/* Asynchronous compute shader compilation. */
static void si_create_compute_state_async(void *job, int thread_index)
{
struct si_compute *program = (struct si_compute *)job;
struct si_shader *shader = &program->shader;
struct si_shader_selector sel;
LLVMTargetMachineRef tm;
struct pipe_debug_callback *debug = &program->compiler_ctx_state.debug;
assert(!debug->debug_message || debug->async);
assert(thread_index >= 0);
assert(thread_index < ARRAY_SIZE(program->screen->tm));
tm = program->screen->tm[thread_index];
memset(&sel, 0, sizeof(sel));
sel.screen = program->screen;
if (program->ir_type == PIPE_SHADER_IR_TGSI) {
tgsi_scan_shader(program->ir.tgsi, &sel.info);
sel.tokens = program->ir.tgsi;
} else {
assert(program->ir_type == PIPE_SHADER_IR_NIR);
sel.nir = program->ir.nir;
si_nir_scan_shader(sel.nir, &sel.info);
si_lower_nir(&sel);
}
sel.type = PIPE_SHADER_COMPUTE;
sel.local_size = program->local_size;
si_get_active_slot_masks(&sel.info,
&program->active_const_and_shader_buffers,
&program->active_samplers_and_images);
program->shader.selector = &sel;
program->shader.is_monolithic = true;
program->uses_grid_size = sel.info.uses_grid_size;
program->uses_block_size = sel.info.uses_block_size;
program->uses_bindless_samplers = sel.info.uses_bindless_samplers;
program->uses_bindless_images = sel.info.uses_bindless_images;
if (si_shader_create(program->screen, tm, &program->shader, debug)) {
program->shader.compilation_failed = true;
} else {
bool scratch_enabled = shader->config.scratch_bytes_per_wave > 0;
unsigned user_sgprs = SI_NUM_RESOURCE_SGPRS +
(sel.info.uses_grid_size ? 3 : 0) +
(sel.info.uses_block_size ? 3 : 0);
shader->config.rsrc1 =
S_00B848_VGPRS((shader->config.num_vgprs - 1) / 4) |
S_00B848_SGPRS((shader->config.num_sgprs - 1) / 8) |
S_00B848_DX10_CLAMP(1) |
S_00B848_FLOAT_MODE(shader->config.float_mode);
shader->config.rsrc2 =
S_00B84C_USER_SGPR(user_sgprs) |
S_00B84C_SCRATCH_EN(scratch_enabled) |
S_00B84C_TGID_X_EN(sel.info.uses_block_id[0]) |
S_00B84C_TGID_Y_EN(sel.info.uses_block_id[1]) |
S_00B84C_TGID_Z_EN(sel.info.uses_block_id[2]) |
S_00B84C_TIDIG_COMP_CNT(sel.info.uses_thread_id[2] ? 2 :
sel.info.uses_thread_id[1] ? 1 : 0) |
S_00B84C_LDS_SIZE(shader->config.lds_size);
program->variable_group_size =
sel.info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0;
}
if (program->ir_type == PIPE_SHADER_IR_TGSI)
FREE(program->ir.tgsi);
program->shader.selector = NULL;
}
/* Asynchronous compute shader compilation. */
static void si_create_compute_state_async(void *job, int thread_index)
{
struct si_compute *program = (struct si_compute *)job;
struct si_shader *shader = &program->shader;
struct si_shader_selector sel;
struct ac_llvm_compiler *compiler;
struct pipe_debug_callback *debug = &program->compiler_ctx_state.debug;
struct si_screen *sscreen = program->screen;
assert(!debug->debug_message || debug->async);
assert(thread_index >= 0);
assert(thread_index < ARRAY_SIZE(sscreen->compiler));
compiler = &sscreen->compiler[thread_index];
memset(&sel, 0, sizeof(sel));
sel.screen = sscreen;
if (program->ir_type == PIPE_SHADER_IR_TGSI) {
tgsi_scan_shader(program->ir.tgsi, &sel.info);
sel.tokens = program->ir.tgsi;
} else {
assert(program->ir_type == PIPE_SHADER_IR_NIR);
sel.nir = program->ir.nir;
si_nir_scan_shader(sel.nir, &sel.info);
si_lower_nir(&sel);
}
/* Store the declared LDS size into tgsi_shader_info for the shader
* cache to include it.
*/
sel.info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE] = program->local_size;
sel.type = PIPE_SHADER_COMPUTE;
si_get_active_slot_masks(&sel.info,
&program->active_const_and_shader_buffers,
&program->active_samplers_and_images);
program->shader.selector = &sel;
program->shader.is_monolithic = true;
program->uses_grid_size = sel.info.uses_grid_size;
program->uses_bindless_samplers = sel.info.uses_bindless_samplers;
program->uses_bindless_images = sel.info.uses_bindless_images;
program->reads_variable_block_size =
sel.info.uses_block_size &&
sel.info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0;
program->num_cs_user_data_dwords =
sel.info.properties[TGSI_PROPERTY_CS_USER_DATA_DWORDS];
void *ir_binary = si_get_ir_binary(&sel);
/* Try to load the shader from the shader cache. */
mtx_lock(&sscreen->shader_cache_mutex);
if (ir_binary &&
si_shader_cache_load_shader(sscreen, ir_binary, shader)) {
mtx_unlock(&sscreen->shader_cache_mutex);
si_shader_dump_stats_for_shader_db(shader, debug);
si_shader_dump(sscreen, shader, debug, PIPE_SHADER_COMPUTE,
stderr, true);
if (si_shader_binary_upload(sscreen, shader))
program->shader.compilation_failed = true;
} else {
mtx_unlock(&sscreen->shader_cache_mutex);
if (si_shader_create(sscreen, compiler, &program->shader, debug)) {
program->shader.compilation_failed = true;
if (program->ir_type == PIPE_SHADER_IR_TGSI)
FREE(program->ir.tgsi);
program->shader.selector = NULL;
return;
}
bool scratch_enabled = shader->config.scratch_bytes_per_wave > 0;
unsigned user_sgprs = SI_NUM_RESOURCE_SGPRS +
(sel.info.uses_grid_size ? 3 : 0) +
(program->reads_variable_block_size ? 3 : 0) +
program->num_cs_user_data_dwords;
shader->config.rsrc1 =
S_00B848_VGPRS((shader->config.num_vgprs - 1) / 4) |
S_00B848_SGPRS((shader->config.num_sgprs - 1) / 8) |
S_00B848_DX10_CLAMP(1) |
S_00B848_FLOAT_MODE(shader->config.float_mode);
shader->config.rsrc2 =
S_00B84C_USER_SGPR(user_sgprs) |
S_00B84C_SCRATCH_EN(scratch_enabled) |
S_00B84C_TGID_X_EN(sel.info.uses_block_id[0]) |
S_00B84C_TGID_Y_EN(sel.info.uses_block_id[1]) |
S_00B84C_TGID_Z_EN(sel.info.uses_block_id[2]) |
S_00B84C_TIDIG_COMP_CNT(sel.info.uses_thread_id[2] ? 2 :
sel.info.uses_thread_id[1] ? 1 : 0) |
S_00B84C_LDS_SIZE(shader->config.lds_size);
if (ir_binary) {
mtx_lock(&sscreen->shader_cache_mutex);
if (!si_shader_cache_insert_shader(sscreen, ir_binary, shader, true))
FREE(ir_binary);
mtx_unlock(&sscreen->shader_cache_mutex);
}
}
if (program->ir_type == PIPE_SHADER_IR_TGSI)
FREE(program->ir.tgsi);
program->shader.selector = NULL;
}
