struct nir_shader *
ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
const struct ir3_shader_key *key)
{
struct nir_lower_tex_options tex_options = {
.lower_rect = 0,
};
if (key) {
switch (shader->type) {
case SHADER_FRAGMENT:
tex_options.saturate_s = key->fsaturate_s;
tex_options.saturate_t = key->fsaturate_t;
tex_options.saturate_r = key->fsaturate_r;
break;
case SHADER_VERTEX:
tex_options.saturate_s = key->vsaturate_s;
tex_options.saturate_t = key->vsaturate_t;
tex_options.saturate_r = key->vsaturate_r;
break;
default:
/* TODO */
break;
}
}
if (shader->compiler->gpu_id >= 400) {
/* a4xx seems to have *no* sam.p */
tex_options.lower_txp = ~0; /* lower all txp */
} else {
/* a3xx just needs to avoid sam.p for 3d tex */
tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
}
if (fd_mesa_debug & FD_DBG_DISASM) {
debug_printf("----------------------\n");
nir_print_shader(s, stdout);
debug_printf("----------------------\n");
}
OPT_V(s, nir_opt_global_to_local);
OPT_V(s, nir_lower_regs_to_ssa);
if (key) {
if (s->info.stage == MESA_SHADER_VERTEX) {
OPT_V(s, nir_lower_clip_vs, key->ucp_enables);
if (key->vclamp_color)
OPT_V(s, nir_lower_clamp_color_outputs);
} else if (s->info.stage == MESA_SHADER_FRAGMENT) {
OPT_V(s, nir_lower_clip_fs, key->ucp_enables);
if (key->fclamp_color)
OPT_V(s, nir_lower_clamp_color_outputs);
}
if (key->color_two_side) {
OPT_V(s, nir_lower_two_sided_color);
}
} else {
/* only want to do this the first time (when key is null)
* and not again on any potential 2nd variant lowering pass:
*/
OPT_V(s, ir3_nir_apply_trig_workarounds);
}
OPT_V(s, nir_lower_tex, &tex_options);
OPT_V(s, nir_lower_load_const_to_scalar);
ir3_optimize_loop(s);
/* do idiv lowering after first opt loop to give a chance for
* divide by immed power-of-two to be caught first:
*/
if (OPT(s, nir_lower_idiv))
ir3_optimize_loop(s);
OPT_V(s, nir_remove_dead_variables, nir_var_local);
if (fd_mesa_debug & FD_DBG_DISASM) {
debug_printf("----------------------\n");
nir_print_shader(s, stdout);
debug_printf("----------------------\n");
}
nir_sweep(s);
return s;
}
nir_shader *
brw_create_nir(struct brw_context *brw,
const struct gl_shader_program *shader_prog,
const struct gl_program *prog,
gl_shader_stage stage,
bool is_scalar)
{
struct gl_context *ctx = &brw->ctx;
const nir_shader_compiler_options *options =
ctx->Const.ShaderCompilerOptions[stage].NirOptions;
struct gl_shader *shader = shader_prog ? shader_prog->_LinkedShaders[stage] : NULL;
bool debug_enabled = INTEL_DEBUG & intel_debug_flag_for_shader_stage(stage);
nir_shader *nir;
/* First, lower the GLSL IR or Mesa IR to NIR */
if (shader_prog) {
nir = glsl_to_nir(shader, options);
} else {
nir = prog_to_nir(prog, options);
nir_convert_to_ssa(nir); /* turn registers into SSA */
}
nir_validate_shader(nir);
nir_lower_global_vars_to_local(nir);
nir_validate_shader(nir);
nir_lower_tex_projector(nir);
nir_validate_shader(nir);
nir_normalize_cubemap_coords(nir);
nir_validate_shader(nir);
nir_split_var_copies(nir);
nir_validate_shader(nir);
nir_optimize(nir, is_scalar);
/* Lower a bunch of stuff */
nir_lower_var_copies(nir);
nir_validate_shader(nir);
/* Get rid of split copies */
nir_optimize(nir, is_scalar);
if (is_scalar) {
nir_assign_var_locations(&nir->uniforms,
&nir->num_uniforms,
type_size_scalar);
nir_assign_var_locations(&nir->inputs, &nir->num_inputs, type_size_scalar);
nir_assign_var_locations(&nir->outputs, &nir->num_outputs, type_size_scalar);
nir_lower_io(nir, type_size_scalar);
} else {
nir_assign_var_locations(&nir->uniforms,
&nir->num_uniforms,
type_size_vec4);
nir_assign_var_locations(&nir->inputs, &nir->num_inputs, type_size_vec4);
foreach_list_typed(nir_variable, var, node, &nir->outputs)
var->data.driver_location = var->data.location;
nir_lower_io(nir, type_size_vec4);
}
nir_validate_shader(nir);
nir_remove_dead_variables(nir);
nir_validate_shader(nir);
if (shader_prog) {
nir_lower_samplers(nir, shader_prog);
nir_validate_shader(nir);
}
nir_lower_system_values(nir);
nir_validate_shader(nir);
nir_lower_atomics(nir);
nir_validate_shader(nir);
nir_optimize(nir, is_scalar);
if (brw->gen >= 6) {
/* Try and fuse multiply-adds */
nir_opt_peephole_ffma(nir);
nir_validate_shader(nir);
}
nir_opt_algebraic_late(nir);
nir_validate_shader(nir);
nir_lower_locals_to_regs(nir);
nir_validate_shader(nir);
nir_lower_to_source_mods(nir);
nir_validate_shader(nir);
nir_copy_prop(nir);
nir_validate_shader(nir);
nir_opt_dce(nir);
nir_validate_shader(nir);
if (unlikely(debug_enabled)) {
/* Re-index SSA defs so we print more sensible numbers. */
nir_foreach_overload(nir, overload) {
if (overload->impl)
nir_index_ssa_defs(overload->impl);
}
fprintf(stderr, "NIR (SSA form) for %s shader:\n",
_mesa_shader_stage_to_string(stage));
nir_print_shader(nir, stderr);
}
nir_convert_from_ssa(nir, is_scalar);
nir_validate_shader(nir);
if (!is_scalar) {
nir_lower_vec_to_movs(nir);
nir_validate_shader(nir);
}
/* This is the last pass we run before we start emitting stuff. It
* determines when we need to insert boolean resolves on Gen <= 5. We
* run it last because it stashes data in instr->pass_flags and we don't
* want that to be squashed by other NIR passes.
*/
if (brw->gen <= 5)
brw_nir_analyze_boolean_resolves(nir);
nir_sweep(nir);
if (unlikely(debug_enabled)) {
fprintf(stderr, "NIR (final form) for %s shader:\n",
_mesa_shader_stage_to_string(stage));
nir_print_shader(nir, stderr);
}
return nir;
}
struct nir_shader *
ir3_tgsi_to_nir(const struct tgsi_token *tokens)
{
static const nir_shader_compiler_options options = {
.lower_fpow = true,
.lower_fsat = true,
.lower_scmp = true,
.lower_flrp32 = true,
.lower_flrp64 = true,
.lower_ffract = true,
.fuse_ffma = true,
.native_integers = true,
.vertex_id_zero_based = true,
.lower_extract_byte = true,
.lower_extract_word = true,
};
return tgsi_to_nir(tokens, &options);
}
/* for given shader key, are any steps handled in nir? */
bool
ir3_key_lowers_nir(const struct ir3_shader_key *key)
{
return key->fsaturate_s | key->fsaturate_t | key->fsaturate_r |
key->vsaturate_s | key->vsaturate_t | key->vsaturate_r |
key->ucp_enables | key->color_two_side |
key->fclamp_color | key->vclamp_color;
}
#define OPT(nir, pass, ...) ({ \
bool this_progress = false; \
NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
this_progress; \
})
#define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
struct nir_shader *
ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
const struct ir3_shader_key *key)
{
struct nir_lower_tex_options tex_options = {
.lower_rect = 0,
};
bool progress;
if (key) {
switch (shader->type) {
case SHADER_FRAGMENT:
case SHADER_COMPUTE:
tex_options.saturate_s = key->fsaturate_s;
tex_options.saturate_t = key->fsaturate_t;
tex_options.saturate_r = key->fsaturate_r;
break;
case SHADER_VERTEX:
tex_options.saturate_s = key->vsaturate_s;
tex_options.saturate_t = key->vsaturate_t;
tex_options.saturate_r = key->vsaturate_r;
break;
}
}
if (shader->compiler->gpu_id >= 400) {
/* a4xx seems to have *no* sam.p */
tex_options.lower_txp = ~0; /* lower all txp */
} else {
/* a3xx just needs to avoid sam.p for 3d tex */
tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
}
if (fd_mesa_debug & FD_DBG_DISASM) {
debug_printf("----------------------\n");
nir_print_shader(s, stdout);
debug_printf("----------------------\n");
}
OPT_V(s, nir_opt_global_to_local);
OPT_V(s, nir_convert_to_ssa);
if (key) {
if (s->stage == MESA_SHADER_VERTEX) {
OPT_V(s, nir_lower_clip_vs, key->ucp_enables);
if (key->vclamp_color)
OPT_V(s, nir_lower_clamp_color_outputs);
} else if (s->stage == MESA_SHADER_FRAGMENT) {
OPT_V(s, nir_lower_clip_fs, key->ucp_enables);
if (key->fclamp_color)
OPT_V(s, nir_lower_clamp_color_outputs);
}
if (key->color_two_side) {
OPT_V(s, nir_lower_two_sided_color);
}
} else {
/* only want to do this the first time (when key is null)
* and not again on any potential 2nd variant lowering pass:
*/
OPT_V(s, ir3_nir_apply_trig_workarounds);
}
OPT_V(s, nir_lower_tex, &tex_options);
OPT_V(s, nir_lower_idiv);
OPT_V(s, nir_lower_load_const_to_scalar);
do {
progress = false;
OPT_V(s, nir_lower_vars_to_ssa);
OPT_V(s, nir_lower_alu_to_scalar);
OPT_V(s, nir_lower_phis_to_scalar);
progress |= OPT(s, nir_copy_prop);
progress |= OPT(s, nir_opt_dce);
progress |= OPT(s, nir_opt_cse);
progress |= OPT(s, ir3_nir_lower_if_else);
progress |= OPT(s, nir_opt_algebraic);
progress |= OPT(s, nir_opt_constant_folding);
} while (progress);
OPT_V(s, nir_remove_dead_variables, nir_var_local);
if (fd_mesa_debug & FD_DBG_DISASM) {
debug_printf("----------------------\n");
nir_print_shader(s, stdout);
debug_printf("----------------------\n");
}
nir_sweep(s);
return s;
}
