* Shader programs refer to most resources via integer handles. These are
* indexes (BTIs) into a "Binding Table", which is simply a list of pointers
* to SURFACE_STATE entries. Each shader stage has its own binding table,
- * set by the 3DSTATE_BINDING_TABLE_POINTERS_* commands. Both the binding
- * table itself and the SURFACE_STATEs are relative to Surface State Base
- * Address, so they all live in IRIS_MEMZONE_SURFACE.
+ * set by the 3DSTATE_BINDING_TABLE_POINTERS_* commands. We stream out
+ * binding tables dynamically, storing them in special BOs we call "binders."
*
* Unfortunately, the hardware designers made 3DSTATE_BINDING_TABLE_POINTERS
* only accept a 16-bit pointer. This means that all binding tables have to
* actual SURFACE_STATE entries can live anywhere in the 4GB zone, as the
* binding table entries are full 32-bit pointers.)
*
- * We stream out binding tables dynamically, storing them in a single 64kB
- * "binder" buffer, located at IRIS_BINDER_ADDRESS. Before emitting a draw
- * call, we reserve space for any new binding tables needed by bound shaders.
- * If there is no space, we flush the batch and swap out the binder for a
- * new empty BO.
+ * To handle this, we split a 4GB region of VMA into two memory zones.
+ * IRIS_MEMZONE_BINDER is a small region at the bottom able to hold a few
+ * binder BOs. IRIS_MEMZONE_SURFACE contains the rest of the 4GB, and is
+ * always at a higher address than the binders. This allows us to program
+ * Surface State Base Address to the binder BO's address, and offset the
+ * values in the binding table to account for the base not starting at the
+ * beginning of the 4GB region.
*
- * XXX: This should be fancier. We currently replace the binder with a
- * fresh BO on every batch, which causes the kernel to stall, trying to
- * pin the new buffer at the same memory address as the old one. We ought
- * to avoid this by using a ringbuffer, tracking the busy section of the BO,
- * and cycling back around where possible to avoid replacing it at all costs.
- *
- * XXX: if we do have to flush, we should emit a performance warning.
+ * This does mean that we have to emit STATE_BASE_ADDRESS and stall when
+ * we run out of space in the binder, which hopefully won't happen too often.
*/
#include <stdlib.h>
#define BTP_ALIGNMENT 32
-/**
- * Reserve a block of space in the binder, given the raw size in bytes.
- */
-uint32_t
-iris_binder_reserve(struct iris_batch *batch, unsigned size)
+/* Avoid using offset 0, tools consider it NULL */
+#define INIT_INSERT_POINT BTP_ALIGNMENT
+
+static bool
+binder_has_space(struct iris_binder *binder, unsigned size)
{
- struct iris_binder *binder = &batch->binder;
+ return binder->insert_point + size <= IRIS_BINDER_SIZE;
+}
+
+static void
+binder_realloc(struct iris_context *ice)
+{
+ struct iris_screen *screen = (void *) ice->ctx.screen;
+ struct iris_bufmgr *bufmgr = screen->bufmgr;
+ struct iris_binder *binder = &ice->state.binder;
+
+ uint64_t next_address = IRIS_MEMZONE_BINDER_START;
+
+ if (binder->bo) {
+ /* Place the new binder just after the old binder, unless we've hit the
+ * end of the memory zone...then wrap around to the start again.
+ */
+ next_address = binder->bo->gtt_offset + IRIS_BINDER_SIZE;
+ if (next_address >= IRIS_MEMZONE_SURFACE_START)
+ next_address = IRIS_MEMZONE_BINDER_START;
+
+ iris_bo_unreference(binder->bo);
+ }
- assert(size > 0);
- assert((binder->insert_point % BTP_ALIGNMENT) == 0);
- /* If we can't fit all stages in the binder, flush the batch which
- * will cause us to gain a new empty binder.
+ binder->bo =
+ iris_bo_alloc(bufmgr, "binder", IRIS_BINDER_SIZE, IRIS_MEMZONE_BINDER);
+ binder->bo->gtt_offset = next_address;
+ binder->map = iris_bo_map(NULL, binder->bo, MAP_WRITE);
+ binder->insert_point = INIT_INSERT_POINT;
+
+ /* Allocating a new binder requires changing Surface State Base Address,
+ * which also invalidates all our previous binding tables - each entry
+ * in those tables is an offset from the old base.
+ *
+ * We do this here so that iris_binder_reserve_3d correctly gets a new
+ * larger total_size when making the updated reservation.
*/
- if (binder->insert_point + size > IRIS_BINDER_SIZE)
- iris_batch_flush(batch);
+ ice->state.dirty |= IRIS_DIRTY_RENDER_BUFFER;
+ ice->state.stage_dirty |= IRIS_ALL_STAGE_DIRTY_BINDINGS;
+}
+static uint32_t
+binder_insert(struct iris_binder *binder, unsigned size)
+{
uint32_t offset = binder->insert_point;
- /* It had better fit now. */
- assert(offset + size <= IRIS_BINDER_SIZE);
-
binder->insert_point = align(binder->insert_point + size, BTP_ALIGNMENT);
- iris_use_pinned_bo(batch, binder->bo, false);
-
return offset;
}
+/**
+ * Reserve a block of space in the binder, given the raw size in bytes.
+ */
+uint32_t
+iris_binder_reserve(struct iris_context *ice,
+ unsigned size)
+{
+ struct iris_binder *binder = &ice->state.binder;
+
+ if (!binder_has_space(binder, size))
+ binder_realloc(ice);
+
+ assert(size > 0);
+ return binder_insert(binder, size);
+}
+
/**
* Reserve and record binder space for 3D pipeline shader stages.
*
* calling this command - the new area is uninitialized.
*/
void
-iris_binder_reserve_3d(struct iris_batch *batch,
- struct iris_context *ice)
+iris_binder_reserve_3d(struct iris_context *ice)
{
struct iris_compiled_shader **shaders = ice->shaders.prog;
- struct iris_binder *binder = &batch->binder;
- unsigned total_size = 0;
+ struct iris_binder *binder = &ice->state.binder;
unsigned sizes[MESA_SHADER_STAGES] = {};
+ unsigned total_size;
- for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
- if (!(ice->state.dirty & (IRIS_DIRTY_BINDINGS_VS << stage)))
- continue;
+ /* If nothing is dirty, skip all this. */
+ if (!(ice->state.dirty & IRIS_DIRTY_RENDER_BUFFER) &&
+ !(ice->state.stage_dirty & IRIS_ALL_STAGE_DIRTY_BINDINGS))
+ return;
+ /* Get the binding table sizes for each stage */
+ for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
if (!shaders[stage])
continue;
- const struct brw_stage_prog_data *prog_data =
- (const void *) shaders[stage]->prog_data;
-
- sizes[stage] = align(prog_data->binding_table.size_bytes, BTP_ALIGNMENT);
- total_size += sizes[stage];
+ /* Round up the size so our next table has an aligned starting offset */
+ sizes[stage] = align(shaders[stage]->bt.size_bytes, BTP_ALIGNMENT);
}
- if (total_size == 0)
- return;
+ /* Make space for the new binding tables...this may take two tries. */
+ while (true) {
+ total_size = 0;
+ for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
+ if (ice->state.stage_dirty & (IRIS_STAGE_DIRTY_BINDINGS_VS << stage))
+ total_size += sizes[stage];
+ }
- uint32_t offset = iris_binder_reserve(batch, total_size);
+ assert(total_size < IRIS_BINDER_SIZE);
- /* Assign space and record the current binding table. */
- for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
- if (!(ice->state.dirty & (IRIS_DIRTY_BINDINGS_VS << stage)))
- continue;
+ if (total_size == 0)
+ return;
+
+ if (binder_has_space(binder, total_size))
+ break;
- binder->bt_offset[stage] = sizes[stage] > 0 ? offset : 0;
- offset += sizes[stage];
+ /* It didn't fit. Allocate a new buffer and try again. Note that
+ * this will flag all bindings dirty, which may increase total_size
+ * on the next iteration.
+ */
+ binder_realloc(ice);
}
-}
-/* Avoid using offset 0, tools consider it NULL */
-#define INIT_INSERT_POINT BTP_ALIGNMENT
+ /* Assign space and record the new binding table offsets. */
+ uint32_t offset = binder_insert(binder, total_size);
+
+ for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
+ if (ice->state.stage_dirty & (IRIS_STAGE_DIRTY_BINDINGS_VS << stage)) {
+ binder->bt_offset[stage] = sizes[stage] > 0 ? offset : 0;
+ iris_record_state_size(ice->state.sizes,
+ binder->bo->gtt_offset + offset, sizes[stage]);
+ offset += sizes[stage];
+ }
+ }
+}
void
-iris_init_binder(struct iris_binder *binder, struct iris_bufmgr *bufmgr)
+iris_binder_reserve_compute(struct iris_context *ice)
{
- binder->bo =
- iris_bo_alloc(bufmgr, "binder", IRIS_BINDER_SIZE, IRIS_MEMZONE_BINDER);
- binder->map = iris_bo_map(NULL, binder->bo, MAP_WRITE);
- binder->insert_point = INIT_INSERT_POINT;
+ if (!(ice->state.stage_dirty & IRIS_STAGE_DIRTY_BINDINGS_CS))
+ return;
+
+ struct iris_binder *binder = &ice->state.binder;
+ struct iris_compiled_shader *shader =
+ ice->shaders.prog[MESA_SHADER_COMPUTE];
+
+ unsigned size = shader->bt.size_bytes;
+
+ if (size == 0)
+ return;
+
+ binder->bt_offset[MESA_SHADER_COMPUTE] = iris_binder_reserve(ice, size);
}
-/**
- * Is the binder empty? (If so, old binding table pointers are stale.)
- */
-bool
-iris_binder_is_empty(struct iris_binder *binder)
+void
+iris_init_binder(struct iris_context *ice)
{
- return binder->insert_point <= INIT_INSERT_POINT;
+ memset(&ice->state.binder, 0, sizeof(struct iris_binder));
+ binder_realloc(ice);
}
void
projects / mesa.git / blobdiff