manual merge

src/include/migraphx/op/sigmoid.hpp

@@ -17,9 +17,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct sigmoid : unary
+struct sigmoid : unary<sigmoid>
 {
-    std::string name() const { return "sigmoid"; }
+    auto apply() const
+    {
+        return [](auto x) { return 1.f / (1.f + std::exp(-x)); };
+    }
 };
 
 } // namespace op

src/include/migraphx/op/sin.hpp

@@ -17,9 +17,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct sin : unary
+struct sin : unary<sin>
 {
-    std::string name() const { return "sin"; }
+    auto apply() const
+    {
+        return [](auto x) { return std::sin(x); };
+    }
 };
 
 } // namespace op

src/include/migraphx/op/sinh.hpp

@@ -17,9 +17,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct sinh : unary
+struct sinh : unary<sinh>
 {
-    std::string name() const { return "sinh"; }
+    auto apply() const
+    {
+        return [](auto x) { return std::sinh(x); };
+    }
 };
 
 } // namespace op

src/include/migraphx/op/sub.hpp

@@ -17,9 +17,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct sub : binary
+struct sub : binary<sub>
 {
-    std::string name() const { return "sub"; }
+    auto apply() const
+    {
+        return [](auto x, auto y) { return x - y; };
+    }
 };
 
 } // namespace op

src/include/migraphx/op/tan.hpp

@@ -17,9 +17,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct tan : unary
+struct tan : unary<tan>
 {
-    std::string name() const { return "tan"; }
+    auto apply() const
+    {
+        return [](auto x) { return std::tan(x); };
+    }
 };
 
 } // namespace op

src/include/migraphx/op/tanh.hpp

@@ -17,9 +17,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct tanh : unary
+struct tanh : unary<tanh>
 {
-    std::string name() const { return "tanh"; }
+    auto apply() const
+    {
+        return [](auto x) { return std::tanh(x); };
+    }
 };
 
 } // namespace op

src/include/migraphx/op/unary.hpp

 #ifndef MIGRAPHX_GUARD_OPERATORS_UNARY_HPP
 #define MIGRAPHX_GUARD_OPERATORS_UNARY_HPP
 
-#include <array>
-#include <migraphx/operation.hpp>
-#include <migraphx/check_shapes.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/streamutils.hpp>
-#include <migraphx/literal.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
-#include <cmath>
-#include <utility>
+#include <migraphx/op/name.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct unary
+template <class Derived>
+struct unary : op_name<Derived>
 {
     shape compute_shape(std::vector<shape> inputs) const
     {
         check_shapes{inputs}.has(1);
         return inputs.at(0);
     }
+    argument compute(const shape& output_shape, std::vector<argument> args) const
+    {
+        argument result{output_shape};
+        visit_all(result, args[0])([&](auto output, auto input) {
+            if(input.get_shape().standard())
+            {
+                std::transform(input.begin(),
+                               input.end(),
+                               output.begin(),
+                               static_cast<const Derived&>(*this).apply());
+            }
+            else
+            {
+                shape_for_each(output.get_shape(), [&](const auto& idx) {
+                    output(idx.begin(), idx.end()) =
+                        static_cast<const Derived&>(*this).apply()(input(idx.begin(), idx.end()));
+                });
+            }
+        });
+        return result;
+    }
 };
 
 } // namespace op

src/include/migraphx/constant_propagate.hpp → src/include/migraphx/propagate_constant.hpp

-#ifndef MIGRAPHX_GUARD_RTGLIB_CONSTANT_PROPAGATE_HPP
-#define MIGRAPHX_GUARD_RTGLIB_CONSTANT_PROPAGATE_HPP
+#ifndef MIGRAPHX_GUARD_RTGLIB_PROPAGATE_CONSTANT_HPP
+#define MIGRAPHX_GUARD_RTGLIB_PROPAGATE_CONSTANT_HPP
 
 #include <string>
 #include <migraphx/config.hpp>
@@ -12,9 +12,9 @@ struct program;
 /**
  * Replace instructions which take all literals with a literal of the computation.
  */
-struct constant_propagate
+struct propagate_constant
 {
-    std::string name() const { return "constant_propagate"; }
+    std::string name() const { return "propagate_constant"; }
     void apply(program& p) const;
 };
 

src/propagate_constant.cpp

+#include <migraphx/propagate_constant.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/literal.hpp>
+#include <migraphx/functional.hpp>
+#include <unordered_set>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+bool skip_propogate(instruction_ref ins)
+{
+    if(ins->name() == "@literal")
+        return true;
+    auto&& s = ins->get_shape();
+    if(s.broadcasted() and not s.scalar())
+        return true;
+    if(s.scalar() and s.elements() != 1)
+        return true;
+    return false;
+}
+
+void propagate_constant::apply(program& p) const
+{
+    fix([&](auto self, auto ins) {
+        if(not skip_propogate(ins))
+        {
+            auto r = ins->eval();
+            if(not r.empty())
+            {
+                assert(r.get_shape() == ins->get_shape());
+                auto l = p.add_literal(r.get_shape(), r.data());
+                p.replace_instruction(ins, l);
+                return;
+            }
+        }
+        std::unordered_set<instruction_ref> children(ins->inputs().begin(), ins->inputs().end());
+        for(auto child : children)
+            self(child);
+    })(std::prev(p.end()));
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/cpu/lowering.cpp

@@ -614,6 +614,7 @@ struct cpu_unary
                 std::transform(input.begin(), input.end(), output.begin(), op.fcn());
             });
         });
+
         return result;
     }
 };

src/targets/gpu/CMakeLists.txt

@@ -66,6 +66,7 @@ add_library(migraphx_gpu
     gather.cpp
     lrn.cpp
     schedule_model.cpp
+    adjust_allocation.cpp
     clip.cpp
 )
 set_target_properties(migraphx_gpu PROPERTIES EXPORT_NAME gpu)

src/targets/gpu/abs.cpp

@@ -8,7 +8,7 @@ namespace gpu {
 shape miopen_abs::compute_shape(const std::vector<shape>& inputs) const
 {
     check_shapes{inputs, *this}.has(2).not_broadcasted();
-    return inputs.at(1);
+    return inputs.at(0);
 }
 
 argument miopen_abs::compute(context& ctx,

src/targets/gpu/adjust_allocation.cpp

+#include <migraphx/gpu/adjust_allocation.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <algorithm>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+void adjust_allocation::apply(program& p) const
+{
+    for(auto ins : iterator_for(p))
+    {
+        // skip instruction with no input
+        if(ins->inputs().empty())
+            continue;
+
+        if(ins->name() == "load")
+            continue;
+
+        auto alias_ins = instruction::get_output_alias(ins, true);
+        if(alias_ins->name() == "hip::allocate")
+        {
+            // shape allocated is different from actual shape
+            // of the instruction, reallocate and replace the previous one
+            if(alias_ins->get_shape() != ins->get_shape())
+            {
+                auto alloc_ins = p.insert_instruction(ins, hip_allocate{ins->get_shape()});
+                p.replace_instruction(alias_ins, alloc_ins);
+            }
+        }
+    }
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/include/migraphx/gpu/adjust_allocation.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_ADJUST_ALLOCATION_HPP
+#define MIGRAPHX_GUARD_RTGLIB_ADJUST_ALLOCATION_HPP
+
+#include <migraphx/program.hpp>
+#include <migraphx/config.hpp>
+#include <migraphx/gpu/context.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+namespace gpu {
+
+struct adjust_allocation
+{
+    std::string name() const { return "gpu::adjust_allocation"; }
+    void apply(program& p) const;
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/oper.hpp

@@ -45,7 +45,7 @@ struct unary_device : oper<Derived>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(2);
-        return inputs.at(0);
+        return inputs.at(1);
     }
 
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
@@ -63,7 +63,7 @@ struct binary_device : oper<Derived>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(3);
-        return inputs.at(0);
+        return inputs.at(2);
     }
 
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const

src/targets/gpu/tanh.cpp

@@ -8,7 +8,7 @@ namespace gpu {
 shape miopen_tanh::compute_shape(const std::vector<shape>& inputs) const
 {
     check_shapes{inputs, *this}.has(2).not_broadcasted();
-    return inputs.at(1);
+    return inputs.at(0);
 }
 
 argument miopen_tanh::compute(context& ctx,

src/targets/gpu/target.cpp

@@ -11,7 +11,7 @@
 #include <migraphx/dead_code_elimination.hpp>
 #include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/simplify_algebra.hpp>
-#include <migraphx/constant_propagate.hpp>
+#include <migraphx/propagate_constant.hpp>
 #include <migraphx/eliminate_contiguous.hpp>
 #include <migraphx/common_subexpression_elimination.hpp>
 #include <migraphx/fwd_conv_batchnorm_rewrite.hpp>
@@ -20,6 +20,7 @@
 #include <migraphx/eliminate_identity.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
 #include <migraphx/gpu/schedule_model.hpp>
+#include <migraphx/gpu/adjust_allocation.hpp>
 #include <migraphx/eliminate_pad.hpp>
 #include <migraphx/schedule.hpp>
 
@@ -47,7 +48,7 @@ std::vector<pass> target::get_passes(migraphx::context& gctx) const
         //dead_code_elimination{},
         simplify_algebra{},
         dead_code_elimination{},
-        constant_propagate{},
+        propagate_constant{},
         dead_code_elimination{},
         auto_contiguous{},
         //simplify_reshapes{},
@@ -57,6 +58,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx) const
         dead_code_elimination{},
         eliminate_contiguous{},
         dead_code_elimination{},
+        adjust_allocation{},
+        dead_code_elimination{},
         fuse_ops{&ctx},
         dead_code_elimination{},
         write_literals{&ctx},

test/gpu/adjust_allocation.cpp

+#include <migraphx/gpu/adjust_allocation.hpp>
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/gpu/lowering.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/auto_contiguous.hpp>
+#include <migraphx/eliminate_contiguous.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/op/add.hpp>
+#include <migraphx/op/transpose.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/op/tanh.hpp>
+#include <basic_ops.hpp>
+#include <test.hpp>
+
+struct lowering_target
+{
+    std::string name() const { return "gpu::lowering"; }
+    std::vector<migraphx::pass> get_passes(migraphx::context& gctx) const
+    {
+        auto& ctx = migraphx::any_cast<migraphx::gpu::context>(gctx);
+        return {migraphx::auto_contiguous{},
+                migraphx::gpu::lowering{ctx},
+                migraphx::dead_code_elimination{},
+                migraphx::eliminate_contiguous{},
+                migraphx::dead_code_elimination{}};
+    }
+    migraphx::gpu::context get_context() const { return migraphx::gpu::context{}; }
+};
+
+TEST_CASE(tanh_shape)
+{
+    auto create_program = [] {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        auto x   = p.add_parameter("x", s);
+        auto tx  = p.add_instruction(migraphx::op::transpose{{1, 0}}, x);
+        auto txh = p.add_instruction(migraphx::op::tanh{}, tx);
+        p.add_instruction(migraphx::op::add{}, txh, txh);
+
+        return p;
+    };
+
+    auto p1 = create_program();
+    auto p2 = create_program();
+    EXPECT(p1 == p2);
+
+    p1.compile(lowering_target{});
+    p2.compile(lowering_target());
+
+    EXPECT(p1 == p2);
+
+    for(auto ins : iterator_for(p1))
+    {
+        if(ins->name() == "hip::allocate")
+        {
+            migraphx::shape wrong_s{migraphx::shape::float_type, {3, 2}, {1, 3}};
+            ins->replace(wrong_s);
+        }
+    }
+    EXPECT(p1 != p2);
+
+    migraphx::run_passes(p2,
+                         {migraphx::gpu::adjust_allocation{}, migraphx::dead_code_elimination{}});
+    EXPECT(p1 == p2);
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/miopen.cpp

@@ -236,8 +236,7 @@ struct test_exp : verify_program<test_exp>
     {
         migraphx::program p;
         migraphx::shape s{migraphx::shape::float_type, {6}};
-        std::vector<float> data{0.1f, 0.2f, 1.f, 2.f, 0.6f, 10.f};
-        auto x = p.add_literal(s, data);
+        auto x = p.add_instruction(migraphx::op::abs{}, p.add_parameter("x", s));
         p.add_instruction(migraphx::op::exp{}, x);
         return p;
     }
@@ -249,8 +248,7 @@ struct test_log : verify_program<test_log>
     {
         migraphx::program p;
         migraphx::shape s{migraphx::shape::float_type, {6}};
-        std::vector<float> data{0.1f, 0.2f, 1.f, 2.f, 0.6f, 100.f};
-        auto x = p.add_literal(s, data);
+        auto x = p.add_instruction(migraphx::op::abs{}, p.add_parameter("x", s));
         p.add_instruction(migraphx::op::log{}, x);
         return p;
     }
@@ -327,6 +325,19 @@ struct test_tanh : verify_program<test_tanh>
     }
 };
 
+struct test_trans_tanh : verify_program<test_trans_tanh>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x  = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        auto tx = p.add_instruction(migraphx::op::transpose{{0, 1, 3, 2}}, x);
+        auto tanhx = p.add_instruction(migraphx::op::tanh{}, tx);
+        p.add_instruction(migraphx::op::add{}, tanhx, tanhx);
+        return p;
+    }
+};
+
 struct test_asin : verify_program<test_asin>
 {
     migraphx::program create_program() const
@@ -674,6 +685,19 @@ struct test_abs : verify_program<test_abs>
     }
 };
 
+struct test_trans_abs : verify_program<test_trans_abs>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x  = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        auto tx = p.add_instruction(migraphx::op::transpose{{0, 1, 3, 2}}, x);
+        auto tanhx = p.add_instruction(migraphx::op::abs{}, tx);
+        p.add_instruction(migraphx::op::add{}, tanhx, tanhx);
+        return p;
+    }
+};
+
 struct test_leaky_relu : verify_program<test_leaky_relu>
 {
     migraphx::program create_program() const

test/constant_propagate_test.cpp → test/propagate_constant_test.cpp

-#include <migraphx/constant_propagate.hpp>
+#include <migraphx/propagate_constant.hpp>
 #include <migraphx/dead_code_elimination.hpp>
 #include <migraphx/op/add.hpp>
+#include <migraphx/op/scalar.hpp>
+#include <migraphx/op/mul.hpp>
 #include <basic_ops.hpp>
 #include <test.hpp>
 
@@ -9,12 +11,12 @@ struct const_prop_target
     std::string name() const { return "const_prop"; }
     std::vector<migraphx::pass> get_passes(migraphx::context&) const
     {
-        return {migraphx::constant_propagate{}, migraphx::dead_code_elimination{}};
+        return {migraphx::propagate_constant{}, migraphx::dead_code_elimination{}};
     }
     migraphx::context get_context() const { return {}; }
 };
 
-TEST_CASE(const_add1)
+TEST_CASE(const_add)
 {
     migraphx::program p1;
     auto one = p1.add_literal(1);
@@ -29,7 +31,7 @@ TEST_CASE(const_add1)
     EXPECT(p1 == p2);
 }
 
-TEST_CASE(const_add2)
+TEST_CASE(const_add_parameter)
 {
     migraphx::program p1;
     auto one = p1.add_parameter("one", {migraphx::shape::int32_type, {1}});
@@ -44,7 +46,7 @@ TEST_CASE(const_add2)
     EXPECT(p1 != p2);
 }
 
-TEST_CASE(const_add3)
+TEST_CASE(const_multiadd)
 {
     migraphx::program p1;
     auto one  = p1.add_literal(1);
@@ -60,4 +62,54 @@ TEST_CASE(const_add3)
     EXPECT(p1 == p2);
 }
 
+TEST_CASE(const_add_mul)
+{
+    migraphx::program p1;
+    auto one  = p1.add_literal(1);
+    auto two  = p1.add_literal(2);
+    auto mul  = p1.add_instruction(migraphx::op::mul{}, two, two);
+    auto sum1 = p1.add_instruction(migraphx::op::add{}, one, mul);
+    auto sum2 = p1.add_instruction(migraphx::op::add{}, sum1, two);
+    p1.add_instruction(pass_op{}, sum2);
+    p1.compile(const_prop_target{});
+
+    migraphx::program p2;
+    auto total = p2.add_literal(7);
+    p2.add_instruction(pass_op{}, total);
+    EXPECT(p1 == p2);
+}
+
+TEST_CASE(const_add_scalar)
+{
+    migraphx::program p1;
+    auto one = p1.add_instruction(migraphx::op::scalar{{2, 2}}, p1.add_literal(1));
+    auto two = p1.add_instruction(migraphx::op::scalar{{2, 2}}, p1.add_literal(2));
+    auto sum = p1.add_instruction(migraphx::op::add{}, one, two);
+    p1.add_instruction(pass_op{}, sum);
+    p1.compile(const_prop_target{});
+
+    migraphx::program p2;
+    auto total =
+        p2.add_literal(migraphx::literal{{migraphx::shape::int32_type, {2, 2}}, {3, 3, 3, 3}});
+    p2.add_instruction(pass_op{}, total);
+    EXPECT(p1 == p2);
+}
+
+TEST_CASE(const_scalar)
+{
+    migraphx::program p1;
+    {
+        auto one = p1.add_instruction(migraphx::op::scalar{{2, 2}}, p1.add_literal(1));
+        p1.add_instruction(pass_op{}, one);
+    }
+    p1.compile(const_prop_target{});
+
+    migraphx::program p2;
+    {
+        auto one = p2.add_instruction(migraphx::op::scalar{{2, 2}}, p2.add_literal(1));
+        p2.add_instruction(pass_op{}, one);
+    }
+    EXPECT(p1 == p2);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }