*/
#include "main/imports.h"
+#include "intel_batchbuffer.h"
#include "brw_state.h"
+#include "brw_vs.h"
+#include "brw_wm.h"
+#include "brw_vs.h"
#define FILE_DEBUG_FLAG DEBUG_STATE
return hash;
}
-
-/**
- * Marks a new buffer as being chosen for the given cache id.
- */
-static void
-update_cache_last(struct brw_cache *cache, enum brw_cache_id cache_id,
- drm_intel_bo *bo)
-{
- if (bo == cache->last_bo[cache_id])
- return; /* no change */
-
- drm_intel_bo_unreference(cache->last_bo[cache_id]);
- cache->last_bo[cache_id] = bo;
- drm_intel_bo_reference(cache->last_bo[cache_id]);
- cache->brw->state.dirty.cache |= 1 << cache_id;
-}
-
static int
brw_cache_item_equals(const struct brw_cache_item *a,
const struct brw_cache_item *b)
GLuint size, i;
size = cache->size * 3;
- items = (struct brw_cache_item**) calloc(1, size * sizeof(*items));
+ items = calloc(1, size * sizeof(*items));
for (i = 0; i < cache->size; i++)
for (c = cache->items[i]; c; c = next) {
items[c->hash % size] = c;
}
- FREE(cache->items);
+ free(cache->items);
cache->items = items;
cache->size = size;
}
/**
* Returns the buffer object matching cache_id and key, or NULL.
*/
-drm_intel_bo *
+bool
brw_search_cache(struct brw_cache *cache,
enum brw_cache_id cache_id,
const void *key, GLuint key_size,
- void *aux_return)
+ uint32_t *inout_offset, void *out_aux)
{
+ struct brw_context *brw = cache->brw;
struct brw_cache_item *item;
struct brw_cache_item lookup;
GLuint hash;
item = search_cache(cache, hash, &lookup);
if (item == NULL)
- return NULL;
+ return false;
- if (aux_return)
- *(void **)aux_return = (void *)((char *)item->key + item->key_size);
+ *(void **)out_aux = ((char *)item->key + item->key_size);
- update_cache_last(cache, cache_id, item->bo);
+ if (item->offset != *inout_offset) {
+ brw->state.dirty.cache |= (1 << cache_id);
+ *inout_offset = item->offset;
+ }
- drm_intel_bo_reference(item->bo);
- return item->bo;
+ return true;
}
+static void
+brw_cache_new_bo(struct brw_cache *cache, uint32_t new_size)
+{
+ struct brw_context *brw = cache->brw;
+ drm_intel_bo *new_bo;
+
+ new_bo = drm_intel_bo_alloc(brw->bufmgr, "program cache", new_size, 64);
+
+ /* Copy any existing data that needs to be saved. */
+ if (cache->next_offset != 0) {
+ drm_intel_bo_map(cache->bo, false);
+ drm_intel_bo_subdata(new_bo, 0, cache->next_offset, cache->bo->virtual);
+ drm_intel_bo_unmap(cache->bo);
+ }
+
+ drm_intel_bo_unreference(cache->bo);
+ cache->bo = new_bo;
+ cache->bo_used_by_gpu = false;
+
+ /* Since we have a new BO in place, we need to signal the units
+ * that depend on it (state base address on gen5+, or unit state before).
+ */
+ brw->state.dirty.brw |= BRW_NEW_PROGRAM_CACHE;
+}
+
+/**
+ * Attempts to find an item in the cache with identical data and aux
+ * data to use
+ */
+static bool
+brw_try_upload_using_copy(struct brw_cache *cache,
+ struct brw_cache_item *result_item,
+ const void *data,
+ const void *aux)
+{
+ int i;
+ struct brw_cache_item *item;
+
+ for (i = 0; i < cache->size; i++) {
+ for (item = cache->items[i]; item; item = item->next) {
+ const void *item_aux = item->key + item->key_size;
+ int ret;
+
+ if (item->cache_id != result_item->cache_id ||
+ item->size != result_item->size ||
+ item->aux_size != result_item->aux_size) {
+ continue;
+ }
+
+ if (cache->aux_compare[result_item->cache_id]) {
+ if (!cache->aux_compare[result_item->cache_id](item_aux, aux,
+ item->aux_size,
+ item->key))
+ continue;
+ } else if (memcmp(item_aux, aux, item->aux_size) != 0) {
+ continue;
+ }
+
+ drm_intel_bo_map(cache->bo, false);
+ ret = memcmp(cache->bo->virtual + item->offset, data, item->size);
+ drm_intel_bo_unmap(cache->bo);
+ if (ret)
+ continue;
+
+ result_item->offset = item->offset;
+
+ return true;
+ }
+ }
+
+ return false;
+}
+
+static void
+brw_upload_item_data(struct brw_cache *cache,
+ struct brw_cache_item *item,
+ const void *data)
+{
+ /* Allocate space in the cache BO for our new program. */
+ if (cache->next_offset + item->size > cache->bo->size) {
+ uint32_t new_size = cache->bo->size * 2;
+
+ while (cache->next_offset + item->size > new_size)
+ new_size *= 2;
-drm_intel_bo *
+ brw_cache_new_bo(cache, new_size);
+ }
+
+ /* If we would block on writing to an in-use program BO, just
+ * recreate it.
+ */
+ if (cache->bo_used_by_gpu) {
+ brw_cache_new_bo(cache, cache->bo->size);
+ }
+
+ item->offset = cache->next_offset;
+
+ /* Programs are always 64-byte aligned, so set up the next one now */
+ cache->next_offset = ALIGN(item->offset + item->size, 64);
+}
+
+void
brw_upload_cache(struct brw_cache *cache,
enum brw_cache_id cache_id,
const void *key,
GLuint data_size,
const void *aux,
GLuint aux_size,
- void *aux_return)
+ uint32_t *out_offset,
+ void *out_aux)
{
struct brw_cache_item *item = CALLOC_STRUCT(brw_cache_item);
GLuint hash;
void *tmp;
- drm_intel_bo *bo;
item->cache_id = cache_id;
+ item->size = data_size;
item->key = key;
item->key_size = key_size;
+ item->aux_size = aux_size;
hash = hash_key(item);
item->hash = hash;
- /* Create the buffer object to contain the data */
- bo = drm_intel_bo_alloc(cache->brw->intel.bufmgr,
- cache->name[cache_id], data_size, 1 << 6);
-
+ /* If we can find a matching prog/prog_data combo in the cache
+ * already, then reuse the existing stuff. This will mean not
+ * flagging CACHE_NEW_* when transitioning between the two
+ * equivalent hash keys. This is notably useful for programs
+ * generating shaders at runtime, where multiple shaders may
+ * compile to the thing in our backend.
+ */
+ if (!brw_try_upload_using_copy(cache, item, data, aux)) {
+ brw_upload_item_data(cache, item, data);
+ }
/* Set up the memory containing the key and aux_data */
tmp = malloc(key_size + aux_size);
item->key = tmp;
- item->bo = bo;
- drm_intel_bo_reference(bo);
-
if (cache->n_items > cache->size * 1.5)
rehash(cache);
cache->items[hash] = item;
cache->n_items++;
- if (aux_return) {
- *(void **)aux_return = (void *)((char *)item->key + item->key_size);
- }
-
- DBG("upload %s: %d bytes to cache id %d\n",
- cache->name[cache_id],
- data_size, cache_id);
-
/* Copy data to the buffer */
- drm_intel_bo_subdata(bo, 0, data_size, data);
+ drm_intel_bo_subdata(cache->bo, item->offset, data_size, data);
- update_cache_last(cache, cache_id, bo);
-
- return bo;
-}
-
-static void
-brw_init_cache_id(struct brw_cache *cache,
- const char *name,
- enum brw_cache_id id)
-{
- cache->name[id] = strdup(name);
+ *out_offset = item->offset;
+ *(void **)out_aux = (void *)((char *)item->key + item->key_size);
+ cache->brw->state.dirty.cache |= 1 << cache_id;
}
-
void
brw_init_caches(struct brw_context *brw)
{
cache->size = 7;
cache->n_items = 0;
- cache->items = (struct brw_cache_item **)
- calloc(1, cache->size * sizeof(struct brw_cache_item));
-
- brw_init_cache_id(cache, "CC_VP", BRW_CC_VP);
- brw_init_cache_id(cache, "CC_UNIT", BRW_CC_UNIT);
- brw_init_cache_id(cache, "WM_PROG", BRW_WM_PROG);
- brw_init_cache_id(cache, "SAMPLER", BRW_SAMPLER);
- brw_init_cache_id(cache, "WM_UNIT", BRW_WM_UNIT);
- brw_init_cache_id(cache, "SF_PROG", BRW_SF_PROG);
- brw_init_cache_id(cache, "SF_VP", BRW_SF_VP);
-
- brw_init_cache_id(cache, "SF_UNIT", BRW_SF_UNIT);
-
- brw_init_cache_id(cache, "VS_UNIT", BRW_VS_UNIT);
-
- brw_init_cache_id(cache, "VS_PROG", BRW_VS_PROG);
+ cache->items =
+ calloc(1, cache->size * sizeof(struct brw_cache_item *));
- brw_init_cache_id(cache, "CLIP_UNIT", BRW_CLIP_UNIT);
+ cache->bo = drm_intel_bo_alloc(brw->bufmgr,
+ "program cache",
+ 4096, 64);
- brw_init_cache_id(cache, "CLIP_PROG", BRW_CLIP_PROG);
- brw_init_cache_id(cache, "CLIP_VP", BRW_CLIP_VP);
-
- brw_init_cache_id(cache, "GS_UNIT", BRW_GS_UNIT);
-
- brw_init_cache_id(cache, "GS_PROG", BRW_GS_PROG);
- brw_init_cache_id(cache, "BLEND_STATE", BRW_BLEND_STATE);
- brw_init_cache_id(cache, "COLOR_CALC_STATE", BRW_COLOR_CALC_STATE);
- brw_init_cache_id(cache, "DEPTH_STENCIL_STATE", BRW_DEPTH_STENCIL_STATE);
+ cache->aux_compare[BRW_VS_PROG] = brw_vs_prog_data_compare;
+ cache->aux_compare[BRW_WM_PROG] = brw_wm_prog_data_compare;
+ cache->aux_free[BRW_VS_PROG] = brw_vs_prog_data_free;
+ cache->aux_free[BRW_WM_PROG] = brw_wm_prog_data_free;
}
static void
for (i = 0; i < cache->size; i++) {
for (c = cache->items[i]; c; c = next) {
next = c->next;
- drm_intel_bo_unreference(c->bo);
+ if (cache->aux_free[c->cache_id]) {
+ const void *item_aux = c->key + c->key_size;
+ cache->aux_free[c->cache_id](item_aux);
+ }
free((void *)c->key);
free(c);
}
cache->n_items = 0;
+ /* Start putting programs into the start of the BO again, since
+ * we'll never find the old results.
+ */
+ cache->next_offset = 0;
+
+ /* We need to make sure that the programs get regenerated, since
+ * any offsets leftover in brw_context will no longer be valid.
+ */
brw->state.dirty.mesa |= ~0;
brw->state.dirty.brw |= ~0;
brw->state.dirty.cache |= ~0;
+ intel_batchbuffer_flush(brw);
}
void
brw_state_cache_check_size(struct brw_context *brw)
{
- /* un-tuned guess. Each object is generally a page, so 1000 of them is 4 MB of
+ /* un-tuned guess. Each object is generally a page, so 2000 of them is 8 MB of
* state cache.
*/
- if (brw->cache.n_items > 1000)
+ if (brw->cache.n_items > 2000) {
+ perf_debug("Exceeded state cache size limit. Clearing the set "
+ "of compiled programs, which will trigger recompiles\n");
brw_clear_cache(brw, &brw->cache);
+ }
}
static void
brw_destroy_cache(struct brw_context *brw, struct brw_cache *cache)
{
- GLuint i;
DBG("%s\n", __FUNCTION__);
+ drm_intel_bo_unreference(cache->bo);
+ cache->bo = NULL;
brw_clear_cache(brw, cache);
- for (i = 0; i < BRW_MAX_CACHE; i++) {
- drm_intel_bo_unreference(cache->last_bo[i]);
- free(cache->name[i]);
- }
free(cache->items);
cache->items = NULL;
cache->size = 0;