nv50,nvc0: optimize coherent buffer checking at draw time

Instead of iterating over all the buffer resources looking for coherent
buffers, we keep track of a context-wide count. This will save some
iterations (and CPU cycles) in 99.99% case because usually coherent
buffers are not so used.

Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Ilia Mirkin <[email protected]>

3 files changed, 47 insertions, 33 deletions

diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_context.h b/src/gallium/drivers/nouveau/nvc0/nvc0_context.h
index 39b73ecb0c2..12195489691 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_context.h
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_context.h
@@ -134,10 +134,12 @@ struct nvc0_context {
    struct nvc0_constbuf constbuf[6][NVC0_MAX_PIPE_CONSTBUFS];
    uint16_t constbuf_dirty[6];
    uint16_t constbuf_valid[6];
+   uint16_t constbuf_coherent[6];
    bool cb_dirty;
 
    struct pipe_vertex_buffer vtxbuf[PIPE_MAX_ATTRIBS];
    unsigned num_vtxbufs;
+   uint32_t vtxbufs_coherent;
    struct pipe_index_buffer idxbuf;
    uint32_t constant_vbos;
    uint32_t vbo_user; /* bitmask of vertex buffers pointing to user memory */
@@ -149,6 +151,7 @@ struct nvc0_context {
    struct pipe_sampler_view *textures[6][PIPE_MAX_SAMPLERS];
    unsigned num_textures[6];
    uint32_t textures_dirty[6];
+   uint32_t textures_coherent[6];
    struct nv50_tsc_entry *samplers[6][PIPE_MAX_SAMPLERS];
    unsigned num_samplers[6];
    uint16_t samplers_dirty[6];
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_state.c b/src/gallium/drivers/nouveau/nvc0/nvc0_state.c
index 41a824a97a0..24a6c222dd5 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_state.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_state.c
@@ -554,6 +554,17 @@ nvc0_stage_set_sampler_views(struct nvc0_context *nvc0, int s,
          continue;
       nvc0->textures_dirty[s] |= 1 << i;
 
+      if (views[i] && views[i]->texture) {
+         struct pipe_resource *res = views[i]->texture;
+         if (res->target == PIPE_BUFFER &&
+             (res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT))
+            nvc0->textures_coherent[s] |= 1 << i;
+         else
+            nvc0->textures_coherent[s] &= ~(1 << i);
+      } else {
+         nvc0->textures_coherent[s] &= ~(1 << i);
+      }
+
       if (old) {
          nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_TEX(s, i));
          nvc0_screen_tic_unlock(nvc0->screen, old);
@@ -596,6 +607,17 @@ nvc0_stage_set_sampler_views_range(struct nvc0_context *nvc0, const unsigned s,
             continue;
          nvc0->textures_dirty[s] |= 1 << i;
 
+         if (views[p] && views[p]->texture) {
+            struct pipe_resource *res = views[p]->texture;
+            if (res->target == PIPE_BUFFER &&
+                (res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT))
+               nvc0->textures_coherent[s] |= 1 << i;
+            else
+               nvc0->textures_coherent[s] &= ~(1 << i);
+         } else {
+            nvc0->textures_coherent[s] &= ~(1 << i);
+         }
+
          if (nvc0->textures[s][i]) {
             struct nv50_tic_entry *old = nv50_tic_entry(nvc0->textures[s][i]);
             nouveau_bufctx_reset(bctx, bin + i);
@@ -842,14 +864,20 @@ nvc0_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
       nvc0->constbuf[s][i].u.data = cb->user_buffer;
       nvc0->constbuf[s][i].size = MIN2(cb->buffer_size, 0x10000);
       nvc0->constbuf_valid[s] |= 1 << i;
+      nvc0->constbuf_coherent[s] &= ~(1 << i);
    } else
    if (cb) {
       nvc0->constbuf[s][i].offset = cb->buffer_offset;
       nvc0->constbuf[s][i].size = MIN2(align(cb->buffer_size, 0x100), 0x10000);
       nvc0->constbuf_valid[s] |= 1 << i;
+      if (res && res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
+         nvc0->constbuf_coherent[s] |= 1 << i;
+      else
+         nvc0->constbuf_coherent[s] &= ~(1 << i);
    }
    else {
       nvc0->constbuf_valid[s] &= ~(1 << i);
+      nvc0->constbuf_coherent[s] &= ~(1 << i);
    }
 }
 
@@ -1009,6 +1037,7 @@ nvc0_set_vertex_buffers(struct pipe_context *pipe,
     if (!vb) {
        nvc0->vbo_user &= ~(((1ull << count) - 1) << start_slot);
        nvc0->constant_vbos &= ~(((1ull << count) - 1) << start_slot);
+       nvc0->vtxbufs_coherent &= ~(((1ull << count) - 1) << start_slot);
        return;
     }
 
@@ -1021,9 +1050,16 @@ nvc0_set_vertex_buffers(struct pipe_context *pipe,
              nvc0->constant_vbos |= 1 << dst_index;
           else
              nvc0->constant_vbos &= ~(1 << dst_index);
+          nvc0->vtxbufs_coherent &= ~(1 << dst_index);
        } else {
           nvc0->vbo_user &= ~(1 << dst_index);
           nvc0->constant_vbos &= ~(1 << dst_index);
+
+          if (vb[i].buffer &&
+              vb[i].buffer->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
+             nvc0->vtxbufs_coherent |= (1 << dst_index);
+          else
+             nvc0->vtxbufs_coherent &= ~(1 << dst_index);
        }
     }
 }
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
index 235b1afc24b..251753357eb 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
@@ -871,7 +871,7 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
 {
    struct nvc0_context *nvc0 = nvc0_context(pipe);
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
-   int i, s;
+   int s;
 
    /* NOTE: caller must ensure that (min_index + index_bias) is >= 0 */
    nvc0->vb_elt_first = info->min_index + info->index_bias;
@@ -922,27 +922,9 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
 
    push->kick_notify = nvc0_draw_vbo_kick_notify;
 
-   /* TODO: Instead of iterating over all the buffer resources looking for
-    * coherent buffers, keep track of a context-wide count.
-    */
    for (s = 0; s < 5 && !nvc0->cb_dirty; ++s) {
-      uint32_t valid = nvc0->constbuf_valid[s];
-
-      while (valid && !nvc0->cb_dirty) {
-         const unsigned i = ffs(valid) - 1;
-         struct pipe_resource *res;
-
-         valid &= ~(1 << i);
-         if (nvc0->constbuf[s][i].user)
-            continue;
-
-         res = nvc0->constbuf[s][i].u.buf;
-         if (!res)
-            continue;
-
-         if (res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
-            nvc0->cb_dirty = true;
-      }
+      if (nvc0->constbuf_coherent[s])
+         nvc0->cb_dirty = true;
    }
 
    if (nvc0->cb_dirty) {
@@ -951,14 +933,12 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    }
 
    for (s = 0; s < 5; ++s) {
+      if (!nvc0->textures_coherent[s])
+         continue;
+
       for (int i = 0; i < nvc0->num_textures[s]; ++i) {
          struct nv50_tic_entry *tic = nv50_tic_entry(nvc0->textures[s][i]);
-         struct pipe_resource *res;
-         if (!tic)
-            continue;
-         res = nvc0->textures[s][i]->texture;
-         if (res->target != PIPE_BUFFER ||
-             !(res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT))
+         if (!(nvc0->textures_coherent[s] & (1 << i)))
             continue;
 
          BEGIN_NVC0(push, NVC0_3D(TEX_CACHE_CTL), 1);
@@ -984,12 +964,7 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
       PUSH_DATA (push, info->start_instance);
    }
 
-   for (i = 0; i < nvc0->num_vtxbufs && !nvc0->base.vbo_dirty; ++i) {
-      if (!nvc0->vtxbuf[i].buffer)
-         continue;
-      if (nvc0->vtxbuf[i].buffer->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
-         nvc0->base.vbo_dirty = true;
-   }
+   nvc0->base.vbo_dirty |= !!nvc0->vtxbufs_coherent;
 
    if (!nvc0->base.vbo_dirty && nvc0->idxbuf.buffer &&
        nvc0->idxbuf.buffer->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)