[microNPU] Tweak a layout transform matrix (apache#10763)

* [microNPU] Fix layout transform matrix

One of the layout transforms currently causes the cascader to stripe
across B16 axis (which is not allowed), so change that and deal with
the implications to the get_valid_block_configs.

Change-Id: I04199f9f35fcc31618581567483cfb80d3b5aad2

* Reduce the duplication of layout transfrom matrices

* Change the nhcwb16_to_nhwc matrix for binary and unary elementwise
  such that it matches the other NPU ops
* Reduce the number of places where the same layout transform matrices are
  defined

* Add documentation to the layout transform matrices

python/tvm/contrib/ethosu/cascader/device_config.py

@@ -439,6 +439,23 @@ def is_partkernel(
 
         return part_kernel_first_utilization > depth_first_utilization or ifm_channels <= 8
 
+    def _get_input_banks(self, input_block_shape, input_bytewidth):
+        input_bytes = input_block_shape.area() * self._align(
+            input_block_shape.depth * input_bytewidth, 8
+        )
+        input_banks = _round_up_div(input_bytes, self._bank_size_bytes) * 2
+        input_banks = _round_up(input_banks, self._input_granularity)
+
+        return input_banks
+
+    def _get_accumulator_banks(self, output_block_shape, acc_bytewidth, depth):
+        acc_depth = _round_up(min(output_block_shape.depth, depth), 8)
+        acc_bytes = output_block_shape.area() * self._align(acc_depth, 8) * acc_bytewidth
+        acc_banks = _round_up_div(acc_bytes, self._bank_size_bytes) * 2
+        acc_banks = _round_up(acc_banks, self._accumulator_granularity[acc_bytewidth])
+
+        return acc_banks
+
     def get_elementwise_block_config(
         self,
         ifm_propagator: Propagator,
@@ -533,16 +550,9 @@ def get_elementwise_block_config(
             input2_block.round_up(self._input_micro_block)
 
             # Banks required for input block
-            input_bytes = input_block.area() * self._align(input_block.depth * input_bytewidth, 8)
-            input_banks = _round_up_div(input_bytes, self._bank_size_bytes) * 2
-            input_banks = _round_up(input_banks, self._input_granularity)
-
+            input_banks = self._get_input_banks(input_block, input_bytewidth)
             # Banks required for input2 block
-            input2_bytes = input2_block.area() * self._align(
-                input2_block.depth * input_bytewidth, 8
-            )
-            input2_banks = _round_up_div(input2_bytes, self._bank_size_bytes) * 2
-            input2_banks = _round_up(input2_banks, self._input_granularity)
+            input2_banks = self._get_input_banks(input2_block, input_bytewidth)
 
             # Check whether or not both IFMs fit into SHRAM
             if (input_banks + input2_banks) <= banks_available:
@@ -561,6 +571,29 @@ def get_elementwise_block_config(
 
         return block_config
 
+    def _get_subkernel_propagator(
+        self, op_attrs, ifm_propagator, input_layout, output_layout, depth
+    ):
+        op_type = op_attrs.get("op")
+        stride_h = int(op_attrs.get("stride_h", 1))
+        stride_w = int(op_attrs.get("stride_w", 1))
+        transform = ifm_propagator.transform
+
+        if input_layout == "NHCWB16":
+            transform[1][-1] = min(transform[1][-1], self._subkernel_limits[0] - stride_h)
+            transform[3][-1] = min(transform[3][-1], self._subkernel_limits[1] - stride_w)
+        else:
+            transform[1][-1] = min(transform[1][-1], self._subkernel_limits[0] - stride_h)
+            transform[2][-1] = min(transform[2][-1], self._subkernel_limits[1] - stride_w)
+
+        if op_type in ("ethosu_pooling", "ethosu_depthwise_conv2d"):
+            if output_layout == "NHCWB16" and input_layout == "NHWC":
+                transform[3][-1] = depth
+            elif output_layout == "NHCWB16" and input_layout == "NHCWB16":
+                transform[2][-1] = depth // 16
+
+        return Propagator(transform, ifm_propagator.offset)
+
     def get_valid_block_configs(
         self,
         ifm_propagator: Propagator,
@@ -612,33 +645,13 @@ def get_valid_block_configs(
         op_type = op_attrs.get("op")
         op_str = op_attrs.get("op_str")
         activation = op_attrs.get("activation", "NONE")
-        stride_h = int(op_attrs.get("stride_h", 1))
-        stride_w = int(op_attrs.get("stride_w", 1))
         upscaling_factor = 1 if op_attrs.get("upscale", "NONE") == "NONE" else 2
 
-        subkernel_transform = ifm_propagator.transform
         if output_layout == "NHCWB16":
             output_shape = _Shape([1, ofm_shape[1], ofm_shape[3], ofm_channels])
         else:
             output_shape = _Shape(ofm_shape)
 
-        if input_layout == "NHCWB16":
-            subkernel_transform[1][-1] = min(
-                subkernel_transform[1][-1], self._subkernel_limits[0] - stride_h
-            )
-            subkernel_transform[3][-1] = min(
-                subkernel_transform[3][-1], self._subkernel_limits[1] - stride_w
-            )
-        else:
-            subkernel_transform[1][-1] = min(
-                subkernel_transform[1][-1], self._subkernel_limits[0] - stride_h
-            )
-            subkernel_transform[2][-1] = min(
-                subkernel_transform[2][-1], self._subkernel_limits[1] - stride_w
-            )
-
-        subkernel_propagator = Propagator(subkernel_transform, ifm_propagator.offset)
-
         # Define search space
         max_height = min(output_shape.height, self._max_block_shape.height)
         min_height = max(self._micro_block.height, upscaling_factor)
@@ -655,7 +668,7 @@ def get_valid_block_configs(
         if activation == "LUT" and not self._lut_reserved:
             banks_available -= 2
 
-        # Input block depth has additional limitations for Operators that require full input depth
+        # Input block depth has additional limitations for operators that require full input depth
         input_block_depth = 0
         is_partkernel = self.is_partkernel(op_type, ifm_channels, ifm_dtype, kernel_h * kernel_w)
         if op_type == "ethosu_conv2d":
@@ -669,6 +682,10 @@ def get_valid_block_configs(
                 # Block depth has to be less than full depth or a multiple of the split depth
                 continue
 
+            subkernel_propagator = self._get_subkernel_propagator(
+                op_attrs, ifm_propagator, input_layout, output_layout, depth
+            )
+
             for width in range(min_width, max_width + min_width, min_width):
                 for height in range(min_height, max_height + min_height, min_height):
                     if output_layout == "NHCWB16":
@@ -709,19 +726,11 @@ def get_valid_block_configs(
                         input_block_shape.depth = input_block_depth
 
                     # Banks required for input block
-                    input_bytes = input_block_shape.area() * self._align(
-                        input_block_shape.depth * input_bytewidth, 8
-                    )
-                    input_banks = _round_up_div(input_bytes, self._bank_size_bytes) * 2
-                    input_banks = _round_up(input_banks, self._input_granularity)
-
+                    input_banks = self._get_input_banks(input_block_shape, input_bytewidth)
                     # Banks required for accumulation
-                    acc_depth = _round_up(min(output_block_shape.depth, ofm_channels), 8)
-                    acc_bytes = (
-                        output_block_shape.area() * self._align(acc_depth, 8) * acc_bytewidth
+                    acc_banks = self._get_accumulator_banks(
+                        output_block_shape, acc_bytewidth, depth
                     )
-                    acc_banks = _round_up_div(acc_bytes, self._bank_size_bytes) * 2
-                    acc_banks = _round_up(acc_banks, self._accumulator_granularity[acc_bytewidth])
 
                     if (input_banks + acc_banks) <= banks_available:
                         output_cycles = self._get_output_cycles(

python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py

@@ -22,6 +22,7 @@
 from tvm.contrib.ethosu.cascader import TESubgraph, EthosuPart, Propagator, register_matcher
 
 from .dma import dma_ofm_compute, dma_ifm_compute
+from .common import get_layout_transform_matrices
 
 
 def binary_elementwise_compute(
@@ -196,21 +197,8 @@ def binary_elementwise_compute(
             attrs=binary_elementwise_attrs,
         )
 
-    nhwc_to_nhcwb16 = [
-        [1, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0],
-        [0, 0, 0, 1 / 16, 0],
-        [0, 0, 1, 0, 0],
-        [0, 0, 0, 0, 16],
-        [0, 0, 0, 0, 1],
-    ]
-    nhcwb16_to_nhwc = [
-        [1, 0, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0, 0],
-        [0, 0, 0, 1, 0, 0],
-        [0, 0, 16, 0, 1, -16],
-        [0, 0, 0, 0, 0, 1],
-    ]
+    nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(int(ifm_channels))
+
     ifm_matrix = [
         [1, 0, 0, 0, 0],
         [0, 1, 0, 0, 0],

python/tvm/relay/backend/contrib/ethosu/te/common.py

@@ -0,0 +1,60 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""Common methods for the NPU tensor expressions"""
+
+from typing import Tuple, List
+
+
+def get_layout_transform_matrices(ofm_channels: int) -> Tuple[List[List[float]], List[List[float]]]:
+    """Get the NHWC->NHCWB16 and NHCWB16->NHWC layout transform matrices.
+    For information about the supported layouts see https://developer.arm.com/documentation/102420/
+    0200/Functional-description/Control-and-data-flow/Supported-memory-formats-for-feature-maps
+
+    Parameters
+    ----------
+    ofm_channels : int
+        The number of output channels in a NHWC layout
+
+    Returns
+    -------
+    nhwc_to_nhcwb16, nhcwb16_to_nhwc : Tuple[List[List[float]], List[List[float]]]
+        The layout transformation matrices
+    """
+
+    # The value of the last dimension (B16) is always 16.
+    nhwc_to_nhcwb16 = [
+        [1, 0, 0, 0, 0],
+        [0, 1, 0, 0, 0],
+        [0, 0, 0, 1 / 16, 0],
+        [0, 0, 1, 0, 0],
+        [0, 0, 0, 0, 16],
+        [0, 0, 0, 0, 1],
+    ]
+
+    # When we convert from NHWC to NHCWB16, the new C value is given by
+    # (ofm_channels - 1) // 16 + 1, which is a lossy operation, so we need to use
+    # the actual value of channels in the transform matrix to accurately recover
+    # the C in NHWC when we convert from NHCWB16 to NHWC.
+    nhcwb16_to_nhwc = [
+        [1, 0, 0, 0, 0, 0],
+        [0, 1, 0, 0, 0, 0],
+        [0, 0, 0, 1, 0, 0],
+        [0, 0, 0, 0, 0, ofm_channels],
+        [0, 0, 0, 0, 0, 1],
+    ]
+
+    return nhwc_to_nhcwb16, nhcwb16_to_nhwc

python/tvm/relay/backend/contrib/ethosu/te/convolution.py

@@ -23,6 +23,7 @@
 from tvm.contrib.ethosu.cascader import TESubgraph, EthosuPart, Propagator, register_matcher
 
 from .dma import dma_ofm_compute, dma_ifm_compute
+from .common import get_layout_transform_matrices
 
 
 def conv2d_compute(
@@ -175,21 +176,8 @@ def conv2d_compute(
         attrs=conv2d_attrs,
     )
 
-    nhwc_to_nhcwb16 = [
-        [1, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0],
-        [0, 0, 0, 1 / 16, 0],
-        [0, 0, 1, 0, 0],
-        [0, 0, 0, 0, 16],
-        [0, 0, 0, 0, 1],
-    ]
-    nhcwb16_to_nhwc = [
-        [1, 0, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0, 0],
-        [0, 0, 0, 1, 0, 0],
-        [0, 0, 16, 0, 1, -16],
-        [0, 0, 0, 0, 0, 1],
-    ]
+    nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(ofm_channels)
+
     ifm_matrix = [
         [1, 0, 0, 0, 0],
         [0, stride_h, 0, 0, (dilated_kernel_h - stride_h)],

python/tvm/relay/backend/contrib/ethosu/te/depthwise.py

@@ -23,6 +23,7 @@
 from tvm.contrib.ethosu.cascader import TESubgraph, EthosuPart, Propagator, register_matcher
 
 from .dma import dma_ofm_compute, dma_ifm_compute
+from .common import get_layout_transform_matrices
 
 
 def depthwise_conv2d_compute(
@@ -169,21 +170,8 @@ def depthwise_conv2d_compute(
         attrs=depthwise_conv2d_attrs,
     )
 
-    nhwc_to_nhcwb16 = [
-        [1, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0],
-        [0, 0, 0, 1 / 16, 0],
-        [0, 0, 1, 0, 0],
-        [0, 0, 0, 0, 16],
-        [0, 0, 0, 0, 1],
-    ]
-    nhcwb16_to_nhwc = [
-        [1, 0, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0, 0],
-        [0, 0, 0, 1, 0, 0],
-        [0, 0, 16, 0, 1, -16],
-        [0, 0, 0, 0, 0, 1],
-    ]
+    nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(channels)
+
     ifm_matrix = [
         [1, 0, 0, 0, 0],
         [0, stride_h, 0, 0, (dilated_kernel_h - stride_h)],

python/tvm/relay/backend/contrib/ethosu/te/pooling.py

@@ -23,6 +23,7 @@
 from tvm.contrib.ethosu.cascader import TESubgraph, EthosuPart, Propagator, register_matcher
 
 from .dma import dma_ofm_compute, dma_ifm_compute
+from .common import get_layout_transform_matrices
 
 
 def pooling_compute(
@@ -157,21 +158,8 @@ def pooling_compute(
         attrs=pooling_attrs,
     )
 
-    nhwc_to_nhcwb16 = [
-        [1, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0],
-        [0, 0, 0, 1 / 16, 0],
-        [0, 0, 1, 0, 0],
-        [0, 0, 0, 0, 16],
-        [0, 0, 0, 0, 1],
-    ]
-    nhcwb16_to_nhwc = [
-        [1, 0, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0, 0],
-        [0, 0, 0, 1, 0, 0],
-        [0, 0, 16, 0, 1, -16],
-        [0, 0, 0, 0, 0, 1],
-    ]
+    nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(int(ofm_channels))
+
     ifm_matrix = [
         [1, 0, 0, 0, 0],
         [0, stride_h, 0, 0, (pool_shape_h - stride_h)],

python/tvm/relay/backend/contrib/ethosu/te/unary_elementwise.py

@@ -21,6 +21,7 @@
 from tvm import te
 from tvm.contrib.ethosu.cascader import TESubgraph, EthosuPart, Propagator, register_matcher
 from .dma import dma_ofm_compute, dma_ifm_compute
+from .common import get_layout_transform_matrices
 
 
 def unary_elementwise_compute(
@@ -129,21 +130,8 @@ def clz_imp(inp):
         attrs=unary_elementwise_attrs,
     )
 
-    nhwc_to_nhcwb16 = [
-        [1, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0],
-        [0, 0, 0, 1 / 16, 0],
-        [0, 0, 1, 0, 0],
-        [0, 0, 0, 0, 16],
-        [0, 0, 0, 0, 1],
-    ]
-    nhcwb16_to_nhwc = [
-        [1, 0, 0, 0, 0, 0],
-        [0, 1, 0, 0, 0, 0],
-        [0, 0, 0, 1, 0, 0],
-        [0, 0, 16, 0, 1, -16],
-        [0, 0, 0, 0, 0, 1],
-    ]
+    nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(int(ofm_channels))
+
     ifm_matrix = [
         [1, 0, 0, 0, 0],
         [0, 1, 0, 0, 0],

src/contrib/ethosu/cascader/block_config.cc

@@ -37,6 +37,8 @@ void BlockConfigNode::VisitAttrs(AttrVisitor* v) {
   v->Visit("_input_shape", &tmp_arr);
   tmp_arr = make_array(output_shape_);
   v->Visit("_output_shape", &tmp_arr);
+  v->Visit("_compute_cycles", &compute_cycles_);
+  v->Visit("_output_cycles", &output_cycles_);
 }
 
 BlockConfig::BlockConfig(const std::vector<int>& input_shape, const std::vector<int>& output_shape,