glthread,gallium: add a CAP to disable glBufferSubData optimization in glthread

it regresses performance on iris

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20379>

docs/gallium/screen.rst

@@ -638,6 +638,7 @@ The integer capabilities:
 * ``PIPE_CAP_MAX_CONSTANT_BUFFER_SIZE_UINT``: Maximum bound constant buffer size in bytes. This is unsigned integer with the maximum of 4GB - 1. This applies to all constant buffers used by UBOs, unlike ``PIPE_SHADER_CAP_MAX_CONST_BUFFER0_SIZE``, which is specifically for GLSL uniforms.
 * ``PIPE_CAP_HARDWARE_GL_SELECT``: Enable hardware accelerated GL_SELECT for this driver.
 * ``PIPE_CAP_DEVICE_PROTECTED_CONTEXT``: Whether the device supports protected / encrypted context which can manipulate protected / encrypted content (some devices might need protected contexts to access protected content, whereas ``PIPE_CAP_DEVICE_PROTECTED_SURFACE`` does not require any particular context to do so).
+* ``PIPE_CAP_ALLOW_GLTHREAD_BUFFER_SUBDATA_OPT``: Whether to allow glthread to convert glBufferSubData to glCopyBufferSubData. This may improve or worsen performance depending on your driver.
 
 .. _pipe_capf:
 

src/gallium/auxiliary/util/u_screen.c

@@ -497,6 +497,7 @@ u_pipe_screen_get_param_defaults(struct pipe_screen *pscreen,
    case PIPE_CAP_QUERY_SPARSE_TEXTURE_RESIDENCY:
    case PIPE_CAP_CLAMP_SPARSE_TEXTURE_LOD:
    case PIPE_CAP_TIMELINE_SEMAPHORE_IMPORT:
+   case PIPE_CAP_ALLOW_GLTHREAD_BUFFER_SUBDATA_OPT:
       return 0;
 
    case PIPE_CAP_MAX_CONSTANT_BUFFER_SIZE_UINT:

src/gallium/drivers/radeonsi/si_get.c

@@ -165,6 +165,7 @@ static int si_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_GLSL_TESS_LEVELS_AS_INPUTS:
    case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
    case PIPE_CAP_TEXTURE_MULTISAMPLE:
+   case PIPE_CAP_ALLOW_GLTHREAD_BUFFER_SUBDATA_OPT: /* TODO: remove if it's slow */
       return 1;
 
    case PIPE_CAP_TEXTURE_TRANSFER_MODES:

src/gallium/include/pipe/p_defines.h

@@ -1012,6 +1012,7 @@ enum pipe_cap
    PIPE_CAP_QUERY_TIMESTAMP_BITS,
    /** For EGL_EXT_protected_content */
    PIPE_CAP_DEVICE_PROTECTED_CONTEXT,
+   PIPE_CAP_ALLOW_GLTHREAD_BUFFER_SUBDATA_OPT,
 
    PIPE_CAP_LAST,
    /* XXX do not add caps after PIPE_CAP_LAST! */

src/mesa/main/consts_exts.h

@@ -992,5 +992,8 @@ struct gl_constants
 
    /** Use hardware accelerated GL_SELECT */
    bool HardwareAcceleratedSelect;
+
+   /** Allow GLThread to convert glBuffer */
+   bool AllowGLThreadBufferSubDataOpt;
 };
 #endif

src/mesa/main/glthread_bufferobj.c

@@ -498,6 +498,7 @@ _mesa_marshal_BufferSubData_merged(GLuint target_or_name, GLintptr offset,
     *       the buffer storage, but we don't know the buffer size in glthread.
     */
    if (ctx->GLThread.SupportsBufferUploads &&
+       ctx->Const.AllowGLThreadBufferSubDataOpt &&
        data && offset > 0 && size > 0) {
       struct gl_buffer_object *upload_buffer = NULL;
       unsigned upload_offset = 0;

src/mesa/state_tracker/st_extensions.c

@@ -628,6 +628,9 @@ void st_init_limits(struct pipe_screen *screen,
 
    c->HardwareAcceleratedSelect =
       screen->get_param(screen, PIPE_CAP_HARDWARE_GL_SELECT);
+
+   c->AllowGLThreadBufferSubDataOpt =
+      screen->get_param(screen, PIPE_CAP_ALLOW_GLTHREAD_BUFFER_SUBDATA_OPT);
 }