From 43f5ce1d20dac94d83d6d6c31b88b4227316877d Mon Sep 17 00:00:00 2001 From: =?utf8?q?Nicolai=20H=C3=A4hnle?= Date: Sun, 13 Mar 2016 11:37:27 -0500 Subject: [PATCH] radeonsi: implement MemoryBarrier (v2) MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit v2: invalidate both constant and VMEM/TC L1 for constant buffers (Marek) Reviewed-by: Marek Olšák --- src/gallium/drivers/radeonsi/si_state.c | 37 +++++++++++++++++++++++++ 1 file changed, 37 insertions(+) diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c index 0c3fbdc9b31..b9bdd47c496 100644 --- a/src/gallium/drivers/radeonsi/si_state.c +++ b/src/gallium/drivers/radeonsi/si_state.c @@ -3521,6 +3521,42 @@ static void si_texture_barrier(struct pipe_context *ctx) SI_CONTEXT_FLUSH_AND_INV_CB; } +static void si_memory_barrier(struct pipe_context *ctx, unsigned flags) +{ + struct si_context *sctx = (struct si_context *)ctx; + + /* Subsequent commands must wait for all shader invocations to + * complete. */ + sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH; + + if (flags & PIPE_BARRIER_CONSTANT_BUFFER) + sctx->b.flags |= SI_CONTEXT_INV_SMEM_L1 | + SI_CONTEXT_INV_VMEM_L1; + + if (flags & (PIPE_BARRIER_VERTEX_BUFFER | + PIPE_BARRIER_SHADER_BUFFER | + PIPE_BARRIER_TEXTURE | + PIPE_BARRIER_IMAGE)) { + /* As far as I can tell, L1 contents are written back to L2 + * automatically at end of shader, but the contents of other + * L1 caches might still be stale. */ + sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1; + } + + if (flags & PIPE_BARRIER_FRAMEBUFFER) + sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER; + + if (flags & (PIPE_BARRIER_MAPPED_BUFFER | + PIPE_BARRIER_FRAMEBUFFER)) { + /* Not sure if INV_GLOBAL_L2 is the best thing here. + * + * We need to make sure that TC L1 & L2 are written back to + * memory, because neither CPU accesses nor CB fetches consider + * TC, but there's no need to invalidate any TC cache lines. */ + sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2; + } +} + static void *si_create_blend_custom(struct si_context *sctx, unsigned mode) { struct pipe_blend_state blend; @@ -3601,6 +3637,7 @@ void si_init_state_functions(struct si_context *sctx) sctx->b.b.set_index_buffer = si_set_index_buffer; sctx->b.b.texture_barrier = si_texture_barrier; + sctx->b.b.memory_barrier = si_memory_barrier; sctx->b.b.set_polygon_stipple = si_set_polygon_stipple; sctx->b.b.set_min_samples = si_set_min_samples; sctx->b.b.set_tess_state = si_set_tess_state; -- 2.30.2