static void si_dump_shader(struct si_screen *sscreen,
struct si_shader_ctx_state *state, FILE *f)
{
if (!state->cso || !state->current)
return;
si_dump_shader_key(state->cso->type, &state->current->key, f);
si_shader_dump(sscreen, state->current, NULL,
state->cso->info.processor, f);
}
static void *si_create_compute_state(
struct pipe_context *ctx,
const struct pipe_compute_state *cso)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_screen *sscreen = (struct si_screen *)ctx->screen;
struct si_compute *program = CALLOC_STRUCT(si_compute);
struct si_shader *shader = &program->shader;
program->ir_type = cso->ir_type;
program->local_size = cso->req_local_mem;
program->private_size = cso->req_private_mem;
program->input_size = cso->req_input_mem;
program->use_code_object_v2 = HAVE_LLVM >= 0x0400 &&
cso->ir_type == PIPE_SHADER_IR_NATIVE;
if (cso->ir_type == PIPE_SHADER_IR_TGSI) {
struct si_shader_selector sel;
bool scratch_enabled;
memset(&sel, 0, sizeof(sel));
sel.tokens = tgsi_dup_tokens(cso->prog);
if (!sel.tokens) {
FREE(program);
return NULL;
}
tgsi_scan_shader(cso->prog, &sel.info);
sel.type = PIPE_SHADER_COMPUTE;
sel.local_size = cso->req_local_mem;
p_atomic_inc(&sscreen->b.num_shaders_created);
program->shader.selector = &sel;
if (si_shader_create(sscreen, sctx->tm, &program->shader,
&sctx->b.debug)) {
FREE(sel.tokens);
FREE(program);
return NULL;
}
scratch_enabled = shader->config.scratch_bytes_per_wave > 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(SI_CS_NUM_USER_SGPR) |
S_00B84C_SCRATCH_EN(scratch_enabled) |
S_00B84C_TGID_X_EN(1) | S_00B84C_TGID_Y_EN(1) |
S_00B84C_TGID_Z_EN(1) | S_00B84C_TIDIG_COMP_CNT(2) |
S_00B84C_LDS_SIZE(shader->config.lds_size);
FREE(sel.tokens);
} else {
const struct pipe_llvm_program_header *header;
const char *code;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
radeon_elf_read(code, header->num_bytes, &program->shader.binary);
if (program->use_code_object_v2) {
const amd_kernel_code_t *code_object =
si_compute_get_code_object(program, 0);
code_object_to_config(code_object, &program->shader.config);
} else {
si_shader_binary_read_config(&program->shader.binary,
&program->shader.config, 0);
}
si_shader_dump(sctx->screen, &program->shader, &sctx->b.debug,
PIPE_SHADER_COMPUTE, stderr);
si_shader_binary_upload(sctx->screen, &program->shader);
}
return program;
}
static void *si_create_compute_state(
struct pipe_context *ctx,
const struct pipe_compute_state *cso)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_screen *sscreen = (struct si_screen *)ctx->screen;
struct si_compute *program = CALLOC_STRUCT(si_compute);
pipe_reference_init(&program->reference, 1);
program->screen = (struct si_screen *)ctx->screen;
program->ir_type = cso->ir_type;
program->local_size = cso->req_local_mem;
program->private_size = cso->req_private_mem;
program->input_size = cso->req_input_mem;
program->use_code_object_v2 = cso->ir_type == PIPE_SHADER_IR_NATIVE;
if (cso->ir_type != PIPE_SHADER_IR_NATIVE) {
if (cso->ir_type == PIPE_SHADER_IR_TGSI) {
program->ir.tgsi = tgsi_dup_tokens(cso->prog);
if (!program->ir.tgsi) {
FREE(program);
return NULL;
}
} else {
assert(cso->ir_type == PIPE_SHADER_IR_NIR);
program->ir.nir = (struct nir_shader *) cso->prog;
}
program->compiler_ctx_state.debug = sctx->debug;
program->compiler_ctx_state.is_debug_context = sctx->is_debug;
p_atomic_inc(&sscreen->num_shaders_created);
util_queue_fence_init(&program->ready);
struct util_async_debug_callback async_debug;
bool wait =
(sctx->debug.debug_message && !sctx->debug.async) ||
sctx->is_debug ||
si_can_dump_shader(sscreen, PIPE_SHADER_COMPUTE);
if (wait) {
u_async_debug_init(&async_debug);
program->compiler_ctx_state.debug = async_debug.base;
}
util_queue_add_job(&sscreen->shader_compiler_queue,
program, &program->ready,
si_create_compute_state_async, NULL);
if (wait) {
util_queue_fence_wait(&program->ready);
u_async_debug_drain(&async_debug, &sctx->debug);
u_async_debug_cleanup(&async_debug);
}
} else {
const struct pipe_llvm_program_header *header;
const char *code;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
ac_elf_read(code, header->num_bytes, &program->shader.binary);
if (program->use_code_object_v2) {
const amd_kernel_code_t *code_object =
si_compute_get_code_object(program, 0);
code_object_to_config(code_object, &program->shader.config);
} else {
si_shader_binary_read_config(&program->shader.binary,
&program->shader.config, 0);
}
si_shader_dump(sctx->screen, &program->shader, &sctx->debug,
PIPE_SHADER_COMPUTE, stderr, true);
if (si_shader_binary_upload(sctx->screen, &program->shader) < 0) {
fprintf(stderr, "LLVM failed to upload shader\n");
FREE(program);
return NULL;
}
}
return program;
}
/* 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;
}
static void *si_create_compute_state(
struct pipe_context *ctx,
const struct pipe_compute_state *cso)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_screen *sscreen = (struct si_screen *)ctx->screen;
struct si_compute *program = CALLOC_STRUCT(si_compute);
pipe_reference_init(&program->reference, 1);
program->screen = (struct si_screen *)ctx->screen;
program->ir_type = cso->ir_type;
program->local_size = cso->req_local_mem;
program->private_size = cso->req_private_mem;
program->input_size = cso->req_input_mem;
program->use_code_object_v2 = cso->ir_type == PIPE_SHADER_IR_NATIVE;
if (cso->ir_type != PIPE_SHADER_IR_NATIVE) {
if (cso->ir_type == PIPE_SHADER_IR_TGSI) {
program->ir.tgsi = tgsi_dup_tokens(cso->prog);
if (!program->ir.tgsi) {
FREE(program);
return NULL;
}
} else {
assert(cso->ir_type == PIPE_SHADER_IR_NIR);
program->ir.nir = (struct nir_shader *) cso->prog;
}
program->compiler_ctx_state.debug = sctx->debug;
program->compiler_ctx_state.is_debug_context = sctx->is_debug;
p_atomic_inc(&sscreen->num_shaders_created);
si_schedule_initial_compile(sctx, PIPE_SHADER_COMPUTE,
&program->ready,
&program->compiler_ctx_state,
program, si_create_compute_state_async);
} else {
const struct pipe_llvm_program_header *header;
const char *code;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
ac_elf_read(code, header->num_bytes, &program->shader.binary);
if (program->use_code_object_v2) {
const amd_kernel_code_t *code_object =
si_compute_get_code_object(program, 0);
code_object_to_config(code_object, &program->shader.config);
if (program->shader.binary.reloc_count != 0) {
fprintf(stderr, "Error: %d unsupported relocations\n",
program->shader.binary.reloc_count);
FREE(program);
return NULL;
}
} else {
si_shader_binary_read_config(&program->shader.binary,
&program->shader.config, 0);
}
si_shader_dump(sctx->screen, &program->shader, &sctx->debug,
PIPE_SHADER_COMPUTE, stderr, true);
if (si_shader_binary_upload(sctx->screen, &program->shader) < 0) {
fprintf(stderr, "LLVM failed to upload shader\n");
FREE(program);
return NULL;
}
}
return program;
}
static void *si_create_compute_state(
struct pipe_context *ctx,
const struct pipe_compute_state *cso)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_screen *sscreen = (struct si_screen *)ctx->screen;
struct si_compute *program = CALLOC_STRUCT(si_compute);
program->screen = (struct si_screen *)ctx->screen;
program->ir_type = cso->ir_type;
program->local_size = cso->req_local_mem;
program->private_size = cso->req_private_mem;
program->input_size = cso->req_input_mem;
program->use_code_object_v2 = HAVE_LLVM >= 0x0400 &&
cso->ir_type == PIPE_SHADER_IR_NATIVE;
if (cso->ir_type == PIPE_SHADER_IR_TGSI) {
program->tokens = tgsi_dup_tokens(cso->prog);
if (!program->tokens) {
FREE(program);
return NULL;
}
program->compiler_ctx_state.tm = sctx->tm;
program->compiler_ctx_state.debug = sctx->b.debug;
program->compiler_ctx_state.is_debug_context = sctx->is_debug;
p_atomic_inc(&sscreen->b.num_shaders_created);
util_queue_fence_init(&program->ready);
if ((sctx->b.debug.debug_message && !sctx->b.debug.async) ||
sctx->is_debug ||
r600_can_dump_shader(&sscreen->b, PIPE_SHADER_COMPUTE))
si_create_compute_state_async(program, -1);
else
util_queue_add_job(&sscreen->shader_compiler_queue,
program, &program->ready,
si_create_compute_state_async, NULL);
} else {
const struct pipe_llvm_program_header *header;
const char *code;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
ac_elf_read(code, header->num_bytes, &program->shader.binary);
if (program->use_code_object_v2) {
const amd_kernel_code_t *code_object =
si_compute_get_code_object(program, 0);
code_object_to_config(code_object, &program->shader.config);
} else {
si_shader_binary_read_config(&program->shader.binary,
&program->shader.config, 0);
}
si_shader_dump(sctx->screen, &program->shader, &sctx->b.debug,
PIPE_SHADER_COMPUTE, stderr, true);
if (si_shader_binary_upload(sctx->screen, &program->shader) < 0) {
fprintf(stderr, "LLVM failed to upload shader\n");
FREE(program);
return NULL;
}
}
return program;
}
