graph: dnnl: comments fix

Author: xiang1guo

src/graph/backend/dnnl/kernels/sdp_primitive_v1.cpp

@@ -192,10 +192,6 @@ status_t sdp_primitive_v1_kernel_t::ocl_execute_impl(const stream_t *g_stream,
         const std::vector<tensor_t> &inputs,
         const std::vector<tensor_t> &outputs,
         const std::vector<cl_event> &cl_deps, cl_event *ret_event) {
-// sdp_primitive_v1_kernel_t only supports Intel GPU.
-#if DNNL_GPU_VENDOR != DNNL_VENDOR_INTEL
-    return status::unimplemented;
-#endif
     auto deps = cl_deps;
     cl_event returned_event {};
 

src/graph/backend/dnnl/op_executable.hpp

@@ -2657,16 +2657,14 @@ struct sdpa_executable_t : public op_executable_t {
 
         auto scale_dt = impl::data_type::undef;
         size_t idx = 3;
-        with_scale_ = op->get_attr<bool>(
-                dnnl::impl::graph::dnnl_impl::op_attr::with_scale);
+        with_scale_ = op->get_attr<bool>(op_attr::with_scale);
         if (with_scale_)
             scale_dt = op->get_input_value(idx++)
                                ->get_logical_tensor()
                                .data_type;
 
         dnnl::memory::desc md_mask;
-        with_mask_ = op->get_attr<bool>(
-                dnnl::impl::graph::dnnl_impl::op_attr::with_mask);
+        with_mask_ = op->get_attr<bool>(op_attr::with_mask);
         if (with_mask_)
             md_mask = make_dnnl_memory_desc(
                     op->get_input_value(idx++)->get_logical_tensor());
@@ -2675,13 +2673,10 @@ struct sdpa_executable_t : public op_executable_t {
         attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
         attr.set_fpmath_mode(
                 static_cast<dnnl::fpmath_mode>(mgr.get_fpmath_mode().mode_));
-        if (op->has_attr(
-                    dnnl::impl::graph::dnnl_impl::op_attr::is_invert_scale))
-            is_invert_scale_ = op->get_attr<bool>(
-                    dnnl::impl::graph::dnnl_impl::op_attr::is_invert_scale);
+        if (op->has_attr(op_attr::is_invert_scale))
+            is_invert_scale_ = op->get_attr<bool>(op_attr::is_invert_scale);
 
-        is_causal_mask_ = op->get_attr<bool>(
-                dnnl::impl::graph::dnnl_impl::op_attr::with_causal);
+        is_causal_mask_ = op->get_attr<bool>(op_attr::with_causal);
 
         dim_t kv_head_number
                 = op->get_input_value(1)->get_logical_tensor().dims[1];
@@ -2692,15 +2687,11 @@ struct sdpa_executable_t : public op_executable_t {
             is_initialized_ = false;
         } else {
             status_t s = sdpa_pd_->create_primitive(sdpa_prim_, p_engine.get());
-            if (s != dnnl::impl::status::success) {
-                is_initialized_ = false;
-            } else {
-                is_initialized_ = true;
-            }
+            is_initialized_ = s == status::success ? true : false;
         }
     }
 
-    bool is_initialized() { return is_initialized_; }
+    bool is_initialized() const { return is_initialized_; }
 
     void execute(const stream &stream,
             const std::unordered_map<int, memory> &args) const override {

src/graph/backend/dnnl/passes/transform.cpp

@@ -4272,7 +4272,7 @@ status_t fuse_sdpa(std::shared_ptr<subgraph_t> &sg) {
                 sdpa_op->set_attr<bool>(op_attr::is_invert_scale,
                         (alg == dnnl::algorithm::binary_div));
             }
-            // hanlde explicit mask
+            // handle explicit mask
             else if (alg == dnnl::algorithm::binary_add) {
                 auto mask_val = op->get_input_value(1);
                 mask_val->remove_consumer(*op, 1);