/**
* Go through the entire state and see if the resource is bound
* anywhere. If it is, mark the relevant state as dirty. This is
- * called on realloc_bo to ensure the neccessary state is re-
+ * called on realloc_bo to ensure the necessary state is re-
* emitted so the GPU looks at the new backing bo.
*/
static void
-rebind_resource(struct fd_context *ctx, struct pipe_resource *prsc)
+rebind_resource_in_ctx(struct fd_context *ctx, struct fd_resource *rsc)
{
+ struct pipe_resource *prsc = &rsc->base;
+
/* VBOs */
- for (unsigned i = 0; i < ctx->vtx.vertexbuf.count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
- if (ctx->vtx.vertexbuf.vb[i].buffer.resource == prsc)
- ctx->dirty |= FD_DIRTY_VTXBUF;
+ if (rsc->dirty & FD_DIRTY_VTXBUF) {
+ struct fd_vertexbuf_stateobj *vb = &ctx->vtx.vertexbuf;
+ for (unsigned i = 0; i < vb->count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
+ if (vb->vb[i].buffer.resource == prsc)
+ ctx->dirty |= FD_DIRTY_VTXBUF;
+ }
}
+ const enum fd_dirty_3d_state per_stage_dirty =
+ FD_DIRTY_CONST | FD_DIRTY_TEX | FD_DIRTY_IMAGE | FD_DIRTY_SSBO;
+
+ if (!(rsc->dirty & per_stage_dirty))
+ return;
+
/* per-shader-stage resources: */
for (unsigned stage = 0; stage < PIPE_SHADER_TYPES; stage++) {
/* Constbufs.. note that constbuf[0] is normal uniforms emitted in
* cmdstream rather than by pointer..
*/
- const unsigned num_ubos = util_last_bit(ctx->constbuf[stage].enabled_mask);
- for (unsigned i = 1; i < num_ubos; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_CONST)
- break;
- if (ctx->constbuf[stage].cb[i].buffer == prsc)
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_CONST;
+ if ((rsc->dirty & FD_DIRTY_CONST) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_CONST)) {
+ struct fd_constbuf_stateobj *cb = &ctx->constbuf[stage];
+ const unsigned num_ubos = util_last_bit(cb->enabled_mask);
+ for (unsigned i = 1; i < num_ubos; i++) {
+ if (cb->cb[i].buffer == prsc) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_CONST;
+ ctx->dirty |= FD_DIRTY_CONST;
+ break;
+ }
+ }
}
/* Textures */
- for (unsigned i = 0; i < ctx->tex[stage].num_textures; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_TEX)
- break;
- if (ctx->tex[stage].textures[i] && (ctx->tex[stage].textures[i]->texture == prsc))
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
+ if ((rsc->dirty & FD_DIRTY_TEX) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_TEX)) {
+ struct fd_texture_stateobj *tex = &ctx->tex[stage];
+ for (unsigned i = 0; i < tex->num_textures; i++) {
+ if (tex->textures[i] && (tex->textures[i]->texture == prsc)) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
+ ctx->dirty |= FD_DIRTY_TEX;
+ break;
+ }
+ }
}
/* Images */
- const unsigned num_images = util_last_bit(ctx->shaderimg[stage].enabled_mask);
- for (unsigned i = 0; i < num_images; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_IMAGE)
- break;
- if (ctx->shaderimg[stage].si[i].resource == prsc)
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE;
+ if ((rsc->dirty & FD_DIRTY_IMAGE) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_IMAGE)) {
+ struct fd_shaderimg_stateobj *si = &ctx->shaderimg[stage];
+ const unsigned num_images = util_last_bit(si->enabled_mask);
+ for (unsigned i = 0; i < num_images; i++) {
+ if (si->si[i].resource == prsc) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE;
+ ctx->dirty |= FD_DIRTY_IMAGE;
+ break;
+ }
+ }
}
/* SSBOs */
- const unsigned num_ssbos = util_last_bit(ctx->shaderbuf[stage].enabled_mask);
- for (unsigned i = 0; i < num_ssbos; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_SSBO)
- break;
- if (ctx->shaderbuf[stage].sb[i].buffer == prsc)
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_SSBO;
+ if ((rsc->dirty & FD_DIRTY_SSBO) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_SSBO)) {
+ struct fd_shaderbuf_stateobj *sb = &ctx->shaderbuf[stage];
+ const unsigned num_ssbos = util_last_bit(sb->enabled_mask);
+ for (unsigned i = 0; i < num_ssbos; i++) {
+ if (sb->sb[i].buffer == prsc) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_SSBO;
+ ctx->dirty |= FD_DIRTY_SSBO;
+ break;
+ }
+ }
}
}
}
+static void
+rebind_resource(struct fd_resource *rsc)
+{
+ struct fd_screen *screen = fd_screen(rsc->base.screen);
+
+ mtx_lock(&screen->lock);
+ fd_resource_lock(rsc);
+
+ if (rsc->dirty)
+ list_for_each_entry (struct fd_context, ctx, &screen->context_list, node)
+ rebind_resource_in_ctx(ctx, rsc);
+
+ fd_resource_unlock(rsc);
+ mtx_unlock(&screen->lock);
+}
+
static void
realloc_bo(struct fd_resource *rsc, uint32_t size)
{
/* shadow should not fail in any cases where we need to uncompress: */
debug_assert(success);
- /*
- * TODO what if rsc is used in other contexts, we don't currently
- * have a good way to rebind_resource() in other contexts. And an
- * app that is reading one resource in multiple contexts, isn't
- * going to expect that the resource is modified.
- *
- * Hopefully the edge cases where we need to uncompress are rare
- * enough that they mostly only show up in deqp.
- */
-
- rebind_resource(ctx, &rsc->base);
+ rebind_resource(rsc);
}
static struct fd_resource *
ptrans->level = level;
ptrans->usage = usage;
ptrans->box = *box;
- ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->layout.cpp;
+ ptrans->stride = slice->pitch;
ptrans->layer_stride = fd_resource_layer_stride(rsc, level);
/* we always need a staging texture for tiled buffers:
fd_resource_slice(staging_rsc, 0);
// TODO for PIPE_TRANSFER_READ, need to do untiling blit..
trans->staging_prsc = &staging_rsc->base;
- trans->base.stride = util_format_get_nblocksx(format,
- staging_slice->pitch) * staging_rsc->layout.cpp;
+ trans->base.stride = staging_slice->pitch;
trans->base.layer_stride = fd_resource_layer_stride(staging_rsc, 0);
trans->staging_box = *box;
trans->staging_box.x = 0;
if (usage & PIPE_TRANSFER_WRITE)
op |= DRM_FREEDRENO_PREP_WRITE;
+ bool needs_flush = pending(rsc, !!(usage & PIPE_TRANSFER_WRITE));
+
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
- realloc_bo(rsc, fd_bo_size(rsc->bo));
- rebind_resource(ctx, prsc);
+ if (needs_flush || fd_resource_busy(rsc, op)) {
+ realloc_bo(rsc, fd_bo_size(rsc->bo));
+ rebind_resource(rsc);
+ }
} else if ((usage & PIPE_TRANSFER_WRITE) &&
prsc->target == PIPE_BUFFER &&
!util_ranges_intersect(&rsc->valid_buffer_range,
/* 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.
*/
- bool needs_flush = pending(rsc, !!(usage & PIPE_TRANSFER_WRITE));
- bool busy = needs_flush || (0 != fd_bo_cpu_prep(rsc->bo,
- ctx->pipe, op | DRM_FREEDRENO_PREP_NOSYNC));
+ bool busy = needs_flush || fd_resource_busy(rsc, op);
/* if we need to flush/stall, see if we can make a shadow buffer
* to avoid this:
if (needs_flush && fd_try_shadow_resource(ctx, rsc, level,
box, DRM_FORMAT_MOD_LINEAR)) {
needs_flush = busy = false;
- rebind_resource(ctx, prsc);
+ rebind_resource(rsc);
ctx->stats.shadow_uploads++;
} else {
struct fd_resource *staging_rsc;
struct fdl_slice *staging_slice =
fd_resource_slice(staging_rsc, 0);
trans->staging_prsc = &staging_rsc->base;
- trans->base.stride = util_format_get_nblocksx(format,
- staging_slice->pitch) * staging_rsc->layout.cpp;
+ trans->base.stride = staging_slice->pitch;
trans->base.layer_stride =
fd_resource_layer_stride(staging_rsc, 0);
trans->staging_box = *box;
renderonly_scanout_destroy(rsc->scanout, fd_screen(pscreen)->ro);
util_range_destroy(&rsc->valid_buffer_range);
+ simple_mtx_destroy(&rsc->lock);
FREE(rsc);
}
handle->modifier = fd_resource_modifier(rsc);
return fd_screen_bo_get_handle(pscreen, rsc->bo, rsc->scanout,
- fd_resource_slice(rsc, 0)->pitch * rsc->layout.cpp, handle);
+ fd_resource_slice(rsc, 0)->pitch, handle);
}
static uint32_t
uint32_t blocks;
if (layout == UTIL_FORMAT_LAYOUT_ASTC)
- slice->pitch = width =
- util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
+ width = util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
else
- slice->pitch = width = align(width, pitchalign);
+ width = align(width, pitchalign);
+ slice->pitch = util_format_get_nblocksx(format, width) * rsc->layout.cpp;
slice->offset = size;
blocks = util_format_get_nblocks(format, width, height);
/* 1d array and 2d array textures must all have the same layer size
util_range_init(&rsc->valid_buffer_range);
+ simple_mtx_init(&rsc->lock, mtx_plain);
+
rsc->internal_format = format;
rsc->layout.ubwc = rsc->layout.tile_mode && is_a6xx(screen) && allow_ubwc;
struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
struct fdl_slice *slice = fd_resource_slice(rsc, 0);
struct pipe_resource *prsc = &rsc->base;
- uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw;
+ uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw * rsc->layout.cpp;
DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
"nr_samples=%u, usage=%u, bind=%x, flags=%x",
util_range_init(&rsc->valid_buffer_range);
+ simple_mtx_init(&rsc->lock, mtx_plain);
+
rsc->bo = fd_screen_bo_from_handle(pscreen, handle);
if (!rsc->bo)
goto fail;
rsc->internal_format = tmpl->format;
- slice->pitch = handle->stride / rsc->layout.cpp;
+ slice->pitch = handle->stride;
slice->offset = handle->offset;
slice->size0 = handle->stride * prsc->height0;
- if ((slice->pitch < align(prsc->width0, pitchalign)) ||
+ if ((slice->pitch < align(prsc->width0 * rsc->layout.cpp, pitchalign)) ||
(slice->pitch & (pitchalign - 1)))
goto fail;