On 08/02/2017 13:20, Chris Wilson wrote:
On Wed, Feb 08, 2017 at 01:13:48PM +0000, Tvrtko Ursulin wrote:From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> -static int gen9_init_perctx_bb(struct intel_engine_cs *engine, - struct i915_wa_ctx_bb *wa_ctx, - uint32_t *batch, - uint32_t *offset) +static u32 *gen9_init_perctx_bb(struct intel_engine_cs *engine, u32 *batch) { - uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS); - - wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END); + *batch++ = MI_BATCH_BUFFER_END; - return wa_ctx_end(wa_ctx, *offset = index, 1); + return batch; }Transformation looks reasonable, but I'd like to omit this per-ctx bb when empty.
Don't know if that is possible. It is always programming the offset at the moment so documentation digging is required.
+ /* + * Emit the two workaround batch buffers, recording the offset from the + * start of the workaround batch buffer object for each and their + * respective sizes. + */ + for (i = 0; i < ARRAY_SIZE(wa_bb_f); i++) { + wa_bb[i]->offset = ALIGN(batch_ptr - batch, CACHELINE_DWORDS); + batch_ptr = wa_bb_f[i](engine, batch_ptr); + wa_bb[i]->size = batch_ptr - &batch[wa_bb[i]->offset];Which will break this pattern. At the least we could do if (!ww_bb_fn[i]) continue; And then skip loading into the context image if size==0. I'll double check the mechanical aspects in a bit.
This should have been an RFC, just a mistake on my part for not marking it as such. It saves around 500-600 bytes AFAIR, but also I was thinking of adding a helper to emit the often used pipe control (gen8_emit_flush_render) which can bring more gains.
Regards, Tvrtko _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
