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_store_ssbo
:
61 case nir_intrinsic_load_ssbo
:
62 /* easy case, keep same opcode and just remap SSBO buffer index: */
63 op
= instr
->intrinsic
;
64 idx_src
= (op
== nir_intrinsic_store_ssbo
) ? 1 : 0;
65 nir_ssa_def
*old_idx
= nir_ssa_for_src(b
, instr
->src
[idx_src
], 1);
66 nir_ssa_def
*new_idx
= nir_iadd(b
, old_idx
, nir_imm_int(b
, ssbo_offset
));
67 nir_instr_rewrite_src(&instr
->instr
,
69 nir_src_for_ssa(new_idx
));
71 case nir_intrinsic_atomic_counter_inc
:
72 case nir_intrinsic_atomic_counter_add
:
73 case nir_intrinsic_atomic_counter_dec
:
74 /* inc and dec get remapped to add: */
75 op
= nir_intrinsic_ssbo_atomic_add
;
77 case nir_intrinsic_atomic_counter_read
:
78 op
= nir_intrinsic_load_ssbo
;
80 case nir_intrinsic_atomic_counter_min
:
81 op
= nir_intrinsic_ssbo_atomic_umin
;
83 case nir_intrinsic_atomic_counter_max
:
84 op
= nir_intrinsic_ssbo_atomic_umax
;
86 case nir_intrinsic_atomic_counter_and
:
87 op
= nir_intrinsic_ssbo_atomic_and
;
89 case nir_intrinsic_atomic_counter_or
:
90 op
= nir_intrinsic_ssbo_atomic_or
;
92 case nir_intrinsic_atomic_counter_xor
:
93 op
= nir_intrinsic_ssbo_atomic_xor
;
95 case nir_intrinsic_atomic_counter_exchange
:
96 op
= nir_intrinsic_ssbo_atomic_exchange
;
98 case nir_intrinsic_atomic_counter_comp_swap
:
99 op
= nir_intrinsic_ssbo_atomic_comp_swap
;
105 nir_ssa_def
*buffer
= nir_imm_int(b
, nir_intrinsic_base(instr
));
106 nir_ssa_def
*temp
= NULL
;
107 nir_intrinsic_instr
*new_instr
=
108 nir_intrinsic_instr_create(ralloc_parent(instr
), op
);
110 /* a couple instructions need special handling since they don't map
111 * 1:1 with ssbo atomics
113 switch (instr
->intrinsic
) {
114 case nir_intrinsic_atomic_counter_inc
:
115 /* remapped to ssbo_atomic_add: { buffer_idx, offset, +1 } */
116 temp
= nir_imm_int(b
, +1);
117 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
118 new_instr
->src
[1] = instr
->src
[0];
119 new_instr
->src
[2] = nir_src_for_ssa(temp
);
121 case nir_intrinsic_atomic_counter_dec
:
122 /* remapped to ssbo_atomic_add: { buffer_idx, offset, -1 } */
123 /* NOTE semantic difference so we adjust the return value below */
124 temp
= nir_imm_int(b
, -1);
125 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
126 new_instr
->src
[1] = instr
->src
[0];
127 new_instr
->src
[2] = nir_src_for_ssa(temp
);
129 case nir_intrinsic_atomic_counter_read
:
130 /* remapped to load_ssbo: { buffer_idx, offset } */
131 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
132 new_instr
->src
[1] = instr
->src
[0];
135 /* remapped to ssbo_atomic_x: { buffer_idx, offset, data, (compare)? } */
136 new_instr
->src
[0] = nir_src_for_ssa(buffer
);
137 new_instr
->src
[1] = instr
->src
[0];
138 new_instr
->src
[2] = instr
->src
[1];
139 if (op
== nir_intrinsic_ssbo_atomic_comp_swap
)
140 new_instr
->src
[3] = instr
->src
[2];
144 nir_ssa_dest_init(&new_instr
->instr
, &new_instr
->dest
,
145 instr
->dest
.ssa
.num_components
,
146 instr
->dest
.ssa
.bit_size
, NULL
);
147 nir_instr_insert_before(&instr
->instr
, &new_instr
->instr
);
148 nir_instr_remove(&instr
->instr
);
150 if (instr
->intrinsic
== nir_intrinsic_atomic_counter_dec
) {
151 b
->cursor
= nir_after_instr(&new_instr
->instr
);
152 nir_ssa_def
*result
= nir_iadd(b
, &new_instr
->dest
.ssa
, temp
);
153 nir_ssa_def_rewrite_uses(&instr
->dest
.ssa
, nir_src_for_ssa(result
));
155 nir_ssa_def_rewrite_uses(&instr
->dest
.ssa
, nir_src_for_ssa(&new_instr
->dest
.ssa
));
158 /* we could be replacing an intrinsic with fixed # of dest num_components
159 * with one that has variable number. So best to take this from the dest:
161 new_instr
->num_components
= instr
->dest
.ssa
.num_components
;
167 nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
)
169 bool progress
= false;
171 nir_foreach_function(function
, shader
) {
172 if (function
->impl
) {
174 nir_builder_init(&builder
, function
->impl
);
175 nir_foreach_block(block
, function
->impl
) {
176 nir_foreach_instr_safe(instr
, block
) {
177 if (instr
->type
== nir_instr_type_intrinsic
)
178 progress
|= lower_instr(nir_instr_as_intrinsic(instr
),
179 ssbo_offset
, &builder
);
183 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
184 nir_metadata_dominance
);
189 /* replace atomic_uint uniforms with ssbo's: */
190 unsigned replaced
= 0;
191 nir_foreach_variable_safe(var
, &shader
->uniforms
) {
192 if (glsl_get_base_type(var
->type
) == GLSL_TYPE_ATOMIC_UINT
) {
193 exec_node_remove(&var
->node
);
195 if (replaced
& (1 << var
->data
.binding
))
201 /* A length of 0 is used to denote unsized arrays */
202 const struct glsl_type
*type
= glsl_array_type(glsl_uint_type(), 0);
204 snprintf(name
, sizeof(name
), "counter%d", var
->data
.binding
);
206 ssbo
= nir_variable_create(shader
, nir_var_shader_storage
,
208 ssbo
->data
.binding
= var
->data
.binding
;
210 struct glsl_struct_field field
= {
216 ssbo
->interface_type
=
217 glsl_interface_type(&field
, 1, GLSL_INTERFACE_PACKING_STD430
,
220 replaced
|= (1 << var
->data
.binding
);