#include "lp_cs_tpool.h"
#include "frontend/sw_winsys.h"
#include "nir/nir_to_tgsi_info.h"
+#include "util/mesa-sha1.h"
#include "nir_serialize.h"
+
+/** Fragment shader number (for debugging) */
+static unsigned cs_no = 0;
+
struct lp_cs_job_info {
unsigned grid_size[3];
unsigned block_size[3];
cs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
cs_type.width = 32; /* 32-bit float */
cs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
- snprintf(func_name, sizeof(func_name), "cs%u_variant%u",
- shader->no, variant->no);
+ snprintf(func_name, sizeof(func_name), "cs_variant");
- snprintf(func_name_coro, sizeof(func_name), "cs_co_%u_variant%u",
- shader->no, variant->no);
+ snprintf(func_name_coro, sizeof(func_name), "cs_co_variant");
arg_types[0] = variant->jit_cs_context_ptr_type; /* context */
arg_types[1] = int32_type; /* block_x_size */
}
}
+ lp_build_coro_declare_malloc_hooks(gallivm);
+
+ if (variant->gallivm->cache->data_size)
+ return;
+
context_ptr = LLVMGetParam(function, 0);
x_size_arg = LLVMGetParam(function, 1);
y_size_arg = LLVMGetParam(function, 2);
builder = gallivm->builder;
assert(builder);
LLVMPositionBuilderAtEnd(builder, block);
- sampler = lp_llvm_sampler_soa_create(key->state);
- image = lp_llvm_image_soa_create(key->image_state);
+ sampler = lp_llvm_sampler_soa_create(key->samplers, key->nr_samplers);
+ image = lp_llvm_image_soa_create(lp_cs_variant_key_images(key));
struct lp_build_loop_state loop_state[4];
LLVMValueRef num_x_loop;
{
struct lp_compute_shader *shader;
int nr_samplers, nr_sampler_views;
+
shader = CALLOC_STRUCT(lp_compute_shader);
if (!shader)
return NULL;
+ shader->no = cs_no++;
+
shader->base.type = templ->ir_type;
if (templ->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
struct blob_reader reader;
nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
- shader->variant_key_size = Offset(struct lp_compute_shader_variant_key,
- state[MAX2(nr_samplers, nr_sampler_views)]);
+ int nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
+ shader->variant_key_size = lp_cs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
+
return shader;
}
FREE(shader);
}
-static void
+static struct lp_compute_shader_variant_key *
make_variant_key(struct llvmpipe_context *lp,
struct lp_compute_shader *shader,
- struct lp_compute_shader_variant_key *key)
+ char *store)
{
int i;
-
- memset(key, 0, shader->variant_key_size);
+ struct lp_compute_shader_variant_key *key;
+ key = (struct lp_compute_shader_variant_key *)store;
+ memset(key, 0, offsetof(struct lp_compute_shader_variant_key, samplers[1]));
/* This value will be the same for all the variants of a given shader:
*/
key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
+ struct lp_sampler_static_state *cs_sampler;
+
+ cs_sampler = key->samplers;
for(i = 0; i < key->nr_samplers; ++i) {
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
- lp_sampler_static_sampler_state(&key->state[i].sampler_state,
+ lp_sampler_static_sampler_state(&cs_sampler[i].sampler_state,
lp->samplers[PIPE_SHADER_COMPUTE][i]);
}
}
* used views may be included in the shader key.
*/
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
- lp_sampler_static_texture_state(&key->state[i].texture_state,
+ lp_sampler_static_texture_state(&cs_sampler[i].texture_state,
lp->sampler_views[PIPE_SHADER_COMPUTE][i]);
}
}
key->nr_sampler_views = key->nr_samplers;
for(i = 0; i < key->nr_sampler_views; ++i) {
if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
- lp_sampler_static_texture_state(&key->state[i].texture_state,
+ lp_sampler_static_texture_state(&cs_sampler[i].texture_state,
lp->sampler_views[PIPE_SHADER_COMPUTE][i]);
}
}
}
+ struct lp_image_static_state *lp_image;
+ lp_image = lp_cs_variant_key_images(key);
key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
for (i = 0; i < key->nr_images; ++i) {
if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
- lp_sampler_static_texture_state_image(&key->image_state[i].image_state,
+ lp_sampler_static_texture_state_image(&lp_image[i].image_state,
&lp->images[PIPE_SHADER_COMPUTE][i]);
}
}
+ return key;
}
static void
debug_printf("cs variant %p:\n", (void *) key);
for (i = 0; i < key->nr_samplers; ++i) {
- const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
+ const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
debug_printf("sampler[%u] = \n", i);
debug_printf(" .wrap = %s %s %s\n",
util_str_tex_wrap(sampler->wrap_s, TRUE),
debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
}
for (i = 0; i < key->nr_sampler_views; ++i) {
- const struct lp_static_texture_state *texture = &key->state[i].texture_state;
+ const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
debug_printf("texture[%u] = \n", i);
debug_printf(" .format = %s\n",
util_format_name(texture->format));
texture->pot_height,
texture->pot_depth);
}
+ struct lp_image_static_state *images = lp_cs_variant_key_images(key);
for (i = 0; i < key->nr_images; ++i) {
- const struct lp_static_texture_state *image = &key->image_state[i].image_state;
+ const struct lp_static_texture_state *image = &images[i].image_state;
debug_printf("image[%u] = \n", i);
debug_printf(" .format = %s\n",
util_format_name(image->format));
debug_printf("\n");
}
+static void
+lp_cs_get_ir_cache_key(struct lp_compute_shader_variant *variant,
+ unsigned char ir_sha1_cache_key[20])
+{
+ struct blob blob = { 0 };
+ unsigned ir_size;
+ void *ir_binary;
+
+ blob_init(&blob);
+ nir_serialize(&blob, variant->shader->base.ir.nir, true);
+ ir_binary = blob.data;
+ ir_size = blob.size;
+
+ struct mesa_sha1 ctx;
+ _mesa_sha1_init(&ctx);
+ _mesa_sha1_update(&ctx, &variant->key, variant->shader->variant_key_size);
+ _mesa_sha1_update(&ctx, ir_binary, ir_size);
+ _mesa_sha1_final(&ctx, ir_sha1_cache_key);
+
+ blob_finish(&blob);
+}
+
static struct lp_compute_shader_variant *
generate_variant(struct llvmpipe_context *lp,
struct lp_compute_shader *shader,
const struct lp_compute_shader_variant_key *key)
{
+ struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
struct lp_compute_shader_variant *variant;
char module_name[64];
-
- variant = CALLOC_STRUCT(lp_compute_shader_variant);
+ unsigned char ir_sha1_cache_key[20];
+ struct lp_cached_code cached = { 0 };
+ bool needs_caching = false;
+ variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
if (!variant)
return NULL;
+ memset(variant, 0, sizeof(*variant));
snprintf(module_name, sizeof(module_name), "cs%u_variant%u",
shader->no, shader->variants_created);
- variant->gallivm = gallivm_create(module_name, lp->context);
+ variant->shader = shader;
+ memcpy(&variant->key, key, shader->variant_key_size);
+
+ if (shader->base.ir.nir) {
+ lp_cs_get_ir_cache_key(variant, ir_sha1_cache_key);
+
+ lp_disk_cache_find_shader(screen, &cached, ir_sha1_cache_key);
+ if (!cached.data_size)
+ needs_caching = true;
+ }
+ variant->gallivm = gallivm_create(module_name, lp->context, &cached);
if (!variant->gallivm) {
FREE(variant);
return NULL;
}
- variant->shader = shader;
variant->list_item_global.base = variant;
variant->list_item_local.base = variant;
variant->no = shader->variants_created++;
- memcpy(&variant->key, key, shader->variant_key_size);
+
if ((LP_DEBUG & DEBUG_CS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
lp_debug_cs_variant(variant);
gallivm_compile_module(variant->gallivm);
+ lp_build_coro_add_malloc_hooks(variant->gallivm);
variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
variant->jit_function = (lp_jit_cs_func)gallivm_jit_function(variant->gallivm, variant->function);
+ if (needs_caching) {
+ lp_disk_cache_insert_shader(screen, &cached, ir_sha1_cache_key);
+ }
gallivm_free_ir(variant->gallivm);
return variant;
}
{
struct lp_compute_shader *shader = lp->cs;
- struct lp_compute_shader_variant_key key;
+ struct lp_compute_shader_variant_key *key;
struct lp_compute_shader_variant *variant = NULL;
struct lp_cs_variant_list_item *li;
+ char store[LP_CS_MAX_VARIANT_KEY_SIZE];
- make_variant_key(lp, shader, &key);
+ key = make_variant_key(lp, shader, store);
/* Search the variants for one which matches the key */
li = first_elem(&shader->variants);
while(!at_end(&shader->variants, li)) {
- if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
+ if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
variant = li->base;
break;
}
* Generate the new variant.
*/
t0 = os_time_get();
- variant = generate_variant(lp, shader, &key);
+ variant = generate_variant(lp, shader, key);
t1 = os_time_get();
dt = t1 - t0;
LP_COUNT_ADD(llvm_compile_time, dt);
projects / mesa.git / blobdiff