Handle broadcasts across dot and concat (#1689) (#1731)

Improves the constant propagation for bert models. Larger batch size no longer use as large of constants.  Also improves the speed of model compilation
Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>

src/normalize_attributes.cpp

@@ -40,10 +40,12 @@ inline namespace MIGRAPHX_INLINE_NS {
  *
  * See normalize_attribute.hpp for explaining the options.
  */
+template <class Message>
 auto tune_attribute(const std::vector<int64_t>& vec,
                     const std::vector<int64_t>& axes,
                     const value& val,
-                    const std::vector<std::size_t>& lens)
+                    const std::vector<std::size_t>& lens,
+                    Message m)
 {
     std::vector<int64_t> result(vec);
     int64_t n_rank                                 = lens.size();
@@ -84,14 +86,14 @@ auto tune_attribute(const std::vector<int64_t>& vec,
         {
             if(not std::equal(result.begin(), result.end(), max_vals.begin(), std::less_equal<>{}))
             {
-                MIGRAPHX_THROW("TUNE_VECTOR: value out of range!");
+                MIGRAPHX_THROW(m() + "value out of range!");
             }
         }
         else
         {
             if(not std::equal(result.begin(), result.end(), max_vals.begin(), std::less<>{}))
             {
-                MIGRAPHX_THROW("TUNE_VECTOR: value out of range!");
+                MIGRAPHX_THROW(m() + "value out of range!");
             }
         }
     }
@@ -124,14 +126,14 @@ auto tune_attribute(const std::vector<int64_t>& vec,
             if(not std::equal(
                    min_vals.begin(), min_vals.end(), result.begin(), std::less_equal<>{}))
             {
-                MIGRAPHX_THROW("TUNE_VECTOR: attribute out of range!");
+                MIGRAPHX_THROW(m() + "attribute out of range!");
             }
         }
         else
         {
             if(not std::equal(result.begin(), result.end(), min_vals.begin(), std::less<>{}))
             {
-                MIGRAPHX_THROW("TUNE_VECTOR: attribute out of range!");
+                MIGRAPHX_THROW(m() + "attribute out of range!");
             }
         }
     }
@@ -193,6 +195,7 @@ bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens)
         const auto& key = rv.get_key();
         if(val.contains(key))
         {
+            auto message = [&] { return op.name() + ": " + key + ": "; };
             auto vv      = val.at(key).without_key();
             if(vv.is_array())
             {
@@ -202,7 +205,7 @@ bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens)
                     axes = val.at("axes").without_key().to_vector<int64_t>();
                 }
                 auto vec    = vv.to_vector<int64_t>();
-                auto result = tune_attribute(vec, axes, rv.without_key(), lens);
+                auto result = tune_attribute(vec, axes, rv.without_key(), lens, message);
                 val[key]    = result;
                 op.from_value(val);
                 val   = op.to_value();
@@ -211,7 +214,7 @@ bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens)
             else
             {
                 auto num    = vv.to<int64_t>();
-                auto result = tune_attribute({num}, {num}, rv.without_key(), lens);
+                auto result = tune_attribute({num}, {num}, rv.without_key(), lens, message);
                 val[key]    = result.front();
                 op.from_value(val);
                 val   = op.to_value();

src/simplify_algebra.cpp

@@ -361,30 +361,118 @@ struct find_inner_broadcast
 {
     auto matcher() const { return pointwise(match::all_of[match::inputs()](match::broadcast())); }
 
+    static auto non_scalar_op(const std::string& name)
+    {
+        return [=](instruction_ref ins) {
+            if(ins->get_shape().scalar())
+                return false;
+            return ins->name() == name;
+        };
+    }
+
     void apply(module& m, const match::matcher_result& r) const
     {
         auto ins        = r.result;
         auto broadcasts = ins->inputs();
         if(broadcasts.empty())
             return;
+        bool mixed_broadcasts = any_of(broadcasts, non_scalar_op("broadcast")) and
+                                any_of(broadcasts, non_scalar_op("multibroadcast"));
+        // If the broadcast is not a single dimension, then dont perform inner_broadcast
+        if(mixed_broadcasts and any_of(broadcasts, [&](instruction_ref i) {
+               if(i->get_shape().scalar())
+                   return false;
+               if(i->name() == "multibroadcast")
+                   return false;
+               auto input       = i->inputs().at(0);
+               const auto& lens = input->get_shape().lens();
+               return std::count_if(lens.begin(), lens.end(), [&](std::size_t d) {
+                          return d == 1;
+                      }) < (lens.size() - 1);
+           }))
+            return;
         std::vector<instruction_ref> inputs;
         std::transform(broadcasts.begin(),
                        broadcasts.end(),
                        std::back_inserter(inputs),
-                       [](auto i) { return i->inputs().front(); });
-        if(std::any_of(inputs.begin(), inputs.end(), [&](auto i) {
-               return i->get_shape() != inputs.front()->get_shape() and
-                      i->get_shape().elements() != 1;
-           }))
-            return;
-
-        auto b_it = std::find_if(broadcasts.begin(), broadcasts.end(), [&](auto i) {
-            return not i->get_shape().scalar();
+                       [&](instruction_ref i) {
+                           auto input = i->inputs().front();
+                           if(mixed_broadcasts and not i->get_shape().scalar() and
+                              i->get_shape().lens().size() > 1)
+                               return m.insert_instruction(i, make_op("squeeze"), input);
+                           return input;
                        });
-        if(b_it == broadcasts.end())
-            b_it = broadcasts.begin();
+
+        std::sort(broadcasts.begin(), broadcasts.end(), by(std::less<>{}, [](instruction_ref i) {
+                      if(i->get_shape().scalar())
+                          return 2;
+                      else if(i->name() == "broadcast")
+                          return 0;
+                      if(i->name() == "multibroadcast")
+                          return 1;
+                      return 3;
+                  }));
         auto op = insert_common_op(m, ins, ins->get_operator(), inputs);
-        m.replace_instruction(ins, (*b_it)->get_operator(), op);
+        m.replace_instruction(ins, broadcasts.front()->get_operator(), op);
+    }
+};
+
+struct find_dot_broadcast
+{
+    auto matcher() const
+    {
+        return match::name("dot")(match::all_of[match::inputs()](match::broadcast()));
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins = r.result;
+        auto a   = ins->inputs()[0];
+        auto b   = ins->inputs()[1];
+        if(a->get_operator().name() != b->get_operator().name())
+            return;
+        if(ins->get_shape().lens().size() < 3)
+            return;
+        auto nbatch_axes      = ins->get_shape().lens().size() - 2;
+        const auto& a_strides = a->get_shape().strides();
+        const auto& b_strides = b->get_shape().strides();
+        // Find leading batch axes that are broadcasted
+        auto p =
+            std::mismatch(a_strides.begin(),
+                          a_strides.begin() + nbatch_axes,
+                          b_strides.begin(),
+                          b_strides.begin() + nbatch_axes,
+                          [](auto astride, auto bstride) { return astride == 0 and bstride == 0; });
+        auto naxes = p.first - a_strides.begin();
+        assert(naxes <= nbatch_axes);
+        std::vector<std::size_t> axes(naxes);
+        std::iota(axes.begin(), axes.end(), 0);
+
+        auto insert_broadcast = [&](instruction_ref b_ins) -> instruction_ref {
+            auto input = b_ins->inputs()[0];
+            std::vector<std::size_t> lens(b_ins->get_shape().lens().begin() + naxes,
+                                          b_ins->get_shape().lens().end());
+            if(b_ins->name() == "multibroadcast")
+            {
+                return m.insert_instruction(
+                    ins, make_op("multibroadcast", {{"out_lens", lens}}), input);
+            }
+            else if(b_ins->name() == "broadcast")
+            {
+                auto v    = b_ins->get_operator().to_value();
+                auto axis = v.at("axis").to<std::size_t>() - naxes;
+                return m.insert_instruction(
+                    ins, make_op("broadcast", {{"axis", axis}, {"out_lens", lens}}), input);
+            }
+            assert(false);
+            return m.end();
+        };
+        auto a1        = insert_broadcast(a);
+        auto b1        = insert_broadcast(b);
+        auto dot       = m.insert_instruction(ins, make_op("dot"), a1, b1);
+        auto broadcast = m.insert_instruction(
+            ins, make_op("multibroadcast", {{"out_lens", ins->get_shape().lens()}}), dot);
+        m.replace_instruction(ins, broadcast);
     }
 };
 
@@ -393,7 +481,8 @@ struct find_concat_op
     auto matcher() const
     {
         return match::name("concat")(match::any_of[match::inputs()](
-            match::any_of(match::pointwise(), match::name("broadcast")), match::used_once()));
+            match::any_of(match::pointwise(), match::name("broadcast", "multibroadcast")),
+            match::used_once()));
     }
 
     template <class Iterator>
@@ -412,7 +501,8 @@ struct find_concat_op
 
     static bool is_valid_op(const operation& op)
     {
-        return op.name() == "broadcast" or op.attributes().contains("pointwise");
+        return contains({"broadcast", "multibroadcast"}, op.name()) or
+               op.attributes().contains("pointwise");
     }
 
     void apply(module& m, const match::matcher_result& r) const
@@ -440,6 +530,16 @@ struct find_concat_op
                 op               = b;
                 iaxis            = 0;
             }
+            else if(op.name() == "multibroadcast")
+            {
+                shape bshape = (*start)->get_shape();
+                auto input   = (*start)->inputs()[0];
+                if(iaxis >= bshape.strides().size() or bshape.strides()[iaxis] == 0)
+                    return {start, last};
+                op.from_value({{"out_lens", get_output_lens(start, last, iaxis)}});
+                auto delta = bshape.lens().size() - input->get_shape().lens().size();
+                iaxis -= delta;
+            }
 
             std::vector<instruction_ref> concats;
             for(std::size_t i = 0; i < x->inputs().size(); i++)
@@ -1260,6 +1360,7 @@ void simplify_algebra::apply(module& m) const
     {
         match::find_matches(m,
                             find_inner_broadcast{},
+                            find_dot_broadcast{},
                             find_double_add_lit_broadcast{},
                             find_add_lit_broadcast{},
                             find_add_convs{},

test/simplify_algebra_test.cpp

@@ -613,6 +613,60 @@ TEST_CASE(simplify_inner_broadcast_scalar)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(simplify_inner_broadcast_different_dims)
+{
+    auto b = migraphx::op::multibroadcast{{2, 384, 768}};
+    migraphx::module m1;
+    {
+        auto x   = m1.add_parameter("x", {migraphx::shape::int32_type, {384, 768}});
+        auto y   = m1.add_parameter("y", {migraphx::shape::int32_type, {768}});
+        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);
+    }
+    run_pass(m1);
+
+    migraphx::module m2;
+    {
+        auto x    = m2.add_parameter("x", {migraphx::shape::int32_type, {384, 768}});
+        auto y    = m2.add_parameter("y", {migraphx::shape::int32_type, {768}});
+        auto yb   = m2.add_instruction(migraphx::op::multibroadcast{{384, 768}}, y);
+        auto sum  = m2.add_instruction(migraphx::make_op("add"), x, yb);
+        auto sumb = m2.add_instruction(b, sum);
+        m2.add_instruction(pass_op{}, sumb);
+    }
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(simplify_inner_broadcast_different_broadcasts)
+{
+    auto b  = migraphx::op::broadcast{1, {1, 24, 112, 112}};
+    auto mb = migraphx::op::multibroadcast{{1, 24, 112, 112}};
+    migraphx::module m1;
+    {
+        auto x   = m1.add_parameter("x", {migraphx::shape::int32_type, {24}});
+        auto y   = m1.add_parameter("y", {migraphx::shape::int32_type, {24, 1, 1}});
+        auto xb  = m1.add_instruction(b, x);
+        auto yb  = m1.add_instruction(mb, y);
+        auto sum = m1.add_instruction(migraphx::make_op("add"), xb, yb);
+        m1.add_instruction(pass_op{}, sum);
+    }
+    run_pass(m1);
+
+    migraphx::module m2;
+    {
+        auto x    = m2.add_parameter("x", {migraphx::shape::int32_type, {24}});
+        auto y    = m2.add_parameter("y", {migraphx::shape::int32_type, {24, 1, 1}});
+        auto xs   = m2.add_instruction(migraphx::make_op("squeeze"), x);
+        auto ys   = m2.add_instruction(migraphx::make_op("squeeze"), y);
+        auto sum  = m2.add_instruction(migraphx::make_op("add"), xs, ys);
+        auto sumb = m2.add_instruction(b, sum);
+        m2.add_instruction(pass_op{}, sumb);
+    }
+    EXPECT(m1 == m2);
+}
+
 TEST_CASE(simplify_add_conv1)
 {
     migraphx::module m;
@@ -3003,6 +3057,38 @@ TEST_CASE(reorder_slice_ins_deps)
     EXPECT(m == create_module());
 }
 
+TEST_CASE(dot_broadcast_different_rank)
+{
+    migraphx::module m1;
+    {
+        auto x  = m1.add_parameter("x", {migraphx::shape::float_type, {768}});
+        auto y  = m1.add_parameter("y", {migraphx::shape::float_type, {768, 3072}});
+        auto xb = m1.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", {2, 384, 768}}}), x);
+        auto yb = m1.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", {2, 768, 3072}}}), y);
+        auto dot = m1.add_instruction(migraphx::make_op("dot"), xb, yb);
+        m1.add_return({dot});
+    };
+
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {768}});
+        auto y = m2.add_parameter("y", {migraphx::shape::float_type, {768, 3072}});
+        auto xb =
+            m2.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {384, 768}}}), x);
+        auto yb =
+            m2.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {768, 3072}}}), y);
+        auto dot       = m2.add_instruction(migraphx::make_op("dot"), xb, yb);
+        auto broadcast = m2.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", {2, 384, 3072}}}), dot);
+        m2.add_return({broadcast});
+    };
+
+    run_pass(m1);
+    EXPECT(m1.sort() == m2.sort());
+}
+
 TEST_CASE(dot_fusion_reshape)
 {
     migraphx::module m1;