Merge pull request #334 from ROCmSoftwarePlatform/bert_ops

Bert ops

src/onnx/onnx.cpp

@@ -504,15 +504,32 @@ struct onnx_parser
     parse_slice(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
     {
         op::slice op;
+        std::vector<size_t> dims = args[0]->get_shape().lens();
+        size_t num_dims          = dims.size();
         if(contains(attributes, "axes"))
         {
             literal s = parse_value(attributes.at("axes"));
             s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
         }
+        else
+        {
+            op.axes = std::vector<int64_t>(num_dims);
+            std::iota(op.axes.begin(), op.axes.end(), 0);
+        }
+
+        if(contains(attributes, "ends"))
         {
             literal s = parse_value(attributes.at("ends"));
             s.visit([&](auto v) { copy(v, std::back_inserter(op.ends)); });
+            for(size_t i = 0; i < num_dims; i++)
+            {
+                if(static_cast<size_t>(op.ends[i]) > dims[i])
+                {
+                    op.ends[i] = dims[i];
+                }
+            }
         }
+        if(contains(attributes, "starts"))
         {
             literal s = parse_value(attributes.at("starts"));
             s.visit([&](auto v) { copy(v, std::back_inserter(op.starts)); });

src/tf/tf.cpp

@@ -26,7 +26,6 @@ struct tf_parser
 {
     using attribute_map = std::unordered_map<std::string, tensorflow::AttrValue>;
     using node_map      = std::map<std::string, tensorflow::NodeDef>;
-    // using input_node_map = std::unordered_map<std::string, std::unordered_set<std::string>>;
     using op_func = std::function<instruction_ref(attribute_map, std::vector<instruction_ref>)>;
 
     node_map nodes;
@@ -149,9 +148,26 @@ struct tf_parser
         return axes;
     }
 
+    std::vector<int64_t> get_axes_from_mask(const size_t num_axes, const uint32_t mask)
+    {
+        uint32_t bitwise_compare = 1;
+        std::vector<int64_t> axes;
+        for(size_t i = 0; i < num_axes; i++)
+        {
+            // the LSB corresponds to axis 0 when determining which axes to begin
+            if(((mask >> i) & bitwise_compare) == 1)
+                axes.push_back(1);
+            else
+                axes.push_back(0);
+        }
+        return axes;
+    }
+
     tf_parser()
     {
+        add_generic_op("All", op::identity{});
         add_generic_op("Identity", op::identity{});
+        add_generic_op("LessEqual", op::identity{});
         add_generic_op("Relu", op::relu{});
         add_generic_op("Relu6", op::clip{6.0, 0.0});
         add_generic_op("Rsqrt", op::rsqrt{});
@@ -166,6 +182,7 @@ struct tf_parser
 
         add_mem_op("AvgPool", &tf_parser::parse_pooling);
         add_mem_op("BatchMatMul", &tf_parser::parse_matmul, false);
+        add_mem_op("BatchMatMulV2", &tf_parser::parse_matmul, false);
         add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
         add_mem_op("Cast", &tf_parser::parse_cast, false);
         add_mem_op("ConcatV2", &tf_parser::parse_concat, false);
@@ -177,14 +194,15 @@ struct tf_parser
         add_mem_op("GatherV2", &tf_parser::parse_gather, false);
         add_mem_op("MatMul", &tf_parser::parse_matmul, false);
         add_mem_op("MaxPool", &tf_parser::parse_pooling);
-        add_mem_op("Mean", &tf_parser::parse_mean);
+        add_mem_op("Mean", &tf_parser::parse_mean, false);
+        add_mem_op("OneHot", &tf_parser::parse_onehot, false);
         add_mem_op("Pack", &tf_parser::parse_pack, false);
         add_mem_op("Pad", &tf_parser::parse_pad);
         add_mem_op("Reshape", &tf_parser::parse_reshape, false);
         add_mem_op("Slice", &tf_parser::parse_slice, false);
-        add_mem_op("Softmax", &tf_parser::parse_softmax<op::softmax>);
+        add_mem_op("Softmax", &tf_parser::parse_softmax<op::softmax>, false);
         add_mem_op("Squeeze", &tf_parser::parse_squeeze, false);
-        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice);
+        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice, false);
         add_mem_op("Transpose", &tf_parser::parse_transpose, false);
     }
 
@@ -547,7 +565,7 @@ struct tf_parser
         }
         if(contains(attributes, "transpose_b"))
         {
-            transb = attributes.at("transpose_a").b();
+            transb = attributes.at("transpose_b").b();
         }
 
         if(contains(attributes, "adj_x"))
@@ -574,8 +592,7 @@ struct tf_parser
     parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
     {
         bool keep_dims = attributes.at("keep_dims").b();
-        auto lens      = args[0]->get_shape().lens();
-        auto axes = parse_axes(args[1]->eval().get<int32_t>().to_vector<int64_t>(), lens.size());
+        auto axes      = args[1]->eval().get<int32_t>().to_vector<int64_t>();
 
         if(keep_dims)
         {
@@ -588,6 +605,32 @@ struct tf_parser
         }
     }
 
+    instruction_ref
+    parse_onehot(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
+    {
+        size_t depth = static_cast<size_t>(args[1]->eval().at<int32_t>());
+
+        int64_t axis    = -1;
+        float on_value  = args[2]->eval().at<float>();
+        float off_value = args[3]->eval().at<float>();
+
+        std::vector<float> depth_input(depth * depth, off_value);
+        for(int i = 0; i < depth; i++)
+        {
+            depth_input[depth * i + i] = on_value;
+        }
+
+        if(contains(attributes, "axis"))
+            axis = attributes.at("axis").i();
+        if(axis == -1)
+        {
+            shape s{shape::float_type, {depth, depth}};
+            auto l0 = prog.add_literal({s, depth_input});
+            return prog.add_instruction(op::gather{0}, {l0, args[0]});
+        }
+        MIGRAPHX_THROW("MIGraphX does not support axis != -1");
+    }
+
     instruction_ref parse_pack(const std::string&,
                                const attribute_map& attributes,
                                std::vector<instruction_ref> args)
@@ -799,21 +842,50 @@ struct tf_parser
                                        std::vector<instruction_ref> args)
     {
         op::slice op;
-        auto starts     = args[1]->eval().get<int32_t>().to_vector();
-        auto ends       = args[2]->eval().get<int32_t>().to_vector();
-        size_t num_axes = args[0]->get_shape().lens().size();
+        auto starts              = args[1]->eval().get<int32_t>().to_vector();
+        auto ends                = args[2]->eval().get<int32_t>().to_vector();
+        auto l0                  = args[0];
+        size_t num_axes          = l0->get_shape().lens().size();
+        std::vector<size_t> axes = l0->get_shape().lens();
 
         op.starts = std::vector<int64_t>(starts.begin(), starts.end());
         op.ends   = std::vector<int64_t>(ends.begin(), ends.end());
         op.axes   = std::vector<int64_t>(num_axes);
         std::iota(op.axes.begin(), op.axes.end(), 0);
+        uint32_t begin_mask       = 0;
+        uint32_t end_mask         = 0;
         uint32_t shrink_axis_mask = 0;
         uint32_t bitwise_compare  = 1;
         std::vector<int64_t> squeeze_axes;
 
+        if(contains(attributes, "begin_mask"))
+            begin_mask = static_cast<uint32_t>(attributes.at("begin_mask").i());
+
+        if(contains(attributes, "end_mask"))
+            end_mask = static_cast<uint32_t>(attributes.at("end_mask").i());
+
         if(contains(attributes, "shrink_axis_mask"))
             shrink_axis_mask = static_cast<uint32_t>(attributes.at("shrink_axis_mask").i());
 
+        std::vector<int64_t> begin_axes = get_axes_from_mask(num_axes, begin_mask);
+        std::vector<int64_t> end_axes   = get_axes_from_mask(num_axes, end_mask);
+
+        for(size_t i = 0; i < num_axes; i++)
+        {
+            if(begin_axes.at(i) == 1)
+            {
+                op.starts.at(i) = 0;
+            }
+            if(end_axes.at(i) == 1)
+            {
+                op.ends.at(i) = axes.at(i);
+            }
+        }
+
+        auto l1 = prog.add_instruction(op, l0);
+        if(shrink_axis_mask == 0)
+            return l1;
+
         for(size_t i = 0; i < num_axes; i++)
         {
             // the LSB corresponds to axis 0 when determining which axes to squeeze
@@ -821,8 +893,7 @@ struct tf_parser
                 squeeze_axes.push_back(i);
         }
 
-        auto l0 = prog.add_instruction(op, make_contiguous(args[0]));
-        return to_nhwc(prog.add_instruction(op::squeeze{squeeze_axes}, l0));
+        return prog.add_instruction(op::squeeze{squeeze_axes}, l1);
     }
 
     instruction_ref
@@ -862,10 +933,16 @@ struct tf_parser
         if(instructions.count(name) == 0)
         {
             auto&& node = nodes.at(name);
+            // assert ops ignored
+            if(node.op() == "Assert" or contains(name, "Assert"))
+                return;
             std::vector<instruction_ref> args;
 
             for(auto&& input : node.input())
             {
+                // control dependencies (signified by ^ before the name) are ignored
+                if(contains(input, "^"))
+                    continue;
                 if(nodes.count(input) > 0)
                 {
                     auto&& iname = get_name(nodes.at(input));

test/tf/tf_test.cpp

@@ -48,6 +48,22 @@ TEST_CASE(add_bcast_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(assert_less_equal_test)
+{
+    migraphx::program p;
+    migraphx::shape s0{migraphx::shape::float_type, {2, 3}};
+    auto l0 = p.add_parameter("0", s0);
+    auto l1 = p.add_parameter("1", s0);
+    migraphx::literal l{migraphx::shape{migraphx::shape::int32_type, {2}}, {0, 1}};
+    auto l2 = p.add_literal(l);
+    p.add_instruction(migraphx::op::add{}, l0, l1);
+    auto l3 = p.add_instruction(migraphx::op::identity{}, l0, l1);
+    p.add_instruction(migraphx::op::identity{}, l3, l2);
+    auto prog = optimize_tf("assert_less_equal_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(batchmatmul_test)
 {
     migraphx::program p;
@@ -100,6 +116,16 @@ TEST_CASE(biasadd_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(cast_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_instruction(migraphx::op::convert{migraphx::shape::int32_type}, l0);
+    auto prog = optimize_tf("cast_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(concat_test)
 {
     migraphx::program p;
@@ -118,16 +144,6 @@ TEST_CASE(concat_test)
     EXPECT(p == prog);
 }
 
-TEST_CASE(cast_test)
-{
-    migraphx::program p;
-    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
-    p.add_instruction(migraphx::op::convert{migraphx::shape::int32_type}, l0);
-    auto prog = optimize_tf("cast_test.pb", false);
-
-    EXPECT(p == prog);
-}
-
 TEST_CASE(const_test)
 {
     migraphx::program p;
@@ -271,9 +287,10 @@ TEST_CASE(mean_test_nhwc)
     migraphx::program p;
     migraphx::literal l{migraphx::shape{migraphx::shape::int32_type, {2}}, {1, 2}};
     auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
-    migraphx::op::reduce_mean op{{2, 3}};
-    auto l3 = p.add_instruction(op, l0);
-    p.add_instruction(migraphx::op::squeeze{{2, 3}}, l3);
+    auto l1 = p.add_instruction(migraphx::op::transpose{{0, 2, 3, 1}}, l0);
+    migraphx::op::reduce_mean op{{1, 2}};
+    auto l2 = p.add_instruction(op, l1);
+    p.add_instruction(migraphx::op::squeeze{{1, 2}}, l2);
     auto prog = optimize_tf("mean_test_nhwc.pb", true);
 
     EXPECT(p == prog);
@@ -291,6 +308,23 @@ TEST_CASE(mul_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(onehot_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_literal(
+        migraphx::literal{migraphx::shape{migraphx::shape::int32_type, {5}}, {1, 1, 1, 1, 1}});
+    p.add_literal(2);
+    p.add_literal(1.0f);
+    p.add_literal(0.0f);
+    auto l1 = p.add_literal(
+        migraphx::literal{migraphx::shape{migraphx::shape::float_type, {2, 2}}, {1, 0, 0, 1}});
+    int axis = 0;
+    p.add_instruction(migraphx::op::gather{axis}, l1, l0);
+    auto prog = optimize_tf("onehot_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(pack_test)
 {
     migraphx::program p;
@@ -475,20 +509,44 @@ TEST_CASE(stridedslice_test)
 {
     migraphx::program p;
     auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 10, 1, 1}});
+    auto l1 = p.add_instruction(migraphx::op::transpose{{0, 2, 3, 1}}, l0);
     std::size_t num_axes = 4;
     migraphx::op::slice op;
     op.starts = {0, 0, 0, 0};
     op.ends   = {1, 1, 1, 5};
     op.axes   = std::vector<int64_t>(num_axes);
     std::iota(op.axes.begin(), op.axes.end(), 0);
-    auto l1          = p.add_instruction(op, l0);
+    auto l2          = p.add_instruction(op, l1);
     auto shrink_axis = 1;
-    p.add_instruction(migraphx::op::squeeze{{shrink_axis}}, l1);
+    p.add_instruction(migraphx::op::squeeze{{shrink_axis}}, l2);
     auto prog = optimize_tf("stridedslice_test.pb", true);
 
     EXPECT(p == prog);
 }
 
+TEST_CASE(stridedslice_masks_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 10, 3, 3}});
+    std::size_t num_axes = 4;
+    migraphx::op::slice op;
+    op.starts = {0, 1, 1, 0};
+    op.ends   = {1, 3, 3, 10};
+    op.axes   = std::vector<int64_t>(num_axes);
+    std::iota(op.axes.begin(), op.axes.end(), 0);
+    // add literals for starts, ends, and strides in tf (NHWC format)
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type, {4}}, std::vector<int>{0, 1, 1, 0});
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type, {4}}, std::vector<int>{0, 0, 0, 0});
+    p.add_literal(migraphx::shape{migraphx::shape::int32_type, {4}}, std::vector<int>{1, 1, 1, 1});
+
+    auto l1 = p.add_instruction(migraphx::op::transpose{{0, 2, 3, 1}}, l0);
+    auto l2 = p.add_instruction(op, l1);
+    p.add_instruction(migraphx::op::transpose{{0, 3, 1, 2}}, l2);
+    auto prog = migraphx::parse_tf("stridedslice_masks_test.pb", true);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(sub_test)
 {
     migraphx::program p;