Merge remote-tracking branch 'origin/fuse-horiz-contiguous' into attention-plus

src/eliminate_contiguous.cpp

@@ -70,16 +70,22 @@ static bool try_compute_shape(instruction_ref ins,
     return try_compute_shape(ins, inputs, mods);
 }
 
-void eliminate_contiguous::apply(module& m) const
+template <class F>
+static void remove_contiguous(const std::string& op_name, module& m, F f)
 {
-    std::vector<instruction_ref> const_instruction;
-
+    auto last = std::prev(m.end());
     for(auto ins : iterator_for(m))
     {
         // return instruction should have inputs with standard shape
         if(ins->name() == "@return")
             continue;
 
+        if(ins != last and ins->outputs().empty())
+            continue;
+
+        if(not f(ins))
+            continue;
+
         // Make a copy so we can modify it while we iterate
         auto args     = ins->inputs();
         auto new_args = args;
@@ -87,8 +93,8 @@ void eliminate_contiguous::apply(module& m) const
 
         for(auto arg : ins->inputs())
         {
-            if(arg->name() == op_name)
-            {
+            if(arg->name() != op_name)
+                continue;
             auto prev = arg->inputs().front();
             replace(new_args, arg, prev);
             if(try_compute_shape(ins, new_args, mod_args))
@@ -97,13 +103,16 @@ void eliminate_contiguous::apply(module& m) const
             }
             else if(prev->can_eval())
             {
-                    const_instruction.push_back(arg);
-                }
+                auto c = op::contiguous{};
+                auto r = c.compute(c.compute_shape({prev->get_shape()}), {prev->eval()});
+
+                auto l = m.add_literal(r.get_shape(), r.data());
+                m.replace_instruction(arg, l);
             }
         }
     }
 
-    // Perform evaluations in parallel
+    /* // Perform evaluations in parallel
     std::vector<argument> literals(const_instruction.size());
     par_for(const_instruction.size(), 1, [&](const auto i) {
         auto c      = op::contiguous{};
@@ -115,7 +124,18 @@ void eliminate_contiguous::apply(module& m) const
     {
         auto l = m.add_literal(literals[i].get_shape(), literals[i].data());
         m.replace_instruction(const_instruction[i], l);
-    }
+    } */
+}
+
+void eliminate_contiguous::apply(module& m) const
+{
+    // Skip contiguous from splits first
+    remove_contiguous(op_name, m, [](auto ins) {
+        if(ins->name() != "slice")
+            return true;
+        return (ins->inputs().front()->outputs().size() == 1);
+    });
+    remove_contiguous(op_name, m, [](auto) { return true; });
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/simplify_reshapes.cpp

@@ -578,6 +578,70 @@ struct find_transpose_contiguous_reshaper_unary
     }
 };
 
+struct find_transpose_slice
+{
+    auto matcher() const
+    {
+        return match::any(match::any_of[match::outputs()](
+            match::name("slice")(match::output(match::name("transpose")))));
+    }
+
+    static std::vector<int64_t> find_common_perm(const std::vector<instruction_ref>& transposes)
+    {
+        std::map<std::vector<int64_t>, int64_t> count;
+        for(auto t : transposes)
+        {
+            auto perm = t->get_operator().to_value()["permutation"].to_vector<int64_t>();
+            count[perm]++;
+        }
+        return std::max_element(
+                   count.begin(), count.end(), by(std::less<>{}, [](auto&& p) { return p.second; }))
+            ->first;
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins = r.result;
+        std::vector<instruction_ref> splits;
+        std::copy_if(ins->outputs().begin(),
+                     ins->outputs().end(),
+                     std::back_inserter(splits),
+                     [&](instruction_ref out) {
+                         return out->name() == "slice" and out->outputs().size() == 1 and
+                                out->outputs().front()->name() == "transpose";
+                     });
+        if(splits.size() < 2)
+            return;
+        std::vector<instruction_ref> transposes;
+        std::transform(splits.begin(),
+                       splits.end(),
+                       std::back_inserter(transposes),
+                       [](auto split) { return split->outputs().front(); });
+        auto perm  = find_common_perm(transposes);
+        auto iperm = invert_permutation(perm);
+        auto pre   = m.insert_instruction(
+            std::next(ins), make_op("transpose", {{"permutation", perm}}), ins);
+        for(auto i : range(transposes.size()))
+        {
+            auto split = splits[i];
+            auto t     = transposes[i];
+            auto op    = any_cast<op::slice>(split->get_operator());
+            for(auto& axis : op.axes)
+            {
+                axis = iperm[axis];
+            }
+            auto new_ins = m.insert_instruction(t, op, pre);
+            if(t->get_operator() != pre->get_operator())
+            {
+                auto curr = t->get_operator().to_value()["permutation"].to_vector<int64_t>();
+                new_ins   = m.insert_instruction(
+                    t, make_op("transpose", {{"permutation", reorder_dims(iperm, curr)}}), new_ins);
+            }
+            m.replace_instruction(t, new_ins);
+        }
+    }
+};
+
 void simplify_reshapes::apply(module& m) const
 {
     for(int i = 0; i < 2; i++)
@@ -593,6 +657,7 @@ void simplify_reshapes::apply(module& m) const
                             find_nested_convert{},
                             find_nested_slice{},
                             find_nested_concat{},
+                            find_transpose_slice{},
                             find_transpose_contiguous_reshaper_unary{});
         dead_code_elimination{}.apply(m);
     }

test/simplify_reshapes_test.cpp

@@ -1118,4 +1118,106 @@ TEST_CASE(transpose_contiguous_reshape_binary_broadcast)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(transpose_slice)
+{
+    migraphx::module m1;
+    {
+        auto x      = m1.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {12}}}), x);
+        auto transpose1 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice1);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {12}}, {"ends", {24}}}), x);
+        auto transpose2 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {24}}, {"ends", {36}}}), x);
+        auto transpose3 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice3);
+        m1.add_return({transpose1, transpose2, transpose3});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto transpose =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto slice2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {12}}, {"ends", {24}}}),
+            transpose);
+        auto slice3 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {24}}, {"ends", {36}}}),
+            transpose);
+        m2.add_return({slice1, slice2, slice3});
+    }
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(transpose_slice_diff_perm)
+{
+    migraphx::module m1;
+    {
+        auto x      = m1.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {12}}}), x);
+        auto transpose1 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice1);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {12}}, {"ends", {24}}}), x);
+        auto transpose2 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 3, 1}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {24}}, {"ends", {36}}}), x);
+        auto transpose3 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice3);
+        m1.add_return({transpose1, transpose2, transpose3});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto transpose =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto slice2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {12}}, {"ends", {24}}}),
+            transpose);
+        auto transpose2 = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), slice2);
+        auto slice3 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {24}}, {"ends", {36}}}),
+            transpose);
+        m2.add_return({slice1, transpose2, slice3});
+    }
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(transpose_slice_single_transpose)
+{
+    migraphx::module m1;
+    {
+        auto x      = m1.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {12}}}), x);
+        auto sqrt1  = m1.add_instruction(migraphx::make_op("sqrt"), slice1);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {12}}, {"ends", {24}}}), x);
+        auto transpose2 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {24}}, {"ends", {36}}}), x);
+        auto sqrt3 = m1.add_instruction(migraphx::make_op("sqrt"), slice3);
+        m1.add_return({sqrt1, transpose2, sqrt3});
+    }
+    migraphx::module m2 = m1;
+    run_pass(m1);
+    EXPECT(m1 == m2);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }