Add 4D transpose for weights (onnx#557)

Signed-off-by: Kevin Chen <[email protected]>

Author: kevinch-nv

ShapedWeights.cpp

@@ -95,43 +95,81 @@ ShapedWeights::operator nvinfer1::Weights() const
 }
 
 template <typename DType>
-void transpose2DWeights(ShapedWeights const& weights, nvinfer1::Dims const& new_shape, ShapedWeights* result)
+void transpose4DWeights(ShapedWeights const& weights, nvinfer1::Permutation const perm, ShapedWeights* result)
 {
+    nvinfer1::Dims original_shape = weights.shape;
+    nvinfer1::Dims new_shape = result->shape;
+    int nbDims = new_shape.nbDims;
     DType const* src = reinterpret_cast<DType*>(weights.values);
     DType* dst = reinterpret_cast<DType*>(result->values);
-    int src_stride = weights.shape.d[1];
-    int dst_stride = result->shape.d[1];
-    for (int i = 0; i < new_shape.d[0]; ++i)
+
+    nvinfer1::Dims expanded_original_shape{4, {1, 1, 1, 1}};
+    nvinfer1::Dims expanded_new_shape{4, {1, 1, 1, 1}};
+    nvinfer1::Permutation expanded_perm{0, 1, 2, 3};
+
+    int pad = 4 - nbDims;
+    for (int i = 0; i < nbDims; ++i)
+    {
+        expanded_original_shape.d[pad + i] = original_shape.d[i];
+        expanded_new_shape.d[pad + i] = new_shape.d[i];
+        expanded_perm.order[pad + i] = perm.order[i] + pad;
+    }
+
+    int src_strides[4] = {1, 1, 1, 1};
+    int dst_strides[4] = {1, 1, 1, 1};
+
+    for (int i = 2; i >= 0; --i)
+    {
+        src_strides[i] = expanded_original_shape.d[i + 1] * src_strides[i + 1];
+        dst_strides[i] = expanded_new_shape.d[i + 1] * dst_strides[i + 1];
+    }
+
+    for (int n = 0; n < expanded_original_shape.d[0]; ++n)
     {
-        for (int j = 0; j < new_shape.d[1]; ++j)
+        for (int c = 0; c < expanded_original_shape.d[1]; ++c)
         {
-            dst[i * dst_stride + j] = src[j * src_stride + i];
+            for (int h = 0; h < expanded_original_shape.d[2]; ++h)
+            {
+                for (int w = 0; w < expanded_original_shape.d[3]; ++w)
+                {
+                    int src_index = 0;
+                    int dst_index = 0;
+                    int src_coord[4] = {n, c, h, w};
+                    int dst_coord[4];
+                    for (int i  = 0 ; i < 4; ++i)
+                    {
+                        dst_coord[i] = src_coord[expanded_perm.order[i]];
+                        src_index += src_coord[i] * src_strides[i];
+                        dst_index += dst_coord[i] * dst_strides[i];
+                    }
+                    dst[dst_index] = src[src_index];
+                }
+            }
         }
     }
 }
 
 bool transposeWeights(ShapedWeights const& weights, nvinfer1::Permutation const& perm, ShapedWeights* result)
 {
     nvinfer1::Dims shape = weights.shape;
+    int nbDims = shape.nbDims;
     nvinfer1::Dims new_shape;
-    new_shape.nbDims = shape.nbDims;
-    for (int d = 0; d < shape.nbDims; ++d)
+    new_shape.nbDims = nbDims;
+    for (int d = 0; d < nbDims; ++d)
     {
         new_shape.d[d] = shape.d[perm.order[d]];
         result->shape.d[d] = new_shape.d[d];
     }
-    // TODO: Need to generalize this transpose implementation
-    assert(perm.order[0] == 1 && perm.order[1] == 0);
 
-    if (shape.nbDims == 2)
+    if (shape.nbDims <= 4)
     {
         if (weights.type == ::ONNX_NAMESPACE::TensorProto::FLOAT)
         {
-            transpose2DWeights<float>(weights, new_shape, result);
+            transpose4DWeights<float>(weights, perm, result);
         }
         else if (weights.type == ::ONNX_NAMESPACE::TensorProto::FLOAT16)
         {
-            transpose2DWeights<uint16_t>(weights, new_shape, result);
+            transpose4DWeights<uint16_t>(weights, perm, result);
         }
         else
         {