2 * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Rob Clark <robclark@freedesktop.org>
28 #include "util/debug.h"
29 #include "util/u_math.h"
32 #include "ir3_compiler.h"
33 #include "ir3_shader.h"
35 static void ir3_setup_const_state(struct ir3_shader
*shader
, nir_shader
*nir
);
37 static const nir_shader_compiler_options options
= {
48 .lower_uadd_carry
= true,
49 .lower_usub_borrow
= true,
50 .lower_mul_high
= true,
51 .lower_mul_2x32_64
= true,
53 .vertex_id_zero_based
= true,
54 .lower_extract_byte
= true,
55 .lower_extract_word
= true,
56 .lower_all_io_to_elements
= true,
57 .lower_helper_invocation
= true,
58 .lower_bitfield_insert_to_shifts
= true,
59 .lower_bitfield_extract_to_shifts
= true,
60 .lower_pack_half_2x16
= true,
61 .lower_pack_half_2x16_split
= true,
62 .lower_pack_snorm_4x8
= true,
63 .lower_pack_snorm_2x16
= true,
64 .lower_pack_unorm_4x8
= true,
65 .lower_pack_unorm_2x16
= true,
66 .lower_unpack_half_2x16
= true,
67 .lower_unpack_half_2x16_split
= true,
68 .lower_unpack_snorm_4x8
= true,
69 .lower_unpack_snorm_2x16
= true,
70 .lower_unpack_unorm_4x8
= true,
71 .lower_unpack_unorm_2x16
= true,
72 .use_interpolated_input_intrinsics
= true,
74 .lower_to_scalar
= true,
78 /* we don't want to lower vertex_id to _zero_based on newer gpus: */
79 static const nir_shader_compiler_options options_a6xx
= {
90 .lower_uadd_carry
= true,
91 .lower_usub_borrow
= true,
92 .lower_mul_high
= true,
93 .lower_mul_2x32_64
= true,
95 .vertex_id_zero_based
= false,
96 .lower_extract_byte
= true,
97 .lower_extract_word
= true,
98 .lower_all_io_to_elements
= true,
99 .lower_helper_invocation
= true,
100 .lower_bitfield_insert_to_shifts
= true,
101 .lower_bitfield_extract_to_shifts
= true,
102 .lower_pack_half_2x16
= true,
103 .lower_pack_half_2x16_split
= true,
104 .lower_pack_snorm_4x8
= true,
105 .lower_pack_snorm_2x16
= true,
106 .lower_pack_unorm_4x8
= true,
107 .lower_pack_unorm_2x16
= true,
108 .lower_unpack_half_2x16
= true,
109 .lower_unpack_half_2x16_split
= true,
110 .lower_unpack_snorm_4x8
= true,
111 .lower_unpack_snorm_2x16
= true,
112 .lower_unpack_unorm_4x8
= true,
113 .lower_unpack_unorm_2x16
= true,
114 .use_interpolated_input_intrinsics
= true,
115 .lower_rotate
= true,
116 .vectorize_io
= true,
117 .lower_to_scalar
= true,
121 const nir_shader_compiler_options
*
122 ir3_get_compiler_options(struct ir3_compiler
*compiler
)
124 if (compiler
->gpu_id
>= 600)
125 return &options_a6xx
;
129 /* for given shader key, are any steps handled in nir? */
131 ir3_key_lowers_nir(const struct ir3_shader_key
*key
)
133 return key
->fsaturate_s
| key
->fsaturate_t
| key
->fsaturate_r
|
134 key
->vsaturate_s
| key
->vsaturate_t
| key
->vsaturate_r
|
135 key
->ucp_enables
| key
->color_two_side
|
136 key
->fclamp_color
| key
->vclamp_color
|
137 key
->tessellation
| key
->has_gs
;
140 #define OPT(nir, pass, ...) ({ \
141 bool this_progress = false; \
142 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
146 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
149 ir3_optimize_loop(nir_shader
*s
)
152 unsigned lower_flrp
=
153 (s
->options
->lower_flrp16
? 16 : 0) |
154 (s
->options
->lower_flrp32
? 32 : 0) |
155 (s
->options
->lower_flrp64
? 64 : 0);
160 OPT_V(s
, nir_lower_vars_to_ssa
);
161 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
162 progress
|= OPT(s
, nir_opt_dead_write_vars
);
163 progress
|= OPT(s
, nir_lower_alu_to_scalar
, NULL
, NULL
);
164 progress
|= OPT(s
, nir_lower_phis_to_scalar
);
166 progress
|= OPT(s
, nir_copy_prop
);
167 progress
|= OPT(s
, nir_opt_dce
);
168 progress
|= OPT(s
, nir_opt_cse
);
171 gcm
= env_var_as_unsigned("GCM", 0);
173 progress
|= OPT(s
, nir_opt_gcm
, true);
175 progress
|= OPT(s
, nir_opt_gcm
, false);
176 progress
|= OPT(s
, nir_opt_peephole_select
, 16, true, true);
177 progress
|= OPT(s
, nir_opt_intrinsics
);
178 progress
|= OPT(s
, nir_opt_algebraic
);
179 progress
|= OPT(s
, nir_lower_alu
);
180 progress
|= OPT(s
, nir_opt_constant_folding
);
182 if (lower_flrp
!= 0) {
183 if (OPT(s
, nir_lower_flrp
,
185 false /* always_precise */,
186 s
->options
->lower_ffma
)) {
187 OPT(s
, nir_opt_constant_folding
);
191 /* Nothing should rematerialize any flrps, so we only
192 * need to do this lowering once.
197 progress
|= OPT(s
, nir_opt_dead_cf
);
198 if (OPT(s
, nir_opt_trivial_continues
)) {
200 /* If nir_opt_trivial_continues makes progress, then we need to clean
201 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
204 OPT(s
, nir_copy_prop
);
207 progress
|= OPT(s
, nir_opt_if
, false);
208 progress
|= OPT(s
, nir_opt_remove_phis
);
209 progress
|= OPT(s
, nir_opt_undef
);
215 ir3_optimize_nir(struct ir3_shader
*shader
, nir_shader
*s
,
216 const struct ir3_shader_key
*key
)
218 struct nir_lower_tex_options tex_options
= {
220 .lower_tg4_offsets
= true,
223 if (key
&& (key
->has_gs
|| key
->tessellation
)) {
224 switch (shader
->type
) {
225 case MESA_SHADER_VERTEX
:
226 NIR_PASS_V(s
, ir3_nir_lower_to_explicit_io
, shader
, key
->tessellation
);
228 case MESA_SHADER_TESS_CTRL
:
229 NIR_PASS_V(s
, ir3_nir_lower_tess_ctrl
, shader
, key
->tessellation
);
231 case MESA_SHADER_TESS_EVAL
:
232 NIR_PASS_V(s
, ir3_nir_lower_tess_eval
, key
->tessellation
);
234 NIR_PASS_V(s
, ir3_nir_lower_to_explicit_io
, shader
, key
->tessellation
);
236 case MESA_SHADER_GEOMETRY
:
237 NIR_PASS_V(s
, ir3_nir_lower_gs
, shader
);
245 switch (shader
->type
) {
246 case MESA_SHADER_FRAGMENT
:
247 tex_options
.saturate_s
= key
->fsaturate_s
;
248 tex_options
.saturate_t
= key
->fsaturate_t
;
249 tex_options
.saturate_r
= key
->fsaturate_r
;
251 case MESA_SHADER_VERTEX
:
252 tex_options
.saturate_s
= key
->vsaturate_s
;
253 tex_options
.saturate_t
= key
->vsaturate_t
;
254 tex_options
.saturate_r
= key
->vsaturate_r
;
262 if (shader
->compiler
->gpu_id
>= 400) {
263 /* a4xx seems to have *no* sam.p */
264 tex_options
.lower_txp
= ~0; /* lower all txp */
266 /* a3xx just needs to avoid sam.p for 3d tex */
267 tex_options
.lower_txp
= (1 << GLSL_SAMPLER_DIM_3D
);
270 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
271 debug_printf("----------------------\n");
272 nir_print_shader(s
, stdout
);
273 debug_printf("----------------------\n");
276 OPT_V(s
, nir_lower_regs_to_ssa
);
277 OPT_V(s
, ir3_nir_lower_io_offsets
);
280 if (s
->info
.stage
== MESA_SHADER_VERTEX
) {
281 OPT_V(s
, nir_lower_clip_vs
, key
->ucp_enables
, false, false, NULL
);
282 if (key
->vclamp_color
)
283 OPT_V(s
, nir_lower_clamp_color_outputs
);
284 } else if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
285 OPT_V(s
, nir_lower_clip_fs
, key
->ucp_enables
, false);
286 if (key
->fclamp_color
)
287 OPT_V(s
, nir_lower_clamp_color_outputs
);
289 if (key
->color_two_side
) {
290 OPT_V(s
, nir_lower_two_sided_color
);
293 /* only want to do this the first time (when key is null)
294 * and not again on any potential 2nd variant lowering pass:
296 OPT_V(s
, ir3_nir_apply_trig_workarounds
);
298 /* This wouldn't hurt to run multiple times, but there is
301 if (shader
->type
== MESA_SHADER_FRAGMENT
)
302 OPT_V(s
, nir_lower_fb_read
);
305 OPT_V(s
, nir_lower_tex
, &tex_options
);
306 OPT_V(s
, nir_lower_load_const_to_scalar
);
307 if (shader
->compiler
->gpu_id
< 500)
308 OPT_V(s
, ir3_nir_lower_tg4_to_tex
);
310 ir3_optimize_loop(s
);
312 /* do ubo load and idiv lowering after first opt loop to get a chance to
313 * propagate constants for divide by immed power-of-two and constant ubo
316 * NOTE that UBO analysis pass should only be done once, before variants
318 const bool ubo_progress
= !key
&& OPT(s
, ir3_nir_analyze_ubo_ranges
, shader
);
319 const bool idiv_progress
= OPT(s
, nir_lower_idiv
, nir_lower_idiv_fast
);
320 if (ubo_progress
|| idiv_progress
)
321 ir3_optimize_loop(s
);
323 /* Do late algebraic optimization to turn add(a, neg(b)) back into
324 * subs, then the mandatory cleanup after algebraic. Note that it may
325 * produce fnegs, and if so then we need to keep running to squash
328 bool more_late_algebraic
= true;
329 while (more_late_algebraic
) {
330 more_late_algebraic
= OPT(s
, nir_opt_algebraic_late
);
331 OPT_V(s
, nir_opt_constant_folding
);
332 OPT_V(s
, nir_copy_prop
);
333 OPT_V(s
, nir_opt_dce
);
334 OPT_V(s
, nir_opt_cse
);
337 OPT_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
339 OPT_V(s
, nir_opt_sink
, nir_move_const_undef
);
341 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
342 debug_printf("----------------------\n");
343 nir_print_shader(s
, stdout
);
344 debug_printf("----------------------\n");
349 /* The first time thru, when not creating variant, do the one-time
350 * const_state layout setup. This should be done after ubo range
354 ir3_setup_const_state(shader
, s
);
359 ir3_nir_scan_driver_consts(nir_shader
*shader
,
360 struct ir3_const_state
*layout
)
362 nir_foreach_function(function
, shader
) {
366 nir_foreach_block(block
, function
->impl
) {
367 nir_foreach_instr(instr
, block
) {
368 if (instr
->type
!= nir_instr_type_intrinsic
)
371 nir_intrinsic_instr
*intr
=
372 nir_instr_as_intrinsic(instr
);
375 switch (intr
->intrinsic
) {
376 case nir_intrinsic_get_buffer_size
:
377 idx
= nir_src_as_uint(intr
->src
[0]);
378 if (layout
->ssbo_size
.mask
& (1 << idx
))
380 layout
->ssbo_size
.mask
|= (1 << idx
);
381 layout
->ssbo_size
.off
[idx
] =
382 layout
->ssbo_size
.count
;
383 layout
->ssbo_size
.count
+= 1; /* one const per */
385 case nir_intrinsic_image_deref_atomic_add
:
386 case nir_intrinsic_image_deref_atomic_imin
:
387 case nir_intrinsic_image_deref_atomic_umin
:
388 case nir_intrinsic_image_deref_atomic_imax
:
389 case nir_intrinsic_image_deref_atomic_umax
:
390 case nir_intrinsic_image_deref_atomic_and
:
391 case nir_intrinsic_image_deref_atomic_or
:
392 case nir_intrinsic_image_deref_atomic_xor
:
393 case nir_intrinsic_image_deref_atomic_exchange
:
394 case nir_intrinsic_image_deref_atomic_comp_swap
:
395 case nir_intrinsic_image_deref_store
:
396 case nir_intrinsic_image_deref_size
:
397 idx
= nir_intrinsic_get_var(intr
, 0)->data
.driver_location
;
398 if (layout
->image_dims
.mask
& (1 << idx
))
400 layout
->image_dims
.mask
|= (1 << idx
);
401 layout
->image_dims
.off
[idx
] =
402 layout
->image_dims
.count
;
403 layout
->image_dims
.count
+= 3; /* three const per */
405 case nir_intrinsic_load_ubo
:
406 if (nir_src_is_const(intr
->src
[0])) {
407 layout
->num_ubos
= MAX2(layout
->num_ubos
,
408 nir_src_as_uint(intr
->src
[0]) + 1);
410 layout
->num_ubos
= shader
->info
.num_ubos
;
413 case nir_intrinsic_load_base_vertex
:
414 case nir_intrinsic_load_first_vertex
:
415 layout
->num_driver_params
=
416 MAX2(layout
->num_driver_params
, IR3_DP_VTXID_BASE
+ 1);
418 case nir_intrinsic_load_user_clip_plane
:
419 layout
->num_driver_params
=
420 MAX2(layout
->num_driver_params
, IR3_DP_UCP7_W
+ 1);
422 case nir_intrinsic_load_num_work_groups
:
423 layout
->num_driver_params
=
424 MAX2(layout
->num_driver_params
, IR3_DP_NUM_WORK_GROUPS_Z
+ 1);
426 case nir_intrinsic_load_local_group_size
:
427 layout
->num_driver_params
=
428 MAX2(layout
->num_driver_params
, IR3_DP_LOCAL_GROUP_SIZE_Z
+ 1);
439 ir3_setup_const_state(struct ir3_shader
*shader
, nir_shader
*nir
)
441 struct ir3_compiler
*compiler
= shader
->compiler
;
442 struct ir3_const_state
*const_state
= &shader
->const_state
;
444 memset(&const_state
->offsets
, ~0, sizeof(const_state
->offsets
));
446 ir3_nir_scan_driver_consts(nir
, const_state
);
448 if ((compiler
->gpu_id
< 500) &&
449 (shader
->stream_output
.num_outputs
> 0)) {
450 const_state
->num_driver_params
=
451 MAX2(const_state
->num_driver_params
, IR3_DP_VTXCNT_MAX
+ 1);
454 /* num_driver_params is scalar, align to vec4: */
455 const_state
->num_driver_params
= align(const_state
->num_driver_params
, 4);
457 debug_assert((shader
->ubo_state
.size
% 16) == 0);
458 unsigned constoff
= align(shader
->ubo_state
.size
/ 16, 8);
459 unsigned ptrsz
= ir3_pointer_size(compiler
);
461 if (const_state
->num_ubos
> 0) {
462 const_state
->offsets
.ubo
= constoff
;
463 constoff
+= align(nir
->info
.num_ubos
* ptrsz
, 4) / 4;
466 if (const_state
->ssbo_size
.count
> 0) {
467 unsigned cnt
= const_state
->ssbo_size
.count
;
468 const_state
->offsets
.ssbo_sizes
= constoff
;
469 constoff
+= align(cnt
, 4) / 4;
472 if (const_state
->image_dims
.count
> 0) {
473 unsigned cnt
= const_state
->image_dims
.count
;
474 const_state
->offsets
.image_dims
= constoff
;
475 constoff
+= align(cnt
, 4) / 4;
478 if (const_state
->num_driver_params
> 0)
479 const_state
->offsets
.driver_param
= constoff
;
480 constoff
+= const_state
->num_driver_params
/ 4;
482 if ((shader
->type
== MESA_SHADER_VERTEX
) &&
483 (compiler
->gpu_id
< 500) &&
484 shader
->stream_output
.num_outputs
> 0) {
485 const_state
->offsets
.tfbo
= constoff
;
486 constoff
+= align(IR3_MAX_SO_BUFFERS
* ptrsz
, 4) / 4;
489 switch (shader
->type
) {
490 case MESA_SHADER_VERTEX
:
491 const_state
->offsets
.primitive_param
= constoff
;
494 case MESA_SHADER_TESS_CTRL
:
495 case MESA_SHADER_TESS_EVAL
:
496 constoff
= align(constoff
- 1, 4) + 3;
497 const_state
->offsets
.primitive_param
= constoff
;
498 const_state
->offsets
.primitive_map
= constoff
+ 5;
499 constoff
+= 5 + DIV_ROUND_UP(nir
->num_inputs
, 4);
501 case MESA_SHADER_GEOMETRY
:
502 const_state
->offsets
.primitive_param
= constoff
;
503 const_state
->offsets
.primitive_map
= constoff
+ 1;
504 constoff
+= 1 + DIV_ROUND_UP(nir
->num_inputs
, 4);
510 const_state
->offsets
.immediate
= constoff
;