return (struct nouveau_transfer *)transfer;
}
-static INLINE boolean
+static INLINE bool
nouveau_buffer_malloc(struct nv04_resource *buf)
{
if (!buf->data)
return !!buf->data;
}
-static INLINE boolean
+static INLINE bool
nouveau_buffer_allocate(struct nouveau_screen *screen,
struct nv04_resource *buf, unsigned domain)
{
- uint32_t size = buf->base.width0;
-
- if (buf->base.bind & (PIPE_BIND_CONSTANT_BUFFER |
- PIPE_BIND_COMPUTE_RESOURCE |
- PIPE_BIND_SHADER_RESOURCE))
- size = align(size, 0x100);
+ uint32_t size = align(buf->base.width0, 0x100);
if (domain == NOUVEAU_BO_VRAM) {
buf->mm = nouveau_mm_allocate(screen->mm_VRAM, size,
buf->mm = nouveau_mm_allocate(screen->mm_GART, size,
&buf->bo, &buf->offset);
if (!buf->bo)
- return FALSE;
+ return false;
NOUVEAU_DRV_STAT(screen, buf_obj_current_bytes_sys, buf->base.width0);
} else {
assert(domain == 0);
if (!nouveau_buffer_malloc(buf))
- return FALSE;
+ return false;
}
buf->domain = domain;
if (buf->bo)
buf->address = buf->bo->offset + buf->offset;
- return TRUE;
+ util_range_set_empty(&buf->valid_buffer_range);
+
+ return true;
}
static INLINE void
buf->domain = 0;
}
-static INLINE boolean
+static INLINE bool
nouveau_buffer_reallocate(struct nouveau_screen *screen,
struct nv04_resource *buf, unsigned domain)
{
nouveau_fence_ref(NULL, &res->fence);
nouveau_fence_ref(NULL, &res->fence_wr);
+ util_range_destroy(&res->valid_buffer_range);
+
FREE(res);
NOUVEAU_DRV_STAT(nouveau_screen(pscreen), buf_obj_current_count, -1);
}
+/* Set up a staging area for the transfer. This is either done in "regular"
+ * system memory if the driver supports push_data (nv50+) and the data is
+ * small enough (and permit_pb == true), or in GART memory.
+ */
static uint8_t *
nouveau_transfer_staging(struct nouveau_context *nv,
- struct nouveau_transfer *tx, boolean permit_pb)
+ struct nouveau_transfer *tx, bool permit_pb)
{
const unsigned adj = tx->base.box.x & NOUVEAU_MIN_BUFFER_MAP_ALIGN_MASK;
const unsigned size = align(tx->base.box.width, 4) + adj;
if (!nv->push_data)
- permit_pb = FALSE;
+ permit_pb = false;
if ((size <= NOUVEAU_TRANSFER_PUSHBUF_THRESHOLD) && permit_pb) {
tx->map = align_malloc(size, NOUVEAU_MIN_BUFFER_MAP_ALIGN);
return tx->map;
}
-/* Maybe just migrate to GART right away if we actually need to do this. */
-static boolean
+/* Copies data from the resource into the the transfer's temporary GART
+ * buffer. Also updates buf->data if present.
+ *
+ * Maybe just migrate to GART right away if we actually need to do this. */
+static bool
nouveau_transfer_read(struct nouveau_context *nv, struct nouveau_transfer *tx)
{
struct nv04_resource *buf = nv04_resource(tx->base.resource);
buf->bo, buf->offset + base, buf->domain, size);
if (nouveau_bo_wait(tx->bo, NOUVEAU_BO_RD, nv->client))
- return FALSE;
+ return false;
if (buf->data)
memcpy(buf->data + base, tx->map, size);
- return TRUE;
+ return true;
}
static void
struct nv04_resource *buf = nv04_resource(tx->base.resource);
uint8_t *data = tx->map + offset;
const unsigned base = tx->base.box.x + offset;
- const boolean can_cb = !((base | size) & 3);
+ const bool can_cb = !((base | size) & 3);
if (buf->data)
memcpy(data, buf->data + base, size);
base, size / 4, (const uint32_t *)data);
else
nv->push_data(nv, buf->bo, buf->offset + base, buf->domain, size, data);
-}
+ nouveau_fence_ref(nv->screen->fence.current, &buf->fence);
+ nouveau_fence_ref(nv->screen->fence.current, &buf->fence_wr);
+}
-static INLINE boolean
+/* Does a CPU wait for the buffer's backing data to become reliably accessible
+ * for write/read by waiting on the buffer's relevant fences.
+ */
+static INLINE bool
nouveau_buffer_sync(struct nv04_resource *buf, unsigned rw)
{
if (rw == PIPE_TRANSFER_READ) {
if (!buf->fence_wr)
- return TRUE;
+ return true;
NOUVEAU_DRV_STAT_RES(buf, buf_non_kernel_fence_sync_count,
!nouveau_fence_signalled(buf->fence_wr));
if (!nouveau_fence_wait(buf->fence_wr))
- return FALSE;
+ return false;
} else {
if (!buf->fence)
- return TRUE;
+ return true;
NOUVEAU_DRV_STAT_RES(buf, buf_non_kernel_fence_sync_count,
!nouveau_fence_signalled(buf->fence));
if (!nouveau_fence_wait(buf->fence))
- return FALSE;
+ return false;
nouveau_fence_ref(NULL, &buf->fence);
}
nouveau_fence_ref(NULL, &buf->fence_wr);
- return TRUE;
+ return true;
}
-static INLINE boolean
+static INLINE bool
nouveau_buffer_busy(struct nv04_resource *buf, unsigned rw)
{
if (rw == PIPE_TRANSFER_READ)
}
}
-static boolean
+/* Creates a cache in system memory of the buffer data. */
+static bool
nouveau_buffer_cache(struct nouveau_context *nv, struct nv04_resource *buf)
{
struct nouveau_transfer tx;
- boolean ret;
+ bool ret;
tx.base.resource = &buf->base;
tx.base.box.x = 0;
tx.base.box.width = buf->base.width0;
if (!buf->data)
if (!nouveau_buffer_malloc(buf))
- return FALSE;
+ return false;
if (!(buf->status & NOUVEAU_BUFFER_STATUS_DIRTY))
- return TRUE;
+ return true;
nv->stats.buf_cache_count++;
- if (!nouveau_transfer_staging(nv, &tx, FALSE))
- return FALSE;
+ if (!nouveau_transfer_staging(nv, &tx, false))
+ return false;
ret = nouveau_transfer_read(nv, &tx);
if (ret) {
#define NOUVEAU_TRANSFER_DISCARD \
(PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
-static INLINE boolean
+/* Checks whether it is possible to completely discard the memory backing this
+ * resource. This can be useful if we would otherwise have to wait for a read
+ * operation to complete on this data.
+ */
+static INLINE bool
nouveau_buffer_should_discard(struct nv04_resource *buf, unsigned usage)
{
if (!(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE))
- return FALSE;
+ return false;
if (unlikely(buf->base.bind & PIPE_BIND_SHARED))
- return FALSE;
+ return false;
+ if (unlikely(usage & PIPE_TRANSFER_PERSISTENT))
+ return false;
return buf->mm && nouveau_buffer_busy(buf, PIPE_TRANSFER_WRITE);
}
+/* Returns a pointer to a memory area representing a window into the
+ * resource's data.
+ *
+ * This may or may not be the _actual_ memory area of the resource. However
+ * when calling nouveau_buffer_transfer_unmap, if it wasn't the actual memory
+ * area, the contents of the returned map are copied over to the resource.
+ *
+ * The usage indicates what the caller plans to do with the map:
+ *
+ * WRITE means that the user plans to write to it
+ *
+ * READ means that the user plans on reading from it
+ *
+ * DISCARD_WHOLE_RESOURCE means that the whole resource is going to be
+ * potentially overwritten, and even if it isn't, the bits that aren't don't
+ * need to be maintained.
+ *
+ * DISCARD_RANGE means that all the data in the specified range is going to
+ * be overwritten.
+ *
+ * The strategy for determining what kind of memory area to return is complex,
+ * see comments inside of the function.
+ */
static void *
nouveau_buffer_transfer_map(struct pipe_context *pipe,
struct pipe_resource *resource,
if (usage & PIPE_TRANSFER_WRITE)
NOUVEAU_DRV_STAT(nv->screen, buf_transfers_wr, 1);
+ /* If we are trying to write to an uninitialized range, the user shouldn't
+ * care what was there before. So we can treat the write as if the target
+ * range were being discarded. Furthermore, since we know that even if this
+ * buffer is busy due to GPU activity, because the contents were
+ * uninitialized, the GPU can't care what was there, and so we can treat
+ * the write as being unsynchronized.
+ */
+ if ((usage & PIPE_TRANSFER_WRITE) &&
+ !util_ranges_intersect(&buf->valid_buffer_range, box->x, box->x + box->width))
+ usage |= PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_UNSYNCHRONIZED;
+
+ if (usage & PIPE_TRANSFER_PERSISTENT)
+ usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
+
if (buf->domain == NOUVEAU_BO_VRAM) {
if (usage & NOUVEAU_TRANSFER_DISCARD) {
+ /* Set up a staging area for the user to write to. It will be copied
+ * back into VRAM on unmap. */
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
buf->status &= NOUVEAU_BUFFER_STATUS_REALLOC_MASK;
- nouveau_transfer_staging(nv, tx, TRUE);
+ nouveau_transfer_staging(nv, tx, true);
} else {
if (buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
+ /* The GPU is currently writing to this buffer. Copy its current
+ * contents to a staging area in the GART. This is necessary since
+ * not the whole area being mapped is being discarded.
+ */
if (buf->data) {
align_free(buf->data);
buf->data = NULL;
}
- nouveau_transfer_staging(nv, tx, FALSE);
+ nouveau_transfer_staging(nv, tx, false);
nouveau_transfer_read(nv, tx);
} else {
+ /* The buffer is currently idle. Create a staging area for writes,
+ * and make sure that the cached data is up-to-date. */
if (usage & PIPE_TRANSFER_WRITE)
- nouveau_transfer_staging(nv, tx, TRUE);
+ nouveau_transfer_staging(nv, tx, true);
if (!buf->data)
nouveau_buffer_cache(nv, buf);
}
return buf->data + box->x;
}
+ /* At this point, buf->domain == GART */
+
if (nouveau_buffer_should_discard(buf, usage)) {
int ref = buf->base.reference.count - 1;
nouveau_buffer_reallocate(nv->screen, buf, buf->domain);
nv->invalidate_resource_storage(nv, &buf->base, ref);
}
+ /* Note that nouveau_bo_map ends up doing a nouveau_bo_wait with the
+ * relevant flags. If buf->mm is set, that means this resource is part of a
+ * larger slab bo that holds multiple resources. So in that case, don't
+ * wait on the whole slab and instead use the logic below to return a
+ * reasonable buffer for that case.
+ */
ret = nouveau_bo_map(buf->bo,
buf->mm ? 0 : nouveau_screen_transfer_flags(usage),
nv->client);
if ((usage & PIPE_TRANSFER_UNSYNCHRONIZED) || !buf->mm)
return map;
+ /* If the GPU is currently reading/writing this buffer, we shouldn't
+ * interfere with its progress. So instead we either wait for the GPU to
+ * complete its operation, or set up a staging area to perform our work in.
+ */
if (nouveau_buffer_busy(buf, usage & PIPE_TRANSFER_READ_WRITE)) {
if (unlikely(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)) {
/* Discarding was not possible, must sync because
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else
if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
- nouveau_transfer_staging(nv, tx, TRUE);
+ /* The whole range is being discarded, so it doesn't matter what was
+ * there before. No need to copy anything over. */
+ nouveau_transfer_staging(nv, tx, true);
map = tx->map;
} else
if (nouveau_buffer_busy(buf, PIPE_TRANSFER_READ)) {
else
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else {
- nouveau_transfer_staging(nv, tx, TRUE);
+ /* It is expected that the returned buffer be a representation of the
+ * data in question, so we must copy it over from the buffer. */
+ nouveau_transfer_staging(nv, tx, true);
if (tx->map)
memcpy(tx->map, map, box->width);
map = tx->map;
const struct pipe_box *box)
{
struct nouveau_transfer *tx = nouveau_transfer(transfer);
+ struct nv04_resource *buf = nv04_resource(transfer->resource);
+
if (tx->map)
nouveau_transfer_write(nouveau_context(pipe), tx, box->x, box->width);
+
+ util_range_add(&buf->valid_buffer_range,
+ tx->base.box.x + box->x,
+ tx->base.box.x + box->x + box->width);
}
+/* Unmap stage of the transfer. If it was a WRITE transfer and the map that
+ * was returned was not the real resource's data, this needs to transfer the
+ * data back to the resource.
+ *
+ * Also marks vbo dirty based on the buffer's binding
+ */
static void
nouveau_buffer_transfer_unmap(struct pipe_context *pipe,
struct pipe_transfer *transfer)
const uint8_t bind = buf->base.bind;
/* make sure we invalidate dedicated caches */
if (bind & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
- nv->vbo_dirty = TRUE;
- if (bind & (PIPE_BIND_CONSTANT_BUFFER))
- nv->cb_dirty = TRUE;
+ nv->vbo_dirty = true;
}
+
+ util_range_add(&buf->valid_buffer_range,
+ tx->base.box.x, tx->base.box.x + tx->base.box.width);
}
if (!tx->bo && (tx->base.usage & PIPE_TRANSFER_WRITE))
&dst->base, 0, dstx, 0, 0,
&src->base, 0, &src_box);
}
+
+ util_range_add(&dst->valid_buffer_range, dstx, dstx + size);
}
{
struct nouveau_screen *screen = nouveau_screen(pscreen);
struct nv04_resource *buffer;
- boolean ret;
+ bool ret;
buffer = CALLOC_STRUCT(nv04_resource);
if (!buffer)
pipe_reference_init(&buffer->base.reference, 1);
buffer->base.screen = pscreen;
- if (buffer->base.bind &
- (screen->vidmem_bindings & screen->sysmem_bindings)) {
+ if (buffer->base.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT |
+ PIPE_RESOURCE_FLAG_MAP_COHERENT)) {
+ buffer->domain = NOUVEAU_BO_GART;
+ } else if (buffer->base.bind &
+ (screen->vidmem_bindings & screen->sysmem_bindings)) {
switch (buffer->base.usage) {
case PIPE_USAGE_DEFAULT:
case PIPE_USAGE_IMMUTABLE:
- case PIPE_USAGE_STATIC:
- buffer->domain = NOUVEAU_BO_VRAM;
+ buffer->domain = NV_VRAM_DOMAIN(screen);
break;
case PIPE_USAGE_DYNAMIC:
/* For most apps, we'd have to do staging transfers to avoid sync
* with this usage, and GART -> GART copies would be suboptimal.
*/
- buffer->domain = NOUVEAU_BO_VRAM;
+ buffer->domain = NV_VRAM_DOMAIN(screen);
break;
case PIPE_USAGE_STAGING:
case PIPE_USAGE_STREAM:
}
} else {
if (buffer->base.bind & screen->vidmem_bindings)
- buffer->domain = NOUVEAU_BO_VRAM;
+ buffer->domain = NV_VRAM_DOMAIN(screen);
else
if (buffer->base.bind & screen->sysmem_bindings)
buffer->domain = NOUVEAU_BO_GART;
}
ret = nouveau_buffer_allocate(screen, buffer, buffer->domain);
- if (ret == FALSE)
+ if (ret == false)
goto fail;
if (buffer->domain == NOUVEAU_BO_VRAM && screen->hint_buf_keep_sysmem_copy)
NOUVEAU_DRV_STAT(screen, buf_obj_current_count, 1);
+ util_range_init(&buffer->valid_buffer_range);
+
return &buffer->base;
fail:
buffer->data = ptr;
buffer->status = NOUVEAU_BUFFER_STATUS_USER_MEMORY;
+ util_range_init(&buffer->valid_buffer_range);
+ util_range_add(&buffer->valid_buffer_range, 0, bytes);
+
return &buffer->base;
}
-static INLINE boolean
+static INLINE bool
nouveau_buffer_data_fetch(struct nouveau_context *nv, struct nv04_resource *buf,
struct nouveau_bo *bo, unsigned offset, unsigned size)
{
if (!nouveau_buffer_malloc(buf))
- return FALSE;
+ return false;
if (nouveau_bo_map(bo, NOUVEAU_BO_RD, nv->client))
- return FALSE;
+ return false;
memcpy(buf->data, (uint8_t *)bo->map + offset, size);
- return TRUE;
+ return true;
}
/* Migrate a linear buffer (vertex, index, constants) USER -> GART -> VRAM. */
-boolean
+bool
nouveau_buffer_migrate(struct nouveau_context *nv,
struct nv04_resource *buf, const unsigned new_domain)
{
if (new_domain == NOUVEAU_BO_GART && old_domain == 0) {
if (!nouveau_buffer_allocate(screen, buf, new_domain))
- return FALSE;
+ return false;
ret = nouveau_bo_map(buf->bo, 0, nv->client);
if (ret)
return ret;
if (new_domain == NOUVEAU_BO_VRAM) {
/* keep a system memory copy of our data in case we hit a fallback */
if (!nouveau_buffer_data_fetch(nv, buf, buf->bo, buf->offset, size))
- return FALSE;
+ return false;
if (nouveau_mesa_debug)
debug_printf("migrating %u KiB to VRAM\n", size / 1024);
}
if (new_domain == NOUVEAU_BO_VRAM && old_domain == 0) {
struct nouveau_transfer tx;
if (!nouveau_buffer_allocate(screen, buf, NOUVEAU_BO_VRAM))
- return FALSE;
+ return false;
tx.base.resource = &buf->base;
tx.base.box.x = 0;
tx.base.box.width = buf->base.width0;
tx.bo = NULL;
tx.map = NULL;
- if (!nouveau_transfer_staging(nv, &tx, FALSE))
- return FALSE;
+ if (!nouveau_transfer_staging(nv, &tx, false))
+ return false;
nouveau_transfer_write(nv, &tx, 0, tx.base.box.width);
nouveau_buffer_transfer_del(nv, &tx);
} else
- return FALSE;
+ return false;
assert(buf->domain == new_domain);
- return TRUE;
+ return true;
}
/* Migrate data from glVertexAttribPointer(non-VBO) user buffers to GART.
* We'd like to only allocate @size bytes here, but then we'd have to rebase
* the vertex indices ...
*/
-boolean
+bool
nouveau_user_buffer_upload(struct nouveau_context *nv,
struct nv04_resource *buf,
unsigned base, unsigned size)
buf->base.width0 = base + size;
if (!nouveau_buffer_reallocate(screen, buf, NOUVEAU_BO_GART))
- return FALSE;
+ return false;
ret = nouveau_bo_map(buf->bo, 0, nv->client);
if (ret)
- return FALSE;
+ return false;
memcpy((uint8_t *)buf->bo->map + buf->offset + base, buf->data + base, size);
- return TRUE;
+ return true;
}
4096, size, NULL, pbo);
}
+static void
+nouveau_scratch_unref_bos(void *d)
+{
+ struct runout *b = d;
+ int i;
+
+ for (i = 0; i < b->nr; ++i)
+ nouveau_bo_ref(NULL, &b->bo[i]);
+
+ FREE(b);
+}
+
void
nouveau_scratch_runout_release(struct nouveau_context *nv)
{
- if (!nv->scratch.nr_runout)
+ if (!nv->scratch.runout)
+ return;
+
+ if (!nouveau_fence_work(nv->screen->fence.current, nouveau_scratch_unref_bos,
+ nv->scratch.runout))
return;
- do {
- --nv->scratch.nr_runout;
- nouveau_bo_ref(NULL, &nv->scratch.runout[nv->scratch.nr_runout]);
- } while (nv->scratch.nr_runout);
- FREE(nv->scratch.runout);
nv->scratch.end = 0;
nv->scratch.runout = NULL;
}
/* Allocate an extra bo if we can't fit everything we need simultaneously.
* (Could happen for very large user arrays.)
*/
-static INLINE boolean
+static INLINE bool
nouveau_scratch_runout(struct nouveau_context *nv, unsigned size)
{
int ret;
- const unsigned n = nv->scratch.nr_runout++;
-
- nv->scratch.runout = REALLOC(nv->scratch.runout,
- (n + 0) * sizeof(*nv->scratch.runout),
- (n + 1) * sizeof(*nv->scratch.runout));
- nv->scratch.runout[n] = NULL;
+ unsigned n;
- ret = nouveau_scratch_bo_alloc(nv, &nv->scratch.runout[n], size);
+ if (nv->scratch.runout)
+ n = nv->scratch.runout->nr;
+ else
+ n = 0;
+ nv->scratch.runout = REALLOC(nv->scratch.runout, n == 0 ? 0 :
+ (sizeof(*nv->scratch.runout) + (n + 0) * sizeof(void *)),
+ sizeof(*nv->scratch.runout) + (n + 1) * sizeof(void *));
+ nv->scratch.runout->nr = n + 1;
+ nv->scratch.runout->bo[n] = NULL;
+
+ ret = nouveau_scratch_bo_alloc(nv, &nv->scratch.runout->bo[n], size);
if (!ret) {
- ret = nouveau_bo_map(nv->scratch.runout[n], 0, NULL);
+ ret = nouveau_bo_map(nv->scratch.runout->bo[n], 0, NULL);
if (ret)
- nouveau_bo_ref(NULL, &nv->scratch.runout[--nv->scratch.nr_runout]);
+ nouveau_bo_ref(NULL, &nv->scratch.runout->bo[--nv->scratch.runout->nr]);
}
if (!ret) {
- nv->scratch.current = nv->scratch.runout[n];
+ nv->scratch.current = nv->scratch.runout->bo[n];
nv->scratch.offset = 0;
nv->scratch.end = size;
nv->scratch.map = nv->scratch.current->map;
/* Continue to next scratch buffer, if available (no wrapping, large enough).
* Allocate it if it has not yet been created.
*/
-static INLINE boolean
+static INLINE bool
nouveau_scratch_next(struct nouveau_context *nv, unsigned size)
{
struct nouveau_bo *bo;
const unsigned i = (nv->scratch.id + 1) % NOUVEAU_MAX_SCRATCH_BUFS;
if ((size > nv->scratch.bo_size) || (i == nv->scratch.wrap))
- return FALSE;
+ return false;
nv->scratch.id = i;
bo = nv->scratch.bo[i];
if (!bo) {
ret = nouveau_scratch_bo_alloc(nv, &bo, nv->scratch.bo_size);
if (ret)
- return FALSE;
+ return false;
nv->scratch.bo[i] = bo;
}
nv->scratch.current = bo;
return !ret;
}
-static boolean
+static bool
nouveau_scratch_more(struct nouveau_context *nv, unsigned min_size)
{
- boolean ret;
+ bool ret;
ret = nouveau_scratch_next(nv, min_size);
if (!ret)