Convert Fp16 instance-norm to FP32 temporarily (#1779)

By converting to fp32 : fp16 3d-unet model accuracy comes out the same as FP32 accuracy.

By using reduce_sum method on Fp16 : accuracy comes out ~0.9% lower compared to fp32 while keeping entire model in fp16.

src/onnx/parse_instancenorm.cpp

@@ -21,10 +21,14 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
+#include <iterator>
 #include <migraphx/onnx/op_parser.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/env.hpp>
+
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_FP16_INSTANCENORM_CONVERT);
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -39,22 +43,43 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
     instruction_ref parse(const op_desc& opd,
                           const onnx_parser& parser,
                           onnx_parser::node_info info,
-                          std::vector<instruction_ref> args) const
+                          std::vector<instruction_ref> oargs) const
     {
         // y = scale * ( x - mean ) / sqrt ( variance + epsilon ) + bias
         // mean = reduce_mean({D1, D2, ... Dk}, x)
         // variance = reduce_mean({D1, D2, ... Dk}, (x - mean)^2)
-
+        // Convert fp16 to fp32 to workaround for FP16 accuracy issues with reduce_mean/variance.
+        bool convert_fp16 = true;
+        if(enabled(MIGRAPHX_DISABLE_FP16_INSTANCENORM_CONVERT{}))
+        {
+            convert_fp16 = false;
+        }
         float epsilon = 1e-5f;
         if(contains(info.attributes, "epsilon"))
         {
             epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
         }
+        auto dtype         = oargs[0]->get_shape().type();
+        auto literal_dtype = dtype;
+        std::vector<instruction_ref> args;
+        // cppcheck-suppress knownConditionTrueFalse
+        if(dtype == shape::half_type and convert_fp16)
+        {
+            std::transform(oargs.begin(), oargs.end(), std::back_inserter(args), [&](const auto i) {
+                return info.add_instruction(
+                    make_op("convert", {{"target_type", shape::float_type}}), i);
+            });
+            literal_dtype = shape::float_type;
+        }
+        else
+        {
+            args = oargs;
+        }
+
         auto x     = args[0];
         auto scale = args[1];
         auto bias  = args[2];
         auto dims  = x->get_shape().lens();
-        auto dtype = x->get_shape().type();
         if(not contains(valid_types, dtype))
             MIGRAPHX_THROW(opd.op_name + ": invalid output type: " + std::to_string(dtype) +
                            ". Valid types are 1 (float), 10 (half), and 11 (double).");
@@ -65,14 +90,29 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
 
         std::vector<int64_t> axes(kdims);
         std::iota(axes.begin(), axes.end(), 2);
-
         auto mean = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x);
         auto mean_bcast =
             info.add_instruction(make_op("multibroadcast", {{"out_lens", dims}}), mean);
+        auto l1 = info.add_instruction(make_op("sub"), x, mean_bcast);
+        // for the fp16, if not converting to fp32 then divide `x` and `mean` by `sqrt(n)` and take
+        // reduce_sum to calculate variance i.e.
+        // var =  reduce_sum((x/s_n - mean/s_n)^2) where s_n = sqrt(n)
+        std::string reduce_op_name =
+            (dtype == shape::half_type and not convert_fp16) ? "reduce_sum" : "reduce_mean";
+        if(dtype == shape::half_type and not convert_fp16)
+        {
+            double n =
+                std::accumulate(dims.begin() + 2, dims.end(), 1, [&](const auto& i, const auto& j) {
+                    return i * j;
+                });
+            n              = 1.0 / std::sqrt(n);
+            auto n_literal = info.add_literal(literal{dtype, {n}});
+            mean_bcast     = info.add_common_op("mul", {mean_bcast, n_literal});
+            x              = info.add_common_op("mul", {x, n_literal});
+        }
         auto l0              = info.add_instruction(make_op("sqdiff"), x, mean_bcast);
-        auto variance        = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), l0);
-        auto l1              = info.add_instruction(make_op("sub"), x, mean_bcast);
-        auto epsilon_literal = info.add_literal(literal{shape{dtype}, {epsilon}});
+        auto variance        = info.add_instruction(make_op(reduce_op_name, {{"axes", axes}}), l0);
+        auto epsilon_literal = info.add_literal(literal{shape{literal_dtype}, {epsilon}});
         auto epsilon_bcast =
             info.add_instruction(make_op("multibroadcast", {{"out_lens", dims}}), epsilon_literal);
         auto variance_bcast =
@@ -82,11 +122,16 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
         auto l4 = info.add_instruction(make_op("mul"), l1, l3);
         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 l5 = info.add_instruction(make_op("mul"), l4, scale_bcast);
-        return info.add_instruction(make_op("add"), l5, bias_bcast);
+        auto l5  = info.add_instruction(make_op("mul"), l4, scale_bcast);
+        auto ret = info.add_instruction(make_op("add"), l5, bias_bcast);
+        if(dtype == shape::half_type and convert_fp16)
+        {
+            return info.add_instruction(make_op("convert", {{"target_type", shape::half_type}}),
+                                        ret);
+        }
+        return ret;
     }
 };
 

test/onnx/onnx_test.cpp

@@ -3176,9 +3176,9 @@ TEST_CASE(instance_norm_test)
     auto mean = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), x);
     auto mean_bcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", dims}}), mean);
-    auto l0       = mm->add_instruction(migraphx::make_op("sqdiff"), x, mean_bcast);
-    auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l0);
-    auto l1       = mm->add_instruction(migraphx::make_op("sub"), x, mean_bcast);
+    auto l0       = mm->add_instruction(migraphx::make_op("sub"), x, mean_bcast);
+    auto l1       = mm->add_instruction(migraphx::make_op("sqdiff"), x, mean_bcast);
+    auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l1);
     auto epsilon_literal =
         mm->add_literal(migraphx::literal{migraphx::shape{migraphx::shape::float_type}, {1e-5}});
     auto epsilon_bcast = mm->add_instruction(
@@ -3187,7 +3187,7 @@ TEST_CASE(instance_norm_test)
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", dims}}), variance);
     auto l2          = mm->add_instruction(migraphx::make_op("add"), variance_bcast, epsilon_bcast);
     auto l3          = mm->add_instruction(migraphx::make_op("rsqrt"), l2);
-    auto l4          = mm->add_instruction(migraphx::make_op("mul"), l1, l3);
+    auto l4          = mm->add_instruction(migraphx::make_op("mul"), l0, l3);
     auto scale_bcast = mm->add_instruction(
         migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), scale);
     auto bias_bcast = mm->add_instruction(
@@ -3207,33 +3207,41 @@ TEST_CASE(instance_norm_half_test)
     migraphx::shape s2{migraphx::shape::half_type, {2}};
 
     migraphx::program p;
-    auto* mm   = p.get_main_module();
-    auto x     = mm->add_parameter("0", s1);
-    auto scale = mm->add_parameter("1", s2);
-    auto bias  = mm->add_parameter("2", s2);
+    auto* mm        = p.get_main_module();
+    auto x_fp16     = mm->add_parameter("0", s1);
+    auto scale_fp16 = mm->add_parameter("1", s2);
+    auto bias_fp16  = mm->add_parameter("2", s2);
+
+    auto x = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::float_type}}), x_fp16);
+    auto scale = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::float_type}}), scale_fp16);
+    auto bias = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::float_type}}), bias_fp16);
 
     auto mean = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), x);
     auto mean_bcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", dims}}), mean);
-    auto l0       = mm->add_instruction(migraphx::make_op("sqdiff"), x, mean_bcast);
-    auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l0);
-    auto l1       = mm->add_instruction(migraphx::make_op("sub"), x, mean_bcast);
+    auto l0       = mm->add_instruction(migraphx::make_op("sub"), x, mean_bcast);
+    auto l1       = mm->add_instruction(migraphx::make_op("sqdiff"), x, mean_bcast);
+    auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l1);
     auto epsilon_literal =
-        mm->add_literal(migraphx::literal{migraphx::shape{migraphx::shape::half_type}, {1e-5}});
+        mm->add_literal(migraphx::literal{migraphx::shape{migraphx::shape::float_type}, {1e-5}});
     auto epsilon_bcast = mm->add_instruction(
         migraphx::make_op("multibroadcast", {{"out_lens", dims}}), epsilon_literal);
     auto variance_bcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", dims}}), variance);
     auto l2          = mm->add_instruction(migraphx::make_op("add"), variance_bcast, epsilon_bcast);
     auto l3          = mm->add_instruction(migraphx::make_op("rsqrt"), l2);
-    auto l4          = mm->add_instruction(migraphx::make_op("mul"), l1, l3);
+    auto l4          = mm->add_instruction(migraphx::make_op("mul"), l0, l3);
     auto scale_bcast = mm->add_instruction(
         migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), scale);
     auto bias_bcast = mm->add_instruction(
         migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), bias);
-    auto l5 = mm->add_instruction(migraphx::make_op("mul"), l4, scale_bcast);
-    mm->add_instruction(migraphx::make_op("add"), l5, bias_bcast);
-
+    auto l5                 = mm->add_instruction(migraphx::make_op("mul"), l4, scale_bcast);
+    auto instance_norm_fp32 = mm->add_instruction(migraphx::make_op("add"), l5, bias_bcast);
+    mm->add_instruction(migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}),
+                        instance_norm_fp32);
     auto prog = optimize_onnx("instance_norm_half_test.onnx");
 
     EXPECT(p == prog);

test/verify/main.cpp

@@ -74,7 +74,9 @@ int main(int argc, const char* argv[])
                          "test_select_module_add",
                          "test_select_module_reduce",
                          "test_select_module_conv",
-                         "test_split_single_dyn_dim"});
+                         "test_split_single_dyn_dim",
+                         "test_instancenorm_large_3d<migraphx::shape::float_type>",
+                         "test_instancenorm_large_3d<migraphx::shape::half_type>"});
     rv.disable_test_for("gpu", {"test_conv_bn_add"});
     rv.run(argc, argv);
 }

test/verify/test_instancenorm.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 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 "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/common.hpp>
+#include <migraphx/make_op.hpp>
+
+migraphx::instruction_ref add_instancenorm(migraphx::module& m,
+                                           migraphx::instruction_ref x,
+                                           const std::vector<size_t>& dims,
+                                           float eps = 1e-5f)
+{
+    auto mgx_type = x->get_shape().type();
+    auto x_lens   = x->get_shape().lens();
+    std::vector<size_t> axes(x_lens.size() - 2);
+    std::iota(axes.begin(), axes.end(), 2);
+    auto scale   = m.add_parameter("scale", migraphx::shape{mgx_type, dims});
+    auto bias    = m.add_parameter("bias", migraphx::shape{mgx_type, dims});
+    auto epsilon = m.add_literal(migraphx::literal{migraphx::shape{mgx_type}, {eps}});
+
+    auto mean = m.add_instruction(migraphx::make_op("reduce_mean", {{"axes", axes}}), x);
+    auto mean_mbcast =
+        m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", x_lens}}), mean);
+    auto sub = m.add_instruction(migraphx::make_op("sub"), x, mean_mbcast);
+    auto l0  = m.add_instruction(migraphx::make_op("sqdiff"), {x, mean_mbcast});
+    auto var = m.add_instruction(migraphx::make_op("reduce_mean", {{"axes", axes}}), {l0});
+    auto epsilon_mbcast =
+        m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", x_lens}}), epsilon);
+    auto var_mbcast =
+        m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", x_lens}}), var);
+    auto add_epsilon = m.add_instruction(migraphx::make_op("add"), var_mbcast, epsilon_mbcast);
+    auto rsqrt       = m.add_instruction(migraphx::make_op("rsqrt"), add_epsilon);
+    auto l1          = m.add_instruction(migraphx::make_op("mul"), {sub, rsqrt});
+    auto scale_mbcast =
+        m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", x_lens}}), scale);
+    auto mul = m.add_instruction(migraphx::make_op("mul"), scale_mbcast, l1);
+    auto bias_mbcast =
+        m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", x_lens}}), bias);
+    return m.add_instruction(migraphx::make_op("add"), mul, bias_mbcast);
+}
+
+template <migraphx::shape::type_t TYPE>
+struct test_instancenorm : verify_program<test_instancenorm<TYPE>>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm                 = p.get_main_module();
+        std::vector<size_t> dims = {1, 2, 5, 5};
+        auto x                   = mm->add_parameter("x", migraphx::shape{TYPE, dims});
+        add_instancenorm(*mm, x, {1, 2, 1, 1});
+        return p;
+    }
+};
+template struct test_instancenorm<migraphx::shape::float_type>;
+template struct test_instancenorm<migraphx::shape::half_type>;
+
+template <migraphx::shape::type_t TYPE>
+struct test_instancenorm_large_3d : verify_program<test_instancenorm_large_3d<TYPE>>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm                 = p.get_main_module();
+        std::vector<size_t> dims = {1, 32, 64, 64, 64};
+        auto x                   = mm->add_parameter("x", migraphx::shape{TYPE, dims});
+        add_instancenorm(*mm, x, {1, 32, 1, 1, 1});
+        return p;
+    }
+};
+
+template struct test_instancenorm_large_3d<migraphx::shape::float_type>;
+template struct test_instancenorm_large_3d<migraphx::shape::half_type>;

test/verify/verify_program.hpp

@@ -27,7 +27,9 @@
 #include <functional>
 #include <migraphx/auto_register.hpp>
 #include <migraphx/program.hpp>
+#include <migraphx/stringutils.hpp>
 #include <migraphx/compile_options.hpp>
+#include <migraphx/ranges.hpp>
 
 struct program_info
 {
@@ -47,7 +49,18 @@ struct register_verify_program_action
     {
         T x;
         program_info pi;
-        pi.name            = migraphx::get_type_name<T>();
+        const std::string& test_type_name             = migraphx::get_type_name<T>();
+        const auto& split_name                        = migraphx::split_string(test_type_name, ':');
+        std::vector<std::string> name_without_version = {};
+        // test_type_name could contain internal namespace name with version_x_y_z i.e.
+        // test_instancenorm<migraphx::version_1::shape::float_type> remove version and construct
+        // test_name such as test_instancenorm<migraphx::shape::float_type>
+        std::copy_if(
+            split_name.begin(),
+            split_name.end(),
+            std::back_inserter(name_without_version),
+            [&](const auto& i) { return not i.empty() and not migraphx::contains(i, "version"); });
+        pi.name            = migraphx::trim(migraphx::join_strings(name_without_version, "::"));
         pi.section         = x.section();
         pi.get_program     = [x] { return x.create_program(); };
         pi.compile_options = x.get_compile_options();