2 * Copyright © 2017 Red Hat
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Rob Clark <robclark@freedesktop.org>
28 #include "nir_builder.h"
30 #if defined(_WIN32) && !defined(snprintf)
31 #define snprintf _snprintf
35 * Remap atomic counters to SSBOs. Atomic counters get remapped to
36 * SSBO binding points [0..ssbo_offset) and the original SSBOs are
37 * remapped to [ssbo_offset..n) (mostly to align with what mesa/st
42 lower_instr(nir_intrinsic_instr
*instr
, unsigned ssbo_offset
, nir_builder
*b
)
47 b
->cursor
= nir_before_instr(&instr
->instr
);
49 switch (instr
->intrinsic
) {
50 case nir_intrinsic_ssbo_atomic_add
:
51 case nir_intrinsic_ssbo_atomic_imin
:
52 case nir_intrinsic_ssbo_atomic_umin
:
53 case nir_intrinsic_ssbo_atomic_imax
:
54 case nir_intrinsic_ssbo_atomic_umax
:
55 case nir_intrinsic_ssbo_atomic_and
:
56 case nir_intrinsic_ssbo_atomic_or
:
57 case nir_intrinsic_ssbo_atomic_xor
:
58 case nir_intrinsic_ssbo_atomic_exchange
:
59 case nir_intrinsic_ssbo_atomic_comp_swap
:
60 case nir_intrinsic_ssbo_atomic_fadd
:
61 case nir_intrinsic_ssbo_atomic_fmin
:
62 case nir_intrinsic_ssbo_atomic_fmax
:
63 case nir_intrinsic_ssbo_atomic_fcomp_swap
:
64 case nir_intrinsic_store_ssbo
:
65 case nir_intrinsic_load_ssbo
:
66 case nir_intrinsic_get_buffer_size
:
67 /* easy case, keep same opcode and just remap SSBO buffer index: */
68 op
= instr
->intrinsic
;
69 idx_src
= (op
== nir_intrinsic_store_ssbo
) ? 1 : 0;
70 nir_ssa_def
*old_idx
= nir_ssa_for_src(b
, instr
->src
[idx_src
], 1);
71 nir_ssa_def
*new_idx
= nir_iadd(b
, old_idx
, nir_imm_int(b
, ssbo_offset
));
72 nir_instr_rewrite_src(&instr
->instr
,
74 nir_src_for_ssa(new_idx
));
76 case nir_intrinsic_atomic_counter_inc
:
77 case nir_intrinsic_atomic_counter_add
:
78 case nir_intrinsic_atomic_counter_pre_dec
:
79 case nir_intrinsic_atomic_counter_post_dec
:
80 /* inc and dec get remapped to add: */
81 op
= nir_intrinsic_ssbo_atomic_add
;
83 case nir_intrinsic_atomic_counter_read
:
84 op
= nir_intrinsic_load_ssbo
;
86 case nir_intrinsic_atomic_counter_min
:
87 op
= nir_intrinsic_ssbo_atomic_umin
;
89 case nir_intrinsic_atomic_counter_max
:
90 op
= nir_intrinsic_ssbo_atomic_umax
;
92 case nir_intrinsic_atomic_counter_and
:
93 op
= nir_intrinsic_ssbo_atomic_and
;
95 case nir_intrinsic_atomic_counter_or
:
96 op
= nir_intrinsic_ssbo_atomic_or
;
98 case nir_intrinsic_atomic_counter_xor
:
99 op
= nir_intrinsic_ssbo_atomic_xor
;
101 case nir_intrinsic_atomic_counter_exchange
:
102 op
= nir_intrinsic_ssbo_atomic_exchange
;
104 case nir_intrinsic_atomic_counter_comp_swap
:
105 op
= nir_intrinsic_ssbo_atomic_comp_swap
;
111 nir_ssa_def
*buffer
= nir_imm_int(b
, nir_intrinsic_base(instr
));
112 nir_ssa_def
*temp
= NULL
;
113 nir_intrinsic_instr
*new_instr
=
114 nir_intrinsic_instr_create(ralloc_parent(instr
), op
);
116 /* a couple instructions need special handling since they don't map
117 * 1:1 with ssbo atomics
119 switch (instr
->intrinsic
) {
120 case nir_intrinsic_atomic_counter_inc
:
121 /* remapped to ssbo_atomic_add: { buffer_idx, offset, +1 } */
122 temp
= nir_imm_int(b
, +1);
123 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
124 nir_src_copy(&new_instr
->src
[1], &instr
->src
[0], new_instr
);
125 new_instr
->src
[2] = nir_src_for_ssa(temp
);
127 case nir_intrinsic_atomic_counter_pre_dec
:
128 case nir_intrinsic_atomic_counter_post_dec
:
129 /* remapped to ssbo_atomic_add: { buffer_idx, offset, -1 } */
130 /* NOTE semantic difference so we adjust the return value below */
131 temp
= nir_imm_int(b
, -1);
132 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
133 nir_src_copy(&new_instr
->src
[1], &instr
->src
[0], new_instr
);
134 new_instr
->src
[2] = nir_src_for_ssa(temp
);
136 case nir_intrinsic_atomic_counter_read
:
137 /* remapped to load_ssbo: { buffer_idx, offset } */
138 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
139 nir_src_copy(&new_instr
->src
[1], &instr
->src
[0], new_instr
);
142 /* remapped to ssbo_atomic_x: { buffer_idx, offset, data, (compare)? } */
143 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
144 nir_src_copy(&new_instr
->src
[1], &instr
->src
[0], new_instr
);
145 nir_src_copy(&new_instr
->src
[2], &instr
->src
[1], new_instr
);
146 if (op
== nir_intrinsic_ssbo_atomic_comp_swap
||
147 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
)
148 nir_src_copy(&new_instr
->src
[3], &instr
->src
[2], new_instr
);
152 if (new_instr
->intrinsic
== nir_intrinsic_load_ssbo
||
153 new_instr
->intrinsic
== nir_intrinsic_store_ssbo
)
154 nir_intrinsic_set_align(new_instr
, 4, 0);
156 nir_ssa_dest_init(&new_instr
->instr
, &new_instr
->dest
,
157 instr
->dest
.ssa
.num_components
,
158 instr
->dest
.ssa
.bit_size
, NULL
);
159 nir_instr_insert_before(&instr
->instr
, &new_instr
->instr
);
160 nir_instr_remove(&instr
->instr
);
162 if (instr
->intrinsic
== nir_intrinsic_atomic_counter_pre_dec
) {
163 b
->cursor
= nir_after_instr(&new_instr
->instr
);
164 nir_ssa_def
*result
= nir_iadd(b
, &new_instr
->dest
.ssa
, temp
);
165 nir_ssa_def_rewrite_uses(&instr
->dest
.ssa
, nir_src_for_ssa(result
));
167 nir_ssa_def_rewrite_uses(&instr
->dest
.ssa
, nir_src_for_ssa(&new_instr
->dest
.ssa
));
170 /* we could be replacing an intrinsic with fixed # of dest num_components
171 * with one that has variable number. So best to take this from the dest:
173 new_instr
->num_components
= instr
->dest
.ssa
.num_components
;
179 is_atomic_uint(const struct glsl_type
*type
)
181 if (glsl_get_base_type(type
) == GLSL_TYPE_ARRAY
)
182 return is_atomic_uint(glsl_get_array_element(type
));
183 return glsl_get_base_type(type
) == GLSL_TYPE_ATOMIC_UINT
;
187 nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
)
189 bool progress
= false;
191 nir_foreach_function(function
, shader
) {
192 if (function
->impl
) {
194 nir_builder_init(&builder
, function
->impl
);
195 nir_foreach_block(block
, function
->impl
) {
196 nir_foreach_instr_safe(instr
, block
) {
197 if (instr
->type
== nir_instr_type_intrinsic
)
198 progress
|= lower_instr(nir_instr_as_intrinsic(instr
),
199 ssbo_offset
, &builder
);
203 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
204 nir_metadata_dominance
);
209 /* replace atomic_uint uniforms with ssbo's: */
210 unsigned replaced
= 0;
211 nir_foreach_variable_safe(var
, &shader
->uniforms
) {
212 if (is_atomic_uint(var
->type
)) {
213 exec_node_remove(&var
->node
);
215 if (replaced
& (1 << var
->data
.binding
))
221 /* A length of 0 is used to denote unsized arrays */
222 const struct glsl_type
*type
= glsl_array_type(glsl_uint_type(), 0, 0);
224 snprintf(name
, sizeof(name
), "counter%d", var
->data
.binding
);
226 ssbo
= nir_variable_create(shader
, nir_var_mem_ssbo
, type
, name
);
227 ssbo
->data
.binding
= var
->data
.binding
;
229 struct glsl_struct_field field
= {
235 ssbo
->interface_type
=
236 glsl_interface_type(&field
, 1, GLSL_INTERFACE_PACKING_STD430
,
239 replaced
|= (1 << var
->data
.binding
);
243 /* Make sure that shader->info.num_ssbos still reflects the maximum SSBO
244 * index that can be used in the shader.
246 if (shader
->info
.num_ssbos
> 0) {
247 shader
->info
.num_ssbos
+= ssbo_offset
;
249 /* We can't use num_abos, because it only represents the number of
250 * active atomic counters, and currently unlike SSBO's they aren't
251 * compacted so num_abos actually isn't a bound on the index passed
252 * to nir_intrinsic_atomic_counter_*. e.g. if we have a single atomic
253 * counter declared like:
255 * layout(binding=1) atomic_uint counter0;
257 * then when we lower accesses to it the atomic_counter_* intrinsics
258 * will have 1 as the index but num_abos will still be 1.
260 shader
->info
.num_ssbos
= util_last_bit(replaced
);
262 shader
->info
.num_abos
= 0;