[Mesa-dev] [PATCH 01/12] radeonsi: implement SDMA-based buffer clearing for CIK-VI
Marek Olšák
maraeo at gmail.com
Mon Jan 2 22:54:06 UTC 2017
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/drivers/radeon/r600_pipe_common.c | 9 ++++++
src/gallium/drivers/radeon/r600_pipe_common.h | 3 ++
src/gallium/drivers/radeonsi/cik_sdma.c | 42 +++++++++++++++++++++++++++
3 files changed, 54 insertions(+)
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index d45a385..53e0ed6 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -530,20 +530,28 @@ bool r600_check_device_reset(struct r600_common_context *rctx)
return false;
status = rctx->b.get_device_reset_status(&rctx->b);
if (status == PIPE_NO_RESET)
return false;
rctx->device_reset_callback.reset(rctx->device_reset_callback.data, status);
return true;
}
+static void r600_dma_clear_buffer_fallback(struct pipe_context *ctx,
+ struct pipe_resource *dst,
+ uint64_t offset, uint64_t size,
+ unsigned value)
+{
+ ctx->clear_buffer(ctx, dst, offset, size, &value, 4);
+}
+
bool r600_common_context_init(struct r600_common_context *rctx,
struct r600_common_screen *rscreen,
unsigned context_flags)
{
slab_create_child(&rctx->pool_transfers, &rscreen->pool_transfers);
rctx->screen = rscreen;
rctx->ws = rscreen->ws;
rctx->family = rscreen->family;
rctx->chip_class = rscreen->chip_class;
@@ -556,20 +564,21 @@ bool r600_common_context_init(struct r600_common_context *rctx,
rctx->max_db = 4;
rctx->b.invalidate_resource = r600_invalidate_resource;
rctx->b.transfer_map = u_transfer_map_vtbl;
rctx->b.transfer_flush_region = u_transfer_flush_region_vtbl;
rctx->b.transfer_unmap = u_transfer_unmap_vtbl;
rctx->b.texture_subdata = u_default_texture_subdata;
rctx->b.memory_barrier = r600_memory_barrier;
rctx->b.flush = r600_flush_from_st;
rctx->b.set_debug_callback = r600_set_debug_callback;
+ rctx->dma_clear_buffer = r600_dma_clear_buffer_fallback;
/* evergreen_compute.c has a special codepath for global buffers.
* Everything else can use the direct path.
*/
if ((rscreen->chip_class == EVERGREEN || rscreen->chip_class == CAYMAN) &&
(context_flags & PIPE_CONTEXT_COMPUTE_ONLY))
rctx->b.buffer_subdata = u_default_buffer_subdata;
else
rctx->b.buffer_subdata = r600_buffer_subdata;
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index 8ebaed7..faf763c 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -622,20 +622,23 @@ struct r600_common_context {
/* Copy one resource to another using async DMA. */
void (*dma_copy)(struct pipe_context *ctx,
struct pipe_resource *dst,
unsigned dst_level,
unsigned dst_x, unsigned dst_y, unsigned dst_z,
struct pipe_resource *src,
unsigned src_level,
const struct pipe_box *src_box);
+ void (*dma_clear_buffer)(struct pipe_context *ctx, struct pipe_resource *dst,
+ uint64_t offset, uint64_t size, unsigned value);
+
void (*clear_buffer)(struct pipe_context *ctx, struct pipe_resource *dst,
uint64_t offset, uint64_t size, unsigned value,
enum r600_coherency coher);
void (*blit_decompress_depth)(struct pipe_context *ctx,
struct r600_texture *texture,
struct r600_texture *staging,
unsigned first_level, unsigned last_level,
unsigned first_layer, unsigned last_layer,
unsigned first_sample, unsigned last_sample);
diff --git a/src/gallium/drivers/radeonsi/cik_sdma.c b/src/gallium/drivers/radeonsi/cik_sdma.c
index b3eaed5..698f8f6 100644
--- a/src/gallium/drivers/radeonsi/cik_sdma.c
+++ b/src/gallium/drivers/radeonsi/cik_sdma.c
@@ -75,20 +75,61 @@ static void cik_sdma_copy_buffer(struct si_context *ctx,
/* Mark the buffer range of destination as valid (initialized),
* so that transfer_map knows it should wait for the GPU when mapping
* that range. */
util_range_add(&rdst->valid_buffer_range, dst_offset,
dst_offset + size);
cik_sdma_do_copy_buffer(ctx, dst, src, dst_offset, src_offset, size);
r600_dma_emit_wait_idle(&ctx->b);
}
+static void cik_sdma_clear_buffer(struct pipe_context *ctx,
+ struct pipe_resource *dst,
+ uint64_t offset,
+ uint64_t size,
+ unsigned clear_value)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct radeon_winsys_cs *cs = sctx->b.dma.cs;
+ unsigned i, ncopy, csize;
+ struct r600_resource *rdst = r600_resource(dst);
+
+ if (!cs || offset % 4 != 0 || size % 4 != 0) {
+ ctx->clear_buffer(ctx, dst, offset, size, &clear_value, 4);
+ return;
+ }
+
+ /* Mark the buffer range of destination as valid (initialized),
+ * so that transfer_map knows it should wait for the GPU when mapping
+ * that range. */
+ util_range_add(&rdst->valid_buffer_range, offset, offset + size);
+
+ offset += rdst->gpu_address;
+
+ /* the same maximum size as for copying */
+ ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
+ r600_need_dma_space(&sctx->b, ncopy * 5, rdst, NULL);
+
+ for (i = 0; i < ncopy; i++) {
+ csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
+ radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_PACKET_CONSTANT_FILL, 0,
+ 0x8000 /* dword copy */));
+ radeon_emit(cs, offset);
+ radeon_emit(cs, offset >> 32);
+ radeon_emit(cs, clear_value);
+ radeon_emit(cs, csize);
+ offset += csize;
+ size -= csize;
+ }
+ r600_dma_emit_wait_idle(&sctx->b);
+}
+
static unsigned minify_as_blocks(unsigned width, unsigned level, unsigned blk_w)
{
width = u_minify(width, level);
return DIV_ROUND_UP(width, blk_w);
}
static unsigned encode_tile_info(struct si_context *sctx,
struct r600_texture *tex, unsigned level,
bool set_bpp)
{
@@ -518,11 +559,12 @@ static void cik_sdma_copy(struct pipe_context *ctx,
return;
fallback:
si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
src, src_level, src_box);
}
void cik_init_sdma_functions(struct si_context *sctx)
{
sctx->b.dma_copy = cik_sdma_copy;
+ sctx->b.dma_clear_buffer = cik_sdma_clear_buffer;
}
--
2.7.4
More information about the mesa-dev
mailing list