Skip to content

GitLab

"examples/vscode:/vscode.git/clone" did not exist on "5d8eb93eeec0476b9f0fddc96f2960be0ce782b6"
Commit 2ab23275 authored by Alan Turner's avatar Alan Turner
Browse files

Merge remote-tracking branch 'origin/develop' into ck-integration-tuning

parents 2c3563dd 013d4829
Show whitespace changes
Inline Side-by-side
Showing with 282 additions and 26 deletions
test/onnx/onnx_test.cpp
View file @ 2ab23275
......@@ -3174,28 +3174,64 @@ TEST_CASE(instance_norm_test)
auto bias = mm->add_parameter("2", s2);
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("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 l1 = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto l0 = add_common_op(*mm, migraphx::make_op("sqdiff"), {x, mean});
auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l0);
auto epsilon_literal =
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 l2 = add_common_op(*mm, migraphx::make_op("add"), {variance, epsilon_literal});
auto l3 = mm->add_instruction(migraphx::make_op("rsqrt"), l2);
auto l4 = mm->add_instruction(migraphx::make_op("mul"), l0, l3);
auto l4 = add_common_op(*mm, migraphx::make_op("mul"), {l1, 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 ret = mm->add_instruction(migraphx::make_op("add"), l5, bias_bcast);
mm->add_return({ret});
migraphx::onnx_options options;
auto prog = migraphx::parse_onnx("instance_norm_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(instance_norm_dyn_batch_test)
{
// instancenorm with dynamic input in the 0'th (batch) dimension
migraphx::shape s1{migraphx::shape::float_type, {{1, 2, {2}}, {2, 2}, {3, 3}, {3, 3}}};
migraphx::shape s2{migraphx::shape::float_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 mean = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), x);
auto l1 = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto l0 = add_common_op(*mm, migraphx::make_op("sqdiff"), {x, mean});
auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l0);
auto epsilon_literal =
mm->add_literal(migraphx::literal{migraphx::shape{migraphx::shape::float_type}, {1e-5}});
auto l2 = add_common_op(*mm, migraphx::make_op("add"), {variance, epsilon_literal});
auto l3 = mm->add_instruction(migraphx::make_op("rsqrt"), l2);
auto l4 = add_common_op(*mm, migraphx::make_op("mul"), {l1, l3});
auto scale_bcast = mm->add_instruction(migraphx::make_op("broadcast", {{"axis", 1}}), scale, x);
auto bias_bcast = mm->add_instruction(migraphx::make_op("broadcast", {{"axis", 1}}), bias, x);
auto l5 = mm->add_instruction(migraphx::make_op("mul"), l4, scale_bcast);
auto ret = mm->add_instruction(migraphx::make_op("add"), l5, bias_bcast);
mm->add_return({ret});
auto prog = optimize_onnx("instance_norm_test.onnx");
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 2, {2}};
auto prog = migraphx::parse_onnx("instance_norm_dyn_batch_test.onnx", options);
EXPECT(p == prog);
}
......@@ -3212,6 +3248,7 @@ TEST_CASE(instance_norm_half_test)
auto scale_fp16 = mm->add_parameter("1", s2);
auto bias_fp16 = mm->add_parameter("2", s2);
// conversion of half type to float is enabled by default
auto x = mm->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::float_type}}), x_fp16);
auto scale = mm->add_instruction(
......@@ -3220,20 +3257,19 @@ TEST_CASE(instance_norm_half_test)
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("sub"), x, mean_bcast);
auto l1 = mm->add_instruction(migraphx::make_op("sqdiff"), x, mean_bcast);
auto l0 = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto l1 = add_common_op(*mm, migraphx::make_op("sqdiff"), {x, mean});
auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l1);
// type of epsilon_literal is same as 0'th input; convert instruction will be added by
// add_common_op
auto epsilon_literal =
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 l2 = add_common_op(*mm, migraphx::make_op("add"), {variance, epsilon_literal});
auto l3 = mm->add_instruction(migraphx::make_op("rsqrt"), l2);
auto l4 = mm->add_instruction(migraphx::make_op("mul"), l0, l3);
auto l4 = add_common_op(*mm, 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(
......@@ -3247,6 +3283,55 @@ TEST_CASE(instance_norm_half_test)
EXPECT(p == prog);
}
TEST_CASE(instance_norm_dyn_batch_half_test)
{
// instancenorm with half type, dynamic input in the 0'th (batch) dimension
migraphx::shape s1{migraphx::shape::half_type, {{1, 2, {2}}, {2, 2}, {3, 3}, {3, 3}}};
migraphx::shape s2{migraphx::shape::half_type, {2}};
migraphx::program p;
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);
// conversion of half type to float is enabled by default
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 l0 = add_common_op(*mm, migraphx::make_op("sub"), {x, mean});
auto l1 = add_common_op(*mm, migraphx::make_op("sqdiff"), {x, mean});
auto variance = mm->add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2, 3}}}), l1);
// type of epsilon_literal is same as 0'th input; convert instruction will be added by
// add_common_op
auto epsilon_literal =
mm->add_literal(migraphx::literal{migraphx::shape{migraphx::shape::float_type}, {1e-5}});
auto l2 = add_common_op(*mm, migraphx::make_op("add"), {variance, epsilon_literal});
auto l3 = mm->add_instruction(migraphx::make_op("rsqrt"), l2);
auto l4 = add_common_op(*mm, migraphx::make_op("mul"), {l0, l3});
auto scale_bcast = mm->add_instruction(migraphx::make_op("broadcast", {{"axis", 1}}), scale, x);
auto bias_bcast = mm->add_instruction(migraphx::make_op("broadcast", {{"axis", 1}}), bias, x);
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);
auto ret = mm->add_instruction(
migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}),
instance_norm_fp32);
mm->add_return({ret});
migraphx::onnx_options options;
options.default_dyn_dim_value = {1, 2, {2}};
auto prog = migraphx::parse_onnx("instance_norm_dyn_batch_half_test.onnx", options);
EXPECT(p == prog);
}
TEST_CASE(instance_norm_type_mismatch_test)
{
EXPECT(test::throws([&] { migraphx::parse_onnx("instance_norm_type_mismatch_test.onnx"); }));
......
test/onnx/verify_onnx.cpp
View file @ 2ab23275
......@@ -880,6 +880,48 @@ TEST_CASE(instance_norm_test)
EXPECT(migraphx::verify_range(result_vector, gold));
}
TEST_CASE(instance_norm_dyn_batch_test)
{
migraphx::program p = migraphx::parse_onnx("instance_norm_dyn_batch_test.onnx");
p.compile(migraphx::make_target("ref"));
migraphx::shape s0{migraphx::shape::float_type, {1, 2, 3, 3}};
std::vector<float> data0 = {0, 1, 2, 3, 4, 5, 6, 7, 8, 0, 1, 2, 3, 4, 5, 6, 7, 8};
migraphx::shape s1{migraphx::shape::float_type, {2}};
std::vector<float> data1 = {1, 2};
migraphx::shape s2{migraphx::shape::float_type, {2}};
std::vector<float> data2 = {0, 1};
migraphx::parameter_map pp;
pp["0"] = migraphx::argument(s0, data0.data());
pp["1"] = migraphx::argument(s1, data1.data());
pp["2"] = migraphx::argument(s2, data2.data());
auto result = p.eval(pp).back();
std::vector<float> result_vector;
result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
std::vector<float> gold = {-1.54919,
-1.16189,
-0.774596,
-0.387298,
0,
0.387298,
0.774596,
1.16189,
1.54919,
-2.09838,
-1.32379,
-0.549192,
0.225404,
1,
1.7746,
2.54919,
3.32379,
4.09838};
EXPECT(migraphx::verify_range(result_vector, gold));
}
TEST_CASE(instance_norm_3d_test)
{
migraphx::program p = migraphx::parse_onnx("instance_norm_val_3d_test.onnx");
......
test/op_shape_test.cpp
View file @ 2ab23275
......@@ -187,6 +187,14 @@ TEST_CASE(broadcast_axis_out_of_range_error)
throws_shape(migraphx::make_op("broadcast", {{"axis", 4}, {"out_lens", lens}}), input);
}
TEST_CASE(broadcast_1in_dyn_error)
{
// broadcast doesn't support single dynamic shape input
std::vector<std::size_t> lens{3, 2, 4, 3};
migraphx::shape input{migraphx::shape::float_type, {{1, 4}, {4, 4}, {2, 2}}};
throws_shape(migraphx::make_op("broadcast", {{"axis", 2}, {"out_lens", lens}}), input);
}
TEST_CASE(broadcast_2in_static_static)
{
migraphx::shape a_input{migraphx::shape::float_type, {4}, {1}};
......@@ -1434,6 +1442,14 @@ TEST_CASE(multibroadcast)
}
}
TEST_CASE(multibroadcast_1in_dyn_error_0)
{
// multibroadcast doesn't support single dynamic shape input
std::vector<std::size_t> lens{4, 4, 1, 3};
migraphx::shape input{migraphx::shape::float_type, {{1, 4}, {4, 4}, {4, 4}}};
throws_shape(migraphx::make_op("multibroadcast", {{"out_lens", lens}}), input);
}
TEST_CASE(multibroadcast_2in_static_dyn0)
{
migraphx::shape a_shape{migraphx::shape::float_type, {4, 4}};
......@@ -2208,6 +2224,119 @@ TEST_CASE(reshape_shape)
}
}
// This uses the permutation to compute the reshape since its simpler than
// trying to calculate strides. As we collapse or expand dimensions, we
// remove the collapsed dimensions or duplicate the expanded dimensions in
// the permutation. Then we renumber the permutation. So for dimensions of 4,
// 24, 1, 1, 1 with a permutation of 1, 0, 2, 3, 4 that reshapes to 4, 1, 3,
// 4, 2, we first remove the collapsed dimensions or duplicate the expanded
// dimensions which gives 1, 0, 0, 0, 0. Then after renumbering we get a
// final permutation of 4, 0, 1, 2, 3.
TEST_CASE(reshape_nonstandard)
{
auto input = migraphx::shape::from_permutation(migraphx::shape::float_type,
{4, 24, 1, 1, 1},
migraphx::invert_permutation({1, 0, 2, 3, 4}));
std::vector<std::pair<std::vector<std::size_t>, std::vector<int64_t>>> tests{
{{4, 24}, {1, 0}},
{{4, 24, 1, 1, 1, 1}, {1, 0, 2, 3, 4, 5}},
{{4, 8, 3, 1, 1}, {2, 0, 1, 3, 4}},
{{4, 1, 3, 4, 2}, {4, 0, 1, 2, 3}},
{{4, 1, 4, 3, 2}, {4, 0, 1, 2, 3}},
{{4, 2, 4, 3}, {3, 0, 1, 2}},
{{4, 2, 12, 1}, {2, 0, 1, 3}},
{{4, 2, 1, 12}, {3, 0, 1, 2}},
{{4, 4, 2, 3}, {3, 0, 1, 2}},
{{4, 8, 1, 3}, {3, 0, 1, 2}},
{{4, 8, 3, 1}, {2, 0, 1, 3}}};
for(const auto& [dims, perm] : tests)
{
migraphx::shape output = migraphx::shape::from_permutation(
migraphx::shape::float_type, dims, migraphx::invert_permutation(perm));
expect_shape(output, migraphx::make_op("reshape", {{"dims", dims}}), input);
}
}
TEST_CASE(reshape_nonstandard_squeeze)
{
auto input = migraphx::shape::from_permutation(
migraphx::shape::float_type, {2, 16, 16, 1280}, migraphx::invert_permutation({0, 2, 3, 1}));
std::vector<std::size_t> lens = {2, 256, 1280};
migraphx::shape output = migraphx::shape::from_permutation(
migraphx::shape::float_type, lens, migraphx::invert_permutation({0, 2, 1}));
expect_shape(output, migraphx::make_op("reshape", {{"dims", lens}}), input);
}
TEST_CASE(reshape_nonstandard_error)
{
auto input = migraphx::shape::from_permutation(migraphx::shape::float_type,
{4, 24, 1, 1, 1},
migraphx::invert_permutation({1, 0, 2, 3, 4}));
for(auto&& new_shape : std::vector<std::vector<int64_t>>{{4, 8, 3, 2, 2},
{1},
{4, 8, 4},
{4, 24, 1, 1, 1, 1, 2},
{8, 4, 4},
{4, 1, 3, -1, -1},
{4, 3, 0},
{4, 3, 2},
{3, 0},
{3, 2}})
{
throws_shape(migraphx::make_op("reshape", {{"dims", new_shape}}), input);
}
}
TEST_CASE(reshape_nonpacked_unsqueeze1)
{
migraphx::shape input{migraphx::shape::float_type, {4, 16}, {32, 2}};
migraphx::shape output{migraphx::shape::float_type, {4, 2, 8}, {32, 16, 2}};
expect_shape(output, migraphx::make_op("reshape", {{"dims", output.lens()}}), input);
}
TEST_CASE(reshape_nonpacked_unsqueeze2)
{
migraphx::shape input{migraphx::shape::float_type, {4, 16}, {32, 2}};
migraphx::shape output{migraphx::shape::float_type, {2, 2, 16}, {64, 32, 2}};
expect_shape(output, migraphx::make_op("reshape", {{"dims", output.lens()}}), input);
}
TEST_CASE(reshape_nonpacked_squeeze)
{
migraphx::shape input{migraphx::shape::float_type, {4, 16}, {32, 2}};
migraphx::shape output{migraphx::shape::float_type, {64}, {2}};
expect_shape(output, migraphx::make_op("reshape", {{"dims", output.lens()}}), input);
}
TEST_CASE(reshape_broadcast_unsqueeze1)
{
migraphx::shape input{migraphx::shape::float_type, {2, 256, 1280}, {0, 0, 1}};
migraphx::shape output{migraphx::shape::float_type, {2, 16, 16, 1280}, {0, 0, 0, 1}};
expect_shape(output, migraphx::make_op("reshape", {{"dims", output.lens()}}), input);
}
TEST_CASE(reshape_broadcast_unsqueeze2)
{
migraphx::shape input{migraphx::shape::float_type, {2, 256, 1280}, {0, 0, 1}};
migraphx::shape output{migraphx::shape::float_type, {2, 256, 16, 80}, {0, 0, 80, 1}};
expect_shape(output, migraphx::make_op("reshape", {{"dims", output.lens()}}), input);
}
TEST_CASE(reshape_broadcast_squeeze)
{
migraphx::shape input{migraphx::shape::float_type, {2, 16, 16, 1280}, {0, 0, 0, 1}};
migraphx::shape output{migraphx::shape::float_type, {2, 256, 1280}, {0, 0, 1}};
expect_shape(output, migraphx::make_op("reshape", {{"dims", output.lens()}}), input);
}
TEST_CASE(reshape_broadcast_squeeze_error)
{
migraphx::shape input{migraphx::shape::float_type, {2, 16, 16, 1280}, {0, 0, 0, 1}};
std::vector<int64_t> new_shape = {2, 16, 20480};
throws_shape(migraphx::make_op("reshape", {{"dims", new_shape}}), input);
}
TEST_CASE(reshape_dyn_shape)
{
migraphx::shape input{migraphx::shape::float_type, {{1, 4}, {24, 24}, {1, 1}, {1, 1}}};
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment