OpenCores
URL https://opencores.org/ocsvn/neorv32/neorv32/trunk

Subversion Repositories neorv32

[/] [neorv32/] [trunk/] [sw/] [example/] [floating_point_test/] [neorv32_zfinx_extension_intrinsics.h] - Blame information for rev 63

Go to most recent revision | Details | Compare with Previous | View Log

Line No. Rev Author Line
1 55 zero_gravi 
// #################################################################################################
2 
// # << NEORV32 - Intrinsics + Emulation Functions for the RISC-V "Zfinx" CPU extension >>         #
3 
// # ********************************************************************************************* #
4 
// # The intrinsics provided by this library allow to use the hardware floating-point unit of the  #
5 
// # RISC-V Zfinx CPU extension without the need for Zfinx support by the compiler / toolchain.    #
6 
// # ********************************************************************************************* #
7 
// # BSD 3-Clause License                                                                          #
8 
// #                                                                                               #
9 
// # Copyright (c) 2021, Stephan Nolting. All rights reserved.                                     #
10 
// #                                                                                               #
11 
// # Redistribution and use in source and binary forms, with or without modification, are          #
12 
// # permitted provided that the following conditions are met:                                     #
13 
// #                                                                                               #
14 
// # 1. Redistributions of source code must retain the above copyright notice, this list of        #
15 
// #    conditions and the following disclaimer.                                                   #
16 
// #                                                                                               #
17 
// # 2. Redistributions in binary form must reproduce the above copyright notice, this list of     #
18 
// #    conditions and the following disclaimer in the documentation and/or other materials        #
19 
// #    provided with the distribution.                                                            #
20 
// #                                                                                               #
21 
// # 3. Neither the name of the copyright holder nor the names of its contributors may be used to  #
22 
// #    endorse or promote products derived from this software without specific prior written      #
23 
// #    permission.                                                                                #
24 
// #                                                                                               #
25 
// # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS   #
26 
// # OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF               #
27 
// # MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE    #
28 
// # COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,     #
29 
// # EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE #
30 
// # GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED    #
31 
// # AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING     #
32 
// # NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED  #
33 
// # OF THE POSSIBILITY OF SUCH DAMAGE.                                                            #
34 
// # ********************************************************************************************* #
35 
// # The NEORV32 Processor - https://github.com/stnolting/neorv32              (c) Stephan Nolting #
36 
// #################################################################################################
37 
 
38 
 
39 
/**********************************************************************//**
40 
 * @file floating_point_test/neorv32_zfinx_extension_intrinsics.h
41 
 * @author Stephan Nolting
42 
 *
43 
 * @brief "Intrinsic" library for the NEORV32 single-precision floating-point in x registers (Zfinx) extension
44 
 * @brief Also provides emulation functions for all intrinsics (functionality re-built in pure software). The functionality of the emulation
45 
 * @brief functions is based on the RISC-V floating-point spec.
46 
 *
47 
 * @note All operations from this library use the default GCC "round to nearest, ties to even" rounding mode.
48 
 *
49 
 * @warning This library is just a temporary fall-back until the Zfinx extensions are supported by the upstream RISC-V GCC port.
50 
 **************************************************************************/
51 
 
52 
#ifndef neorv32_zfinx_extension_intrinsics_h
53 
#define neorv32_zfinx_extension_intrinsics_h
54 
 
55 
#define __USE_GNU
56 
 
57 
#include <fenv.h>
58 
//#pragma STDC FENV_ACCESS ON
59 
 
60 
#define _GNU_SOURCE
61 
 
62 
#include <float.h>
63 
#include <math.h>
64 
 
65 
 
66 
/**********************************************************************//**
67 
 * Sanity check
68 
 **************************************************************************/
69 
#if defined __riscv_f || (__riscv_flen == 32)
70 
  #error Application programs using the Zfinx intrinsic library have to be compiled WITHOUT the <F> MARCH ISA attribute!
71 
#endif
72 
 
73 
 
74 
/**********************************************************************//**
75 
 * Custom data type to access floating-point values as native floats and in binary representation
76 
 **************************************************************************/
77 
typedef union
78 
{
79 
  uint32_t binary_value; /**< Access as native float */
80 
  float    float_value;  /**< Access in binary representation */
81 
} float_conv_t;
82 
 
83 
 
84 
// ################################################################################################
85 
// Helper functions
86 
// ################################################################################################
87 
 
88 
/**********************************************************************//**
89 
 * Flush to zero if denormal number.
90 
 *
91 
 * @warning Subnormal numbers are not supported yet! Flush them to zero.
92 
 *
93 56 zero_gravi 
 * @param[in] tmp Source operand.
94 55 zero_gravi 
 * @return Result.
95 
 **************************************************************************/
96 
float subnormal_flush(float tmp) {
97 
 
98 
  float res = tmp;
99 
 
100 
  if (fpclassify(tmp) == FP_SUBNORMAL) {
101 
    if (signbit(tmp) != 0) {
102 
      res = -0.0f;
103 
    }
104 
    else {
105 
      res = +0.0f;
106 
    }
107 
  }
108 
 
109 
  return res;
110 
}
111 
 
112 
 
113 
// ################################################################################################
114 
// Exception access
115 
// ################################################################################################
116 
 
117 
/**********************************************************************//**
118 
 * Get exception flags from fflags CSR (floating-point hardware).
119 
 *
120 
 * @return Floating point exception status word.
121 
 **************************************************************************/
122 
uint32_t get_hw_exceptions(void) {
123 
 
124 
  uint32_t res = neorv32_cpu_csr_read(CSR_FFLAGS);
125 
 
126 
  neorv32_cpu_csr_write(CSR_FFLAGS, 0); // clear status word
127 
 
128 
  return res;
129 
}
130 
 
131 
 
132 
/**********************************************************************//**
133 
 * Get exception flags from C runtime (floating-point emulation).
134 
 *
135 
 * @warning WORK-IN-PROGRESS!
136 
 *
137 
 * @return Floating point exception status word.
138 
 **************************************************************************/
139 
uint32_t get_sw_exceptions(void) {
140 
 
141 
  const uint32_t FP_EXC_NV_C = 1 << 0; // invalid operation
142 
  const uint32_t FP_EXC_DZ_C = 1 << 1; // divide by zero
143 
  const uint32_t FP_EXC_OF_C = 1 << 2; // overflow
144 
  const uint32_t FP_EXC_UF_C = 1 << 3; // underflow
145 
  const uint32_t FP_EXC_NX_C = 1 << 4; // inexact
146 
 
147 
  int fpeRaised = fetestexcept(FE_ALL_EXCEPT);
148 
 
149 
  uint32_t res = 0;
150 
 
151 
  if (fpeRaised & FE_INVALID)   { res |= FP_EXC_NV_C; }
152 
  if (fpeRaised & FE_DIVBYZERO) { res |= FP_EXC_DZ_C; }
153 
  if (fpeRaised & FE_OVERFLOW)  { res |= FP_EXC_OF_C; }
154 
  if (fpeRaised & FE_UNDERFLOW) { res |= FP_EXC_UF_C; }
155 
  if (fpeRaised & FE_INEXACT)   { res |= FP_EXC_NX_C; }
156 
 
157 
  feclearexcept(FE_ALL_EXCEPT);
158 
 
159 
  return res;
160 
}
161 
 
162 
 
163 
// ################################################################################################
164 
// "Intrinsics"
165 
// ################################################################################################
166 
 
167 
/**********************************************************************//**
168 
 * Single-precision floating-point addition
169 
 *
170 
 * @param[in] rs1 Source operand 1 (a0).
171 
 * @param[in] rs2 Source operand 2 (a1).
172 
 * @return Result.
173 
 **************************************************************************/
174 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fadds(float rs1, float rs2) {
175 55 zero_gravi 
 
176 
  float_conv_t opa, opb, res;
177 
  opa.float_value = rs1;
178 
  opb.float_value = rs2;
179 
 
180 
  register uint32_t result __asm__ ("a0");
181 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
182 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
183 
 
184 
  // dummy instruction to prevent GCC "constprop" optimization
185 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
186 55 zero_gravi 
 
187 
  // fadd.s a0, a0, a1
188 
  CUSTOM_INSTR_R2_TYPE(0b0000000, a1, a0, 0b000, a0, 0b1010011);
189 
 
190 
  res.binary_value = result;
191 
  return res.float_value;
192 
}
193 
 
194 
 
195 
/**********************************************************************//**
196 
 * Single-precision floating-point subtraction
197 
 *
198 
 * @param[in] rs1 Source operand 1 (a0).
199 
 * @param[in] rs2 Source operand 2 (a1).
200 
 * @return Result.
201 
 **************************************************************************/
202 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fsubs(float rs1, float rs2) {
203 55 zero_gravi 
 
204 
  float_conv_t opa, opb, res;
205 
  opa.float_value = rs1;
206 
  opb.float_value = rs2;
207 
 
208 
  register uint32_t result __asm__ ("a0");
209 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
210 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
211 
 
212 
  // dummy instruction to prevent GCC "constprop" optimization
213 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
214 55 zero_gravi 
 
215 
  // fsub.s a0, a0, a1
216 
  CUSTOM_INSTR_R2_TYPE(0b0000100, a1, a0, 0b000, a0, 0b1010011);
217 
 
218 
  res.binary_value = result;
219 
  return res.float_value;
220 
}
221 
 
222 
 
223 
/**********************************************************************//**
224 
 * Single-precision floating-point multiplication
225 
 *
226 
 * @param[in] rs1 Source operand 1 (a0).
227 
 * @param[in] rs2 Source operand 2 (a1).
228 
 * @return Result.
229 
 **************************************************************************/
230 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fmuls(float rs1, float rs2) {
231 55 zero_gravi 
 
232 
  float_conv_t opa, opb, res;
233 
  opa.float_value = rs1;
234 
  opb.float_value = rs2;
235 
 
236 
  register uint32_t result __asm__ ("a0");
237 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
238 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
239 
 
240 
  // dummy instruction to prevent GCC "constprop" optimization
241 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
242 55 zero_gravi 
 
243 
  // fmul.s a0, a0, a1
244 
  CUSTOM_INSTR_R2_TYPE(0b0001000, a1, a0, 0b000, a0, 0b1010011);
245 
 
246 
  res.binary_value = result;
247 
  return res.float_value;
248 
}
249 
 
250 
 
251 
/**********************************************************************//**
252 
 * Single-precision floating-point minimum
253 
 *
254 
 * @param[in] rs1 Source operand 1 (a0).
255 
 * @param[in] rs2 Source operand 2 (a1).
256 
 * @return Result.
257 
 **************************************************************************/
258 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fmins(float rs1, float rs2) {
259 55 zero_gravi 
 
260 
  float_conv_t opa, opb, res;
261 
  opa.float_value = rs1;
262 
  opb.float_value = rs2;
263 
 
264 
  register uint32_t result __asm__ ("a0");
265 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
266 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
267 
 
268 
  // dummy instruction to prevent GCC "constprop" optimization
269 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
270 55 zero_gravi 
 
271 
  // fmin.s a0, a0, a1
272 
  CUSTOM_INSTR_R2_TYPE(0b0010100, a1, a0, 0b000, a0, 0b1010011);
273 
 
274 
  res.binary_value = result;
275 
  return res.float_value;
276 
}
277 
 
278 
 
279 
/**********************************************************************//**
280 
 * Single-precision floating-point maximum
281 
 *
282 
 * @param[in] rs1 Source operand 1 (a0).
283 
 * @param[in] rs2 Source operand 2 (a1).
284 
 * @return Result.
285 
 **************************************************************************/
286 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fmaxs(float rs1, float rs2) {
287 55 zero_gravi 
 
288 
  float_conv_t opa, opb, res;
289 
  opa.float_value = rs1;
290 
  opb.float_value = rs2;
291 
 
292 
  register uint32_t result __asm__ ("a0");
293 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
294 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
295 
 
296 
  // dummy instruction to prevent GCC "constprop" optimization
297 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
298 55 zero_gravi 
 
299 
  // fmax.s a0, a0, a1
300 
  CUSTOM_INSTR_R2_TYPE(0b0010100, a1, a0, 0b001, a0, 0b1010011);
301 
 
302 
  res.binary_value = result;
303 
  return res.float_value;
304 
}
305 
 
306 
 
307 
/**********************************************************************//**
308 
 * Single-precision floating-point convert float to unsigned integer
309 
 *
310 
 * @param[in] rs1 Source operand 1 (a0).
311 
 * @return Result.
312 
 **************************************************************************/
313 56 zero_gravi 
inline uint32_t __attribute__ ((always_inline)) riscv_intrinsic_fcvt_wus(float rs1) {
314 55 zero_gravi 
 
315 
  float_conv_t opa;
316 
  opa.float_value = rs1;
317 
 
318 
  register uint32_t result __asm__ ("a0");
319 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
320 
 
321 
  // dummy instruction to prevent GCC "constprop" optimization
322 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a));
323 55 zero_gravi 
 
324 
  // fcvt.wu.s a0, a0
325 
  CUSTOM_INSTR_R2_TYPE(0b1100000, x1, a0, 0b000, a0, 0b1010011);
326 
 
327 
  return result;
328 
}
329 
 
330 
 
331 
/**********************************************************************//**
332 
 * Single-precision floating-point convert float to signed integer
333 
 *
334 
 * @param[in] rs1 Source operand 1 (a0).
335 
 * @return Result.
336 
 **************************************************************************/
337 56 zero_gravi 
inline int32_t __attribute__ ((always_inline)) riscv_intrinsic_fcvt_ws(float rs1) {
338 55 zero_gravi 
 
339 
  float_conv_t opa;
340 
  opa.float_value = rs1;
341 
 
342 
  register uint32_t result __asm__ ("a0");
343 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
344 
 
345 
  // dummy instruction to prevent GCC "constprop" optimization
346 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a));
347 55 zero_gravi 
 
348 
  // fcvt.w.s a0, a0
349 
  CUSTOM_INSTR_R2_TYPE(0b1100000, x0, a0, 0b000, a0, 0b1010011);
350 
 
351 
  return (int32_t)result;
352 
}
353 
 
354 
 
355 
/**********************************************************************//**
356 
 * Single-precision floating-point convert unsigned integer to float
357 
 *
358 
 * @param[in] rs1 Source operand 1 (a0).
359 
 * @return Result.
360 
 **************************************************************************/
361 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fcvt_swu(uint32_t rs1) {
362 55 zero_gravi 
 
363 
  float_conv_t res;
364 
 
365 
  register uint32_t result __asm__ ("a0");
366 
  register uint32_t tmp_a  __asm__ ("a0") = rs1;
367 
 
368 
  // dummy instruction to prevent GCC "constprop" optimization
369 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a));
370 55 zero_gravi 
 
371 
  // fcvt.s.wu a0, a0
372 
  CUSTOM_INSTR_R2_TYPE(0b1101000, x1, a0, 0b000, a0, 0b1010011);
373 
 
374 
  res.binary_value = result;
375 
  return res.float_value;
376 
}
377 
 
378 
 
379 
/**********************************************************************//**
380 
 * Single-precision floating-point convert signed integer to float
381 
 *
382 
 * @param[in] rs1 Source operand 1 (a0).
383 
 * @return Result.
384 
 **************************************************************************/
385 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fcvt_sw(int32_t rs1) {
386 55 zero_gravi 
 
387 
  float_conv_t res;
388 
 
389 
  register uint32_t result __asm__ ("a0");
390 
  register uint32_t tmp_a  __asm__ ("a0") = (uint32_t)rs1;
391 
 
392 
  // dummy instruction to prevent GCC "constprop" optimization
393 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a));
394 55 zero_gravi 
 
395 
  // fcvt.s.w a0, a0
396 
  CUSTOM_INSTR_R2_TYPE(0b1101000, x0, a0, 0b000, a0, 0b1010011);
397 
 
398 
  res.binary_value = result;
399 
  return res.float_value;
400 
}
401 
 
402 
 
403 
/**********************************************************************//**
404 
 * Single-precision floating-point equal comparison
405 
 *
406 
 * @param[in] rs1 Source operand 1 (a0).
407 
 * @param[in] rs2 Source operand 2 (a1).
408 
 * @return Result.
409 
 **************************************************************************/
410 56 zero_gravi 
inline uint32_t __attribute__ ((always_inline)) riscv_intrinsic_feqs(float rs1, float rs2) {
411 55 zero_gravi 
 
412 
  float_conv_t opa, opb;
413 
  opa.float_value = rs1;
414 
  opb.float_value = rs2;
415 
 
416 
  register uint32_t result __asm__ ("a0");
417 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
418 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
419 
 
420 
  // dummy instruction to prevent GCC "constprop" optimization
421 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
422 55 zero_gravi 
 
423 
  // feq.s a0, a0, a1
424 
  CUSTOM_INSTR_R2_TYPE(0b1010000, a1, a0, 0b010, a0, 0b1010011);
425 
 
426 
  return result;
427 
}
428 
 
429 
 
430 
/**********************************************************************//**
431 
 * Single-precision floating-point less-than comparison
432 
 *
433 
 * @param[in] rs1 Source operand 1 (a0).
434 
 * @param[in] rs2 Source operand 2 (a1).
435 
 * @return Result.
436 
 **************************************************************************/
437 56 zero_gravi 
inline uint32_t __attribute__ ((always_inline)) riscv_intrinsic_flts(float rs1, float rs2) {
438 55 zero_gravi 
 
439 
  float_conv_t opa, opb;
440 
  opa.float_value = rs1;
441 
  opb.float_value = rs2;
442 
 
443 
  register uint32_t result __asm__ ("a0");
444 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
445 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
446 
 
447 
  // dummy instruction to prevent GCC "constprop" optimization
448 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
449 55 zero_gravi 
 
450 
  // flt.s a0, a0, a1
451 
  CUSTOM_INSTR_R2_TYPE(0b1010000, a1, a0, 0b001, a0, 0b1010011);
452 
 
453 
  return result;
454 
}
455 
 
456 
 
457 
/**********************************************************************//**
458 
 * Single-precision floating-point less-than-or-equal comparison
459 
 *
460 
 * @param[in] rs1 Source operand 1 (a0).
461 
 * @param[in] rs2 Source operand 2 (a1).
462 
 * @return Result.
463 
 **************************************************************************/
464 56 zero_gravi 
inline uint32_t __attribute__ ((always_inline)) riscv_intrinsic_fles(float rs1, float rs2) {
465 55 zero_gravi 
 
466 
  float_conv_t opa, opb;
467 
  opa.float_value = rs1;
468 
  opb.float_value = rs2;
469 
 
470 
  register uint32_t result __asm__ ("a0");
471 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
472 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
473 
 
474 
  // dummy instruction to prevent GCC "constprop" optimization
475 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
476 55 zero_gravi 
 
477 
  // fle.s a0, a0, a1
478 
  CUSTOM_INSTR_R2_TYPE(0b1010000, a1, a0, 0b000, a0, 0b1010011);
479 
 
480 
  return result;
481 
}
482 
 
483 
 
484 
/**********************************************************************//**
485 
 * Single-precision floating-point sign-injection
486 
 *
487 
 * @param[in] rs1 Source operand 1 (a0).
488 
 * @param[in] rs2 Source operand 2 (a1).
489 
 * @return Result.
490 
 **************************************************************************/
491 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fsgnjs(float rs1, float rs2) {
492 55 zero_gravi 
 
493 
  float_conv_t opa, opb, res;
494 
  opa.float_value = rs1;
495 
  opb.float_value = rs2;
496 
 
497 
  register uint32_t result __asm__ ("a0");
498 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
499 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
500 
 
501 
  // dummy instruction to prevent GCC "constprop" optimization
502 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
503 55 zero_gravi 
 
504 
  // fsgnj.s a0, a0, a1
505 
  CUSTOM_INSTR_R2_TYPE(0b0010000, a1, a0, 0b000, a0, 0b1010011);
506 
 
507 
  res.binary_value = result;
508 
  return res.float_value;
509 
}
510 
 
511 
 
512 
/**********************************************************************//**
513 
 * Single-precision floating-point sign-injection NOT
514 
 *
515 
 * @param[in] rs1 Source operand 1 (a0).
516 
 * @param[in] rs2 Source operand 2 (a1).
517 
 * @return Result.
518 
 **************************************************************************/
519 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fsgnjns(float rs1, float rs2) {
520 55 zero_gravi 
 
521 
  float_conv_t opa, opb, res;
522 
  opa.float_value = rs1;
523 
  opb.float_value = rs2;
524 
 
525 
  register uint32_t result __asm__ ("a0");
526 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
527 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
528 
 
529 
  // dummy instruction to prevent GCC "constprop" optimization
530 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
531 55 zero_gravi 
 
532 
  // fsgnjn.s a0, a0, a1
533 
  CUSTOM_INSTR_R2_TYPE(0b0010000, a1, a0, 0b001, a0, 0b1010011);
534 
 
535 
  res.binary_value = result;
536 
  return res.float_value;
537 
}
538 
 
539 
 
540 
/**********************************************************************//**
541 
 * Single-precision floating-point sign-injection XOR
542 
 *
543 
 * @param[in] rs1 Source operand 1 (a0).
544 
 * @param[in] rs2 Source operand 2 (a1).
545 
 * @return Result.
546 
 **************************************************************************/
547 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fsgnjxs(float rs1, float rs2) {
548 55 zero_gravi 
 
549 
  float_conv_t opa, opb, res;
550 
  opa.float_value = rs1;
551 
  opb.float_value = rs2;
552 
 
553 
  register uint32_t result __asm__ ("a0");
554 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
555 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
556 
 
557 
  // dummy instruction to prevent GCC "constprop" optimization
558 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
559 55 zero_gravi 
 
560 
  // fsgnjx.s a0, a0, a1
561 
  CUSTOM_INSTR_R2_TYPE(0b0010000, a1, a0, 0b010, a0, 0b1010011);
562 
 
563 
  res.binary_value = result;
564 
  return res.float_value;
565 
}
566 
 
567 
 
568 
/**********************************************************************//**
569 
 * Single-precision floating-point number classification
570 
 *
571 
 * @param[in] rs1 Source operand 1 (a0).
572 
 * @return Result.
573 
 **************************************************************************/
574 56 zero_gravi 
inline uint32_t __attribute__ ((always_inline)) riscv_intrinsic_fclasss(float rs1) {
575 55 zero_gravi 
 
576 
  float_conv_t opa;
577 
  opa.float_value = rs1;
578 
 
579 
  register uint32_t result __asm__ ("a0");
580 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
581 
 
582 
  // dummy instruction to prevent GCC "constprop" optimization
583 63 zero_gravi 
  asm volatile ("" : [output] "=r" (result) : [input_i] "r" (tmp_a));
584 55 zero_gravi 
 
585 
  // fclass.s a0, a0
586 
  CUSTOM_INSTR_R2_TYPE(0b1110000, x0, a0, 0b001, a0, 0b1010011);
587 
 
588 
  return result;
589 
}
590 
 
591 
 
592 
// ################################################################################################
593 
// !!! UNSUPPORTED instructions !!!
594 
// ################################################################################################
595 
 
596 
/**********************************************************************//**
597 
 * Single-precision floating-point division
598 
 *
599 
 * @warning This instruction is not supported and should raise an illegal instruction exception when executed.
600 
 *
601 
 * @param[in] rs1 Source operand 1 (a0).
602 
 * @param[in] rs2 Source operand 2 (a1).
603 
 * @return Result.
604 
 **************************************************************************/
605 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fdivs(float rs1, float rs2) {
606 55 zero_gravi 
 
607 
  float_conv_t opa, opb, res;
608 
  opa.float_value = rs1;
609 
  opb.float_value = rs2;
610 
 
611 
  register uint32_t result __asm__ ("a0");
612 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
613 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
614 
 
615 
  // dummy instruction to prevent GCC "constprop" optimization
616 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
617 
 
618 
  // fdiv.s a0, a0, x1
619 
  CUSTOM_INSTR_R2_TYPE(0b0001100, a1, a0, 0b000, a0, 0b1010011);
620 
 
621 56 zero_gravi 
  // dummy instruction to prevent GCC "constprop" optimization
622 
  asm volatile ("add %[res], %[input], x0" : [res] "=r" (result) : [input] "r" (result) );
623 
 
624 55 zero_gravi 
  res.binary_value = result;
625 
  return res.float_value;
626 
}
627 
 
628 
 
629 
/**********************************************************************//**
630 
 * Single-precision floating-point square root
631 
 *
632 
 * @warning This instruction is not supported and should raise an illegal instruction exception when executed.
633 
 *
634 
 * @param[in] rs1 Source operand 1 (a0).
635 
 * @return Result.
636 
 **************************************************************************/
637 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fsqrts(float rs1) {
638 55 zero_gravi 
 
639 
  float_conv_t opa, res;
640 
  opa.float_value = rs1;
641 
 
642 
  register uint32_t result __asm__ ("a0");
643 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
644 
 
645 
  // dummy instruction to prevent GCC "constprop" optimization
646 
  asm volatile ("add x0, %[input_i], x0" : : [input_i] "r" (tmp_a));
647 
 
648 
  // fsqrt.s a0, a0, a1
649 
  CUSTOM_INSTR_R2_TYPE(0b0101100, a1, a0, 0b000, a0, 0b1010011);
650 
 
651 56 zero_gravi 
  // dummy instruction to prevent GCC "constprop" optimization
652 
  asm volatile ("add %[res], %[input], x0" : [res] "=r" (result) : [input] "r" (result) );
653 
 
654 55 zero_gravi 
  res.binary_value = result;
655 
  return res.float_value;
656 
}
657 
 
658 
 
659 
/**********************************************************************//**
660 
 * Single-precision floating-point fused multiply-add
661 
 *
662 
 * @warning This instruction is not supported and should raise an illegal instruction exception when executed.
663 
 *
664 
 * @param[in] rs1 Source operand 1 (a0)
665 
 * @param[in] rs2 Source operand 2 (a1)
666 
 * @param[in] rs3 Source operand 3 (a2)
667 
 * @return Result.
668 
 **************************************************************************/
669 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fmadds(float rs1, float rs2, float rs3) {
670 55 zero_gravi 
 
671 
  float_conv_t opa, opb, opc, res;
672 
  opa.float_value = rs1;
673 
  opb.float_value = rs2;
674 
  opc.float_value = rs3;
675 
 
676 
  register uint32_t result __asm__ ("a0");
677 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
678 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
679 
  register uint32_t tmp_c  __asm__ ("a2") = opc.binary_value;
680 
 
681 
  // dummy instruction to prevent GCC "constprop" optimization
682 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
683 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_b), [input_j] "r" (tmp_c));
684 
 
685 
  // fmadd.s a0, a0, a1, a2
686 
  CUSTOM_INSTR_R3_TYPE(a2, a1, a0, 0b000, a0, 0b1000011);
687 
 
688 56 zero_gravi 
  // dummy instruction to prevent GCC "constprop" optimization
689 
  asm volatile ("add %[res], %[input], x0" : [res] "=r" (result) : [input] "r" (result) );
690 
 
691 55 zero_gravi 
  res.binary_value = result;
692 
  return res.float_value;
693 
}
694 
 
695 
 
696 
/**********************************************************************//**
697 
 * Single-precision floating-point fused multiply-sub
698 
 *
699 
 * @warning This instruction is not supported and should raise an illegal instruction exception when executed.
700 
 *
701 
 * @param[in] rs1 Source operand 1 (a0)
702 
 * @param[in] rs2 Source operand 2 (a1)
703 
 * @param[in] rs3 Source operand 3 (a2)
704 
 * @return Result.
705 
 **************************************************************************/
706 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fmsubs(float rs1, float rs2, float rs3) {
707 55 zero_gravi 
 
708 
  float_conv_t opa, opb, opc, res;
709 
  opa.float_value = rs1;
710 
  opb.float_value = rs2;
711 
  opc.float_value = rs3;
712 
 
713 
  register uint32_t result __asm__ ("a0");
714 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
715 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
716 
  register uint32_t tmp_c  __asm__ ("a2") = opc.binary_value;
717 
 
718 
  // dummy instruction to prevent GCC "constprop" optimization
719 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
720 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_b), [input_j] "r" (tmp_c));
721 
 
722 
  // fmsub.s a0, a0, a1, a2
723 
  CUSTOM_INSTR_R3_TYPE(a2, a1, a0, 0b000, a0, 0b1000111);
724 
 
725 56 zero_gravi 
  // dummy instruction to prevent GCC "constprop" optimization
726 
  asm volatile ("add %[res], %[input], x0" : [res] "=r" (result) : [input] "r" (result) );
727 
 
728 55 zero_gravi 
  res.binary_value = result;
729 
  return res.float_value;
730 
}
731 
 
732 
 
733 
/**********************************************************************//**
734 
 * Single-precision floating-point fused negated multiply-sub
735 
 *
736 
 * @warning This instruction is not supported and should raise an illegal instruction exception when executed.
737 
 *
738 
 * @param[in] rs1 Source operand 1 (a0)
739 
 * @param[in] rs2 Source operand 2 (a1)
740 
 * @param[in] rs3 Source operand 3 (a2)
741 
 * @return Result.
742 
 **************************************************************************/
743 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fnmsubs(float rs1, float rs2, float rs3) {
744 55 zero_gravi 
 
745 
  float_conv_t opa, opb, opc, res;
746 
  opa.float_value = rs1;
747 
  opb.float_value = rs2;
748 
  opc.float_value = rs3;
749 
 
750 
  register uint32_t result __asm__ ("a0");
751 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
752 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
753 
  register uint32_t tmp_c  __asm__ ("a2") = opc.binary_value;
754 
 
755 
  // dummy instruction to prevent GCC "constprop" optimization
756 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
757 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_b), [input_j] "r" (tmp_c));
758 
 
759 
  // fnmsub.s a0, a0, a1, a2
760 
  CUSTOM_INSTR_R3_TYPE(a2, a1, a0, 0b000, a0, 0b1001011);
761 
 
762 56 zero_gravi 
  // dummy instruction to prevent GCC "constprop" optimization
763 
  asm volatile ("add %[res], %[input], x0" : [res] "=r" (result) : [input] "r" (result) );
764 
 
765 55 zero_gravi 
  res.binary_value = result;
766 
  return res.float_value;
767 
}
768 
 
769 
 
770 
/**********************************************************************//**
771 
 * Single-precision floating-point fused negated multiply-add
772 
 *
773 
 * @warning This instruction is not supported and should raise an illegal instruction exception when executed.
774 
 *
775 
 * @param[in] rs1 Source operand 1 (a0)
776 
 * @param[in] rs2 Source operand 2 (a1)
777 
 * @param[in] rs3 Source operand 3 (a2)
778 
 * @return Result.
779 
 **************************************************************************/
780 56 zero_gravi 
inline float __attribute__ ((always_inline)) riscv_intrinsic_fnmadds(float rs1, float rs2, float rs3) {
781 55 zero_gravi 
 
782 
  float_conv_t opa, opb, opc, res;
783 
  opa.float_value = rs1;
784 
  opb.float_value = rs2;
785 
  opc.float_value = rs3;
786 
 
787 
  register uint32_t result __asm__ ("a0");
788 
  register uint32_t tmp_a  __asm__ ("a0") = opa.binary_value;
789 
  register uint32_t tmp_b  __asm__ ("a1") = opb.binary_value;
790 
  register uint32_t tmp_c  __asm__ ("a2") = opc.binary_value;
791 
 
792 
  // dummy instruction to prevent GCC "constprop" optimization
793 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_a), [input_j] "r" (tmp_b));
794 
  asm volatile ("add x0, %[input_i], %[input_j]" : : [input_i] "r" (tmp_b), [input_j] "r" (tmp_c));
795 
 
796 
  // fnmadd.s a0, a0, a1, a2
797 
  CUSTOM_INSTR_R3_TYPE(a2, a1, a0, 0b000, a0, 0b1001111);
798 
 
799 56 zero_gravi 
  // dummy instruction to prevent GCC "constprop" optimization
800 
  asm volatile ("add %[res], %[input], x0" : [res] "=r" (result) : [input] "r" (result) );
801 
 
802 55 zero_gravi 
  res.binary_value = result;
803 
  return res.float_value;
804 
}
805 
 
806 
 
807 
// ################################################################################################
808 
// Emulation functions
809 
// ################################################################################################
810 
 
811 
/**********************************************************************//**
812 
 * Single-precision floating-point addition
813 
 *
814 
 * @param[in] rs1 Source operand 1.
815 
 * @param[in] rs2 Source operand 2.
816 
 * @return Result.
817 
 **************************************************************************/
818 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fadds(float rs1, float rs2) {
819 55 zero_gravi 
 
820 
  float opa = subnormal_flush(rs1);
821 
  float opb = subnormal_flush(rs2);
822 
 
823 
  float res = opa + opb;
824 
  return subnormal_flush(res);
825 
}
826 
 
827 
 
828 
/**********************************************************************//**
829 
 * Single-precision floating-point subtraction
830 
 *
831 
 * @param[in] rs1 Source operand 1.
832 
 * @param[in] rs2 Source operand 2.
833 
 * @return Result.
834 
 **************************************************************************/
835 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fsubs(float rs1, float rs2) {
836 55 zero_gravi 
 
837 
  float opa = subnormal_flush(rs1);
838 
  float opb = subnormal_flush(rs2);
839 
 
840 
  float res = opa - opb;
841 
  return subnormal_flush(res);
842 
}
843 
 
844 
 
845 
/**********************************************************************//**
846 
 * Single-precision floating-point multiplication
847 
 *
848 
 * @param[in] rs1 Source operand 1.
849 
 * @param[in] rs2 Source operand 2.
850 
 * @return Result.
851 
 **************************************************************************/
852 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fmuls(float rs1, float rs2) {
853 55 zero_gravi 
 
854 
  float opa = subnormal_flush(rs1);
855 
  float opb = subnormal_flush(rs2);
856 
 
857 
  float res = opa * opb;
858 
  return subnormal_flush(res);
859 
}
860 
 
861 
 
862 
/**********************************************************************//**
863 
 * Single-precision floating-point minimum
864 
 *
865 
 * @param[in] rs1 Source operand 1.
866 
 * @param[in] rs2 Source operand 2.
867 
 * @return Result.
868 
 **************************************************************************/
869 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fmins(float rs1, float rs2) {
870 55 zero_gravi 
 
871 
  float opa = subnormal_flush(rs1);
872 
  float opb = subnormal_flush(rs2);
873 
 
874 
  union {
875 
  uint32_t binary_value; /**< Access as native float */
876 
  float    float_value;  /**< Access in binary representation */
877 
  } tmp_a, tmp_b;
878 
 
879 
  if ((fpclassify(opa) == FP_NAN) && (fpclassify(opb) == FP_NAN)) {
880 
    return nanf("");
881 
  }
882 
 
883 
  if (fpclassify(opa) == FP_NAN) {
884 
    return opb;
885 
  }
886 
 
887 
  if (fpclassify(opb) == FP_NAN) {
888 
    return opa;
889 
  }
890 
 
891 
  // RISC-V spec: -0 < +0
892 
  tmp_a.float_value = opa;
893 
  tmp_b.float_value = opb;
894 
  if (((tmp_a.binary_value == 0x80000000) && (tmp_b.binary_value == 0x00000000)) ||
895 
      ((tmp_a.binary_value == 0x00000000) && (tmp_b.binary_value == 0x80000000))) {
896 
    return -0.0f;
897 
  }
898 
 
899 
  return fmin(opa, opb);
900 
}
901 
 
902 
 
903 
/**********************************************************************//**
904 
 * Single-precision floating-point maximum
905 
 *
906 
 * @param[in] rs1 Source operand 1.
907 
 * @param[in] rs2 Source operand 2.
908 
 * @return Result.
909 
 **************************************************************************/
910 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fmaxs(float rs1, float rs2) {
911 55 zero_gravi 
 
912 
  float opa = subnormal_flush(rs1);
913 
  float opb = subnormal_flush(rs2);
914 
 
915 
  union {
916 
  uint32_t binary_value; /**< Access as native float */
917 
  float    float_value;  /**< Access in binary representation */
918 
  } tmp_a, tmp_b;
919 
 
920 
 
921 
  if ((fpclassify(opa) == FP_NAN) && (fpclassify(opb) == FP_NAN)) {
922 
    return nanf("");
923 
  }
924 
 
925 
  if (fpclassify(opa) == FP_NAN) {
926 
    return opb;
927 
  }
928 
 
929 
  if (fpclassify(opb) == FP_NAN) {
930 
    return opa;
931 
  }
932 
 
933 
  // RISC-V spec: -0 < +0
934 
  tmp_a.float_value = opa;
935 
  tmp_b.float_value = opb;
936 
  if (((tmp_a.binary_value == 0x80000000) && (tmp_b.binary_value == 0x00000000)) ||
937 
      ((tmp_a.binary_value == 0x00000000) && (tmp_b.binary_value == 0x80000000))) {
938 
    return +0.0f;
939 
  }
940 
 
941 
  return fmax(opa, opb);
942 
}
943 
 
944 
 
945 
/**********************************************************************//**
946 
 * Single-precision floating-point float to unsigned integer
947 
 *
948 
 * @param[in] rs1 Source operand 1.
949 
 * @return Result.
950 
 **************************************************************************/
951 56 zero_gravi 
uint32_t __attribute__ ((noinline)) riscv_emulate_fcvt_wus(float rs1) {
952 55 zero_gravi 
 
953 
  float opa = subnormal_flush(rs1);
954 
 
955 
  return (uint32_t)roundf(opa);
956 
}
957 
 
958 
 
959 
/**********************************************************************//**
960 
 * Single-precision floating-point float to signed integer
961 
 *
962 
 * @param[in] rs1 Source operand 1.
963 
 * @return Result.
964 
 **************************************************************************/
965 56 zero_gravi 
int32_t __attribute__ ((noinline)) riscv_emulate_fcvt_ws(float rs1) {
966 55 zero_gravi 
 
967 
  float opa = subnormal_flush(rs1);
968 
 
969 
  return (int32_t)roundf(opa);
970 
}
971 
 
972 
 
973 
/**********************************************************************//**
974 
 * Single-precision floating-point unsigned integer to float
975 
 *
976 
 * @param[in] rs1 Source operand 1.
977 
 * @return Result.
978 
 **************************************************************************/
979 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fcvt_swu(uint32_t rs1) {
980 55 zero_gravi 
 
981 
  return (float)rs1;
982 
}
983 
 
984 
 
985 
/**********************************************************************//**
986 
 * Single-precision floating-point signed integer to float
987 
 *
988 
 * @param[in] rs1 Source operand 1.
989 
 * @return Result.
990 
 **************************************************************************/
991 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fcvt_sw(int32_t rs1) {
992 55 zero_gravi 
 
993 
  return (float)rs1;
994 
}
995 
 
996 
 
997 
/**********************************************************************//**
998 
 * Single-precision floating-point equal comparison
999 
 *
1000 
 * @param[in] rs1 Source operand 1.
1001 
 * @param[in] rs2 Source operand 2.
1002 
 * @return Result.
1003 
 **************************************************************************/
1004 56 zero_gravi 
uint32_t __attribute__ ((noinline)) riscv_emulate_feqs(float rs1, float rs2) {
1005 55 zero_gravi 
 
1006 
  float opa = subnormal_flush(rs1);
1007 
  float opb = subnormal_flush(rs2);
1008 
 
1009 
  if ((fpclassify(opa) == FP_NAN) || (fpclassify(opb) == FP_NAN)) {
1010 
    return 0;
1011 
  }
1012 
 
1013 
  if isless(opa, opb) {
1014 
    return 0;
1015 
  }
1016 
  else if isgreater(opa, opb) {
1017 
    return 0;
1018 
  }
1019 
  else {
1020 
    return 1;
1021 
  }
1022 
}
1023 
 
1024 
 
1025 
/**********************************************************************//**
1026 
 * Single-precision floating-point less-than comparison
1027 
 *
1028 
 * @param[in] rs1 Source operand 1.
1029 
 * @param[in] rs2 Source operand 2.
1030 
 * @return Result.
1031 
 **************************************************************************/
1032 56 zero_gravi 
uint32_t __attribute__ ((noinline)) riscv_emulate_flts(float rs1, float rs2) {
1033 55 zero_gravi 
 
1034 
  float opa = subnormal_flush(rs1);
1035 
  float opb = subnormal_flush(rs2);
1036 
 
1037 
  if ((fpclassify(opa) == FP_NAN) || (fpclassify(opb) == FP_NAN)) {
1038 
    return 0;
1039 
  }
1040 
 
1041 
  if isless(opa, opb) {
1042 
    return 1;
1043 
  }
1044 
  else {
1045 
    return 0;
1046 
  }
1047 
}
1048 
 
1049 
 
1050 
/**********************************************************************//**
1051 
 * Single-precision floating-point less-than-or-equal comparison
1052 
 *
1053 
 * @param[in] rs1 Source operand 1.
1054 
 * @param[in] rs2 Source operand 2.
1055 
 * @return Result.
1056 
 **************************************************************************/
1057 56 zero_gravi 
uint32_t __attribute__ ((noinline)) riscv_emulate_fles(float rs1, float rs2) {
1058 55 zero_gravi 
 
1059 
  float opa = subnormal_flush(rs1);
1060 
  float opb = subnormal_flush(rs2);
1061 
 
1062 
  if ((fpclassify(opa) == FP_NAN) || (fpclassify(opb) == FP_NAN)) {
1063 
    return 0;
1064 
  }
1065 
 
1066 
  if islessequal(opa, opb) {
1067 
    return 1;
1068 
  }
1069 
  else {
1070 
    return 0;
1071 
  }
1072 
}
1073 
 
1074 
 
1075 
/**********************************************************************//**
1076 
 * Single-precision floating-point sign-injection
1077 
 *
1078 
 * @param[in] rs1 Source operand 1.
1079 
 * @param[in] rs2 Source operand 2.
1080 
 * @return Result.
1081 
 **************************************************************************/
1082 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fsgnjs(float rs1, float rs2) {
1083 55 zero_gravi 
 
1084 
  float opa = subnormal_flush(rs1);
1085 
  float opb = subnormal_flush(rs2);
1086 
 
1087 
  int sign_1 = (int)signbit(opa);
1088 
  int sign_2 = (int)signbit(opb);
1089 
  float res = 0;
1090 
 
1091 
  if (sign_2 != 0) { // opb is negative
1092 
    if (sign_1 == 0) {
1093 
      res = -opa;
1094 
    }
1095 
    else {
1096 
      res = opa;
1097 
    }
1098 
  }
1099 
  else { // opb is positive
1100 
    if (sign_1 == 0) {
1101 
      res = opa;
1102 
    }
1103 
    else {
1104 
      res = -opa;
1105 
    }
1106 
  }
1107 
 
1108 
  return res;
1109 
}
1110 
 
1111 
 
1112 
/**********************************************************************//**
1113 
 * Single-precision floating-point sign-injection NOT
1114 
 *
1115 
 * @param[in] rs1 Source operand 1.
1116 
 * @param[in] rs2 Source operand 2.
1117 
 * @return Result.
1118 
 **************************************************************************/
1119 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fsgnjns(float rs1, float rs2) {
1120 55 zero_gravi 
 
1121 
  float opa = subnormal_flush(rs1);
1122 
  float opb = subnormal_flush(rs2);
1123 
 
1124 
  int sign_1 = (int)signbit(opa);
1125 
  int sign_2 = (int)signbit(opb);
1126 
  float res = 0;
1127 
 
1128 
  if (sign_2 != 0) { // opb is negative
1129 
    if (sign_1 == 0) {
1130 
      res = opa;
1131 
    }
1132 
    else {
1133 
      res = -opa;
1134 
    }
1135 
  }
1136 
  else { // opb is positive
1137 
    if (sign_1 == 0) {
1138 
      res = -opa;
1139 
    }
1140 
    else {
1141 
      res = opa;
1142 
    }
1143 
  }
1144 
 
1145 
  return res;
1146 
}
1147 
 
1148 
 
1149 
/**********************************************************************//**
1150 
 * Single-precision floating-point sign-injection XOR
1151 
 *
1152 
 * @param[in] rs1 Source operand 1.
1153 
 * @param[in] rs2 Source operand 2.
1154 
 * @return Result.
1155 
 **************************************************************************/
1156 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fsgnjxs(float rs1, float rs2) {
1157 55 zero_gravi 
 
1158 
  float opa = subnormal_flush(rs1);
1159 
  float opb = subnormal_flush(rs2);
1160 
 
1161 
  int sign_1 = (int)signbit(opa);
1162 
  int sign_2 = (int)signbit(opb);
1163 
  float res = 0;
1164 
 
1165 
  if (((sign_1 == 0) && (sign_2 != 0)) || ((sign_1 != 0) && (sign_2 == 0))) {
1166 
    if (sign_1 == 0) {
1167 
      res = -opa;
1168 
    }
1169 
    else {
1170 
      res = opa;
1171 
    }
1172 
  }
1173 
  else {
1174 
    if (sign_1 == 0) {
1175 
      res = opa;
1176 
    }
1177 
    else {
1178 
      res = -opa;
1179 
    }
1180 
  }
1181 
 
1182 
  return res;
1183 
}
1184 
 
1185 
 
1186 
/**********************************************************************//**
1187 
 * Single-precision floating-point number classification
1188 
 *
1189 
 * @param[in] rs1 Source operand 1.
1190 
 * @return Result.
1191 
 **************************************************************************/
1192 56 zero_gravi 
uint32_t __attribute__ ((noinline)) riscv_emulate_fclasss(float rs1) {
1193 55 zero_gravi 
 
1194 
  float opa = subnormal_flush(rs1);
1195 
 
1196 
  union {
1197 
    uint32_t binary_value; /**< Access as native float */
1198 
    float    float_value;  /**< Access in binary representation */
1199 
  } aux;
1200 
 
1201 
  // RISC-V classify result layout
1202 
  const uint32_t CLASS_NEG_INF    = 1 << 0; // negative infinity
1203 
  const uint32_t CLASS_NEG_NORM   = 1 << 1; // negative normal number
1204 
  const uint32_t CLASS_NEG_DENORM = 1 << 2; // negative subnormal number
1205 
  const uint32_t CLASS_NEG_ZERO   = 1 << 3; // negative zero
1206 
  const uint32_t CLASS_POS_ZERO   = 1 << 4; // positive zero
1207 
  const uint32_t CLASS_POS_DENORM = 1 << 5; // positive subnormal number
1208 
  const uint32_t CLASS_POS_NORM   = 1 << 6; // positive normal number
1209 
  const uint32_t CLASS_POS_INF    = 1 << 7; // positive infinity
1210 
  const uint32_t CLASS_SNAN       = 1 << 8; // signaling NaN (sNaN)
1211 
  const uint32_t CLASS_QNAN       = 1 << 9; // quiet NaN (qNaN)
1212 
 
1213 
  int tmp = fpclassify(opa);
1214 
  int sgn = (int)signbit(opa);
1215 
 
1216 
  uint32_t res = 0;
1217 
 
1218 
  // infinity
1219 
  if (tmp == FP_INFINITE) {
1220 
    if (sgn) { res |= CLASS_NEG_INF; }
1221 
    else     { res |= CLASS_POS_INF; }
1222 
  }
1223 
 
1224 
  // zero
1225 
  if (tmp == FP_ZERO) {
1226 
    if (sgn) { res |= CLASS_NEG_ZERO; }
1227 
    else     { res |= CLASS_POS_ZERO; }
1228 
  }
1229 
 
1230 
  // normal
1231 
  if (tmp == FP_NORMAL) {
1232 
    if (sgn) { res |= CLASS_NEG_NORM; }
1233 
    else     { res |= CLASS_POS_NORM; }
1234 
  }
1235 
 
1236 
  // subnormal
1237 
  if (tmp == FP_SUBNORMAL) {
1238 
    if (sgn) { res |= CLASS_NEG_DENORM; }
1239 
    else     { res |= CLASS_POS_DENORM; }
1240 
  }
1241 
 
1242 
  // NaN
1243 
  if (tmp == FP_NAN) {
1244 
    aux.float_value = opa;
1245 
    if ((aux.binary_value >> 22) & 0b1) { // bit 22 (mantissa's MSB) is set -> canonical (quiet) NAN
1246 
      res |= CLASS_QNAN;
1247 
    }
1248 
    else {
1249 
      res |= CLASS_SNAN;
1250 
    }
1251 
  }
1252 
 
1253 
  return res;
1254 
}
1255 
 
1256 
 
1257 
/**********************************************************************//**
1258 
 * Single-precision floating-point division
1259 
 *
1260 
 * @param[in] rs1 Source operand 1.
1261 
 * @param[in] rs2 Source operand 2.
1262 
 * @return Result.
1263 
 **************************************************************************/
1264 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fdivs(float rs1, float rs2) {
1265 55 zero_gravi 
 
1266 
  float opa = subnormal_flush(rs1);
1267 
  float opb = subnormal_flush(rs2);
1268 
 
1269 
  float res = opa / opb;
1270 
  return subnormal_flush(res);
1271 
}
1272 
 
1273 
 
1274 
/**********************************************************************//**
1275 
 * Single-precision floating-point square root
1276 
 *
1277 
 * @param[in] rs1 Source operand 1.
1278 
 * @return Result.
1279 
 **************************************************************************/
1280 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fsqrts(float rs1) {
1281 55 zero_gravi 
 
1282 
  float opa = subnormal_flush(rs1);
1283 
 
1284 
  float res = sqrtf(opa);
1285 
  return subnormal_flush(res);
1286 
}
1287 
 
1288 
 
1289 
/**********************************************************************//**
1290 
 * Single-precision floating-point fused multiply-add
1291 
 *
1292 
 * @note "noinline" attributed to make sure arguments/return values are in a0 and a1.
1293 
 *
1294 
 * @warning This instruction is not supported!
1295 
 *
1296 
 * @param[in] rs1 Source operand 1
1297 
 * @param[in] rs2 Source operand 2
1298 
 * @param[in] rs3 Source operand 3
1299 
 * @return Result.
1300 
 **************************************************************************/
1301 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fmadds(float rs1, float rs2, float rs3) {
1302 55 zero_gravi 
 
1303 
  float opa = subnormal_flush(rs1);
1304 
  float opb = subnormal_flush(rs2);
1305 
  float opc = subnormal_flush(rs3);
1306 
 
1307 
  float res = (opa * opb) + opc;
1308 
  return subnormal_flush(res);
1309 
}
1310 
 
1311 
 
1312 
/**********************************************************************//**
1313 
 * Single-precision floating-point fused multiply-sub
1314 
 *
1315 
 * @param[in] rs1 Source operand 1
1316 
 * @param[in] rs2 Source operand 2
1317 
 * @param[in] rs3 Source operand 3
1318 
 * @return Result.
1319 
 **************************************************************************/
1320 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fmsubs(float rs1, float rs2, float rs3) {
1321 55 zero_gravi 
 
1322 
  float opa = subnormal_flush(rs1);
1323 
  float opb = subnormal_flush(rs2);
1324 
  float opc = subnormal_flush(rs3);
1325 
 
1326 
  float res = (opa * opb) - opc;
1327 
  return subnormal_flush(res);
1328 
}
1329 
 
1330 
 
1331 
/**********************************************************************//**
1332 
 * Single-precision floating-point fused negated multiply-sub
1333 
 *
1334 
 * @param[in] rs1 Source operand 1
1335 
 * @param[in] rs2 Source operand 2
1336 
 * @param[in] rs3 Source operand 3
1337 
 * @return Result.
1338 
 **************************************************************************/
1339 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fnmsubs(float rs1, float rs2, float rs3) {
1340 55 zero_gravi 
 
1341 
  float opa = subnormal_flush(rs1);
1342 
  float opb = subnormal_flush(rs2);
1343 
  float opc = subnormal_flush(rs3);
1344 
 
1345 
  float res = -(opa * opb) + opc;
1346 
  return subnormal_flush(res);
1347 
}
1348 
 
1349 
 
1350 
/**********************************************************************//**
1351 
 * Single-precision floating-point fused negated multiply-add
1352 
 *
1353 
 * @param[in] rs1 Source operand 1
1354 
 * @param[in] rs2 Source operand 2
1355 
 * @param[in] rs3 Source operand 3
1356 
 * @return Result.
1357 
 **************************************************************************/
1358 56 zero_gravi 
float __attribute__ ((noinline)) riscv_emulate_fnmadds(float rs1, float rs2, float rs3) {
1359 55 zero_gravi 
 
1360 
  float opa = subnormal_flush(rs1);
1361 
  float opb = subnormal_flush(rs2);
1362 
  float opc = subnormal_flush(rs3);
1363 
 
1364 
  float res = -(opa * opb) - opc;
1365 
  return subnormal_flush(res);
1366 
}
1367 
 
1368 
 
1369 
#endif // neorv32_zfinx_extension_intrinsics_h
1370 
 

powered by: WebSVN 2.1.0

© copyright 1999-2025 OpenCores.org, equivalent to Oliscience, all rights reserved. OpenCores®, registered trademark.