Merge branch 'develop' into ck-flash-attn

test/onnx/const_of_shape_no_value_attr_test.onnx

-constant-of-shape:

+!const_of_shape_no_value_attr_test:

 6shape"Constant*#
-value**Bshape_tensor 

+value*:Bshape_tensor

 
-shape
y"ConstantOfShapeconstant_of_shapeb
+shape
y"ConstantOfShape!const_of_shape_no_value_attr_testb
 
y
 

 
 
-B
+B
\ No newline at end of file

test/onnx/gen_onnx.py

@@ -582,6 +582,29 @@ def cast_test():
     return ([node], [x], [y])
 
 
+@onnx_test()
+def castlike_test():
+    input = helper.make_tensor_value_info('0', TensorProto.FLOAT16, [10])
+    target_type = helper.make_tensor_value_info('1', TensorProto.FLOAT, [10])
+    output = helper.make_tensor_value_info('out', TensorProto.FLOAT, [10])
+
+    node = onnx.helper.make_node('CastLike',
+                                 inputs=['0', '1'],
+                                 outputs=['out'])
+
+    return ([node], [input, target_type], [output])
+
+
+@onnx_test()
+def castlike_error_test():
+    input = helper.make_tensor_value_info('0', TensorProto.FLOAT16, [10])
+    output = helper.make_tensor_value_info('out', TensorProto.FLOAT, [10])
+
+    node = onnx.helper.make_node('CastLike', inputs=['0'], outputs=['out'])
+
+    return ([node], [input], [output])
+
+
 @onnx_test()
 def ceil_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10])
@@ -1007,9 +1030,9 @@ def const_of_shape_empty_input_test():
                                          [10])
     empty_val = np.array([]).astype(np.int64)
     empty_ts = helper.make_tensor(name='empty_tensor',
-                                  data_type=TensorProto.INT32,
+                                  data_type=TensorProto.INT64,
                                   dims=empty_val.shape,
-                                  vals=empty_val.flatten().astype(int))
+                                  vals=empty_val.flatten().astype(np.int64))
     shape_const = helper.make_node(
         'Constant',
         inputs=[],
@@ -1035,9 +1058,9 @@ def const_of_shape_float_test():
 
     shape_val = np.array([2, 3, 4]).astype(np.int64)
     shape_ts = helper.make_tensor(name='shape_tensor',
-                                  data_type=TensorProto.INT32,
+                                  data_type=TensorProto.INT64,
                                   dims=shape_val.shape,
-                                  vals=shape_val.flatten().astype(int))
+                                  vals=shape_val.flatten().astype(np.int64))
 
     shape_const = helper.make_node(
         'Constant',
@@ -1055,22 +1078,44 @@ def const_of_shape_float_test():
     return ([shape_const, node], [], [y])
 
 
+@onnx_test()
+def const_of_shape_default_test():
+    shape_val = np.array([2, 3, 4]).astype(np.int64)
+    shape_ts = helper.make_tensor(name='shape_tensor',
+                                  data_type=TensorProto.INT64,
+                                  dims=shape_val.shape,
+                                  vals=shape_val.flatten().astype(np.int64))
+    shape_const = helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=['shape'],
+        value=shape_ts,
+    )
+    y = helper.make_tensor_value_info('y', TensorProto.INT64, [2, 3, 4])
+
+    node = onnx.helper.make_node('ConstantOfShape',
+                                 inputs=['shape'],
+                                 outputs=['y'])
+
+    return ([shape_const, node], [], [y])
+
+
 @onnx_test()
 def const_of_shape_int64_test():
     tensor_val = onnx.helper.make_tensor('value', onnx.TensorProto.INT64, [1],
                                          [10])
     shape_val = np.array([2, 3, 4]).astype(np.int64)
     shape_ts = helper.make_tensor(name='shape_tensor',
-                                  data_type=TensorProto.INT32,
+                                  data_type=TensorProto.INT64,
                                   dims=shape_val.shape,
-                                  vals=shape_val.flatten().astype(int))
+                                  vals=shape_val.flatten().astype(np.int64))
     shape_const = helper.make_node(
         'Constant',
         inputs=[],
         outputs=['shape'],
         value=shape_ts,
     )
-    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3, 4])
+    y = helper.make_tensor_value_info('y', TensorProto.INT64, [2, 3, 4])
 
     node = onnx.helper.make_node('ConstantOfShape',
                                  inputs=['shape'],
@@ -1084,9 +1129,9 @@ def const_of_shape_int64_test():
 def const_of_shape_no_value_attr_test():
     shape_val = np.array([2, 3, 4]).astype(np.int64)
     shape_ts = helper.make_tensor(name='shape_tensor',
-                                  data_type=TensorProto.INT32,
+                                  data_type=TensorProto.INT64,
                                   dims=shape_val.shape,
-                                  vals=shape_val.flatten().astype(int))
+                                  vals=shape_val.flatten().astype(np.int64))
     shape_const = helper.make_node(
         'Constant',
         inputs=[],
@@ -1104,6 +1149,40 @@ def const_of_shape_no_value_attr_test():
     return ([shape_const, node], [], [y])
 
 
+@onnx_test()
+def const_of_shape_dyn_float_test():
+    tensor_val = onnx.helper.make_tensor('value', onnx.TensorProto.FLOAT, [1],
+                                         [10])
+
+    output_dims = helper.make_tensor_value_info('output_dims',
+                                                TensorProto.INT64, [3])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 3, 4])
+
+    node = onnx.helper.make_node('ConstantOfShape',
+                                 inputs=['output_dims'],
+                                 outputs=['y'],
+                                 value=tensor_val)
+
+    return ([node], [output_dims], [y])
+
+
+@onnx_test()
+def const_of_shape_dyn_int64_test():
+    tensor_val = onnx.helper.make_tensor('value', onnx.TensorProto.INT64, [1],
+                                         [10])
+
+    output_dims = helper.make_tensor_value_info('output_dims',
+                                                TensorProto.INT64, [3])
+    y = helper.make_tensor_value_info('y', TensorProto.INT64, [2, 3, 4])
+
+    node = onnx.helper.make_node('ConstantOfShape',
+                                 inputs=['output_dims'],
+                                 outputs=['y'],
+                                 value=tensor_val)
+
+    return ([node], [output_dims], [y])
+
+
 @onnx_test()
 def conv_1d_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 5])

test/onnx/onnx_test.cpp

@@ -687,6 +687,26 @@ TEST_CASE(cast_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(castlike_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l   = mm->add_parameter("0", migraphx::shape{migraphx::shape::half_type, {10}});
+    mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {10}});
+    mm->add_instruction(
+        migraphx::make_op("convert",
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
+        l);
+
+    auto prog = optimize_onnx("castlike_test.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(castlike_error_test)
+{
+    EXPECT(test::throws([&] { migraphx::parse_onnx("castlike_error_test.onnx"); }));
+}
+
 TEST_CASE(ceil_test)
 {
     migraphx::program p;
@@ -1040,11 +1060,25 @@ TEST_CASE(constant_one_val_int64_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(const_of_shape_default_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape output_dims_shape(migraphx::shape::int64_type, {3});
+    mm->add_literal(migraphx::literal(output_dims_shape, {2, 3, 4}));
+    migraphx::shape output_shape{migraphx::shape::float_type, {2, 3, 4}};
+    std::vector<float> vec(output_shape.elements(), 0.0);
+    mm->add_literal(migraphx::literal(output_shape, vec));
+
+    auto prog = optimize_onnx("const_of_shape_default_test.onnx");
+    EXPECT(p == prog);
+}
+
 TEST_CASE(const_of_shape_empty_input_test)
 {
     migraphx::program p;
     auto* mm = p.get_main_module();
-    mm->add_literal(migraphx::literal(migraphx::shape::int32_type));
+    mm->add_literal(migraphx::literal(migraphx::shape::int64_type));
     migraphx::shape s(migraphx::shape::int64_type, {1}, {0});
     std::vector<int64_t> vec(s.elements(), 10);
     mm->add_literal(migraphx::literal(s, vec));
@@ -1057,7 +1091,7 @@ TEST_CASE(const_of_shape_float_test)
 {
     migraphx::program p;
     auto* mm = p.get_main_module();
-    migraphx::shape ss(migraphx::shape::int32_type, {3});
+    migraphx::shape ss(migraphx::shape::int64_type, {3});
     mm->add_literal(migraphx::literal(ss, {2, 3, 4}));
     migraphx::shape s(migraphx::shape::float_type, {2, 3, 4});
     std::vector<float> vec(s.elements(), 10.0f);
@@ -1071,8 +1105,10 @@ TEST_CASE(const_of_shape_int64_test)
 {
     migraphx::program p;
     auto* mm = p.get_main_module();
-    migraphx::shape ss(migraphx::shape::int32_type, {3});
+    // output_dims
+    migraphx::shape ss(migraphx::shape::int64_type, {3});
     mm->add_literal(migraphx::literal(ss, {2, 3, 4}));
+    // constant shape literal
     migraphx::shape s(migraphx::shape::int64_type, {2, 3, 4});
     std::vector<int64_t> vec(s.elements(), 10);
     mm->add_literal(migraphx::literal(s, vec));
@@ -1085,7 +1121,7 @@ TEST_CASE(const_of_shape_no_value_attr_test)
 {
     migraphx::program p;
     auto* mm = p.get_main_module();
-    migraphx::shape ss(migraphx::shape::int32_type, {3});
+    migraphx::shape ss(migraphx::shape::int64_type, {3});
     mm->add_literal(migraphx::literal(ss, {2, 3, 4}));
     migraphx::shape s(migraphx::shape::float_type, {2, 3, 4});
     std::vector<float> vec(s.elements(), 0.0f);
@@ -1095,6 +1131,42 @@ TEST_CASE(const_of_shape_no_value_attr_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(const_of_shape_dyn_float_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto od_param =
+        mm->add_parameter("output_dims", migraphx::shape{migraphx::shape::int64_type, {3}});
+    auto alloc_ins = mm->add_instruction(
+        migraphx::make_op("allocate", {{"buf_type", migraphx::shape::float_type}}), od_param);
+    migraphx::shape dv_shape(migraphx::shape::float_type, {1}, {0});
+    auto dv_lit   = mm->add_literal(migraphx::literal(dv_shape, {10}));
+    auto fill_ins = mm->add_instruction(migraphx::make_op("fill"), dv_lit, alloc_ins);
+    mm->add_return({fill_ins});
+
+    migraphx::onnx_options options;
+    auto prog = parse_onnx("const_of_shape_dyn_float_test.onnx", options);
+    EXPECT(p == prog);
+}
+
+TEST_CASE(const_of_shape_dyn_int64_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto od_param =
+        mm->add_parameter("output_dims", migraphx::shape{migraphx::shape::int64_type, {3}});
+    auto alloc_ins = mm->add_instruction(
+        migraphx::make_op("allocate", {{"buf_type", migraphx::shape::int64_type}}), od_param);
+    migraphx::shape dv_shape(migraphx::shape::int64_type, {1}, {0});
+    auto dv_lit   = mm->add_literal(migraphx::literal(dv_shape, {10}));
+    auto fill_ins = mm->add_instruction(migraphx::make_op("fill"), dv_lit, alloc_ins);
+    mm->add_return({fill_ins});
+
+    migraphx::onnx_options options;
+    auto prog = parse_onnx("const_of_shape_dyn_int64_test.onnx", options);
+    EXPECT(p == prog);
+}
+
 TEST_CASE(conv_autopad_fail_test)
 {
     EXPECT(test::throws([&] { optimize_onnx("conv_autopad_fail_test.onnx"); }));

test/py/test_gpu_async.py

 #####################################################################################
 # The MIT License (MIT)
 #
-# Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+# 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
@@ -23,10 +23,55 @@
 #####################################################################################
 import migraphx
 import ctypes
+import os
+import glob
 
 
 def test_conv_relu():
-    hip = ctypes.cdll.LoadLibrary("libamdhip64.so")
+    # Full path of the library is needed to fix an issue on sles
+    # where the library is not loaded otherwise.
+    # We check for the presence of library at the following paths,
+    # in the order listed below:
+    #
+    # 1. 'rocm_path' environment variable
+    # 2. /opt/rocm
+    # 3. /opt/rocm-*
+    #
+    # If the library is not found at any of these paths, we fall back
+    # to the library path being detected automatically.
+
+    library = "libamdhip64.so"
+
+    # Environment variable containing path to rocm
+    rocm_path_env_var = "rocm_path"
+
+    # Check for rocm_path, default to /opt/rocm if it does not exist.
+    rocm_path_var = os.getenv(rocm_path_env_var, default="/opt/rocm")
+
+    # Join the paths to the library to get full path,
+    # e.g. /opt/rocm/lib/libamdhip64.so
+    library_file = os.path.join(rocm_path_var, "lib", library)
+
+    # Check if the library file exists at the specified path
+    if os.path.exists(library_file):
+        # Replace library name by full path to the library
+        library = library_file
+    else:
+        # Pattern match to look for path to different
+        # rocm versions: /opt/rocm-*
+        rocm_path_pattern = "/opt/rocm-*/lib/libamdhip64.so"
+        matching_libraries = glob.glob(rocm_path_pattern)
+
+        if matching_libraries:
+            # Replace library name by full path to the first
+            # library found.
+            library = matching_libraries[0]
+
+    # Loads library either by using the full path to the
+    # library, if it has been detected earlier,
+    # or, proceeds to load the library based on the name
+    # of the library.
+    hip = ctypes.cdll.LoadLibrary(library)
 
     p = migraphx.parse_onnx("conv_relu_maxpool_test.onnx")
     print(p)

tools/accuracy/accuracy_checker.py

@@ -82,6 +82,27 @@ def parse_args():
                         default=False,
                         help='Turn on ort VERBOSE logging via session options')
 
+    parser.add_argument(
+        '--disable-offload-copy',
+        dest="offload_copy",
+        action='store_false',
+        default=True,
+        help=
+        'Disable offload copying (user must handle copy to and from device)')
+
+    parser.add_argument(
+        '--disable-fast-math',
+        dest="fast_math",
+        action='store_false',
+        default=True,
+        help='Disable fast math optimizations (etc: rewrite_gelu)')
+
+    parser.add_argument('--exhaustive_tune',
+                        dest="exhaustive_tune",
+                        action='store_true',
+                        default=False,
+                        help='Enable exhaustive tuning for solutions')
+
     args = parser.parse_args()
 
     return args
@@ -177,7 +198,12 @@ def main():
         print(model)
 
     if not args.ort_run:
-        model.compile(migraphx.get_target(args.target))
+        model.compile(
+            migraphx.get_target(args.target),
+            offload_copy=args.offload_copy,
+            fast_math=args.fast_math,
+            exhaustive_tune=args.exhaustive_tune,
+        )
 
     params = {}
     test_inputs = {}