0001-softfloat-add-APIs-to-handle-alternative-sNaN-propag.patch 9.2 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214
  1. From 7616e27c73841ac1e8fa031d843bde1325307bab Mon Sep 17 00:00:00 2001
  2. From: Chih-Min Chao <chihmin.chao@sifive.com>
  3. Date: Thu, 30 Jul 2020 02:52:23 -0700
  4. Subject: [PATCH 001/107] softfloat: add APIs to handle alternative sNaN
  5. propagation
  6. For "fmax/fmin ft0, ft1, ft2" and if one of the inputs is sNaN,
  7. The original logic
  8. return NaN and set invalid flag if ft1 == sNaN || ft2 == sNan
  9. The alternative path
  10. set invalid flag if ft1 == sNaN || ft2 == sNaN
  11. return NaN if ft1 == sNaN && ft2 == sNaN
  12. The ieee754 spec allows both implementation and some architecture such
  13. as riscv choose differenct defintion in two spec versions.
  14. (riscv-spec-v2.2 use original version, riscv-spec-20191213 changes to
  15. alternative)
  16. Signed-off-by: Chih-Min Chao <chihmin.chao@sifive.com>
  17. Message-Id: <1596102747-20226-3-git-send-email-chihmin.chao@sifive.com>
  18. ---
  19. fpu/softfloat.c | 92 +++++++++++++++++++++++++----------------
  20. include/fpu/softfloat.h | 6 +++
  21. 2 files changed, 63 insertions(+), 35 deletions(-)
  22. diff --git a/fpu/softfloat.c b/fpu/softfloat.c
  23. index 67cfa0fd82..edaa760310 100644
  24. --- a/fpu/softfloat.c
  25. +++ b/fpu/softfloat.c
  26. @@ -894,11 +894,16 @@ static FloatParts return_nan(FloatParts a, float_status *s)
  27. return a;
  28. }
  29. -static FloatParts pick_nan(FloatParts a, FloatParts b, float_status *s)
  30. +static void set_snan_flag(FloatParts a, FloatParts b, float_status *s)
  31. {
  32. if (is_snan(a.cls) || is_snan(b.cls)) {
  33. s->float_exception_flags |= float_flag_invalid;
  34. }
  35. +}
  36. +
  37. +static FloatParts pick_nan(FloatParts a, FloatParts b, float_status *s)
  38. +{
  39. + set_snan_flag(a, b, s);
  40. if (s->default_nan_mode) {
  41. return parts_default_nan(s);
  42. @@ -3091,23 +3096,32 @@ bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status)
  43. * and minNumMag() from the IEEE-754 2008.
  44. */
  45. static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin,
  46. - bool ieee, bool ismag, float_status *s)
  47. + bool ieee, bool ismag, bool issnan_prop,
  48. + float_status *s)
  49. {
  50. if (unlikely(is_nan(a.cls) || is_nan(b.cls))) {
  51. if (ieee) {
  52. /* Takes two floating-point values `a' and `b', one of
  53. * which is a NaN, and returns the appropriate NaN
  54. * result. If either `a' or `b' is a signaling NaN,
  55. - * the invalid exception is raised.
  56. + * the invalid exception is raised but the NaN
  57. + * propagation is 'shall'.
  58. */
  59. if (is_snan(a.cls) || is_snan(b.cls)) {
  60. - return pick_nan(a, b, s);
  61. - } else if (is_nan(a.cls) && !is_nan(b.cls)) {
  62. + if (issnan_prop) {
  63. + return pick_nan(a, b, s);
  64. + } else {
  65. + set_snan_flag(a, b, s);
  66. + }
  67. + }
  68. +
  69. + if (is_nan(a.cls) && !is_nan(b.cls)) {
  70. return b;
  71. } else if (is_nan(b.cls) && !is_nan(a.cls)) {
  72. return a;
  73. }
  74. }
  75. +
  76. return pick_nan(a, b, s);
  77. } else {
  78. int a_exp, b_exp;
  79. @@ -3161,56 +3175,64 @@ static FloatParts minmax_floats(FloatParts a, FloatParts b, bool ismin,
  80. }
  81. }
  82. -#define MINMAX(sz, name, ismin, isiee, ismag) \
  83. +#define MINMAX(sz, name, ismin, isiee, ismag, issnan_prop) \
  84. float ## sz float ## sz ## _ ## name(float ## sz a, float ## sz b, \
  85. float_status *s) \
  86. { \
  87. FloatParts pa = float ## sz ## _unpack_canonical(a, s); \
  88. FloatParts pb = float ## sz ## _unpack_canonical(b, s); \
  89. - FloatParts pr = minmax_floats(pa, pb, ismin, isiee, ismag, s); \
  90. + FloatParts pr = minmax_floats(pa, pb, ismin, isiee, ismag, \
  91. + issnan_prop, s); \
  92. \
  93. return float ## sz ## _round_pack_canonical(pr, s); \
  94. }
  95. -MINMAX(16, min, true, false, false)
  96. -MINMAX(16, minnum, true, true, false)
  97. -MINMAX(16, minnummag, true, true, true)
  98. -MINMAX(16, max, false, false, false)
  99. -MINMAX(16, maxnum, false, true, false)
  100. -MINMAX(16, maxnummag, false, true, true)
  101. -
  102. -MINMAX(32, min, true, false, false)
  103. -MINMAX(32, minnum, true, true, false)
  104. -MINMAX(32, minnummag, true, true, true)
  105. -MINMAX(32, max, false, false, false)
  106. -MINMAX(32, maxnum, false, true, false)
  107. -MINMAX(32, maxnummag, false, true, true)
  108. -
  109. -MINMAX(64, min, true, false, false)
  110. -MINMAX(64, minnum, true, true, false)
  111. -MINMAX(64, minnummag, true, true, true)
  112. -MINMAX(64, max, false, false, false)
  113. -MINMAX(64, maxnum, false, true, false)
  114. -MINMAX(64, maxnummag, false, true, true)
  115. +MINMAX(16, min, true, false, false, true)
  116. +MINMAX(16, minnum, true, true, false, true)
  117. +MINMAX(16, minnum_noprop, true, true, false, false)
  118. +MINMAX(16, minnummag, true, true, true, true)
  119. +MINMAX(16, max, false, false, false, true)
  120. +MINMAX(16, maxnum, false, true, false, true)
  121. +MINMAX(16, maxnum_noprop, false, true, false, false)
  122. +MINMAX(16, maxnummag, false, true, true, true)
  123. +
  124. +MINMAX(32, min, true, false, false, true)
  125. +MINMAX(32, minnum, true, true, false, true)
  126. +MINMAX(32, minnum_noprop, true, true, false, false)
  127. +MINMAX(32, minnummag, true, true, true, true)
  128. +MINMAX(32, max, false, false, false, true)
  129. +MINMAX(32, maxnum, false, true, false, true)
  130. +MINMAX(32, maxnum_noprop, false, true, false, false)
  131. +MINMAX(32, maxnummag, false, true, true, true)
  132. +
  133. +MINMAX(64, min, true, false, false, true)
  134. +MINMAX(64, minnum, true, true, false, true)
  135. +MINMAX(64, minnum_noprop, true, true, false, false)
  136. +MINMAX(64, minnummag, true, true, true, true)
  137. +MINMAX(64, max, false, false, false, true)
  138. +MINMAX(64, maxnum, false, true, false, true)
  139. +MINMAX(64, maxnum_noprop, false, true, false, false)
  140. +MINMAX(64, maxnummag, false, true, true, true)
  141. #undef MINMAX
  142. -#define BF16_MINMAX(name, ismin, isiee, ismag) \
  143. +#define BF16_MINMAX(name, ismin, isiee, ismag, issnan_prop) \
  144. bfloat16 bfloat16_ ## name(bfloat16 a, bfloat16 b, float_status *s) \
  145. { \
  146. FloatParts pa = bfloat16_unpack_canonical(a, s); \
  147. FloatParts pb = bfloat16_unpack_canonical(b, s); \
  148. - FloatParts pr = minmax_floats(pa, pb, ismin, isiee, ismag, s); \
  149. + FloatParts pr = minmax_floats(pa, pb, ismin, isiee, ismag, \
  150. + issnan_prop, s); \
  151. \
  152. return bfloat16_round_pack_canonical(pr, s); \
  153. }
  154. -BF16_MINMAX(min, true, false, false)
  155. -BF16_MINMAX(minnum, true, true, false)
  156. -BF16_MINMAX(minnummag, true, true, true)
  157. -BF16_MINMAX(max, false, false, false)
  158. -BF16_MINMAX(maxnum, false, true, false)
  159. -BF16_MINMAX(maxnummag, false, true, true)
  160. +BF16_MINMAX(min, true, false, false, true)
  161. +BF16_MINMAX(minnum, true, true, false, true)
  162. +BF16_MINMAX(minnummag, true, true, true, true)
  163. +BF16_MINMAX(max, false, false, false, true)
  164. +BF16_MINMAX(maxnum, false, true, false, true)
  165. +BF16_MINMAX(maxnummag, false, true, true, true)
  166. #undef BF16_MINMAX
  167. diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h
  168. index 78ad5ca738..5cd7d44bc9 100644
  169. --- a/include/fpu/softfloat.h
  170. +++ b/include/fpu/softfloat.h
  171. @@ -240,6 +240,8 @@ float16 float16_minnum(float16, float16, float_status *status);
  172. float16 float16_maxnum(float16, float16, float_status *status);
  173. float16 float16_minnummag(float16, float16, float_status *status);
  174. float16 float16_maxnummag(float16, float16, float_status *status);
  175. +float16 float16_minnum_noprop(float16, float16, float_status *status);
  176. +float16 float16_maxnum_noprop(float16, float16, float_status *status);
  177. float16 float16_sqrt(float16, float_status *status);
  178. FloatRelation float16_compare(float16, float16, float_status *status);
  179. FloatRelation float16_compare_quiet(float16, float16, float_status *status);
  180. @@ -586,6 +588,8 @@ float32 float32_minnum(float32, float32, float_status *status);
  181. float32 float32_maxnum(float32, float32, float_status *status);
  182. float32 float32_minnummag(float32, float32, float_status *status);
  183. float32 float32_maxnummag(float32, float32, float_status *status);
  184. +float32 float32_minnum_noprop(float32, float32, float_status *status);
  185. +float32 float32_maxnum_noprop(float32, float32, float_status *status);
  186. bool float32_is_quiet_nan(float32, float_status *status);
  187. bool float32_is_signaling_nan(float32, float_status *status);
  188. float32 float32_silence_nan(float32, float_status *status);
  189. @@ -775,6 +779,8 @@ float64 float64_minnum(float64, float64, float_status *status);
  190. float64 float64_maxnum(float64, float64, float_status *status);
  191. float64 float64_minnummag(float64, float64, float_status *status);
  192. float64 float64_maxnummag(float64, float64, float_status *status);
  193. +float64 float64_minnum_noprop(float64, float64, float_status *status);
  194. +float64 float64_maxnum_noprop(float64, float64, float_status *status);
  195. bool float64_is_quiet_nan(float64 a, float_status *status);
  196. bool float64_is_signaling_nan(float64, float_status *status);
  197. float64 float64_silence_nan(float64, float_status *status);
  198. --
  199. 2.33.1