Skip to contentSkip to conversation

drm/i915/bdw: Enable resource streamer on Broadwell

PATCH In review Abdiel Janulgue 3 participants 6 messages May 6, 2014 – Jun 11, 2014

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

From: Abdiel Janulgue <[email protected]>

This is a re-spin of my resource streamer patchset from October adapted to enable the feature on Broadwell instead.

The resource streamer is a hw-feature that helps in reducing commands being submitted by the CPU. Haswell initially has this feature. Unfortunately, HSW seems to have problems switching between non-RS and RS-enabled commands[1]. On BDW however it works seamlessly as expected.

The i-g-t test comes next on a separate patch-series.

--
[1] http://lists.freedesktop.org/archives/intel-gfx/2014-May/044577.html

Abdiel Janulgue (2):
drm/i915/bdw: Enable resource streamer bits on MI_BATCH_BUFFER_START drm/i915/bdw: Expose I915_EXEC_RESOURCE_STREAMER flag

 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    8 ++++++++
 drivers/gpu/drm/i915/i915_reg.h            |    1 +
 drivers/gpu/drm/i915/intel_ringbuffer.c    |    3 ++-
 drivers/gpu/drm/i915/intel_ringbuffer.h    |    1 +
 include/uapi/drm/i915_drm.h                |    7 ++++++-
 5 files changed, 18 insertions(+), 2 deletions(-)
#1 lore raw

PATCH 1/2 drm/i915/bdw: Enable resource streamer bits on MI_BATCH_BUFFER_START

Adds support for executing the resource streamer on BDW

---
 drivers/gpu/drm/i915/i915_reg.h         |    1 +
 drivers/gpu/drm/i915/intel_ringbuffer.c |    3 ++-
 drivers/gpu/drm/i915/intel_ringbuffer.h |    1 +
 3 files changed, 4 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_reg.h +1 -0
index 0eff337..2d5d0ab 100644
@@ -286,6 +286,7 @@
 #define   MI_BATCH_NON_SECURE_HSW	(1<<13)
 #define MI_BATCH_BUFFER_START	MI_INSTR(0x31, 0)
 #define   MI_BATCH_GTT		    (2<<6) /* aliased with (1<<7) on gen4 */
+#define   MI_BATCH_RESOURCE_STREAMER (1<<10)
 #define MI_BATCH_BUFFER_START_GEN8	MI_INSTR(0x31, 1)
 
 
drivers/gpu/drm/i915/intel_ringbuffer.c +2 -1
index 40a7aa4..cc0110c 100644
@@ -1810,7 +1810,8 @@ gen8_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
 		return ret;
 
 	/* FIXME(BDW): Address space and security selectors. */
-	intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8));
+	intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8) |
+			(flags & I915_DISPATCH_RS ? MI_BATCH_RESOURCE_STREAMER : 0));
 	intel_ring_emit(ring, lower_32_bits(offset));
 	intel_ring_emit(ring, upper_32_bits(offset));
 	intel_ring_emit(ring, MI_NOOP);
drivers/gpu/drm/i915/intel_ringbuffer.h +1 -0
index 72c3c15..cf9e259 100644
@@ -116,6 +116,7 @@ struct  intel_ring_buffer {
 					       unsigned flags);
 #define I915_DISPATCH_SECURE 0x1
 #define I915_DISPATCH_PINNED 0x2
+#define I915_DISPATCH_RS     0x4
 	void		(*cleanup)(struct intel_ring_buffer *ring);
 
 	struct {
Signature · 1 line
1.7.9.5
#2 lore raw

PATCH 2/2 drm/i915/bdw: Expose I915_EXEC_RESOURCE_STREAMER flag

In reply to Abdiel Janulgue

Ensures that the batch buffer is executed by the resource streamer

---
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    8 ++++++++
 include/uapi/drm/i915_drm.h                |    7 ++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_gem_execbuffer.c +8 -0
index 47fe8ec..68aaa9b 100644
@@ -1074,6 +1074,14 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	if (args->flags & I915_EXEC_IS_PINNED)
 		flags |= I915_DISPATCH_PINNED;
 
+	if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
+		if ((args->flags & I915_EXEC_RING_MASK) != I915_EXEC_RENDER ||
+		    !IS_BROADWELL(dev))
+			return -EINVAL;
+
+		flags |= I915_DISPATCH_RS;
+	}
+
 	if ((args->flags & I915_EXEC_RING_MASK) > LAST_USER_RING) {
 		DRM_DEBUG("execbuf with unknown ring: %d\n",
 			  (int)(args->flags & I915_EXEC_RING_MASK));
include/uapi/drm/i915_drm.h +6 -1
index 8a3e4ef00..adfc45d 100644
@@ -734,7 +734,12 @@ struct drm_i915_gem_execbuffer2 {
  */
 #define I915_EXEC_HANDLE_LUT		(1<<12)
 
-#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_HANDLE_LUT<<1)
+/** Tell the kernel that the batchbuffer is processed by
+ *  the resource streamer.
+ */
+#define I915_EXEC_RESOURCE_STREAMER     (1<<13)
+
+#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_RESOURCE_STREAMER <<1)
 
 #define I915_EXEC_CONTEXT_ID_MASK	(0xffffffff)
 #define i915_execbuffer2_set_context_id(eb2, context) \
Signature · 1 line
1.7.9.5
#3 lore raw

Re: [PATCH 2/2] drm/i915/bdw: Expose I915_EXEC_RESOURCE_STREAMER flag

Quoted 13 lines · Abdiel Janulgue

Ensures that the batch buffer is executed by the resource streamer

---
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    8 ++++++++
 include/uapi/drm/i915_drm.h                |    7 ++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0 -0
index 47fe8ec..68aaa9b 100644

drivers/gpu/drm/i915/i915_gem_execbuffer.c i915_gem_do_execbuffer(struct drm_device *dev, void *data, · line 1080

Quoted 7 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +4 -0
@@ -1074,6 +1074,14 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	if (args->flags & I915_EXEC_IS_PINNED)
 		flags |= I915_DISPATCH_PINNED;
 
+	if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
+		if ((args->flags & I915_EXEC_RING_MASK) != I915_EXEC_RENDER ||
+		    !IS_BROADWELL(dev))
+			return -EINVAL;

This misses I915_EXEC_DEFAULT. If you reorder this testing to after we decide upon which ring is being run, this should look simpler.

 if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
	if (!IS_BROADWELL(dev)) {
		DRM_DEBUG("resource streamer is only enabled for Broadwell\n");
		return -EINVAL;
	}
	if (ring->id != RCS) {
		DRM_DEBUG("resource streamer is only available on the render ring (not %s)\n", ring->name);
		return -EINVAL;
	}
	flags |= I915_DISPATCH_RS;

}
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#4 lore raw
In reply to Abdiel Janulgue
Quoted 20 lines · On Tue, 6 May 2014 22:25:04 +0300 Abdiel Janulgue…

From: Abdiel Janulgue <[email protected]>

This is a re-spin of my resource streamer patchset from October adapted to enable the feature on Broadwell instead.

The resource streamer is a hw-feature that helps in reducing commands being submitted by the CPU. Haswell initially has this feature. Unfortunately, HSW seems to have problems switching between non-RS and RS-enabled commands[1]. On BDW however it works seamlessly as expected.

The i-g-t test comes next on a separate patch-series.

--
[1] http://lists.freedesktop.org/archives/intel-gfx/2014-May/044577.html

Abdiel Janulgue (2):
drm/i915/bdw: Enable resource streamer bits on MI_BATCH_BUFFER_START drm/i915/bdw: Expose I915_EXEC_RESOURCE_STREAMER flag

Quoted 6 lines
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    8 ++++++++
 drivers/gpu/drm/i915/i915_reg.h            |    1 +
 drivers/gpu/drm/i915/intel_ringbuffer.c    |    3 ++-
 drivers/gpu/drm/i915/intel_ringbuffer.h    |    1 +
 include/uapi/drm/i915_drm.h                |    7 ++++++-
 5 files changed, 18 insertions(+), 2 deletions(-)

These seem trivial enough... have you seen cases where you'd like to enable it in Mesa? If so, it probably makes sense to merge this patch so you can do your tuning and enabling on the Mesa side...

Signature · 1 line
Jesse Barnes, Intel Open Source Technology Center
#5 lore raw
Quoted 29 lines · Jesse Barnes
Quoted 20 lines · On Tue, 6 May 2014 22:25:04 +0300 Abdiel Janulgue…

From: Abdiel Janulgue <[email protected]>

This is a re-spin of my resource streamer patchset from October adapted to enable the feature on Broadwell instead.

The resource streamer is a hw-feature that helps in reducing commands being submitted by the CPU. Haswell initially has this feature. Unfortunately, HSW seems to have problems switching between non-RS and RS-enabled commands[1]. On BDW however it works seamlessly as expected.

The i-g-t test comes next on a separate patch-series.

--
[1] http://lists.freedesktop.org/archives/intel-gfx/2014-May/044577.html

Abdiel Janulgue (2):
drm/i915/bdw: Enable resource streamer bits on MI_BATCH_BUFFER_START drm/i915/bdw: Expose I915_EXEC_RESOURCE_STREAMER flag

Quoted 6 lines
  drivers/gpu/drm/i915/i915_gem_execbuffer.c |    8 ++++++++
  drivers/gpu/drm/i915/i915_reg.h            |    1 +
  drivers/gpu/drm/i915/intel_ringbuffer.c    |    3 ++-
  drivers/gpu/drm/i915/intel_ringbuffer.h    |    1 +
  include/uapi/drm/i915_drm.h                |    7 ++++++-
  5 files changed, 18 insertions(+), 2 deletions(-)
Quoted 3 lines

These seem trivial enough... have you seen cases where you'd like to enable it in Mesa? If so, it probably makes sense to merge this patch so you can do your tuning and enabling on the Mesa side...

I've had some Mesa patches since last year. Unfortunately, the changes didn't give any dramatic performance improvements. On the other hand, I admit: 1. I've only run it against GLBenchmark and Unigine benchmarks. Discussion with Portland folks seem to suggest that we need to have a comprehensive run with more benchmarks. 2. I only did the benchmarks in Haswell. Would be nice to see how it does on Broadwell.

I agree though that further tuning of the performance needs to be done on the Mesa side.