remoting: cache the buffer_get_base result

Author: kpouget

ggml/src/ggml-remotingfrontend/ggml-backend-buffer-type.cpp

@@ -15,6 +15,7 @@ ggml_backend_remoting_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
 
   context->gpu = gpu;
   context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size);
+  context->base = NULL;
 
   ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft, ggml_backend_remoting_buffer_interface, (void *) context, size);
 

ggml/src/ggml-remotingfrontend/ggml-backend-buffer.cpp

@@ -4,11 +4,17 @@
   ((struct ggml_backend_remoting_buffer_context *) (name)->context)->gpu
 
 static void * ggml_backend_remoting_buffer_get_base(ggml_backend_buffer_t buffer) {
-  //IMPLEMENTED;
+  IMPLEMENTED_ONCE;
 
-  struct virtgpu *gpu = BUFFER_TO_GPU(buffer);
+  struct ggml_backend_remoting_buffer_context *context = (struct ggml_backend_remoting_buffer_context *) buffer->context;
+  if (context->base) {
+    return context->base;
+  }
+
+  context->base = apir_buffer_get_base(BUFFER_TO_GPU(buffer),
+				       BUFFER_TO_APIR_CONTEXT(buffer));
 
-  return apir_buffer_get_base(gpu, BUFFER_TO_APIR_CONTEXT(buffer));
+  return context->base;
 }
 
 static void ggml_backend_remoting_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {

ggml/src/ggml-remotingfrontend/ggml-remoting.h

@@ -88,6 +88,8 @@ struct ggml_backend_remoting_buffer_context {
   apir_buffer_context_t apir_context;
 
   struct virtgpu *gpu;
+
+  void *base;
 };
 
 extern const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface;