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 conversion (**TODO, which standard?** can it
38 be described in words? how does it differ from the other "standards"?)
39 * standard OpenPower FP -> Integer conversion (saturation with NaN
40 converted to minimum valid integer)
41 * Rust FP -> Integer conversion (saturation with NaN converted to 0)
42 * JavaScript FP -> Integer conversion (modular with Inf/NaN converted to 0)
44 # A bit more research into integer - fp conversion
46 here is a paragraph which explains that there are different semantics
47 for conversion, i don't know what the paragraph should say, but it needs
48 to be here, to give some background. it also acts as a lead-in to the
49 sub-sections, introducing them and explaining why they are here, as
50 justifications and background research as to why the ISA should support
51 the feature being proposed.
53 *nothing* can be left to chance or guesswork.
55 ## standard Integer -> FP conversion
57 TODO, explain this further
59 - rounding mode read from FPSCR
61 # standard OpenPower FP -> Integer conversion
63 TODO, explain this further, make this a complete sentence:
64 "saturation with NaN converted to minimum valid integer"
66 - Matches x86's conversion semantics
67 - Has instructions for both:
68 * rounding mode read from FPSCR
69 * rounding mode is always truncate
71 ## Rust FP -> Integer conversion
73 TODO, explain this further, the following is not a complete sentence,
74 "saturation with NaN converted to 0"
76 Semantics required by all of:
77 (what does this mean, what is "required"?
78 what semantics are being referred to? the sentence needs completing:
79 "For Rust integer conversion, the semantics required are shown by the
80 following, all of which are supported in XYZ" something like that)
82 * Rust's FP -> Integer conversion using the
83 [`as` operator](https://doc.rust-lang.org/reference/expressions/operator-expr.html#semantics)
85 [FP -> Integer conversion](https://docs.oracle.com/javase/specs/jls/se16/html/jls-5.html#jls-5.1.3)
87 [`llvm.fptosi.sat`](https://llvm.org/docs/LangRef.html#llvm-fptosi-sat-intrinsic) and
88 [`llvm.fptoui.sat`](https://llvm.org/docs/LangRef.html#llvm-fptoui-sat-intrinsic) intrinsics
89 * SPIR-V's OpenCL dialect's
90 [`OpConvertFToU`](https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpConvertFToU) and
91 [`OpConvertFToS`](https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpConvertFToS)
92 instructions when decorated with
93 [the `SaturatedConversion` decorator](https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#_a_id_decoration_a_decoration).
95 ## JavaScript FP -> Integer conversion
97 modular with Inf/NaN converted to 0
99 TODO, explain this further, it is not a sentence:
100 "Semantics required by JavaScript"
104 TODO: review and investigate other language semantics
108 * <https://bugs.libre-soc.org/show_bug.cgi?id=650>
109 * <https://bugs.libre-soc.org/show_bug.cgi?id=230#c71>
110 * <https://bugs.libre-soc.org/show_bug.cgi?id=230#c74>
111 * <https://bugs.libre-soc.org/show_bug.cgi?id=230#c76>
113 # Proposed New Scalar Instructions
115 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.
117 This can be overridden by SimpleV, which sets the following
118 operation "reinterpretation" rules:
120 * any operation whose assembler mnemonic does not end in "s"
121 (being defined in v3.0B as a "double" operation) is
122 instead an operation at the overridden elwidth for the
124 * any operation nominally defined as a "single" FP operation
125 is redefined to be **half the elwidth** rather than
130 * `sv.fmvtg/sw=32 RT.v, FRA.v` is defined as treating FRA
131 as a vector of *FP32* source operands each *32* bits wide
132 which are to be placed into *64* bit integer destination elements.
133 * `sv.fmvfgs/dw=32 FRT.v, RA.v` is defined as taking the bottom
134 32 bits of each RA integer source, then performing a **32 bit**
135 FP32 to **FP16** conversion and storing the result in the
136 **32 bits** of an FRT destination element.
138 "Single" is therefore redefined in SVP64 to be "half elwidth"
139 rather than Double width hardcoded to 64 and Single width
140 hardcoded to 32. This allows a full range of conversions
141 between FP64, FP32, FP16 and BF16.
148 move a 64-bit float from a FPR to a GPR, just copying bits directly.
149 Rc=1 tests RT and sets CR0
154 move a 32-bit float from a FPR to a GPR, just copying bits. Converts the
155 64-bit float in `FRA` to a 32-bit float, then writes the 32-bit float to
157 Rc=1 tests RT and sets CR0
163 move a 64-bit float from a GPR to a FPR, just copying bits.
167 move a 32-bit float from a GPR to a FPR, just copying bits. Converts the
168 32-bit float in `RA` to a 64-bit float, then writes the 64-bit float to
173 ### Float load immediate (kinda a variant of `fmvfg`)
177 Reinterprets `FI << 16` as a 32-bit float, which is then converted to a
178 64-bit float and written to `FRT`. This is equivalent to reinterpreting
179 `FI` as a `BF16` and converting to 64-bit float.
185 fmvis f4, 0 # writes +0.0 to f4
186 # loading handy constants
187 fmvis f4, 0x8000 # writes -0.0 to f4
188 fmvis f4, 0x3F80 # writes +1.0 to f4
189 fmvis f4, 0xBF80 # writes -1.0 to f4
190 fmvis f4, 0xBFC0 # writes -1.5 to f4
191 fmvis f4, 0x7FC0 # writes +qNaN to f4
192 fmvis f4, 0x7F80 # writes +Infinity to f4
193 fmvis f4, 0xFF80 # writes -Infinity to f4
194 fmvis f4, 0x3FFF # writes +1.9921875 to f4
196 # clearing 128 FPRs with 2 SVP64 instructions
197 # by issuing 32 vec4 (subvector length 4) ops
199 sv.fmvis/vec4 f0, 0 # writes +0.0 to f0-f127
201 Important: If the float load immediate instruction(s) are left out,
202 change all [GPR to FPR conversion instructions](#GPR-to-FPR-conversions)
203 to instead write `+0.0` if `RA` is register `0`, at least
204 allowing clearing FPRs.
206 | 0-5 | 6-10 | 11-25 | 26-30 | 31 |
207 |--------|------|-------|-------|-----|
208 | Major | FRT | FI | XO | FI0 |
210 The above fits reasonably well with Minor 19 and follows the
211 pattern shown by `addpcis`, which uses an entire column of Minor 19
212 XO. 15 bits of FI fit into bits 11 to 25,
213 the top bit FI0 (MSB0 numbered 0) makes 16.
216 fp32 = bf16 || [0]*16
217 FRT = Single_to_Double(fp32)
219 ## FPR to GPR conversions
221 <div id="fpr-to-gpr-conversion-mode"></div>
225 | 0-5 | 6-10 | 11-15 | 16-25 | 26-30 | 31 |
226 |--------|------|--------|-------|-------|----|
227 | Major | RT | //Mode | FRA | XO | Rc |
228 | Major | FRT | //Mode | RA | XO | Rc |
232 | Mode | `rounding_mode` | Semantics |
233 |------|-----------------|----------------------------------|
234 | 000 | from `FPSCR` | [OpenPower semantics] |
235 | 001 | Truncate | [OpenPower semantics] |
236 | 010 | from `FPSCR` | [Rust semantics] |
237 | 011 | Truncate | [Rust semantics] |
238 | 100 | from `FPSCR` | [JavaScript semantics] |
239 | 101 | Truncate | [JavaScript semantics] |
240 | rest | -- | illegal instruction trap for now |
242 [OpenPower semantics]: #fp-to-int-openpower-conversion-semantics
243 [Rust semantics]: #fp-to-int-rust-conversion-semantics
244 [JavaScript semantics]: #fp-to-int-javascript-conversion-semantics
246 `fcvttgw RT, FRA, Mode`
248 Convert from 64-bit float to 32-bit signed integer, writing the result
249 to the GPR `RT`. Converts using [mode `Mode`]
251 `fcvttguw RT, FRA, Mode`
253 Convert from 64-bit float to 32-bit unsigned integer, writing the result
254 to the GPR `RT`. Converts using [mode `Mode`]
256 `fcvttgd RT, FRA, Mode`
258 Convert from 64-bit float to 64-bit signed integer, writing the result
259 to the GPR `RT`. Converts using [mode `Mode`]
261 `fcvttgud RT, FRA, Mode`
263 Convert from 64-bit float to 64-bit unsigned integer, writing the result
264 to the GPR `RT`. Converts using [mode `Mode`]
266 `fcvtstgw RT, FRA, Mode`
268 Convert from 32-bit float to 32-bit signed integer, writing the result
269 to the GPR `RT`. Converts using [mode `Mode`]
271 `fcvtstguw RT, FRA, Mode`
273 Convert from 32-bit float to 32-bit unsigned integer, writing the result
274 to the GPR `RT`. Converts using [mode `Mode`]
276 `fcvtstgd RT, FRA, Mode`
278 Convert from 32-bit float to 64-bit signed integer, writing the result
279 to the GPR `RT`. Converts using [mode `Mode`]
281 `fcvtstgud RT, FRA, Mode`
283 Convert from 32-bit float to 64-bit unsigned integer, writing the result
284 to the GPR `RT`. Converts using [mode `Mode`]
286 [mode `Mode`]: #fpr-to-gpr-conversion-mode
288 ## GPR to FPR conversions
290 All of the following GPR to FPR conversions use the rounding mode from `FPSCR`.
294 Convert from 32-bit signed integer in the GPR `RA` to 64-bit float in `FRT`.
298 Convert from 32-bit signed integer in the GPR `RA` to 32-bit float in `FRT`.
302 Convert from 32-bit unsigned integer in the GPR `RA` to 64-bit float in `FRT`.
306 Convert from 32-bit unsigned integer in the GPR `RA` to 32-bit float in `FRT`.
310 Convert from 64-bit signed integer in the GPR `RA` to 64-bit float in `FRT`.
314 Convert from 64-bit signed integer in the GPR `RA` to 32-bit float in `FRT`.
318 Convert from 64-bit unsigned integer in the GPR `RA` to 64-bit float in `FRT`.
322 Convert from 64-bit unsigned integer in the GPR `RA` to 32-bit float in `FRT`.
324 # FP to Integer Conversion Pseudo-code
328 | term | result type | definition |
329 |---------------------------|-------------|----------------------------------------------------------------------------------------------------|
330 | `fp` | -- | `f32` or `f64` (or other types from SimpleV) |
331 | `int` | -- | `u32`/`u64`/`i32`/`i64` (or other types from SimpleV) |
332 | `uint` | -- | the unsigned integer of the same bit-width as `int` |
333 | `int::BITS` | `int` | the bit-width of `int` |
334 | `int::MIN_VALUE` | `int` | the minimum value `int` can store (`0` if unsigned, `-2^(int::BITS-1)` if signed) |
335 | `int::MAX_VALUE` | `int` | the maximum value `int` can store (`2^int::BITS - 1` if unsigned, `2^(int::BITS-1) - 1` if signed) |
336 | `int::VALUE_COUNT` | Integer | the number of different values `int` can store (`2^int::BITS`). too big to fit in `int`. |
337 | `rint(fp, rounding_mode)` | `fp` | rounds the floating-point value `fp` to an integer according to rounding mode `rounding_mode` |
339 <div id="fp-to-int-openpower-conversion-semantics"></div>
340 OpenPower conversion semantics (section A.2 page 999 (page 1023) of OpenPower ISA v3.1):
343 def fp_to_int_open_power<fp, int>(v: fp) -> int:
345 return int::MIN_VALUE
346 if v >= int::MAX_VALUE:
347 return int::MAX_VALUE
348 if v <= int::MIN_VALUE:
349 return int::MIN_VALUE
350 return (int)rint(v, rounding_mode)
353 <div id="fp-to-int-rust-conversion-semantics"></div>
354 Rust [conversion semantics](https://doc.rust-lang.org/reference/expressions/operator-expr.html#semantics) (with adjustment to add non-truncate rounding modes):
357 def fp_to_int_rust<fp, int>(v: fp) -> int:
360 if v >= int::MAX_VALUE:
361 return int::MAX_VALUE
362 if v <= int::MIN_VALUE:
363 return int::MIN_VALUE
364 return (int)rint(v, rounding_mode)
367 <div id="fp-to-int-javascript-conversion-semantics"></div>
368 JavaScript [conversion semantics](https://262.ecma-international.org/11.0/#sec-toint32) (with adjustment to add non-truncate rounding modes):
371 def fp_to_int_java_script<fp, int>(v: fp) -> int:
372 if v is NaN or infinite:
374 v = rint(v, rounding_mode)
375 v = v mod int::VALUE_COUNT # 2^32 for i32, 2^64 for i64, result is non-negative
380 # Equivalent OpenPower ISA v3.0 Assembly Language for FP -> Integer Conversion Modes
382 Moved to [[int_fp_mv/appendix]]