ekalda commented on a change in pull request #9384:
URL: https://github.com/apache/tvm/pull/9384#discussion_r738285485
##########
File path: python/tvm/relay/backend/contrib/ethosu/util.py
##########
@@ -75,6 +75,15 @@ class ClipArgs(Enum):
A_MAX = 2
+class MaxPoolArgs(Enum):
+ """
+ This is a helper enums to access the correct index
+ of max pool arguments
+ """
+
+ ifm = 0
+
+
Review comment:
Looks like there isn't an equivalent class for average pooling? I think
it is not a hugely informative class though, so maybe we can do without it and
similarly how it is done for average pooling, just assume we know where the ifm
is...
##########
File path: src/relay/op/contrib/ethosu/pooling.cc
##########
@@ -0,0 +1,186 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/pooling.cc
+ * \brief Pooling operators definitions for the Arm(R) Ethos(TM)-U NPU
convolution ops.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU pooling operator */
+struct EthosuPoolingAttrs : public tvm::AttrsNode<EthosuPoolingAttrs> {
+ String pooling_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ Array<IndexExpr> pool_shape;
+ IndexExpr ofm_channels;
+ Array<IndexExpr> strides;
+ Array<IndexExpr> padding;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String upscale;
+ String ifm_layout;
+ String ofm_layout;
+
+ TVM_DECLARE_ATTRS(EthosuPoolingAttrs, "relay.attrs.EthosuPoolingAttrs") {
+ TVM_ATTR_FIELD(pooling_type)
+ .describe("The type of the pooling. 'AVG' - average pool, 'MAX' - max
pool.");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input
Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map
tensor.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output
Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map
tensor.");
+ TVM_ATTR_FIELD(pool_shape)
+ .describe("The 2 dimensional pool shape as (pool_shape_height,
pool_shape_width).")
+ .set_default(NullValue<Array<IndexExpr> >());
+ TVM_ATTR_FIELD(ofm_channels)
+ .describe(" The number of OFM channels.")
+ .set_default(NullValue<IndexExpr>());
+ TVM_ATTR_FIELD(strides)
+ .set_default(Array<IndexExpr>({1, 1}))
+ .describe("The 2 dimensional strides as (stride_height,
stride_width).");
+ TVM_ATTR_FIELD(padding)
+ .describe("The 4 dimensional padding as (pad_top, pad_left,
pad_bottom, pad_right).")
+ .set_default(Array<IndexExpr>({0, 0, 0, 0}));
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function.")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(upscale)
+ .describe(
+ "The 2x2 upscaling mode to apply to the Input Feature Map tensor. "
+ "'NONE' - no upscaling. "
+ "'NEAREST' - upscale using nearest neighbour. "
+ "'ZEROS' - upscale using zeros.")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC'
or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC'
or 'NHCWB16'.")
+ .set_default("NHWC");
+ }
+};
+
+TVM_REGISTER_NODE_TYPE(EthosuPoolingAttrs);
+
+bool EthosuPoolingRel(const Array<Type>& types, int num_inputs, const Attrs&
attrs,
+ const TypeReporter& reporter) {
+ int ifm_index = 0;
+ int result_index = 2;
+ ICHECK_EQ(types.size(), result_index + 1);
+
+ const auto* ifm = types[ifm_index].as<TensorTypeNode>();
+ if (ifm == nullptr) return false;
+
+ const auto* param = attrs.as<EthosuPoolingAttrs>();
+ ICHECK(param != nullptr) << "EthosuPoolingAttrs cannot be nullptr.";
+
+ bool is_avg_pooling = param->pooling_type == "AVG";
+ ICHECK(is_avg_pooling || param->pooling_type == "MAX")
Review comment:
Maybe just ``` ICHECK(param->pooling_type == "AVG" ||
param->pooling_type == "MAX")```
##########
File path: python/tvm/relay/backend/contrib/ethosu/tir_to_cs_translator.py
##########
@@ -417,20 +418,22 @@ def _create_npu_op_conv2d(
return npu_conv2d_op, weights_zero_point
-def translate_ethosu_depthwise_conv2d(tir_extern_call):
- """This function will translate a tir extern_call
- as produced by Relay to TIR compilation.
+def translate_ethosu_depthwise_conv2d(
+ tir_extern_call: tvm.tir.Call,
+) -> Tuple[vapi.NpuConvDepthWiseOperation, int]:
+ """This function will translate a TIR call_extern
+ as produced by NPU Relay to TIR compilation.
Parameters
----------
tir_extern_call : tvm.tir.Call
- This should be a tir external call that has an agreed upon ordering
Review comment:
I think "an" is correct there
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -331,6 +332,133 @@ def qnn_depthwise_conv2d_pattern() ->
tvm.relay.dataflow_pattern.DFPattern:
return clip_or_req
+class MaxPool2DParams:
+ """
+ This class will parse a call to a ethosu.maxpool2d composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.maxpool2d"
+ # The hardware only supports padding upto the numbers as follows
+ padding_bounds = [127, 127, 128, 128]
+
+ def __init__(self, func_body: Call):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ pool_op = clip.args[0]
+ else:
+ pool_op = func_body
+
+ attrs = pool_op.attrs
+ self.ifm = TensorParams(pool_op.args[MaxPoolArgs.ifm.value],
attrs.layout)
+ self.ofm = TensorParams(pool_op, attrs.layout)
+ self.pool_shape = [int(i) for i in attrs.pool_size]
+ self.strides = attrs.strides
+ self.padding = attrs.padding
+ self.activation = clip
+ self.pooling_type = "MAX"
+
+ def is_valid(self):
+ """
+ This function checks whether MaxPool2D has compatible attributes with
the NPU
+ """
+ tensor_params = [self.ifm, self.ofm]
+ if not check_valid_dtypes(tensor_params):
+ return False
+ if self.ifm.dtype != self.ofm.dtype:
+ return False
+ if not check_strides(self.strides):
+ return False
+ if not check_batch_size(self.ifm):
+ return False
+ if not check_padding(self.padding, self.padding_bounds):
+ return False
+ # Check pool size
+ if (
+ len(self.pool_shape) != 2
+ or self.pool_shape[1] > 256
+ or self.pool_shape[0] * self.pool_shape[1] > 256 * 256
+ ):
+ return False
+ return True
+
+
+def qnn_maxpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for nn.max_pool2d with optional fused
RELU activation.
+ """
+ pattern = is_op("nn.max_pool2d")(wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class AvgPool2DParams:
Review comment:
Looks like `class AvgPool2DParams` and `class MaxPool2DParams` are
identical except the pooling type? Maybe we can turn the identical parts into
one parent class?
##########
File path: python/tvm/relay/backend/contrib/ethosu/util.py
##########
@@ -75,6 +75,15 @@ class ClipArgs(Enum):
A_MAX = 2
+class MaxPoolArgs(Enum):
+ """
+ This is a helper enums to access the correct index
+ of max pool arguments
+ """
+
+ ifm = 0
+
+
Review comment:
Looks like there isn't an equivalent class for average pooling? I think
it is not a hugely informative class though, so maybe we can do without it and
similarly how it is done for average pooling, just assume we know where the ifm
is...
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -254,5 +254,93 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
[email protected](
+ "accel_type",
+ ACCEL_TYPES,
+)
[email protected]("pooling_type", ["MAX", "AVG"])
[email protected]("ifm_shape", [[1, 3, 4, 3], [1, 4, 5, 2]])
[email protected]("pool_shape, strides", [([1, 2], [1, 2]), ([2, 3],
[2, 3])])
[email protected]("activation_function", ["NONE", "RELU"])
[email protected]("padding", ["SAME", "VALID"])
Review comment:
To my counting, this is 256 tests, that will take a while to run...
Maybe we can group together more things that (hopefully) don't interact, e.g.
I'm not sure different combinations of activation functions and paddings give
us any useful information. It would be good to know what others, e.g.
@manupa-arm and @mbaret think about it
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
