#include "util/u_transfer.h"
#include "util/u_string.h"
#include "util/u_surface.h"
+#include "util/set.h"
#include "freedreno_resource.h"
+#include "freedreno_batch_cache.h"
#include "freedreno_screen.h"
#include "freedreno_surface.h"
#include "freedreno_context.h"
#include <errno.h>
-
-static bool
-pending(struct fd_resource *rsc, enum fd_resource_status status)
-{
- return (rsc->status & status) ||
- (rsc->stencil && (rsc->stencil->status & status));
-}
+/* XXX this should go away, needed for 'struct winsys_handle' */
+#include "state_tracker/drm_driver.h"
static void
fd_invalidate_resource(struct fd_context *ctx, struct pipe_resource *prsc)
rsc->bo = fd_bo_new(screen->dev, size, flags);
rsc->timestamp = 0;
- rsc->status = 0;
- rsc->pending_ctx = NULL;
- list_delinit(&rsc->list);
util_range_set_empty(&rsc->valid_buffer_range);
+ fd_bc_invalidate_resource(rsc, true);
+}
+
+static void
+do_blit(struct fd_context *ctx, const struct pipe_blit_info *blit, bool fallback)
+{
+ /* TODO size threshold too?? */
+ if ((blit->src.resource->target != PIPE_BUFFER) && !fallback) {
+ /* do blit on gpu: */
+ fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_BLIT);
+ util_blitter_blit(ctx->blitter, blit);
+ fd_blitter_pipe_end(ctx);
+ } else {
+ /* do blit on cpu: */
+ util_resource_copy_region(&ctx->base,
+ blit->dst.resource, blit->dst.level, blit->dst.box.x,
+ blit->dst.box.y, blit->dst.box.z,
+ blit->src.resource, blit->src.level, &blit->src.box);
+ }
+}
+
+static bool
+fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
+ unsigned level, unsigned usage, const struct pipe_box *box)
+{
+ struct pipe_context *pctx = &ctx->base;
+ struct pipe_resource *prsc = &rsc->base.b;
+ bool fallback = false;
+
+ if (prsc->next)
+ return false;
+
+ /* TODO: somehow munge dimensions and format to copy unsupported
+ * render target format to something that is supported?
+ */
+ if (!pctx->screen->is_format_supported(pctx->screen,
+ prsc->format, prsc->target, prsc->nr_samples,
+ PIPE_BIND_RENDER_TARGET))
+ fallback = true;
+
+ /* these cases should be handled elsewhere.. just for future
+ * reference in case this gets split into a more generic(ish)
+ * helper.
+ */
+ debug_assert(!(usage & PIPE_TRANSFER_READ));
+ debug_assert(!(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE));
+
+ /* if we do a gpu blit to clone the whole resource, we'll just
+ * end up stalling on that.. so only allow if we can discard
+ * current range (and blit, possibly cpu or gpu, the rest)
+ */
+ if (!(usage & PIPE_TRANSFER_DISCARD_RANGE))
+ return false;
+
+ bool whole_level = util_texrange_covers_whole_level(prsc, level,
+ box->x, box->y, box->z, box->width, box->height, box->depth);
+
+ /* TODO need to be more clever about current level */
+ if ((prsc->target >= PIPE_TEXTURE_2D) && !whole_level)
+ return false;
+
+ struct pipe_resource *pshadow =
+ pctx->screen->resource_create(pctx->screen, prsc);
+
+ if (!pshadow)
+ return false;
+
+ assert(!ctx->in_shadow);
+ ctx->in_shadow = true;
+
+ /* get rid of any references that batch-cache might have to us (which
+ * should empty/destroy rsc->batches hashset)
+ */
+ fd_bc_invalidate_resource(rsc, false);
+
+ mtx_lock(&ctx->screen->lock);
+
+ /* Swap the backing bo's, so shadow becomes the old buffer,
+ * blit from shadow to new buffer. From here on out, we
+ * cannot fail.
+ *
+ * Note that we need to do it in this order, otherwise if
+ * we go down cpu blit path, the recursive transfer_map()
+ * sees the wrong status..
+ */
+ struct fd_resource *shadow = fd_resource(pshadow);
+
+ DBG("shadow: %p (%d) -> %p (%d)\n", rsc, rsc->base.b.reference.count,
+ shadow, shadow->base.b.reference.count);
+
+ /* TODO valid_buffer_range?? */
+ swap(rsc->bo, shadow->bo);
+ swap(rsc->timestamp, shadow->timestamp);
+ swap(rsc->write_batch, shadow->write_batch);
+
+ /* at this point, the newly created shadow buffer is not referenced
+ * by any batches, but the existing rsc (probably) is. We need to
+ * transfer those references over:
+ */
+ debug_assert(shadow->batch_mask == 0);
+ struct fd_batch *batch;
+ foreach_batch(batch, &ctx->screen->batch_cache, rsc->batch_mask) {
+ struct set_entry *entry = _mesa_set_search(batch->resources, rsc);
+ _mesa_set_remove(batch->resources, entry);
+ _mesa_set_add(batch->resources, shadow);
+ }
+ swap(rsc->batch_mask, shadow->batch_mask);
+
+ pipe_mutex_unlock(ctx->screen->lock);
+
+ struct pipe_blit_info blit = {0};
+ blit.dst.resource = prsc;
+ blit.dst.format = prsc->format;
+ blit.src.resource = pshadow;
+ blit.src.format = pshadow->format;
+ blit.mask = util_format_get_mask(prsc->format);
+ blit.filter = PIPE_TEX_FILTER_NEAREST;
+
+#define set_box(field, val) do { \
+ blit.dst.field = (val); \
+ blit.src.field = (val); \
+ } while (0)
+
+ /* blit the other levels in their entirety: */
+ for (unsigned l = 0; l <= prsc->last_level; l++) {
+ if (l == level)
+ continue;
+
+ /* just blit whole level: */
+ set_box(level, l);
+ set_box(box.width, u_minify(prsc->width0, l));
+ set_box(box.height, u_minify(prsc->height0, l));
+ set_box(box.depth, u_minify(prsc->depth0, l));
+
+ do_blit(ctx, &blit, fallback);
+ }
+
+ /* deal w/ current level specially, since we might need to split
+ * it up into a couple blits:
+ */
+ if (!whole_level) {
+ set_box(level, level);
+
+ switch (prsc->target) {
+ case PIPE_BUFFER:
+ case PIPE_TEXTURE_1D:
+ set_box(box.y, 0);
+ set_box(box.z, 0);
+ set_box(box.height, 1);
+ set_box(box.depth, 1);
+
+ if (box->x > 0) {
+ set_box(box.x, 0);
+ set_box(box.width, box->x);
+
+ do_blit(ctx, &blit, fallback);
+ }
+ if ((box->x + box->width) < u_minify(prsc->width0, level)) {
+ set_box(box.x, box->x + box->width);
+ set_box(box.width, u_minify(prsc->width0, level) - (box->x + box->width));
+
+ do_blit(ctx, &blit, fallback);
+ }
+ break;
+ case PIPE_TEXTURE_2D:
+ /* TODO */
+ default:
+ unreachable("TODO");
+ }
+ }
+
+ ctx->in_shadow = false;
+
+ pipe_resource_reference(&pshadow, NULL);
+
+ return true;
}
static unsigned
ptrans->box.x + ptrans->box.width);
pipe_resource_reference(&ptrans->resource, NULL);
- util_slab_free(&ctx->transfer_pool, ptrans);
+ slab_free(&ctx->transfer_pool, ptrans);
- if (trans->staging)
- free(trans->staging);
+ free(trans->staging);
}
static void *
DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc, level, usage,
box->width, box->height, box->x, box->y);
- ptrans = util_slab_alloc(&ctx->transfer_pool);
+ ptrans = slab_alloc(&ctx->transfer_pool);
if (!ptrans)
return NULL;
- /* util_slab_alloc() doesn't zero: */
+ /* slab_alloc_st() doesn't zero: */
trans = fd_transfer(ptrans);
memset(trans, 0, sizeof(*trans));
ptrans->usage = usage;
ptrans->box = *box;
ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->cpp;
- ptrans->layer_stride = slice->size0;
+ ptrans->layer_stride = rsc->layer_first ? rsc->layer_size : slice->size0;
+
+ if (ctx->in_shadow && !(usage & PIPE_TRANSFER_READ))
+ usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
if (usage & PIPE_TRANSFER_READ)
op |= DRM_FREEDRENO_PREP_READ;
* to wait.
*/
} else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
+ struct fd_batch *write_batch = NULL;
+
+ /* hold a reference, so it doesn't disappear under us: */
+ fd_batch_reference(&write_batch, rsc->write_batch);
+
+ if ((usage & PIPE_TRANSFER_WRITE) && write_batch &&
+ write_batch->back_blit) {
+ /* if only thing pending is a back-blit, we can discard it: */
+ fd_batch_reset(write_batch);
+ }
+
/* If the GPU is writing to the resource, or if it is reading from the
* resource and we're trying to write to it, flush the renders.
*/
- if (((ptrans->usage & PIPE_TRANSFER_WRITE) &&
- pending(rsc, FD_PENDING_READ | FD_PENDING_WRITE)) ||
- pending(rsc, FD_PENDING_WRITE))
- fd_context_render(pctx);
+ bool needs_flush = pending(rsc, !!(usage & PIPE_TRANSFER_WRITE));
+ bool busy = needs_flush || (0 != fd_bo_cpu_prep(rsc->bo,
+ ctx->screen->pipe, op | DRM_FREEDRENO_PREP_NOSYNC));
+
+ /* if we need to flush/stall, see if we can make a shadow buffer
+ * to avoid this:
+ *
+ * TODO we could go down this path !reorder && !busy_for_read
+ * ie. we only *don't* want to go down this path if the blit
+ * will trigger a flush!
+ */
+ if (ctx->screen->reorder && busy && !(usage & PIPE_TRANSFER_READ)) {
+ if (fd_try_shadow_resource(ctx, rsc, level, usage, box)) {
+ needs_flush = busy = false;
+ fd_invalidate_resource(ctx, prsc);
+ }
+ }
+
+ if (needs_flush) {
+ if (usage & PIPE_TRANSFER_WRITE) {
+ struct fd_batch *batch, *last_batch = NULL;
+ foreach_batch(batch, &ctx->screen->batch_cache, rsc->batch_mask) {
+ fd_batch_reference(&last_batch, batch);
+ fd_batch_flush(batch, false);
+ }
+ if (last_batch) {
+ fd_batch_sync(last_batch);
+ fd_batch_reference(&last_batch, NULL);
+ }
+ assert(rsc->batch_mask == 0);
+ } else {
+ fd_batch_flush(write_batch, true);
+ }
+ assert(!rsc->write_batch);
+ }
+
+ fd_batch_reference(&write_batch, NULL);
/* The GPU keeps track of how the various bo's are being used, and
* will wait if necessary for the proper operation to have
* completed.
*/
- ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
- if (ret)
- goto fail;
+ if (busy) {
+ ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
+ if (ret)
+ goto fail;
+ }
}
buf = fd_bo_map(rsc->bo);
- if (!buf) {
- fd_resource_transfer_unmap(pctx, ptrans);
- return NULL;
- }
+ if (!buf)
+ goto fail;
offset = slice->offset +
box->y / util_format_get_blockheight(format) * ptrans->stride +
struct pipe_resource *prsc)
{
struct fd_resource *rsc = fd_resource(prsc);
+ fd_bc_invalidate_resource(rsc, true);
if (rsc->bo)
fd_bo_del(rsc->bo);
- list_delinit(&rsc->list);
util_range_destroy(&rsc->valid_buffer_range);
FREE(rsc);
}
.transfer_map = fd_resource_transfer_map,
.transfer_flush_region = fd_resource_transfer_flush_region,
.transfer_unmap = fd_resource_transfer_unmap,
- .transfer_inline_write = u_default_transfer_inline_write,
};
static uint32_t
{
struct pipe_resource *prsc = &rsc->base.b;
enum util_format_layout layout = util_format_description(format)->layout;
+ uint32_t pitchalign = fd_screen(prsc->screen)->gmem_alignw;
uint32_t level, size = 0;
uint32_t width = prsc->width0;
uint32_t height = prsc->height0;
if (layout == UTIL_FORMAT_LAYOUT_ASTC)
slice->pitch = width =
- util_align_npot(width, 32 * util_format_get_blockwidth(format));
+ util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
else
- slice->pitch = width = align(width, 32);
+ slice->pitch = width = align(width, pitchalign);
slice->offset = size;
blocks = util_format_get_nblocks(format, width, height);
/* 1d array and 2d array textures must all have the same layer size
}
}
+/* special case to resize query buf after allocated.. */
+void
+fd_resource_resize(struct pipe_resource *prsc, uint32_t sz)
+{
+ struct fd_resource *rsc = fd_resource(prsc);
+
+ debug_assert(prsc->width0 == 0);
+ debug_assert(prsc->target == PIPE_BUFFER);
+ debug_assert(prsc->bind == PIPE_BIND_QUERY_BUFFER);
+
+ prsc->width0 = sz;
+ realloc_bo(rsc, setup_slices(rsc, 1, prsc->format));
+}
+
/**
* Create a new texture object, using the given template info.
*/
enum pipe_format format = tmpl->format;
uint32_t size, alignment;
- DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
- "nr_samples=%u, usage=%u, bind=%x, flags=%x",
+ DBG("%p: target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
+ "nr_samples=%u, usage=%u, bind=%x, flags=%x", prsc,
tmpl->target, util_format_name(format),
tmpl->width0, tmpl->height0, tmpl->depth0,
tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
*prsc = *tmpl;
pipe_reference_init(&prsc->reference, 1);
- list_inithead(&rsc->list);
+
prsc->screen = pscreen;
util_range_init(&rsc->valid_buffer_range);
assert(rsc->cpp);
alignment = slice_alignment(pscreen, tmpl);
- if (is_a4xx(fd_screen(pscreen))) {
+ if (is_a4xx(fd_screen(pscreen)) || is_a5xx(fd_screen(pscreen))) {
switch (tmpl->target) {
case PIPE_TEXTURE_3D:
rsc->layer_first = false;
size = setup_slices(rsc, alignment, format);
+ /* special case for hw-query buffer, which we need to allocate before we
+ * know the size:
+ */
+ if (size == 0) {
+ /* note, semi-intention == instead of & */
+ debug_assert(prsc->bind == PIPE_BIND_QUERY_BUFFER);
+ return prsc;
+ }
+
if (rsc->layer_first) {
rsc->layer_size = align(size, 4096);
size = rsc->layer_size * prsc->array_size;
static struct pipe_resource *
fd_resource_from_handle(struct pipe_screen *pscreen,
const struct pipe_resource *tmpl,
- struct winsys_handle *handle)
+ struct winsys_handle *handle, unsigned usage)
{
struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
struct fd_resource_slice *slice = &rsc->slices[0];
struct pipe_resource *prsc = &rsc->base.b;
+ uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw;
DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
"nr_samples=%u, usage=%u, bind=%x, flags=%x",
*prsc = *tmpl;
pipe_reference_init(&prsc->reference, 1);
- list_inithead(&rsc->list);
+
prsc->screen = pscreen;
util_range_init(&rsc->valid_buffer_range);
- rsc->bo = fd_screen_bo_from_handle(pscreen, handle, &slice->pitch);
+ rsc->bo = fd_screen_bo_from_handle(pscreen, handle);
if (!rsc->bo)
goto fail;
rsc->base.vtbl = &fd_resource_vtbl;
rsc->cpp = util_format_get_blocksize(tmpl->format);
- slice->pitch /= rsc->cpp;
+ slice->pitch = handle->stride / rsc->cpp;
+ slice->offset = handle->offset;
+ slice->size0 = handle->stride * prsc->height0;
+
+ if ((slice->pitch < align(prsc->width0, pitchalign)) ||
+ (slice->pitch & (pitchalign - 1)))
+ goto fail;
assert(rsc->cpp);
return NULL;
}
-static void fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond);
-static void fd_blitter_pipe_end(struct fd_context *ctx);
-
/**
* _copy_region using pipe (3d engine)
*/
if (!util_blitter_is_copy_supported(ctx->blitter, dst, src))
return false;
- fd_blitter_pipe_begin(ctx, false);
+ /* TODO we could discard if dst box covers dst level fully.. */
+ fd_blitter_pipe_begin(ctx, false, false, FD_STAGE_BLIT);
util_blitter_copy_texture(ctx->blitter,
dst, dst_level, dstx, dsty, dstz,
src, src_level, src_box);
{
struct fd_context *ctx = fd_context(pctx);
struct pipe_blit_info info = *blit_info;
+ bool discard = false;
if (info.src.resource->nr_samples > 1 &&
info.dst.resource->nr_samples <= 1 &&
if (info.render_condition_enable && !fd_render_condition_check(pctx))
return;
+ if (!info.scissor_enable && !info.alpha_blend) {
+ discard = util_texrange_covers_whole_level(info.dst.resource,
+ info.dst.level, info.dst.box.x, info.dst.box.y,
+ info.dst.box.z, info.dst.box.width,
+ info.dst.box.height, info.dst.box.depth);
+ }
+
if (util_try_blit_via_copy_region(pctx, &info)) {
return; /* done */
}
return;
}
- fd_blitter_pipe_begin(ctx, info.render_condition_enable);
+ fd_blitter_pipe_begin(ctx, info.render_condition_enable, discard, FD_STAGE_BLIT);
util_blitter_blit(ctx->blitter, &info);
fd_blitter_pipe_end(ctx);
}
-static void
-fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond)
+void
+fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond, bool discard,
+ enum fd_render_stage stage)
{
+ util_blitter_save_fragment_constant_buffer_slot(ctx->blitter,
+ ctx->constbuf[PIPE_SHADER_FRAGMENT].cb);
util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb);
util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx);
util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp);
util_blitter_save_depth_stencil_alpha(ctx->blitter, ctx->zsa);
util_blitter_save_stencil_ref(ctx->blitter, &ctx->stencil_ref);
util_blitter_save_sample_mask(ctx->blitter, ctx->sample_mask);
- util_blitter_save_framebuffer(ctx->blitter, &ctx->framebuffer);
+ util_blitter_save_framebuffer(ctx->blitter,
+ ctx->batch ? &ctx->batch->framebuffer : NULL);
util_blitter_save_fragment_sampler_states(ctx->blitter,
ctx->fragtex.num_samplers,
(void **)ctx->fragtex.samplers);
util_blitter_save_render_condition(ctx->blitter,
ctx->cond_query, ctx->cond_cond, ctx->cond_mode);
- fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_BLIT);
+ if (ctx->batch)
+ fd_hw_query_set_stage(ctx->batch, ctx->batch->draw, stage);
+
+ ctx->in_blit = discard;
}
-static void
+void
fd_blitter_pipe_end(struct fd_context *ctx)
{
- fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
+ if (ctx->batch)
+ fd_hw_query_set_stage(ctx->batch, ctx->batch->draw, FD_STAGE_NULL);
+ ctx->in_blit = false;
}
static void
{
struct fd_resource *rsc = fd_resource(prsc);
- if (pending(rsc, FD_PENDING_WRITE | FD_PENDING_READ))
- fd_context_render(pctx);
+ if (rsc->write_batch)
+ fd_batch_flush(rsc->write_batch, true);
+
+ assert(!rsc->write_batch);
}
void
pctx->transfer_map = u_transfer_map_vtbl;
pctx->transfer_flush_region = u_transfer_flush_region_vtbl;
pctx->transfer_unmap = u_transfer_unmap_vtbl;
- pctx->transfer_inline_write = u_transfer_inline_write_vtbl;
+ pctx->buffer_subdata = u_default_buffer_subdata;
+ pctx->texture_subdata = u_default_texture_subdata;
pctx->create_surface = fd_create_surface;
pctx->surface_destroy = fd_surface_destroy;
pctx->resource_copy_region = fd_resource_copy_region;