f, 0.0f, 0.0f,
+                                    0.0f, 0.0f, 8.0f,
+                                    };
+
+    std::vector<float> output_data(element_count(output_shape));
+
+    EXPECT_TRUE(input_tensor->CopyDataToTensor(input_data.data(), input_data.size()*4));
+    EXPECT_TRUE(kernel_tensor->CopyDataToTensor(kernel_data.data(), kernel_data.size()*4));
+
+    auto add = graph->CreateOperation<tim::vx::ops::DeConv2d>(
+        2,
+        tim::vx::PadType::SAME,
+        std::array<uint32_t, 2>({3, 3}),    /*ksize*/
+        std::array<uint32_t, 2>({1, 1}),    /*stride*/
+        std::array<uint32_t, 2>({1, 1}),    /*dilation*/
+        std::array<uint32_t, 4>({0, 0, 0, 0}), /*pad*/
+        2/*group*/);
+    (*add).BindInputs({input_tensor, kernel_tensor}).BindOutputs({output_tensor});
+
+    EXPECT_TRUE(graph->Compile());
+    EXPECT_TRUE(graph->Run());
+
+    EXPECT_TRUE(output_tensor->CopyDataFromTensor(output_data.data()));
+    std::vector<float> golden = {
+                        27.0f, 72.0f, 18.0f, 24.0f, 3.0f,
+                        81.0f, 45.0f, 90.0f, 15.0f, 21.0f,
+                        30.0f, 26.0f, 43.0f, 22.0f, 11.0f,
+                        9.0f, 5.0f, 25.0f, 10.0f, 14.0f,
+                        3.0f, 2.0f, 9.0f, 4.0f, 6.0f,
+
+                        21.0f, 27.0f, 52.0f, 63.0f, 7.0f,
+                        15.0f, 6.0f, 44.0f, 14.0f, 21.0f,
+                        27.0f, 0.0f, 125.0f, 72.0f, 22.0f,
+                        0.0f, 0.0f, 40.0f, 16.0f, 24.0f,
+                        0.0f, 0.0f, 72.0f, 0.0f, 16.0f};
+
+    EXPECT_EQ(golden, output_data) << "Result mismatch";
+}