Skip to contentSkip to conversation

Nonblocking maps

PATCH New Ben Widawsky 2 participants 8 messages Sep 26, 2011 – Sep 27, 2011

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v3 cover Nonblocking maps

After going back and forth many times, I think Daniel and I have agreed on the solution for non-blocking maps.

This new interface adds a new call to map buffers non-blocking if possible. In actuality it may block, but it will track if the buffer needs flushing or not and does the right thing for the use. This relies on a new IOCTL to determine the cache type of the object.

With the posted benchmark, there are significant improvements on Gen5 with a very synthetic test meant to test unnecessary blocking on architectures. The test is posted in the series for reference, but isn't actually useful in it's current form other than to prove there is a potential performance improvement. The test shows no performance regression on Gen6.

v2:
Some cleanups to libdrm, and mesa patch was messed up. Lot of reworks on mesa (cleanup + usage of nonblocking) Removed the gpu-tools test case since it was just reference

v3:
Mesa cleanups were total bonghits
Added a piglit test

Ben

#1 lore raw

PATCH v2 1/1 drm/i915: IOCTL to query the cache level of a BO.

Querying a BO's cache level can help clients wishing to map the buffer. For example, an Ironlake machine could query the buffer to determine it should use GTT mappings, while a Sandybridge machine running the same code would prefer to use regular CPU mapped buffers since the cache is coherent.

Cc Daniel Vetter <[email protected]>
Cc Eric Anholt <[email protected]>
---
 drivers/gpu/drm/i915/i915_dma.c |    1 +
 drivers/gpu/drm/i915/i915_drv.h |    2 ++
 drivers/gpu/drm/i915/i915_gem.c |   22 ++++++++++++++++++++++
 include/drm/i915_drm.h          |    7 +++++++
 4 files changed, 32 insertions(+), 0 deletions(-)
drivers/gpu/drm/i915/i915_dma.c +1 -0
index 8a3942c..8178cbb 100644
@@ -2292,6 +2292,7 @@ struct drm_ioctl_desc i915_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(I915_GEM_MADVISE, i915_gem_madvise_ioctl, DRM_UNLOCKED),
 	DRM_IOCTL_DEF_DRV(I915_OVERLAY_PUT_IMAGE, intel_overlay_put_image, DRM_MASTER|DRM_CONTROL_ALLOW|DRM_UNLOCKED),
 	DRM_IOCTL_DEF_DRV(I915_OVERLAY_ATTRS, intel_overlay_attrs, DRM_MASTER|DRM_CONTROL_ALLOW|DRM_UNLOCKED),
+	DRM_IOCTL_DEF_DRV(I915_GEM_GET_CACHE_TYPE, i915_gem_get_cache_type_ioctl, DRM_UNLOCKED),
 };
 
 int i915_max_ioctl = DRM_ARRAY_SIZE(i915_ioctls);
drivers/gpu/drm/i915/i915_drv.h +2 -0
index 7916bd9..2dfcb3d 100644
@@ -1118,6 +1118,8 @@ int i915_gem_get_tiling(struct drm_device *dev, void *data,
 			struct drm_file *file_priv);
 int i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
 				struct drm_file *file_priv);
+int i915_gem_get_cache_type_ioctl(struct drm_device *dev, void *data,
+				  struct drm_file *file_priv);
 void i915_gem_load(struct drm_device *dev);
 int i915_gem_init_object(struct drm_gem_object *obj);
 int __must_check i915_gem_flush_ring(struct intel_ring_buffer *ring,
drivers/gpu/drm/i915/i915_gem.c +22 -0
index a546a71..a120a74 100644
@@ -3777,6 +3777,28 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj)
 }
 
 int
+i915_gem_get_cache_type_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *file_priv)
+{
+
+	struct drm_i915_gem_get_cache_type *args = data;
+	struct drm_i915_gem_object *obj;
+	int ret = 0;
+
+	obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
+	if (&obj->base == NULL) {
+		ret = -ENOENT;
+		goto out;
+	}
+
+	args->cache_level = obj->cache_level;
+	drm_gem_object_unreference_unlocked(&obj->base);
+
+out:
+	return ret;
+}
+
+int
 i915_gem_idle(struct drm_device *dev)
 {
 	drm_i915_private_t *dev_priv = dev->dev_private;
include/drm/i915_drm.h +7 -0
index 28c0d11..a0dff8a 100644
@@ -198,6 +198,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_I915_OVERLAY_PUT_IMAGE	0x27
 #define DRM_I915_OVERLAY_ATTRS	0x28
 #define DRM_I915_GEM_EXECBUFFER2	0x29
+#define DRM_I915_GEM_GET_CACHE_TYPE	0x2a
 
 #define DRM_IOCTL_I915_INIT		DRM_IOW( DRM_COMMAND_BASE + DRM_I915_INIT, drm_i915_init_t)
 #define DRM_IOCTL_I915_FLUSH		DRM_IO ( DRM_COMMAND_BASE + DRM_I915_FLUSH)
@@ -239,6 +240,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_IOCTL_I915_GEM_MADVISE	DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_MADVISE, struct drm_i915_gem_madvise)
 #define DRM_IOCTL_I915_OVERLAY_PUT_IMAGE	DRM_IOW(DRM_COMMAND_BASE + DRM_I915_OVERLAY_PUT_IMAGE, struct drm_intel_overlay_put_image)
 #define DRM_IOCTL_I915_OVERLAY_ATTRS	DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_OVERLAY_ATTRS, struct drm_intel_overlay_attrs)
+#define DRM_IOCTL_I915_GEM_GET_CACHE_TYPE	DRM_IOR  (DRM_COMMAND_BASE + DRM_I915_GEM_GET_CACHE_TYPE, struct drm_i915_gem_get_cache_type)
 
 /* Allow drivers to submit batchbuffers directly to hardware, relying
  * on the security mechanisms provided by hardware.
@@ -844,4 +846,9 @@ struct drm_intel_overlay_attrs {
 	__u32 gamma5;
 };
 
+struct drm_i915_gem_get_cache_type {
+	__u32 handle;
+	__u32 cache_level;
+};
+
 #endif				/* _I915_DRM_H_ */
Signature · 1 line
1.7.6.3
#2 lore raw

PATCH 1/1 intel: non-blocking mmaps on the cheap

In reply to Ben Widawsky

From: Daniel Vetter <[email protected]>

This adds a new mode to map gem objects in a non-blocking way.

The new kernel interface required to get the caching level/coherency is not yet wired up. All the code to transparently choose between gtt mappings or (if coherent) cpu mappings is already in place, though.

To make this useable for mesa, we need to allow other buffer acces in between non-blocking modes. It's not a big problem if this accidentaly blocks once when switching back to non-blocking maps, but domains need to be correctly tracked.

I've gone through historical kernel versions and luckily the only place we move a bo from the gtt domain to the cpu domain is in the set_ioctl. pwrite/pread don't move objects into the cpu domain as long as they're untiled or llc cached.

So track whether an object is in the gtt domain and invalidate this only on access to cpu mappings.

[patch reworked by Ben]

---
 include/drm/i915_drm.h   |    7 ++
 intel/intel_bufmgr.h     |    3 +
 intel/intel_bufmgr_gem.c |  175 +++++++++++++++++++++++++++++++++++++++++-----
 3 files changed, 168 insertions(+), 17 deletions(-)
include/drm/i915_drm.h +7 -0
index adc2392..4b62222 100644
@@ -189,6 +189,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_I915_OVERLAY_PUT_IMAGE	0x27
 #define DRM_I915_OVERLAY_ATTRS	0x28
 #define DRM_I915_GEM_EXECBUFFER2	0x29
+#define DRM_I915_GEM_GET_CACHE_TYPE	0x2a
 
 #define DRM_IOCTL_I915_INIT		DRM_IOW( DRM_COMMAND_BASE + DRM_I915_INIT, drm_i915_init_t)
 #define DRM_IOCTL_I915_FLUSH		DRM_IO ( DRM_COMMAND_BASE + DRM_I915_FLUSH)
@@ -230,6 +231,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_IOCTL_I915_GEM_MADVISE	DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_MADVISE, struct drm_i915_gem_madvise)
 #define DRM_IOCTL_I915_OVERLAY_PUT_IMAGE	DRM_IOW(DRM_COMMAND_BASE + DRM_IOCTL_I915_OVERLAY_ATTRS, struct drm_intel_overlay_put_image)
 #define DRM_IOCTL_I915_OVERLAY_ATTRS	DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_OVERLAY_ATTRS, struct drm_intel_overlay_attrs)
+#define DRM_IOCTL_I915_GEM_GET_CACHE_TYPE	DRM_IOR  (DRM_COMMAND_BASE + DRM_I915_GEM_GET_CACHE_TYPE, struct drm_i915_gem_get_cache_type)
 
 /* Allow drivers to submit batchbuffers directly to hardware, relying
  * on the security mechanisms provided by hardware.
@@ -835,4 +837,9 @@ struct drm_intel_overlay_attrs {
 	__u32 gamma5;
 };
 
+struct drm_i915_gem_get_cache_type {
+	__u32 handle;
+	__u32 cache_level;
+};
+
 #endif				/* _I915_DRM_H_ */
intel/intel_bufmgr.h +3 -0
index 889ef46..47f3416 100644
@@ -149,6 +149,9 @@ int drm_intel_gem_bo_map_gtt(drm_intel_bo *bo);
 int drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo);
 void drm_intel_gem_bo_start_gtt_access(drm_intel_bo *bo, int write_enable);
 
+int drm_intel_gem_bo_map_nonblocking(drm_intel_bo *bo);
+int drm_intel_gem_bo_unmap_nonblocking(drm_intel_bo *bo);
+
 int drm_intel_get_pipe_from_crtc_id(drm_intel_bufmgr *bufmgr, int crtc_id);
 
 int drm_intel_get_aperture_sizes(int fd, size_t *mappable, size_t *total);
intel/intel_bufmgr_gem.c +158 -17

252 lines not on this page · raw message

Signature · 1 line
1.7.6.4
#3 lore raw

PATCH 1/1 spec: add on to Yuanhan's arb_map_buffer_range

In reply to Ben Widawsky

Yuanhan recently added tests which check for conformance on the API. This patch goes on top of his to help make a dent in testing functionality.

Cc Yuanhan Liu <[email protected]>
Cc Piglit Devs <[email protected]>
---
 tests/all.tests                                    |    1 +
 tests/spec/arb_map_buffer_range/CMakeLists.gl.txt  |    1 +
 .../arb_map_buffer_range/map_buffer_range_test.c   |  131 ++++++++++++++++++++
 3 files changed, 133 insertions(+), 0 deletions(-)
 create mode 100644 tests/spec/arb_map_buffer_range/map_buffer_range_test.c
tests/all.tests +1 -0
index b24f8e8..667bb34 100644
@@ -1250,6 +1250,7 @@ add_plain_test(arb_instanced_arrays, 'instanced_arrays')
 arb_map_buffer_range = Group()
 spec['ARB_map_buffer_range'] = arb_map_buffer_range
 add_plain_test(arb_map_buffer_range, 'map_buffer_range_error_check')
+add_plain_test(arb_map_buffer_range, 'map_buffer_range_test')
 
 arb_seamless_cube_map = Group()
 spec['ARB_seamless_cube_map'] = arb_seamless_cube_map
tests/spec/arb_map_buffer_range/CMakeLists.gl.txt +1 -0
index a44d0ca..c595d47 100644
@@ -12,5 +12,6 @@ link_libraries (
 )
 
 add_executable (map_buffer_range_error_check map_buffer_range_error_check.c)
+add_executable (map_buffer_range_test map_buffer_range_test.c)
 
 # vim: ft=cmake:
tests/spec/arb_map_buffer_range/map_buffer_range_test.c new file +131 -0

131 lines not on this page · raw message

Signature · 6 lines
1.7.6.4

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw

PATCH v2 1/2 i965: Cleanup MapRangeBuffer

In reply to Ben Widawsky

Clean the code up, and always use a BO when creating a new buffer. I've not seen any regressions but haven't yet tried this on < Gen6.

Cc Chad Versace <[email protected]>
Cc Eric Anholt <[email protected]>
---
 src/mesa/drivers/dri/intel/intel_buffer_objects.c |  114 ++++++++++-----------
 src/mesa/drivers/dri/intel/intel_buffer_objects.h |    4 +-
 2 files changed, 55 insertions(+), 63 deletions(-)
src/mesa/drivers/dri/intel/intel_buffer_objects.c +54 -60
index 4df2d76..78e3468 100644
@@ -291,6 +291,19 @@ intel_bufferobj_map_range(struct gl_context * ctx,
 
    assert(intel_obj);
 
+   /* Catch some errors early to make real logic a bit easier */
+   if ((access & (GL_MAP_FLUSH_EXPLICIT_BIT | GL_MAP_WRITE_BIT)) ==
+      GL_MAP_INVALIDATE_RANGE_BIT)
+	 goto error_out;
+
+   if ((access & (GL_MAP_INVALIDATE_RANGE_BIT | GL_MAP_READ_BIT)) ==
+      (GL_MAP_INVALIDATE_RANGE_BIT|GL_MAP_READ_BIT))
+	 goto error_out;
+
+   if ((access & (GL_MAP_INVALIDATE_BUFFER_BIT | GL_MAP_READ_BIT)) ==
+      (GL_MAP_INVALIDATE_BUFFER_BIT | GL_MAP_READ_BIT))
+	 goto error_out;
+
    /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
     * internally uses our functions directly.
     */
@@ -314,6 +327,9 @@ intel_bufferobj_map_range(struct gl_context * ctx,
       intel_obj->sys_buffer = NULL;
    }
 
+   if (intel_obj->buffer == NULL)
+      goto error_out;
+
    /* If the mapping is synchronized with other GL operations, flush
     * the batchbuffer so that GEM knows about the buffer access for later
     * syncing.
@@ -322,11 +338,6 @@ intel_bufferobj_map_range(struct gl_context * ctx,
        drm_intel_bo_references(intel->batch.bo, intel_obj->buffer))
       intel_flush(ctx);
 
-   if (intel_obj->buffer == NULL) {
-      obj->Pointer = NULL;
-      return NULL;
-   }
-
    /* If the user doesn't care about existing buffer contents and mapping
     * would cause us to block, then throw out the old buffer.
     */
@@ -344,36 +355,30 @@ intel_bufferobj_map_range(struct gl_context * ctx,
     */
    if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
        drm_intel_bo_busy(intel_obj->buffer)) {
-      if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
-	 intel_obj->range_map_buffer = malloc(length);
-	 obj->Pointer = intel_obj->range_map_buffer;
-      } else {
-	 intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
-						      "range map",
-						      length, 64);
-	 if (!(access & GL_MAP_READ_BIT)) {
-	    drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
-	    intel_obj->mapped_gtt = GL_TRUE;
-	 } else {
-	    drm_intel_bo_map(intel_obj->range_map_bo,
-			     (access & GL_MAP_WRITE_BIT) != 0);
-	    intel_obj->mapped_gtt = GL_FALSE;
-	 }
-	 obj->Pointer = intel_obj->range_map_bo->virtual;
-      }
-      return obj->Pointer;
-   }
-
-   if (!(access & GL_MAP_READ_BIT)) {
+      intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
+						   "range map",
+						   length, 64);
+      drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
+      intel_obj->mapped_gtt = true;
+      obj->Pointer = intel_obj->range_map_bo->virtual;
+   } else if (!(access & GL_MAP_READ_BIT)) { /* Write only */
       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
-      intel_obj->mapped_gtt = GL_TRUE;
-   } else {
+      intel_obj->mapped_gtt = true;
+      obj->Pointer = intel_obj->buffer->virtual + offset;
+   } else { /* R/W or RO */
       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
-      intel_obj->mapped_gtt = GL_FALSE;
+      intel_obj->mapped_gtt = false;
+      obj->Pointer = intel_obj->buffer->virtual + offset;
    }
 
-   obj->Pointer = intel_obj->buffer->virtual + offset;
+   if (!(access & GL_MAP_FLUSH_EXPLICIT_BIT))
+      intel_obj->needs_flush_at_unmap = true;
+
    return obj->Pointer;
+
+error_out:
+      obj->Pointer = NULL;
+      return NULL;
 }
 
 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
@@ -388,27 +393,22 @@ intel_bufferobj_flush_mapped_range(struct gl_context *ctx,
 {
    struct intel_context *intel = intel_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
-   drm_intel_bo *temp_bo;
 
-   /* Unless we're in the range map using a temporary system buffer,
-    * there's no work to do.
-    */
-   if (intel_obj->range_map_buffer == NULL)
-      return;
+   assert(intel_obj->needs_flush_at_unmap == false);
 
    if (length == 0)
       return;
 
-   temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
-
-   drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
-
-   intel_emit_linear_blit(intel,
-			  intel_obj->buffer, obj->Offset + offset,
-			  temp_bo, 0,
-			  length);
+   if (!intel_obj->range_map_bo) {
+      intel_batchbuffer_emit_mi_flush(intel);
+   } else {
+      intel_emit_linear_blit(intel,
+			     intel_obj->buffer, obj->Offset + offset,
+			     intel_obj->range_map_bo, 0,
+			     length);
 
-   drm_intel_bo_unreference(temp_bo);
+      drm_intel_bo_unreference(intel_obj->range_map_bo);
+   }
 }
 
 
@@ -425,15 +425,6 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
    assert(obj->Pointer);
    if (intel_obj->sys_buffer != NULL) {
       /* always keep the mapping around. */
-   } else if (intel_obj->range_map_buffer != NULL) {
-      /* Since we've emitted some blits to buffers that will (likely) be used
-       * in rendering operations in other cache domains in this batch, emit a
-       * flush.  Once again, we wish for a domain tracker in libdrm to cover
-       * usage inside of a batchbuffer.
-       */
-      intel_batchbuffer_emit_mi_flush(intel);
-      free(intel_obj->range_map_buffer);
-      intel_obj->range_map_buffer = NULL;
    } else if (intel_obj->range_map_bo != NULL) {
       if (intel_obj->mapped_gtt) {
 	 drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
@@ -441,10 +432,13 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
 	 drm_intel_bo_unmap(intel_obj->range_map_bo);
       }
 
-      intel_emit_linear_blit(intel,
-			     intel_obj->buffer, obj->Offset,
-			     intel_obj->range_map_bo, 0,
-			     obj->Length);
+      if (intel_obj->needs_flush_at_unmap) {
+	 intel_emit_linear_blit(intel,
+				intel_obj->buffer, obj->Offset,
+				intel_obj->range_map_bo, 0,
+				obj->Length);
+	 drm_intel_bo_unreference(intel_obj->range_map_bo);
+      }
 
       /* Since we've emitted some blits to buffers that will (likely) be used
        * in rendering operations in other cache domains in this batch, emit a
@@ -452,8 +446,6 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
        * usage inside of a batchbuffer.
        */
       intel_batchbuffer_emit_mi_flush(intel);
-
-      drm_intel_bo_unreference(intel_obj->range_map_bo);
       intel_obj->range_map_bo = NULL;
    } else if (intel_obj->buffer != NULL) {
       if (intel_obj->mapped_gtt) {
@@ -462,6 +454,8 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
 	 drm_intel_bo_unmap(intel_obj->buffer);
       }
    }
+
+   intel_obj->needs_flush_at_unmap = false;
    obj->Pointer = NULL;
    obj->Offset = 0;
    obj->Length = 0;
src/mesa/drivers/dri/intel/intel_buffer_objects.h +1 -3
index e74d061..24a1636 100644
@@ -47,9 +47,7 @@ struct intel_buffer_object
    void *sys_buffer;
 
    drm_intel_bo *range_map_bo;
-   void *range_map_buffer;
-   unsigned int range_map_offset;
-   GLsizei range_map_size;
+   bool needs_flush_at_unmap;
 
    GLboolean mapped_gtt;
    GLboolean source;
Signature · 1 line
1.7.6.4
#5 lore raw

Re: [PATCH v2 1/2] i965: Cleanup MapRangeBuffer

Quoted 22 lines · Ben Widawsky

Clean the code up, and always use a BO when creating a new buffer. I've not seen any regressions but haven't yet tried this on < Gen6.

Cc Chad Versace <[email protected]>
Cc Eric Anholt <[email protected]>
---
 src/mesa/drivers/dri/intel/intel_buffer_objects.c |  114 ++++++++++-----------
 src/mesa/drivers/dri/intel/intel_buffer_objects.h |    4 +-
 2 files changed, 55 insertions(+), 63 deletions(-)
src/mesa/drivers/dri/intel/intel_buffer_objects.c +3 -0
index 4df2d76..78e3468 100644
@@ -291,6 +291,19 @@ intel_bufferobj_map_range(struct gl_context * ctx,
 
    assert(intel_obj);
 
+   /* Catch some errors early to make real logic a bit easier */
+   if ((access & (GL_MAP_FLUSH_EXPLICIT_BIT | GL_MAP_WRITE_BIT)) ==
+      GL_MAP_INVALIDATE_RANGE_BIT)

src/mesa/drivers/dri/intel/intel_buffer_objects.c intel_bufferobj_map_range(struct gl_context * ctx, · line 306

Quoted 10 lines
src/mesa/drivers/dri/intel/intel_buffer_objects.c +10 -0
+	 goto error_out;
+
+   if ((access & (GL_MAP_INVALIDATE_RANGE_BIT | GL_MAP_READ_BIT)) ==
+      (GL_MAP_INVALIDATE_RANGE_BIT|GL_MAP_READ_BIT))
+	 goto error_out;
+
+   if ((access & (GL_MAP_INVALIDATE_BUFFER_BIT | GL_MAP_READ_BIT)) ==
+      (GL_MAP_INVALIDATE_BUFFER_BIT | GL_MAP_READ_BIT))
+	 goto error_out;
+

Why are you introducing broken error checks for things that are already correctly checked in Mesa core?

Quoted 132 lines
src/mesa/drivers/dri/intel/intel_buffer_objects.c +36 -53
    /* If the user doesn't care about existing buffer contents and mapping
     * would cause us to block, then throw out the old buffer.
     */
@@ -344,36 +355,30 @@ intel_bufferobj_map_range(struct gl_context * ctx,
     */
    if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
        drm_intel_bo_busy(intel_obj->buffer)) {
-      if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
-	 intel_obj->range_map_buffer = malloc(length);
-	 obj->Pointer = intel_obj->range_map_buffer;
-      } else {
-	 intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
-						      "range map",
-						      length, 64);
-	 if (!(access & GL_MAP_READ_BIT)) {
-	    drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
-	    intel_obj->mapped_gtt = GL_TRUE;
-	 } else {
-	    drm_intel_bo_map(intel_obj->range_map_bo,
-			     (access & GL_MAP_WRITE_BIT) != 0);
-	    intel_obj->mapped_gtt = GL_FALSE;
-	 }
-	 obj->Pointer = intel_obj->range_map_bo->virtual;
-      }
-      return obj->Pointer;
-   }
-
-   if (!(access & GL_MAP_READ_BIT)) {
+      intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
+						   "range map",
+						   length, 64);
+      drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
+      intel_obj->mapped_gtt = true;
+      obj->Pointer = intel_obj->range_map_bo->virtual;
+   } else if (!(access & GL_MAP_READ_BIT)) { /* Write only */
       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
-      intel_obj->mapped_gtt = GL_TRUE;
-   } else {
+      intel_obj->mapped_gtt = true;
+      obj->Pointer = intel_obj->buffer->virtual + offset;
+   } else { /* R/W or RO */
       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
-      intel_obj->mapped_gtt = GL_FALSE;
+      intel_obj->mapped_gtt = false;
+      obj->Pointer = intel_obj->buffer->virtual + offset;
    }
 
-   obj->Pointer = intel_obj->buffer->virtual + offset;
+   if (!(access & GL_MAP_FLUSH_EXPLICIT_BIT))
+      intel_obj->needs_flush_at_unmap = true;
+
    return obj->Pointer;
+
+error_out:
+      obj->Pointer = NULL;
+      return NULL;
 }
 
 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
@@ -388,27 +393,22 @@ intel_bufferobj_flush_mapped_range(struct gl_context *ctx,
 {
    struct intel_context *intel = intel_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
-   drm_intel_bo *temp_bo;
 
-   /* Unless we're in the range map using a temporary system buffer,
-    * there's no work to do.
-    */
-   if (intel_obj->range_map_buffer == NULL)
-      return;
+   assert(intel_obj->needs_flush_at_unmap == false);
 
    if (length == 0)
       return;
 
-   temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
-
-   drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
-
-   intel_emit_linear_blit(intel,
-			  intel_obj->buffer, obj->Offset + offset,
-			  temp_bo, 0,
-			  length);
+   if (!intel_obj->range_map_bo) {
+      intel_batchbuffer_emit_mi_flush(intel);
+   } else {
+      intel_emit_linear_blit(intel,
+			     intel_obj->buffer, obj->Offset + offset,
+			     intel_obj->range_map_bo, 0,
+			     length);
 
-   drm_intel_bo_unreference(temp_bo);
+      drm_intel_bo_unreference(intel_obj->range_map_bo);
+   }
 }
 
 
@@ -425,15 +425,6 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
    assert(obj->Pointer);
    if (intel_obj->sys_buffer != NULL) {
       /* always keep the mapping around. */
-   } else if (intel_obj->range_map_buffer != NULL) {
-      /* Since we've emitted some blits to buffers that will (likely) be used
-       * in rendering operations in other cache domains in this batch, emit a
-       * flush.  Once again, we wish for a domain tracker in libdrm to cover
-       * usage inside of a batchbuffer.
-       */
-      intel_batchbuffer_emit_mi_flush(intel);
-      free(intel_obj->range_map_buffer);
-      intel_obj->range_map_buffer = NULL;
    } else if (intel_obj->range_map_bo != NULL) {
       if (intel_obj->mapped_gtt) {
 	 drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
@@ -441,10 +432,13 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
 	 drm_intel_bo_unmap(intel_obj->range_map_bo);
       }
 
-      intel_emit_linear_blit(intel,
-			     intel_obj->buffer, obj->Offset,
-			     intel_obj->range_map_bo, 0,
-			     obj->Length);
+      if (intel_obj->needs_flush_at_unmap) {
+	 intel_emit_linear_blit(intel,
+				intel_obj->buffer, obj->Offset,
+				intel_obj->range_map_bo, 0,
+				obj->Length);
+	 drm_intel_bo_unreference(intel_obj->range_map_bo);
+      }
 
       /* Since we've emitted some blits to buffers that will (likely) be used
        * in rendering operations in other cache domains in this batch, emit a

src/mesa/drivers/dri/intel/intel_buffer_objects.c intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj) · line 448

Quoted 3 lines
src/mesa/drivers/dri/intel/intel_buffer_objects.c +0 -0
@@ -452,8 +446,6 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
        * usage inside of a batchbuffer.
        */
       intel_batchbuffer_emit_mi_flush(intel);

If you have this flush here, you don't need it in the explicit range flushes where you added it.

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#6 lore raw

Re: [PATCH v2 1/2] i965: Cleanup MapRangeBuffer

Quoted 20 lines · Eric Anholt
Quoted 19 lines · Ben Widawsky

Clean the code up, and always use a BO when creating a new buffer. I've not seen any regressions but haven't yet tried this on < Gen6.

Cc Chad Versace <[email protected]>
Cc Eric Anholt <[email protected]>
---
 src/mesa/drivers/dri/intel/intel_buffer_objects.c |  114 ++++++++++-----------
 src/mesa/drivers/dri/intel/intel_buffer_objects.h |    4 +-
 2 files changed, 55 insertions(+), 63 deletions(-)
src/mesa/drivers/dri/intel/intel_buffer_objects.c +0 -0
index 4df2d76..78e3468 100644
@@ -291,6 +291,19 @@ intel_bufferobj_map_range(struct gl_context * ctx,
 
    assert(intel_obj);
 
Quoted 16 lines

src/mesa/drivers/dri/intel/intel_buffer_objects.c intel_bufferobj_map_range(struct gl_context * ctx, · line 306

Quoted 13 lines
src/mesa/drivers/dri/intel/intel_buffer_objects.c +13 -0
+   /* Catch some errors early to make real logic a bit easier */
+   if ((access & (GL_MAP_FLUSH_EXPLICIT_BIT | GL_MAP_WRITE_BIT)) ==
+      GL_MAP_INVALIDATE_RANGE_BIT)
+	 goto error_out;
+
+   if ((access & (GL_MAP_INVALIDATE_RANGE_BIT | GL_MAP_READ_BIT)) ==
+      (GL_MAP_INVALIDATE_RANGE_BIT|GL_MAP_READ_BIT))
+	 goto error_out;
+
+   if ((access & (GL_MAP_INVALIDATE_BUFFER_BIT | GL_MAP_READ_BIT)) ==
+      (GL_MAP_INVALIDATE_BUFFER_BIT | GL_MAP_READ_BIT))
+	 goto error_out;
+

Why are you introducing broken error checks for things that are already correctly checked in Mesa core?

While I agree this shouldn't be there if Mesa core does it, I argue that the checks are broken (they aren't comprehensive). Anyway, I'm convinced it does work correctly in Mesa core because of Yuanhan's tests, but I'll be damned if I can find the code that does it.

Quoted 130 lines
Quoted 130 lines
src/mesa/drivers/dri/intel/intel_buffer_objects.c +36 -53
    /* If the user doesn't care about existing buffer contents and mapping
     * would cause us to block, then throw out the old buffer.
     */
@@ -344,36 +355,30 @@ intel_bufferobj_map_range(struct gl_context * ctx,
     */
    if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
        drm_intel_bo_busy(intel_obj->buffer)) {
-      if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
-	 intel_obj->range_map_buffer = malloc(length);
-	 obj->Pointer = intel_obj->range_map_buffer;
-      } else {
-	 intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
-						      "range map",
-						      length, 64);
-	 if (!(access & GL_MAP_READ_BIT)) {
-	    drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
-	    intel_obj->mapped_gtt = GL_TRUE;
-	 } else {
-	    drm_intel_bo_map(intel_obj->range_map_bo,
-			     (access & GL_MAP_WRITE_BIT) != 0);
-	    intel_obj->mapped_gtt = GL_FALSE;
-	 }
-	 obj->Pointer = intel_obj->range_map_bo->virtual;
-      }
-      return obj->Pointer;
-   }
-
-   if (!(access & GL_MAP_READ_BIT)) {
+      intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
+						   "range map",
+						   length, 64);
+      drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
+      intel_obj->mapped_gtt = true;
+      obj->Pointer = intel_obj->range_map_bo->virtual;
+   } else if (!(access & GL_MAP_READ_BIT)) { /* Write only */
       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
-      intel_obj->mapped_gtt = GL_TRUE;
-   } else {
+      intel_obj->mapped_gtt = true;
+      obj->Pointer = intel_obj->buffer->virtual + offset;
+   } else { /* R/W or RO */
       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
-      intel_obj->mapped_gtt = GL_FALSE;
+      intel_obj->mapped_gtt = false;
+      obj->Pointer = intel_obj->buffer->virtual + offset;
    }
 
-   obj->Pointer = intel_obj->buffer->virtual + offset;
+   if (!(access & GL_MAP_FLUSH_EXPLICIT_BIT))
+      intel_obj->needs_flush_at_unmap = true;
+
    return obj->Pointer;
+
+error_out:
+      obj->Pointer = NULL;
+      return NULL;
 }
 
 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
@@ -388,27 +393,22 @@ intel_bufferobj_flush_mapped_range(struct gl_context *ctx,
 {
    struct intel_context *intel = intel_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
-   drm_intel_bo *temp_bo;
 
-   /* Unless we're in the range map using a temporary system buffer,
-    * there's no work to do.
-    */
-   if (intel_obj->range_map_buffer == NULL)
-      return;
+   assert(intel_obj->needs_flush_at_unmap == false);
 
    if (length == 0)
       return;
 
-   temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
-
-   drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
-
-   intel_emit_linear_blit(intel,
-			  intel_obj->buffer, obj->Offset + offset,
-			  temp_bo, 0,
-			  length);
+   if (!intel_obj->range_map_bo) {
+      intel_batchbuffer_emit_mi_flush(intel);
+   } else {
+      intel_emit_linear_blit(intel,
+			     intel_obj->buffer, obj->Offset + offset,
+			     intel_obj->range_map_bo, 0,
+			     length);
 
-   drm_intel_bo_unreference(temp_bo);
+      drm_intel_bo_unreference(intel_obj->range_map_bo);
+   }
 }
 
 
@@ -425,15 +425,6 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
    assert(obj->Pointer);
    if (intel_obj->sys_buffer != NULL) {
       /* always keep the mapping around. */
-   } else if (intel_obj->range_map_buffer != NULL) {
-      /* Since we've emitted some blits to buffers that will (likely) be used
-       * in rendering operations in other cache domains in this batch, emit a
-       * flush.  Once again, we wish for a domain tracker in libdrm to cover
-       * usage inside of a batchbuffer.
-       */
-      intel_batchbuffer_emit_mi_flush(intel);
-      free(intel_obj->range_map_buffer);
-      intel_obj->range_map_buffer = NULL;
    } else if (intel_obj->range_map_bo != NULL) {
       if (intel_obj->mapped_gtt) {
 	 drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
@@ -441,10 +432,13 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
 	 drm_intel_bo_unmap(intel_obj->range_map_bo);
       }
 
-      intel_emit_linear_blit(intel,
-			     intel_obj->buffer, obj->Offset,
-			     intel_obj->range_map_bo, 0,
-			     obj->Length);
+      if (intel_obj->needs_flush_at_unmap) {
+	 intel_emit_linear_blit(intel,
+				intel_obj->buffer, obj->Offset,
+				intel_obj->range_map_bo, 0,
+				obj->Length);
+	 drm_intel_bo_unreference(intel_obj->range_map_bo);
+      }
 
Quoted 9 lines

src/mesa/drivers/dri/intel/intel_buffer_objects.c intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj) · line 448

Quoted 6 lines
src/mesa/drivers/dri/intel/intel_buffer_objects.c +0 -0
       /* Since we've emitted some blits to buffers that will (likely) be used
        * in rendering operations in other cache domains in this batch, emit a
@@ -452,8 +446,6 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
        * usage inside of a batchbuffer.
        */
       intel_batchbuffer_emit_mi_flush(intel);

If you have this flush here, you don't need it in the explicit range flushes where you added it.

This belonged in the if condition right above, we only need to flush if needs_flush_at_unmap is set.

How do you feel about the patch otherwise?
Ben

#7 lore raw

PATCH v2 2/2 i965: Add calls to nonblocking maps

In reply to Ben Widawsky

When mapping a range of a buffer that has the UNSYNCHRONIZED_BIT, and is only writable, we can take some shortcuts and let people shoot their feet.

---
 src/mesa/drivers/dri/intel/intel_buffer_objects.c |   34 ++++++++++++++------
 src/mesa/drivers/dri/intel/intel_buffer_objects.h |    8 ++++-
 2 files changed, 31 insertions(+), 11 deletions(-)
src/mesa/drivers/dri/intel/intel_buffer_objects.c +24 -10
index 78e3468..13a24b9 100644
@@ -359,15 +359,21 @@ intel_bufferobj_map_range(struct gl_context * ctx,
 						   "range map",
 						   length, 64);
       drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
-      intel_obj->mapped_gtt = true;
+      intel_obj->mapped_type = BO_MAP_TYPE_GTT;
       obj->Pointer = intel_obj->range_map_bo->virtual;
    } else if (!(access & GL_MAP_READ_BIT)) { /* Write only */
-      drm_intel_gem_bo_map_gtt(intel_obj->buffer);
-      intel_obj->mapped_gtt = true;
-      obj->Pointer = intel_obj->buffer->virtual + offset;
+      if (access & GL_MAP_UNSYNCHRONIZED_BIT) {
+	 drm_intel_gem_bo_map_nonblocking(intel_obj->buffer);
+	 intel_obj->mapped_type = BO_MAP_TYPE_NONBLOCKING;
+	 obj->Pointer = intel_obj->buffer->virtual + offset;
+      } else {
+	 drm_intel_gem_bo_map_gtt(intel_obj->buffer);
+	 intel_obj->mapped_type = BO_MAP_TYPE_GTT;
+	 obj->Pointer = intel_obj->buffer->virtual + offset;
+      }
    } else { /* R/W or RO */
       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
-      intel_obj->mapped_gtt = false;
+      intel_obj->mapped_type = BO_MAP_TYPE_CPU;
       obj->Pointer = intel_obj->buffer->virtual + offset;
    }
 
@@ -426,10 +432,12 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
    if (intel_obj->sys_buffer != NULL) {
       /* always keep the mapping around. */
    } else if (intel_obj->range_map_bo != NULL) {
-      if (intel_obj->mapped_gtt) {
+      switch (intel_obj->mapped_type) {
+      case BO_MAP_TYPE_GTT:
 	 drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
-      } else {
-	 drm_intel_bo_unmap(intel_obj->range_map_bo);
+	 break;
+      default:
+	 return GL_FALSE;
       }
 
       if (intel_obj->needs_flush_at_unmap) {
@@ -448,10 +456,16 @@ intel_bufferobj_unmap(struct gl_context * ctx, struct gl_buffer_object *obj)
       intel_batchbuffer_emit_mi_flush(intel);
       intel_obj->range_map_bo = NULL;
    } else if (intel_obj->buffer != NULL) {
-      if (intel_obj->mapped_gtt) {
+      switch (intel_obj->mapped_type) {
+      case BO_MAP_TYPE_GTT:
 	 drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
-      } else {
+	 break;
+      case BO_MAP_TYPE_CPU:
 	 drm_intel_bo_unmap(intel_obj->buffer);
+	 break;
+      case BO_MAP_TYPE_NONBLOCKING:
+	 drm_intel_gem_bo_unmap_nonblocking(intel_obj->buffer);
+	 break;
       }
    }
 
src/mesa/drivers/dri/intel/intel_buffer_objects.h +7 -1
index 24a1636..fd43bd5 100644
@@ -33,6 +33,12 @@
 struct intel_context;
 struct gl_buffer_object;
 
+enum {
+   BO_MAP_TYPE_NONE=0,
+   BO_MAP_TYPE_GTT,
+   BO_MAP_TYPE_CPU,
+   BO_MAP_TYPE_NONBLOCKING
+};
 
 /**
  * Intel vertex/pixel buffer object, derived from Mesa's gl_buffer_object.
@@ -48,8 +54,8 @@ struct intel_buffer_object
 
    drm_intel_bo *range_map_bo;
    bool needs_flush_at_unmap;
+   int mapped_type;
 
-   GLboolean mapped_gtt;
    GLboolean source;
 };
 
Signature · 1 line
1.7.6.4