You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2022/07/06 14:52:07 UTC
[GitHub] [tvm] Mousius commented on a diff in pull request #12006: [CMSIS-NN][Perf] Converted Relay Conv2D into CMSIS-NN Depthwise
Mousius commented on code in PR #12006:
URL: https://github.com/apache/tvm/pull/12006#discussion_r914809595
##########
src/relay/backend/contrib/cmsisnn/relay_to_tir.cc:
##########
@@ -31,6 +30,7 @@
#include "../../../transforms/pattern_utils.h"
#include "buffer_size.h"
#include "compiler_attrs.h"
+#include "utils.h"
Review Comment:
I always find creating these `utils` files leads to creating areas for functions with no relation to one another to exist. What do you think about `convolutions.cc` and `convolutions.h` to represent functions relevant to convolutions?
##########
src/relay/backend/contrib/cmsisnn/utils.cc:
##########
@@ -0,0 +1,47 @@
+/*
+ * 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.
+ */
+
+#include "../../../qnn/utils.h"
+
+#include <tvm/ir/transform.h>
+#include <tvm/relay/attrs/nn.h>
+
+namespace tvm {
+namespace relay {
+namespace contrib {
+namespace cmsisnn {
+
+bool is_cmsisnn_depthwise(const Conv2DAttrs* conv2d_attrs, const Array<PrimExpr>& input_shape,
Review Comment:
```suggestion
bool IsCMSISNNDepthwise(const Conv2DAttrs* conv2d_attrs, const Array<PrimExpr>& input_shape,
```
to match C++ style guide?
##########
tests/python/contrib/test_cmsisnn/test_conv2d.py:
##########
@@ -540,6 +542,111 @@ def test_depthwise_int8(
)
+@tvm.testing.requires_cmsisnn
+@pytest.mark.parametrize("padding", ["SAME", "VALID"])
+@pytest.mark.parametrize("strides, dilation", [((1, 1), (1, 1))])
+@pytest.mark.parametrize("relu_type", ["RELU", "NONE"])
+@pytest.mark.parametrize("depth_multiplier", [1, 3])
+@pytest.mark.parametrize(
+ "input_zero_point, input_scale, kernel_scale",
+ [
+ (
+ 10,
+ 0.0128,
+ [0.11, 0.22],
+ ),
+ (
+ -64,
+ 1,
+ [1, 0.0256, 1.37],
+ ),
+ ],
+)
+def test_relay_conv2d_cmsisnn_depthwise_int8(
+ padding,
+ strides,
+ dilation,
+ relu_type,
+ input_zero_point,
+ input_scale,
+ kernel_scale,
+ depth_multiplier,
+):
+ """Tests QNN Depthwise int8 op via CMSIS-NN"""
+ interface_api = "c"
+ use_unpacked_api = True
+ test_runner = AOT_USMP_CORSTONE300_RUNNER
+
+ dtype = "int8"
+ in_min, in_max = get_range_for_dtype_str(dtype)
+
+ ifm_shape = (1, 24, 24, 1)
+ groups = ifm_shape[3]
+ weight_format = "HWIO"
+ (kernel_h, kernel_w) = (3, 3)
+ kernel_shape = (kernel_h, kernel_w, ifm_shape[3], depth_multiplier)
+ out_channels = ifm_shape[3] * depth_multiplier
+ enable_bias = True
+ ks_len = len(kernel_scale)
+ kernel_zero_point = 0
+ kernel_scale = [kernel_scale[i % ks_len] for i in range(out_channels)]
+
+ output_scale, output_zero_point = get_conv2d_qnn_params(
+ kernel_shape,
+ input_scale,
+ input_zero_point,
+ kernel_scale,
+ kernel_zero_point,
+ dtype,
+ dtype,
+ dtype,
+ True,
+ )
+
+ model, params = make_model(
+ ifm_shape,
+ kernel_shape,
+ input_zero_point,
+ input_scale,
+ kernel_zero_point,
+ kernel_scale,
+ output_zero_point,
+ output_scale,
+ padding,
+ strides,
+ dilation,
+ groups,
+ dtype,
+ dtype,
+ out_channels,
+ weight_format,
+ enable_bias,
+ relu_type,
+ )
+ orig_mod = make_module(model)
+ cmsisnn_mod = cmsisnn.partition_for_cmsisnn(orig_mod, params)
+
+ # validate pattern matching
+ assert_partitioned_function(orig_mod, cmsisnn_mod)
Review Comment:
Can we check we got depthwise rather than conv2d to ensure the logic works?
##########
src/relay/backend/contrib/cmsisnn/utils.cc:
##########
@@ -0,0 +1,47 @@
+/*
+ * 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.
+ */
+
+#include "../../../qnn/utils.h"
+
+#include <tvm/ir/transform.h>
+#include <tvm/relay/attrs/nn.h>
+
+namespace tvm {
+namespace relay {
+namespace contrib {
+namespace cmsisnn {
+
+bool is_cmsisnn_depthwise(const Conv2DAttrs* conv2d_attrs, const Array<PrimExpr>& input_shape,
+ const Array<PrimExpr>& kernel_shape) {
+ std::string kernel_layout = conv2d_attrs->kernel_layout.c_str();
+ int kernel_pos_o = kernel_layout.find("O");
+ int kernel_pos_i = kernel_layout.find("I");
+ int kernel_dim_o_val = qnn::get_const_int(kernel_shape[kernel_pos_o]);
+ int kernel_dim_i_val = qnn::get_const_int(kernel_shape[kernel_pos_i]);
+ int64_t out_channels = conv2d_attrs->channels.as<IntImmNode>()->value;
+ if (out_channels == kernel_dim_o_val * kernel_dim_i_val) {
+ return true;
+ }
+ return false;
Review Comment:
```suggestion
return (out_channels == kernel_dim_o_val * kernel_dim_i_val);
```
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org