Merge

test/simplify_algebra_test.cpp

@@ -2910,6 +2910,179 @@ TEST_CASE(reorder_reshape_slice_not_apply)
     EXPECT(m1.sort() == m2.sort());
 }
 
+TEST_CASE(reorder_reshape_slice_multi_rsp)
+{
+    migraphx::module m1;
+    {
+        migraphx::shape s{migraphx::shape::float_type, {4, 128, 3, 32, 80}};
+        auto input = m1.add_parameter("input", s);
+        auto t1    = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {2, 0, 3, 1, 4}}}), input);
+        auto slc0 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), t1);
+        auto slc1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {1}}, {"ends", {2}}}), t1);
+        auto slc2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {2}}, {"ends", {3}}}), t1);
+
+        auto c1_1 = m1.add_instruction(migraphx::make_op("contiguous"), slc1);
+        auto c2_1 = m1.add_instruction(migraphx::make_op("contiguous"), slc2);
+
+        auto c1 = m1.add_instruction(migraphx::make_op("contiguous"), slc1);
+        auto r1 =
+            m1.add_instruction(migraphx::make_op("reshape", {{"dims", {4, 32, 128, 80}}}), c1);
+
+        auto c2 = m1.add_instruction(migraphx::make_op("contiguous"), slc2);
+        auto r2 =
+            m1.add_instruction(migraphx::make_op("reshape", {{"dims", {4, 32, 128, 80}}}), c2);
+
+        auto r1_1 =
+            m1.add_instruction(migraphx::make_op("reshape", {{"dims", {128, 128, 80}}}), c1_1);
+        auto r2_1 =
+            m1.add_instruction(migraphx::make_op("reshape", {{"dims", {128, 128, 80}}}), c2_1);
+
+        auto c0 = m1.add_instruction(migraphx::make_op("contiguous"), slc0);
+        auto r0 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", {128, 128, 80}}}), c0);
+
+        auto t2 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1}}}), r1_1);
+        auto c_t2 = m1.add_instruction(migraphx::make_op("contiguous"), t2);
+
+        auto dot = m1.add_instruction(migraphx::make_op("dot"), r0, c_t2);
+
+        m1.add_return({r1, r2, dot, r2_1});
+    };
+
+    migraphx::module m2;
+    {
+        migraphx::shape s{migraphx::shape::float_type, {4, 128, 3, 32, 80}};
+        auto input = m2.add_parameter("input", s);
+        auto t1    = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {2, 0, 3, 1, 4}}}), input);
+        auto c_t1 = m2.add_instruction(migraphx::make_op("contiguous"), t1);
+        auto rsp1 =
+            m2.add_instruction(migraphx::make_op("reshape", {{"dims", {384, 128, 80}}}), c_t1);
+        auto slc0 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {256}}, {"ends", {384}}}), rsp1);
+        auto slc1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {128}}, {"ends", {256}}}), rsp1);
+
+        auto t_slc1 =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1}}}), slc1);
+        auto c_t_slc1 = m2.add_instruction(migraphx::make_op("contiguous"), t_slc1);
+
+        auto slc2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {128}}}), rsp1);
+
+        auto dot = m2.add_instruction(migraphx::make_op("dot"), slc2, c_t_slc1);
+
+        auto c_t1_1 = m2.add_instruction(migraphx::make_op("contiguous"), t1);
+        auto rsp2 =
+            m2.add_instruction(migraphx::make_op("reshape", {{"dims", {12, 32, 128, 80}}}), c_t1_1);
+
+        auto slc2_1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {4}}, {"ends", {8}}}), rsp2);
+
+        auto slc2_2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {8}}, {"ends", {12}}}), rsp2);
+
+        m2.add_return({slc2_1, slc2_2, dot, slc0});
+    };
+
+    run_pass(m1);
+    EXPECT(m1.sort() == m2.sort());
+}
+
+TEST_CASE(reorder_reshape_slice_partial)
+{
+    migraphx::module m1;
+    {
+        migraphx::shape s{migraphx::shape::float_type, {128, 96}};
+        auto input = m1.add_parameter("input", s);
+        auto slc0  = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {8}}}), input);
+        auto slc1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {8}}, {"ends", {16}}}), input);
+        auto slc2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {16}}, {"ends", {24}}}), input);
+        auto slc3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {24}}, {"ends", {128}}}), input);
+
+        auto c0 = m1.add_instruction(migraphx::make_op("contiguous"), slc0);
+        auto c1 = m1.add_instruction(migraphx::make_op("contiguous"), slc1);
+        auto c2 = m1.add_instruction(migraphx::make_op("contiguous"), slc2);
+
+        std::vector<int64_t> lens = {2, 4, 96};
+        auto r0 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c0);
+        auto r1 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c1);
+        auto r2 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c2);
+
+        auto sum = m1.add_instruction(migraphx::make_op("add"), r0, r1);
+        auto ret = m1.add_instruction(migraphx::make_op("mul"), sum, r2);
+        m1.add_return({ret, slc3});
+    };
+
+    migraphx::module m2;
+    {
+        migraphx::shape s{migraphx::shape::float_type, {128, 96}};
+        auto input = m2.add_parameter("input", s);
+        auto rsp = m2.add_instruction(migraphx::make_op("reshape", {{"dims", {32, 4, 96}}}), input);
+        auto slc3 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {24}}, {"ends", {128}}}), input);
+
+        auto slc0 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {2}}}), rsp);
+        auto slc1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {2}}, {"ends", {4}}}), rsp);
+        auto slc2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {4}}, {"ends", {6}}}), rsp);
+
+        auto sum = m2.add_instruction(migraphx::make_op("add"), slc0, slc1);
+        auto ret = m2.add_instruction(migraphx::make_op("mul"), sum, slc2);
+        m2.add_return({ret, slc3});
+    };
+
+    run_pass(m1);
+    EXPECT(m1.sort() == m2.sort());
+}
+
+TEST_CASE(reorder_reshape_slice_uneven_slice)
+{
+    auto create_p = [] {
+        migraphx::module m;
+        migraphx::shape s{migraphx::shape::float_type, {128, 96}};
+        auto input = m.add_parameter("input", s);
+        auto slc0  = m.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {31}}}), input);
+        auto slc1 = m.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {31}}, {"ends", {62}}}), input);
+        auto slc2 = m.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {62}}, {"ends", {93}}}), input);
+        auto slc3 = m.add_instruction(
+            migraphx::make_op("slice", {{"axes", {0}}, {"starts", {93}}, {"ends", {128}}}), input);
+
+        auto c0 = m.add_instruction(migraphx::make_op("contiguous"), slc0);
+        auto c1 = m.add_instruction(migraphx::make_op("contiguous"), slc1);
+        auto c2 = m.add_instruction(migraphx::make_op("contiguous"), slc2);
+
+        std::vector<int64_t> lens = {1, 31, 96};
+        auto r0 = m.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c0);
+        auto r1 = m.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c1);
+        auto r2 = m.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c2);
+
+        auto sum = m.add_instruction(migraphx::make_op("add"), r0, r1);
+        auto ret = m.add_instruction(migraphx::make_op("mul"), sum, r2);
+        m.add_return({ret, slc3});
+
+        return m;
+    };
+
+    auto m1 = create_p();
+    auto m2 = m1;
+    run_pass(m1);
+    EXPECT(m1.sort() == m2.sort());
+}
+
 template <std::size_t BS>
 void reorder_reshape_slice_diff_dims()
 {
@@ -2931,13 +3104,32 @@ void reorder_reshape_slice_diff_dims()
         std::vector<int64_t> lens  = {static_cast<int64_t>(BS), 32, 3, 32};
         std::vector<int64_t> lens1 = {static_cast<int64_t>(BS), 48, 2, 32};
         auto r0 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c0);
-        auto r1 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c1);
-        auto r2 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens1}}), c2);
+        auto r1 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens1}}), c1);
+        auto r2 = m1.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), c2);
 
         m1.add_return({r0, r1, r2});
     };
 
-    auto m2 = m1;
+    migraphx::module m2;
+    {
+        auto s     = migraphx::shape{migraphx::shape::float_type, {BS, 96, 96}};
+        auto input = m2.add_parameter("input", s);
+        auto slc1  = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {32}}, {"ends", {64}}}), input);
+        auto c1                    = m2.add_instruction(migraphx::make_op("contiguous"), slc1);
+        std::vector<int64_t> lens1 = {static_cast<int64_t>(BS), 48, 2, 32};
+        auto r1 = m2.add_instruction(migraphx::make_op("reshape", {{"dims", lens1}}), c1);
+
+        std::vector<int64_t> lens = {static_cast<int64_t>(BS), 32, 3, 96};
+        auto r_new = m2.add_instruction(migraphx::make_op("reshape", {{"dims", lens}}), input);
+        auto slc0  = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {3}}, {"starts", {0}}, {"ends", {32}}}), r_new);
+        auto slc2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {3}}, {"starts", {64}}, {"ends", {96}}}), r_new);
+
+        m2.add_return({slc0, r1, slc2});
+    };
+
     run_pass(m1);
     EXPECT(m1.sort() == m2.sort());
 }

test/simplify_qdq_test.cpp

@@ -479,11 +479,11 @@ TEST_CASE(conv_pooling_dot)
         auto q1  = add_quantize_op(m1, "quantizelinear", input, scale, zero);
         auto d5  = add_quantize_op(m1, "dequantizelinear", q1, scale, zero);
         auto c1  = m1.add_instruction(migraphx::make_op("convolution",
-                                                       {{"padding", {0, 0, 0, 0}},
-                                                        {"stride", {1, 1}},
-                                                        {"dilation", {1, 1}},
-                                                        {"group", 1},
-                                                        {"padding_mode", 0}}),
+                                                        {{"padding", {0, 0, 0, 0}},
+                                                         {"stride", {1, 1}},
+                                                         {"dilation", {1, 1}},
+                                                         {"group", 1},
+                                                         {"padding_mode", 0}}),
                                      d5,
                                      d1);
         auto bc1 = m1.add_instruction(
@@ -526,11 +526,11 @@ TEST_CASE(conv_pooling_dot)
         auto d3  = add_quantize_op(m2, "dequantizelinear", ab, scale, zero);
         auto q1  = add_quantize_op(m2, "quantizelinear", input, scale, zero);
         auto c1  = m2.add_instruction(migraphx::make_op("quant_convolution",
-                                                       {{"padding", {0, 0, 0, 0}},
-                                                        {"stride", {1, 1}},
-                                                        {"dilation", {1, 1}},
-                                                        {"group", 1},
-                                                        {"padding_mode", 0}}),
+                                                        {{"padding", {0, 0, 0, 0}},
+                                                         {"stride", {1, 1}},
+                                                         {"dilation", {1, 1}},
+                                                         {"group", 1},
+                                                         {"padding_mode", 0}}),
                                      q1,
                                      weights);
         auto d5  = add_quantize_op(m2, "dequantizelinear", c1, scale1);
@@ -585,11 +585,11 @@ TEST_CASE(mobilenet_snippet)
         auto q1  = add_quantize_op(mm, "quantizelinear", input, scale, zero);
         auto d5  = add_quantize_op(mm, "dequantizelinear", q1, scale, zero);
         auto c1  = mm.add_instruction(migraphx::make_op("convolution",
-                                                       {{"padding", {0, 0, 0, 0}},
-                                                        {"stride", {1, 1}},
-                                                        {"dilation", {1, 1}},
-                                                        {"group", 1},
-                                                        {"padding_mode", 0}}),
+                                                        {{"padding", {0, 0, 0, 0}},
+                                                         {"stride", {1, 1}},
+                                                         {"dilation", {1, 1}},
+                                                         {"group", 1},
+                                                         {"padding_mode", 0}}),
                                      d5,
                                      d1);
         auto bc1 = mm.add_instruction(
@@ -700,7 +700,7 @@ TEST_CASE(conv_correctness)
     auto result2 = p2.eval({{"input", input}, {"weights", weights}}).back();
     std::vector<float> rv2(16);
     result2.visit([&](auto output) { rv2.assign(output.begin(), output.end()); });
-    EXPECT(migraphx::verify::verify_range(rv1, rv2));
+    EXPECT(migraphx::verify::verify_rms_range(rv1, rv2));
 }
 
 TEST_CASE(dot_correctness)
@@ -750,7 +750,7 @@ TEST_CASE(dot_correctness)
     auto result2 = p2.eval({{"a", a}, {"b", b}}).back();
     std::vector<float> rv2(sh3.elements());
     result2.visit([&](auto output) { rv2.assign(output.begin(), output.end()); });
-    EXPECT(migraphx::verify::verify_range(rv1, rv2));
+    EXPECT(migraphx::verify::verify_rms_range(rv1, rv2));
 }
 
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/targets.cpp

 /*
  * 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

test/verify/run_verify.cpp

@@ -251,7 +251,8 @@ void run_verify::verify(const std::string& name,
             std::size_t num = gold.size();
             for(std::size_t i = 0; ((i < num) and passed); ++i)
             {
-                passed &= migraphx::verify_args(tname, gold[i], result[i]);
+                passed &= migraphx::verify_args_with_tolerance(
+                    tname, result[i], migraphx::verify::expected{gold[i]});
             }
 
             if(not passed or migraphx::enabled(MIGRAPHX_TRACE_TEST{}))

test/verify/test_layernorm.cpp

@@ -49,7 +49,8 @@ migraphx::instruction_ref add_layernorm(migraphx::module& m,
     auto pow            = m.add_instruction(migraphx::make_op("pow"), sub, exponent_mbcast);
     auto var            = m.add_instruction(migraphx::make_op("reduce_mean", {{"axes", {2}}}), pow);
     auto epsilon_mbcast = m.add_instruction(
-        migraphx::make_op("multibroadcast", {{"out_lens", {1, dims.at(1), 1}}}), epsilon);
+        migraphx::make_op("multibroadcast", {{"out_lens", {dims.at(0), dims.at(1), 1}}}), epsilon);
+
     auto add_epsilon = m.add_instruction(migraphx::make_op("add"), var, epsilon_mbcast);
     auto sqrt        = m.add_instruction(migraphx::make_op("sqrt"), add_epsilon);
     auto sqrt_mbcast =
@@ -57,7 +58,8 @@ migraphx::instruction_ref add_layernorm(migraphx::module& m,
     auto div = m.add_instruction(migraphx::make_op("div"), sub, sqrt_mbcast);
     auto scale_mbcast =
         m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", dims}}), scale);
-    auto mul = m.add_instruction(migraphx::make_op("mul"), scale_mbcast, div);
+    auto mul = m.add_instruction(migraphx::make_op("mul"), div, scale_mbcast);
+
     auto bias_mbcast =
         m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", dims}}), bias);
     return m.add_instruction(migraphx::make_op("add"), mul, bias_mbcast);
@@ -161,3 +163,21 @@ struct test_layernorm_triadd_large : verify_program<test_layernorm_triadd_large>
         return p;
     }
 };
+
+struct test_add_layernorm_add_gemm_nonstd : verify_program<test_add_layernorm_add_gemm_nonstd>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        auto s =
+            migraphx::shape::from_permutation(migraphx::shape::float_type, {8, 1, 16}, {1, 2, 0});
+        auto x = mm->add_parameter("x", s);
+        auto y = mm->add_parameter("y", s);
+        auto z = mm->add_parameter("z", migraphx::shape{migraphx::shape::float_type, {8, 16, 64}});
+        auto add           = mm->add_instruction(migraphx::make_op("add"), x, y);
+        auto layernorm_ins = add_layernorm(*mm, add, s.lens());
+        mm->add_instruction(migraphx::make_op("dot"), layernorm_ins, z);
+        return p;
+    }
+};

tools/license_stamper.py

@@ -2,7 +2,7 @@
 #####################################################################################
 #  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
@@ -38,7 +38,7 @@ def getipynb_markdownBlockAsList():
         '\t\t"cell_type": "code",\n', '\t\t"execution_count": null,\n',
         '\t\t"metadata": {},\n', '\t\t"outputs": [],\n', '\t\t"source": [\n',
         '\t\t\t\"#  The MIT License (MIT)\\n\",\n', '\t\t\t\"#\\n\",\n',
-        '\t\t\t\"#  Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.\\n\",\n',
+        '\t\t\t\"#  Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.\\n\",\n',
         '\t\t\t\"#\\n\",\n',
         '\t\t\t\"#  Permission is hereby granted, free of charge, to any person obtaining a copy\\n\",\n',
         '\t\t\t\"#  of this software and associated documentation files (the \'Software\'), to deal\\n\",\n',

tools/test_runner.py

@@ -39,6 +39,15 @@ def parse_args():
                         type=str,
                         default='gpu',
                         help='Specify where the tests execute (ref, gpu)')
+    parser.add_argument('--fp16', action='store_true', help='Quantize to fp16')
+    parser.add_argument('--atol',
+                        type=float,
+                        default=1e-3,
+                        help='The absolute tolerance parameter')
+    parser.add_argument('--rtol',
+                        type=float,
+                        default=1e-3,
+                        help='The relative tolerance parameter')
     args = parser.parse_args()
 
     return args
@@ -257,6 +266,8 @@ def main():
 
     # read and compile model
     model = migraphx.parse_onnx(model_path_name, map_input_dims=param_shapes)
+    if args.fp16:
+        migraphx.quantize_fp16(model)
     model.compile(migraphx.get_target(target))
 
     # get test cases
@@ -279,7 +290,10 @@ def main():
         output_data = run_one_case(model, input_data)
 
         # check output correctness
-        ret = check_correctness(gold_outputs, output_data)
+        ret = check_correctness(gold_outputs,
+                                output_data,
+                                atol=args.atol,
+                                rtol=args.rtol)
         if ret:
             correct_num += 1