struct pipe_resource *
nouveau_buffer_create(struct pipe_screen *pscreen,
const struct pipe_resource *templ)
{
struct nouveau_screen *screen = nouveau_screen(pscreen);
struct nv04_resource *buffer;
boolean ret;
buffer = CALLOC_STRUCT(nv04_resource);
if (!buffer)
return NULL;
buffer->base = *templ;
buffer->vtbl = &nouveau_buffer_vtbl;
pipe_reference_init(&buffer->base.reference, 1);
buffer->base.screen = pscreen;
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;
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;
break;
case PIPE_USAGE_STAGING:
case PIPE_USAGE_STREAM:
buffer->domain = NOUVEAU_BO_GART;
break;
default:
assert(0);
break;
}
} else {
if (buffer->base.bind & screen->vidmem_bindings)
buffer->domain = NOUVEAU_BO_VRAM;
else
if (buffer->base.bind & screen->sysmem_bindings)
buffer->domain = NOUVEAU_BO_GART;
}
ret = nouveau_buffer_allocate(screen, buffer, buffer->domain);
if (ret == FALSE)
goto fail;
if (buffer->domain == NOUVEAU_BO_VRAM && screen->hint_buf_keep_sysmem_copy)
nouveau_buffer_cache(NULL, buffer);
return &buffer->base;
fail:
FREE(buffer);
return NULL;
}
void *
nouveau_resource_map_offset(struct nouveau_context *nv,
struct nv04_resource *res, uint32_t offset,
uint32_t flags)
{
if (unlikely(res->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY))
return res->data + offset;
if (res->domain == NOUVEAU_BO_VRAM) {
if (!res->data || (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING))
nouveau_buffer_cache(nv, res);
}
if (res->domain != NOUVEAU_BO_GART)
return res->data + offset;
if (res->mm) {
unsigned rw;
rw = (flags & NOUVEAU_BO_WR) ? PIPE_TRANSFER_WRITE : PIPE_TRANSFER_READ;
nouveau_buffer_sync(res, rw);
if (nouveau_bo_map(res->bo, 0, NULL))
return NULL;
} else {
if (nouveau_bo_map(res->bo, flags, nv->client))
return NULL;
}
return (uint8_t *)res->bo->map + res->offset + offset;
}
/* 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,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer)
{
struct nouveau_context *nv = nouveau_context(pipe);
struct nv04_resource *buf = nv04_resource(resource);
struct nouveau_transfer *tx = MALLOC_STRUCT(nouveau_transfer);
uint8_t *map;
int ret;
if (!tx)
return NULL;
nouveau_buffer_transfer_init(tx, resource, box, usage);
*ptransfer = &tx->base;
if (usage & PIPE_TRANSFER_READ)
NOUVEAU_DRV_STAT(nv->screen, buf_transfers_rd, 1);
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);
} 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_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);
if (!buf->data)
nouveau_buffer_cache(nv, buf);
}
}
return buf->data ? (buf->data + box->x) : tx->map;
} else
if (unlikely(buf->domain == 0)) {
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);
if (ref > 0) /* any references inside context possible ? */
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 (ret) {
FREE(tx);
return NULL;
}
map = (uint8_t *)buf->bo->map + buf->offset + box->x;
/* using kernel fences only if !buf->mm */
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
* subsequent transfers might use UNSYNCHRONIZED. */
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else
if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
/* 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)) {
if (usage & PIPE_TRANSFER_DONTBLOCK)
map = NULL;
else
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else {
/* 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;
}
}
if (!map)
FREE(tx);
return map;
}
static void *
nouveau_buffer_transfer_map(struct pipe_context *pipe,
struct pipe_resource *resource,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer)
{
struct nouveau_context *nv = nouveau_context(pipe);
struct nv04_resource *buf = nv04_resource(resource);
struct nouveau_transfer *tx = MALLOC_STRUCT(nouveau_transfer);
uint8_t *map;
int ret;
if (!tx)
return NULL;
nouveau_buffer_transfer_init(tx, resource, box, usage);
*ptransfer = &tx->base;
if (buf->domain == NOUVEAU_BO_VRAM) {
if (usage & NOUVEAU_TRANSFER_DISCARD) {
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
buf->status &= NOUVEAU_BUFFER_STATUS_REALLOC_MASK;
nouveau_transfer_staging(nv, tx, TRUE);
} else {
if (buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
if (buf->data) {
align_free(buf->data);
buf->data = NULL;
}
nouveau_transfer_staging(nv, tx, FALSE);
nouveau_transfer_read(nv, tx);
} else {
if (usage & PIPE_TRANSFER_WRITE)
nouveau_transfer_staging(nv, tx, TRUE);
if (!buf->data)
nouveau_buffer_cache(nv, buf);
}
}
return buf->data ? (buf->data + box->x) : tx->map;
} else
if (unlikely(buf->domain == 0)) {
return buf->data + box->x;
}
if (nouveau_buffer_should_discard(buf, usage)) {
int ref = buf->base.reference.count - 1;
nouveau_buffer_reallocate(nv->screen, buf, buf->domain);
if (ref > 0) /* any references inside context possible ? */
nv->invalidate_resource_storage(nv, &buf->base, ref);
}
ret = nouveau_bo_map(buf->bo,
buf->mm ? 0 : nouveau_screen_transfer_flags(usage),
nv->client);
if (ret) {
FREE(tx);
return NULL;
}
map = (uint8_t *)buf->bo->map + buf->offset + box->x;
/* using kernel fences only if !buf->mm */
if ((usage & PIPE_TRANSFER_UNSYNCHRONIZED) || !buf->mm)
return map;
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
* subsequent transfers might use UNSYNCHRONIZED. */
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else
if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
nouveau_transfer_staging(nv, tx, TRUE);
map = tx->map;
} else
if (nouveau_buffer_busy(buf, PIPE_TRANSFER_READ)) {
if (usage & PIPE_TRANSFER_DONTBLOCK)
map = NULL;
else
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else {
nouveau_transfer_staging(nv, tx, TRUE);
if (tx->map)
memcpy(tx->map, map, box->width);
map = tx->map;
}
}
if (!map)
FREE(tx);
return map;
}
struct pipe_resource *
nouveau_buffer_create(struct pipe_screen *pscreen,
const struct pipe_resource *templ)
{
struct nouveau_screen *screen = nouveau_screen(pscreen);
struct nv04_resource *buffer;
bool ret;
buffer = CALLOC_STRUCT(nv04_resource);
if (!buffer)
return NULL;
buffer->base = *templ;
buffer->vtbl = &nouveau_buffer_vtbl;
pipe_reference_init(&buffer->base.reference, 1);
buffer->base.screen = pscreen;
if (buffer->base.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT |
PIPE_RESOURCE_FLAG_MAP_COHERENT)) {
buffer->domain = NOUVEAU_BO_GART;
} else if (buffer->base.bind == 0 || (buffer->base.bind &
(screen->vidmem_bindings & screen->sysmem_bindings))) {
switch (buffer->base.usage) {
case PIPE_USAGE_DEFAULT:
case PIPE_USAGE_IMMUTABLE:
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 = NV_VRAM_DOMAIN(screen);
break;
case PIPE_USAGE_STAGING:
case PIPE_USAGE_STREAM:
buffer->domain = NOUVEAU_BO_GART;
break;
default:
assert(0);
break;
}
} else {
if (buffer->base.bind & screen->vidmem_bindings)
buffer->domain = NV_VRAM_DOMAIN(screen);
else
if (buffer->base.bind & screen->sysmem_bindings)
buffer->domain = NOUVEAU_BO_GART;
}
/* There can be very special situations where we want non-gpu-mapped
* buffers, but never through this interface.
*/
assert(buffer->domain);
ret = nouveau_buffer_allocate(screen, buffer, buffer->domain);
if (ret == false)
goto fail;
if (buffer->domain == NOUVEAU_BO_VRAM && screen->hint_buf_keep_sysmem_copy)
nouveau_buffer_cache(NULL, buffer);
NOUVEAU_DRV_STAT(screen, buf_obj_current_count, 1);
util_range_init(&buffer->valid_buffer_range);
return &buffer->base;
fail:
FREE(buffer);
return NULL;
}
