struct ir3_shader *
ir3_shader_create(struct ir3_compiler *compiler,
const struct pipe_shader_state *cso, gl_shader_stage type,
struct pipe_debug_callback *debug,
struct pipe_screen *screen)
{
nir_shader *nir;
if (cso->type == PIPE_SHADER_IR_NIR) {
/* we take ownership of the reference: */
nir = cso->ir.nir;
} else {
debug_assert(cso->type == PIPE_SHADER_IR_TGSI);
if (ir3_shader_debug & IR3_DBG_DISASM) {
tgsi_dump(cso->tokens, 0);
}
nir = tgsi_to_nir(cso->tokens, screen);
}
struct ir3_shader *shader = ir3_shader_from_nir(compiler, nir);
copy_stream_out(&shader->stream_output, &cso->stream_output);
if (fd_mesa_debug & FD_DBG_SHADERDB) {
/* if shader-db run, create a standard variant immediately
* (as otherwise nothing will trigger the shader to be
* actually compiled)
*/
static struct ir3_shader_key key;
memset(&key, 0, sizeof(key));
ir3_shader_variant(shader, key, false, debug);
}
return shader;
}
/* a bit annoying that compute-shader and normal shader state objects
* aren't a bit more aligned.
*/
struct ir3_shader *
ir3_shader_create_compute(struct ir3_compiler *compiler,
const struct pipe_compute_state *cso,
struct pipe_debug_callback *debug,
struct pipe_screen *screen)
{
nir_shader *nir;
if (cso->ir_type == PIPE_SHADER_IR_NIR) {
/* we take ownership of the reference: */
nir = (nir_shader *)cso->prog;
} else {
debug_assert(cso->ir_type == PIPE_SHADER_IR_TGSI);
if (ir3_shader_debug & IR3_DBG_DISASM) {
tgsi_dump(cso->prog, 0);
}
nir = tgsi_to_nir(cso->prog, screen);
}
struct ir3_shader *shader = ir3_shader_from_nir(compiler, nir);
return shader;
}
struct nir_shader *
ir3_tgsi_to_nir(const struct tgsi_token *tokens)
{
return tgsi_to_nir(tokens, &options);
}
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;
}
