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, 35 insertions, 35 deletions

diff --git a/src/gallium/drivers/nouveau/nv50/nv50_context.h b/src/gallium/drivers/nouveau/nv50/nv50_context.h
index 2cebcd99423..712d00ed2d3 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_context.h
+++ b/src/gallium/drivers/nouveau/nv50/nv50_context.h
@@ -134,9 +134,11 @@ struct nv50_context {
    struct nv50_constbuf constbuf[3][NV50_MAX_PIPE_CONSTBUFS];
    uint16_t constbuf_dirty[3];
    uint16_t constbuf_valid[3];
+   uint16_t constbuf_coherent[3];
 
    struct pipe_vertex_buffer vtxbuf[PIPE_MAX_ATTRIBS];
    unsigned num_vtxbufs;
+   uint32_t vtxbufs_coherent;
    struct pipe_index_buffer idxbuf;
    uint32_t vbo_fifo; /* bitmask of vertex elements to be pushed to FIFO */
    uint32_t vbo_user; /* bitmask of vertex buffers pointing to user memory */
@@ -148,6 +150,7 @@ struct nv50_context {
 
    struct pipe_sampler_view *textures[3][PIPE_MAX_SAMPLERS];
    unsigned num_textures[3];
+   uint32_t textures_coherent[3];
    struct nv50_tsc_entry *samplers[3][PIPE_MAX_SAMPLERS];
    unsigned num_samplers[3];
 
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_state.c b/src/gallium/drivers/nouveau/nv50/nv50_state.c
index de655971b66..cb040439139 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_state.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_state.c
@@ -664,6 +664,17 @@ nv50_stage_set_sampler_views(struct nv50_context *nv50, int s,
       if (old)
          nv50_screen_tic_unlock(nv50->screen, old);
 
+      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))
+            nv50->textures_coherent[s] |= 1 << i;
+         else
+            nv50->textures_coherent[s] &= ~(1 << i);
+      } else {
+         nv50->textures_coherent[s] &= ~(1 << i);
+      }
+
       pipe_sampler_view_reference(&nv50->textures[s][i], views[i]);
    }
 
@@ -847,13 +858,19 @@ nv50_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
       nv50->constbuf[s][i].u.data = cb->user_buffer;
       nv50->constbuf[s][i].size = MIN2(cb->buffer_size, 0x10000);
       nv50->constbuf_valid[s] |= 1 << i;
+      nv50->constbuf_coherent[s] &= ~(1 << i);
    } else
    if (res) {
       nv50->constbuf[s][i].offset = cb->buffer_offset;
       nv50->constbuf[s][i].size = MIN2(align(cb->buffer_size, 0x100), 0x10000);
       nv50->constbuf_valid[s] |= 1 << i;
+      if (res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
+         nv50->constbuf_coherent[s] |= 1 << i;
+      else
+         nv50->constbuf_coherent[s] &= ~(1 << i);
    } else {
       nv50->constbuf_valid[s] &= ~(1 << i);
+      nv50->constbuf_coherent[s] &= ~(1 << i);
    }
    nv50->constbuf_dirty[s] |= 1 << i;
 
@@ -1003,6 +1020,7 @@ nv50_set_vertex_buffers(struct pipe_context *pipe,
    if (!vb) {
       nv50->vbo_user &= ~(((1ull << count) - 1) << start_slot);
       nv50->vbo_constant &= ~(((1ull << count) - 1) << start_slot);
+      nv50->vtxbufs_coherent &= ~(((1ull << count) - 1) << start_slot);
       return;
    }
 
@@ -1015,9 +1033,16 @@ nv50_set_vertex_buffers(struct pipe_context *pipe,
             nv50->vbo_constant |= 1 << dst_index;
          else
             nv50->vbo_constant &= ~(1 << dst_index);
+         nv50->vtxbufs_coherent &= ~(1 << dst_index);
       } else {
          nv50->vbo_user &= ~(1 << dst_index);
          nv50->vbo_constant &= ~(1 << dst_index);
+
+         if (vb[i].buffer &&
+             vb[i].buffer->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
+            nv50->vtxbufs_coherent |= (1 << dst_index);
+         else
+            nv50->vtxbufs_coherent &= ~(1 << dst_index);
       }
    }
 }
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
index 2d1aa6abcd2..60fa2bc06a8 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
@@ -765,7 +765,7 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    struct nv50_context *nv50 = nv50_context(pipe);
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    bool tex_dirty = false;
-   int i, s;
+   int s;
 
    /* NOTE: caller must ensure that (min_index + index_bias) is >= 0 */
    nv50->vb_elt_first = info->min_index + info->index_bias;
@@ -794,27 +794,9 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
 
    push->kick_notify = nv50_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 < 3 && !nv50->cb_dirty; ++s) {
-      uint32_t valid = nv50->constbuf_valid[s];
-
-      while (valid && !nv50->cb_dirty) {
-         const unsigned i = ffs(valid) - 1;
-         struct pipe_resource *res;
-
-         valid &= ~(1 << i);
-         if (nv50->constbuf[s][i].user)
-            continue;
-
-         res = nv50->constbuf[s][i].u.buf;
-         if (!res)
-            continue;
-
-         if (res->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
-            nv50->cb_dirty = true;
-      }
+      if (nv50->constbuf_coherent[s])
+         nv50->cb_dirty = true;
    }
 
    /* If there are any coherent constbufs, flush the cache */
@@ -825,15 +807,10 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
    }
 
    for (s = 0; s < 3 && !tex_dirty; ++s) {
-      for (i = 0; i < nv50->num_textures[s] && !tex_dirty; ++i) {
-         if (!nv50->textures[s][i] ||
-             nv50->textures[s][i]->texture->target != PIPE_BUFFER)
-            continue;
-         if (nv50->textures[s][i]->texture->flags &
-             PIPE_RESOURCE_FLAG_MAP_COHERENT)
-            tex_dirty = true;
-      }
+      if (nv50->textures_coherent[s])
+         tex_dirty = true;
    }
+
    if (tex_dirty) {
       BEGIN_NV04(push, NV50_3D(TEX_CACHE_CTL), 1);
       PUSH_DATA (push, 0x20);
@@ -853,12 +830,7 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
       PUSH_DATA (push, info->start_instance);
    }
 
-   for (i = 0; i < nv50->num_vtxbufs && !nv50->base.vbo_dirty; ++i) {
-      if (!nv50->vtxbuf[i].buffer)
-         continue;
-      if (nv50->vtxbuf[i].buffer->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
-         nv50->base.vbo_dirty = true;
-   }
+   nv50->base.vbo_dirty |= !!nv50->vtxbufs_coherent;
 
    if (nv50->base.vbo_dirty) {
       BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FLUSH), 1);