radv: align the LDS size in calculate_tess_lds_size()
[mesa.git] / src / amd / compiler / aco_ir.cpp
1 /*
2 * Copyright © 2020 Valve 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 "aco_ir.h"
25 #include "vulkan/radv_shader.h"
26 #include "c11/threads.h"
27 #include "util/debug.h"
28
29 namespace aco {
30
31 uint64_t debug_flags = 0;
32
33 static const struct debug_control aco_debug_options[] = {
34 {"validateir", DEBUG_VALIDATE},
35 {"validatera", DEBUG_VALIDATE_RA},
36 {"perfwarn", DEBUG_PERFWARN},
37 {NULL, 0}
38 };
39
40 static once_flag init_once_flag = ONCE_FLAG_INIT;
41
42 static void init_once()
43 {
44 debug_flags = parse_debug_string(getenv("ACO_DEBUG"), aco_debug_options);
45
46 #ifndef NDEBUG
47 /* enable some flags by default on debug builds */
48 debug_flags |= aco::DEBUG_VALIDATE;
49 #endif
50 }
51
52 void init()
53 {
54 call_once(&init_once_flag, init_once);
55 }
56
57 void init_program(Program *program, Stage stage, struct radv_shader_info *info,
58 enum chip_class chip_class, enum radeon_family family,
59 ac_shader_config *config)
60 {
61 program->stage = stage;
62 program->config = config;
63 program->info = info;
64 program->chip_class = chip_class;
65 if (family == CHIP_UNKNOWN) {
66 switch (chip_class) {
67 case GFX6:
68 program->family = CHIP_TAHITI;
69 break;
70 case GFX7:
71 program->family = CHIP_BONAIRE;
72 break;
73 case GFX8:
74 program->family = CHIP_POLARIS10;
75 break;
76 case GFX9:
77 program->family = CHIP_VEGA10;
78 break;
79 case GFX10:
80 program->family = CHIP_NAVI10;
81 break;
82 default:
83 program->family = CHIP_UNKNOWN;
84 break;
85 }
86 } else {
87 program->family = family;
88 }
89 program->wave_size = info->wave_size;
90 program->lane_mask = program->wave_size == 32 ? s1 : s2;
91
92 program->lds_alloc_granule = chip_class >= GFX7 ? 512 : 256;
93 program->lds_limit = chip_class >= GFX7 ? 65536 : 32768;
94 /* apparently gfx702 also has 16-bank LDS but I can't find a family for that */
95 program->has_16bank_lds = family == CHIP_KABINI || family == CHIP_STONEY;
96
97 program->vgpr_limit = 256;
98 program->vgpr_alloc_granule = 3;
99
100 if (chip_class >= GFX10) {
101 program->physical_sgprs = 2560; /* doesn't matter as long as it's at least 128 * 20 */
102 program->sgpr_alloc_granule = 127;
103 program->sgpr_limit = 106;
104 program->vgpr_alloc_granule = program->wave_size == 32 ? 7 : 3;
105 } else if (program->chip_class >= GFX8) {
106 program->physical_sgprs = 800;
107 program->sgpr_alloc_granule = 15;
108 if (family == CHIP_TONGA || family == CHIP_ICELAND)
109 program->sgpr_limit = 94; /* workaround hardware bug */
110 else
111 program->sgpr_limit = 102;
112 } else {
113 program->physical_sgprs = 512;
114 program->sgpr_alloc_granule = 7;
115 program->sgpr_limit = 104;
116 }
117
118 program->next_fp_mode.preserve_signed_zero_inf_nan32 = false;
119 program->next_fp_mode.preserve_signed_zero_inf_nan16_64 = false;
120 program->next_fp_mode.must_flush_denorms32 = false;
121 program->next_fp_mode.must_flush_denorms16_64 = false;
122 program->next_fp_mode.care_about_round32 = false;
123 program->next_fp_mode.care_about_round16_64 = false;
124 program->next_fp_mode.denorm16_64 = fp_denorm_keep;
125 program->next_fp_mode.denorm32 = 0;
126 program->next_fp_mode.round16_64 = fp_round_ne;
127 program->next_fp_mode.round32 = fp_round_ne;
128 }
129
130 bool can_use_SDWA(chip_class chip, const aco_ptr<Instruction>& instr)
131 {
132 if (!instr->isVALU())
133 return false;
134
135 if (chip < GFX8 || instr->isDPP())
136 return false;
137
138 if (instr->isSDWA())
139 return true;
140
141 if (instr->isVOP3()) {
142 VOP3A_instruction *vop3 = static_cast<VOP3A_instruction*>(instr.get());
143 if (instr->format == Format::VOP3)
144 return false;
145 if (vop3->clamp && instr->format == asVOP3(Format::VOPC) && chip != GFX8)
146 return false;
147 if (vop3->omod && chip < GFX9)
148 return false;
149
150 //TODO: return true if we know we will use vcc
151 if (instr->definitions.size() >= 2)
152 return false;
153
154 for (unsigned i = 1; i < instr->operands.size(); i++) {
155 if (instr->operands[i].isLiteral())
156 return false;
157 if (chip < GFX9 && !instr->operands[i].isOfType(RegType::vgpr))
158 return false;
159 }
160 }
161
162 if (!instr->operands.empty()) {
163 if (instr->operands[0].isLiteral())
164 return false;
165 if (chip < GFX9 && !instr->operands[0].isOfType(RegType::vgpr))
166 return false;
167 }
168
169 bool is_mac = instr->opcode == aco_opcode::v_mac_f32 ||
170 instr->opcode == aco_opcode::v_mac_f16 ||
171 instr->opcode == aco_opcode::v_fmac_f32 ||
172 instr->opcode == aco_opcode::v_fmac_f16;
173
174 if (chip != GFX8 && is_mac)
175 return false;
176
177 //TODO: return true if we know we will use vcc
178 if ((unsigned)instr->format & (unsigned)Format::VOPC)
179 return false;
180 if (instr->operands.size() >= 3 && !is_mac)
181 return false;
182
183 return instr->opcode != aco_opcode::v_madmk_f32 &&
184 instr->opcode != aco_opcode::v_madak_f32 &&
185 instr->opcode != aco_opcode::v_madmk_f16 &&
186 instr->opcode != aco_opcode::v_madak_f16 &&
187 instr->opcode != aco_opcode::v_readfirstlane_b32 &&
188 instr->opcode != aco_opcode::v_clrexcp &&
189 instr->opcode != aco_opcode::v_swap_b32;
190 }
191
192 /* updates "instr" and returns the old instruction (or NULL if no update was needed) */
193 aco_ptr<Instruction> convert_to_SDWA(chip_class chip, aco_ptr<Instruction>& instr)
194 {
195 if (instr->isSDWA())
196 return NULL;
197
198 aco_ptr<Instruction> tmp = std::move(instr);
199 Format format = (Format)(((uint16_t)tmp->format & ~(uint16_t)Format::VOP3) | (uint16_t)Format::SDWA);
200 instr.reset(create_instruction<SDWA_instruction>(tmp->opcode, format, tmp->operands.size(), tmp->definitions.size()));
201 std::copy(tmp->operands.cbegin(), tmp->operands.cend(), instr->operands.begin());
202 std::copy(tmp->definitions.cbegin(), tmp->definitions.cend(), instr->definitions.begin());
203
204 SDWA_instruction *sdwa = static_cast<SDWA_instruction*>(instr.get());
205
206 if (tmp->isVOP3()) {
207 VOP3A_instruction *vop3 = static_cast<VOP3A_instruction*>(tmp.get());
208 memcpy(sdwa->neg, vop3->neg, sizeof(sdwa->neg));
209 memcpy(sdwa->abs, vop3->abs, sizeof(sdwa->abs));
210 sdwa->omod = vop3->omod;
211 sdwa->clamp = vop3->clamp;
212 }
213
214 for (unsigned i = 0; i < instr->operands.size(); i++) {
215 switch (instr->operands[i].bytes()) {
216 case 1:
217 sdwa->sel[i] = sdwa_ubyte;
218 break;
219 case 2:
220 sdwa->sel[i] = sdwa_uword;
221 break;
222 case 4:
223 sdwa->sel[i] = sdwa_udword;
224 break;
225 }
226 }
227 switch (instr->definitions[0].bytes()) {
228 case 1:
229 sdwa->dst_sel = sdwa_ubyte;
230 sdwa->dst_preserve = true;
231 break;
232 case 2:
233 sdwa->dst_sel = sdwa_uword;
234 sdwa->dst_preserve = true;
235 break;
236 case 4:
237 sdwa->dst_sel = sdwa_udword;
238 break;
239 }
240
241 if (instr->definitions[0].getTemp().type() == RegType::sgpr && chip == GFX8)
242 instr->definitions[0].setFixed(vcc);
243 if (instr->definitions.size() >= 2)
244 instr->definitions[1].setFixed(vcc);
245 if (instr->operands.size() >= 3)
246 instr->operands[2].setFixed(vcc);
247
248 return tmp;
249 }
250
251 bool can_use_opsel(chip_class chip, aco_opcode op, int idx, bool high)
252 {
253 /* opsel is only GFX9+ */
254 if ((high || idx == -1) && chip < GFX9)
255 return false;
256
257 switch (op) {
258 case aco_opcode::v_div_fixup_f16:
259 case aco_opcode::v_fma_f16:
260 case aco_opcode::v_mad_f16:
261 case aco_opcode::v_mad_u16:
262 case aco_opcode::v_mad_i16:
263 case aco_opcode::v_med3_f16:
264 case aco_opcode::v_med3_i16:
265 case aco_opcode::v_med3_u16:
266 case aco_opcode::v_min3_f16:
267 case aco_opcode::v_min3_i16:
268 case aco_opcode::v_min3_u16:
269 case aco_opcode::v_max3_f16:
270 case aco_opcode::v_max3_i16:
271 case aco_opcode::v_max3_u16:
272 case aco_opcode::v_max_u16_e64:
273 case aco_opcode::v_max_i16_e64:
274 case aco_opcode::v_min_u16_e64:
275 case aco_opcode::v_min_i16_e64:
276 case aco_opcode::v_add_i16:
277 case aco_opcode::v_sub_i16:
278 case aco_opcode::v_add_u16_e64:
279 case aco_opcode::v_sub_u16_e64:
280 case aco_opcode::v_cvt_pknorm_i16_f16:
281 case aco_opcode::v_cvt_pknorm_u16_f16:
282 case aco_opcode::v_lshlrev_b16_e64:
283 case aco_opcode::v_lshrrev_b16_e64:
284 case aco_opcode::v_ashrrev_i16_e64:
285 case aco_opcode::v_mul_lo_u16_e64:
286 return true;
287 case aco_opcode::v_pack_b32_f16:
288 return idx != -1;
289 case aco_opcode::v_mad_u32_u16:
290 case aco_opcode::v_mad_i32_i16:
291 return idx >= 0 && idx < 2;
292 default:
293 return false;
294 }
295 }
296
297 }