/*
Copyright (C) Intel Corp. 2006. All Rights Reserved.
- Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
+ Intel funded Tungsten Graphics to
develop this 3D driver.
-
+
Permission is hereby granted, free of charge, to any person obtaining
a copy of this software and associated documentation files (the
"Software"), to deal in the Software without restriction, including
distribute, sublicense, and/or sell copies of the Software, and to
permit persons to whom the Software is furnished to do so, subject to
the following conditions:
-
+
The above copyright notice and this permission notice (including the
next paragraph) shall be included in all copies or substantial
portions of the Software.
-
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
-
+
**********************************************************************/
/*
* Authors:
- * Keith Whitwell <keith@tungstengraphics.com>
+ * Keith Whitwell <keithw@vmware.com>
*/
-
-
-#include "brw_state.h"
-#include "brw_aub.h"
-#include "intel_batchbuffer.h"
-#include "imports.h"
-/* XXX: Fixme - have to include these to get the sizes of the prog_key
- * structs:
+/** @file brw_state_cache.c
+ *
+ * This file implements a simple static state cache for 965. The
+ * consumers can query the hash table of state using a cache_id,
+ * opaque key data, and receive the corresponding state buffer object
+ * of state (plus associated auxiliary data) in return. Objects in
+ * the cache may not have relocations (pointers to other BOs) in them.
+ *
+ * The inner workings are a simple hash table based on a CRC of the
+ * key data.
+ *
+ * Replacement is not implemented. Instead, when the cache gets too
+ * big we throw out all of the cache data and let it get regenerated.
*/
-#include "brw_wm.h"
+
+#include "main/imports.h"
+#include "intel_batchbuffer.h"
+#include "brw_state.h"
#include "brw_vs.h"
-#include "brw_clip.h"
-#include "brw_sf.h"
+#include "brw_wm.h"
#include "brw_gs.h"
+#include "brw_cs.h"
+#define FILE_DEBUG_FLAG DEBUG_STATE
-/***********************************************************************
- * Check cache for uploaded version of struct, else upload new one.
- * Fail when memory is exhausted.
- *
- * XXX: FIXME: Currently search is so slow it would be quicker to
- * regenerate the data every time...
- */
-
-static GLuint hash_key( const void *key, GLuint key_size )
+static GLuint
+hash_key(struct brw_cache_item *item)
{
- GLuint *ikey = (GLuint *)key;
- GLuint hash = 0, i;
+ GLuint *ikey = (GLuint *)item->key;
+ GLuint hash = item->cache_id, i;
- assert(key_size % 4 == 0);
+ assert(item->key_size % 4 == 0);
/* I'm sure this can be improved on:
*/
- for (i = 0; i < key_size/4; i++)
+ for (i = 0; i < item->key_size/4; i++) {
hash ^= ikey[i];
+ hash = (hash << 5) | (hash >> 27);
+ }
return hash;
}
-static struct brw_cache_item *search_cache( struct brw_cache *cache,
- GLuint hash,
- const void *key,
- GLuint key_size)
+static int
+brw_cache_item_equals(const struct brw_cache_item *a,
+ const struct brw_cache_item *b)
+{
+ return a->cache_id == b->cache_id &&
+ a->hash == b->hash &&
+ a->key_size == b->key_size &&
+ (memcmp(a->key, b->key, a->key_size) == 0);
+}
+
+static struct brw_cache_item *
+search_cache(struct brw_cache *cache, GLuint hash,
+ struct brw_cache_item *lookup)
{
struct brw_cache_item *c;
+#if 0
+ int bucketcount = 0;
+
+ for (c = cache->items[hash % cache->size]; c; c = c->next)
+ bucketcount++;
+
+ fprintf(stderr, "bucket %d/%d = %d/%d items\n", hash % cache->size,
+ cache->size, bucketcount, cache->n_items);
+#endif
+
for (c = cache->items[hash % cache->size]; c; c = c->next) {
- if (c->hash == hash &&
- c->key_size == key_size &&
- memcmp(c->key, key, key_size) == 0)
+ if (brw_cache_item_equals(lookup, c))
return c;
}
}
-static void rehash( struct brw_cache *cache )
+static void
+rehash(struct brw_cache *cache)
{
struct brw_cache_item **items;
struct brw_cache_item *c, *next;
GLuint size, i;
size = cache->size * 3;
- items = (struct brw_cache_item**) _mesa_malloc(size * sizeof(*items));
- _mesa_memset(items, 0, size * sizeof(*items));
+ items = calloc(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;
}
-GLboolean brw_search_cache( struct brw_cache *cache,
- const void *key,
- GLuint key_size,
- void *aux_return,
- GLuint *offset_return)
+/**
+ * Returns the buffer object matching cache_id and key, or NULL.
+ */
+bool
+brw_search_cache(struct brw_cache *cache,
+ enum brw_cache_id cache_id,
+ const void *key, GLuint key_size,
+ uint32_t *inout_offset, void *out_aux)
{
+ struct brw_context *brw = cache->brw;
struct brw_cache_item *item;
- GLuint addr = 0;
- GLuint hash = hash_key(key, key_size);
-
- item = search_cache(cache, hash, key, key_size);
-
- if (item) {
- if (aux_return)
- *(void **)aux_return = (void *)((char *)item->key + item->key_size);
-
- *offset_return = addr = item->offset;
- }
-
- if (item == NULL || addr != cache->last_addr) {
- cache->brw->state.dirty.cache |= 1<<cache->id;
- cache->last_addr = addr;
- }
-
- return item != NULL;
-}
+ struct brw_cache_item lookup;
+ GLuint hash;
-GLuint brw_upload_cache( struct brw_cache *cache,
- const void *key,
- GLuint key_size,
- const void *data,
- GLuint data_size,
- const void *aux,
- void *aux_return )
-{
- GLuint offset;
- struct brw_cache_item *item = CALLOC_STRUCT(brw_cache_item);
- GLuint hash = hash_key(key, key_size);
- void *tmp = _mesa_malloc(key_size + cache->aux_size);
-
- if (!brw_pool_alloc(cache->pool, data_size, 6, &offset)) {
- /* Should not be possible:
- */
- _mesa_printf("brw_pool_alloc failed\n");
- exit(1);
- }
+ lookup.cache_id = cache_id;
+ lookup.key = key;
+ lookup.key_size = key_size;
+ hash = hash_key(&lookup);
+ lookup.hash = hash;
- memcpy(tmp, key, key_size);
+ item = search_cache(cache, hash, &lookup);
- if (cache->aux_size)
- memcpy(tmp+key_size, aux, cache->aux_size);
-
- item->key = tmp;
- item->hash = hash;
- item->key_size = key_size;
- item->offset = offset;
- item->data_size = data_size;
+ if (item == NULL)
+ return false;
- if (++cache->n_items > cache->size * 1.5)
- rehash(cache);
-
- hash %= cache->size;
- item->next = cache->items[hash];
- cache->items[hash] = item;
-
- if (aux_return) {
- assert(cache->aux_size);
- *(void **)aux_return = (void *)((char *)item->key + item->key_size);
+ *(void **)out_aux = ((char *)item->key + item->key_size);
+
+ if (item->offset != *inout_offset) {
+ brw->ctx.NewDriverState |= (1 << cache_id);
+ *inout_offset = item->offset;
}
- if (INTEL_DEBUG & DEBUG_STATE)
- _mesa_printf("upload %s: %d bytes to pool buffer %d offset %x\n",
- cache->name,
- data_size,
- cache->pool->buffer,
- offset);
+ return true;
+}
- /* Copy data to the buffer:
- */
- bmBufferSubDataAUB(&cache->brw->intel,
- cache->pool->buffer,
- offset,
- data_size,
- data,
- cache->aub_type,
- cache->aub_sub_type);
-
-
- cache->brw->state.dirty.cache |= 1<<cache->id;
- cache->last_addr = offset;
+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);
+ if (brw->has_llc)
+ drm_intel_gem_bo_map_unsynchronized(new_bo);
+
+ /* Copy any existing data that needs to be saved. */
+ if (cache->next_offset != 0) {
+ if (brw->has_llc) {
+ memcpy(new_bo->virtual, cache->bo->virtual, cache->next_offset);
+ } else {
+ 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);
+ }
+ }
- return offset;
+ if (brw->has_llc)
+ 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->ctx.NewDriverState |= BRW_NEW_PROGRAM_CACHE;
}
-/* This doesn't really work with aux data. Use search/upload instead
+/**
+ * Attempts to find an item in the cache with identical data.
*/
-GLuint brw_cache_data_sz(struct brw_cache *cache,
- const void *data,
- GLuint data_size)
+static const struct brw_cache_item *
+brw_lookup_prog(const struct brw_cache *cache,
+ enum brw_cache_id cache_id,
+ const void *data, unsigned data_size)
{
- GLuint addr;
+ const struct brw_context *brw = cache->brw;
+ unsigned i;
+ const struct brw_cache_item *item;
- if (!brw_search_cache(cache, data, data_size, NULL, &addr)) {
- addr = brw_upload_cache(cache,
- data, data_size,
- data, data_size,
- NULL, NULL);
+ for (i = 0; i < cache->size; i++) {
+ for (item = cache->items[i]; item; item = item->next) {
+ int ret;
+
+ if (item->cache_id != cache_id || item->size != data_size)
+ continue;
+
+ if (!brw->has_llc)
+ drm_intel_bo_map(cache->bo, false);
+ ret = memcmp(cache->bo->virtual + item->offset, data, item->size);
+ if (!brw->has_llc)
+ drm_intel_bo_unmap(cache->bo);
+ if (ret)
+ continue;
+
+ return item;
+ }
}
- return addr;
+ return NULL;
}
-GLuint brw_cache_data(struct brw_cache *cache,
- const void *data)
+static uint32_t
+brw_alloc_item_data(struct brw_cache *cache, uint32_t size)
{
- return brw_cache_data_sz(cache, data, cache->key_size);
-}
+ uint32_t offset;
+ struct brw_context *brw = cache->brw;
+ /* Allocate space in the cache BO for our new program. */
+ if (cache->next_offset + size > cache->bo->size) {
+ uint32_t new_size = cache->bo->size * 2;
+ while (cache->next_offset + size > new_size)
+ new_size *= 2;
+ brw_cache_new_bo(cache, new_size);
+ }
+
+ /* If we would block on writing to an in-use program BO, just
+ * recreate it.
+ */
+ if (!brw->has_llc && cache->bo_used_by_gpu) {
+ perf_debug("Copying busy program cache buffer.\n");
+ brw_cache_new_bo(cache, cache->bo->size);
+ }
+ offset = cache->next_offset;
+
+ /* Programs are always 64-byte aligned, so set up the next one now */
+ cache->next_offset = ALIGN(offset + size, 64);
+
+ return offset;
+}
-static void brw_init_cache( struct brw_context *brw,
- const char *name,
- GLuint id,
- GLuint key_size,
- GLuint aux_size,
- GLuint aub_type,
- GLuint aub_sub_type )
+void
+brw_upload_cache(struct brw_cache *cache,
+ enum brw_cache_id cache_id,
+ const void *key,
+ GLuint key_size,
+ const void *data,
+ GLuint data_size,
+ const void *aux,
+ GLuint aux_size,
+ uint32_t *out_offset,
+ void *out_aux)
{
- struct brw_cache *cache = &brw->cache[id];
- cache->brw = brw;
- cache->id = id;
- cache->name = name;
- cache->items = NULL;
+ struct brw_context *brw = cache->brw;
+ struct brw_cache_item *item = CALLOC_STRUCT(brw_cache_item);
+ const struct brw_cache_item *matching_data =
+ brw_lookup_prog(cache, cache_id, data, data_size);
+ GLuint hash;
+ void *tmp;
+
+ 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;
- cache->size = 7;
- cache->n_items = 0;
- cache->items = (struct brw_cache_item **)
- _mesa_calloc(cache->size *
- sizeof(struct brw_cache_item));
-
-
- cache->key_size = key_size;
- cache->aux_size = aux_size;
- cache->aub_type = aub_type;
- cache->aub_sub_type = aub_sub_type;
- switch (aub_type) {
- case DW_GENERAL_STATE: cache->pool = &brw->pool[BRW_GS_POOL]; break;
- case DW_SURFACE_STATE: cache->pool = &brw->pool[BRW_SS_POOL]; break;
- default: assert(0); break;
+ /* If we can find a matching prog in the cache already, then reuse the
+ * existing stuff without creating new copy into the underlying buffer
+ * object. This is notably useful for programs generating shaders at
+ * runtime, where multiple shaders may compile to the same thing in our
+ * backend.
+ */
+ if (matching_data) {
+ item->offset = matching_data->offset;
+ } else {
+ item->offset = brw_alloc_item_data(cache, data_size);
+
+ /* Copy data to the buffer */
+ if (brw->has_llc) {
+ memcpy((char *)cache->bo->virtual + item->offset, data, data_size);
+ } else {
+ drm_intel_bo_subdata(cache->bo, item->offset, data_size, data);
+ }
}
+
+ /* Set up the memory containing the key and aux_data */
+ tmp = malloc(key_size + aux_size);
+
+ memcpy(tmp, key, key_size);
+ memcpy(tmp + key_size, aux, aux_size);
+
+ item->key = tmp;
+
+ if (cache->n_items > cache->size * 1.5f)
+ rehash(cache);
+
+ hash %= cache->size;
+ item->next = cache->items[hash];
+ cache->items[hash] = item;
+ cache->n_items++;
+
+ *out_offset = item->offset;
+ *(void **)out_aux = (void *)((char *)item->key + item->key_size);
+ cache->brw->ctx.NewDriverState |= 1 << cache_id;
}
-void brw_init_caches( struct brw_context *brw )
+void
+brw_init_caches(struct brw_context *brw)
{
+ struct brw_cache *cache = &brw->cache;
- brw_init_cache(brw,
- "CC_VP",
- BRW_CC_VP,
- sizeof(struct brw_cc_viewport),
- 0,
- DW_GENERAL_STATE,
- DWGS_COLOR_CALC_VIEWPORT_STATE);
-
- brw_init_cache(brw,
- "CC_UNIT",
- BRW_CC_UNIT,
- sizeof(struct brw_cc_unit_state),
- 0,
- DW_GENERAL_STATE,
- DWGS_COLOR_CALC_STATE);
-
- brw_init_cache(brw,
- "WM_PROG",
- BRW_WM_PROG,
- sizeof(struct brw_wm_prog_key),
- sizeof(struct brw_wm_prog_data),
- DW_GENERAL_STATE,
- DWGS_KERNEL_INSTRUCTIONS);
-
- brw_init_cache(brw,
- "SAMPLER_DEFAULT_COLOR",
- BRW_SAMPLER_DEFAULT_COLOR,
- sizeof(struct brw_sampler_default_color),
- 0,
- DW_GENERAL_STATE,
- DWGS_SAMPLER_DEFAULT_COLOR);
-
- brw_init_cache(brw,
- "SAMPLER",
- BRW_SAMPLER,
- 0, /* variable key/data size */
- 0,
- DW_GENERAL_STATE,
- DWGS_SAMPLER_STATE);
-
- brw_init_cache(brw,
- "WM_UNIT",
- BRW_WM_UNIT,
- sizeof(struct brw_wm_unit_state),
- 0,
- DW_GENERAL_STATE,
- DWGS_WINDOWER_IZ_STATE);
-
- brw_init_cache(brw,
- "SF_PROG",
- BRW_SF_PROG,
- sizeof(struct brw_sf_prog_key),
- sizeof(struct brw_sf_prog_data),
- DW_GENERAL_STATE,
- DWGS_KERNEL_INSTRUCTIONS);
-
- brw_init_cache(brw,
- "SF_VP",
- BRW_SF_VP,
- sizeof(struct brw_sf_viewport),
- 0,
- DW_GENERAL_STATE,
- DWGS_STRIPS_FANS_VIEWPORT_STATE);
-
- brw_init_cache(brw,
- "SF_UNIT",
- BRW_SF_UNIT,
- sizeof(struct brw_sf_unit_state),
- 0,
- DW_GENERAL_STATE,
- DWGS_STRIPS_FANS_STATE);
-
- brw_init_cache(brw,
- "VS_UNIT",
- BRW_VS_UNIT,
- sizeof(struct brw_vs_unit_state),
- 0,
- DW_GENERAL_STATE,
- DWGS_VERTEX_SHADER_STATE);
-
- brw_init_cache(brw,
- "VS_PROG",
- BRW_VS_PROG,
- sizeof(struct brw_vs_prog_key),
- sizeof(struct brw_vs_prog_data),
- DW_GENERAL_STATE,
- DWGS_KERNEL_INSTRUCTIONS);
-
- brw_init_cache(brw,
- "CLIP_UNIT",
- BRW_CLIP_UNIT,
- sizeof(struct brw_clip_unit_state),
- 0,
- DW_GENERAL_STATE,
- DWGS_CLIPPER_STATE);
-
- brw_init_cache(brw,
- "CLIP_PROG",
- BRW_CLIP_PROG,
- sizeof(struct brw_clip_prog_key),
- sizeof(struct brw_clip_prog_data),
- DW_GENERAL_STATE,
- DWGS_KERNEL_INSTRUCTIONS);
-
- brw_init_cache(brw,
- "GS_UNIT",
- BRW_GS_UNIT,
- sizeof(struct brw_gs_unit_state),
- 0,
- DW_GENERAL_STATE,
- DWGS_GEOMETRY_SHADER_STATE);
-
- brw_init_cache(brw,
- "GS_PROG",
- BRW_GS_PROG,
- sizeof(struct brw_gs_prog_key),
- sizeof(struct brw_gs_prog_data),
- DW_GENERAL_STATE,
- DWGS_KERNEL_INSTRUCTIONS);
-
- brw_init_cache(brw,
- "SS_SURFACE",
- BRW_SS_SURFACE,
- sizeof(struct brw_surface_state),
- 0,
- DW_SURFACE_STATE,
- DWSS_SURFACE_STATE);
-
- brw_init_cache(brw,
- "SS_SURF_BIND",
- BRW_SS_SURF_BIND,
- sizeof(struct brw_surface_binding_table),
- 0,
- DW_SURFACE_STATE,
- DWSS_BINDING_TABLE_STATE);
-}
+ cache->brw = brw;
+ cache->size = 7;
+ cache->n_items = 0;
+ cache->items =
+ calloc(cache->size, sizeof(struct brw_cache_item *));
+
+ cache->bo = drm_intel_bo_alloc(brw->bufmgr,
+ "program cache",
+ 4096, 64);
+ if (brw->has_llc)
+ drm_intel_gem_bo_map_unsynchronized(cache->bo);
+
+ cache->aux_compare[BRW_CACHE_VS_PROG] = brw_vs_prog_data_compare;
+ cache->aux_compare[BRW_CACHE_GS_PROG] = brw_gs_prog_data_compare;
+ cache->aux_compare[BRW_CACHE_FS_PROG] = brw_wm_prog_data_compare;
+ cache->aux_compare[BRW_CACHE_CS_PROG] = brw_cs_prog_data_compare;
+ cache->aux_free[BRW_CACHE_VS_PROG] = brw_stage_prog_data_free;
+ cache->aux_free[BRW_CACHE_GS_PROG] = brw_stage_prog_data_free;
+ cache->aux_free[BRW_CACHE_FS_PROG] = brw_stage_prog_data_free;
+ cache->aux_free[BRW_CACHE_CS_PROG] = brw_stage_prog_data_free;
+}
-/* When we lose hardware context, need to invalidate the surface cache
- * as these structs must be explicitly re-uploaded. They are subject
- * to fixup by the memory manager as they contain absolute agp
- * offsets, so we need to ensure there is a fresh version of the
- * struct available to receive the fixup.
- *
- * XXX: Need to ensure that there aren't two versions of a surface or
- * bufferobj with different backing data active in the same buffer at
- * once? Otherwise the cache could confuse them. Maybe better not to
- * cache at all?
- *
- * --> Isn't this the same as saying need to ensure batch is flushed
- * before new data is uploaded to an existing buffer? We
- * already try to make sure of that.
- */
-static void clear_cache( struct brw_cache *cache )
+static void
+brw_clear_cache(struct brw_context *brw, struct brw_cache *cache)
{
struct brw_cache_item *c, *next;
GLuint i;
+ DBG("%s\n", __func__);
+
for (i = 0; i < cache->size; i++) {
for (c = cache->items[i]; c; c = next) {
next = c->next;
+ 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;
-}
-
-void brw_clear_all_caches( struct brw_context *brw )
-{
- GLint i;
- if (INTEL_DEBUG & DEBUG_STATE)
- _mesa_printf("%s\n", __FUNCTION__);
+ /* Start putting programs into the start of the BO again, since
+ * we'll never find the old results.
+ */
+ cache->next_offset = 0;
- for (i = 0; i < BRW_MAX_CACHE; i++)
- clear_cache(&brw->cache[i]);
+ /* We need to make sure that the programs get regenerated, since
+ * any offsets leftover in brw_context will no longer be valid.
+ */
+ brw->NewGLState |= ~0;
+ brw->ctx.NewDriverState |= ~0ull;
+ intel_batchbuffer_flush(brw);
+}
- if (brw->curbe.last_buf) {
- _mesa_free(brw->curbe.last_buf);
- brw->curbe.last_buf = NULL;
+void
+brw_state_cache_check_size(struct brw_context *brw)
+{
+ /* un-tuned guess. Each object is generally a page, so 2000 of them is 8 MB of
+ * state cache.
+ */
+ 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);
}
-
- brw->state.dirty.mesa |= ~0;
- brw->state.dirty.brw |= ~0;
- brw->state.dirty.cache |= ~0;
}
+static void
+brw_destroy_cache(struct brw_context *brw, struct brw_cache *cache)
+{
+ DBG("%s\n", __func__);
+ if (cache->bo == NULL)
+ return;
+
+ if (brw->has_llc)
+ drm_intel_bo_unmap(cache->bo);
+ drm_intel_bo_unreference(cache->bo);
+ cache->bo = NULL;
+ brw_clear_cache(brw, cache);
+ free(cache->items);
+ cache->items = NULL;
+ cache->size = 0;
+}
-void brw_destroy_caches( struct brw_context *brw )
-{
- GLuint i;
- for (i = 0; i < BRW_MAX_CACHE; i++)
- clear_cache(&brw->cache[i]);
+void
+brw_destroy_caches(struct brw_context *brw)
+{
+ brw_destroy_cache(brw, &brw->cache);
}
projects / mesa.git / blobdiff