Merge branch 'develop' into sqlite3_windows

test/onnx/gen_onnx.py

@@ -5151,6 +5151,223 @@ def qlinearadd_bcast_test():
             [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
 
 
+@onnx_test()
+def qlinearconv_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__QLinearConv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 7, 7])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [], [0.00369204697])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [132])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 1, 1], [0])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [0.00172794575])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [255])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.00162681262])
+    zero_pt_y = helper.make_tensor('7', TensorProto.UINT8, [], [123])
+
+    out = helper.make_tensor_value_info('out', TensorProto.UINT8, [1, 1, 7, 7])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_pad_1_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 3, 3],
+                            [1, 1, 1, 1, 1, 1, 1, 1, 1])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [1.0])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.UINT8, [], [0])
+
+    out = helper.make_tensor_value_info('out', TensorProto.UINT8, [1, 1, 5, 5])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[1, 1, 1, 1],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_pad_0_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 3, 3],
+                            [1, 1, 1, 1, 1, 1, 1, 1, 1])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [1.0])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.INT8, [], [-128])
+
+    out = helper.make_tensor_value_info('out', TensorProto.INT8, [1, 1, 3, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[0, 0, 0, 0],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_scale_1D_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor(
+        '3', TensorProto.UINT8, [2, 1, 3, 3],
+        [1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [2], [1.0, 0.5])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [2], [0, 0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.INT8, [], [-128])
+
+    out = helper.make_tensor_value_info('out', TensorProto.INT8, [1, 2, 3, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[0, 0, 0, 0],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearglobalavgpool_test():
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 3, 4, 4])
+
+    sc_x = helper.make_tensor('X_scale', TensorProto.FLOAT, [], [0.05])
+    z_pt_x = helper.make_tensor('X_zero_point', TensorProto.UINT8, [], [128])
+
+    y = helper.make_tensor_value_info('Y', TensorProto.UINT8, [1, 3, 1, 1])
+
+    sc_y = helper.make_tensor('Y_scale', TensorProto.FLOAT, [], [0.025])
+    z_pt_y = helper.make_tensor('Y_zero_point', TensorProto.UINT8, [], [64])
+
+    n = onnx.helper.make_node(
+        'QLinearGlobalAveragePool',
+        inputs=['X', 'X_scale', 'X_zero_point', 'Y_scale', 'Y_zero_point'],
+        outputs=['Y'],
+        channels_last=0,
+    )
+
+    return ([n], [x], [y], [sc_x, z_pt_x, sc_y, z_pt_y])
+
+
+def qlinearmatmul_1D_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [8])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [], [0])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [8])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [128])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [], [64])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [1])
+
+    node = onnx.helper.make_node(
+        'QLinearMatMul',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
+@onnx_test()
+def qlinearmatmul_2D_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [1, 8])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [], [0])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [8, 1])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [128])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [], [64])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [1, 1])
+
+    node = onnx.helper.make_node(
+        'QLinearMatMul',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
+@onnx_test()
+def qlinearmatmul_3D_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [2, 2, 4])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.0066])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [],
+                                   [113])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [2, 4, 3])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.00705])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [114])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.0107])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [],
+                                   [118])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [2, 2, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearMatMul',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
 @onnx_test()
 def quantizelinear_test():
     arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT, [5])

test/onnx/onnx_test.cpp

@@ -4904,6 +4904,218 @@ TEST_CASE(qlinearadd_test)
     EXPECT(p.sort() == prog.sort());
 }
 
+TEST_CASE(qlinearconv_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    auto x = mm->add_parameter("X", {migraphx::shape::uint8_type, {1, 1, 7, 7}});
+
+    auto sc_x   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.00369204697}});
+    auto z_pt_x = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {132}});
+
+    auto w = mm->add_literal(
+        migraphx::literal{migraphx::shape{migraphx::shape::uint8_type, {1, 1, 1, 1}}, {0}});
+
+    auto sc_w   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.00172794575}});
+    auto z_pt_w = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {255}});
+
+    auto sc_y   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.00162681262}});
+    auto z_pt_y = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {123}});
+
+    auto scale_x_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 7, 7}}}), sc_x);
+
+    auto z_pt_x_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 7, 7}}}), z_pt_x);
+
+    auto fp_x =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), x, scale_x_bcast, z_pt_x_bcast);
+
+    auto scale_w_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 1, 1}}}), sc_w);
+
+    auto z_pt_w_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 1, 1}}}), z_pt_w);
+
+    auto fp_w =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), w, scale_w_bcast, z_pt_w_bcast);
+
+    auto fp_y = mm->add_instruction(migraphx::make_op("convolution"), fp_x, fp_w);
+
+    auto scale_y_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 7, 7}}}), sc_y);
+
+    auto z_pt_y_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 1, 7, 7}}}), z_pt_y);
+
+    auto y =
+        mm->add_instruction(migraphx::make_op("quantizelinear"), fp_y, scale_y_bcast, z_pt_y_bcast);
+
+    mm->add_return({y});
+
+    auto prog = migraphx::parse_onnx("qlinearconv_test.onnx");
+
+    EXPECT(p.sort() == prog.sort());
+}
+
+TEST_CASE(qlinearglobalavgpool_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    auto x = mm->add_parameter("X", {migraphx::shape::uint8_type, {1, 3, 4, 4}});
+
+    auto sc_x   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_x = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {128}});
+
+    auto sc_y   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.025}});
+    auto z_pt_y = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {64}});
+
+    auto scale_x_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 3, 4, 4}}}), sc_x);
+
+    auto z_pt_x_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 3, 4, 4}}}), z_pt_x);
+
+    auto fp_x =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), x, scale_x_bcast, z_pt_x_bcast);
+
+    auto fp_y =
+        mm->add_instruction(migraphx::make_op("pooling",
+                                              {{"mode", migraphx::op::pooling_mode::average},
+                                               {"padding", {0, 0, 0, 0}},
+                                               {"lengths", {4, 4}}}),
+                            fp_x);
+
+    auto scale_y_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 3, 1, 1}}}), sc_y);
+
+    auto z_pt_y_bcast = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", {1, 3, 1, 1}}}), z_pt_y);
+
+    auto y =
+        mm->add_instruction(migraphx::make_op("quantizelinear"), fp_y, scale_y_bcast, z_pt_y_bcast);
+
+    mm->add_return({y});
+
+    auto prog = migraphx::parse_onnx("qlinearglobalavgpool_test.onnx");
+
+    EXPECT(p.sort() == prog.sort());
+}
+
+TEST_CASE(qlinearmatmul_1D_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    auto a = mm->add_parameter("A", {migraphx::shape::uint8_type, {8}});
+    auto b = mm->add_parameter("B", {migraphx::shape::uint8_type, {8}});
+
+    auto sc_a   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_a = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {0}});
+
+    auto sc_b   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_b = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {128}});
+
+    auto sc_c   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_c = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {64}});
+
+    auto scale_a_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {8}}}), sc_a);
+
+    auto z_pt_a_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {8}}}), z_pt_a);
+
+    auto fp_a =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), a, scale_a_bcast, z_pt_a_bcast);
+
+    auto scale_b_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {8}}}), sc_b);
+
+    auto z_pt_b_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {8}}}), z_pt_b);
+
+    auto fp_b =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), b, scale_b_bcast, z_pt_b_bcast);
+
+    auto sq_a = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {0}}}), fp_a);
+
+    auto sq_b = mm->add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1}}}), fp_b);
+
+    auto fp_c = mm->add_instruction(migraphx::make_op("dot"), sq_a, sq_b);
+
+    auto sq_c = mm->add_instruction(migraphx::make_op("squeeze", {{"axes", {0}}}), fp_c);
+
+    auto scale_c_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1}}}), sc_c);
+
+    auto z_pt_c_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1}}}), z_pt_c);
+
+    auto c =
+        mm->add_instruction(migraphx::make_op("quantizelinear"), sq_c, scale_c_bcast, z_pt_c_bcast);
+
+    mm->add_return({c});
+
+    auto prog = migraphx::parse_onnx("qlinearmatmul_1D_test.onnx");
+
+    EXPECT(p.sort() == prog.sort());
+}
+
+TEST_CASE(qlinearmatmul_2D_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    auto a = mm->add_parameter("A", {migraphx::shape::uint8_type, {1, 8}});
+    auto b = mm->add_parameter("B", {migraphx::shape::uint8_type, {8, 1}});
+
+    auto sc_a   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_a = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {0}});
+
+    auto sc_b   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_b = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {128}});
+
+    auto sc_c   = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0.05}});
+    auto z_pt_c = mm->add_literal(migraphx::literal{migraphx::shape::uint8_type, {64}});
+
+    auto scale_a_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 8}}}), sc_a);
+
+    auto z_pt_a_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 8}}}), z_pt_a);
+
+    auto fp_a =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), a, scale_a_bcast, z_pt_a_bcast);
+
+    auto scale_b_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {8, 1}}}), sc_b);
+
+    auto z_pt_b_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {8, 1}}}), z_pt_b);
+
+    auto fp_b =
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), b, scale_b_bcast, z_pt_b_bcast);
+
+    auto fp_c = mm->add_instruction(migraphx::make_op("dot"), fp_a, fp_b);
+
+    auto scale_c_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 1}}}), sc_c);
+
+    auto z_pt_c_bcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 1}}}), z_pt_c);
+
+    auto c =
+        mm->add_instruction(migraphx::make_op("quantizelinear"), fp_c, scale_c_bcast, z_pt_c_bcast);
+
+    mm->add_return({c});
+
+    auto prog = migraphx::parse_onnx("qlinearmatmul_2D_test.onnx");
+
+    EXPECT(p.sort() == prog.sort());
+}
+
 TEST_CASE(quantizelinear_test)
 {
     migraphx::program p;

test/onnx/qlinearmatmul_3D_test.onnx

+qlinearmatmul_3D_test:
+]
+
A
+A_scale
+A_zero_point
+
B
+B_scale
+B_zero_point
+C_scale
+C_zero_point
C"
QLinearMatMulqlinearmatmul_3D_test*
"D;BA_scale**
qBA_zero_point*
";BB_scale**
rBB_zero_point*
"O/<BC_scale**
vBC_zero_pointZ
+
A
+
+
+
+Z
+
B
+
+
+
+b
+
C
+
+
+
+B
\ No newline at end of file

test/onnx/verify_onnx.cpp

@@ -1270,7 +1270,7 @@ TEST_CASE(qlinearadd_test)
     pp["B"]     = migraphx::argument(b, data_b.data());
     auto result = p.eval(pp).back();
 
-    std::vector<unsigned char> result_vector;
+    std::vector<uint8_t> result_vector;
     result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
 
     std::vector<uint8_t> gold = {64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
@@ -1318,6 +1318,215 @@ TEST_CASE(qlinearadd_bcast_test)
     EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
 }
 
+TEST_CASE(qlinearconv_test)
+{
+    // https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__QLinearConv.html
+    migraphx::program p = migraphx::parse_onnx("qlinearconv_test.onnx");
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape sx{migraphx::shape::uint8_type, {1, 1, 7, 7}};
+    std::vector<uint8_t> x_data = {255, 174, 162, 25,  203, 168, 58,  15,  59,  237, 95,  129, 0,
+                                   64,  56,  242, 153, 221, 168, 12,  166, 232, 178, 186, 195, 237,
+                                   162, 237, 188, 39,  124, 77,  80,  102, 43,  127, 230, 21,  83,
+                                   41,  40,  134, 255, 154, 92,  141, 42,  148, 247};
+
+    migraphx::parameter_map pp;
+    pp["X"]     = migraphx::argument(sx, x_data.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<uint8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    std::vector<uint8_t> gold = {0,   81,  93,  230, 52,  87,  197, 240, 196, 18,  160, 126, 255,
+                                 191, 199, 13,  102, 34,  87,  243, 89,  23,  77,  69,  60,  18,
+                                 93,  18,  67,  216, 131, 178, 175, 153, 212, 128, 25,  234, 172,
+                                 214, 215, 121, 0,   101, 163, 114, 213, 107, 8};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearconv_pad_0_test)
+{
+    // https:xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+
+    migraphx::program p = migraphx::parse_onnx("qlinearconv_pad_0_test.onnx");
+
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape sx{migraphx::shape::uint8_type, {1, 1, 5, 5}};
+
+    std::vector<uint8_t> x_data = {0,   11,  21,  32,  42,  53,  64,  74,  85,  96,  106, 117, 128,
+                                   138, 149, 159, 170, 181, 191, 202, 212, 223, 234, 244, 255};
+
+    migraphx::parameter_map pp;
+    pp["X"]     = migraphx::argument(sx, x_data.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<int8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    // # (1, 1, 3, 3) output tensor
+    std::vector<int8_t> gold = {-43, -29, -15, 28, 42, 56, 99, 113, 127};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearconv_pad_1_test)
+{
+    // https:xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    migraphx::program p = migraphx::parse_onnx("qlinearconv_pad_1_test.onnx");
+
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape sx{migraphx::shape::uint8_type, {1, 1, 5, 5}};
+
+    std::vector<uint8_t> x_data = {0,   11,  21,  32,  42,  53,  64,  74,  85,  96,  106, 117, 128,
+                                   138, 149, 159, 170, 181, 191, 202, 212, 223, 234, 244, 255};
+
+    migraphx::parameter_map pp;
+    pp["X"]     = migraphx::argument(sx, x_data.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<uint8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    // # (1, 1, 5, 5) output tensor
+    std::vector<uint8_t> gold = {19,  33,  43,  52,  38,  52,  85,  99,  113, 80,  99,  156, 170,
+                                 184, 128, 146, 227, 241, 255, 175, 113, 175, 184, 194, 132};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearconv_scale_1D_test)
+{
+    // https:xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+
+    migraphx::program p = migraphx::parse_onnx("qlinearconv_scale_1D_test.onnx");
+
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape sx{migraphx::shape::uint8_type, {1, 1, 5, 5}};
+
+    std::vector<uint8_t> x_data = {0,   11,  21,  32,  42,  53,  64,  74,  85,  96,  106, 117, 128,
+                                   138, 149, 159, 170, 181, 191, 202, 212, 223, 234, 244, 255};
+
+    migraphx::parameter_map pp;
+    pp["X"]     = migraphx::argument(sx, x_data.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<int8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    // # (1, 2, 3, 3) output tensor
+    std::vector<int8_t> gold = {
+        -43, -29, -15, 28, 42, 56, 99, 113, 127, -43, -29, -15, 28, 42, 56, 99, 113, 127};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearglobalavgpool_test)
+{
+    // github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md
+    // #com.microsoft.QLinearGlobalAveragePool
+
+    migraphx::program p = migraphx::parse_onnx("qlinearglobalavgpool_test.onnx");
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape sh_x{migraphx::shape::uint8_type, {1, 3, 4, 4}};
+    std::vector<uint8_t> data_x = {160, 156, 152, 148, 144, 140, 136, 132, 124, 120, 116, 112,
+                                   108, 104, 100, 96,  64,  72,  80,  88,  96,  104, 112, 120,
+                                   136, 144, 152, 160, 168, 176, 184, 192, 120, 121, 122, 123,
+                                   124, 125, 126, 127, 129, 130, 131, 132, 133, 134, 135, 136};
+
+    migraphx::parameter_map pp;
+    pp["X"] = migraphx::argument(sh_x, data_x.data());
+
+    auto result = p.eval(pp).back();
+
+    std::vector<uint8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    std::vector<uint8_t> gold = {64, 64, 64};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearmatmul_1D_test)
+{
+    migraphx::program p = migraphx::parse_onnx("qlinearmatmul_1D_test.onnx");
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape a{migraphx::shape::uint8_type, {8}};
+    std::vector<uint8_t> data_a = {2, 4, 6, 8, 10, 12, 14, 16};
+
+    migraphx::shape b{migraphx::shape::uint8_type, {8}};
+    std::vector<uint8_t> data_b = {126, 130, 124, 132, 122, 134, 120, 136};
+
+    migraphx::parameter_map pp;
+    pp["A"]     = migraphx::argument(a, data_a.data());
+    pp["B"]     = migraphx::argument(b, data_b.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<uint8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    std::vector<uint8_t> gold = {66};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearmatmul_2D_test)
+{
+    migraphx::program p = migraphx::parse_onnx("qlinearmatmul_2D_test.onnx");
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape a{migraphx::shape::uint8_type, {1, 8}};
+    std::vector<uint8_t> data_a = {2, 4, 6, 8, 10, 12, 14, 16};
+
+    migraphx::shape b{migraphx::shape::uint8_type, {8, 1}};
+    std::vector<uint8_t> data_b = {126, 130, 124, 132, 122, 134, 120, 136};
+
+    migraphx::parameter_map pp;
+    pp["A"]     = migraphx::argument(a, data_a.data());
+    pp["B"]     = migraphx::argument(b, data_b.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<uint8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    std::vector<uint8_t> gold = {66};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
+TEST_CASE(qlinearmatmul_3D_test)
+{
+    // https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__QLinearMatMul.html
+
+    migraphx::program p = migraphx::parse_onnx("qlinearmatmul_3D_test.onnx");
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape a{migraphx::shape::uint8_type, {2, 2, 4}};
+    std::vector<uint8_t> data_a = {
+        208, 236, 0, 238, 3, 214, 255, 29, 208, 236, 0, 238, 3, 214, 255, 29};
+
+    migraphx::shape b{migraphx::shape::uint8_type, {2, 4, 3}};
+    std::vector<uint8_t> data_b = {152, 51, 244, 60, 26, 255, 0, 127, 246, 127, 254, 247,
+                                   152, 51, 244, 60, 26, 255, 0, 127, 246, 127, 254, 247};
+
+    migraphx::parameter_map pp;
+    pp["A"]     = migraphx::argument(a, data_a.data());
+    pp["B"]     = migraphx::argument(b, data_b.data());
+    auto result = p.eval(pp).back();
+
+    std::vector<uint8_t> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    std::vector<uint8_t> gold = {168, 115, 255, 1, 66, 151, 168, 115, 255, 1, 66, 151};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
 TEST_CASE(resize_downsample_f_test)
 {
     migraphx::program p = migraphx::parse_onnx("resize_downsample_f_test.onnx");

test/optimize_module_test.cpp

@@ -62,4 +62,41 @@ TEST_CASE(broadcast_transpose_inner_broadcast)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(broadcast_transpose_inner_broadcast_generic)
+{
+    // first optimizes broadcast+transpose to unsqueeze+transpose+broadcast,
+    // then finds inner broadcast to become mul+broadcast
+    migraphx::module m1;
+    {
+        auto l1 = m1.add_parameter("x", {migraphx::shape::float_type, {5, 10}});
+        auto l2 = m1.add_parameter("y", {migraphx::shape::float_type, {5}});
+        auto mb1 =
+            m1.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {3, 5, 10}}}), l1);
+        auto mb2 =
+            m1.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {3, 10, 5}}}), l2);
+        auto t1 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1}}}), mb2);
+        auto mul = m1.add_instruction(migraphx::make_op("mul"), mb1, t1);
+        m1.add_return({mul});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto l1        = m2.add_parameter("x", {migraphx::shape::float_type, {5, 10}});
+        auto l2        = m2.add_parameter("y", {migraphx::shape::float_type, {5}});
+        auto unsqueeze = m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {0, 1}}}), l2);
+        auto transpose = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1}}}), unsqueeze);
+        auto mb1 =
+            m2.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {1, 5, 10}}}), l1);
+        auto mb2 = m2.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", {1, 5, 10}}}), transpose);
+        auto mul = m2.add_instruction(migraphx::make_op("mul"), mb1, mb2);
+        auto mb3 = m2.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", {3, 5, 10}}}), mul);
+        m2.add_return({mb3});
+    }
+    EXPECT(m1 == m2);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/py/CMakeLists.txt

 #####################################################################################
 # 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
@@ -23,8 +23,14 @@
 #####################################################################################
 
 include(PythonModules)
+set(VENV ${CMAKE_BINARY_DIR}/test/py/venv)
+set(VENV_ONNX ${CMAKE_BINARY_DIR}/test/py/venv-onnx)
+set(REQUIREMENTS ${CMAKE_CURRENT_SOURCE_DIR}/requirements.txt)
+set(REQUIREMENTS_ONNX ${CMAKE_CURRENT_SOURCE_DIR}/requirements-onnx.txt)
+set(PYTHON_VERSION_TO_DISABLE_ONNX 3.6)
 
-function(add_py_test NAME SCRIPT)
+
+function(add_py_venv_fixture FIXTURE_NAME VIRTUAL_ENV_DIR REQUIREMENTS_FILE)
     foreach(PYTHON_VERSION ${PYTHON_VERSIONS})
         set (ENV_COMMAND ${CMAKE_COMMAND} -E env 
             "PYTHONPATH=$<TARGET_FILE_DIR:migraphx_pybind_${PYTHON_VERSION}>"
@@ -32,28 +38,57 @@ function(add_py_test NAME SCRIPT)
             "MALLOC_CHECK_=3"
         )
         set(PYTHON_EXECUTABLE ${PYTHON_${PYTHON_VERSION}_EXECUTABLE})
-        add_test(
-            NAME test_py_${PYTHON_VERSION}_${NAME} 
-            COMMAND ${ENV_COMMAND} ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/${SCRIPT} ${ARGN})
-        add_custom_target(test_py_${PYTHON_VERSION}_${NAME} 
-            COMMAND ${ENV_COMMAND} ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/${SCRIPT} ${ARGN} 
-            COMMENT "${PYTHON_EXECUTABLE} ${SCRIPT}")
 
+        if(NOT TEST py_${PYTHON_VERSION}_${FIXTURE_NAME}_initialize_env)
+            if (NOT (${FIXTURE_NAME} STREQUAL "onnx" AND ${PYTHON_VERSION} STREQUAL ${PYTHON_VERSION_TO_DISABLE_ONNX}))
+                add_test(NAME py_${PYTHON_VERSION}_${FIXTURE_NAME}_initialize_env COMMAND ${PYTHON_EXECUTABLE} -m venv ${VIRTUAL_ENV_DIR}/${PYTHON_VERSION} --clear)
+                set_tests_properties(py_${PYTHON_VERSION}_${FIXTURE_NAME}_initialize_env PROPERTIES FIXTURES_SETUP ${FIXTURE_NAME}_${PYTHON_VERSION}_INIT_VENV)
+                set(PYTHON_EXECUTABLE ${VIRTUAL_ENV_DIR}/${PYTHON_VERSION}/bin/python)
+                add_test(
+                    NAME py_${PYTHON_VERSION}_${FIXTURE_NAME}_setup_env
+                    COMMAND ${PYTHON_EXECUTABLE} -m pip install -r ${REQUIREMENTS_FILE})
+                set_tests_properties(py_${PYTHON_VERSION}_${FIXTURE_NAME}_setup_env PROPERTIES FIXTURES_REQUIRED ${FIXTURE_NAME}_${PYTHON_VERSION}_INIT_VENV)
+                set_tests_properties(py_${PYTHON_VERSION}_${FIXTURE_NAME}_setup_env PROPERTIES FIXTURES_SETUP ${FIXTURE_NAME}_${PYTHON_VERSION}_VENV)
+            endif()
+        endif()
+    endforeach()
+endfunction()
+
+function(add_py_test NAME SCRIPT FIXTURE_NAME VENV_DIR)
+    foreach(PYTHON_VERSION ${PYTHON_VERSIONS})
+        set (ENV_COMMAND ${CMAKE_COMMAND} -E env
+            "PYTHONPATH=$<TARGET_FILE_DIR:migraphx_pybind_${PYTHON_VERSION}>"
+            "PYTHONMALLOC=debug"
+            "MALLOC_CHECK_=3"
+        )
+        set(PYTHON_EXECUTABLE ${VENV_DIR}/${PYTHON_VERSION}/bin/python)
+        if(NOT (${FIXTURE_NAME} STREQUAL "onnx" AND ${PYTHON_VERSION} STREQUAL ${PYTHON_VERSION_TO_DISABLE_ONNX}))
+            add_test(
+                NAME test_py_${PYTHON_VERSION}_${NAME}
+                COMMAND ${ENV_COMMAND} ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/${SCRIPT} ${ARGN})
+            set_tests_properties(test_py_${PYTHON_VERSION}_${NAME} PROPERTIES FIXTURES_REQUIRED ${FIXTURE_NAME}_${PYTHON_VERSION}_VENV)
+            add_custom_target(test_py_${PYTHON_VERSION}_${NAME}
+                COMMAND ${ENV_COMMAND} ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/${SCRIPT} ${ARGN}
+                COMMENT "${PYTHON_EXECUTABLE} ${SCRIPT}")
+        endif()
     endforeach()
 endfunction()
 add_dependencies(tests migraphx_py)
 add_dependencies(check migraphx_py)
 
-add_py_test(ref test_cpu.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(save_load test_save_load.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(op test_op.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(shape test_shape.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(module_construct test_module_construct.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(literal test_literal.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_venv_fixture(common ${VENV} ${REQUIREMENTS})
+add_py_venv_fixture(onnx ${VENV_ONNX} ${REQUIREMENTS_ONNX})
+
+add_py_test(ref test_cpu.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(save_load test_save_load.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(op test_op.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(shape test_shape.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(module_construct test_module_construct.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(literal test_literal.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
 if(MIGRAPHX_ENABLE_GPU)
-add_py_test(gpu_offload test_gpu_offload.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(gpu test_gpu.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(array test_array.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(backend onnx_backend_test.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
-add_py_test(gpu_async test_gpu_async.py WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(gpu_offload test_gpu_offload.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(gpu test_gpu.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(array test_array.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(backend onnx_backend_test.py onnx ${VENV_ONNX} WORKING_DIRECTORY ${TEST_ONNX_DIR})
+add_py_test(gpu_async test_gpu_async.py common ${VENV} WORKING_DIRECTORY ${TEST_ONNX_DIR})
 endif()

test/py/requirements-onnx.txt

+#####################################################################################
+# The MIT License (MIT)
+#
+# 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
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in
+# all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+# THE SOFTWARE.
+#####################################################################################
+
+onnx==1.14.1
+protobuf==3.20.2
+numpy==1.21.6
+packaging==23.0
+pytest==6.0.1
\ No newline at end of file

test/py/requirements.txt

+#####################################################################################
+# The MIT License (MIT)
+#
+# 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
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in
+# all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+# THE SOFTWARE.
+#####################################################################################
+
+numpy==1.21.6
\ No newline at end of file

test/simplify_algebra_test.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
@@ -669,6 +669,23 @@ TEST_CASE(simplify_inner_broadcast_different_broadcasts)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(simplify_inner_broadcast_no_common_axis)
+{
+    auto b = migraphx::make_op("multibroadcast", {{"out_lens", {1, 5, 10}}});
+    migraphx::module m1;
+    {
+        auto x   = m1.add_parameter("x", {migraphx::shape::int32_type, {5, 10}});
+        auto y   = m1.add_parameter("y", {migraphx::shape::int32_type, {1, 5, 1}});
+        auto xb  = m1.add_instruction(b, x);
+        auto yb  = m1.add_instruction(b, y);
+        auto sum = m1.add_instruction(migraphx::make_op("add"), xb, yb);
+        m1.add_instruction(pass_op{}, sum);
+    }
+    migraphx::module m2 = m1;
+    run_pass(m1);
+    EXPECT(m1 == m2);
+}
+
 TEST_CASE(simplify_add_conv1)
 {
     migraphx::module m;

test/simplify_reshapes_test.cpp

@@ -67,6 +67,57 @@ migraphx::module make_concat_multibroadcast(const std::vector<size_t>& in_lens,
     return m;
 }
 
+TEST_CASE(broadcast_transpose)
+{
+    migraphx::module m1;
+    {
+        auto l = m1.add_parameter("x", {migraphx::shape::float_type, {5}});
+        auto mb =
+            m1.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {2, 3, 5}}}), l);
+        auto t1 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1}}}), mb);
+        m1.add_return({t1});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto l  = m2.add_parameter("x", {migraphx::shape::float_type, {5}});
+        auto u1 = m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {0, 1}}}), l);
+        auto t1 =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {2, 0, 1}}}), u1);
+        auto mb =
+            m2.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5, 2, 3}}}), t1);
+        m2.add_return({mb});
+    }
+
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(broadcast_transpose_opt)
+{
+    // extra transpose from transformation will be optimized out
+    migraphx::module m1;
+    {
+        auto l = m1.add_parameter("x", {migraphx::shape::float_type, {5}});
+        auto mb =
+            m1.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {2, 3, 5}}}), l);
+        auto t1 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 0, 2}}}), mb);
+        m1.add_return({t1});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto l  = m2.add_parameter("x", {migraphx::shape::float_type, {5}});
+        auto u1 = m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {0, 1}}}), l);
+        auto mb =
+            m2.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {3, 2, 5}}}), u1);
+        m2.add_return({mb});
+    }
+
+    EXPECT(m1 == m2);
+}
+
 TEST_CASE(broadcast_transpose_scalar)
 {
     migraphx::module m1;

test/verify/ck_gemm_softmax_gemm.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 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
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct ck_gemm_softmax_gemm : verify_program<ck_gemm_softmax_gemm>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        migraphx::shape m1_shape{migraphx::shape::half_type, {1, 12, 256, 256}};
+        migraphx::shape m2_shape{migraphx::shape::half_type, {1, 12, 256, 256}};
+        auto m2_elements = m2_shape.elements();
+        auto a           = mm->add_parameter("1", m1_shape);
+        auto b           = mm->add_parameter("2", m1_shape);
+        auto b1          = mm->add_parameter("3", m1_shape);
+        std::vector<float> eights(m2_elements, 0.125);
+        auto eight = mm->add_literal(migraphx::literal{m2_shape, eights});
+        std::vector<float> zeros(m2_elements, 0);
+        auto zero = mm->add_literal(migraphx::literal{m2_shape, zeros});
+
+        b = mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), b);
+        auto gemm1   = mm->add_instruction(migraphx::make_op("dot"), a, b);
+        auto scale   = mm->add_instruction(migraphx::make_op("mul"), gemm1, eight);
+        auto bias    = mm->add_instruction(migraphx::make_op("add"), scale, zero);
+        auto softmax = mm->add_instruction(migraphx::make_op("softmax", {{"axis", -1}}), bias);
+        mm->add_instruction(migraphx::make_op("dot"), softmax, b1);
+
+        return p;
+    }
+};

tools/install_prereqs.sh

@@ -3,7 +3,7 @@
 #####################################################################################
 # 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
@@ -51,6 +51,7 @@ else
     openmp-extras \
     python3-dev \
     python3-pip \
+    python3-venv \
     rocblas-dev \
     rocm-cmake
 fi