GitLab

This MR depends on and contains !8180 (merged), which is squashed into 1 commit here.

This is a very large improvement in CPU overhead mainly due to the elimination of atomic ops for reference counting. When I saw the results in my CPU profiler, my jaw dropped. 😮

The Gallium interface is changed as follows:

• Many pipe_context functions have their parameter types changed from unsigned to ubyte.
• Changes that make resource unbinding more efficient:
  • set_vertex_buffers gets new parameter unbind_num_trailing_slots, which unbinds slots after start + count while binding buffers at the same time.
  • set_shader_images gets new parameter unbind_num_trailing_slots
  • set_sampler_views gets new parameter unbind_num_trailing_slots
• Changes that totally eliminate refcount atomics from st/mesa and u_threaded_context, improving performance significantly on some CPU architectures:
  • set_constant_buffer gets new parameter take_ownership, meaning the driver shouldn't increment the reference count of the buffer. This is used to pass a buffer from st/mesa through u_threaded_context into the driver without touching the reference count on the way.
  • set_vertex_buffers gets new parameter take_ownership, meaning the driver shouldn't increment the reference counts of the buffers.
  • Added pipe_draw_info::take_index_buffer_ownership, meaning the driver should decrement the reference count of the index buffer in draw_vbo. Only used when u_threaded_context is used.

Additionally, all refcount atomic ops in u_upload_mgr are eliminated when returning buffers to callers. There are also other small improvements in this MR.