diff options
Diffstat (limited to 'src/gallium/drivers/ilo/ilo_transfer.c')
-rw-r--r-- | src/gallium/drivers/ilo/ilo_transfer.c | 248 |
1 files changed, 164 insertions, 84 deletions
diff --git a/src/gallium/drivers/ilo/ilo_transfer.c b/src/gallium/drivers/ilo/ilo_transfer.c index 906c20162ce..7645a250173 100644 --- a/src/gallium/drivers/ilo/ilo_transfer.c +++ b/src/gallium/drivers/ilo/ilo_transfer.c @@ -36,6 +36,159 @@ #include "ilo_state.h" #include "ilo_transfer.h" +/* + * For buffers that are not busy, we want to map/unmap them directly. For + * those that are busy, we have to worry about synchronization. We could wait + * for GPU to finish, but there are cases where we could avoid waiting. + * + * - When PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE is set, the contents of the + * buffer can be discarded. We can replace the backing bo by a new one of + * the same size (renaming). + * - When PIPE_TRANSFER_DISCARD_RANGE is set, the contents of the mapped + * range can be discarded. We can allocate and map a staging bo on + * mapping, and (pipelined-)copy it over to the real bo on unmapping. + * - When PIPE_TRANSFER_FLUSH_EXPLICIT is set, there is no reading and only + * flushed regions need to be written. We can still allocate and map a + * staging bo, but should copy only the flushed regions over. + * + * However, there are other flags to consider. + * + * - When PIPE_TRANSFER_UNSYNCHRONIZED is set, we do not need to worry about + * synchronization at all on mapping. + * - When PIPE_TRANSFER_MAP_DIRECTLY is set, no staging area is allowed. + * - When PIPE_TRANSFER_DONTBLOCK is set, we should fail if we have to block. + * + * These also apply to textures, except that we may additionally need to do + * format conversion or tiling/untiling. + */ + +/** + * Return a transfer method suitable for the usage. The returned method will + * correctly block when the resource is busy. + */ +static bool +resource_get_transfer_method(struct pipe_resource *res, unsigned usage, + enum ilo_transfer_map_method *method) +{ + const struct ilo_screen *is = ilo_screen(res->screen); + enum ilo_transfer_map_method m; + bool tiled; + + if (res->target == PIPE_BUFFER) { + tiled = false; + } + else { + struct ilo_texture *tex = ilo_texture(res); + bool need_convert = true; + + /* we may need to convert on the fly */ + if (tex->separate_s8 || tex->bo_format == PIPE_FORMAT_S8_UINT) + m = ILO_TRANSFER_MAP_SW_ZS; + else if (tex->bo_format != tex->base.format) + m = ILO_TRANSFER_MAP_SW_CONVERT; + else + need_convert = false; + + if (need_convert) { + if (usage & PIPE_TRANSFER_MAP_DIRECTLY) + return false; + + *method = m; + return true; + } + + tiled = (tex->tiling != INTEL_TILING_NONE); + } + + if (tiled) + m = ILO_TRANSFER_MAP_GTT; /* to have a linear view */ + else if (is->dev.has_llc) + m = ILO_TRANSFER_MAP_CPU; /* fast and mostly coherent */ + else if (usage & PIPE_TRANSFER_READ) + m = ILO_TRANSFER_MAP_CPU; /* gtt read is too slow */ + else + m = ILO_TRANSFER_MAP_GTT; + + *method = m; + + return true; +} + +/** + * Return the bo of the resource. + */ +static struct intel_bo * +resource_get_bo(struct pipe_resource *res) +{ + return (res->target == PIPE_BUFFER) ? + ilo_buffer(res)->bo : ilo_texture(res)->bo; +} + +/** + * Rename the bo of the resource. + */ +static bool +resource_rename_bo(struct pipe_resource *res) +{ + return (res->target == PIPE_BUFFER) ? + ilo_buffer_rename_bo(ilo_buffer(res)) : + ilo_texture_rename_bo(ilo_texture(res)); +} + +/** + * Return true if usage allows the use of staging bo to avoid blocking. + */ +static bool +usage_allows_staging_bo(unsigned usage) +{ + /* do we know how to write the data back to the resource? */ + const unsigned can_writeback = (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE | + PIPE_TRANSFER_DISCARD_RANGE | + PIPE_TRANSFER_FLUSH_EXPLICIT); + const unsigned reasons_against = (PIPE_TRANSFER_READ | + PIPE_TRANSFER_MAP_DIRECTLY); + + return (usage & can_writeback) && !(usage & reasons_against); +} + +/** + * Use an alternative transfer method or rename the resource to unblock an + * otherwise blocking transfer. + */ +static bool +xfer_unblock(struct ilo_transfer *xfer, bool *resource_renamed) +{ + struct pipe_resource *res = xfer->base.resource; + bool unblocked = false, renamed = false; + + switch (xfer->method) { + case ILO_TRANSFER_MAP_CPU: + case ILO_TRANSFER_MAP_GTT: + if (xfer->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED) { + xfer->method = ILO_TRANSFER_MAP_UNSYNC; + unblocked = true; + } + else if ((xfer->base.usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) && + resource_rename_bo(res)) { + renamed = true; + unblocked = true; + } + else if (usage_allows_staging_bo(xfer->base.usage)) { + /* TODO */ + } + break; + case ILO_TRANSFER_MAP_UNSYNC: + unblocked = true; + break; + default: + break; + } + + *resource_renamed = renamed; + + return unblocked; +} + static bool is_bo_busy(struct ilo_context *ilo, struct intel_bo *bo, bool *need_flush) { @@ -83,99 +236,26 @@ static bool choose_transfer_method(struct ilo_context *ilo, struct ilo_transfer *xfer) { struct pipe_resource *res = xfer->base.resource; - const unsigned usage = xfer->base.usage; - /* prefer map() when there is the last-level cache */ - const bool prefer_cpu = - (ilo->dev->has_llc || (usage & PIPE_TRANSFER_READ)); - struct ilo_texture *tex; - struct ilo_buffer *buf; - struct intel_bo *bo; - bool tiled, need_flush; - - if (res->target == PIPE_BUFFER) { - tex = NULL; - - buf = ilo_buffer(res); - bo = buf->bo; - tiled = false; - } - else { - buf = NULL; - - tex = ilo_texture(res); - bo = tex->bo; - tiled = (tex->tiling != INTEL_TILING_NONE); - } + bool need_flush; - /* choose between mapping through CPU or GTT */ - if (usage & PIPE_TRANSFER_MAP_DIRECTLY) { - /* we do not want fencing */ - if (tiled || prefer_cpu) - xfer->method = ILO_TRANSFER_MAP_CPU; - else - xfer->method = ILO_TRANSFER_MAP_GTT; - } - else { - if (!tiled && prefer_cpu) - xfer->method = ILO_TRANSFER_MAP_CPU; - else - xfer->method = ILO_TRANSFER_MAP_GTT; - } + if (!resource_get_transfer_method(res, xfer->base.usage, &xfer->method)) + return false; - /* see if we can avoid stalling */ - if (is_bo_busy(ilo, bo, &need_flush)) { - bool will_stall = true; + /* see if we can avoid blocking */ + if (is_bo_busy(ilo, resource_get_bo(res), &need_flush)) { + bool resource_renamed; - if (usage & PIPE_TRANSFER_MAP_DIRECTLY) { - /* nothing we can do */ - } - else if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) { - /* unsynchronized gtt mapping does not stall */ - xfer->method = ILO_TRANSFER_MAP_UNSYNC; - will_stall = false; - } - else if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) { - /* discard old bo and allocate a new one for mapping */ - if ((tex && ilo_texture_rename_bo(tex)) || - (buf && ilo_buffer_rename_bo(buf))) { - ilo_mark_states_with_resource_dirty(ilo, res); - will_stall = false; - } - } - else if (usage & PIPE_TRANSFER_FLUSH_EXPLICIT) { - /* - * We could allocate and return a system buffer here. When a region of - * the buffer is explicitly flushed, we pwrite() the region to a - * temporary bo and emit pipelined copy blit. - * - * For now, do nothing. - */ - } - else if (usage & PIPE_TRANSFER_DISCARD_RANGE) { - /* - * We could allocate a temporary bo for mapping, and emit pipelined copy - * blit upon unmapping. - * - * For now, do nothing. - */ - } - - if (will_stall) { - if (usage & PIPE_TRANSFER_DONTBLOCK) + if (!xfer_unblock(xfer, &resource_renamed)) { + if (xfer->base.usage & PIPE_TRANSFER_DONTBLOCK) return false; - /* flush to make bo busy (so that map() stalls as it should be) */ + /* flush to make bo really busy so that map() correctly blocks */ if (need_flush) ilo_cp_flush(ilo->cp, "syncing for transfers"); } - } - if (tex && !(usage & PIPE_TRANSFER_MAP_DIRECTLY)) { - if (tex->separate_s8 || tex->bo_format == PIPE_FORMAT_S8_UINT) - xfer->method = ILO_TRANSFER_MAP_SW_ZS; - /* need to convert on-the-fly */ - else if (tex->bo_format != tex->base.format) - xfer->method = ILO_TRANSFER_MAP_SW_CONVERT; + if (resource_renamed) + ilo_mark_states_with_resource_dirty(ilo, res); } return true; |