/**
* Allocates a block of space in the batchbuffer for indirect state.
*
* We don't want to allocate separate BOs for every bit of indirect
* state in the driver. It means overallocating by a significant
* margin (4096 bytes, even if the object is just a 20-byte surface
* state), and more buffers to walk and count for aperture size checking.
*
* However, due to the restrictions imposed by the aperture size
* checking performance hacks, we can't have the batch point at a
* separate indirect state buffer, because once the batch points at
* it, no more relocations can be added to it. So, we sneak these
* buffers in at the top of the batchbuffer.
*/
void *
brw_state_batch(struct brw_context *brw,
int size,
int alignment,
uint32_t *out_offset)
{
struct intel_batchbuffer *batch = &brw->batch;
uint32_t offset;
assert(size < batch->bo->size);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
/* If allocating from the top would wrap below the batchbuffer, or
* if the batch's used space (plus the reserved pad) collides with our
* space, then flush and try again.
*/
if (batch->state_batch_offset < size ||
offset < 4 * USED_BATCH(*batch) + batch->reserved_space) {
intel_batchbuffer_flush(brw);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
}
batch->state_batch_offset = offset;
if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
_mesa_hash_table_insert(batch->state_batch_sizes,
(void *) (uintptr_t) offset,
(void *) (uintptr_t) size);
}
*out_offset = offset;
return batch->map + (offset>>2);
}
void
brw_blorp_exec(struct brw_context *brw, const brw_blorp_params *params)
{
struct gl_context *ctx = &brw->ctx;
uint32_t estimated_max_batch_usage = 1500;
bool check_aperture_failed_once = false;
/* Flush the sampler and render caches. We definitely need to flush the
* sampler cache so that we get updated contents from the render cache for
* the glBlitFramebuffer() source. Also, we are sometimes warned in the
* docs to flush the cache between reinterpretations of the same surface
* data with different formats, which blorp does for stencil and depth
* data.
*/
brw_emit_mi_flush(brw);
retry:
intel_batchbuffer_require_space(brw, estimated_max_batch_usage, RENDER_RING);
intel_batchbuffer_save_state(brw);
drm_intel_bo *saved_bo = brw->batch.bo;
uint32_t saved_used = USED_BATCH(brw->batch);
uint32_t saved_state_batch_offset = brw->batch.state_batch_offset;
switch (brw->gen) {
case 6:
gen6_blorp_exec(brw, params);
break;
case 7:
gen7_blorp_exec(brw, params);
break;
default:
/* BLORP is not supported before Gen6. */
unreachable("not reached");
}
/* Make sure we didn't wrap the batch unintentionally, and make sure we
* reserved enough space that a wrap will never happen.
*/
assert(brw->batch.bo == saved_bo);
assert((USED_BATCH(brw->batch) - saved_used) * 4 +
(saved_state_batch_offset - brw->batch.state_batch_offset) <
estimated_max_batch_usage);
/* Shut up compiler warnings on release build */
(void)saved_bo;
(void)saved_used;
(void)saved_state_batch_offset;
/* Check if the blorp op we just did would make our batch likely to fail to
* map all the BOs into the GPU at batch exec time later. If so, flush the
* batch and try again with nothing else in the batch.
*/
if (dri_bufmgr_check_aperture_space(&brw->batch.bo, 1)) {
if (!check_aperture_failed_once) {
check_aperture_failed_once = true;
intel_batchbuffer_reset_to_saved(brw);
intel_batchbuffer_flush(brw);
goto retry;
} else {
int ret = intel_batchbuffer_flush(brw);
WARN_ONCE(ret == -ENOSPC,
"i965: blorp emit exceeded available aperture space\n");
}
}
if (unlikely(brw->always_flush_batch))
intel_batchbuffer_flush(brw);
/* We've smashed all state compared to what the normal 3D pipeline
* rendering tracks for GL.
*/
brw->ctx.NewDriverState = ~0ull;
brw->no_depth_or_stencil = false;
brw->ib.type = -1;
/* Flush the sampler cache so any texturing from the destination is
* coherent.
*/
brw_emit_mi_flush(brw);
}
