iris: chaining not growing

This commit is contained in:
Kenneth Graunke 2018-05-02 19:54:23 -07:00
parent 053fb51125
commit 604a1a1614
4 changed files with 93 additions and 94 deletions

View file

@ -37,17 +37,13 @@
#define FILE_DEBUG_FLAG DEBUG_BUFMGR
/**
* Target sizes of the batch and state buffers. We create the initial
* buffers at these sizes, and flush when they're nearly full. If we
* underestimate how close we are to the end, and suddenly need more space
* in the middle of a draw, we can grow the buffers, and finish the draw.
* At that point, we'll be over our target size, so the next operation
* should flush. Each time we flush the batch, we recreate both buffers
* at the original target size, so it doesn't grow without bound.
*/
#define BATCH_SZ (20 * 1024)
#define STATE_SZ (18 * 1024)
/* Terminating the batch takes either 4 bytes for MI_BATCH_BUFFER_END
* or 12 bytes for MI_BATCH_BUFFER_START (when chaining). Plus, we may
* need an extra 4 bytes to pad out to the nearest QWord. So reserve 16.
*/
#define BATCH_RESERVED 16
static void decode_batch(struct iris_batch *batch);
@ -109,17 +105,6 @@ uint_key_hash(const void *key)
return (uintptr_t) key;
}
static void
create_batch_buffer(struct iris_bufmgr *bufmgr,
struct iris_batch_buffer *buf,
const char *name, unsigned size)
{
buf->bo = iris_bo_alloc(bufmgr, name, size, IRIS_MEMZONE_OTHER);
buf->bo->kflags |= EXEC_OBJECT_CAPTURE;
buf->map = iris_bo_map(NULL, buf->bo, MAP_READ | MAP_WRITE);
buf->map_next = buf->map;
}
void
iris_init_batch(struct iris_batch *batch,
struct iris_screen *screen,
@ -208,21 +193,32 @@ add_exec_bo(struct iris_batch *batch, struct iris_bo *bo)
}
static void
iris_batch_reset(struct iris_batch *batch)
create_batch(struct iris_batch *batch)
{
struct iris_screen *screen = batch->screen;
struct iris_bufmgr *bufmgr = screen->bufmgr;
if (batch->last_cmd_bo != NULL) {
iris_bo_unreference(batch->last_cmd_bo);
batch->last_cmd_bo = NULL;
batch->bo = iris_bo_alloc(bufmgr, "command buffer",
BATCH_SZ + BATCH_RESERVED, IRIS_MEMZONE_OTHER);
batch->bo->kflags |= EXEC_OBJECT_CAPTURE;
batch->map = iris_bo_map(NULL, batch->bo, MAP_READ | MAP_WRITE);
batch->map_next = batch->map;
add_exec_bo(batch, batch->bo);
}
static void
iris_batch_reset(struct iris_batch *batch)
{
if (batch->last_bo != NULL) {
iris_bo_unreference(batch->last_bo);
batch->last_bo = NULL;
}
batch->last_cmd_bo = batch->cmdbuf.bo;
batch->last_bo = batch->bo;
batch->primary_batch_size = 0;
create_batch_buffer(bufmgr, &batch->cmdbuf, "command buffer", BATCH_SZ);
add_exec_bo(batch, batch->cmdbuf.bo);
assert(batch->cmdbuf.bo->index == 0);
create_batch(batch);
assert(batch->bo->index == 0);
if (batch->state_sizes)
_mesa_hash_table_clear(batch->state_sizes, NULL);
@ -230,15 +226,6 @@ iris_batch_reset(struct iris_batch *batch)
iris_cache_sets_clear(batch);
}
static void
free_batch_buffer(struct iris_batch_buffer *buf)
{
iris_bo_unreference(buf->bo);
buf->bo = NULL;
buf->map = NULL;
buf->map_next = NULL;
}
void
iris_batch_free(struct iris_batch *batch)
{
@ -247,9 +234,12 @@ iris_batch_free(struct iris_batch *batch)
}
free(batch->exec_bos);
free(batch->validation_list);
free_batch_buffer(&batch->cmdbuf);
iris_bo_unreference(batch->bo);
batch->bo = NULL;
batch->map = NULL;
batch->map_next = NULL;
iris_bo_unreference(batch->last_cmd_bo);
iris_bo_unreference(batch->last_bo);
_mesa_hash_table_destroy(batch->cache.render, NULL);
_mesa_set_destroy(batch->cache.depth, NULL);
@ -261,41 +251,51 @@ iris_batch_free(struct iris_batch *batch)
}
static unsigned
buffer_bytes_used(struct iris_batch_buffer *buf)
batch_bytes_used(struct iris_batch *batch)
{
return buf->map_next - buf->map;
return batch->map_next - batch->map;
}
static void
require_buffer_space(struct iris_batch *batch,
struct iris_batch_buffer *buf,
unsigned size,
unsigned flush_threshold,
unsigned max_buffer_size)
/**
* If we've chained to a secondary batch, or are getting near to the end,
* then flush. This should only be called between draws.
*/
void
iris_batch_maybe_flush(struct iris_batch *batch, unsigned estimate)
{
const unsigned required_bytes = buffer_bytes_used(buf) + size;
if (!batch->no_wrap && required_bytes >= flush_threshold) {
if (batch->bo != batch->exec_bos[0] ||
batch_bytes_used(batch) + estimate >= BATCH_SZ) {
iris_batch_flush(batch);
} else if (required_bytes >= buf->bo->size) {
assert(!"Can't grow");
}
}
void
iris_require_command_space(struct iris_batch *batch, unsigned size)
{
require_buffer_space(batch, &batch->cmdbuf, size, BATCH_SZ, MAX_BATCH_SIZE);
const unsigned required_bytes = batch_bytes_used(batch) + size;
if (required_bytes >= BATCH_SZ) {
/* No longer held by batch->bo, still held by validation list */
iris_bo_unreference(batch->bo);
batch->primary_batch_size = batch_bytes_used(batch);
const uint32_t MI_BATCH_BUFFER_START = (0x31 << 23) | (1 << 8);
uint32_t *cmd = batch->map += sizeof(uint32_t);
uint64_t *addr = batch->map += sizeof(uint64_t);
create_batch(batch);
*cmd = MI_BATCH_BUFFER_START;
*addr = batch->bo->gtt_offset;
}
}
void *
iris_get_command_space(struct iris_batch *batch, unsigned bytes)
{
iris_require_command_space(batch, bytes);
void *map = batch->cmdbuf.map_next;
batch->cmdbuf.map_next += bytes;
return map;
return batch->map_next += bytes;
}
void
@ -315,8 +315,6 @@ iris_batch_emit(struct iris_batch *batch, const void *data, unsigned size)
static void
iris_finish_batch(struct iris_batch *batch)
{
batch->no_wrap = true;
// XXX: ISP DIS
/* Emit MI_BATCH_BUFFER_END to finish our batch. Note that execbuf2
@ -324,16 +322,14 @@ iris_finish_batch(struct iris_batch *batch)
* necessary by emitting an extra MI_NOOP after the end.
*/
const uint32_t MI_BATCH_BUFFER_END_AND_NOOP[2] = { (0xA << 23), 0 };
const bool qword_aligned = (buffer_bytes_used(&batch->cmdbuf) % 8) == 0;
const bool qword_aligned = (batch_bytes_used(batch) % 8) == 0;
iris_batch_emit(batch, MI_BATCH_BUFFER_END_AND_NOOP, qword_aligned ? 8 : 4);
batch->no_wrap = false;
}
static int
submit_batch(struct iris_batch *batch, int in_fence_fd, int *out_fence_fd)
{
iris_bo_unmap(batch->cmdbuf.bo);
iris_bo_unmap(batch->bo);
/* The requirement for using I915_EXEC_NO_RELOC are:
*
@ -351,7 +347,8 @@ submit_batch(struct iris_batch *batch, int in_fence_fd, int *out_fence_fd)
.buffers_ptr = (uintptr_t) batch->validation_list,
.buffer_count = batch->exec_count,
.batch_start_offset = 0,
.batch_len = buffer_bytes_used(&batch->cmdbuf),
.batch_len = batch->bo == batch->exec_bos[0] ? batch_bytes_used(batch)
: batch->primary_batch_size,
.flags = batch->ring |
I915_EXEC_NO_RELOC |
I915_EXEC_BATCH_FIRST |
@ -405,16 +402,15 @@ _iris_batch_flush_fence(struct iris_batch *batch,
int in_fence_fd, int *out_fence_fd,
const char *file, int line)
{
if (buffer_bytes_used(&batch->cmdbuf) == 0)
if (batch_bytes_used(batch) == 0)
return 0;
/* Check that we didn't just wrap our batchbuffer at a bad time. */
assert(!batch->no_wrap);
iris_finish_batch(batch);
if (unlikely(INTEL_DEBUG & (DEBUG_BATCH | DEBUG_SUBMIT))) {
int bytes_for_commands = buffer_bytes_used(&batch->cmdbuf);
int bytes_for_commands = batch_bytes_used(batch);
if (batch->bo != batch->exec_bos[0])
bytes_for_commands += batch->primary_batch_size;
fprintf(stderr, "%19s:%-3d: Batchbuffer flush with %5db (%0.1f%%), "
"%4d BOs (%0.1fMb aperture)\n",
file, line,
@ -424,8 +420,9 @@ _iris_batch_flush_fence(struct iris_batch *batch,
dump_validation_list(batch);
}
if (unlikely(INTEL_DEBUG & DEBUG_BATCH))
if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
decode_batch(batch);
}
int ret = submit_batch(batch, in_fence_fd, out_fence_fd);
@ -439,7 +436,7 @@ _iris_batch_flush_fence(struct iris_batch *batch,
if (unlikely(INTEL_DEBUG & DEBUG_SYNC)) {
dbg_printf("waiting for idle\n");
iris_bo_wait_rendering(batch->cmdbuf.bo);
iris_bo_wait_rendering(batch->bo);
}
/* Clean up after the batch we submitted and prepare for a new one. */
@ -486,7 +483,12 @@ iris_use_pinned_bo(struct iris_batch *batch,
static void
decode_batch(struct iris_batch *batch)
{
gen_print_batch(&batch->decoder, batch->cmdbuf.map,
buffer_bytes_used(&batch->cmdbuf),
batch->cmdbuf.bo->gtt_offset);
if (batch->bo != batch->exec_bos[0]) {
void *map = iris_bo_map(batch->dbg, batch->exec_bos[0], MAP_READ);
gen_print_batch(&batch->decoder, map, batch->primary_batch_size,
batch->exec_bos[0]->gtt_offset);
}
gen_print_batch(&batch->decoder, batch->map, batch_bytes_used(batch),
batch->bo->gtt_offset);
}

View file

@ -38,30 +38,27 @@ struct iris_address {
bool write;
};
struct iris_batch_buffer {
struct iris_bo *bo;
void *map;
void *map_next;
};
struct iris_batch {
struct iris_screen *screen;
struct iris_vtable *vtbl;
struct pipe_debug_callback *dbg;
/** Current batchbuffer being queued up. */
struct iris_batch_buffer cmdbuf;
struct iris_bo *bo;
void *map;
void *map_next;
/** Size of the primary batch if we've moved on to a secondary. */
unsigned primary_batch_size;
/** Last BO submitted to the hardware. Used for glFinish(). */
struct iris_bo *last_cmd_bo;
struct iris_bo *last_bo;
uint32_t hw_ctx_id;
/** Which ring this batch targets - a I915_EXEC_RING_MASK value */
uint8_t ring;
bool no_wrap;
/** The validation list */
struct drm_i915_gem_exec_object2 *validation_list;
struct iris_bo **exec_bos;
@ -101,6 +98,7 @@ void iris_init_batch(struct iris_batch *batch,
struct pipe_debug_callback *dbg,
uint8_t ring);
void iris_batch_free(struct iris_batch *batch);
void iris_batch_maybe_flush(struct iris_batch *batch, unsigned estimate);
void iris_require_command_space(struct iris_batch *batch, unsigned size);
void *iris_get_command_space(struct iris_batch *batch, unsigned bytes);
void iris_batch_emit(struct iris_batch *batch, const void *data, unsigned size);

View file

@ -274,8 +274,6 @@ iris_blorp_exec(struct blorp_batch *blorp_batch,
iris_cache_flush_for_depth(batch, params->stencil.addr.buffer);
iris_require_command_space(batch, 1400);
//iris_require_statebuffer_space(ice, 600); // XXX: THIS. Need this.
batch->no_wrap = true;
// XXX: Emit L3 state
@ -287,8 +285,6 @@ iris_blorp_exec(struct blorp_batch *blorp_batch,
blorp_exec(blorp_batch, params);
batch->no_wrap = false;
// XXX: aperture checks?
/* We've smashed all state compared to what the normal 3D pipeline

View file

@ -35,21 +35,24 @@ void
iris_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
{
struct iris_context *ice = (struct iris_context *) ctx;
struct iris_batch *batch = &ice->render_batch;
iris_batch_maybe_flush(batch, 1500);
// XXX: actually do brw_cache_flush_for_*
iris_emit_pipe_control_flush(&ice->render_batch,
iris_emit_pipe_control_flush(batch,
PIPE_CONTROL_DEPTH_CACHE_FLUSH |
PIPE_CONTROL_RENDER_TARGET_FLUSH |
PIPE_CONTROL_CS_STALL);
iris_emit_pipe_control_flush(&ice->render_batch,
iris_emit_pipe_control_flush(batch,
PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
PIPE_CONTROL_CONST_CACHE_INVALIDATE);
iris_cache_sets_clear(&ice->render_batch);
iris_cache_sets_clear(batch);
// XXX: ^^^
iris_update_compiled_shaders(ice);
ice->vtbl.upload_render_state(ice, &ice->render_batch, info);
ice->vtbl.upload_render_state(ice, batch, info);
}