[Mesa-dev] [PATCH v2 17/24] anv: Use blorp_ccs_ambiguate instead of fast-clears
Nanley Chery
nanleychery at gmail.com
Wed Jan 31 01:03:11 UTC 2018
On Tue, Jan 30, 2018 at 04:25:59PM -0800, Jason Ekstrand wrote:
> On Tue, Jan 30, 2018 at 2:54 PM, Nanley Chery <nanleychery at gmail.com> wrote:
>
> > On Fri, Jan 19, 2018 at 03:47:34PM -0800, Jason Ekstrand wrote:
> > > Even though the blorp pass looks a bit on the sketchy side, the end
> > > result in the Vulkan driver is very nice. Instead of having this weird
> > > case where you do a fast clear and then maybe have to resolve, we just
> > > do the ambiguate and are done with it. The ambiguate does exactly what
> > > we want of setting all the CCS values to 0 which puts it inot the
> > ^
> > in
Typo.
> > > pass-through state.
> > >
> > > This should also improve performance a bit in certain cases. For
> > > instance, if we did a transition from UNDEFINED to GENERAL for a surface
> > > that doesn't have CCS enabled all the time, we would end up doing a
> > > fast-clear and then a full resolve which ends up touching every byte in
> > > the main surface as well as the CCS. With the ambiguate pass, that
> > > transition only touches the CCS.
> > > ---
> > > src/intel/vulkan/anv_blorp.c | 5 ++++
> > > src/intel/vulkan/genX_cmd_buffer.c | 54 +++++++++---------------------
> > --------
> > > 2 files changed, 17 insertions(+), 42 deletions(-)
> > >
> > > diff --git a/src/intel/vulkan/anv_blorp.c b/src/intel/vulkan/anv_blorp.c
> > > index 05efc6d..3698543 100644
> > > --- a/src/intel/vulkan/anv_blorp.c
> > > +++ b/src/intel/vulkan/anv_blorp.c
> > > @@ -1792,6 +1792,11 @@ anv_image_ccs_op(struct anv_cmd_buffer
> > *cmd_buffer,
> > > surf.surf->format, isl_to_blorp_fast_clear_op(
> > ccs_op));
> > > break;
> > > case ISL_AUX_OP_AMBIGUATE:
> > > + for (uint32_t a = 0; a < layer_count; a++) {
> > > + const uint32_t layer = base_layer + a;
> > > + blorp_ccs_ambiguate(&batch, &surf, level, layer);
> > > + }
> > > + break;
> > > default:
> > > unreachable("Unsupported CCS operation");
> > > }
> > > diff --git a/src/intel/vulkan/genX_cmd_buffer.c
> > b/src/intel/vulkan/genX_cmd_buffer.c
> > > index 77fdadf..9e2eba3 100644
> > > --- a/src/intel/vulkan/genX_cmd_buffer.c
> > > +++ b/src/intel/vulkan/genX_cmd_buffer.c
> > > @@ -486,15 +486,6 @@ init_fast_clear_state_entry(struct anv_cmd_buffer
> > *cmd_buffer,
> > > uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);
> > > enum isl_aux_usage aux_usage = image->planes[plane].aux_usage;
> > >
> > > - /* The resolve flag should updated to signify that
> > fast-clear/compression
> > > - * data needs to be removed when leaving the undefined layout. Such
> > data
> > > - * may need to be removed if it would cause accesses to the color
> > buffer
> > > - * to return incorrect data. The fast clear data in CCS_D buffers
> > should
> > > - * be removed because CCS_D isn't enabled all the time.
> > > - */
> > > - genX(set_image_needs_resolve)(cmd_buffer, image, aspect, level,
> > > - aux_usage == ISL_AUX_USAGE_NONE);
> > > -
> > > /* The fast clear value dword(s) will be copied into a surface state
> > object.
> > > * Ensure that the restrictions of the fields in the dword(s) are
> > followed.
> > > *
> > > @@ -677,10 +668,9 @@ transition_color_buffer(struct anv_cmd_buffer
> > *cmd_buffer,
> > > for (unsigned level = base_level; level < last_level_num; level++)
> > > init_fast_clear_state_entry(cmd_buffer, image, aspect, level);
> > >
> > > - /* Initialize the aux buffers to enable correct rendering. This
> > operation
> > > - * requires up to two steps: one to rid the aux buffer of data
> > that may
> > > - * cause GPU hangs, and another to ensure that writes done
> > without aux
> > > - * will be visible to reads done with aux.
> > > + /* Initialize the aux buffers to enable correct rendering. In
> > order to
> > > + * ensure that things such as storage images work correctly, aux
> > buffers
> > > + * are initialized to the pass-through state.
> >
> > Only CCS is initialized to the pass-through state while MCS is
> > fast-cleared. We may also want to update the comment below since we're
> > no longer fast-clearing CCS.
> >
>
> Right. I've replaced this entire comment with:
>
> /* Initialize the aux buffers to enable correct rendering. In order
> to
> * ensure that things such as storage images work correctly, aux
> buffers
> * need to be initialized to valid data.
> *
> * Having an aux buffer with invalid data is a problem for two
> reasons:
> *
> * 1) Having an invalid value in the buffer can confuse the hardware.
> * For instance, with CCS_E on SKL, a two-bit CCS value of 2 is
> * invalid and leads to the hardware doing strange things. It
> * doesn't hang as far as we can tell but rendering corruption can
> * occur.
> *
> * 2) If this transition is into the GENERAL layout and we then use
> the
> * image as a storage image, then we must have the aux buffer in
> the
> * pass-through state so that, if we then go to texture from the
> * image, we get the results of our storage image writes and not
> the
> * fast clear color or other random data.
This reads better. Is 2) an actual issue? When we create surface states
for storage images, we set the aux_usage to NONE.
- Nanley
> *
> * For CCS both of the problems above are real demonstrable issues.
> In
> * that case, the only thing we can do is to perform an ambiguate to
> * transition the aux surface into the pass-through state.
> *
> * For MCS, (2) is never an issue because we don't support
> multisampled
> * storage images. In theory, issue (1) is a problem with MCS but
> we've
> * never seen it in the wild. For 4x and 16x, all bit patters could,
> in
> * theory, be interpreted as something but we don't know that all bit
> * patterns are actually valid. For 2x and 8x, you could easily end
> up
> * with the MCS referring to an invalid plane because not all bits of
> * the MCS value are actually used. Even though we've never seen
> issues
> * in the wild, it's best to play it safe and initialize the MCS. We
> * can use a fast-clear for MCS because we only ever touch from render
> * and texture (no image load store).
> */
>
> > *
> > > * Having an aux buffer with invalid data is possible for CCS
> > buffers
> > > * SKL+ and for MCS buffers with certain sample counts (2x and
> > 8x). One
> > > @@ -692,16 +682,18 @@ transition_color_buffer(struct anv_cmd_buffer
> > *cmd_buffer,
> > > * We don't have any data to show that this is a problem, but we
> > want to
> > > * avoid causing difficult-to-debug problems.
> > > */
> > > - if (GEN_GEN >= 9 && image->samples == 1) {
> > > + if (image->samples == 1) {
> > > for (uint32_t l = 0; l < level_count; l++) {
> > > const uint32_t level = base_level + l;
> > > const uint32_t level_layer_count =
> > > MIN2(layer_count, anv_image_aux_layers(image, aspect,
> > level));
> > > anv_image_ccs_op(cmd_buffer, image, aspect, level,
> > > base_layer, level_layer_count,
> > > - ISL_AUX_OP_FAST_CLEAR, false);
> > > + ISL_AUX_OP_AMBIGUATE, false);
> > > + genX(set_image_needs_resolve)(cmd_buffer, image,
> > > + aspect, level, false);
> > > }
> > > - } else if (image->samples > 1) {
> > > + } else {
> > > if (image->samples == 4 || image->samples == 16) {
> > > anv_perf_warn(cmd_buffer->device->instance, image,
> > > "Doing a potentially unnecessary fast-clear
> > to "
> > > @@ -712,32 +704,10 @@ transition_color_buffer(struct anv_cmd_buffer
> > *cmd_buffer,
> > > anv_image_mcs_op(cmd_buffer, image, aspect,
> > > base_layer, layer_count,
> > > ISL_AUX_OP_FAST_CLEAR, false);
> > > - }
> > > - /* At this point, some elements of the CCS buffer may have the
> > fast-clear
> > > - * bit-arrangement. As the user writes to a subresource, we need
> > to have
> > > - * the associated CCS elements enter the ambiguated state. This
> > enables
> > > - * reads (implicit or explicit) to reflect the user-written data
> > instead
> > > - * of the clear color. The only time such elements will not
> > change their
> > > - * state as described above, is in a final layout that doesn't
> > have CCS
> > > - * enabled. In this case, we must force the associated CCS
> > buffers of the
> > > - * specified range to enter the ambiguated state in advance.
> > > - */
> > > - if (image->samples == 1 &&
> > > - image->planes[plane].aux_usage != ISL_AUX_USAGE_CCS_E &&
> > > - final_layout != VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL) {
> > > - /* The CCS_D buffer may not be enabled in the final layout.
> > Call this
> > > - * function again with a initial layout of
> > COLOR_ATTACHMENT_OPTIMAL
> > > - * to perform a resolve.
> > > - */
> > > - anv_perf_warn(cmd_buffer->device->instance, image,
> > > - "Performing an additional resolve for CCS_D
> > layout "
> > > - "transition. Consider always leaving it on or "
> > > - "performing an ambiguation pass.");
> > > - transition_color_buffer(cmd_buffer, image, aspect,
> > > - base_level, level_count,
> > > - base_layer, layer_count,
> > > - VK_IMAGE_LAYOUT_COLOR_
> > ATTACHMENT_OPTIMAL,
> > > - final_layout);
> >
> > Nice to be rid of this.
> >
>
> Yup.
>
>
> > > + for (unsigned level = base_level; level < last_level_num;
> > level++) {
> > > + genX(set_image_needs_resolve)(cmd_buffer, image,
> > > + aspect, level, true);
> > > + }
> >
> > Why set needs resolve true here?
> >
>
> There's no particularly good reason. I'll delete it. We never do MCS
> resolves anyway. When the time comes for me to go hook up MCS fast-clears
> again, we can sort it out then.
More information about the mesa-dev
mailing list