#include "lp_perf.h"
#include "lp_screen.h"
#include "lp_memory.h"
+#include "lp_query.h"
#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) */
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->samplers);
- image = lp_llvm_image_soa_create(lp_cs_variant_key_images(key));
+ sampler = lp_llvm_sampler_soa_create(key->samplers, key->nr_samplers);
+ image = lp_llvm_image_soa_create(lp_cs_variant_key_images(key), key->nr_images);
struct lp_build_loop_state loop_state[4];
LLVMValueRef num_x_loop;
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];
+ 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;
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;
}
for (i = 0; i < ARRAY_SIZE(csctx->constants); ++i) {
struct pipe_resource *buffer = csctx->constants[i].current.buffer;
const ubyte *current_data = NULL;
-
+ unsigned current_size = csctx->constants[i].current.buffer_size;
if (buffer) {
/* resource buffer */
current_data = (ubyte *) llvmpipe_resource_data(buffer);
current_data = (ubyte *) csctx->constants[i].current.user_buffer;
}
- if (current_data) {
+ if (current_data && current_size >= sizeof(float)) {
current_data += csctx->constants[i].current.buffer_offset;
-
csctx->cs.current.jit_context.constants[i] = (const float *)current_data;
- csctx->cs.current.jit_context.num_constants[i] = csctx->constants[i].current.buffer_size;
+ csctx->cs.current.jit_context.num_constants[i] =
+ DIV_ROUND_UP(csctx->constants[i].current.buffer_size,
+ lp_get_constant_buffer_stride(llvmpipe->pipe.screen));
} else {
- csctx->cs.current.jit_context.constants[i] = NULL;
+ static const float fake_const_buf[4];
+ csctx->cs.current.jit_context.constants[i] = fake_const_buf;
csctx->cs.current.jit_context.num_constants[i] = 0;
}
}
static void
llvmpipe_cs_update_derived(struct llvmpipe_context *llvmpipe, void *input)
{
- if (llvmpipe->cs_dirty & (LP_CSNEW_CS))
- llvmpipe_update_cs(llvmpipe);
-
if (llvmpipe->cs_dirty & LP_CSNEW_CONSTANTS) {
lp_csctx_set_cs_constants(llvmpipe->csctx,
ARRAY_SIZE(llvmpipe->constants[PIPE_SHADER_COMPUTE]),
csctx->cs.current.jit_context.kernel_args = input;
}
+ if (llvmpipe->cs_dirty & (LP_CSNEW_CS |
+ LP_CSNEW_IMAGES |
+ LP_CSNEW_SAMPLER_VIEW |
+ LP_CSNEW_SAMPLER))
+ llvmpipe_update_cs(llvmpipe);
+
+
llvmpipe->cs_dirty = 0;
}
struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
struct lp_cs_job_info job_info;
+ if (!llvmpipe_check_render_cond(llvmpipe))
+ return;
+
memset(&job_info, 0, sizeof(job_info));
llvmpipe_cs_update_derived(llvmpipe, info->input);