#include "util/debug.h"
+#include "util/u_math.h"
#include "ir3_nir.h"
#include "ir3_compiler.h"
#include "ir3_shader.h"
+static void ir3_setup_const_state(struct ir3_shader *shader, nir_shader *nir);
+
static const nir_shader_compiler_options options = {
.lower_fpow = true,
.lower_scmp = true,
.lower_flrp32 = true,
.lower_flrp64 = true,
.lower_ffract = true,
- .lower_fmod32 = true,
- .lower_fmod64 = true,
+ .lower_fmod = true,
.lower_fdiv = true,
.lower_isign = true,
.lower_ldexp = true,
.lower_uadd_carry = true,
.lower_mul_high = true,
.fuse_ffma = true,
- .native_integers = true,
.vertex_id_zero_based = true,
.lower_extract_byte = true,
.lower_extract_word = true,
- .lower_all_io_to_temps = true,
+ .lower_all_io_to_elements = true,
.lower_helper_invocation = true,
.lower_bitfield_insert_to_shifts = true,
.lower_bitfield_extract_to_shifts = true,
- .lower_bfm = true,
+ .use_interpolated_input_intrinsics = true,
+ .lower_rotate = true,
+ .lower_to_scalar = true,
};
/* we don't want to lower vertex_id to _zero_based on newer gpus: */
.lower_flrp32 = true,
.lower_flrp64 = true,
.lower_ffract = true,
- .lower_fmod32 = true,
- .lower_fmod64 = true,
+ .lower_fmod = true,
.lower_fdiv = true,
.lower_isign = true,
.lower_ldexp = true,
.lower_uadd_carry = true,
.lower_mul_high = true,
.fuse_ffma = true,
- .native_integers = true,
.vertex_id_zero_based = false,
.lower_extract_byte = true,
.lower_extract_word = true,
- .lower_all_io_to_temps = true,
+ .lower_all_io_to_elements = true,
.lower_helper_invocation = true,
.lower_bitfield_insert_to_shifts = true,
.lower_bitfield_extract_to_shifts = true,
- .lower_bfm = true,
+ .use_interpolated_input_intrinsics = true,
+ .lower_rotate = true,
+ .vectorize_io = true,
+ .lower_to_scalar = true,
};
const nir_shader_compiler_options *
ir3_optimize_loop(nir_shader *s)
{
bool progress;
+ unsigned lower_flrp =
+ (s->options->lower_flrp16 ? 16 : 0) |
+ (s->options->lower_flrp32 ? 32 : 0) |
+ (s->options->lower_flrp64 ? 64 : 0);
+
do {
progress = false;
OPT_V(s, nir_lower_vars_to_ssa);
progress |= OPT(s, nir_opt_copy_prop_vars);
progress |= OPT(s, nir_opt_dead_write_vars);
- progress |= OPT(s, nir_lower_alu_to_scalar);
+ progress |= OPT(s, nir_lower_alu_to_scalar, NULL, NULL);
progress |= OPT(s, nir_lower_phis_to_scalar);
progress |= OPT(s, nir_copy_prop);
progress |= OPT(s, nir_opt_intrinsics);
progress |= OPT(s, nir_opt_algebraic);
progress |= OPT(s, nir_opt_constant_folding);
+
+ if (lower_flrp != 0) {
+ if (OPT(s, nir_lower_flrp,
+ lower_flrp,
+ false /* always_precise */,
+ s->options->lower_ffma)) {
+ OPT(s, nir_opt_constant_folding);
+ progress = true;
+ }
+
+ /* Nothing should rematerialize any flrps, so we only
+ * need to do this lowering once.
+ */
+ lower_flrp = 0;
+ }
+
progress |= OPT(s, nir_opt_dead_cf);
if (OPT(s, nir_opt_trivial_continues)) {
progress |= true;
OPT(s, nir_copy_prop);
OPT(s, nir_opt_dce);
}
- progress |= OPT(s, nir_opt_if);
+ progress |= OPT(s, nir_opt_if, false);
progress |= OPT(s, nir_opt_remove_phis);
progress |= OPT(s, nir_opt_undef);
} while (progress);
}
-struct nir_shader *
+void
ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
const struct ir3_shader_key *key)
{
debug_printf("----------------------\n");
}
- OPT_V(s, nir_opt_global_to_local);
OPT_V(s, nir_lower_regs_to_ssa);
OPT_V(s, ir3_nir_lower_io_offsets);
if (key) {
if (s->info.stage == MESA_SHADER_VERTEX) {
- OPT_V(s, nir_lower_clip_vs, key->ucp_enables, false);
+ OPT_V(s, nir_lower_clip_vs, key->ucp_enables, false, false);
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);
+ OPT_V(s, nir_lower_clip_fs, key->ucp_enables, false);
if (key->fclamp_color)
OPT_V(s, nir_lower_clamp_color_outputs);
}
* and not again on any potential 2nd variant lowering pass:
*/
OPT_V(s, ir3_nir_apply_trig_workarounds);
+
+ /* This wouldn't hurt to run multiple times, but there is
+ * no need to:
+ */
+ if (shader->type == MESA_SHADER_FRAGMENT)
+ OPT_V(s, nir_lower_fb_read);
}
OPT_V(s, nir_lower_tex, &tex_options);
/* do ubo load and idiv lowering after first opt loop to get a chance to
* propagate constants for divide by immed power-of-two and constant ubo
* block/offsets:
+ *
+ * NOTE that UBO analysis pass should only be done once, before variants
*/
- const bool ubo_progress = OPT(s, ir3_nir_analyze_ubo_ranges, shader);
+ const bool ubo_progress = !key && OPT(s, ir3_nir_analyze_ubo_ranges, shader);
const bool idiv_progress = OPT(s, nir_lower_idiv);
if (ubo_progress || idiv_progress)
ir3_optimize_loop(s);
+ /* Do late algebraic optimization to turn add(a, neg(b)) back into
+ * subs, then the mandatory cleanup after algebraic. Note that it may
+ * produce fnegs, and if so then we need to keep running to squash
+ * fneg(fneg(a)).
+ */
+ bool more_late_algebraic = true;
+ while (more_late_algebraic) {
+ more_late_algebraic = OPT(s, nir_opt_algebraic_late);
+ OPT_V(s, nir_opt_constant_folding);
+ OPT_V(s, nir_copy_prop);
+ OPT_V(s, nir_opt_dce);
+ OPT_V(s, nir_opt_cse);
+ }
+
OPT_V(s, nir_remove_dead_variables, nir_var_function_temp);
- OPT_V(s, nir_move_load_const);
+ OPT_V(s, nir_opt_sink, nir_move_const_undef);
if (ir3_shader_debug & IR3_DBG_DISASM) {
debug_printf("----------------------\n");
nir_sweep(s);
- return s;
+ /* The first time thru, when not creating variant, do the one-time
+ * const_state layout setup. This should be done after ubo range
+ * analysis.
+ */
+ if (!key) {
+ ir3_setup_const_state(shader, s);
+ }
}
-void
+static void
ir3_nir_scan_driver_consts(nir_shader *shader,
- struct ir3_driver_const_layout *layout)
+ struct ir3_const_state *layout)
{
nir_foreach_function(function, shader) {
if (!function->impl)
switch (intr->intrinsic) {
case nir_intrinsic_get_buffer_size:
- idx = nir_src_as_const_value(intr->src[0])->u32[0];
+ idx = nir_src_as_uint(intr->src[0]);
if (layout->ssbo_size.mask & (1 << idx))
break;
layout->ssbo_size.mask |= (1 << idx);
layout->ssbo_size.count += 1; /* one const per */
break;
case nir_intrinsic_image_deref_atomic_add:
- case nir_intrinsic_image_deref_atomic_min:
- case nir_intrinsic_image_deref_atomic_max:
+ case nir_intrinsic_image_deref_atomic_imin:
+ case nir_intrinsic_image_deref_atomic_umin:
+ case nir_intrinsic_image_deref_atomic_imax:
+ case nir_intrinsic_image_deref_atomic_umax:
case nir_intrinsic_image_deref_atomic_and:
case nir_intrinsic_image_deref_atomic_or:
case nir_intrinsic_image_deref_atomic_xor:
layout->image_dims.count;
layout->image_dims.count += 3; /* three const per */
break;
+ case nir_intrinsic_load_ubo:
+ if (nir_src_is_const(intr->src[0])) {
+ layout->num_ubos = MAX2(layout->num_ubos,
+ nir_src_as_uint(intr->src[0]) + 1);
+ } else {
+ layout->num_ubos = shader->info.num_ubos;
+ }
+ break;
+ case nir_intrinsic_load_base_vertex:
+ case nir_intrinsic_load_first_vertex:
+ layout->num_driver_params =
+ MAX2(layout->num_driver_params, IR3_DP_VTXID_BASE + 1);
+ break;
+ case nir_intrinsic_load_user_clip_plane:
+ layout->num_driver_params =
+ MAX2(layout->num_driver_params, IR3_DP_UCP7_W + 1);
+ break;
+ case nir_intrinsic_load_num_work_groups:
+ layout->num_driver_params =
+ MAX2(layout->num_driver_params, IR3_DP_NUM_WORK_GROUPS_Z + 1);
+ break;
+ case nir_intrinsic_load_local_group_size:
+ layout->num_driver_params =
+ MAX2(layout->num_driver_params, IR3_DP_LOCAL_GROUP_SIZE_Z + 1);
+ break;
default:
break;
}
}
}
}
+
+static void
+ir3_setup_const_state(struct ir3_shader *shader, nir_shader *nir)
+{
+ struct ir3_compiler *compiler = shader->compiler;
+ struct ir3_const_state *const_state = &shader->const_state;
+
+ memset(&const_state->offsets, ~0, sizeof(const_state->offsets));
+
+ ir3_nir_scan_driver_consts(nir, const_state);
+
+ if ((compiler->gpu_id < 500) &&
+ (shader->stream_output.num_outputs > 0)) {
+ const_state->num_driver_params =
+ MAX2(const_state->num_driver_params, IR3_DP_VTXCNT_MAX + 1);
+ }
+
+ /* num_driver_params is scalar, align to vec4: */
+ const_state->num_driver_params = align(const_state->num_driver_params, 4);
+
+ debug_assert((shader->ubo_state.size % 16) == 0);
+ unsigned constoff = align(shader->ubo_state.size / 16, 8);
+ unsigned ptrsz = ir3_pointer_size(compiler);
+
+ if (const_state->num_ubos > 0) {
+ const_state->offsets.ubo = constoff;
+ constoff += align(nir->info.num_ubos * ptrsz, 4) / 4;
+ }
+
+ if (const_state->ssbo_size.count > 0) {
+ unsigned cnt = const_state->ssbo_size.count;
+ const_state->offsets.ssbo_sizes = constoff;
+ constoff += align(cnt, 4) / 4;
+ }
+
+ if (const_state->image_dims.count > 0) {
+ unsigned cnt = const_state->image_dims.count;
+ const_state->offsets.image_dims = constoff;
+ constoff += align(cnt, 4) / 4;
+ }
+
+ if (const_state->num_driver_params > 0)
+ const_state->offsets.driver_param = constoff;
+ constoff += const_state->num_driver_params / 4;
+
+ if ((shader->type == MESA_SHADER_VERTEX) &&
+ (compiler->gpu_id < 500) &&
+ shader->stream_output.num_outputs > 0) {
+ const_state->offsets.tfbo = constoff;
+ constoff += align(IR3_MAX_SO_BUFFERS * ptrsz, 4) / 4;
+ }
+
+ const_state->offsets.immediate = constoff;
+}