[Mesa-dev] [PATCH v2] nv50, nvc0: optimize coherent buffer checking at draw time
Ilia Mirkin
imirkin at alum.mit.edu
Wed Dec 9 14:48:24 PST 2015
On Wed, Dec 9, 2015 at 5:40 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
> 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.
>
> Changes from v2:
> - forgot to apply some changes for nv50 (texture/vertex bufs)
>
> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> ---
> src/gallium/drivers/nouveau/nv50/nv50_context.h | 3 ++
> src/gallium/drivers/nouveau/nv50/nv50_state.c | 19 +++++++++++
> src/gallium/drivers/nouveau/nv50/nv50_vbo.c | 44 +++----------------------
> 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, 64 insertions(+), 72 deletions(-)
>
> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_context.h b/src/gallium/drivers/nouveau/nv50/nv50_context.h
> index 2cebcd9..712d00e 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 fd7c7cd..b6e5c75 100644
> --- a/src/gallium/drivers/nouveau/nv50/nv50_state.c
> +++ b/src/gallium/drivers/nouveau/nv50/nv50_state.c
> @@ -661,9 +661,16 @@ nv50_stage_set_sampler_views(struct nv50_context *nv50, int s,
> assert(nr <= PIPE_MAX_SAMPLERS);
> for (i = 0; i < nr; ++i) {
> struct nv50_tic_entry *old = nv50_tic_entry(nv50->textures[s][i]);
> + struct pipe_resource *res = views[i]->texture;
I'm moderately sure either views[i] or texture can be null. [Same for nvc0.]
> if (old)
> nv50_screen_tic_unlock(nv50->screen, old);
>
> + 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);
> +
> pipe_sampler_view_reference(&nv50->textures[s][i], views[i]);
> }
>
> @@ -852,8 +859,13 @@ 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);
You also need to clear it out in the user case.
> } else {
> nv50->constbuf_valid[s] &= ~(1 << i);
> + nv50->constbuf_coherent[s] &= ~(1 << i);
> }
> nv50->constbuf_dirty[s] |= 1 << i;
>
> @@ -1000,6 +1012,7 @@ 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;
> }
>
> @@ -1012,9 +1025,15 @@ 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->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT)
I wonder if vb[i].buffer might be null here... Not sure if that's
allowed or not...
> + 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 85878d5..b6ba803 100644
> --- a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
> +++ b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
> @@ -761,8 +761,7 @@ 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;
> @@ -791,27 +790,9 @@ 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 */
> @@ -821,17 +802,7 @@ nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
> nv50->cb_dirty = false;
> }
>
> - 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 (tex_dirty) {
> + if (nv50->textures_coherent) {
nv50->textures_coherent[s]
> BEGIN_NV04(push, NV50_3D(TEX_CACHE_CTL), 1);
> PUSH_DATA (push, 0x20);
> }
> @@ -850,12 +821,7 @@ 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);
> 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;
views[p] might be null (see below where it's being checked for null)
> 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);
Don't forget about the user_buffer case here.
> }
> 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)
The rest looks fine to me, thanks for doing this!
-ilia
More information about the mesa-dev
mailing list