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"
31 #include "ir3_compiler.h"
32 #include "ir3_shader.h"
34 static const nir_shader_compiler_options options
= {
45 .lower_uadd_carry
= true,
46 .lower_mul_high
= true,
48 .vertex_id_zero_based
= true,
49 .lower_extract_byte
= true,
50 .lower_extract_word
= true,
51 .lower_all_io_to_elements
= true,
52 .lower_helper_invocation
= true,
53 .lower_bitfield_insert_to_shifts
= true,
54 .lower_bitfield_extract_to_shifts
= true,
56 .use_interpolated_input_intrinsics
= true,
59 /* we don't want to lower vertex_id to _zero_based on newer gpus: */
60 static const nir_shader_compiler_options options_a6xx
= {
71 .lower_uadd_carry
= true,
72 .lower_mul_high
= true,
74 .vertex_id_zero_based
= false,
75 .lower_extract_byte
= true,
76 .lower_extract_word
= true,
77 .lower_all_io_to_elements
= true,
78 .lower_helper_invocation
= true,
79 .lower_bitfield_insert_to_shifts
= true,
80 .lower_bitfield_extract_to_shifts
= true,
82 .use_interpolated_input_intrinsics
= true,
85 const nir_shader_compiler_options
*
86 ir3_get_compiler_options(struct ir3_compiler
*compiler
)
88 if (compiler
->gpu_id
>= 600)
93 /* for given shader key, are any steps handled in nir? */
95 ir3_key_lowers_nir(const struct ir3_shader_key
*key
)
97 return key
->fsaturate_s
| key
->fsaturate_t
| key
->fsaturate_r
|
98 key
->vsaturate_s
| key
->vsaturate_t
| key
->vsaturate_r
|
99 key
->ucp_enables
| key
->color_two_side
|
100 key
->fclamp_color
| key
->vclamp_color
;
103 #define OPT(nir, pass, ...) ({ \
104 bool this_progress = false; \
105 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
109 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
112 ir3_optimize_loop(nir_shader
*s
)
118 OPT_V(s
, nir_lower_vars_to_ssa
);
119 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
120 progress
|= OPT(s
, nir_opt_dead_write_vars
);
121 progress
|= OPT(s
, nir_lower_alu_to_scalar
);
122 progress
|= OPT(s
, nir_lower_phis_to_scalar
);
124 progress
|= OPT(s
, nir_copy_prop
);
125 progress
|= OPT(s
, nir_opt_dce
);
126 progress
|= OPT(s
, nir_opt_cse
);
129 gcm
= env_var_as_unsigned("GCM", 0);
131 progress
|= OPT(s
, nir_opt_gcm
, true);
133 progress
|= OPT(s
, nir_opt_gcm
, false);
134 progress
|= OPT(s
, nir_opt_peephole_select
, 16, true, true);
135 progress
|= OPT(s
, nir_opt_intrinsics
);
136 progress
|= OPT(s
, nir_opt_algebraic
);
137 progress
|= OPT(s
, nir_opt_constant_folding
);
138 progress
|= OPT(s
, nir_opt_dead_cf
);
139 if (OPT(s
, nir_opt_trivial_continues
)) {
141 /* If nir_opt_trivial_continues makes progress, then we need to clean
142 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
145 OPT(s
, nir_copy_prop
);
148 progress
|= OPT(s
, nir_opt_if
, false);
149 progress
|= OPT(s
, nir_opt_remove_phis
);
150 progress
|= OPT(s
, nir_opt_undef
);
156 ir3_optimize_nir(struct ir3_shader
*shader
, nir_shader
*s
,
157 const struct ir3_shader_key
*key
)
159 struct nir_lower_tex_options tex_options
= {
161 .lower_tg4_offsets
= true,
165 switch (shader
->type
) {
166 case MESA_SHADER_FRAGMENT
:
167 tex_options
.saturate_s
= key
->fsaturate_s
;
168 tex_options
.saturate_t
= key
->fsaturate_t
;
169 tex_options
.saturate_r
= key
->fsaturate_r
;
171 case MESA_SHADER_VERTEX
:
172 tex_options
.saturate_s
= key
->vsaturate_s
;
173 tex_options
.saturate_t
= key
->vsaturate_t
;
174 tex_options
.saturate_r
= key
->vsaturate_r
;
182 if (shader
->compiler
->gpu_id
>= 400) {
183 /* a4xx seems to have *no* sam.p */
184 tex_options
.lower_txp
= ~0; /* lower all txp */
186 /* a3xx just needs to avoid sam.p for 3d tex */
187 tex_options
.lower_txp
= (1 << GLSL_SAMPLER_DIM_3D
);
190 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
191 debug_printf("----------------------\n");
192 nir_print_shader(s
, stdout
);
193 debug_printf("----------------------\n");
196 OPT_V(s
, nir_lower_regs_to_ssa
);
197 OPT_V(s
, ir3_nir_lower_io_offsets
);
200 if (s
->info
.stage
== MESA_SHADER_VERTEX
) {
201 OPT_V(s
, nir_lower_clip_vs
, key
->ucp_enables
, false);
202 if (key
->vclamp_color
)
203 OPT_V(s
, nir_lower_clamp_color_outputs
);
204 } else if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
205 OPT_V(s
, nir_lower_clip_fs
, key
->ucp_enables
);
206 if (key
->fclamp_color
)
207 OPT_V(s
, nir_lower_clamp_color_outputs
);
209 if (key
->color_two_side
) {
210 OPT_V(s
, nir_lower_two_sided_color
);
213 /* only want to do this the first time (when key is null)
214 * and not again on any potential 2nd variant lowering pass:
216 OPT_V(s
, ir3_nir_apply_trig_workarounds
);
218 /* This wouldn't hurt to run multiple times, but there is
221 if (shader
->type
== MESA_SHADER_FRAGMENT
)
222 OPT_V(s
, nir_lower_fb_read
);
225 OPT_V(s
, nir_lower_tex
, &tex_options
);
226 OPT_V(s
, nir_lower_load_const_to_scalar
);
227 if (shader
->compiler
->gpu_id
< 500)
228 OPT_V(s
, ir3_nir_lower_tg4_to_tex
);
230 ir3_optimize_loop(s
);
232 /* do ubo load and idiv lowering after first opt loop to get a chance to
233 * propagate constants for divide by immed power-of-two and constant ubo
236 * NOTE that UBO analysis pass should only be done once, before variants
238 const bool ubo_progress
= !key
&& OPT(s
, ir3_nir_analyze_ubo_ranges
, shader
);
239 const bool idiv_progress
= OPT(s
, nir_lower_idiv
);
240 if (ubo_progress
|| idiv_progress
)
241 ir3_optimize_loop(s
);
243 OPT_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
245 OPT_V(s
, nir_move_load_const
);
247 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
248 debug_printf("----------------------\n");
249 nir_print_shader(s
, stdout
);
250 debug_printf("----------------------\n");
259 ir3_nir_scan_driver_consts(nir_shader
*shader
,
260 struct ir3_driver_const_layout
*layout
)
262 nir_foreach_function(function
, shader
) {
266 nir_foreach_block(block
, function
->impl
) {
267 nir_foreach_instr(instr
, block
) {
268 if (instr
->type
!= nir_instr_type_intrinsic
)
271 nir_intrinsic_instr
*intr
=
272 nir_instr_as_intrinsic(instr
);
275 switch (intr
->intrinsic
) {
276 case nir_intrinsic_get_buffer_size
:
277 idx
= nir_src_as_uint(intr
->src
[0]);
278 if (layout
->ssbo_size
.mask
& (1 << idx
))
280 layout
->ssbo_size
.mask
|= (1 << idx
);
281 layout
->ssbo_size
.off
[idx
] =
282 layout
->ssbo_size
.count
;
283 layout
->ssbo_size
.count
+= 1; /* one const per */
285 case nir_intrinsic_image_deref_atomic_add
:
286 case nir_intrinsic_image_deref_atomic_min
:
287 case nir_intrinsic_image_deref_atomic_max
:
288 case nir_intrinsic_image_deref_atomic_and
:
289 case nir_intrinsic_image_deref_atomic_or
:
290 case nir_intrinsic_image_deref_atomic_xor
:
291 case nir_intrinsic_image_deref_atomic_exchange
:
292 case nir_intrinsic_image_deref_atomic_comp_swap
:
293 case nir_intrinsic_image_deref_store
:
294 case nir_intrinsic_image_deref_size
:
295 idx
= nir_intrinsic_get_var(intr
, 0)->data
.driver_location
;
296 if (layout
->image_dims
.mask
& (1 << idx
))
298 layout
->image_dims
.mask
|= (1 << idx
);
299 layout
->image_dims
.off
[idx
] =
300 layout
->image_dims
.count
;
301 layout
->image_dims
.count
+= 3; /* three const per */