merge develop branch

src/targets/gpu/lowering.cpp

@@ -26,7 +26,9 @@
 #include <migraphx/gpu/logsoftmax.hpp>
 #include <migraphx/gpu/add.hpp>
 #include <migraphx/gpu/sub.hpp>
+#include <migraphx/gpu/div.hpp>
 #include <migraphx/gpu/exp.hpp>
+#include <migraphx/gpu/erf.hpp>
 #include <migraphx/gpu/log.hpp>
 #include <migraphx/gpu/sin.hpp>
 #include <migraphx/gpu/cos.hpp>
@@ -51,6 +53,7 @@
 #include <migraphx/gpu/clip.hpp>
 #include <migraphx/gpu/reduce_sum.hpp>
 #include <migraphx/gpu/pow.hpp>
+#include <migraphx/gpu/reduce_mean.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>
@@ -87,6 +90,7 @@ struct miopen_apply
         add_generic_op<hip_add>("add");
         add_generic_op<hip_sub>("sub");
         add_generic_op<hip_exp>("exp");
+        add_generic_op<hip_erf>("erf");
         add_generic_op<hip_log>("log");
         add_generic_op<hip_sin>("sin");
         add_generic_op<hip_cos>("cos");
@@ -97,6 +101,7 @@ struct miopen_apply
         add_generic_op<hip_acos>("acos");
         add_generic_op<hip_atan>("atan");
         add_generic_op<hip_mul>("mul");
+        add_generic_op<hip_div>("div");
         add_generic_op<hip_max>("max");
         add_generic_op<hip_min>("min");
         add_generic_op<hip_pow>("pow");
@@ -113,6 +118,7 @@ struct miopen_apply
         add_extend_op<hip_convert, op::convert>("convert");
         add_extend_op<hip_clip, op::clip>("clip");
         add_extend_op<hip_reduce_sum, op::reduce_sum>("reduce_sum");
+        add_extend_op<hip_reduce_mean, op::reduce_mean>("reduce_mean");
 
         add_lrn_op();
         add_convolution_op();

src/targets/gpu/reduce_mean.cpp

+#include <migraphx/gpu/reduce_mean.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device/reduce_mean.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+shape hip_reduce_mean::compute_shape(std::vector<shape> inputs) const
+{
+    inputs.pop_back();
+    return op.compute_shape(inputs);
+}
+
+argument
+hip_reduce_mean::compute(context& ctx, const shape&, const std::vector<argument>& args) const
+{
+    device::reduce_mean(ctx.get_stream().get(), args.back(), args.front());
+    return args.back();
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

test/cpu_ops_test.cpp

@@ -527,6 +527,21 @@ TEST_CASE(exp_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(erf_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {4}};
+    auto l =
+        p.add_literal(migraphx::literal{s, {0.73785057, 1.58165966, -0.43597795, -0.01677432}});
+    p.add_instruction(migraphx::op::erf{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {0.70327317, 0.97470088, -0.46247893, -0.01892602};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 TEST_CASE(log_test)
 {
     migraphx::program p;
@@ -1703,7 +1718,7 @@ TEST_CASE(clip_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
-TEST_CASE(reduce_sum_test0)
+TEST_CASE(reduce_sum_axis0)
 {
     migraphx::program p;
     migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
@@ -1718,7 +1733,7 @@ TEST_CASE(reduce_sum_test0)
     EXPECT(results_vector == gold);
 }
 
-TEST_CASE(reduce_sum_test1)
+TEST_CASE(reduce_sum_axis1)
 {
     migraphx::program p;
     migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
@@ -1733,7 +1748,7 @@ TEST_CASE(reduce_sum_test1)
     EXPECT(results_vector == gold);
 }
 
-TEST_CASE(reduce_sum_test2)
+TEST_CASE(reduce_sum_axis2)
 {
     migraphx::program p;
     migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
@@ -1748,7 +1763,7 @@ TEST_CASE(reduce_sum_test2)
     EXPECT(results_vector == gold);
 }
 
-TEST_CASE(reduce_sum_test02)
+TEST_CASE(reduce_sum_axis02)
 {
     migraphx::program p;
     migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
@@ -1763,7 +1778,7 @@ TEST_CASE(reduce_sum_test02)
     EXPECT(results_vector == gold);
 }
 
-TEST_CASE(reduce_sum_test12)
+TEST_CASE(reduce_sum_axis12)
 {
     migraphx::program p;
     migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
@@ -1778,4 +1793,79 @@ TEST_CASE(reduce_sum_test12)
     EXPECT(results_vector == gold);
 }
 
+TEST_CASE(reduce_mean_axis1)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
+    auto input = migraphx::literal{s, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}};
+    auto l0    = p.add_literal(input);
+    p.add_instruction(migraphx::op::reduce_mean{{1}}, l0);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{2, 3, 6, 7, 10, 11};
+    EXPECT(results_vector == gold);
+}
+
+TEST_CASE(reduce_mean_axis2)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
+    auto input = migraphx::literal{s, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}};
+    auto l0    = p.add_literal(input);
+    p.add_instruction(migraphx::op::reduce_mean{{2}}, l0);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{1.5f, 3.5f, 5.5f, 7.5f, 9.5f, 11.5f};
+    EXPECT(results_vector == gold);
+}
+
+TEST_CASE(reduce_mean_axis02)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
+    auto input = migraphx::literal{s, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}};
+    auto l0    = p.add_literal(input);
+    p.add_instruction(migraphx::op::reduce_mean{{0, 2}}, l0);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{5.5, 7.5};
+    EXPECT(results_vector == gold);
+}
+
+TEST_CASE(reduce_mean_axis12)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3, 2, 2}};
+    auto input = migraphx::literal{s, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}};
+    auto l0    = p.add_literal(input);
+    p.add_instruction(migraphx::op::reduce_mean{{1, 2}}, l0);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{2.5f, 6.5f, 10.5f};
+    EXPECT(results_vector == gold);
+}
+
+TEST_CASE(reduce_mean_int)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::int32_type, {3, 2, 2}};
+    auto input = migraphx::literal{s, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}};
+    auto l0    = p.add_literal(input);
+    p.add_instruction(migraphx::op::reduce_mean{{1, 2}}, l0);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<int> gold{2, 6, 10};
+    EXPECT(results_vector == gold);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/miopen.cpp

@@ -243,6 +243,18 @@ struct test_exp : verify_program<test_exp>
     }
 };
 
+struct test_erf : verify_program<test_erf>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3, 4, 6}};
+        auto param = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::erf{}, param);
+        return p;
+    }
+};
+
 struct test_log : verify_program<test_log>
 {
     migraphx::program create_program() const
@@ -583,6 +595,38 @@ struct test_sub2 : verify_program<test_sub2>
     }
 };
 
+struct test_div : verify_program<test_div>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {3}};
+        auto x    = p.add_parameter("x", s);
+        auto y    = p.add_parameter("y", s);
+        auto z    = p.add_parameter("z", s);
+        auto diff = p.add_instruction(migraphx::op::div{}, x, y);
+        p.add_instruction(migraphx::op::div{}, diff, z);
+        return p;
+    }
+};
+
+struct test_div2 : verify_program<test_div2>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        migraphx::shape b{migraphx::shape::float_type, {3}};
+        auto x    = p.add_parameter("x", s);
+        auto y    = p.add_parameter("y", s);
+        auto z    = p.add_parameter("z", b);
+        auto zb   = p.add_instruction(migraphx::op::broadcast{1, s.lens()}, z);
+        auto diff = p.add_instruction(migraphx::op::div{}, x, y);
+        p.add_instruction(migraphx::op::div{}, diff, zb);
+        return p;
+    }
+};
+
 struct test_softmax1 : verify_program<test_softmax1>
 {
     migraphx::program create_program() const
@@ -3513,4 +3557,40 @@ struct test_reduce_sum_half : verify_program<test_reduce_sum_half>
     };
 };
 
+struct test_reduce_mean : verify_program<test_reduce_mean>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {3, 9, 4, 3}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::reduce_mean{{1}}, x);
+        return p;
+    };
+};
+
+struct test_reduce_mean_int : verify_program<test_reduce_mean_int>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::int32_type, {3, 1024, 8, 8}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::reduce_mean{{1}}, x);
+        return p;
+    };
+};
+
+struct test_reduce_mean_half : verify_program<test_reduce_mean_half>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::half_type, {3, 1024, 8, 8}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::reduce_mean{{2}}, x);
+        return p;
+    };
+};
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/onnx/erf_test.onnx

+erf-example:A
+
+
x
y"Erftest_erfZ
+
x
+

+
+
+b
+
y
+

+
+
+B

test/onnx/onnx_test.cpp

@@ -192,6 +192,16 @@ TEST_CASE(exp_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(erf_test)
+{
+    migraphx::program p;
+    auto input = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {10, 15}});
+    p.add_instruction(migraphx::op::erf{}, input);
+
+    auto prog = migraphx::parse_onnx("erf_test.onnx");
+    EXPECT(p == prog);
+}
+
 TEST_CASE(log_test)
 {
     migraphx::program p;
@@ -848,6 +858,27 @@ TEST_CASE(reducesum_test3)
     EXPECT(p == prog);
 }
 
+TEST_CASE(reducemean_test1)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {3, 4, 5, 6}});
+    auto l1 = p.add_instruction(migraphx::op::reduce_mean{{2, 3}}, l0);
+    p.add_instruction(migraphx::op::squeeze{{2, 3}}, l1);
+    auto prog = migraphx::parse_onnx("reducemean_test1.onnx");
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(reducemean_test2)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {3, 4, 5, 6}});
+    p.add_instruction(migraphx::op::reduce_mean{{2}}, l0);
+    auto prog = migraphx::parse_onnx("reducemean_test2.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(clip_test)
 {
     migraphx::program p;

test/onnx/reducemean_test2.onnx

+reducemean-example:}
+0
+
x
y"
+ReduceMean*
+axes@
*
+keepdims

test_reducemeanZ
+
x
+

+
+
+
+b
+
y
+

+
+
+

+B

test/op_shape_test.cpp

@@ -390,37 +390,111 @@ TEST_CASE(softmax) { test_softmax_variations<migraphx::op::softmax>(); }
 
 TEST_CASE(logsoftmax) { test_softmax_variations<migraphx::op::logsoftmax>(); }
 
-template <class T>
-void test_argop_var()
+TEST_CASE(test_argmax)
 {
     {
         migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
-        expect_shape(migraphx::shape{migraphx::shape::int64_type, {1, 3, 4, 5}}, T{0}, input);
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {1, 3, 4, 5}},
+                     migraphx::op::argmax{0},
+                     input);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 1, 4, 5}},
+                     migraphx::op::argmax{1},
+                     input);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 3, 1, 5}},
+                     migraphx::op::argmax{2},
+                     input);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 3, 4, 1}},
+                     migraphx::op::argmax{3},
+                     input);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        throws_shape(migraphx::op::argmax{4}, input);
     }
+}
 
+TEST_CASE(test_argmin)
+{
     {
         migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
-        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 1, 4, 5}}, T{1}, input);
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {1, 3, 4, 5}},
+                     migraphx::op::argmin{0},
+                     input);
     }
 
     {
         migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
-        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 3, 1, 5}}, T{2}, input);
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 1, 4, 5}},
+                     migraphx::op::argmin{1},
+                     input);
     }
 
     {
         migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
-        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 3, 4, 1}}, T{3}, input);
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 3, 1, 5}},
+                     migraphx::op::argmin{2},
+                     input);
     }
 
+    {
+        migraphx::shape input{migraphx::shape::half_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::int64_type, {2, 3, 4, 1}},
+                     migraphx::op::argmin{3},
+                     input);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        throws_shape(migraphx::op::argmin{4}, input);
+    }
+}
+
+template <class T>
+void test_reduce_ops()
+{
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::float_type, {1, 1, 1, 1}}, T{}, input);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        expect_shape(
+            migraphx::shape{migraphx::shape::float_type, {1, 1, 1, 1}}, T{{0, 1, 2, 3}}, input);
+    }
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::float_type, {2, 3, 1, 1}}, T{{2, 3}}, input);
+    }
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::float_type, {1, 3, 4, 5}}, T{{0}}, input);
+    }
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        expect_shape(migraphx::shape{migraphx::shape::float_type, {2, 3, 4, 1}}, T{{-1}}, input);
+    }
     {
         migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
-        throws_shape(T{4}, input);
+        throws_shape(T{{4}}, input);
     }
 }
 
-TEST_CASE(argmax) { test_argop_var<migraphx::op::argmax>(); }
-TEST_CASE(argmin) { test_argop_var<migraphx::op::argmin>(); }
+TEST_CASE(reduce_sum) { test_reduce_ops<migraphx::op::reduce_sum>(); }
+TEST_CASE(reduce_mean) { test_reduce_ops<migraphx::op::reduce_mean>(); }
 
 // 2 inputs arguments
 TEST_CASE(matmul)