[Mesa-dev] [PATCH 08/13] gallium/radeon: subclass threaded_resource

Marek Olšák maraeo at gmail.com
Wed May 10 22:45:41 UTC 2017


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

---
 src/gallium/drivers/radeon/r600_buffer_common.c | 16 ++++++++++------
 src/gallium/drivers/radeon/r600_pipe_common.h   |  4 ++--
 src/gallium/drivers/radeon/r600_texture.c       | 20 ++++++++++----------
 3 files changed, 22 insertions(+), 18 deletions(-)

diff --git a/src/gallium/drivers/radeon/r600_buffer_common.c b/src/gallium/drivers/radeon/r600_buffer_common.c
index 5d984ea..b57632e 100644
--- a/src/gallium/drivers/radeon/r600_buffer_common.c
+++ b/src/gallium/drivers/radeon/r600_buffer_common.c
@@ -20,21 +20,20 @@
  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  *
  * Authors:
  *      Marek Olšák
  */
 
 #include "r600_cs.h"
 #include "util/u_memory.h"
 #include "util/u_upload_mgr.h"
-#include "util/u_threaded_context.h"
 #include <inttypes.h>
 #include <stdio.h>
 
 bool r600_rings_is_buffer_referenced(struct r600_common_context *ctx,
 				     struct pb_buffer *buf,
 				     enum radeon_bo_usage usage)
 {
 	if (ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, buf, usage)) {
 		return true;
 	}
@@ -232,41 +231,42 @@ bool r600_alloc_resource(struct r600_common_screen *rscreen,
 			res->buf->size);
 	}
 	return true;
 }
 
 static void r600_buffer_destroy(struct pipe_screen *screen,
 				struct pipe_resource *buf)
 {
 	struct r600_resource *rbuffer = r600_resource(buf);
 
+	threaded_resource_deinit(buf);
 	util_range_destroy(&rbuffer->valid_buffer_range);
 	pb_reference(&rbuffer->buf, NULL);
 	FREE(rbuffer);
 }
 
 static bool
 r600_invalidate_buffer(struct r600_common_context *rctx,
 		       struct r600_resource *rbuffer)
 {
 	/* Shared buffers can't be reallocated. */
-	if (rbuffer->is_shared)
+	if (rbuffer->b.is_shared)
 		return false;
 
 	/* Sparse buffers can't be reallocated. */
 	if (rbuffer->flags & RADEON_FLAG_SPARSE)
 		return false;
 
 	/* In AMD_pinned_memory, the user pointer association only gets
 	 * broken when the buffer is explicitly re-allocated.
 	 */
-	if (rctx->ws->buffer_is_user_ptr(rbuffer->buf))
+	if (rbuffer->b.is_user_ptr)
 		return false;
 
 	/* Check if mapping this buffer would cause waiting for the GPU. */
 	if (r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
 	    !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) {
 		rctx->invalidate_buffer(&rctx->b, &rbuffer->b.b);
 	} else {
 		util_range_set_empty(&rbuffer->valid_buffer_range);
 	}
 
@@ -342,29 +342,29 @@ static void *r600_buffer_transfer_map(struct pipe_context *ctx,
 	 *
 	 *     4) Is glMapBuffer on a shared buffer guaranteed to return the
 	 *        same system address which was specified at creation time?
 	 *
 	 *        RESOLVED: NO. The GL implementation might return a different
 	 *        virtual mapping of that memory, although the same physical
 	 *        page will be used.
 	 *
 	 * So don't ever use staging buffers.
 	 */
-	if (rscreen->ws->buffer_is_user_ptr(rbuffer->buf))
+	if (rbuffer->b.is_user_ptr)
 		usage |= PIPE_TRANSFER_PERSISTENT;
 
 	/* See if the buffer range being mapped has never been initialized,
 	 * in which case it can be mapped unsynchronized. */
 	if (!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
 		       TC_TRANSFER_MAP_IGNORE_VALID_RANGE)) &&
 	    usage & PIPE_TRANSFER_WRITE &&
-	    !rbuffer->is_shared &&
+	    !rbuffer->b.is_shared &&
 	    !util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
 		usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
 	}
 
 	/* If discarding the entire range, discard the whole resource instead. */
 	if (usage & PIPE_TRANSFER_DISCARD_RANGE &&
 	    box->x == 0 && box->width == resource->width0) {
 		usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
 	}
 
@@ -558,25 +558,27 @@ r600_alloc_buffer_struct(struct pipe_screen *screen,
 			 const struct pipe_resource *templ)
 {
 	struct r600_resource *rbuffer;
 
 	rbuffer = MALLOC_STRUCT(r600_resource);
 
 	rbuffer->b.b = *templ;
 	rbuffer->b.b.next = NULL;
 	pipe_reference_init(&rbuffer->b.b.reference, 1);
 	rbuffer->b.b.screen = screen;
+
 	rbuffer->b.vtbl = &r600_buffer_vtbl;
+	threaded_resource_init(&rbuffer->b.b);
+
 	rbuffer->buf = NULL;
 	rbuffer->bind_history = 0;
 	rbuffer->TC_L2_dirty = false;
-	rbuffer->is_shared = false;
 	util_range_init(&rbuffer->valid_buffer_range);
 	return rbuffer;
 }
 
 struct pipe_resource *r600_buffer_create(struct pipe_screen *screen,
 					 const struct pipe_resource *templ,
 					 unsigned alignment)
 {
 	struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 	struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
@@ -620,21 +622,23 @@ struct pipe_resource *
 r600_buffer_from_user_memory(struct pipe_screen *screen,
 			     const struct pipe_resource *templ,
 			     void *user_memory)
 {
 	struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 	struct radeon_winsys *ws = rscreen->ws;
 	struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
 
 	rbuffer->domains = RADEON_DOMAIN_GTT;
 	rbuffer->flags = 0;
+	rbuffer->b.is_user_ptr = true;
 	util_range_add(&rbuffer->valid_buffer_range, 0, templ->width0);
+	util_range_add(&rbuffer->b.valid_buffer_range, 0, templ->width0);
 
 	/* Convert a user pointer to a buffer. */
 	rbuffer->buf = ws->buffer_from_ptr(ws, user_memory, templ->width0);
 	if (!rbuffer->buf) {
 		FREE(rbuffer);
 		return NULL;
 	}
 
 	if (rscreen->info.has_virtual_memory)
 		rbuffer->gpu_address =
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index 51d797a..3449786 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -38,20 +38,21 @@
 
 #include "radeon/radeon_winsys.h"
 
 #include "util/disk_cache.h"
 #include "util/u_blitter.h"
 #include "util/list.h"
 #include "util/u_range.h"
 #include "util/slab.h"
 #include "util/u_suballoc.h"
 #include "util/u_transfer.h"
+#include "util/u_threaded_context.h"
 
 #define ATI_VENDOR_ID 0x1002
 
 #define R600_RESOURCE_FLAG_TRANSFER		(PIPE_RESOURCE_FLAG_DRV_PRIV << 0)
 #define R600_RESOURCE_FLAG_FLUSHED_DEPTH	(PIPE_RESOURCE_FLAG_DRV_PRIV << 1)
 #define R600_RESOURCE_FLAG_FORCE_TILING		(PIPE_RESOURCE_FLAG_DRV_PRIV << 2)
 #define R600_RESOURCE_FLAG_DISABLE_DCC		(PIPE_RESOURCE_FLAG_DRV_PRIV << 3)
 #define R600_RESOURCE_FLAG_UNMAPPABLE		(PIPE_RESOURCE_FLAG_DRV_PRIV << 4)
 
 #define R600_CONTEXT_STREAMOUT_FLUSH		(1u << 0)
@@ -133,21 +134,21 @@ struct r600_perfcounters;
 struct tgsi_shader_info;
 struct r600_qbo_state;
 
 void radeon_shader_binary_init(struct ac_shader_binary *b);
 void radeon_shader_binary_clean(struct ac_shader_binary *b);
 
 /* Only 32-bit buffer allocations are supported, gallium doesn't support more
  * at the moment.
  */
 struct r600_resource {
-	struct u_resource		b;
+	struct threaded_resource	b;
 
 	/* Winsys objects. */
 	struct pb_buffer		*buf;
 	uint64_t			gpu_address;
 	/* Memory usage if the buffer placement is optimal. */
 	uint64_t			vram_usage;
 	uint64_t			gart_usage;
 
 	/* Resource properties. */
 	uint64_t			bo_size;
@@ -172,21 +173,20 @@ struct r600_resource {
 	 * flag and flush the cache before using the buffer.
 	 *
 	 * For example, TC L2 must be flushed if a buffer which has been
 	 * modified by a shader store instruction is about to be used as
 	 * an index buffer. The reason is that VGT DMA index fetching doesn't
 	 * use TC L2.
 	 */
 	bool				TC_L2_dirty;
 
 	/* Whether the resource has been exported via resource_get_handle. */
-	bool				is_shared;
 	unsigned			external_usage; /* PIPE_HANDLE_USAGE_* */
 };
 
 struct r600_transfer {
 	struct pipe_transfer		transfer;
 	struct r600_resource		*staging;
 	unsigned			offset;
 };
 
 struct r600_fmask_info {
diff --git a/src/gallium/drivers/radeon/r600_texture.c b/src/gallium/drivers/radeon/r600_texture.c
index 449d7bc..3bcf966 100644
--- a/src/gallium/drivers/radeon/r600_texture.c
+++ b/src/gallium/drivers/radeon/r600_texture.c
@@ -382,21 +382,21 @@ static void r600_texture_discard_cmask(struct r600_common_screen *rscreen,
 
 	/* Notify all contexts about the change. */
 	p_atomic_inc(&rscreen->dirty_tex_counter);
 	p_atomic_inc(&rscreen->compressed_colortex_counter);
 }
 
 static bool r600_can_disable_dcc(struct r600_texture *rtex)
 {
 	/* We can't disable DCC if it can be written by another process. */
 	return rtex->dcc_offset &&
-	       (!rtex->resource.is_shared ||
+	       (!rtex->resource.b.is_shared ||
 		!(rtex->resource.external_usage & PIPE_HANDLE_USAGE_WRITE));
 }
 
 static bool r600_texture_discard_dcc(struct r600_common_screen *rscreen,
 				     struct r600_texture *rtex)
 {
 	if (!r600_can_disable_dcc(rtex))
 		return false;
 
 	assert(rtex->dcc_separate_buffer == NULL);
@@ -459,21 +459,21 @@ static void r600_degrade_tile_mode_to_linear(struct r600_common_context *rctx,
 	struct r600_texture *new_tex;
 	struct pipe_resource templ = rtex->resource.b.b;
 	unsigned i;
 
 	templ.bind |= PIPE_BIND_LINEAR;
 
 	/* r600g doesn't react to dirty_tex_descriptor_counter */
 	if (rctx->chip_class < SI)
 		return;
 
-	if (rtex->resource.is_shared ||
+	if (rtex->resource.b.is_shared ||
 	    rtex->surface.is_linear)
 		return;
 
 	/* This fails with MSAA, depth, and compressed textures. */
 	if (r600_choose_tiling(rctx->screen, &templ) !=
 	    RADEON_SURF_MODE_LINEAR_ALIGNED)
 		return;
 
 	new_tex = (struct r600_texture*)screen->resource_create(screen, &templ);
 	if (!new_tex)
@@ -563,39 +563,39 @@ static boolean r600_texture_get_handle(struct pipe_screen* screen,
 			r600_eliminate_fast_color_clear(rctx, rtex);
 
 			/* Disable CMASK if flush_resource isn't going
 			 * to be called.
 			 */
 			if (rtex->cmask.size)
 				r600_texture_discard_cmask(rscreen, rtex);
 		}
 
 		/* Set metadata. */
-		if (!res->is_shared || update_metadata) {
+		if (!res->b.is_shared || update_metadata) {
 			r600_texture_init_metadata(rscreen, rtex, &metadata);
 			if (rscreen->query_opaque_metadata)
 				rscreen->query_opaque_metadata(rscreen, rtex,
 							       &metadata);
 
 			rscreen->ws->buffer_set_metadata(res->buf, &metadata);
 		}
 	}
 
-	if (res->is_shared) {
+	if (res->b.is_shared) {
 		/* USAGE_EXPLICIT_FLUSH must be cleared if at least one user
 		 * doesn't set it.
 		 */
 		res->external_usage |= usage & ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH;
 		if (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
 			res->external_usage &= ~PIPE_HANDLE_USAGE_EXPLICIT_FLUSH;
 	} else {
-		res->is_shared = true;
+		res->b.is_shared = true;
 		res->external_usage = usage;
 	}
 
 	if (rscreen->chip_class >= GFX9) {
 		offset = rtex->surface.u.gfx9.surf_offset;
 		stride = rtex->surface.u.gfx9.surf_pitch *
 			 rtex->surface.bpe;
 		slice_size = rtex->surface.u.gfx9.surf_slice_size;
 	} else {
 		offset = rtex->surface.u.legacy.level[0].offset;
@@ -1425,21 +1425,21 @@ static struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen
 	r = r600_init_surface(rscreen, &surface, templ, array_mode, stride,
 			      offset, true, is_scanout, false, false);
 	if (r) {
 		return NULL;
 	}
 
 	rtex = r600_texture_create_object(screen, templ, buf, &surface);
 	if (!rtex)
 		return NULL;
 
-	rtex->resource.is_shared = true;
+	rtex->resource.b.is_shared = true;
 	rtex->resource.external_usage = usage;
 
 	if (rscreen->apply_opaque_metadata)
 		rscreen->apply_opaque_metadata(rscreen, rtex, &metadata);
 
 	/* Validate that addrlib arrived at the same surface parameters. */
 	if (rscreen->chip_class >= GFX9) {
 		assert(metadata.u.gfx9.swizzle_mode == surface.u.gfx9.surf.swizzle_mode);
 	}
 
@@ -1543,21 +1543,21 @@ static void r600_init_temp_resource_from_box(struct pipe_resource *res,
 	}
 }
 
 static bool r600_can_invalidate_texture(struct r600_common_screen *rscreen,
 					struct r600_texture *rtex,
 					unsigned transfer_usage,
 					const struct pipe_box *box)
 {
 	/* r600g doesn't react to dirty_tex_descriptor_counter */
 	return rscreen->chip_class >= SI &&
-		!rtex->resource.is_shared &&
+		!rtex->resource.b.is_shared &&
 		!(transfer_usage & PIPE_TRANSFER_READ) &&
 		rtex->resource.b.b.last_level == 0 &&
 		util_texrange_covers_whole_level(&rtex->resource.b.b, 0,
 						 box->x, box->y, box->z,
 						 box->width, box->height,
 						 box->depth);
 }
 
 static void r600_texture_invalidate_storage(struct r600_common_context *rctx,
 					    struct r600_texture *rtex)
@@ -2250,21 +2250,21 @@ static bool vi_should_enable_separate_dcc(struct r600_texture *tex)
 	return tex->ps_draw_ratio + tex->num_slow_clears >= 5;
 }
 
 /* Called by fast clear. */
 static void vi_separate_dcc_try_enable(struct r600_common_context *rctx,
 				       struct r600_texture *tex)
 {
 	/* The intent is to use this with shared displayable back buffers,
 	 * but it's not strictly limited only to them.
 	 */
-	if (!tex->resource.is_shared ||
+	if (!tex->resource.b.is_shared ||
 	    !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) ||
 	    tex->resource.b.b.target != PIPE_TEXTURE_2D ||
 	    tex->resource.b.b.last_level > 0 ||
 	    !tex->surface.dcc_size)
 		return;
 
 	if (tex->dcc_offset)
 		return; /* already enabled */
 
 	/* Enable the DCC stat gathering. */
@@ -2517,21 +2517,21 @@ void vi_dcc_clear_level(struct r600_common_context *rctx,
 			   clear_value, R600_COHERENCY_CB_META);
 }
 
 /* Set the same micro tile mode as the destination of the last MSAA resolve.
  * This allows hitting the MSAA resolve fast path, which requires that both
  * src and dst micro tile modes match.
  */
 static void si_set_optimal_micro_tile_mode(struct r600_common_screen *rscreen,
 					   struct r600_texture *rtex)
 {
-	if (rtex->resource.is_shared ||
+	if (rtex->resource.b.is_shared ||
 	    rtex->resource.b.b.nr_samples <= 1 ||
 	    rtex->surface.micro_tile_mode == rtex->last_msaa_resolve_target_micro_mode)
 		return;
 
 	assert(rscreen->chip_class >= GFX9 ||
 	       rtex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
 	assert(rtex->resource.b.b.last_level == 0);
 
 	if (rscreen->chip_class >= GFX9) {
 		/* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
@@ -2667,21 +2667,21 @@ void evergreen_do_fast_color_clear(struct r600_common_context *rctx,
 
 		/* only supported on tiled surfaces */
 		if (tex->surface.is_linear) {
 			continue;
 		}
 
 		/* shared textures can't use fast clear without an explicit flush,
 		 * because there is no way to communicate the clear color among
 		 * all clients
 		 */
-		if (tex->resource.is_shared &&
+		if (tex->resource.b.is_shared &&
 		    !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
 			continue;
 
 		/* fast color clear with 1D tiling doesn't work on old kernels and CIK */
 		if (rctx->chip_class == CIK &&
 		    tex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_1D &&
 		    rctx->screen->info.drm_major == 2 &&
 		    rctx->screen->info.drm_minor < 38) {
 			continue;
 		}
-- 
2.7.4



More information about the mesa-dev mailing list