Merge branch 'propogate-broadcast' into multiply-add

src/targets/gpu/include/migraphx/gpu/device/argmin.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_ARGMIN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_ARGMIN_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void argmin(hipStream_t stream, const argument& result, const argument& arg, int64_t axis);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/device/erf.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_ERF_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_ERF_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void erf(hipStream_t stream, const argument& result, const argument& arg);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/erf.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_ERF_HPP
+#define MIGRAPHX_GUARD_RTGLIB_ERF_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/erf.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_erf : unary_device<hip_erf, device::erf>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/lowering.cpp

@@ -11,6 +11,8 @@
 #include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/add.hpp>
 #include <migraphx/iterator_for.hpp>
+#include <migraphx/gpu/argmax.hpp>
+#include <migraphx/gpu/argmin.hpp>
 #include <migraphx/gpu/rocblas.hpp>
 #include <migraphx/gpu/context.hpp>
 #include <migraphx/gpu/convolution.hpp>
@@ -25,6 +27,7 @@
 #include <migraphx/gpu/add.hpp>
 #include <migraphx/gpu/sub.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>
@@ -84,6 +87,7 @@ 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");
@@ -102,6 +106,8 @@ struct miopen_apply
         add_extend_op<hip_concat, op::concat>("concat");
         add_extend_op<hip_softmax, op::softmax>("softmax");
         add_extend_op<hip_logsoftmax, op::logsoftmax>("logsoftmax");
+        add_extend_op<hip_argmax, op::argmax>("argmax");
+        add_extend_op<hip_argmin, op::argmin>("argmin");
         add_extend_op<hip_gather, op::gather>("gather");
         add_extend_op<hip_pad, op::pad>("pad");
         add_extend_op<hip_convert, op::convert>("convert");

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;
@@ -941,9 +956,6 @@ TEST_CASE(softmax_simple_test)
     auto result = p.eval({});
     std::vector<float> results_vector(2);
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
-    for(auto v : results_vector)
-        std::cout << v << "\t";
-    std::cout << std::endl;
     EXPECT(migraphx::verify_range(results_vector, s));
 }
 
@@ -1138,6 +1150,114 @@ TEST_CASE(logsoftmax_test_axis_3)
     EXPECT(migraphx::verify_range(results_vector, s));
 }
 
+TEST_CASE(argmax_test_0)
+{
+    migraphx::program p;
+    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
+                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
+                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
+    std::vector<int64_t> res_gold = {0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 1};
+    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
+    auto dl = p.add_literal(migraphx::literal{data_shape, data});
+    p.add_instruction(migraphx::op::argmax{0}, dl);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int64_t> result_vec;
+    result.visit([&](auto output) { result_vec.assign(output.begin(), output.end()); });
+
+    EXPECT(migraphx::verify_range(result_vec, res_gold));
+}
+
+TEST_CASE(argmax_test_1)
+{
+    migraphx::program p;
+    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
+                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
+                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
+    std::vector<int64_t> res_gold = {0, 0, 2, 1, 2, 0, 0, 2};
+    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
+    auto dl = p.add_literal(migraphx::literal{data_shape, data});
+    p.add_instruction(migraphx::op::argmax{1}, dl);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int64_t> result_vec;
+    result.visit([&](auto output) { result_vec.assign(output.begin(), output.end()); });
+
+    EXPECT(migraphx::verify_range(result_vec, res_gold));
+}
+
+TEST_CASE(argmax_test_2)
+{
+    migraphx::program p;
+    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
+                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
+                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
+    std::vector<int64_t> res_gold = {1, 3, 2, 2, 2, 3};
+    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
+    auto dl = p.add_literal(migraphx::literal{data_shape, data});
+    p.add_instruction(migraphx::op::argmax{2}, dl);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int64_t> result_vec;
+    result.visit([&](auto output) { result_vec.assign(output.begin(), output.end()); });
+
+    EXPECT(migraphx::verify_range(result_vec, res_gold));
+}
+
+TEST_CASE(argmin_test_0)
+{
+    migraphx::program p;
+    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
+                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
+                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
+    std::vector<int64_t> res_gold = {1, 1, 0, 1, 0, 1, 1, 1, 0, 0, 1, 0};
+    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
+    auto dl = p.add_literal(migraphx::literal{data_shape, data});
+    p.add_instruction(migraphx::op::argmin{0}, dl);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int64_t> result_vec;
+    result.visit([&](auto output) { result_vec.assign(output.begin(), output.end()); });
+
+    EXPECT(migraphx::verify_range(result_vec, res_gold));
+}
+
+TEST_CASE(argmin_test_1)
+{
+    migraphx::program p;
+    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
+                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
+                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
+    std::vector<int64_t> res_gold = {2, 2, 0, 2, 0, 1, 2, 0};
+    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
+    auto dl = p.add_literal(migraphx::literal{data_shape, data});
+    p.add_instruction(migraphx::op::argmin{1}, dl);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int64_t> result_vec;
+    result.visit([&](auto output) { result_vec.assign(output.begin(), output.end()); });
+
+    EXPECT(migraphx::verify_range(result_vec, res_gold));
+}
+
+TEST_CASE(argmin_test_2)
+{
+    migraphx::program p;
+    std::vector<float> data = {1.2255,  1.6834,  -2.0305, -0.3221, 0.4701,  0.2583, 0.7545, 2.5758,
+                               -1.6849, 0.0928,  0.9022,  -0.8765, -0.4090, 0.9301, 2.0724, -1.5706,
+                               0.4867,  -0.1493, 0.6957,  -0.2179, 0.7142,  0.7177, 0.0183, 1.3497};
+    std::vector<int64_t> res_gold = {2, 1, 0, 3, 3, 2};
+    migraphx::shape data_shape{migraphx::shape::float_type, {2, 3, 4}};
+    auto dl = p.add_literal(migraphx::literal{data_shape, data});
+    p.add_instruction(migraphx::op::argmin{2}, dl);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<int64_t> result_vec;
+    result.visit([&](auto output) { result_vec.assign(output.begin(), output.end()); });
+
+    EXPECT(migraphx::verify_range(result_vec, res_gold));
+}
+
 TEST_CASE(conv2d_test)
 {
     migraphx::program p;

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
@@ -615,6 +627,30 @@ template struct test_softmax<1, migraphx::shape::half_type>;
 template struct test_softmax<2, migraphx::shape::half_type>;
 template struct test_softmax<3, migraphx::shape::half_type>;
 
+template <class T, int Axis>
+struct test_arg_ops : verify_program<test_arg_ops<T, Axis>>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3, 4, 1025}};
+        auto param = p.add_parameter("data", s);
+        p.add_instruction(T{Axis}, param);
+
+        return p;
+    }
+};
+
+template struct test_arg_ops<migraphx::op::argmax, 0>;
+template struct test_arg_ops<migraphx::op::argmax, 1>;
+template struct test_arg_ops<migraphx::op::argmax, 2>;
+template struct test_arg_ops<migraphx::op::argmax, 3>;
+
+template struct test_arg_ops<migraphx::op::argmin, 0>;
+template struct test_arg_ops<migraphx::op::argmin, 1>;
+template struct test_arg_ops<migraphx::op::argmin, 2>;
+template struct test_arg_ops<migraphx::op::argmin, 3>;
+
 struct test_conv : verify_program<test_conv>
 {
     migraphx::program create_program() const
@@ -3348,22 +3384,6 @@ struct test_lstm_bidirct_default_actv2 : verify_program<test_lstm_bidirct_defaul
     }
 };
 
-template <int Axis>
-struct test_logsoftmax_1 : verify_program<test_logsoftmax_1<Axis>>
-{
-    migraphx::program create_program() const
-    {
-        migraphx::program p;
-        migraphx::shape s{migraphx::shape::float_type, {3}};
-        auto param = p.add_parameter("0", s);
-        p.add_instruction(migraphx::op::logsoftmax{Axis}, param);
-
-        return p;
-    }
-};
-
-template struct test_logsoftmax_1<0>;
-
 template <int Axis, migraphx::shape::type_t T>
 struct test_logsoftmax : verify_program<test_logsoftmax<Axis, T>>
 {

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;
@@ -784,6 +794,28 @@ TEST_CASE(logsoftmax)
     EXPECT(p == prog);
 }
 
+TEST_CASE(argmax)
+{
+    migraphx::program p;
+    auto l0  = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {3, 4, 5, 6}});
+    auto ins = p.add_instruction(migraphx::op::argmax{2}, l0);
+    p.add_instruction(migraphx::op::squeeze{{2}}, ins);
+    auto prog = migraphx::parse_onnx("argmax_test.onnx");
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(argmin)
+{
+    migraphx::program p;
+    auto l0  = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {3, 4, 5, 6}});
+    auto ins = p.add_instruction(migraphx::op::argmin{3}, l0);
+    p.add_instruction(migraphx::op::squeeze{{3}}, ins);
+    auto prog = migraphx::parse_onnx("argmin_test.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(no_pad_test)
 {
     migraphx::program p;
@@ -794,6 +826,38 @@ TEST_CASE(no_pad_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(reducesum_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_sum{{2}}, l0);
+    p.add_instruction(migraphx::op::squeeze{{2}}, l1);
+    auto prog = migraphx::parse_onnx("reducesum_test1.onnx");
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(reducesum_test2)
+{
+    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_sum{{2, 3}}, l0);
+    p.add_instruction(migraphx::op::squeeze{{2, 3}}, l1);
+    auto prog = migraphx::parse_onnx("reducesum_test2.onnx");
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(reducesum_test3)
+{
+    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_sum{{2, 3}}, l0);
+    auto prog = migraphx::parse_onnx("reducesum_test3.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(clip_test)
 {
     migraphx::program p;

test/onnx/reducesum_test3.onnx

+reducesum-example:}
+1
+
x
y"	ReduceSum*
+axes@@
*
+keepdims

test_reducesumZ
+
x
+

+
+
+
+b
+
y
+

+
+
+

+
B

test/op_shape_test.cpp

@@ -390,6 +390,38 @@ 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()
+{
+    {
+        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);
+    }
+
+    {
+        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);
+    }
+
+    {
+        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);
+    }
+
+    {
+        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);
+    }
+
+    {
+        migraphx::shape input{migraphx::shape::float_type, {2, 3, 4, 5}};
+        throws_shape(T{4}, input);
+    }
+}
+
+TEST_CASE(argmax) { test_argop_var<migraphx::op::argmax>(); }
+TEST_CASE(argmin) { test_argop_var<migraphx::op::argmin>(); }
+
 // 2 inputs arguments
 TEST_CASE(matmul)
 {