remoting: implement the free_buffer function

Author: kpouget

ggml/src/ggml-remotingbackend/backend-dispatched-buffer.cpp

@@ -114,3 +114,16 @@ backend_buffer_clear(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struc
 
   return 0;
 }
+
+uint32_t
+backend_buffer_free_buffer(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struct virgl_apir_context *ctx) {
+  UNUSED(ctx);
+  UNUSED(enc);
+
+  ggml_backend_buffer_t buffer;
+  buffer = vn_decode_ggml_buffer(dec);
+
+  buffer->iface.free_buffer(buffer);
+
+  return 0;
+}

ggml/src/ggml-remotingbackend/backend-dispatched.h

@@ -40,6 +40,7 @@ uint32_t backend_buffer_get_base(struct vn_cs_encoder *enc, struct vn_cs_decoder
 uint32_t backend_buffer_set_tensor(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struct virgl_apir_context *ctx);
 uint32_t backend_buffer_get_tensor(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struct virgl_apir_context *ctx);
 uint32_t backend_buffer_clear(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struct virgl_apir_context *ctx);
+uint32_t backend_buffer_free_buffer(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struct virgl_apir_context *ctx);
 
 /* backend */
 uint32_t backend_graph_compute(struct vn_cs_encoder *enc, struct vn_cs_decoder *dec, struct virgl_apir_context *ctx);
@@ -69,6 +70,7 @@ static inline const char *backend_dispatch_command_name(ApirBackendCommandType t
   case APIR_COMMAND_TYPE_BUFFER_SET_TENSOR: return "backend_buffer_set_tensor";
   case APIR_COMMAND_TYPE_BUFFER_GET_TENSOR: return "backend_buffer_get_tensor";
   case APIR_COMMAND_TYPE_BUFFER_CLEAR: return "backend_buffer_clear";
+  case APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER: return "backend_buffer_free_buffer";
 
   /* backend */
   case APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE: return "backend_graph_compute";
@@ -99,6 +101,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
   [APIR_COMMAND_TYPE_BUFFER_SET_TENSOR] = backend_buffer_set_tensor,
   [APIR_COMMAND_TYPE_BUFFER_GET_TENSOR] = backend_buffer_get_tensor,
   [APIR_COMMAND_TYPE_BUFFER_CLEAR] = backend_buffer_clear,
+  [APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER] = backend_buffer_free_buffer,
 
   /* backend */
   [APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE] = backend_graph_compute,

ggml/src/ggml-remotingbackend/shared/apir_backend.h

@@ -52,12 +52,13 @@ typedef enum ApirBackendCommandType {
   APIR_COMMAND_TYPE_BUFFER_SET_TENSOR = 14,
   APIR_COMMAND_TYPE_BUFFER_GET_TENSOR = 15,
   APIR_COMMAND_TYPE_BUFFER_CLEAR = 16,
+  APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER = 17,
 
   /* backend */
-  APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE = 17,
+  APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE = 18,
 
   // last command_type index + 1
-  APIR_BACKEND_DISPATCH_TABLE_COUNT = 18,
+  APIR_BACKEND_DISPATCH_TABLE_COUNT = 19,
 } ApirBackendCommandType;
 
 

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

@@ -76,9 +76,11 @@ static void ggml_backend_remoting_buffer_clear(ggml_backend_buffer_t buffer, uin
 static void ggml_backend_remoting_buffer_free_buffer(ggml_backend_buffer_t buffer) {
   UNUSED(buffer);
 
-  NOT_IMPLEMENTED;
+  IMPLEMENTED_ONCE;
 
-  STOP_HERE;
+  struct virtgpu *gpu = BUFFER_TO_GPU(buffer);
+
+  apir_buffer_free_buffer(gpu, BUFFER_TO_HANDLE(buffer));
 }
 
 const ggml_backend_buffer_i ggml_backend_remoting_buffer_interface = {

ggml/src/ggml-remotingfrontend/virtgpu-forward-buffer.cpp

@@ -99,3 +99,18 @@ apir_buffer_clear(struct virtgpu *gpu, apir_buffer_handle_t buffer_handle,
 
   REMOTE_CALL_FINISH(gpu, encoder, decoder);
 }
+
+
+void
+apir_buffer_free_buffer(struct virtgpu *gpu, apir_buffer_handle_t buffer_handle) {
+  struct vn_cs_encoder *encoder;
+  struct vn_cs_decoder *decoder;
+
+  REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER);
+
+  vn_encode_apir_buffer_handle_t(encoder, &buffer_handle);
+
+  REMOTE_CALL(gpu, encoder, decoder);
+
+  REMOTE_CALL_FINISH(gpu, encoder, decoder);
+}

ggml/src/ggml-remotingfrontend/virtgpu-forward.h

@@ -35,6 +35,7 @@ void apir_buffer_get_tensor(struct virtgpu *gpu, apir_buffer_handle_t buffer_han
 			    const ggml_tensor *tensor, void *data, size_t offset, size_t size);
 void apir_buffer_clear(struct virtgpu *gpu, apir_buffer_handle_t buffer_handle,
 		       uint8_t value);
+void apir_buffer_free_buffer(struct virtgpu *gpu, apir_buffer_handle_t buffer_handle);
 
 /* backend */