* Jerome Glisse
* Corbin Simpson <MostAwesomeDude@gmail.com>
*/
-#include <pipe/p_screen.h>
-#include <util/u_format.h>
-#include <util/u_math.h>
-#include <util/u_inlines.h>
-#include <util/u_memory.h>
-#include <util/u_upload_mgr.h>
-#include "state_tracker/drm_driver.h"
-#include <xf86drm.h>
-#include "radeon_drm.h"
-#include "r600.h"
#include "r600_pipe.h"
+#include "util/u_upload_mgr.h"
+#include "util/u_memory.h"
+#include "util/u_surface.h"
-extern struct u_resource_vtbl r600_buffer_vtbl;
-
-
-struct pipe_resource *r600_buffer_create(struct pipe_screen *screen,
- const struct pipe_resource *templ)
+static void r600_buffer_destroy(struct pipe_screen *screen,
+ struct pipe_resource *buf)
{
- struct r600_resource_buffer *rbuffer;
- struct r600_bo *bo;
- /* XXX We probably want a different alignment for buffers and textures. */
- unsigned alignment = 4096;
+ struct r600_resource *rbuffer = r600_resource(buf);
- rbuffer = CALLOC_STRUCT(r600_resource_buffer);
- if (rbuffer == NULL)
- return NULL;
+ util_range_destroy(&rbuffer->valid_buffer_range);
+ pb_reference(&rbuffer->buf, NULL);
+ FREE(rbuffer);
+}
- rbuffer->magic = R600_BUFFER_MAGIC;
- rbuffer->user_buffer = NULL;
- rbuffer->num_ranges = 0;
- rbuffer->r.base.b = *templ;
- pipe_reference_init(&rbuffer->r.base.b.reference, 1);
- rbuffer->r.base.b.screen = screen;
- rbuffer->r.base.vtbl = &r600_buffer_vtbl;
- rbuffer->r.size = rbuffer->r.base.b.width0;
- bo = r600_bo((struct radeon*)screen->winsys, rbuffer->r.base.b.width0, alignment, rbuffer->r.base.b.bind, rbuffer->r.base.b.usage);
- if (bo == NULL) {
- FREE(rbuffer);
- return NULL;
+static void r600_set_constants_dirty_if_bound(struct r600_context *rctx,
+ struct r600_resource *rbuffer)
+{
+ unsigned shader;
+
+ for (shader = 0; shader < PIPE_SHADER_TYPES; shader++) {
+ struct r600_constbuf_state *state = &rctx->constbuf_state[shader];
+ bool found = false;
+ uint32_t mask = state->enabled_mask;
+
+ while (mask) {
+ unsigned i = u_bit_scan(&mask);
+ if (state->cb[i].buffer == &rbuffer->b.b) {
+ found = true;
+ state->dirty_mask |= 1 << i;
+ }
+ }
+ if (found) {
+ r600_constant_buffers_dirty(rctx, state);
+ }
}
- rbuffer->r.bo = bo;
- return &rbuffer->r.base.b;
}
-struct pipe_resource *r600_user_buffer_create(struct pipe_screen *screen,
- void *ptr, unsigned bytes,
- unsigned bind)
+static void *r600_buffer_get_transfer(struct pipe_context *ctx,
+ struct pipe_resource *resource,
+ unsigned level,
+ unsigned usage,
+ const struct pipe_box *box,
+ struct pipe_transfer **ptransfer,
+ void *data, struct r600_resource *staging,
+ unsigned offset)
{
- struct r600_resource_buffer *rbuffer;
-
- rbuffer = CALLOC_STRUCT(r600_resource_buffer);
- if (rbuffer == NULL)
- return NULL;
-
- rbuffer->magic = R600_BUFFER_MAGIC;
- pipe_reference_init(&rbuffer->r.base.b.reference, 1);
- rbuffer->r.base.vtbl = &r600_buffer_vtbl;
- rbuffer->r.base.b.screen = screen;
- rbuffer->r.base.b.target = PIPE_BUFFER;
- rbuffer->r.base.b.format = PIPE_FORMAT_R8_UNORM;
- rbuffer->r.base.b.usage = PIPE_USAGE_IMMUTABLE;
- rbuffer->r.base.b.bind = bind;
- rbuffer->r.base.b.width0 = bytes;
- rbuffer->r.base.b.height0 = 1;
- rbuffer->r.base.b.depth0 = 1;
- rbuffer->r.base.b.flags = 0;
- rbuffer->num_ranges = 0;
- rbuffer->r.bo = NULL;
- rbuffer->user_buffer = ptr;
- return &rbuffer->r.base.b;
+ struct r600_context *rctx = (struct r600_context*)ctx;
+ struct r600_transfer *transfer = util_slab_alloc(&rctx->pool_transfers);
+
+ transfer->transfer.resource = resource;
+ transfer->transfer.level = level;
+ transfer->transfer.usage = usage;
+ transfer->transfer.box = *box;
+ transfer->transfer.stride = 0;
+ transfer->transfer.layer_stride = 0;
+ transfer->offset = offset;
+ transfer->staging = staging;
+ *ptransfer = &transfer->transfer;
+ return data;
}
-static void r600_buffer_destroy(struct pipe_screen *screen,
- struct pipe_resource *buf)
+static void *r600_buffer_transfer_map(struct pipe_context *ctx,
+ struct pipe_resource *resource,
+ unsigned level,
+ unsigned usage,
+ const struct pipe_box *box,
+ struct pipe_transfer **ptransfer)
{
- struct r600_resource_buffer *rbuffer = r600_buffer(buf);
+ struct r600_context *rctx = (struct r600_context*)ctx;
+ struct r600_resource *rbuffer = r600_resource(resource);
+ uint8_t *data;
- if (rbuffer->r.bo) {
- r600_bo_reference((struct radeon*)screen->winsys, &rbuffer->r.bo, NULL);
+ assert(box->x + box->width <= resource->width0);
+
+ /* See if the buffer range being mapped has never been initialized,
+ * in which case it can be mapped unsynchronized. */
+ if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
+ usage & PIPE_TRANSFER_WRITE &&
+ !util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
+ usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
- FREE(rbuffer);
-}
-static void *r600_buffer_transfer_map(struct pipe_context *pipe,
- struct pipe_transfer *transfer)
-{
- struct r600_resource_buffer *rbuffer = r600_buffer(transfer->resource);
- int write = 0;
- uint8_t *data;
- int i;
- boolean flush = FALSE;
-
- if (rbuffer->user_buffer)
- return (uint8_t*)rbuffer->user_buffer + transfer->box.x;
-
- if (transfer->usage & PIPE_TRANSFER_DISCARD) {
- for (i = 0; i < rbuffer->num_ranges; i++) {
- if ((transfer->box.x >= rbuffer->ranges[i].start) &&
- (transfer->box.x < rbuffer->ranges[i].end))
- flush = TRUE;
-
- if (flush) {
- r600_bo_reference((struct radeon*)pipe->winsys, &rbuffer->r.bo, NULL);
- rbuffer->num_ranges = 0;
- rbuffer->r.bo = r600_bo((struct radeon*)pipe->winsys,
- rbuffer->r.base.b.width0, 0,
- rbuffer->r.base.b.bind,
- rbuffer->r.base.b.usage);
- break;
+ if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
+ !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
+ assert(usage & PIPE_TRANSFER_WRITE);
+
+ /* Check if mapping this buffer would cause waiting for the GPU. */
+ if (r600_rings_is_buffer_referenced(rctx, rbuffer->cs_buf, RADEON_USAGE_READWRITE) ||
+ rctx->b.ws->buffer_is_busy(rbuffer->buf, RADEON_USAGE_READWRITE)) {
+ unsigned i, mask;
+
+ /* Discard the buffer. */
+ pb_reference(&rbuffer->buf, NULL);
+
+ /* Create a new one in the same pipe_resource. */
+ /* XXX We probably want a different alignment for buffers and textures. */
+ r600_init_resource(rctx->screen, rbuffer, rbuffer->b.b.width0, 4096,
+ TRUE, rbuffer->b.b.usage);
+
+ /* We changed the buffer, now we need to bind it where the old one was bound. */
+ /* Vertex buffers. */
+ mask = rctx->vertex_buffer_state.enabled_mask;
+ while (mask) {
+ i = u_bit_scan(&mask);
+ if (rctx->vertex_buffer_state.vb[i].buffer == &rbuffer->b.b) {
+ rctx->vertex_buffer_state.dirty_mask |= 1 << i;
+ r600_vertex_buffers_dirty(rctx);
+ }
+ }
+ /* Streamout buffers. */
+ for (i = 0; i < rctx->b.streamout.num_targets; i++) {
+ if (rctx->b.streamout.targets[i]->b.buffer == &rbuffer->b.b) {
+ if (rctx->b.streamout.begin_emitted) {
+ r600_emit_streamout_end(&rctx->b);
+ }
+ rctx->b.streamout.append_bitmask = rctx->b.streamout.enabled_mask;
+ r600_streamout_buffers_dirty(&rctx->b);
+ }
}
+ /* Constant buffers. */
+ r600_set_constants_dirty_if_bound(rctx, rbuffer);
}
}
- if (transfer->usage & PIPE_TRANSFER_DONTBLOCK) {
- /* FIXME */
- }
- if (transfer->usage & PIPE_TRANSFER_WRITE) {
- write = 1;
+ else if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
+ !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
+ !(rctx->screen->debug_flags & DBG_NO_DISCARD_RANGE) &&
+ (rctx->screen->has_cp_dma ||
+ (rctx->screen->has_streamout &&
+ /* The buffer range must be aligned to 4 with streamout. */
+ box->x % 4 == 0 && box->width % 4 == 0))) {
+ assert(usage & PIPE_TRANSFER_WRITE);
+
+ /* Check if mapping this buffer would cause waiting for the GPU. */
+ if (r600_rings_is_buffer_referenced(rctx, rbuffer->cs_buf, RADEON_USAGE_READWRITE) ||
+ rctx->b.ws->buffer_is_busy(rbuffer->buf, RADEON_USAGE_READWRITE)) {
+ /* Do a wait-free write-only transfer using a temporary buffer. */
+ unsigned offset;
+ struct r600_resource *staging = NULL;
+
+ u_upload_alloc(rctx->uploader, 0, box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
+ &offset, (struct pipe_resource**)&staging, (void**)&data);
+
+ if (staging) {
+ data += box->x % R600_MAP_BUFFER_ALIGNMENT;
+ return r600_buffer_get_transfer(ctx, resource, level, usage, box,
+ ptransfer, data, staging, offset);
+ }
+ }
}
- data = r600_bo_map((struct radeon*)pipe->winsys, rbuffer->r.bo, transfer->usage, pipe);
- if (!data)
+
+ /* mmap and synchronize with rings */
+ data = r600_buffer_mmap_sync_with_rings(rctx, rbuffer, usage);
+ if (!data) {
return NULL;
+ }
+ data += box->x;
- return (uint8_t*)data + transfer->box.x;
+ return r600_buffer_get_transfer(ctx, resource, level, usage, box,
+ ptransfer, data, NULL, 0);
}
static void r600_buffer_transfer_unmap(struct pipe_context *pipe,
struct pipe_transfer *transfer)
{
- struct r600_resource_buffer *rbuffer = r600_buffer(transfer->resource);
-
- if (rbuffer->r.bo)
- r600_bo_unmap((struct radeon*)pipe->winsys, rbuffer->r.bo);
-}
+ struct r600_context *rctx = (struct r600_context*)pipe;
+ struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
+ struct r600_resource *rbuffer = r600_resource(transfer->resource);
+
+ if (rtransfer->staging) {
+ struct pipe_resource *dst, *src;
+ unsigned soffset, doffset, size;
+
+ dst = transfer->resource;
+ src = &rtransfer->staging->b.b;
+ size = transfer->box.width;
+ doffset = transfer->box.x;
+ soffset = rtransfer->offset + transfer->box.x % R600_MAP_BUFFER_ALIGNMENT;
+ /* Copy the staging buffer into the original one. */
+ if (rctx->b.rings.dma.cs && !(size % 4) && !(doffset % 4) && !(soffset % 4)) {
+ if (rctx->screen->b.chip_class >= EVERGREEN) {
+ evergreen_dma_copy(rctx, dst, src, doffset, soffset, size);
+ } else {
+ r600_dma_copy(rctx, dst, src, doffset, soffset, size);
+ }
+ } else {
+ struct pipe_box box;
-static void r600_buffer_transfer_flush_region(struct pipe_context *pipe,
- struct pipe_transfer *transfer,
- const struct pipe_box *box)
-{
- struct r600_resource_buffer *rbuffer = r600_buffer(transfer->resource);
- unsigned i;
- unsigned offset = transfer->box.x + box->x;
- unsigned length = box->width;
-
- assert(box->x + box->width <= transfer->box.width);
-
- if (rbuffer->user_buffer)
- return;
-
- /* mark the range as used */
- for(i = 0; i < rbuffer->num_ranges; ++i) {
- if(offset <= rbuffer->ranges[i].end && rbuffer->ranges[i].start <= (offset+box->width)) {
- rbuffer->ranges[i].start = MIN2(rbuffer->ranges[i].start, offset);
- rbuffer->ranges[i].end = MAX2(rbuffer->ranges[i].end, (offset+length));
- return;
+ u_box_1d(soffset, size, &box);
+ r600_copy_buffer(pipe, dst, doffset, src, &box);
}
+ pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
}
- rbuffer->ranges[rbuffer->num_ranges].start = offset;
- rbuffer->ranges[rbuffer->num_ranges].end = offset+length;
- rbuffer->num_ranges++;
-}
-
-unsigned r600_buffer_is_referenced_by_cs(struct pipe_context *context,
- struct pipe_resource *buf,
- unsigned face, unsigned level)
-{
- /* FIXME */
- return PIPE_REFERENCED_FOR_READ | PIPE_REFERENCED_FOR_WRITE;
-}
-
-struct pipe_resource *r600_buffer_from_handle(struct pipe_screen *screen,
- struct winsys_handle *whandle)
-{
- struct radeon *rw = (struct radeon*)screen->winsys;
- struct r600_resource *rbuffer;
- struct r600_bo *bo = NULL;
-
- bo = r600_bo_handle(rw, whandle->handle, NULL);
- if (bo == NULL) {
- return NULL;
- }
-
- rbuffer = CALLOC_STRUCT(r600_resource);
- if (rbuffer == NULL) {
- r600_bo_reference(rw, &bo, NULL);
- return NULL;
+ if (transfer->usage & PIPE_TRANSFER_WRITE) {
+ util_range_add(&rbuffer->valid_buffer_range, transfer->box.x,
+ transfer->box.x + transfer->box.width);
}
-
- pipe_reference_init(&rbuffer->base.b.reference, 1);
- rbuffer->base.b.target = PIPE_BUFFER;
- rbuffer->base.b.screen = screen;
- rbuffer->base.vtbl = &r600_buffer_vtbl;
- rbuffer->bo = bo;
- return &rbuffer->base.b;
+ util_slab_free(&rctx->pool_transfers, transfer);
}
-struct u_resource_vtbl r600_buffer_vtbl =
+static const struct u_resource_vtbl r600_buffer_vtbl =
{
u_default_resource_get_handle, /* get_handle */
r600_buffer_destroy, /* resource_destroy */
- r600_buffer_is_referenced_by_cs, /* is_buffer_referenced */
- u_default_get_transfer, /* get_transfer */
- u_default_transfer_destroy, /* transfer_destroy */
r600_buffer_transfer_map, /* transfer_map */
- r600_buffer_transfer_flush_region, /* transfer_flush_region */
+ NULL, /* transfer_flush_region */
r600_buffer_transfer_unmap, /* transfer_unmap */
- u_default_transfer_inline_write /* transfer_inline_write */
+ NULL /* transfer_inline_write */
};
-int r600_upload_index_buffer(struct r600_pipe_context *rctx, struct r600_drawl *draw)
+bool r600_init_resource(struct r600_screen *rscreen,
+ struct r600_resource *res,
+ unsigned size, unsigned alignment,
+ bool use_reusable_pool, unsigned usage)
{
- struct pipe_resource *upload_buffer = NULL;
- unsigned index_offset = draw->index_buffer_offset;
- int ret = 0;
-
- if (r600_buffer_is_user_buffer(draw->index_buffer)) {
- ret = u_upload_buffer(rctx->upload_ib,
- index_offset,
- draw->count * draw->index_size,
- draw->index_buffer,
- &index_offset,
- &upload_buffer);
- if (ret) {
- goto done;
- }
- draw->index_buffer_offset = index_offset;
+ uint32_t initial_domain, domains;
+
+ switch(usage) {
+ case PIPE_USAGE_STAGING:
+ /* Staging resources participate in transfers, i.e. are used
+ * for uploads and downloads from regular resources.
+ * We generate them internally for some transfers.
+ */
+ initial_domain = RADEON_DOMAIN_GTT;
+ domains = RADEON_DOMAIN_GTT;
+ break;
+ case PIPE_USAGE_DYNAMIC:
+ case PIPE_USAGE_STREAM:
+ /* Default to GTT, but allow the memory manager to move it to VRAM. */
+ initial_domain = RADEON_DOMAIN_GTT;
+ domains = RADEON_DOMAIN_GTT | RADEON_DOMAIN_VRAM;
+ break;
+ case PIPE_USAGE_DEFAULT:
+ case PIPE_USAGE_STATIC:
+ case PIPE_USAGE_IMMUTABLE:
+ default:
+ /* Don't list GTT here, because the memory manager would put some
+ * resources to GTT no matter what the initial domain is.
+ * Not listing GTT in the domains improves performance a lot. */
+ initial_domain = RADEON_DOMAIN_VRAM;
+ domains = RADEON_DOMAIN_VRAM;
+ break;
+ }
- /* Transfer ownership. */
- pipe_resource_reference(&draw->index_buffer, upload_buffer);
- pipe_resource_reference(&upload_buffer, NULL);
+ res->buf = rscreen->b.ws->buffer_create(rscreen->b.ws, size, alignment,
+ use_reusable_pool,
+ initial_domain);
+ if (!res->buf) {
+ return false;
}
-done:
- return ret;
+ res->cs_buf = rscreen->b.ws->buffer_get_cs_handle(res->buf);
+ res->domains = domains;
+ util_range_set_empty(&res->valid_buffer_range);
+
+ if (rscreen->debug_flags & DBG_VM && res->b.b.target == PIPE_BUFFER) {
+ fprintf(stderr, "VM start=0x%llX end=0x%llX | Buffer %u bytes\n",
+ r600_resource_va(&rscreen->b.b, &res->b.b),
+ r600_resource_va(&rscreen->b.b, &res->b.b) + res->buf->size,
+ res->buf->size);
+ }
+ return true;
}
-int r600_upload_user_buffers(struct r600_pipe_context *rctx)
+struct pipe_resource *r600_buffer_create(struct pipe_screen *screen,
+ const struct pipe_resource *templ,
+ unsigned alignment)
{
- enum pipe_error ret = PIPE_OK;
- int i, nr;
-
- nr = rctx->vertex_elements->count;
-
- for (i = 0; i < nr; i++) {
- struct pipe_vertex_buffer *vb =
- &rctx->vertex_buffer[rctx->vertex_elements->elements[i].vertex_buffer_index];
-
- if (r600_buffer_is_user_buffer(vb->buffer)) {
- struct pipe_resource *upload_buffer = NULL;
- unsigned offset = 0; /*vb->buffer_offset * 4;*/
- unsigned size = vb->buffer->width0;
- unsigned upload_offset;
- ret = u_upload_buffer(rctx->upload_vb,
- offset, size,
- vb->buffer,
- &upload_offset, &upload_buffer);
- if (ret)
- return ret;
-
- pipe_resource_reference(&vb->buffer, NULL);
- vb->buffer = upload_buffer;
- vb->buffer_offset = upload_offset;
- }
+ struct r600_screen *rscreen = (struct r600_screen*)screen;
+ struct r600_resource *rbuffer;
+
+ rbuffer = MALLOC_STRUCT(r600_resource);
+
+ rbuffer->b.b = *templ;
+ pipe_reference_init(&rbuffer->b.b.reference, 1);
+ rbuffer->b.b.screen = screen;
+ rbuffer->b.vtbl = &r600_buffer_vtbl;
+ util_range_init(&rbuffer->valid_buffer_range);
+
+ if (!r600_init_resource(rscreen, rbuffer, templ->width0, alignment, TRUE, templ->usage)) {
+ FREE(rbuffer);
+ return NULL;
}
- return ret;
+ return &rbuffer->b.b;
}
projects / mesa.git / blobdiff