[Intel-gfx] [PATCH 3/3] drm/i915: Micro-optimise hotpath through intel_ring_begin()

[Chris Wilson]

On Thu, May 04, 2017 at 03:59:05PM +0300, Mika Kuoppala wrote:
> Chris Wilson <chris at chris-wilson.co.uk> writes:
> 
> > On Thu, May 04, 2017 at 03:11:45PM +0300, Mika Kuoppala wrote:
> >> Chris Wilson <chris at chris-wilson.co.uk> writes:
> >> 
> >> > Typically, there is space available within the ring and if not we have
> >> > to wait (by definition a slow path). Rearrange the code to reduce the
> >> > number of branches and stack size for the hotpath, accomodating a slight
> >> > growth for the wait.
> >> >
> >> > v2: Fix the new assert that packets are not larger than the actual ring.
> >> >
> >> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> >> > ---
> >> >  drivers/gpu/drm/i915/intel_ringbuffer.c | 63 +++++++++++++++++----------------
> >> >  1 file changed, 33 insertions(+), 30 deletions(-)
> >> >
> >> > diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
> >> > index c46e5439d379..53123c1cfcc5 100644
> >> > --- a/drivers/gpu/drm/i915/intel_ringbuffer.c
> >> > +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
> >> > @@ -1654,7 +1654,7 @@ static int ring_request_alloc(struct drm_i915_gem_request *request)
> >> >  	return 0;
> >> >  }
> >> >  
> >> > -static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
> >> > +static noinline int wait_for_space(struct drm_i915_gem_request *req, int bytes)
> >> >  {
> >> >  	struct intel_ring *ring = req->ring;
> >> >  	struct drm_i915_gem_request *target;
> >> > @@ -1702,49 +1702,52 @@ static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
> >> >  u32 *intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
> >> >  {
> >> >  	struct intel_ring *ring = req->ring;
> >> > -	int remain_actual = ring->size - ring->emit;
> >> > -	int remain_usable = ring->effective_size - ring->emit;
> >> > -	int bytes = num_dwords * sizeof(u32);
> >> > -	int total_bytes, wait_bytes;
> >> > -	bool need_wrap = false;
> >> > +	const unsigned int remain_usable = ring->effective_size - ring->emit;
> >> > +	const unsigned int bytes = num_dwords * sizeof(u32);
> >> > +	unsigned int need_wrap = 0;
> >> > +	unsigned int total_bytes;
> >> >  	u32 *cs;
> >> >  
> >> >  	total_bytes = bytes + req->reserved_space;
> >> > +	GEM_BUG_ON(total_bytes > ring->effective_size);
> >> >  
> >> > -	if (unlikely(bytes > remain_usable)) {
> >> > -		/*
> >> > -		 * Not enough space for the basic request. So need to flush
> >> > -		 * out the remainder and then wait for base + reserved.
> >> > -		 */
> >> > -		wait_bytes = remain_actual + total_bytes;
> >> > -		need_wrap = true;
> >> > -	} else if (unlikely(total_bytes > remain_usable)) {
> >> > -		/*
> >> > -		 * The base request will fit but the reserved space
> >> > -		 * falls off the end. So we don't need an immediate wrap
> >> > -		 * and only need to effectively wait for the reserved
> >> > -		 * size space from the start of ringbuffer.
> >> > -		 */
> >> > -		wait_bytes = remain_actual + req->reserved_space;
> >> > -	} else {
> >> > -		/* No wrapping required, just waiting. */
> >> > -		wait_bytes = total_bytes;
> >> > +	if (unlikely(total_bytes > remain_usable)) {
> >> > +		const int remain_actual = ring->size - ring->emit;
> >> > +
> >> > +		if (bytes > remain_usable) {
> >> > +			/*
> >> > +			 * Not enough space for the basic request. So need to
> >> > +			 * flush out the remainder and then wait for
> >> > +			 * base + reserved.
> >> > +			 */
> >> > +			total_bytes += remain_actual;
> >> > +			need_wrap = remain_actual | 1;
> >> 
> >> Your remain_actual should never reach zero. So in here
> >> forcing the lowest bit on, and later off, seems superfluous.
> >
> > Why can't we fill up to the last byte with commands? remain_actual is
> > just (size - tail) and we don't force a wrap until emit crosses the
> > boundary (and not before). We hit remain_actual == 0 in practice.
> > -Chris
> 
> My mistake, was thinking postwrap.
> 
> num_dwords and second parameter to wait_for_space should be unsigned.

You predictive algorithm is working fine though. Applied after your
suggestion from patch 1.

Thanks,
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre