Merge branch 'develop' into pybuf

src/opt/memory_coloring_impl.cpp

@@ -203,9 +203,8 @@ void memory_coloring_impl::rewrite()
 
             if(is_allocate(ins))
             {
-                assert(!ins->inputs().empty());
                 p_program->replace_instruction(
-                    ins, op::load{ins->inputs().at(0)->get_shape(), offset}, scratch_param);
+                    ins, op::load{ins->get_shape(), offset}, scratch_param);
             }
             else if(is_literal(ins))
             {

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

@@ -11,7 +11,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_DISABLE_NULL_STREAM)
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_ENABLE_NULL_STREAM)
 
 struct hip_device
 {
@@ -40,7 +40,7 @@ struct hip_device
 
         hipStream_t get()
         {
-            if(enabled(MIGRAPHX_DISABLE_NULL_STREAM{}))
+            if(not enabled(MIGRAPHX_ENABLE_NULL_STREAM{}))
             {
                 setup();
                 if(s == nullptr)
@@ -53,7 +53,7 @@ struct hip_device
 
         auto create_miopen_handle()
         {
-            if(enabled(MIGRAPHX_DISABLE_NULL_STREAM{}))
+            if(not enabled(MIGRAPHX_ENABLE_NULL_STREAM{}))
                 return make_obj<miopen_handle>(&miopenCreateWithStream, get());
             else
                 return make_obj<miopen_handle>(&miopenCreate);

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

@@ -23,12 +23,13 @@ void copy_to_gpu(const argument& src, const argument& dst);
 
 struct hip_allocate
 {
+    shape s;
     std::string tag{};
     std::string name() const { return "hip::allocate"; }
     shape compute_shape(const std::vector<shape>& inputs) const
     {
-        check_shapes{inputs}.has(1);
-        return inputs.front();
+        check_shapes{inputs}.has(0);
+        return s;
     }
     argument compute(context&, const shape& output_shape, const std::vector<argument>&) const
     {

src/targets/gpu/lowering.cpp

@@ -127,8 +127,7 @@ struct miopen_apply
         }
         else
         {
-            auto is     = prog->add_outline(s);
-            auto result = prog->insert_instruction(ins, hip_allocate{std::move(tag)}, is);
+            auto result = prog->insert_instruction(ins, hip_allocate{s, std::move(tag)});
             return result;
         }
     }

src/targets/gpu/write_literals.cpp

@@ -37,8 +37,7 @@ void write_literals::apply(program& p) const
             {
                 literal l  = ins->get_literal();
                 auto pre   = p.add_literal(l);
-                auto s     = p.add_outline(l.get_shape());
-                auto alloc = p.insert_instruction(std::next(pre), hip_allocate{}, s);
+                auto alloc = p.insert_instruction(std::next(pre), hip_allocate{l.get_shape()});
                 p.replace_instruction(ins, hip_copy{}, pre, alloc);
             }
             else

test/memory_coloring_test.cpp

@@ -21,8 +21,8 @@ struct allocate
     std::string name() const { return "allocate"; }
     migraphx::shape compute_shape(const std::vector<migraphx::shape>& inputs) const
     {
-        migraphx::check_shapes{inputs, *this}.has(1);
-        return inputs.front();
+        migraphx::check_shapes{inputs, *this}.has(0);
+        return s;
     }
     migraphx::argument compute(migraphx::context&,
                                const migraphx::shape& output_shape,
@@ -34,8 +34,7 @@ struct allocate
 
 migraphx::instruction_ref add_alloc(migraphx::program& p, const migraphx::shape& s)
 {
-    auto a0 = p.add_outline(s);
-    return p.add_instruction(allocate{}, a0);
+    return p.add_instruction(allocate{s});
 }
 
 bool no_allocate(const migraphx::program& p)