1 /*
2 * Copyright © 2018 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include "brw_nir.h"
25 #include "compiler/nir/nir_builder.h"
26
27 static nir_op
get_conversion_op(nir_alu_type src_type,unsigned src_bit_size,nir_alu_type dst_type,unsigned dst_bit_size,nir_rounding_mode rounding_mode)28 get_conversion_op(nir_alu_type src_type,
29 unsigned src_bit_size,
30 nir_alu_type dst_type,
31 unsigned dst_bit_size,
32 nir_rounding_mode rounding_mode)
33 {
34 nir_alu_type src_full_type = (nir_alu_type) (src_type | src_bit_size);
35 nir_alu_type dst_full_type = (nir_alu_type) (dst_type | dst_bit_size);
36
37 return nir_type_conversion_op(src_full_type, dst_full_type, rounding_mode);
38 }
39
40 static nir_rounding_mode
get_opcode_rounding_mode(nir_op op)41 get_opcode_rounding_mode(nir_op op)
42 {
43 switch (op) {
44 case nir_op_f2f16_rtz:
45 return nir_rounding_mode_rtz;
46 case nir_op_f2f16_rtne:
47 return nir_rounding_mode_rtne;
48 default:
49 return nir_rounding_mode_undef;
50 }
51 }
52
53 static void
split_conversion(nir_builder * b,nir_alu_instr * alu,nir_op op1,nir_op op2)54 split_conversion(nir_builder *b, nir_alu_instr *alu, nir_op op1, nir_op op2)
55 {
56 b->cursor = nir_before_instr(&alu->instr);
57 assert(alu->dest.write_mask == 1);
58 nir_ssa_def *src = nir_ssa_for_alu_src(b, alu, 0);
59 nir_ssa_def *tmp = nir_build_alu(b, op1, src, NULL, NULL, NULL);
60 nir_ssa_def *res = nir_build_alu(b, op2, tmp, NULL, NULL, NULL);
61 nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(res));
62 nir_instr_remove(&alu->instr);
63 }
64
65 static bool
lower_instr(nir_builder * b,nir_alu_instr * alu)66 lower_instr(nir_builder *b, nir_alu_instr *alu)
67 {
68 unsigned src_bit_size = nir_src_bit_size(alu->src[0].src);
69 nir_alu_type src_type = nir_op_infos[alu->op].input_types[0];
70 nir_alu_type src_full_type = (nir_alu_type) (src_type | src_bit_size);
71
72 unsigned dst_bit_size = nir_dest_bit_size(alu->dest.dest);
73 nir_alu_type dst_full_type = nir_op_infos[alu->op].output_type;
74 nir_alu_type dst_type = nir_alu_type_get_base_type(dst_full_type);
75
76 /* BDW PRM, vol02, Command Reference Instructions, mov - MOVE:
77 *
78 * "There is no direct conversion from HF to DF or DF to HF.
79 * Use two instructions and F (Float) as an intermediate type.
80 *
81 * There is no direct conversion from HF to Q/UQ or Q/UQ to HF.
82 * Use two instructions and F (Float) or a word integer type
83 * or a DWord integer type as an intermediate type."
84 *
85 * It is important that the intermediate conversion happens through a
86 * 32-bit float type so we don't lose range when we convert from
87 * a 64-bit integer.
88 */
89 if ((src_full_type == nir_type_float16 && dst_bit_size == 64) ||
90 (src_bit_size == 64 && dst_full_type == nir_type_float16)) {
91 nir_op op1 = get_conversion_op(src_type, src_bit_size,
92 nir_type_float, 32,
93 nir_rounding_mode_undef);
94 nir_op op2 = get_conversion_op(nir_type_float, 32,
95 dst_type, dst_bit_size,
96 get_opcode_rounding_mode(alu->op));
97 split_conversion(b, alu, op1, op2);
98 return true;
99 }
100
101 /* SKL PRM, vol 02a, Command Reference: Instructions, Move:
102 *
103 * "There is no direct conversion from B/UB to DF or DF to B/UB. Use
104 * two instructions and a word or DWord intermediate type."
105 *
106 * "There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB.
107 * Use two instructions and a word or DWord intermediate integer
108 * type."
109 *
110 * It is important that we use a 32-bit integer matching the sign of the
111 * destination as the intermediate type so we avoid any chance of rtne
112 * rounding happening before the conversion to integer (which is expected
113 * to round towards zero) in double to byte conversions.
114 */
115 if ((src_bit_size == 8 && dst_bit_size == 64) ||
116 (src_bit_size == 64 && dst_bit_size == 8)) {
117 nir_op op1 = get_conversion_op(src_type, src_bit_size, dst_type, 32,
118 nir_rounding_mode_undef);
119 nir_op op2 = get_conversion_op(dst_type, 32, dst_type, dst_bit_size,
120 nir_rounding_mode_undef);
121 split_conversion(b, alu, op1, op2);
122 return true;
123 }
124
125 return false;
126 }
127
128 static bool
lower_impl(nir_function_impl * impl)129 lower_impl(nir_function_impl *impl)
130 {
131 nir_builder b;
132 nir_builder_init(&b, impl);
133 bool progress = false;
134
135 nir_foreach_block(block, impl) {
136 nir_foreach_instr_safe(instr, block) {
137 if (instr->type != nir_instr_type_alu)
138 continue;
139
140 nir_alu_instr *alu = nir_instr_as_alu(instr);
141 assert(alu->dest.dest.is_ssa);
142
143 if (!nir_op_infos[alu->op].is_conversion)
144 continue;
145
146 progress = lower_instr(&b, alu) || progress;
147 }
148 }
149
150 if (progress) {
151 nir_metadata_preserve(impl, nir_metadata_block_index |
152 nir_metadata_dominance);
153 } else {
154 nir_metadata_preserve(impl, nir_metadata_all);
155 }
156
157 return progress;
158 }
159
160 bool
brw_nir_lower_conversions(nir_shader * shader)161 brw_nir_lower_conversions(nir_shader *shader)
162 {
163 bool progress = false;
164
165 nir_foreach_function(function, shader) {
166 if (function->impl)
167 progress |= lower_impl(function->impl);
168 }
169
170 return progress;
171 }
172