Merge branch 'main' into upstream_irmodule_parser_from_unity_pr_14487

Author: Lunderberg

docker/Dockerfile.ci_cpu

@@ -47,6 +47,9 @@ RUN bash /install/ubuntu_install_python_package.sh
 COPY install/ubuntu1804_install_llvm.sh /install/ubuntu1804_install_llvm.sh
 RUN bash /install/ubuntu1804_install_llvm.sh
 
+COPY install/ubuntu_install_llvm_from_source.sh /install/ubuntu_install_llvm_from_source.sh
+RUN bash /install/ubuntu_install_llvm_from_source.sh 15.0.7 8b5fcb24b4128cf04df1b0b9410ce8b1a729cb3c544e6da885d234280dedeac6
+
 COPY install/ubuntu_install_dnnl.sh /install/ubuntu_install_dnnl.sh
 RUN bash /install/ubuntu_install_dnnl.sh
 

docker/install/ubuntu_install_llvm_from_source.sh

@@ -60,6 +60,7 @@ cmake \
     -DLLVM_ENABLE_ASSERTIONS=ON \
     -DLLVM_ENABLE_RTTI=ON \
     -DLLVM_ENABLE_OCAMLDOC=OFF \
+    -DLLVM_ENABLE_PROJECTS=mlir \
     -DLLVM_USE_INTEL_JITEVENTS=ON \
     -DLLVM_TEMPORARILY_ALLOW_OLD_TOOLCHAIN=ON \
     -DPYTHON_EXECUTABLE="$(cpython_path 3.7)/bin/python" \

include/tvm/relay/attrs/nn.h

@@ -671,10 +671,10 @@ struct Conv1DTransposeAttrs : public tvm::AttrsNode<Conv1DTransposeAttrs> {
             "dimensions respectively. Convolution is applied on the"
             "'W' dimension.");
     TVM_ATTR_FIELD(kernel_layout)
-        .set_default("OIW")
+        .set_default("IOW")
         .describe(
-            "Dimension ordering of data and weight. Can be 'OIW', 'OIW16o16i', etc."
-            "'O', 'I', 'W' stands for num_filter, input_channel, and width"
+            "Dimension ordering of data and weight. Can be 'IOW', 'IOW16o16i', etc."
+            "'I', 'O', 'W' stands for input_channel, num_filter and width"
             "dimensions respectively.");
     TVM_ATTR_FIELD(out_layout)
         .set_default("")

include/tvm/tir/schedule/schedule.h

@@ -292,6 +292,16 @@ class ScheduleNode : public runtime::Object {
    */
   virtual Array<BlockRV> GetConsumers(const BlockRV& block_rv) = 0;
   /******** Schedule: Transform loops ********/
+  /*!
+   * \brief Merge a list of loops into one. The loops under their LCA requires:
+   * 1) Under the same scope
+   * 2) Can't have annotations or thread bindings
+   * 3) Start with 0 and have same extent and same nesting depth
+   * 4) From target loop to their LCA, the inner loop must be the only child of the outer loop
+   * \param loop_rvs The loops to be merged
+   * \return The new loop after merge
+   */
+  virtual LoopRV Merge(const Array<LoopRV>& loop_rvs) = 0;
   /*!
    * \brief Fuse a list of consecutive loops into one. It requires:
    * 1) The loops can't have annotations or thread bindings.
@@ -328,6 +338,12 @@ class ScheduleNode : public runtime::Object {
    * \param ordered_loop_rvs The loops in the new order
    */
   virtual void Reorder(const Array<LoopRV>& ordered_loop_rvs) = 0;
+  /*!
+   * \brief Reorder the itervars inside a block.
+   * \param block_rv The block to be transformed.
+   * \param new_order The new itervar order.
+   */
+  virtual void ReorderBlockIterVar(const BlockRV& block_rv, const Array<Integer> new_order) = 0;
   /*!
    * \brief Create a new unit loop on top of the specific block.
    * \param block_rv The block above which the new loop is created

python/tvm/relay/frontend/keras.py

@@ -282,6 +282,8 @@ def _convert_dense(
 
 
 def _convert_convolution1d(inexpr, keras_layer, etab, data_layout, input_shape=None):
+    is_deconv = type(keras_layer).__name__ == "Conv1DTranspose"
+
     if input_shape is None:
         input_shape = keras_layer.input_shape
     _check_data_format(keras_layer)
@@ -290,19 +292,21 @@ def _convert_convolution1d(inexpr, keras_layer, etab, data_layout, input_shape=N
 
     if data_layout == "NWC":
         kernel_layout = "WIO"
+        if is_deconv:
+            kernel_layout = "WOI"
     else:
         kernel_layout = "OIW"
+        if is_deconv:
+            kernel_layout = "IOW"
         msg = (
             "Kernel layout with {} is not supported for operator Convolution1D "
             "in frontend Keras."
         )
         raise tvm.error.OpAttributeUnImplemented(msg.format(data_layout))
 
-    is_deconv = type(keras_layer).__name__ == "Conv1DTranspose"
-
     if is_deconv:
-        if kernel_layout == "OIW":
-            weight = weight.transpose([2, 0, 1])
+        if kernel_layout == "IOW":
+            weight = weight.transpose([2, 1, 0])
         kernel_w, n_filters, _ = weight.shape
     else:
         kernel_w, _, n_filters = weight.shape

python/tvm/relay/frontend/mxnet.py

@@ -304,7 +304,7 @@ def _mx_conv1d_transpose(inputs, attrs):
     if data_layout != "NCW":
         raise tvm.error.OpAttributeInvalid('Only "NCW" data layout is supported for 1D Convolution')
     channel_axis = 1
-    kernel_layout = "OIW"
+    kernel_layout = "IOW"
     new_attrs = {}
     new_attrs["channels"] = attrs.get_int("num_filter")
     new_attrs["kernel_size"] = attrs.get_int_tuple("kernel")

python/tvm/relay/frontend/oneflow.py

@@ -84,8 +84,8 @@ def get_node_info(node):
     shape = tuple(node.input_conf.blob_conf.shape.dim)
     # get data type
     dtype = node.input_conf.blob_conf.data_type
-    if dtype in list(FLOW_2_NP_DTYPE.keys()):
-        data_type = FLOW_2_NP_DTYPE[dtype]
+    if dtype in list(FLOW_2_STR_DTYPE.keys()):
+        data_type = FLOW_2_STR_DTYPE[dtype]
     else:
         raise IndexError("Please check the data type of your node: %s" % node.name)
 

python/tvm/relay/frontend/pytorch.py

@@ -1263,6 +1263,9 @@ def convolution(self, inputs, input_types):
         else:
             data_layout = "NCW"
             kernel_layout = "OIW"
+            if use_transpose:
+                # Transposed convolutions have IOW layout.
+                kernel_layout = "IOW"
 
         # Conv1d does not currently support grouped convolution so we convert it to conv2d
         is_grouped_conv1d = False

python/tvm/relay/op/contrib/arm_compute_lib.py

@@ -359,6 +359,10 @@ def qnn_conv2d(expr):
     kernel_typ = args[1].checked_type
     if len(kernel_typ.shape) != 4 or kernel_typ.dtype not in qnn_dtypes:
         return False
+    if is_per_channel_quantization(
+        zero_point=args[2], scale=args[4]
+    ) or is_per_channel_quantization(zero_point=args[3], scale=args[5]):
+        return False
     is_depthwise = is_depthwise_conv2d(
         data_typ.shape,
         attrs["data_layout"],
@@ -422,6 +426,10 @@ def qnn_dense(expr):
         return False
     if attrs.out_dtype != "int32":
         return False
+    if is_per_channel_quantization(
+        zero_point=args[2], scale=args[4]
+    ) or is_per_channel_quantization(zero_point=args[3], scale=args[5]):
+        return False
     return True
 
 
@@ -514,10 +522,24 @@ def qnn_add(expr):
     for typ in [args[0].checked_type, args[1].checked_type]:
         if typ.dtype not in ["int8", "uint8"]:
             return False
-
+    if (
+        is_per_channel_quantization(zero_point=args[3], scale=args[2])
+        or is_per_channel_quantization(zero_point=args[5], scale=args[4])
+        or is_per_channel_quantization(zero_point=args[7], scale=args[6])
+    ):
+        return False
     return True
 
 
+def is_per_channel_quantization(zero_point, scale):
+    """Check if the quantization is per-channel"""
+    for value in [zero_point, scale]:
+        shape = value.checked_type.shape
+        if len(shape) != 0 and shape[0] != 1:
+            return True
+    return False
+
+
 class OpAttrContext(object):
     """Temporarily changes the attr of an op."""
 

python/tvm/relay/op/contrib/ethosu.py

@@ -1671,7 +1671,18 @@ def check_compatible_size(mode, method, upscale_size, ifm_size):
             return False
         if self.method not in ("nearest_neighbor", "linear"):
             return False
-        if self.coordinate_transformation_mode not in ("asymmetric", "align_corners"):
+        if self.coordinate_transformation_mode not in (
+            "asymmetric",
+            "align_corners",
+            "half_pixel",
+        ):
+            return False
+        if (
+            self.coordinate_transformation_mode == "half_pixel"
+            and self.rounding_method != "round_prefer_ceil"
+            or self.coordinate_transformation_mode != "half_pixel"
+            and self.rounding_method != ""
+        ):
             return False
         if not check_compatible_size(
             self.coordinate_transformation_mode,
@@ -1680,8 +1691,6 @@ def check_compatible_size(mode, method, upscale_size, ifm_size):
             self.ifm.shape[1:3],
         ):
             return False
-        if self.rounding_method != "":
-            return False
         if self.out_dtype and self.out_dtype != "int8":
             return False
         return True