panfrost: Stop using sparse_array for batch BOs

Iterating over a util_sparse_array is very expensive; replace this
with a standard dynarray.

Using the sparse 'nodearray' datastructure instead was tested, but
found to be slower in some cases.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16988>
This commit is contained in:
Icecream95 2022-06-10 23:41:48 +12:00 committed by Marge Bot
parent 169e03800d
commit ec70291da9
3 changed files with 32 additions and 17 deletions

View File

@ -72,9 +72,7 @@ panfrost_batch_init(struct panfrost_context *ctx,
batch->seqnum = ++ctx->batches.seqnum;
batch->first_bo = INT32_MAX;
batch->last_bo = INT32_MIN;
util_sparse_array_init(&batch->bos, sizeof(uint32_t), 64);
util_dynarray_init(&batch->bos, NULL);
batch->minx = batch->miny = ~0;
batch->maxx = batch->maxy = 0;
@ -113,10 +111,11 @@ panfrost_batch_cleanup(struct panfrost_context *ctx, struct panfrost_batch *batc
unsigned batch_idx = panfrost_batch_idx(batch);
for (int i = batch->first_bo; i <= batch->last_bo; i++) {
uint32_t *flags = util_sparse_array_get(&batch->bos, i);
pan_bo_access *flags = util_dynarray_begin(&batch->bos);
unsigned end_bo = util_dynarray_num_elements(&batch->bos, pan_bo_access);
if (!*flags)
for (int i = 0; i < end_bo; ++i) {
if (!flags[i])
continue;
struct panfrost_bo *bo = pan_lookup_bo(dev, i);
@ -142,7 +141,7 @@ panfrost_batch_cleanup(struct panfrost_context *ctx, struct panfrost_batch *batc
util_unreference_framebuffer_state(&batch->key);
util_sparse_array_finish(&batch->bos);
util_dynarray_fini(&batch->bos);
memset(batch, 0, sizeof(*batch));
BITSET_CLEAR(ctx->batches.active, batch_idx);
@ -275,6 +274,21 @@ panfrost_batch_update_access(struct panfrost_batch *batch,
}
}
static pan_bo_access *
panfrost_batch_get_bo_access(struct panfrost_batch *batch, unsigned handle)
{
unsigned size = util_dynarray_num_elements(&batch->bos, pan_bo_access);
if (handle >= size) {
unsigned grow = handle + 1 - size;
memset(util_dynarray_grow(&batch->bos, pan_bo_access, grow),
0, grow * sizeof(pan_bo_access));
}
return util_dynarray_element(&batch->bos, pan_bo_access, handle);
}
static void
panfrost_batch_add_bo_old(struct panfrost_batch *batch,
struct panfrost_bo *bo, uint32_t flags)
@ -282,13 +296,12 @@ panfrost_batch_add_bo_old(struct panfrost_batch *batch,
if (!bo)
return;
uint32_t *entry = util_sparse_array_get(&batch->bos, bo->gem_handle);
uint32_t old_flags = *entry;
pan_bo_access *entry =
panfrost_batch_get_bo_access(batch, bo->gem_handle);
pan_bo_access old_flags = *entry;
if (!old_flags) {
batch->num_bos++;
batch->first_bo = MIN2(batch->first_bo, bo->gem_handle);
batch->last_bo = MAX2(batch->last_bo, bo->gem_handle);
panfrost_bo_reference(bo);
}
@ -590,10 +603,11 @@ panfrost_batch_submit_ioctl(struct panfrost_batch *batch,
sizeof(*bo_handles));
assert(bo_handles);
for (int i = batch->first_bo; i <= batch->last_bo; i++) {
uint32_t *flags = util_sparse_array_get(&batch->bos, i);
pan_bo_access *flags = util_dynarray_begin(&batch->bos);
unsigned end_bo = util_dynarray_num_elements(&batch->bos, pan_bo_access);
if (!*flags)
for (int i = 0; i < end_bo; ++i) {
if (!flags[i])
continue;
assert(submit.bo_handle_count < batch->num_bos);
@ -608,7 +622,7 @@ panfrost_batch_submit_ioctl(struct panfrost_batch *batch,
*/
struct panfrost_bo *bo = pan_lookup_bo(dev, i);
bo->gpu_access |= *flags & (PAN_BO_ACCESS_RW);
bo->gpu_access |= flags[i] & (PAN_BO_ACCESS_RW);
}
panfrost_pool_get_bo_handles(&batch->pool, bo_handles + submit.bo_handle_count);

View File

@ -126,9 +126,8 @@ struct panfrost_batch {
bool scissor_culls_everything;
/* BOs referenced not in the pool */
int first_bo, last_bo;
unsigned num_bos;
struct util_sparse_array bos;
struct util_dynarray bos;
/* Pool owned by this batch (released when the batch is released) used for temporary descriptors */
struct panfrost_pool pool;

View File

@ -68,6 +68,8 @@
/* BO is accessed by the fragment job. */
#define PAN_BO_ACCESS_FRAGMENT (1 << 4)
typedef uint8_t pan_bo_access;
struct panfrost_device;
struct panfrost_ptr {