[Mesa-dev] [PATCH 08/12] gallium/radeon: move unrelated code from dma_emit_wait_idle to need_dma_space

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

r600_dma_emit_wait_idle is going away in its current form.
The only difference is that the moved code is executed before DMA calls
instead of after them.
---
 src/gallium/drivers/radeon/r600_pipe_common.c | 33 ++++++++++++---------------
 1 file changed, 15 insertions(+), 18 deletions(-)

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 1fe3af0..6b7bbaf 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -220,21 +220,22 @@ void r600_draw_rectangle(struct blitter_context *blitter,
 
 	/* draw */
 	util_draw_vertex_buffer(&rctx->b, NULL, buf, blitter->vb_slot, offset,
 				R600_PRIM_RECTANGLE_LIST, 3, 2);
 	pipe_resource_reference(&buf, NULL);
 }
 
 void r600_need_dma_space(struct r600_common_context *ctx, unsigned num_dw,
                          struct r600_resource *dst, struct r600_resource *src)
 {
-	uint64_t vram = 0, gtt = 0;
+	uint64_t vram = ctx->dma.cs->used_vram;
+	uint64_t gtt = ctx->dma.cs->used_gart;
 
 	if (dst) {
 		vram += dst->vram_usage;
 		gtt += dst->gart_usage;
 	}
 	if (src) {
 		vram += src->vram_usage;
 		gtt += src->gart_usage;
 	}
 
@@ -243,64 +244,60 @@ void r600_need_dma_space(struct r600_common_context *ctx, unsigned num_dw,
 	    ((dst &&
 	      ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, dst->buf,
 					       RADEON_USAGE_READWRITE)) ||
 	     (src &&
 	      ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, src->buf,
 					       RADEON_USAGE_WRITE))))
 		ctx->gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
 
 	/* Flush if there's not enough space, or if the memory usage per IB
 	 * is too large.
+	 *
+	 * IBs using too little memory are limited by the IB submission overhead.
+	 * IBs using too much memory are limited by the kernel/TTM overhead.
+	 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
+	 *
+	 * This heuristic makes sure that DMA requests are executed
+	 * very soon after the call is made and lowers memory usage.
+	 * It improves texture upload performance by keeping the DMA
+	 * engine busy while uploads are being submitted.
 	 */
 	if (!ctx->ws->cs_check_space(ctx->dma.cs, num_dw) ||
+	    ctx->dma.cs->used_vram + ctx->dma.cs->used_gart > 64 * 1024 * 1024 ||
 	    !radeon_cs_memory_below_limit(ctx->screen, ctx->dma.cs, vram, gtt)) {
 		ctx->dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
 		assert((num_dw + ctx->dma.cs->current.cdw) <= ctx->dma.cs->current.max_dw);
 	}
 
 	/* If GPUVM is not supported, the CS checker needs 2 entries
 	 * in the buffer list per packet, which has to be done manually.
 	 */
 	if (ctx->screen->info.has_virtual_memory) {
 		if (dst)
 			radeon_add_to_buffer_list(ctx, &ctx->dma, dst,
 						  RADEON_USAGE_WRITE,
 						  RADEON_PRIO_SDMA_BUFFER);
 		if (src)
 			radeon_add_to_buffer_list(ctx, &ctx->dma, src,
 						  RADEON_USAGE_READ,
 						  RADEON_PRIO_SDMA_BUFFER);
 	}
+
+	/* this function is called before all DMA calls, so increment this. */
+	ctx->num_dma_calls++;
 }
 
 /* This is required to prevent read-after-write hazards. */
 void r600_dma_emit_wait_idle(struct r600_common_context *rctx)
 {
 	struct radeon_winsys_cs *cs = rctx->dma.cs;
 
-	/* done at the end of DMA calls, so increment this. */
-	rctx->num_dma_calls++;
-
-	/* IBs using too little memory are limited by the IB submission overhead.
-	 * IBs using too much memory are limited by the kernel/TTM overhead.
-	 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
-	 *
-	 * This heuristic makes sure that DMA requests are executed
-	 * very soon after the call is made and lowers memory usage.
-	 * It improves texture upload performance by keeping the DMA
-	 * engine busy while uploads are being submitted.
-	 */
-	if (cs->used_vram + cs->used_gart > 64 * 1024 * 1024) {
-		rctx->dma.flush(rctx, RADEON_FLUSH_ASYNC, NULL);
-		return;
-	}
-
 	r600_need_dma_space(rctx, 1, NULL, NULL);
 
 	if (!radeon_emitted(cs, 0)) /* empty queue */
 		return;
 
 	/* NOP waits for idle on Evergreen and later. */
 	if (rctx->chip_class >= CIK)
 		radeon_emit(cs, 0x00000000); /* NOP */
 	else if (rctx->chip_class >= EVERGREEN)
 		radeon_emit(cs, 0xf0000000); /* NOP */
-- 
2.7.4