1 # FPR-to-GPR and GPR-to-FPR
5 High-performance CPU/GPU software needs to often convert between integers
6 and floating-point, therefore fast conversion/data-movement instructions
7 are needed. Also given that initialisation of floats tends to take up
8 considerable space (even to just load 0.0) the inclusion of float immediate
9 is up for consideration (BF16 as immediates)
11 Libre-SOC will be compliant with the
12 **Scalar Floating-Point Subset** (SFFS) i.e. is not implementing VMX/VSX,
13 and with its focus on modern 3D GPU hybrid workloads represents an
14 important new potential use-case for OpenPOWER.
15 With VMX/VSX not available in the SFFS Compliancy Level, the
16 existing non-VSX conversion/data-movement instructions require load/store
17 instructions (slow and expensive) to transfer data between the FPRs and
18 the GPRs. Also, because SimpleV needs efficient scalar instructions in
19 order to generate efficient vector instructions, adding new instructions
20 for data-transfer/conversion between FPRs and GPRs seems necessary.
22 In addition, the vast majority of GPR <-> FPR data-transfers are as part
23 of a FP <-> Integer conversion sequence, therefore reducing the number
24 of instructions required to the minimum seems necessary.
26 Therefore, we are proposing adding:
28 * FPR load-immediate using `BF16` as the constant
29 * FPR <-> GPR data-transfer instructions that just copy bits without conversion
30 * FPR <-> GPR combined data-transfer/conversion instructions that do
31 Integer <-> FP conversions
33 If we're adding new Integer <-> FP conversion instructions, we may
34 as well take this opportunity to modernise the instructions and make them
35 well suited for common/important conversion sequences:
37 * standard Integer -> FP IEEE754 conversion (used by most languages and CPUs)
38 * standard OpenPower FP -> Integer conversion (saturation with NaN
39 converted to minimum valid integer)
40 * Rust FP -> Integer conversion (saturation with NaN converted to 0)
41 * JavaScript FP -> Integer conversion (modular with Inf/NaN converted to 0)
43 # A bit more research into integer - fp conversion
45 here is a paragraph which explains that there are different semantics
46 for conversion, i don't know what the paragraph should say, but it needs
47 to be here, to give some background. it also acts as a lead-in to the
48 sub-sections, introducing them and explaining why they are here, as
49 justifications and background research as to why the ISA should support
50 the feature being proposed.
52 *nothing* can be left to chance or guesswork.
54 ## standard Integer -> FP conversion
56 This conversion is outlined in the IEEE754 specification. It is used
57 by nearly all programming languages and CPUs. In the case of OpenPOWER,
58 the rounding mode is read from FPSCR
60 # standard OpenPower FP -> Integer conversion
62 TODO, explain this further, make this a complete sentence:
63 "saturation with NaN converted to minimum valid integer"
65 - Matches x86's conversion semantics
66 - Has instructions for both:
67 * rounding mode read from FPSCR
68 * rounding mode is always truncate
70 ## Rust FP -> Integer conversion
72 TODO, explain this further, the following is not a complete sentence,
73 "saturation with NaN converted to 0"
75 Semantics required by all of:
76 (what does this mean, what is "required"?
77 what semantics are being referred to? the sentence needs completing:
78 "For Rust integer conversion, the semantics required are shown by the
79 following, all of which are supported in XYZ" something like that)
81 * Rust's FP -> Integer conversion using the
82 [`as` operator](https://doc.rust-lang.org/reference/expressions/operator-expr.html#semantics)
84 [FP -> Integer conversion](https://docs.oracle.com/javase/specs/jls/se16/html/jls-5.html#jls-5.1.3)
86 [`llvm.fptosi.sat`](https://llvm.org/docs/LangRef.html#llvm-fptosi-sat-intrinsic) and
87 [`llvm.fptoui.sat`](https://llvm.org/docs/LangRef.html#llvm-fptoui-sat-intrinsic) intrinsics
88 * SPIR-V's OpenCL dialect's
89 [`OpConvertFToU`](https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpConvertFToU) and
90 [`OpConvertFToS`](https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpConvertFToS)
91 instructions when decorated with
92 [the `SaturatedConversion` decorator](https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#_a_id_decoration_a_decoration).
94 ## JavaScript FP -> Integer conversion
96 modular with Inf/NaN converted to 0
98 TODO, explain this further, it is not a sentence:
99 "Semantics required by JavaScript"
103 TODO: review and investigate other language semantics
107 * <https://bugs.libre-soc.org/show_bug.cgi?id=650>
108 * <https://bugs.libre-soc.org/show_bug.cgi?id=230#c71>
109 * <https://bugs.libre-soc.org/show_bug.cgi?id=230#c74>
110 * <https://bugs.libre-soc.org/show_bug.cgi?id=230#c76>
112 # Proposed New Scalar Instructions
114 All of the following instructions use the standard OpenPower conversion to/from 64-bit float format when reading/writing a 32-bit float from/to a FPR.
116 This can be overridden by SimpleV, which sets the following
117 operation "reinterpretation" rules:
119 * any operation whose assembler mnemonic does not end in "s"
120 (being defined in v3.0B as a "double" operation) is
121 instead an operation at the overridden elwidth for the
123 * any operation nominally defined as a "single" FP operation
124 is redefined to be **half the elwidth** rather than
129 * `sv.fmvtg/sw=32 RT.v, FRA.v` is defined as treating FRA
130 as a vector of *FP32* source operands each *32* bits wide
131 which are to be placed into *64* bit integer destination elements.
132 * `sv.fmvfgs/dw=32 FRT.v, RA.v` is defined as taking the bottom
133 32 bits of each RA integer source, then performing a **32 bit**
134 FP32 to **FP16** conversion and storing the result in the
135 **32 bits** of an FRT destination element.
137 "Single" is therefore redefined in SVP64 to be "half elwidth"
138 rather than Double width hardcoded to 64 and Single width
139 hardcoded to 32. This allows a full range of conversions
140 between FP64, FP32, FP16 and BF16.
147 move a 64-bit float from a FPR to a GPR, just copying bits directly.
148 Rc=1 tests RT and sets CR0
153 move a 32-bit float from a FPR to a GPR, just copying bits. Converts the
154 64-bit float in `FRA` to a 32-bit float, then writes the 32-bit float to
156 Rc=1 tests RT and sets CR0
162 move a 64-bit float from a GPR to a FPR, just copying bits. No exceptions
163 are raised, no flags are altered of any kind.
169 move a 32-bit float from a GPR to a FPR, just copying bits. Converts the
170 32-bit float in `RA` to a 64-bit float, then writes the 64-bit float to
175 TODO: clear statement on evaluation as to whethrer exceptions or flags raised as part of the **FP** conversion (not the int bitcopy part, the conversion part. the semantics should really be the same as frsp)
177 v3.0C section 4.6.7.1 states:
179 FPRF is set to the class and sign of the result, except for Invalid Operation Exceptions when VE=1.
181 Special Registers Altered:
186 ### Float load immediate (kinda a variant of `fmvfg`)
190 Reinterprets `FI << 16` as a 32-bit float, which is then converted to a
191 64-bit float and written to `FRT`. This is equivalent to reinterpreting
192 `FI` as a `BF16` and converting to 64-bit float.
198 fmvis f4, 0 # writes +0.0 to f4
199 # loading handy constants
200 fmvis f4, 0x8000 # writes -0.0 to f4
201 fmvis f4, 0x3F80 # writes +1.0 to f4
202 fmvis f4, 0xBF80 # writes -1.0 to f4
203 fmvis f4, 0xBFC0 # writes -1.5 to f4
204 fmvis f4, 0x7FC0 # writes +qNaN to f4
205 fmvis f4, 0x7F80 # writes +Infinity to f4
206 fmvis f4, 0xFF80 # writes -Infinity to f4
207 fmvis f4, 0x3FFF # writes +1.9921875 to f4
209 # clearing 128 FPRs with 2 SVP64 instructions
210 # by issuing 32 vec4 (subvector length 4) ops
212 sv.fmvis/vec4 f0, 0 # writes +0.0 to f0-f127
214 Important: If the float load immediate instruction(s) are left out,
215 change all [GPR to FPR conversion instructions](#GPR-to-FPR-conversions)
216 to instead write `+0.0` if `RA` is register `0`, at least
217 allowing clearing FPRs.
219 | 0-5 | 6-10 | 11-25 | 26-30 | 31 |
220 |--------|------|-------|-------|-----|
221 | Major | FRT | FI | XO | FI0 |
223 The above fits reasonably well with Minor 19 and follows the
224 pattern shown by `addpcis`, which uses an entire column of Minor 19
225 XO. 15 bits of FI fit into bits 11 to 25,
226 the top bit FI0 (MSB0 numbered 0) makes 16.
229 fp32 = bf16 || [0]*16
230 FRT = Single_to_Double(fp32)
232 ## FPR to GPR conversions
234 <div id="fpr-to-gpr-conversion-mode"></div>
238 | 0-5 | 6-10 | 11-15 | 16-25 | 26-30 | 31 |
239 |--------|------|--------|-------|-------|----|
240 | Major | RT | //Mode | FRA | XO | Rc |
241 | Major | FRT | //Mode | RA | XO | Rc |
245 | Mode | `rounding_mode` | Semantics |
246 |------|-----------------|----------------------------------|
247 | 000 | from `FPSCR` | [OpenPower semantics] |
248 | 001 | Truncate | [OpenPower semantics] |
249 | 010 | from `FPSCR` | [Rust semantics] |
250 | 011 | Truncate | [Rust semantics] |
251 | 100 | from `FPSCR` | [JavaScript semantics] |
252 | 101 | Truncate | [JavaScript semantics] |
253 | rest | -- | illegal instruction trap for now |
255 [OpenPower semantics]: #fp-to-int-openpower-conversion-semantics
256 [Rust semantics]: #fp-to-int-rust-conversion-semantics
257 [JavaScript semantics]: #fp-to-int-javascript-conversion-semantics
259 `fcvttgw RT, FRA, Mode`
261 Convert from 64-bit float to 32-bit signed integer, writing the result
262 to the GPR `RT`. Converts using [mode `Mode`]
264 `fcvttguw RT, FRA, Mode`
266 Convert from 64-bit float to 32-bit unsigned integer, writing the result
267 to the GPR `RT`. Converts using [mode `Mode`]
269 `fcvttgd RT, FRA, Mode`
271 Convert from 64-bit float to 64-bit signed integer, writing the result
272 to the GPR `RT`. Converts using [mode `Mode`]
274 `fcvttgud RT, FRA, Mode`
276 Convert from 64-bit float to 64-bit unsigned integer, writing the result
277 to the GPR `RT`. Converts using [mode `Mode`]
279 `fcvtstgw RT, FRA, Mode`
281 Convert from 32-bit float to 32-bit signed integer, writing the result
282 to the GPR `RT`. Converts using [mode `Mode`]
284 `fcvtstguw RT, FRA, Mode`
286 Convert from 32-bit float to 32-bit unsigned integer, writing the result
287 to the GPR `RT`. Converts using [mode `Mode`]
289 `fcvtstgd RT, FRA, Mode`
291 Convert from 32-bit float to 64-bit signed integer, writing the result
292 to the GPR `RT`. Converts using [mode `Mode`]
294 `fcvtstgud RT, FRA, Mode`
296 Convert from 32-bit float to 64-bit unsigned integer, writing the result
297 to the GPR `RT`. Converts using [mode `Mode`]
299 [mode `Mode`]: #fpr-to-gpr-conversion-mode
301 ## GPR to FPR conversions
303 All of the following GPR to FPR conversions use the rounding mode from `FPSCR`.
307 Convert from 32-bit signed integer in the GPR `RA` to 64-bit float in `FRT`.
311 Convert from 32-bit signed integer in the GPR `RA` to 32-bit float in `FRT`.
315 Convert from 32-bit unsigned integer in the GPR `RA` to 64-bit float in `FRT`.
319 Convert from 32-bit unsigned integer in the GPR `RA` to 32-bit float in `FRT`.
323 Convert from 64-bit signed integer in the GPR `RA` to 64-bit float in `FRT`.
327 Convert from 64-bit signed integer in the GPR `RA` to 32-bit float in `FRT`.
331 Convert from 64-bit unsigned integer in the GPR `RA` to 64-bit float in `FRT`.
335 Convert from 64-bit unsigned integer in the GPR `RA` to 32-bit float in `FRT`.
337 # FP to Integer Conversion Pseudo-code
341 | term | result type | definition |
342 |---------------------------|-------------|----------------------------------------------------------------------------------------------------|
343 | `fp` | -- | `f32` or `f64` (or other types from SimpleV) |
344 | `int` | -- | `u32`/`u64`/`i32`/`i64` (or other types from SimpleV) |
345 | `uint` | -- | the unsigned integer of the same bit-width as `int` |
346 | `int::BITS` | `int` | the bit-width of `int` |
347 | `int::MIN_VALUE` | `int` | the minimum value `int` can store (`0` if unsigned, `-2^(int::BITS-1)` if signed) |
348 | `int::MAX_VALUE` | `int` | the maximum value `int` can store (`2^int::BITS - 1` if unsigned, `2^(int::BITS-1) - 1` if signed) |
349 | `int::VALUE_COUNT` | Integer | the number of different values `int` can store (`2^int::BITS`). too big to fit in `int`. |
350 | `rint(fp, rounding_mode)` | `fp` | rounds the floating-point value `fp` to an integer according to rounding mode `rounding_mode` |
352 <div id="fp-to-int-openpower-conversion-semantics"></div>
353 OpenPower conversion semantics (section A.2 page 999 (page 1023) of OpenPower ISA v3.1):
356 def fp_to_int_open_power<fp, int>(v: fp) -> int:
358 return int::MIN_VALUE
359 if v >= int::MAX_VALUE:
360 return int::MAX_VALUE
361 if v <= int::MIN_VALUE:
362 return int::MIN_VALUE
363 return (int)rint(v, rounding_mode)
366 <div id="fp-to-int-rust-conversion-semantics"></div>
367 Rust [conversion semantics](https://doc.rust-lang.org/reference/expressions/operator-expr.html#semantics) (with adjustment to add non-truncate rounding modes):
370 def fp_to_int_rust<fp, int>(v: fp) -> int:
373 if v >= int::MAX_VALUE:
374 return int::MAX_VALUE
375 if v <= int::MIN_VALUE:
376 return int::MIN_VALUE
377 return (int)rint(v, rounding_mode)
380 <div id="fp-to-int-javascript-conversion-semantics"></div>
381 JavaScript [conversion semantics](https://262.ecma-international.org/11.0/#sec-toint32) (with adjustment to add non-truncate rounding modes):
384 def fp_to_int_java_script<fp, int>(v: fp) -> int:
385 if v is NaN or infinite:
387 v = rint(v, rounding_mode)
388 v = v mod int::VALUE_COUNT # 2^32 for i32, 2^64 for i64, result is non-negative
393 # Equivalent OpenPower ISA v3.0 Assembly Language for FP -> Integer Conversion Modes
395 Moved to [[int_fp_mv/appendix]]