2 * Copyright (C) 2019 Connor Abbott <cwabbott0@gmail.com>
3 * Copyright (C) 2019 Lyude Paul <thatslyude@gmail.com>
4 * Copyright (C) 2019 Ryan Houdek <Sonicadvance1@gmail.com>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 struct bifrost_header
{
34 // If true, convert any infinite result of any floating-point operation to
35 // the biggest representable number.
36 unsigned suppress_inf
: 1;
37 // Convert any NaN results to 0.
38 unsigned suppress_nan
: 1;
40 // true if the execution mask of the next clause is the same as the mask of
41 // the current clause.
42 unsigned back_to_back
: 1;
43 unsigned no_end_of_shader
: 1;
45 // Set to true for fragment shaders, to implement this bit of spec text
46 // from section 7.1.5 of the GLSL ES spec:
48 // "Stores to image and buffer variables performed by helper invocations
49 // have no effect on the underlying image or buffer memory."
51 // Helper invocations are threads (invocations) corresponding to pixels in
52 // a quad that aren't actually part of the triangle, but are included to
53 // make derivatives work correctly. They're usually turned on, but they
54 // need to be masked off for GLSL-level stores. This bit seems to be the
55 // only bit that's actually different between fragment shaders and other
56 // shaders, so this is probably what it's doing.
57 unsigned elide_writes
: 1;
58 // If backToBack is off:
59 // - true for conditional branches and fallthrough
60 // - false for unconditional branches
61 // The blob seems to always set it to true if back-to-back is on.
62 unsigned branch_cond
: 1;
63 // This bit is set when the next clause writes to the data register of some
65 unsigned datareg_writebarrier
: 1;
67 unsigned scoreboard_deps
: 8;
68 unsigned scoreboard_index
: 3;
69 unsigned clause_type
: 4;
70 unsigned unk3
: 1; // part of clauseType?
71 unsigned next_clause_type
: 4;
72 unsigned unk4
: 1; // part of nextClauseType?
75 struct bifrost_fma_inst
{
80 struct bifrost_add_inst
{
88 BIFROST_SAT_SIGNED
= 0x2,
92 enum bifrost_roundmode
{
99 struct bifrost_fma_add
{
102 unsigned src1_abs
: 1;
103 unsigned src0_neg
: 1;
104 unsigned src1_neg
: 1;
106 unsigned src0_abs
: 1;
107 enum bifrost_outmod outmod
: 2;
108 enum bifrost_roundmode roundmode
: 2;
112 enum bifrost_csel_cond
{
123 struct bifrost_csel4
{
128 enum bifrost_csel_cond cond
: 3;
132 struct bifrost_shift_fma
{
136 unsigned half
: 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
137 unsigned unk
: 1; /* always set? */
138 unsigned invert_1
: 1; /* Inverts sources to combining op */
139 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
140 unsigned invert_2
: 1;
144 struct bifrost_shift_add
{
150 unsigned invert_1
: 1;
151 unsigned invert_2
: 1;
156 enum bifrost_ldst_type
{
157 BIFROST_LDST_F16
= 0,
158 BIFROST_LDST_F32
= 1,
159 BIFROST_LDST_I32
= 2,
163 struct bifrost_ld_var_addr
{
166 unsigned location
: 5;
167 enum bifrost_ldst_type type
: 2;
171 struct bifrost_ld_attr
{
174 unsigned location
: 5;
175 unsigned channels
: 2; /* MALI_POSITIVE */
176 enum bifrost_ldst_type type
: 2;
180 struct bifrost_tex_ctrl
{
181 unsigned sampler_index
: 4; // also used to signal indirects
182 unsigned tex_index
: 7;
183 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
184 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
186 bool texel_offset
: 1; // *Offset()
189 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
190 bool compute_lod
: 1; // 0 for *Lod()
191 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
192 bool calc_gradients
: 1; // 0 for *Grad()
194 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
198 struct bifrost_dual_tex_ctrl
{
199 unsigned sampler_index0
: 2;
201 unsigned tex_index0
: 2;
202 unsigned sampler_index1
: 2;
203 unsigned tex_index1
: 2;
207 enum branch_bit_size
{
211 // For the above combinations of bitsize and location, an extra bit is
212 // encoded via comparing the sources. The only possible source of ambiguity
213 // would be if the sources were the same, but then the branch condition
214 // would be always true or always false anyways, so we can ignore it. But
215 // this no longer works when comparing the y component to the x component,
216 // since it's valid to compare the y component of a source against its own
217 // x component. Instead, the extra bit is encoded via an extra bitsize.
220 BR_SIZE_32_AND_16X
= 5,
221 BR_SIZE_32_AND_16Y
= 6,
222 // Used for comparisons with zero and always-true, see below. I think this
223 // only works for integer comparisons.
227 enum bifrost_reg_write_unit
{
228 REG_WRITE_NONE
= 0, // don't write
229 REG_WRITE_TWO
, // write using reg2
230 REG_WRITE_THREE
, // write using reg3
233 struct bifrost_regs
{
234 unsigned uniform_const
: 8;
242 enum bifrost_branch_cond
{
247 // Equal vs. not-equal determined by src0/src1 comparison
249 // floating-point comparisons
250 // Becomes UNE when you flip the arguments
252 // TODO what happens when you flip the arguments?
257 enum bifrost_branch_code
{