zink: rework buffer mapping

this collects a bunch of changes which need to be made but which can't
be split up incrementally without breaking things:
* move map offset to zink_transfer::offset for consistent handling
* fix stream uploader map rules to only apply for discard cases so
  the stream uploader doesn't explode
* fix the staging buffer map rules to apply to any non-staging readback
* stop making huge staging buffers for readback
* break out the DONTBLOCK case since this is only coming from qbo readback
* add explicit read-only sync for maps to stall on writes

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12090>
This commit is contained in:
Mike Blumenkrantz 2021-05-17 08:44:02 -04:00 committed by Marge Bot
parent de2da3dadf
commit 6bfbce0447

View file

@ -1158,9 +1158,10 @@ buffer_transfer_map(struct zink_context *ctx, struct zink_resource *res, unsigne
} }
} }
if ((usage & PIPE_MAP_WRITE) && if (usage & PIPE_MAP_DISCARD_RANGE &&
(usage & PIPE_MAP_DISCARD_RANGE || (!(usage & PIPE_MAP_READ) && zink_resource_has_usage(res))) && (!res->obj->host_visible ||
((!res->obj->host_visible) || !(usage & (PIPE_MAP_UNSYNCHRONIZED | PIPE_MAP_PERSISTENT)))) { !(usage & (PIPE_MAP_UNSYNCHRONIZED | PIPE_MAP_PERSISTENT)))) {
/* Check if mapping this buffer would cause waiting for the GPU. /* Check if mapping this buffer would cause waiting for the GPU.
*/ */
@ -1182,29 +1183,43 @@ buffer_transfer_map(struct zink_context *ctx, struct zink_resource *res, unsigne
screen->info.props.limits.minMemoryMapAlignment, &offset, screen->info.props.limits.minMemoryMapAlignment, &offset,
(struct pipe_resource **)&trans->staging_res, (void **)&ptr); (struct pipe_resource **)&trans->staging_res, (void **)&ptr);
res = zink_resource(trans->staging_res); res = zink_resource(trans->staging_res);
trans->offset = offset; trans->offset = offset + box->x;
usage |= PIPE_MAP_UNSYNCHRONIZED;
ptr = ((uint8_t *)ptr) + box->x;
} else { } else {
/* At this point, the buffer is always idle (we checked it above). */ /* At this point, the buffer is always idle (we checked it above). */
usage |= PIPE_MAP_UNSYNCHRONIZED; usage |= PIPE_MAP_UNSYNCHRONIZED;
} }
} else if ((usage & PIPE_MAP_READ) && !(usage & PIPE_MAP_PERSISTENT)) { } else if (((usage & PIPE_MAP_READ) && !(usage & PIPE_MAP_PERSISTENT) && res->base.b.usage != PIPE_USAGE_STAGING) || !res->obj->host_visible) {
assert(!(usage & (TC_TRANSFER_MAP_THREADED_UNSYNC | PIPE_MAP_THREAD_SAFE))); assert(!(usage & (TC_TRANSFER_MAP_THREADED_UNSYNC | PIPE_MAP_THREAD_SAFE)));
if (usage & PIPE_MAP_DONTBLOCK) { if (!res->obj->host_visible || !(usage & PIPE_MAP_ONCE)) {
/* sparse/device-local will always need to wait since it has to copy */ trans->offset = box->x % screen->info.props.limits.minMemoryMapAlignment;
if (!res->obj->host_visible) trans->staging_res = pipe_buffer_create(&screen->base, PIPE_BIND_LINEAR, PIPE_USAGE_STAGING, box->width + trans->offset);
return NULL;
if (!zink_resource_usage_check_completion(screen, res, ZINK_RESOURCE_ACCESS_WRITE))
return NULL;
} else if (!res->obj->host_visible) {
trans->staging_res = pipe_buffer_create(&screen->base, PIPE_BIND_LINEAR, PIPE_USAGE_STAGING, box->x + box->width);
if (!trans->staging_res) if (!trans->staging_res)
return NULL; return NULL;
struct zink_resource *staging_res = zink_resource(trans->staging_res); struct zink_resource *staging_res = zink_resource(trans->staging_res);
zink_copy_buffer(ctx, NULL, staging_res, res, box->x, box->x, box->width); zink_copy_buffer(ctx, NULL, staging_res, res, trans->offset, box->x, box->width);
res = staging_res; res = staging_res;
zink_fence_wait(&ctx->base); usage &= ~PIPE_MAP_UNSYNCHRONIZED;
} else ptr = map_resource(screen, res);
ptr = ((uint8_t *)ptr) + trans->offset;
}
} else if (usage & PIPE_MAP_DONTBLOCK) {
/* sparse/device-local will always need to wait since it has to copy */
if (!res->obj->host_visible)
return NULL;
if (!zink_resource_usage_check_completion(screen, res, ZINK_RESOURCE_ACCESS_WRITE))
return NULL;
usage |= PIPE_MAP_UNSYNCHRONIZED;
}
if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) {
if (usage & PIPE_MAP_WRITE)
zink_resource_usage_wait(ctx, res, ZINK_RESOURCE_ACCESS_RW);
else
zink_resource_usage_wait(ctx, res, ZINK_RESOURCE_ACCESS_WRITE); zink_resource_usage_wait(ctx, res, ZINK_RESOURCE_ACCESS_WRITE);
res->access = 0;
res->access_stage = 0;
} }
if (!ptr) { if (!ptr) {
@ -1217,6 +1232,7 @@ buffer_transfer_map(struct zink_context *ctx, struct zink_resource *res, unsigne
ptr = map_resource(screen, res); ptr = map_resource(screen, res);
if (!ptr) if (!ptr)
return NULL; return NULL;
ptr = ((uint8_t *)ptr) + box->x;
} }
if (!res->obj->coherent if (!res->obj->coherent
@ -1230,7 +1246,7 @@ buffer_transfer_map(struct zink_context *ctx, struct zink_resource *res, unsigne
#endif #endif
) { ) {
VkDeviceSize size = box->width; VkDeviceSize size = box->width;
VkDeviceSize offset = res->obj->offset + trans->offset + box->x; VkDeviceSize offset = res->obj->offset + trans->offset;
VkMappedMemoryRange range = zink_resource_init_mem_range(screen, res->obj, offset, size); VkMappedMemoryRange range = zink_resource_init_mem_range(screen, res->obj, offset, size);
if (vkInvalidateMappedMemoryRanges(screen->dev, 1, &range) != VK_SUCCESS) { if (vkInvalidateMappedMemoryRanges(screen->dev, 1, &range) != VK_SUCCESS) {
vkUnmapMemory(screen->dev, res->obj->mem); vkUnmapMemory(screen->dev, res->obj->mem);
@ -1276,8 +1292,7 @@ zink_transfer_map(struct pipe_context *pctx,
void *ptr, *base; void *ptr, *base;
if (pres->target == PIPE_BUFFER) { if (pres->target == PIPE_BUFFER) {
base = buffer_transfer_map(ctx, res, usage, box, trans); ptr = base = buffer_transfer_map(ctx, res, usage, box, trans);
ptr = ((uint8_t *)base) + box->x;
} else { } else {
if (usage & PIPE_MAP_WRITE && !(usage & PIPE_MAP_READ)) if (usage & PIPE_MAP_WRITE && !(usage & PIPE_MAP_READ))
/* this is like a blit, so we can potentially dump some clears or maybe we have to */ /* this is like a blit, so we can potentially dump some clears or maybe we have to */
@ -1391,7 +1406,7 @@ zink_transfer_flush_region(struct pipe_context *pctx,
ASSERTED VkDeviceSize size, offset; ASSERTED VkDeviceSize size, offset;
if (m->obj->is_buffer) { if (m->obj->is_buffer) {
size = box->width; size = box->width;
offset = trans->offset + box->x; offset = trans->offset;
} else { } else {
size = box->width * box->height * util_format_get_blocksize(m->base.b.format); size = box->width * box->height * util_format_get_blocksize(m->base.b.format);
offset = trans->offset + offset = trans->offset +