Commit 62229e89 authored by Marek Olšák's avatar Marek Olšák

radeonsi: use IR SHA1 as the cache key for the in-memory shader cache

instead of using whole IR binaries. This saves some memory.
Reviewed-by: Timothy Arceri's avatarTimothy Arceri <tarceri@itsqueeze.com>
parent 65a5b24a
......@@ -148,13 +148,13 @@ static void si_create_compute_state_async(void *job, int thread_index)
program->num_cs_user_data_dwords =
sel->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
void *ir_binary = si_get_ir_binary(sel, false, false);
unsigned char ir_sha1_cache_key[20];
si_get_ir_cache_key(sel, false, false, ir_sha1_cache_key);
/* Try to load the shader from the shader cache. */
simple_mtx_lock(&sscreen->shader_cache_mutex);
if (ir_binary &&
si_shader_cache_load_shader(sscreen, ir_binary, shader)) {
if (si_shader_cache_load_shader(sscreen, ir_sha1_cache_key, shader)) {
simple_mtx_unlock(&sscreen->shader_cache_mutex);
si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
......@@ -202,12 +202,10 @@ static void si_create_compute_state_async(void *job, int thread_index)
sel->info.uses_thread_id[1] ? 1 : 0) |
S_00B84C_LDS_SIZE(shader->config.lds_size);
if (ir_binary) {
simple_mtx_lock(&sscreen->shader_cache_mutex);
if (!si_shader_cache_insert_shader(sscreen, ir_binary, shader, true))
FREE(ir_binary);
simple_mtx_unlock(&sscreen->shader_cache_mutex);
}
simple_mtx_lock(&sscreen->shader_cache_mutex);
si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key,
shader, true);
simple_mtx_unlock(&sscreen->shader_cache_mutex);
}
if (program->ir_type == PIPE_SHADER_IR_TGSI)
......
......@@ -583,10 +583,13 @@ si_compute_fast_udiv_info32(uint32_t D, unsigned num_bits);
void si_emit_dpbb_state(struct si_context *sctx);
/* si_state_shaders.c */
void *si_get_ir_binary(struct si_shader_selector *sel, bool ngg, bool es);
bool si_shader_cache_load_shader(struct si_screen *sscreen, void *ir_binary,
void si_get_ir_cache_key(struct si_shader_selector *sel, bool ngg, bool es,
unsigned char ir_sha1_cache_key[20]);
bool si_shader_cache_load_shader(struct si_screen *sscreen,
unsigned char ir_sha1_cache_key[20],
struct si_shader *shader);
bool si_shader_cache_insert_shader(struct si_screen *sscreen, void *ir_binary,
void si_shader_cache_insert_shader(struct si_screen *sscreen,
unsigned char ir_sha1_cache_key[20],
struct si_shader *shader,
bool insert_into_disk_cache);
bool si_update_shaders(struct si_context *sctx);
......
......@@ -42,10 +42,10 @@
/* SHADER_CACHE */
/**
* Return the IR binary in a buffer. For TGSI the first 4 bytes contain its
* size as integer.
* Return the IR key for the shader cache.
*/
void *si_get_ir_binary(struct si_shader_selector *sel, bool ngg, bool es)
void si_get_ir_cache_key(struct si_shader_selector *sel, bool ngg, bool es,
unsigned char ir_sha1_cache_key[20])
{
struct blob blob;
unsigned ir_size;
......@@ -54,7 +54,7 @@ void *si_get_ir_binary(struct si_shader_selector *sel, bool ngg, bool es)
if (sel->tokens) {
ir_binary = sel->tokens;
ir_size = tgsi_num_tokens(sel->tokens) *
sizeof(struct tgsi_token);
sizeof(struct tgsi_token);
} else {
assert(sel->nir);
......@@ -78,20 +78,18 @@ void *si_get_ir_binary(struct si_shader_selector *sel, bool ngg, bool es)
if (sel->force_correct_derivs_after_kill)
shader_variant_flags |= 1 << 3;
unsigned size = 4 + 4 + ir_size + sizeof(sel->so);
char *result = (char*)MALLOC(size);
if (!result)
return NULL;
((uint32_t*)result)[0] = size;
((uint32_t*)result)[1] = shader_variant_flags;
memcpy(result + 8, ir_binary, ir_size);
memcpy(result + 8 + ir_size, &sel->so, sizeof(sel->so));
struct mesa_sha1 ctx;
_mesa_sha1_init(&ctx);
_mesa_sha1_update(&ctx, &shader_variant_flags, 4);
_mesa_sha1_update(&ctx, ir_binary, ir_size);
if (sel->type == PIPE_SHADER_VERTEX ||
sel->type == PIPE_SHADER_TESS_EVAL ||
sel->type == PIPE_SHADER_GEOMETRY)
_mesa_sha1_update(&ctx, &sel->so, sizeof(sel->so));
_mesa_sha1_final(&ctx, ir_sha1_cache_key);
if (sel->nir)
blob_finish(&blob);
return result;
}
/** Copy "data" to "ptr" and return the next dword following copied data. */
......@@ -208,10 +206,9 @@ static bool si_load_shader_binary(struct si_shader *shader, void *binary)
/**
* Insert a shader into the cache. It's assumed the shader is not in the cache.
* Use si_shader_cache_load_shader before calling this.
*
* Returns false on failure, in which case the ir_binary should be freed.
*/
bool si_shader_cache_insert_shader(struct si_screen *sscreen, void *ir_binary,
void si_shader_cache_insert_shader(struct si_screen *sscreen,
unsigned char ir_sha1_cache_key[20],
struct si_shader *shader,
bool insert_into_disk_cache)
{
......@@ -219,42 +216,41 @@ bool si_shader_cache_insert_shader(struct si_screen *sscreen, void *ir_binary,
struct hash_entry *entry;
uint8_t key[CACHE_KEY_SIZE];
entry = _mesa_hash_table_search(sscreen->shader_cache, ir_binary);
entry = _mesa_hash_table_search(sscreen->shader_cache, ir_sha1_cache_key);
if (entry)
return false; /* already added */
return; /* already added */
hw_binary = si_get_shader_binary(shader);
if (!hw_binary)
return false;
return;
if (_mesa_hash_table_insert(sscreen->shader_cache, ir_binary,
if (_mesa_hash_table_insert(sscreen->shader_cache,
mem_dup(ir_sha1_cache_key, 20),
hw_binary) == NULL) {
FREE(hw_binary);
return false;
return;
}
if (sscreen->disk_shader_cache && insert_into_disk_cache) {
disk_cache_compute_key(sscreen->disk_shader_cache, ir_binary,
*((uint32_t *)ir_binary), key);
disk_cache_compute_key(sscreen->disk_shader_cache,
ir_sha1_cache_key, 20, key);
disk_cache_put(sscreen->disk_shader_cache, key, hw_binary,
*((uint32_t *) hw_binary), NULL);
}
return true;
}
bool si_shader_cache_load_shader(struct si_screen *sscreen, void *ir_binary,
bool si_shader_cache_load_shader(struct si_screen *sscreen,
unsigned char ir_sha1_cache_key[20],
struct si_shader *shader)
{
struct hash_entry *entry =
_mesa_hash_table_search(sscreen->shader_cache, ir_binary);
_mesa_hash_table_search(sscreen->shader_cache, ir_sha1_cache_key);
if (!entry) {
if (sscreen->disk_shader_cache) {
unsigned char sha1[CACHE_KEY_SIZE];
size_t tg_size = *((uint32_t *) ir_binary);
disk_cache_compute_key(sscreen->disk_shader_cache,
ir_binary, tg_size, sha1);
ir_sha1_cache_key, 20, sha1);
size_t binary_size;
uint8_t *buffer =
......@@ -285,16 +281,13 @@ bool si_shader_cache_load_shader(struct si_screen *sscreen, void *ir_binary,
}
free(buffer);
if (!si_shader_cache_insert_shader(sscreen, ir_binary,
shader, false))
FREE(ir_binary);
si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key,
shader, false);
} else {
return false;
}
} else {
if (si_load_shader_binary(shader, entry->data))
FREE(ir_binary);
else
if (!si_load_shader_binary(shader, entry->data))
return false;
}
p_atomic_inc(&sscreen->num_shader_cache_hits);
......@@ -303,20 +296,14 @@ bool si_shader_cache_load_shader(struct si_screen *sscreen, void *ir_binary,
static uint32_t si_shader_cache_key_hash(const void *key)
{
/* The first dword is the key size. */
return util_hash_crc32(key, *(uint32_t*)key);
/* Take the first dword of SHA1. */
return *(uint32_t*)key;
}
static bool si_shader_cache_key_equals(const void *a, const void *b)
{
uint32_t *keya = (uint32_t*)a;
uint32_t *keyb = (uint32_t*)b;
/* The first dword is the key size. */
if (*keya != *keyb)
return false;
return memcmp(keya, keyb, *keya) == 0;
/* Compare SHA1s. */
return memcmp(a, b, 20) == 0;
}
static void si_destroy_shader_cache_entry(struct hash_entry *entry)
......@@ -2484,7 +2471,7 @@ static void si_init_shader_selector_async(void *job, int thread_index)
*/
if (!sscreen->use_monolithic_shaders) {
struct si_shader *shader = CALLOC_STRUCT(si_shader);
void *ir_binary = NULL;
unsigned char ir_sha1_cache_key[20];
if (!shader) {
fprintf(stderr, "radeonsi: can't allocate a main shader part\n");
......@@ -2509,15 +2496,14 @@ static void si_init_shader_selector_async(void *job, int thread_index)
shader->key.as_ngg = 1;
if (sel->tokens || sel->nir) {
ir_binary = si_get_ir_binary(sel, shader->key.as_ngg,
shader->key.as_es);
si_get_ir_cache_key(sel, shader->key.as_ngg,
shader->key.as_es, ir_sha1_cache_key);
}
/* Try to load the shader from the shader cache. */
simple_mtx_lock(&sscreen->shader_cache_mutex);
if (ir_binary &&
si_shader_cache_load_shader(sscreen, ir_binary, shader)) {
if (si_shader_cache_load_shader(sscreen, ir_sha1_cache_key, shader)) {
simple_mtx_unlock(&sscreen->shader_cache_mutex);
si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
} else {
......@@ -2527,17 +2513,14 @@ static void si_init_shader_selector_async(void *job, int thread_index)
if (si_compile_tgsi_shader(sscreen, compiler, shader,
debug) != 0) {
FREE(shader);
FREE(ir_binary);
fprintf(stderr, "radeonsi: can't compile a main shader part\n");
return;
}
if (ir_binary) {
simple_mtx_lock(&sscreen->shader_cache_mutex);
if (!si_shader_cache_insert_shader(sscreen, ir_binary, shader, true))
FREE(ir_binary);
simple_mtx_unlock(&sscreen->shader_cache_mutex);
}
simple_mtx_lock(&sscreen->shader_cache_mutex);
si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key,
shader, true);
simple_mtx_unlock(&sscreen->shader_cache_mutex);
}
*si_get_main_shader_part(sel, &shader->key) = shader;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment