[Mesa-dev] [PATCH] nvc0: optimize coherent buffer checking at draw time
Samuel Pitoiset
samuel.pitoiset at gmail.com
Wed Dec 9 13:00:55 PST 2015
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 <samuel.pitoiset at gmail.com>
---
I didn't test the patch, but I will run a full piglit tonight.
src/gallium/drivers/nouveau/nv50/nv50_context.h | 1 +
src/gallium/drivers/nouveau/nv50/nv50_state.c | 5 +++
src/gallium/drivers/nouveau/nv50/nv50_vbo.c | 22 ++-----------
src/gallium/drivers/nouveau/nvc0/nvc0_context.h | 3 ++
src/gallium/drivers/nouveau/nvc0/nvc0_state.c | 26 ++++++++++++++++
src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c | 41 +++++--------------------
6 files changed, 45 insertions(+), 53 deletions(-)
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_context.h b/src/gallium/drivers/nouveau/nv50/nv50_context.h
index 2cebcd9..dd5b5e3 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_context.h
+++ b/src/gallium/drivers/nouveau/nv50/nv50_context.h
@@ -134,6 +134,7 @@ 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;
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_state.c b/src/gallium/drivers/nouveau/nv50/nv50_state.c
index fd7c7cd..c08a63e 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_state.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_state.c
@@ -852,8 +852,13 @@ nv50_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
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;
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
index 85878d5..4ca8f11 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
@@ -791,27 +791,9 @@ 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 */
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_context.h b/src/gallium/drivers/nouveau/nvc0/nvc0_context.h
index 39b73ec..1219548 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 5da0ea8..b498b5d 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_state.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_state.c
@@ -549,11 +549,18 @@ nvc0_stage_set_sampler_views(struct nvc0_context *nvc0, int s,
for (i = 0; i < nr; ++i) {
struct nv50_tic_entry *old = nv50_tic_entry(nvc0->textures[s][i]);
+ struct pipe_resource *res = views[i]->texture;
if (views[i] == nvc0->textures[s][i])
continue;
nvc0->textures_dirty[s] |= 1 << i;
+ 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);
+
if (old) {
nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_TEX(s, i));
nvc0_screen_tic_unlock(nvc0->screen, old);
@@ -590,12 +597,19 @@ nvc0_stage_set_sampler_views_range(struct nvc0_context *nvc0, const unsigned s,
if (views) {
for (i = start; i < end; ++i) {
const unsigned p = i - start;
+ struct pipe_resource *res = views[p]->texture;
if (views[p])
last_valid = i;
if (views[p] == nvc0->textures[s][i])
continue;
nvc0->textures_dirty[s] |= 1 << i;
+ 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);
+
if (nvc0->textures[s][i]) {
struct nv50_tic_entry *old = nv50_tic_entry(nvc0->textures[s][i]);
nouveau_bufctx_reset(bctx, bin + i);
@@ -847,9 +861,14 @@ nvc0_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
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->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);
}
}
@@ -1006,6 +1025,7 @@ 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;
}
@@ -1018,9 +1038,15 @@ 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->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 c464904..585f786 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
@@ -861,7 +861,7 @@ 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;
@@ -900,27 +900,9 @@ 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) {
@@ -929,14 +911,12 @@ 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);
@@ -962,12 +942,7 @@ 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)
--
2.6.2
More information about the mesa-dev
mailing list