drm/i915: Micro-optimise hotpath through intel_ring_begin()
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 4 May 2017 13:08:46 +0000 (14:08 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 4 May 2017 14:40:38 +0000 (15:40 +0100)
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.
v3: Make the parameters unsigned as well to make usage.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170504130846.4807-3-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/intel_ringbuffer.c
drivers/gpu/drm/i915/intel_ringbuffer.h

index b308e73fcfaeced264097a4547a423d943df3863..acd1da9b62a3f2d5ed2f4620c130c33c684d6042 100644 (file)
@@ -1656,7 +1656,8 @@ 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,
+                                  unsigned int bytes)
 {
        struct intel_ring *ring = req->ring;
        struct drm_i915_gem_request *target;
@@ -1701,52 +1702,56 @@ static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
        return 0;
 }
 
-u32 *intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
+u32 *intel_ring_begin(struct drm_i915_gem_request *req,
+                     unsigned 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;
+               } else  {
+                       /*
+                        * 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 from the start of ringbuffer.
+                        */
+                       total_bytes = req->reserved_space + remain_actual;
+               }
        }
 
-       if (wait_bytes > ring->space) {
-               int ret = wait_for_space(req, wait_bytes);
+       if (unlikely(total_bytes > ring->space)) {
+               int ret = wait_for_space(req, total_bytes);
                if (unlikely(ret))
                        return ERR_PTR(ret);
        }
 
        if (unlikely(need_wrap)) {
-               GEM_BUG_ON(remain_actual > ring->space);
-               GEM_BUG_ON(ring->emit + remain_actual > ring->size);
+               need_wrap &= ~1;
+               GEM_BUG_ON(need_wrap > ring->space);
+               GEM_BUG_ON(ring->emit + need_wrap > ring->size);
 
                /* Fill the tail with MI_NOOP */
-               memset(ring->vaddr + ring->emit, 0, remain_actual);
+               memset(ring->vaddr + ring->emit, 0, need_wrap);
                ring->emit = 0;
-               ring->space -= remain_actual;
+               ring->space -= need_wrap;
        }
 
        GEM_BUG_ON(ring->emit > ring->size - bytes);
index 3e343b09eeb63895347feedd8294b39b8304d59a..ec16fb6fde625f5015136040a5cf6d16d1fcf5b2 100644 (file)
@@ -497,7 +497,8 @@ void intel_legacy_submission_resume(struct drm_i915_private *dev_priv);
 
 int __must_check intel_ring_cacheline_align(struct drm_i915_gem_request *req);
 
-u32 __must_check *intel_ring_begin(struct drm_i915_gem_request *req, int n);
+u32 __must_check *intel_ring_begin(struct drm_i915_gem_request *req,
+                                  unsigned int n);
 
 static inline void
 intel_ring_advance(struct drm_i915_gem_request *req, u32 *cs)