Merge branch 'develop' into simplify_1_mul_div_ops

src/targets/gpu/target.cpp

@@ -42,6 +42,7 @@
 #include <migraphx/register_target.hpp>
 #include <migraphx/replace_allocate.hpp>
 #include <migraphx/rewrite_batchnorm.hpp>
+#include <migraphx/rewrite_gelu.hpp>
 #include <migraphx/rewrite_pooling.hpp>
 #include <migraphx/rewrite_quantization.hpp>
 #include <migraphx/rewrite_rnn.hpp>
@@ -116,6 +117,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         inline_module{},
         rewrite_pooling{},
         dead_code_elimination{},
+        rewrite_gelu{},
+        dead_code_elimination{},
         eliminate_common_subexpression{},
         dead_code_elimination{},
         simplify_algebra{},
@@ -134,8 +137,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         lowering{&ctx, options.offload_copy},
         eliminate_contiguous{"gpu::contiguous"},
         dead_code_elimination{},
-        replace_allocate{gpu_allocation_model{}, options.offload_copy},
-        dead_code_elimination{},
         eliminate_concat{concat_gpu_optimization{}},
         dead_code_elimination{},
         pack_int8_args{},
@@ -144,6 +145,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         fuse_ops{&ctx, options.fast_math},
         dead_code_elimination{},
+        replace_allocate{gpu_allocation_model{}, options.offload_copy},
+        dead_code_elimination{},
         compile_ops{&ctx},
         dead_code_elimination{},
         write_literals{&ctx},

src/targets/ref/lowering.cpp

@@ -244,7 +244,6 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
             auto weights_lens = args[1].get_shape().lens();
             std::vector<std::size_t> k_lens{weights_lens.begin() + 2, weights_lens.end()};
             padding = calc_dyn_auto_pad(img_lens, k_lens, op.stride, op.dilation);
-            std::cout << "[ ";
             output_shape =
                 compute_padded_shape({args.at(0).get_shape(), args.at(1).get_shape()}, padding);
         }

src/tf/parse_conv.cpp

@@ -100,7 +100,7 @@ struct parse_conv : op_parser<parse_conv>
                 {
                     MIGRAPHX_THROW("padding should have 4 values");
                 }
-                if(padding[0] != padding[2] || padding[1] != padding[3])
+                if(padding[0] != padding[2] or padding[1] != padding[3])
                 {
                     MIGRAPHX_THROW("migraphx does not support asymetric padding");
                 }

src/tf/parse_depthwiseconv.cpp

@@ -90,7 +90,7 @@ struct parse_depthwiseconv : op_parser<parse_depthwiseconv>
                 calculate_padding(0, pads, input_dims[2], op.stride[0], op.dilation[0], weight_h);
                 calculate_padding(1, pads, input_dims[3], op.stride[1], op.dilation[1], weight_w);
 
-                if(pads[0] != pads[2] || pads[1] != pads[3])
+                if(pads[0] != pads[2] or pads[1] != pads[3])
                 {
                     std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
                     l0 = info.add_instruction(migraphx::make_op("pad", {{"pads", padding}}), l0);

src/tf/parse_pooling.cpp

@@ -42,7 +42,7 @@ struct parse_pooling : op_parser<parse_pooling>
                           tf_parser::node_info info,
                           std::vector<instruction_ref> args) const
     {
-        if(!starts_with(opd.tf_name, "Max") && !starts_with(opd.tf_name, "Av"))
+        if(not starts_with(opd.tf_name, "Max") and not starts_with(opd.tf_name, "Av"))
         {
             MIGRAPHX_THROW("tf pooling mode must be Max or Average");
         }

src/tf/parse_relu6.cpp

@@ -41,8 +41,9 @@ struct parse_relu6 : op_parser<parse_relu6>
                           const tf_parser::node_info& info,
                           std::vector<instruction_ref> args) const
     {
-        auto min_val = info.add_literal(0.0f);
-        auto max_val = info.add_literal(6.0f);
+        shape::type_t output_type = args[0]->get_shape().type();
+        auto min_val = info.add_literal(migraphx::literal{migraphx::shape{output_type}, {0.0f}});
+        auto max_val = info.add_literal(migraphx::literal{migraphx::shape{output_type}, {6.0f}});
 
         return info.add_common_op("clip", args[0], min_val, max_val);
     }

src/tf/tf_parser.cpp

@@ -371,7 +371,7 @@ void tf_parser::parse_node(const std::string& name)
         {
             result = ops[node.op()](*this, {get_attributes(node), node.op(), mm}, args);
         }
-        assert(!result.empty());
+        assert(not result.empty());
         // First output has no ":" delimiter
         instructions[name] = result.front();
         for(size_t i = 1; i < result.size(); i++)
@@ -458,7 +458,7 @@ literal tf_parser::parse_tensor(const tensorflow::TensorProto& t) const
 {
     std::vector<size_t> dims = parse_dims(t.tensor_shape());
     size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
-    if(!t.tensor_content().empty()) // has raw data
+    if(not t.tensor_content().empty()) // has raw data
     {
         const std::string& s = t.tensor_content();
         switch(t.dtype())

src/tmp_dir.cpp

@@ -78,7 +78,7 @@ void tmp_dir::execute(const std::string& exe, const std::string& args) const
 
 tmp_dir::~tmp_dir()
 {
-    if(!enabled(MIGRAPHX_DEBUG_SAVE_TEMP_DIR{}))
+    if(not enabled(MIGRAPHX_DEBUG_SAVE_TEMP_DIR{}))
     {
         fs::remove_all(this->path);
     }

src/value.cpp

@@ -400,7 +400,7 @@ std::pair<value*, bool> value::insert(const value& v)
 {
     if(v.key.empty())
     {
-        if(!x)
+        if(not x)
             x = std::make_shared<array_value_holder>();
         get_array_impl(x).push_back(v);
         assert(this->if_array());
@@ -408,7 +408,7 @@ std::pair<value*, bool> value::insert(const value& v)
     }
     else
     {
-        if(!x)
+        if(not x)
             x = std::make_shared<object_value_holder>();
         auto p = x->if_object()->emplace(v.key, get_array_impl(x).size());
         if(p.second)
@@ -420,7 +420,7 @@ std::pair<value*, bool> value::insert(const value& v)
 value* value::insert(const value* pos, const value& v)
 {
     assert(v.key.empty());
-    if(!x)
+    if(not x)
         x = std::make_shared<array_value_holder>();
     auto&& a = get_array_impl(x);
     auto it  = a.insert(a.begin() + (pos - begin()), v);
@@ -466,7 +466,7 @@ bool compare(const value& x, const value& y, F f)
 
 value::type_t value::get_type() const
 {
-    if(!x)
+    if(not x)
         return null_type;
     return x->get_type();
 }

test/api/test_custom_op_gpu.cpp

@@ -55,7 +55,7 @@ struct simple_custom_op final : migraphx::experimental_custom_op_base
 
     virtual migraphx::shape compute_shape(migraphx::shapes inputs) const override
     {
-        if(!inputs[0].standard())
+        if(not inputs[0].standard())
         {
             throw std::runtime_error("first arg must be standard shaped");
         }

test/check_shapes_test.cpp

@@ -49,6 +49,6 @@ bool create_shapes(bool dynamic_allowed)
 
 TEST_CASE(allow_dynamic_shape) { EXPECT(create_shapes(true)); }
 
-TEST_CASE(fail_dynamic_shape) { EXPECT(!create_shapes(false)); }
+TEST_CASE(fail_dynamic_shape) { EXPECT(not create_shapes(false)); }
 
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/eval_test.cpp

@@ -187,7 +187,7 @@ TEST_CASE(print_test)
     std::stringstream ss;
     ss << p;
     std::string s = ss.str();
-    EXPECT(!s.empty());
+    EXPECT(not s.empty());
 }
 
 TEST_CASE(param_test)

test/get_target_assignments.cpp → test/fpga/get_target_assignments.cpp

@@ -26,8 +26,9 @@
 #include <migraphx/make_op.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/register_target.hpp>
-#include <migraphx/ref/target.hpp>
+#include <migraphx/fpga/target.hpp>
 #include <migraphx/target_assignments.hpp>
+#include <migraphx/iterator_for.hpp>
 
 migraphx::program create_program()
 {
@@ -37,8 +38,8 @@ migraphx::program create_program()
     auto x    = mm->add_parameter("x", s);
     auto y    = mm->add_parameter("y", s);
     auto z    = mm->add_parameter("z", s);
-    auto diff = mm->add_instruction(migraphx::make_op("div"), x, y);
-    mm->add_instruction(migraphx::make_op("div"), diff, z);
+    auto diff = mm->add_instruction(migraphx::make_op("add"), x, y);
+    mm->add_instruction(migraphx::make_op("add"), diff, z);
     return p;
 }
 
@@ -46,15 +47,17 @@ TEST_CASE(is_supported)
 {
     auto p       = create_program();
     auto targets = migraphx::get_targets();
-    EXPECT(!targets.empty());
-    auto first_target = targets[0];
-    auto t            = migraphx::make_target(first_target);
+    EXPECT(not targets.empty());
+    auto t = migraphx::make_target("fpga");
 
     const auto assignments = p.get_target_assignments({t});
-    for(const auto& [ins, target] : assignments)
+    const auto* mod        = p.get_main_module();
+    EXPECT(mod->size() == assignments.size());
+
+    for(const auto ins : iterator_for(*mod))
     {
-        (void)ins;
-        EXPECT(target == first_target);
+        const auto& target = assignments.at(ins);
+        EXPECT(target == "fpga");
     }
 }
 

test/include/basic_ops.hpp

@@ -112,12 +112,12 @@ struct mod_pass_op
     migraphx::shape compute_shape(std::vector<migraphx::shape> inputs,
                                   std::vector<migraphx::module_ref> mods) const
     {
-        if(!mods.empty())
+        if(not mods.empty())
         {
             auto out_shapes = mods[0]->get_output_shapes();
             return out_shapes[0];
         }
-        if(!inputs.empty())
+        if(not inputs.empty())
         {
             return inputs.front();
         }
@@ -186,9 +186,10 @@ struct nop
     migraphx::shape compute_shape(const std::vector<migraphx::shape>&) const { return {}; }
 };
 
-inline migraphx::literal get_2x2()
+inline migraphx::literal get_2x2(int base = 0)
 {
-    return migraphx::literal{{migraphx::shape::float_type, {2, 2}}, {1, 2, 3, 4}};
+    return migraphx::literal{{migraphx::shape::float_type, {2, 2}},
+                             {base + 1, base + 2, base + 3, base + 4}};
 }
 
 inline migraphx::literal get_2x2_transposed()

test/include/test.hpp

@@ -108,15 +108,7 @@ struct function
 };
 
 template <class Stream, class Iterator>
-inline Stream& stream_range(Stream& s, Iterator start, Iterator last)
-{
-    if(start != last)
-    {
-        s << *start;
-        std::for_each(std::next(start), last, [&](auto&& x) { s << ", " << x; });
-    }
-    return s;
-}
+Stream& stream_range(Stream& s, Iterator start, Iterator last);
 
 template <class Stream>
 inline Stream& operator<<(Stream& s, std::nullptr_t)
@@ -136,6 +128,17 @@ inline auto operator<<(Stream& s, const Range& v) -> decltype(stream_range(s, v.
     return s;
 }
 
+template <class Stream, class Iterator>
+inline Stream& stream_range(Stream& s, Iterator start, Iterator last)
+{
+    if(start != last)
+    {
+        s << *start;
+        std::for_each(std::next(start), last, [&](auto&& x) { s << ", " << x; });
+    }
+    return s;
+}
+
 template <class T>
 const T& get_value(const T& x)
 {
@@ -342,7 +345,7 @@ inline std::ostream& operator<<(std::ostream& os, const color& c)
 template <class T, class F>
 void failed(T x, const char* msg, const char* func, const char* file, int line, F f)
 {
-    if(!bool(x.value()))
+    if(not bool(x.value()))
     {
         std::cout << func << std::endl;
         std::cout << file << ":" << line << ":" << std::endl;

test/literal_test.cpp

@@ -39,8 +39,8 @@ TEST_CASE(literal_test)
     migraphx::literal l2 = l1; // NOLINT
     EXPECT(l1 == l2);
     EXPECT(l1.at<int>(0) == 1);
-    EXPECT(!l1.empty());
-    EXPECT(!l2.empty());
+    EXPECT(not l1.empty());
+    EXPECT(not l2.empty());
 
     migraphx::literal l3{};
     migraphx::literal l4{};

test/onnx/gen_onnx.py

@@ -3589,7 +3589,7 @@ def nms_test():
     st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
                                        [1])
     out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
-                                        [6, 3])
+                                        [None, 3])
 
     node = onnx.helper.make_node('NonMaxSuppression',
                                  inputs=[
@@ -3603,6 +3603,108 @@ def nms_test():
     return ([node], [b, s, mo, iou, st], [out])
 
 
+@onnx_test
+def nms_use_dyn_output_false_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, 6, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT, [1, 1, 6])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'],
+                                 use_dyn_output=0)
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
+@onnx_test
+def nms_dynamic_batch_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [None, 6, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT,
+                                      [None, 1, 6])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'],
+                                 center_point_box=1,
+                                 use_dyn_output=1)
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
+@onnx_test
+def nms_dynamic_boxes_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, None, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT,
+                                      [1, 1, None])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'])
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
+@onnx_test
+def nms_dynamic_classes_test():
+    b = helper.make_tensor_value_info('boxes', TensorProto.FLOAT, [1, 6, 4])
+    s = helper.make_tensor_value_info('scores', TensorProto.FLOAT,
+                                      [1, None, 6])
+    mo = helper.make_tensor_value_info('max_output_boxes_per_class',
+                                       TensorProto.INT64, [1])
+    iou = helper.make_tensor_value_info('iou_threshold', TensorProto.FLOAT,
+                                        [1])
+    st = helper.make_tensor_value_info('score_threshold', TensorProto.FLOAT,
+                                       [1])
+    out = helper.make_tensor_value_info('selected_indices', TensorProto.INT64,
+                                        [None, 3])
+
+    node = onnx.helper.make_node('NonMaxSuppression',
+                                 inputs=[
+                                     'boxes', 'scores',
+                                     'max_output_boxes_per_class',
+                                     'iou_threshold', 'score_threshold'
+                                 ],
+                                 outputs=['selected_indices'])
+
+    return ([node], [b, s, mo, iou, st], [out])
+
+
 @onnx_test
 def not_test():
     x = helper.make_tensor_value_info('0', TensorProto.INT32, [4])