Add GroupNorm and LayerNorm onnx parsing (#2242)

src/onnx/parse_groupnorm.cpp

@@ -0,0 +1,130 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_groupnorm : op_parser<parse_groupnorm>
+{
+    std::vector<op_desc> operators() const { return {{"GroupNormalization"}}; }
+
+    instruction_ref parse(const op_desc& /*opd*/,
+                          const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          std::vector<instruction_ref> args) const
+    {
+        float epsilon = 1e-5f;
+        if(contains(info.attributes, "epsilon"))
+        {
+            epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
+        }
+        size_t num_groups;
+        if(contains(info.attributes, "num_groups"))
+        {
+            num_groups = parser.parse_value(info.attributes.at("num_groups")).at<size_t>();
+        }
+        else
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: num_groups must be available");
+        }
+
+        if(args.size() != 3)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: invalid input count");
+        }
+
+        auto x     = args.at(0);
+        auto scale = args.at(1);
+        auto bias  = args.at(2);
+
+        auto x_shape = x->get_shape();
+        auto x_dtype = x_shape.type();
+        auto x_dims  = x_shape.lens();
+
+        if(x_shape.ndim() <= 2)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: invalid input shape");
+        }
+
+        auto c = x_shape.lens().at(1);
+        if(c % num_groups != 0)
+        {
+            MIGRAPHX_THROW(
+                "PARSE_GROUPNORM: num_groups should be a divisor of the number of channels");
+        }
+        auto group_size = c / num_groups;
+        if(scale->get_shape().ndim() != 1 or scale->get_shape().lens().at(0) != num_groups)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: scale tensor shape should be num_groups");
+        }
+        if(bias->get_shape().ndim() != 1 or bias->get_shape().lens().at(0) != num_groups)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: bias tensor shape should be num_groups");
+        }
+
+        // Original shape: N x C x D1 x ... x Dn
+        // New shape: N x num_groups x C // num_groups x D1 x ... x Dn
+
+        std::vector<size_t> dims = {x_dims.at(0), num_groups, group_size};
+        std::copy(x_dims.begin() + 2, x_dims.end(), std::back_inserter(dims));
+        auto x_reshaped = info.add_instruction(make_op("reshape", {{"dims", dims}}), x);
+
+        // Axes for D1 x ... x Dn
+        std::vector<size_t> axes(dims.size() - 2);
+        std::iota(axes.begin(), axes.end(), 2);
+
+        // y = (x - mean) * rsqrt(variance + epsilon) * scale + bias
+        // mean = reduce_mean({D1, D2, ... Dk}, x)
+        // variance = reduce_mean({D1, D2, ... Dk}, (x - mean)^2)
+
+        auto mean = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x_reshaped);
+        auto x_sub_mean    = info.add_common_op("sub", x_reshaped, mean);
+        auto x_sqdiff_mean = info.add_common_op("sqdiff", x_reshaped, mean);
+        auto variance =
+            info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x_sqdiff_mean);
+        epsilon =
+            (x_dtype == migraphx::shape::half_type and std::abs(epsilon) < 1e-7) ? 1e-7 : epsilon;
+        auto eps     = info.add_literal(migraphx::literal{migraphx::shape{x_dtype}, {epsilon}});
+        auto var_eps = info.add_common_op("add", variance, eps);
+        auto rsqrt   = info.add_instruction(make_op("rsqrt"), var_eps);
+        auto result  = info.add_common_op("mul", x_sub_mean, rsqrt);
+        auto scale_bcast =
+            info.add_instruction(make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), scale);
+        auto bias_bcast =
+            info.add_instruction(make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), bias);
+        auto scaled     = info.add_instruction(make_op("mul"), result, scale_bcast);
+        auto y          = info.add_instruction(make_op("add"), scaled, bias_bcast);
+        auto y_reshaped = info.add_instruction(make_op("reshape", {{"dims", x_dims}}), y);
+        return y_reshaped;
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_layernorm.cpp

@@ -0,0 +1,131 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_layernorm : op_parser<parse_layernorm>
+{
+    std::vector<op_desc> operators() const { return {{"LayerNormalization"}}; }
+
+    std::vector<instruction_ref> parse(const op_desc& /*opd*/,
+                                       const onnx_parser& parser,
+                                       const onnx_parser::node_info& info,
+                                       std::vector<instruction_ref> args) const
+    {
+        int64_t axis = -1;
+        if(contains(info.attributes, "axis"))
+        {
+            axis = parser.parse_value(info.attributes.at("axis")).at<int64_t>();
+        }
+        float epsilon = 1e-5f;
+        if(contains(info.attributes, "epsilon"))
+        {
+            epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
+        }
+        if(contains(info.attributes, "stash_type"))
+        {
+            std::cerr << "WARNING: LAYERNORM does not support stash_type, it will be ignored.\n";
+        }
+
+        if(args.size() < 2 or args.size() > 3)
+        {
+            MIGRAPHX_THROW("PARSE_LAYERNORM: invalid input count");
+        }
+
+        auto x         = args.at(0);
+        auto scale     = args.at(1);
+        bool skip_bias = args.size() == 2;
+        instruction_ref bias;
+        if(not skip_bias)
+        {
+            bias = args.at(2);
+        }
+
+        auto x_shape   = x->get_shape();
+        auto x_dtype   = x_shape.type();
+        int64_t x_rank = x_shape.ndim();
+
+        if(x_rank < 2)
+        {
+            MIGRAPHX_THROW("PARSE_LAYERNORM: invalid input shape");
+        }
+
+        // If rank(X) is r, axis' allowed range is [-r, r)
+        if(axis < -x_rank or axis >= x_rank)
+        {
+            MIGRAPHX_THROW("PARSE_LAYERNORM: invalid axis");
+        }
+
+        // y = (x - mean) * rsqrt(variance + epsilon) * scale + bias
+        // mean = reduce_mean({D1, D2, ... Dk}, x)
+        // variance = reduce_mean({D1, D2, ... Dk}, (x - mean)^2)
+
+        // axis can be negative
+        axis = axis < 0 ? axis + x_rank : axis;
+
+        auto kdims = x_rank - axis;
+        std::vector<int64_t> axes(kdims);
+        std::iota(axes.begin(), axes.end(), axis);
+        auto skipped_axes = x_rank - kdims;
+
+        auto mean          = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x);
+        auto x_sub_mean    = info.add_common_op("sub", x, mean);
+        auto x_sqdiff_mean = info.add_common_op("sqdiff", x, mean);
+        auto variance =
+            info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x_sqdiff_mean);
+        epsilon =
+            (x_dtype == migraphx::shape::half_type and std::abs(epsilon) < 1e-7) ? 1e-7 : epsilon;
+        auto eps     = info.add_literal(migraphx::literal{migraphx::shape{x_dtype}, {epsilon}});
+        auto var_eps = info.add_common_op("add", variance, eps);
+        auto rsqrt   = info.add_instruction(make_op("rsqrt"), var_eps);
+        auto result  = info.add_common_op("mul", x_sub_mean, rsqrt);
+
+        instruction_ref scale_bcast = scale;
+        instruction_ref bias_bcast  = bias;
+        if(skipped_axes > 0)
+        {
+            auto x_dims = x_shape.lens();
+            scale_bcast = info.add_instruction(
+                make_op("broadcast", {{"axis", skipped_axes}, {"out_lens", x_dims}}), scale);
+            if(not skip_bias)
+            {
+                bias_bcast = info.add_instruction(
+                    make_op("broadcast", {{"axis", skipped_axes}, {"out_lens", x_dims}}), bias);
+            }
+        }
+        auto scaled = info.add_instruction(make_op("mul"), result, scale_bcast);
+        auto y      = skip_bias ? scaled : info.add_instruction(make_op("add"), scaled, bias_bcast);
+        return {y, mean, rsqrt};
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx