LLVM 19.0.0git
SIModeRegisterDefaults.cpp
Go to the documentation of this file.
1//===-- SIModeRegisterDefaults.cpp ------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9#include "SIModeRegisterDefaults.h"
10#include "GCNSubtarget.h"
11
12using namespace llvm;
13
14SIModeRegisterDefaults::SIModeRegisterDefaults(const Function &F,
15 const GCNSubtarget &ST) {
16 *this = getDefaultForCallingConv(F.getCallingConv());
17
18 if (ST.hasIEEEMode()) {
19 StringRef IEEEAttr = F.getFnAttribute("amdgpu-ieee").getValueAsString();
20 if (!IEEEAttr.empty())
21 IEEE = IEEEAttr == "true";
22 }
23
24 if (ST.hasDX10ClampMode()) {
25 StringRef DX10ClampAttr =
26 F.getFnAttribute("amdgpu-dx10-clamp").getValueAsString();
27 if (!DX10ClampAttr.empty())
28 DX10Clamp = DX10ClampAttr == "true";
29 }
30
31 StringRef DenormF32Attr =
32 F.getFnAttribute("denormal-fp-math-f32").getValueAsString();
33 if (!DenormF32Attr.empty())
34 FP32Denormals = parseDenormalFPAttribute(DenormF32Attr);
35
36 StringRef DenormAttr =
37 F.getFnAttribute("denormal-fp-math").getValueAsString();
38 if (!DenormAttr.empty()) {
39 DenormalMode DenormMode = parseDenormalFPAttribute(DenormAttr);
40 if (DenormF32Attr.empty())
41 FP32Denormals = DenormMode;
42 FP64FP16Denormals = DenormMode;
43 }
44}
45
46using namespace AMDGPU;
47
48/// Combine f32 and f64 rounding modes into a combined rounding mode value.
49static constexpr uint32_t getModeRegisterRoundMode(uint32_t HWFP32Val,
50 uint32_t HWFP64Val) {
51 return HWFP32Val << F32FltRoundOffset | HWFP64Val << F64FltRoundOffset;
52}
53
54static constexpr uint64_t encodeFltRoundsTable(uint32_t FltRoundsVal,
55 uint32_t HWF32Val,
56 uint32_t HWF64Val) {
57 uint32_t ModeVal = getModeRegisterRoundMode(HWF32Val, HWF64Val);
58 if (FltRoundsVal > TowardNegative)
59 FltRoundsVal -= ExtendedFltRoundOffset;
60
61 uint32_t BitIndex = ModeVal << 2;
62 return static_cast<uint64_t>(FltRoundsVal) << BitIndex;
63}
64
65// Encode FLT_ROUNDS value where the two rounding modes are the same and use a
66// standard value
67static constexpr uint64_t
68encodeFltRoundsTableSame(AMDGPUFltRounds FltRoundsMode, uint32_t HWVal) {
69 return encodeFltRoundsTable(FltRoundsMode, HWVal, HWVal);
70}
71
72// Convert mode register encoded rounding mode to AMDGPUFltRounds
73static constexpr AMDGPUFltRounds
74decodeIndexFltRoundConversionTable(uint32_t HWMode) {
75 uint32_t TableRead = (FltRoundConversionTable >> (HWMode << 2)) & 0xf;
76 if (TableRead > TowardNegative)
77 TableRead += ExtendedFltRoundOffset;
78 return static_cast<AMDGPUFltRounds>(TableRead);
79}
80
81static constexpr uint32_t HWTowardZero = FP_ROUND_ROUND_TO_ZERO;
82static constexpr uint32_t HWNearestTiesToEven = FP_ROUND_ROUND_TO_NEAREST;
83static constexpr uint32_t HWTowardPositive = FP_ROUND_ROUND_TO_INF;
84static constexpr uint32_t HWTowardNegative = FP_ROUND_ROUND_TO_NEGINF;
85
86const uint64_t AMDGPU::FltRoundConversionTable =
87 encodeFltRoundsTableSame(TowardZeroF32_TowardZeroF64, HWTowardZero) |
88 encodeFltRoundsTableSame(NearestTiesToEvenF32_NearestTiesToEvenF64,
89 HWNearestTiesToEven) |
90 encodeFltRoundsTableSame(TowardPositiveF32_TowardPositiveF64,
91 HWTowardPositive) |
92 encodeFltRoundsTableSame(TowardNegativeF32_TowardNegativeF64,
93 HWTowardNegative) |
94
95 encodeFltRoundsTable(TowardZeroF32_NearestTiesToEvenF64, HWTowardZero,
96 HWNearestTiesToEven) |
97 encodeFltRoundsTable(TowardZeroF32_TowardPositiveF64, HWTowardZero,
98 HWTowardPositive) |
99 encodeFltRoundsTable(TowardZeroF32_TowardNegativeF64, HWTowardZero,
100 HWTowardNegative) |
101
102 encodeFltRoundsTable(NearestTiesToEvenF32_TowardZeroF64,
103 HWNearestTiesToEven, HWTowardZero) |
104 encodeFltRoundsTable(NearestTiesToEvenF32_TowardPositiveF64,
105 HWNearestTiesToEven, HWTowardPositive) |
106 encodeFltRoundsTable(NearestTiesToEvenF32_TowardNegativeF64,
107 HWNearestTiesToEven, HWTowardNegative) |
108
109 encodeFltRoundsTable(TowardPositiveF32_TowardZeroF64, HWTowardPositive,
110 HWTowardZero) |
111 encodeFltRoundsTable(TowardPositiveF32_NearestTiesToEvenF64,
112 HWTowardPositive, HWNearestTiesToEven) |
113 encodeFltRoundsTable(TowardPositiveF32_TowardNegativeF64, HWTowardPositive,
114 HWTowardNegative) |
115
116 encodeFltRoundsTable(TowardNegativeF32_TowardZeroF64, HWTowardNegative,
117 HWTowardZero) |
118 encodeFltRoundsTable(TowardNegativeF32_NearestTiesToEvenF64,
119 HWTowardNegative, HWNearestTiesToEven) |
120 encodeFltRoundsTable(TowardNegativeF32_TowardPositiveF64, HWTowardNegative,
121 HWTowardPositive);
122
123// Verify evaluation of FltRoundConversionTable
124
125// If both modes are the same, should return the standard values.
126static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
127 HWTowardZero, HWTowardZero)) == AMDGPUFltRounds::TowardZero);
128static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
129 HWNearestTiesToEven, HWNearestTiesToEven)) ==
130 AMDGPUFltRounds::NearestTiesToEven);
131static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
132 HWTowardPositive, HWTowardPositive)) ==
133 AMDGPUFltRounds::TowardPositive);
134static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
135 HWTowardNegative, HWTowardNegative)) ==
136 AMDGPUFltRounds::TowardNegative);
137
138static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
139 HWTowardZero, HWNearestTiesToEven)) ==
140 TowardZeroF32_NearestTiesToEvenF64);
141static_assert(decodeIndexFltRoundConversionTable(
142 getModeRegisterRoundMode(HWTowardZero, HWTowardPositive)) ==
143 TowardZeroF32_TowardPositiveF64);
144static_assert(decodeIndexFltRoundConversionTable(
145 getModeRegisterRoundMode(HWTowardZero, HWTowardNegative)) ==
146 TowardZeroF32_TowardNegativeF64);
147
148static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
149 HWNearestTiesToEven, HWTowardZero)) ==
150 NearestTiesToEvenF32_TowardZeroF64);
151static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
152 HWNearestTiesToEven, HWTowardPositive)) ==
153 NearestTiesToEvenF32_TowardPositiveF64);
154static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
155 HWNearestTiesToEven, HWTowardNegative)) ==
156 NearestTiesToEvenF32_TowardNegativeF64);
157
158static_assert(decodeIndexFltRoundConversionTable(
159 getModeRegisterRoundMode(HWTowardPositive, HWTowardZero)) ==
160 TowardPositiveF32_TowardZeroF64);
161static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
162 HWTowardPositive, HWNearestTiesToEven)) ==
163 TowardPositiveF32_NearestTiesToEvenF64);
164static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
165 HWTowardPositive, HWTowardNegative)) ==
166 TowardPositiveF32_TowardNegativeF64);
167
168static_assert(decodeIndexFltRoundConversionTable(
169 getModeRegisterRoundMode(HWTowardNegative, HWTowardZero)) ==
170 TowardNegativeF32_TowardZeroF64);
171static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
172 HWTowardNegative, HWNearestTiesToEven)) ==
173 TowardNegativeF32_NearestTiesToEvenF64);
174static_assert(decodeIndexFltRoundConversionTable(getModeRegisterRoundMode(
175 HWTowardNegative, HWTowardPositive)) ==
176 TowardNegativeF32_TowardPositiveF64);
177
178// Decode FLT_ROUNDS into the hardware value where the two rounding modes are
179// the same and use a standard value
180static constexpr uint64_t encodeFltRoundsToHWTableSame(uint32_t HWVal,
181 uint32_t FltRoundsVal) {
182 if (FltRoundsVal > TowardNegative)
183 FltRoundsVal -= ExtendedFltRoundOffset;
184
185 return static_cast<uint64_t>(getModeRegisterRoundMode(HWVal, HWVal))
186 << (FltRoundsVal << 2);
187}
188
189/// Decode FLT_ROUNDS into the hardware value where the two rounding modes
190/// different and use an extended value.
191static constexpr uint64_t encodeFltRoundsToHWTable(uint32_t HWF32Val,
192 uint32_t HWF64Val,
193 uint32_t FltRoundsVal) {
194 if (FltRoundsVal > TowardNegative)
195 FltRoundsVal -= ExtendedFltRoundOffset;
196 return static_cast<uint64_t>(getModeRegisterRoundMode(HWF32Val, HWF64Val))
197 << (FltRoundsVal << 2);
198}
199
200const uint64_t AMDGPU::FltRoundToHWConversionTable =
201 encodeFltRoundsToHWTableSame(HWTowardZero, TowardZeroF32_TowardZeroF64) |
202 encodeFltRoundsToHWTableSame(HWNearestTiesToEven,
203 NearestTiesToEvenF32_NearestTiesToEvenF64) |
204 encodeFltRoundsToHWTableSame(HWTowardPositive,
205 TowardPositiveF32_TowardPositiveF64) |
206 encodeFltRoundsToHWTableSame(HWTowardNegative,
207 TowardNegativeF32_TowardNegativeF64) |
208
209 encodeFltRoundsToHWTable(HWTowardZero, HWNearestTiesToEven,
210 TowardZeroF32_NearestTiesToEvenF64) |
211 encodeFltRoundsToHWTable(HWTowardZero, HWTowardPositive,
212 TowardZeroF32_TowardPositiveF64) |
213 encodeFltRoundsToHWTable(HWTowardZero, HWTowardNegative,
214 TowardZeroF32_TowardNegativeF64) |
215
216 encodeFltRoundsToHWTable(HWNearestTiesToEven, HWTowardZero,
217 NearestTiesToEvenF32_TowardZeroF64) |
218 encodeFltRoundsToHWTable(HWNearestTiesToEven, HWTowardPositive,
219 NearestTiesToEvenF32_TowardPositiveF64) |
220 encodeFltRoundsToHWTable(HWNearestTiesToEven, HWTowardNegative,
221 NearestTiesToEvenF32_TowardNegativeF64) |
222
223 encodeFltRoundsToHWTable(HWTowardPositive, HWTowardZero,
224 TowardPositiveF32_TowardZeroF64) |
225 encodeFltRoundsToHWTable(HWTowardPositive, HWNearestTiesToEven,
226 TowardPositiveF32_NearestTiesToEvenF64) |
227 encodeFltRoundsToHWTable(HWTowardPositive, HWTowardNegative,
228 TowardPositiveF32_TowardNegativeF64) |
229
230 encodeFltRoundsToHWTable(HWTowardNegative, HWTowardZero,
231 TowardNegativeF32_TowardZeroF64) |
232 encodeFltRoundsToHWTable(HWTowardNegative, HWNearestTiesToEven,
233 TowardNegativeF32_NearestTiesToEvenF64) |
234 encodeFltRoundsToHWTable(HWTowardNegative, HWTowardPositive,
235 TowardNegativeF32_TowardPositiveF64);
236
237/// Read the hardware rounding mode equivalent of a AMDGPUFltRounds value.
238static constexpr uint32_t
239decodeFltRoundToHWConversionTable(uint64_t FltRoundToHWConversionTable,
240 uint32_t FltRounds) {
241 uint32_t IndexVal = FltRounds;
242 if (IndexVal > TowardNegative)
243 IndexVal -= ExtendedFltRoundOffset;
244 return (FltRoundToHWConversionTable >> (IndexVal << 2)) & 0xf;
245}
246
247uint32_t AMDGPU::decodeFltRoundToHWConversionTable(uint32_t FltRounds) {
248 return ::decodeFltRoundToHWConversionTable(FltRoundToHWConversionTable,
249 FltRounds);
250}
251
252static constexpr uint32_t decodeFltRoundToHW(uint32_t FltRounds) {
253 return ::decodeFltRoundToHWConversionTable(FltRoundToHWConversionTable,
254 FltRounds);
255}
256
257// Verify evaluation of FltRoundToHWConversionTable
258
259static_assert(decodeFltRoundToHW(AMDGPUFltRounds::TowardZero) ==
260 getModeRegisterRoundMode(HWTowardZero, HWTowardZero));
261static_assert(decodeFltRoundToHW(AMDGPUFltRounds::NearestTiesToEven) ==
262 getModeRegisterRoundMode(HWNearestTiesToEven,
263 HWNearestTiesToEven));
264static_assert(decodeFltRoundToHW(AMDGPUFltRounds::TowardPositive) ==
265 getModeRegisterRoundMode(HWTowardPositive, HWTowardPositive));
266static_assert(decodeFltRoundToHW(AMDGPUFltRounds::TowardNegative) ==
267 getModeRegisterRoundMode(HWTowardNegative, HWTowardNegative));
268
269static_assert(decodeFltRoundToHW(NearestTiesToEvenF32_TowardPositiveF64) ==
270 getModeRegisterRoundMode(HWNearestTiesToEven, HWTowardPositive));
271static_assert(decodeFltRoundToHW(NearestTiesToEvenF32_TowardNegativeF64) ==
272 getModeRegisterRoundMode(HWNearestTiesToEven, HWTowardNegative));
273static_assert(decodeFltRoundToHW(NearestTiesToEvenF32_TowardZeroF64) ==
274 getModeRegisterRoundMode(HWNearestTiesToEven, HWTowardZero));
275
276static_assert(decodeFltRoundToHW(TowardPositiveF32_NearestTiesToEvenF64) ==
277 getModeRegisterRoundMode(HWTowardPositive, HWNearestTiesToEven));
278static_assert(decodeFltRoundToHW(TowardPositiveF32_TowardNegativeF64) ==
279 getModeRegisterRoundMode(HWTowardPositive, HWTowardNegative));
280static_assert(decodeFltRoundToHW(TowardPositiveF32_TowardZeroF64) ==
281 getModeRegisterRoundMode(HWTowardPositive, HWTowardZero));
282
283static_assert(decodeFltRoundToHW(TowardNegativeF32_NearestTiesToEvenF64) ==
284 getModeRegisterRoundMode(HWTowardNegative, HWNearestTiesToEven));
285static_assert(decodeFltRoundToHW(TowardNegativeF32_TowardPositiveF64) ==
286 getModeRegisterRoundMode(HWTowardNegative, HWTowardPositive));
287static_assert(decodeFltRoundToHW(TowardNegativeF32_TowardZeroF64) ==
288 getModeRegisterRoundMode(HWTowardNegative, HWTowardZero));
289
290static_assert(decodeFltRoundToHW(TowardZeroF32_NearestTiesToEvenF64) ==
291 getModeRegisterRoundMode(HWTowardZero, HWNearestTiesToEven));
292static_assert(decodeFltRoundToHW(TowardZeroF32_TowardPositiveF64) ==
293 getModeRegisterRoundMode(HWTowardZero, HWTowardPositive));
294static_assert(decodeFltRoundToHW(TowardZeroF32_TowardNegativeF64) ==
295 getModeRegisterRoundMode(HWTowardZero, HWTowardNegative));
GCNSubtarget.h
AMD GCN specific subclass of TargetSubtarget.
F
#define F(x, y, z)
Definition: MD5.cpp:55
FP_ROUND_ROUND_TO_INF
#define FP_ROUND_ROUND_TO_INF
Definition: SIDefines.h:1159
FP_ROUND_ROUND_TO_NEAREST
#define FP_ROUND_ROUND_TO_NEAREST
Definition: SIDefines.h:1158
FP_ROUND_ROUND_TO_ZERO
#define FP_ROUND_ROUND_TO_ZERO
Definition: SIDefines.h:1161
FP_ROUND_ROUND_TO_NEGINF
#define FP_ROUND_ROUND_TO_NEGINF
Definition: SIDefines.h:1160
encodeFltRoundsToHWTable
static constexpr uint64_t encodeFltRoundsToHWTable(uint32_t HWF32Val, uint32_t HWF64Val, uint32_t FltRoundsVal)
Decode FLT_ROUNDS into the hardware value where the two rounding modes different and use an extended ...
Definition: SIModeRegisterDefaults.cpp:191
decodeFltRoundToHW
static constexpr uint32_t decodeFltRoundToHW(uint32_t FltRounds)
Definition: SIModeRegisterDefaults.cpp:252
decodeIndexFltRoundConversionTable
static constexpr AMDGPUFltRounds decodeIndexFltRoundConversionTable(uint32_t HWMode)
Definition: SIModeRegisterDefaults.cpp:74
HWTowardNegative
static constexpr uint32_t HWTowardNegative
Definition: SIModeRegisterDefaults.cpp:84
encodeFltRoundsTableSame
static constexpr uint64_t encodeFltRoundsTableSame(AMDGPUFltRounds FltRoundsMode, uint32_t HWVal)
Definition: SIModeRegisterDefaults.cpp:68
HWTowardPositive
static constexpr uint32_t HWTowardPositive
Definition: SIModeRegisterDefaults.cpp:83
HWTowardZero
static constexpr uint32_t HWTowardZero
Definition: SIModeRegisterDefaults.cpp:81
encodeFltRoundsToHWTableSame
static constexpr uint64_t encodeFltRoundsToHWTableSame(uint32_t HWVal, uint32_t FltRoundsVal)
Definition: SIModeRegisterDefaults.cpp:180
encodeFltRoundsTable
static constexpr uint64_t encodeFltRoundsTable(uint32_t FltRoundsVal, uint32_t HWF32Val, uint32_t HWF64Val)
Definition: SIModeRegisterDefaults.cpp:54
decodeFltRoundToHWConversionTable
static constexpr uint32_t decodeFltRoundToHWConversionTable(uint64_t FltRoundToHWConversionTable, uint32_t FltRounds)
Read the hardware rounding mode equivalent of a AMDGPUFltRounds value.
Definition: SIModeRegisterDefaults.cpp:239
HWNearestTiesToEven
static constexpr uint32_t HWNearestTiesToEven
Definition: SIModeRegisterDefaults.cpp:82
getModeRegisterRoundMode
static constexpr uint32_t getModeRegisterRoundMode(uint32_t HWFP32Val, uint32_t HWFP64Val)
Combine f32 and f64 rounding modes into a combined rounding mode value.
Definition: SIModeRegisterDefaults.cpp:49
llvm::StringRef
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
llvm::StringRef::empty
constexpr bool empty() const
empty - Check if the string is empty.
Definition: StringRef.h:134
uint32_t
uint64_t
llvm::AMDGPU::FltRoundToHWConversionTable
const uint64_t FltRoundToHWConversionTable
Definition: SIModeRegisterDefaults.cpp:200
llvm::AMDGPU::F64FltRoundOffset
static constexpr uint32_t F64FltRoundOffset
Offset in mode register of f64/f16 rounding mode.
Definition: SIModeRegisterDefaults.h:141
llvm::AMDGPU::ExtendedFltRoundOffset
static constexpr uint32_t ExtendedFltRoundOffset
Offset of nonstandard values for llvm.get.rounding results from the largest supported mode.
Definition: SIModeRegisterDefaults.h:135
llvm::AMDGPU::AMDGPUFltRounds
AMDGPUFltRounds
Return values used for llvm.get.rounding.
Definition: SIModeRegisterDefaults.h:96
llvm::AMDGPU::decodeFltRoundToHWConversionTable
uint32_t decodeFltRoundToHWConversionTable(uint32_t FltRounds)
Read the hardware rounding mode equivalent of a AMDGPUFltRounds value.
Definition: SIModeRegisterDefaults.cpp:247
llvm::AMDGPU::F32FltRoundOffset
static constexpr uint32_t F32FltRoundOffset
Offset in mode register of f32 rounding mode.
Definition: SIModeRegisterDefaults.h:138
llvm::AMDGPU::FltRoundConversionTable
const uint64_t FltRoundConversionTable
Definition: SIModeRegisterDefaults.cpp:86
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
llvm::RoundingMode::TowardNegative
@ TowardNegative
roundTowardNegative.
llvm::parseDenormalFPAttribute
DenormalMode parseDenormalFPAttribute(StringRef Str)
Returns the denormal mode to use for inputs and outputs.
Definition: FloatingPointMode.h:218
llvm::DenormalMode
Represent subnormal handling kind for floating point instruction inputs and outputs.
Definition: FloatingPointMode.h:70
llvm::SIModeRegisterDefaults::DX10Clamp
bool DX10Clamp
Used by the vector ALU to force DX10-style treatment of NaNs: when set, clamp NaN to zero; otherwise,...
Definition: SIModeRegisterDefaults.h:29
llvm::SIModeRegisterDefaults::FP64FP16Denormals
DenormalMode FP64FP16Denormals
If this is set, neither input or output denormals are flushed for both f64 and f16/v2f16 instructions...
Definition: SIModeRegisterDefaults.h:37
llvm::SIModeRegisterDefaults::IEEE
bool IEEE
Floating point opcodes that support exception flag gathering quiet and propagate signaling NaN inputs...
Definition: SIModeRegisterDefaults.h:25
llvm::SIModeRegisterDefaults::getDefaultForCallingConv
static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC)
Definition: SIModeRegisterDefaults.h:47
llvm::SIModeRegisterDefaults::FP32Denormals
DenormalMode FP32Denormals
If this is set, neither input or output denormals are flushed for most f32 instructions.
Definition: SIModeRegisterDefaults.h:33
Generated on Wed May 8 2024 00:29:49 for LLVM by doxygen 1.9.6