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
= {
47 .lower_uadd_carry
= true,
48 .lower_mul_high
= true,
50 .vertex_id_zero_based
= true,
51 .lower_extract_byte
= true,
52 .lower_extract_word
= true,
53 .lower_all_io_to_elements
= true,
54 .lower_helper_invocation
= true,
55 .lower_bitfield_insert_to_shifts
= true,
56 .lower_bitfield_extract_to_shifts
= true,
57 .use_interpolated_input_intrinsics
= true,
59 .lower_to_scalar
= true,
63 /* we don't want to lower vertex_id to _zero_based on newer gpus: */
64 static const nir_shader_compiler_options options_a6xx
= {
74 .lower_uadd_carry
= true,
75 .lower_mul_high
= true,
77 .vertex_id_zero_based
= false,
78 .lower_extract_byte
= true,
79 .lower_extract_word
= true,
80 .lower_all_io_to_elements
= true,
81 .lower_helper_invocation
= true,
82 .lower_bitfield_insert_to_shifts
= true,
83 .lower_bitfield_extract_to_shifts
= true,
84 .use_interpolated_input_intrinsics
= true,
87 .lower_to_scalar
= true,
91 const nir_shader_compiler_options
*
92 ir3_get_compiler_options(struct ir3_compiler
*compiler
)
94 if (compiler
->gpu_id
>= 600)
99 /* for given shader key, are any steps handled in nir? */
101 ir3_key_lowers_nir(const struct ir3_shader_key
*key
)
103 return key
->fsaturate_s
| key
->fsaturate_t
| key
->fsaturate_r
|
104 key
->vsaturate_s
| key
->vsaturate_t
| key
->vsaturate_r
|
105 key
->ucp_enables
| key
->color_two_side
|
106 key
->fclamp_color
| key
->vclamp_color
|
110 #define OPT(nir, pass, ...) ({ \
111 bool this_progress = false; \
112 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
116 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
119 ir3_optimize_loop(nir_shader
*s
)
122 unsigned lower_flrp
=
123 (s
->options
->lower_flrp16
? 16 : 0) |
124 (s
->options
->lower_flrp32
? 32 : 0) |
125 (s
->options
->lower_flrp64
? 64 : 0);
130 OPT_V(s
, nir_lower_vars_to_ssa
);
131 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
132 progress
|= OPT(s
, nir_opt_dead_write_vars
);
133 progress
|= OPT(s
, nir_lower_alu_to_scalar
, NULL
, NULL
);
134 progress
|= OPT(s
, nir_lower_phis_to_scalar
);
136 progress
|= OPT(s
, nir_copy_prop
);
137 progress
|= OPT(s
, nir_opt_dce
);
138 progress
|= OPT(s
, nir_opt_cse
);
141 gcm
= env_var_as_unsigned("GCM", 0);
143 progress
|= OPT(s
, nir_opt_gcm
, true);
145 progress
|= OPT(s
, nir_opt_gcm
, false);
146 progress
|= OPT(s
, nir_opt_peephole_select
, 16, true, true);
147 progress
|= OPT(s
, nir_opt_intrinsics
);
148 progress
|= OPT(s
, nir_opt_algebraic
);
149 progress
|= OPT(s
, nir_opt_constant_folding
);
151 if (lower_flrp
!= 0) {
152 if (OPT(s
, nir_lower_flrp
,
154 false /* always_precise */,
155 s
->options
->lower_ffma
)) {
156 OPT(s
, nir_opt_constant_folding
);
160 /* Nothing should rematerialize any flrps, so we only
161 * need to do this lowering once.
166 progress
|= OPT(s
, nir_opt_dead_cf
);
167 if (OPT(s
, nir_opt_trivial_continues
)) {
169 /* If nir_opt_trivial_continues makes progress, then we need to clean
170 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
173 OPT(s
, nir_copy_prop
);
176 progress
|= OPT(s
, nir_opt_if
, false);
177 progress
|= OPT(s
, nir_opt_remove_phis
);
178 progress
|= OPT(s
, nir_opt_undef
);
184 ir3_optimize_nir(struct ir3_shader
*shader
, nir_shader
*s
,
185 const struct ir3_shader_key
*key
)
187 struct nir_lower_tex_options tex_options
= {
189 .lower_tg4_offsets
= true,
192 if (key
&& key
->has_gs
) {
193 switch (shader
->type
) {
194 case MESA_SHADER_VERTEX
:
195 NIR_PASS_V(s
, ir3_nir_lower_vs_to_explicit_io
, shader
);
197 case MESA_SHADER_GEOMETRY
:
198 NIR_PASS_V(s
, ir3_nir_lower_gs
, shader
);
206 switch (shader
->type
) {
207 case MESA_SHADER_FRAGMENT
:
208 tex_options
.saturate_s
= key
->fsaturate_s
;
209 tex_options
.saturate_t
= key
->fsaturate_t
;
210 tex_options
.saturate_r
= key
->fsaturate_r
;
212 case MESA_SHADER_VERTEX
:
213 tex_options
.saturate_s
= key
->vsaturate_s
;
214 tex_options
.saturate_t
= key
->vsaturate_t
;
215 tex_options
.saturate_r
= key
->vsaturate_r
;
223 if (shader
->compiler
->gpu_id
>= 400) {
224 /* a4xx seems to have *no* sam.p */
225 tex_options
.lower_txp
= ~0; /* lower all txp */
227 /* a3xx just needs to avoid sam.p for 3d tex */
228 tex_options
.lower_txp
= (1 << GLSL_SAMPLER_DIM_3D
);
231 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
232 debug_printf("----------------------\n");
233 nir_print_shader(s
, stdout
);
234 debug_printf("----------------------\n");
237 OPT_V(s
, nir_lower_regs_to_ssa
);
238 OPT_V(s
, ir3_nir_lower_io_offsets
);
241 if (s
->info
.stage
== MESA_SHADER_VERTEX
) {
242 OPT_V(s
, nir_lower_clip_vs
, key
->ucp_enables
, false, false, NULL
);
243 if (key
->vclamp_color
)
244 OPT_V(s
, nir_lower_clamp_color_outputs
);
245 } else if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
246 OPT_V(s
, nir_lower_clip_fs
, key
->ucp_enables
, false);
247 if (key
->fclamp_color
)
248 OPT_V(s
, nir_lower_clamp_color_outputs
);
250 if (key
->color_two_side
) {
251 OPT_V(s
, nir_lower_two_sided_color
);
254 /* only want to do this the first time (when key is null)
255 * and not again on any potential 2nd variant lowering pass:
257 OPT_V(s
, ir3_nir_apply_trig_workarounds
);
259 /* This wouldn't hurt to run multiple times, but there is
262 if (shader
->type
== MESA_SHADER_FRAGMENT
)
263 OPT_V(s
, nir_lower_fb_read
);
266 OPT_V(s
, nir_lower_tex
, &tex_options
);
267 OPT_V(s
, nir_lower_load_const_to_scalar
);
268 if (shader
->compiler
->gpu_id
< 500)
269 OPT_V(s
, ir3_nir_lower_tg4_to_tex
);
271 ir3_optimize_loop(s
);
273 /* do ubo load and idiv lowering after first opt loop to get a chance to
274 * propagate constants for divide by immed power-of-two and constant ubo
277 * NOTE that UBO analysis pass should only be done once, before variants
279 const bool ubo_progress
= !key
&& OPT(s
, ir3_nir_analyze_ubo_ranges
, shader
);
280 const bool idiv_progress
= OPT(s
, nir_lower_idiv
, nir_lower_idiv_fast
);
281 if (ubo_progress
|| idiv_progress
)
282 ir3_optimize_loop(s
);
284 /* Do late algebraic optimization to turn add(a, neg(b)) back into
285 * subs, then the mandatory cleanup after algebraic. Note that it may
286 * produce fnegs, and if so then we need to keep running to squash
289 bool more_late_algebraic
= true;
290 while (more_late_algebraic
) {
291 more_late_algebraic
= OPT(s
, nir_opt_algebraic_late
);
292 OPT_V(s
, nir_opt_constant_folding
);
293 OPT_V(s
, nir_copy_prop
);
294 OPT_V(s
, nir_opt_dce
);
295 OPT_V(s
, nir_opt_cse
);
298 OPT_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
300 OPT_V(s
, nir_opt_sink
, nir_move_const_undef
);
302 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
303 debug_printf("----------------------\n");
304 nir_print_shader(s
, stdout
);
305 debug_printf("----------------------\n");
310 /* The first time thru, when not creating variant, do the one-time
311 * const_state layout setup. This should be done after ubo range
315 ir3_setup_const_state(shader
, s
);
320 ir3_nir_scan_driver_consts(nir_shader
*shader
,
321 struct ir3_const_state
*layout
)
323 nir_foreach_function(function
, shader
) {
327 nir_foreach_block(block
, function
->impl
) {
328 nir_foreach_instr(instr
, block
) {
329 if (instr
->type
!= nir_instr_type_intrinsic
)
332 nir_intrinsic_instr
*intr
=
333 nir_instr_as_intrinsic(instr
);
336 switch (intr
->intrinsic
) {
337 case nir_intrinsic_get_buffer_size
:
338 idx
= nir_src_as_uint(intr
->src
[0]);
339 if (layout
->ssbo_size
.mask
& (1 << idx
))
341 layout
->ssbo_size
.mask
|= (1 << idx
);
342 layout
->ssbo_size
.off
[idx
] =
343 layout
->ssbo_size
.count
;
344 layout
->ssbo_size
.count
+= 1; /* one const per */
346 case nir_intrinsic_image_deref_atomic_add
:
347 case nir_intrinsic_image_deref_atomic_imin
:
348 case nir_intrinsic_image_deref_atomic_umin
:
349 case nir_intrinsic_image_deref_atomic_imax
:
350 case nir_intrinsic_image_deref_atomic_umax
:
351 case nir_intrinsic_image_deref_atomic_and
:
352 case nir_intrinsic_image_deref_atomic_or
:
353 case nir_intrinsic_image_deref_atomic_xor
:
354 case nir_intrinsic_image_deref_atomic_exchange
:
355 case nir_intrinsic_image_deref_atomic_comp_swap
:
356 case nir_intrinsic_image_deref_store
:
357 case nir_intrinsic_image_deref_size
:
358 idx
= nir_intrinsic_get_var(intr
, 0)->data
.driver_location
;
359 if (layout
->image_dims
.mask
& (1 << idx
))
361 layout
->image_dims
.mask
|= (1 << idx
);
362 layout
->image_dims
.off
[idx
] =
363 layout
->image_dims
.count
;
364 layout
->image_dims
.count
+= 3; /* three const per */
366 case nir_intrinsic_load_ubo
:
367 if (nir_src_is_const(intr
->src
[0])) {
368 layout
->num_ubos
= MAX2(layout
->num_ubos
,
369 nir_src_as_uint(intr
->src
[0]) + 1);
371 layout
->num_ubos
= shader
->info
.num_ubos
;
374 case nir_intrinsic_load_base_vertex
:
375 case nir_intrinsic_load_first_vertex
:
376 layout
->num_driver_params
=
377 MAX2(layout
->num_driver_params
, IR3_DP_VTXID_BASE
+ 1);
379 case nir_intrinsic_load_user_clip_plane
:
380 layout
->num_driver_params
=
381 MAX2(layout
->num_driver_params
, IR3_DP_UCP7_W
+ 1);
383 case nir_intrinsic_load_num_work_groups
:
384 layout
->num_driver_params
=
385 MAX2(layout
->num_driver_params
, IR3_DP_NUM_WORK_GROUPS_Z
+ 1);
387 case nir_intrinsic_load_local_group_size
:
388 layout
->num_driver_params
=
389 MAX2(layout
->num_driver_params
, IR3_DP_LOCAL_GROUP_SIZE_Z
+ 1);
400 ir3_setup_const_state(struct ir3_shader
*shader
, nir_shader
*nir
)
402 struct ir3_compiler
*compiler
= shader
->compiler
;
403 struct ir3_const_state
*const_state
= &shader
->const_state
;
405 memset(&const_state
->offsets
, ~0, sizeof(const_state
->offsets
));
407 ir3_nir_scan_driver_consts(nir
, const_state
);
409 if ((compiler
->gpu_id
< 500) &&
410 (shader
->stream_output
.num_outputs
> 0)) {
411 const_state
->num_driver_params
=
412 MAX2(const_state
->num_driver_params
, IR3_DP_VTXCNT_MAX
+ 1);
415 /* num_driver_params is scalar, align to vec4: */
416 const_state
->num_driver_params
= align(const_state
->num_driver_params
, 4);
418 debug_assert((shader
->ubo_state
.size
% 16) == 0);
419 unsigned constoff
= align(shader
->ubo_state
.size
/ 16, 8);
420 unsigned ptrsz
= ir3_pointer_size(compiler
);
422 if (const_state
->num_ubos
> 0) {
423 const_state
->offsets
.ubo
= constoff
;
424 constoff
+= align(nir
->info
.num_ubos
* ptrsz
, 4) / 4;
427 if (const_state
->ssbo_size
.count
> 0) {
428 unsigned cnt
= const_state
->ssbo_size
.count
;
429 const_state
->offsets
.ssbo_sizes
= constoff
;
430 constoff
+= align(cnt
, 4) / 4;
433 if (const_state
->image_dims
.count
> 0) {
434 unsigned cnt
= const_state
->image_dims
.count
;
435 const_state
->offsets
.image_dims
= constoff
;
436 constoff
+= align(cnt
, 4) / 4;
439 if (const_state
->num_driver_params
> 0)
440 const_state
->offsets
.driver_param
= constoff
;
441 constoff
+= const_state
->num_driver_params
/ 4;
443 if ((shader
->type
== MESA_SHADER_VERTEX
) &&
444 (compiler
->gpu_id
< 500) &&
445 shader
->stream_output
.num_outputs
> 0) {
446 const_state
->offsets
.tfbo
= constoff
;
447 constoff
+= align(IR3_MAX_SO_BUFFERS
* ptrsz
, 4) / 4;
450 switch (shader
->type
) {
451 case MESA_SHADER_VERTEX
:
452 const_state
->offsets
.primitive_param
= constoff
;
455 case MESA_SHADER_GEOMETRY
:
456 const_state
->offsets
.primitive_param
= constoff
;
457 const_state
->offsets
.primitive_map
= constoff
+ 1;
458 constoff
+= 1 + DIV_ROUND_UP(nir
->num_inputs
, 4);
464 const_state
->offsets
.immediate
= constoff
;