diff options
Diffstat (limited to 'src')
-rw-r--r-- | src/H5T.c | 190 | ||||
-rw-r--r-- | src/H5Tconv.c | 377 | ||||
-rw-r--r-- | src/H5Tinit_float.c | 48 | ||||
-rw-r--r-- | src/H5Tmodule.h | 15 | ||||
-rw-r--r-- | src/H5Tnative.c | 28 | ||||
-rw-r--r-- | src/H5Tpkg.h | 60 | ||||
-rw-r--r-- | src/H5Tpublic.h | 19 | ||||
-rw-r--r-- | src/H5private.h | 20 | ||||
-rw-r--r-- | src/H5trace.c | 8 |
9 files changed, 721 insertions, 44 deletions
@@ -100,6 +100,30 @@ dt->shared->u.atomic.msb_pad = H5T_PAD_ZERO; \ } +/* Define the code templates for standard 16-bit floats for the "GUTS" in the H5T_INIT_TYPE macro */ +#define H5T_INIT_TYPE_FLOAT16_COMMON(ENDIANNESS) \ + { \ + H5T_INIT_TYPE_NUM_COMMON(ENDIANNESS) \ + dt->shared->u.atomic.u.f.sign = 15; \ + dt->shared->u.atomic.u.f.epos = 10; \ + dt->shared->u.atomic.u.f.esize = 5; \ + dt->shared->u.atomic.u.f.ebias = 0xf; \ + dt->shared->u.atomic.u.f.mpos = 0; \ + dt->shared->u.atomic.u.f.msize = 10; \ + dt->shared->u.atomic.u.f.norm = H5T_NORM_IMPLIED; \ + dt->shared->u.atomic.u.f.pad = H5T_PAD_ZERO; \ + } + +#define H5T_INIT_TYPE_FLOAT16LE_CORE \ + { \ + H5T_INIT_TYPE_FLOAT16_COMMON(H5T_ORDER_LE) \ + } + +#define H5T_INIT_TYPE_FLOAT16BE_CORE \ + { \ + H5T_INIT_TYPE_FLOAT16_COMMON(H5T_ORDER_BE) \ + } + /* Define the code templates for standard floats for the "GUTS" in the H5T_INIT_TYPE macro */ #define H5T_INIT_TYPE_FLOAT_COMMON(ENDIANNESS) \ { \ @@ -374,6 +398,8 @@ H5T_order_t H5T_native_order_g = H5T_ORDER_ERROR; * If more of these are added, the new ones must be added to the list of * types to reset in H5T_term_package(). */ +hid_t H5T_IEEE_F16BE_g = FAIL; +hid_t H5T_IEEE_F16LE_g = FAIL; hid_t H5T_IEEE_F32BE_g = FAIL; hid_t H5T_IEEE_F32LE_g = FAIL; hid_t H5T_IEEE_F64BE_g = FAIL; @@ -429,6 +455,7 @@ hid_t H5T_NATIVE_LONG_g = FAIL; hid_t H5T_NATIVE_ULONG_g = FAIL; hid_t H5T_NATIVE_LLONG_g = FAIL; hid_t H5T_NATIVE_ULLONG_g = FAIL; +hid_t H5T_NATIVE_FLOAT16_g = FAIL; hid_t H5T_NATIVE_FLOAT_g = FAIL; hid_t H5T_NATIVE_DOUBLE_g = FAIL; hid_t H5T_NATIVE_LDOUBLE_g = FAIL; @@ -497,6 +524,7 @@ size_t H5T_NATIVE_LONG_ALIGN_g = 0; size_t H5T_NATIVE_ULONG_ALIGN_g = 0; size_t H5T_NATIVE_LLONG_ALIGN_g = 0; size_t H5T_NATIVE_ULLONG_ALIGN_g = 0; +size_t H5T_NATIVE_FLOAT16_ALIGN_g = 0; size_t H5T_NATIVE_FLOAT_ALIGN_g = 0; size_t H5T_NATIVE_DOUBLE_ALIGN_g = 0; size_t H5T_NATIVE_LDOUBLE_ALIGN_g = 0; @@ -532,6 +560,15 @@ size_t H5T_NATIVE_UINT_FAST64_ALIGN_g = 0; /* Useful floating-point values for conversion routines */ /* (+/- Inf for all floating-point types) */ +#ifdef H5_HAVE__FLOAT16 +/* Initialize these with a float literal since the f16 suffix + * is non-standard C and gives warnings when compiling the + * library with the -pedantic flag. These values will be + * overwritten anyway. + */ +H5__Float16 H5T_NATIVE_FLOAT16_POS_INF_g = 0.0f; +H5__Float16 H5T_NATIVE_FLOAT16_NEG_INF_g = 0.0f; +#endif float H5T_NATIVE_FLOAT_POS_INF_g = 0.0F; float H5T_NATIVE_FLOAT_NEG_INF_g = 0.0F; double H5T_NATIVE_DOUBLE_POS_INF_g = 0.0; @@ -684,6 +721,49 @@ H5T__init_inf(void) } /* end for */ } /* end if */ +#ifdef H5_HAVE__FLOAT16 + /* Get the _Float16 datatype */ + if (NULL == (dst_p = (H5T_t *)H5I_object(H5T_NATIVE_FLOAT16_g))) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a datatype"); + dst = &dst_p->shared->u.atomic; + + /* Check that we can re-order the bytes correctly */ + if (H5T_ORDER_LE != H5T_native_order_g && H5T_ORDER_BE != H5T_native_order_g) + HGOTO_ERROR(H5E_DATATYPE, H5E_UNSUPPORTED, FAIL, "unsupported byte order"); + + /* +Inf */ + d = (uint8_t *)&H5T_NATIVE_FLOAT16_POS_INF_g; + H5T__bit_set(d, dst->u.f.sign, (size_t)1, false); + H5T__bit_set(d, dst->u.f.epos, dst->u.f.esize, true); + H5T__bit_set(d, dst->u.f.mpos, dst->u.f.msize, false); + + /* Swap the bytes if the machine architecture is big-endian */ + if (H5T_ORDER_BE == H5T_native_order_g) { + half_size = dst_p->shared->size / 2; + for (u = 0; u < half_size; u++) { + uint8_t tmp = d[dst_p->shared->size - (u + 1)]; + d[dst_p->shared->size - (u + 1)] = d[u]; + d[u] = tmp; + } /* end for */ + } /* end if */ + + /* -Inf */ + d = (uint8_t *)&H5T_NATIVE_FLOAT16_NEG_INF_g; + H5T__bit_set(d, dst->u.f.sign, (size_t)1, true); + H5T__bit_set(d, dst->u.f.epos, dst->u.f.esize, true); + H5T__bit_set(d, dst->u.f.mpos, dst->u.f.msize, false); + + /* Swap the bytes if the machine architecture is big-endian */ + if (H5T_ORDER_BE == H5T_native_order_g) { + half_size = dst_p->shared->size / 2; + for (u = 0; u < half_size; u++) { + uint8_t tmp = d[dst_p->shared->size - (u + 1)]; + d[dst_p->shared->size - (u + 1)] = d[u]; + d[u] = tmp; + } /* end for */ + } /* end if */ +#endif + done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5T__init_inf() */ @@ -737,6 +817,9 @@ H5T_init(void) herr_t status; bool copied_dtype = true; /* Flag to indicate whether datatype was copied or allocated (for error cleanup) */ +#ifdef H5_HAVE__FLOAT16 + H5T_t *native_float16 = NULL; /* Datatype structure for native _Float16 type */ +#endif herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_NOAPI(FAIL) @@ -778,6 +861,10 @@ H5T_init(void) HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a datatype object"); if (NULL == (native_ullong = (H5T_t *)H5I_object(H5T_NATIVE_ULLONG_g))) HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a datatype object"); +#ifdef H5_HAVE__FLOAT16 + if (NULL == (native_float16 = (H5T_t *)H5I_object(H5T_NATIVE_FLOAT16_g))) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a datatype object"); +#endif if (NULL == (native_float = (H5T_t *)H5I_object(H5T_NATIVE_FLOAT_g))) HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a datatype object"); if (NULL == (native_double = (H5T_t *)H5I_object(H5T_NATIVE_DOUBLE_g))) @@ -822,6 +909,12 @@ H5T_init(void) *------------------------------------------------------------ */ + /* IEEE 2-byte little-endian float */ + H5T_INIT_TYPE(FLOAT16LE, H5T_IEEE_F16LE_g, COPY, native_double, SET, 2) + + /* IEEE 2-byte big-endian float */ + H5T_INIT_TYPE(FLOAT16BE, H5T_IEEE_F16BE_g, COPY, native_double, SET, 2) + /* IEEE 4-byte little-endian float */ H5T_INIT_TYPE(FLOATLE, H5T_IEEE_F32LE_g, COPY, native_double, SET, 4) @@ -1061,6 +1154,20 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "ldbl_flt", native_ldouble, native_float, H5T__conv_ldouble_float); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_dbl", native_ldouble, native_double, H5T__conv_ldouble_double); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "flt16_flt", native_float16, native_float, H5T__conv__Float16_float); + status |= H5T__register_int(H5T_PERS_HARD, "flt16_dbl", native_float16, native_double, + H5T__conv__Float16_double); + status |= H5T__register_int(H5T_PERS_HARD, "flt16_ldbl", native_float16, native_ldouble, + H5T__conv__Float16_ldouble); + status |= + H5T__register_int(H5T_PERS_HARD, "flt_flt16", native_float, native_float16, H5T__conv_float__Float16); + status |= H5T__register_int(H5T_PERS_HARD, "dbl_flt16", native_double, native_float16, + H5T__conv_double__Float16); + status |= H5T__register_int(H5T_PERS_HARD, "ldbl_flt16", native_ldouble, native_float16, + H5T__conv_ldouble__Float16); +#endif /* from long long */ status |= @@ -1219,6 +1326,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "schar_dbl", native_schar, native_double, H5T__conv_schar_double); status |= H5T__register_int(H5T_PERS_HARD, "schar_ldbl", native_schar, native_ldouble, H5T__conv_schar_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "schar_flt16", native_schar, native_float16, + H5T__conv_schar__Float16); +#endif /* From unsigned char to floats */ status |= @@ -1227,6 +1338,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "uchar_dbl", native_uchar, native_double, H5T__conv_uchar_double); status |= H5T__register_int(H5T_PERS_HARD, "uchar_ldbl", native_uchar, native_ldouble, H5T__conv_uchar_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "uchar_flt16", native_uchar, native_float16, + H5T__conv_uchar__Float16); +#endif /* From short to floats */ status |= @@ -1235,6 +1350,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "short_dbl", native_short, native_double, H5T__conv_short_double); status |= H5T__register_int(H5T_PERS_HARD, "short_ldbl", native_short, native_ldouble, H5T__conv_short_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "short_flt16", native_short, native_float16, + H5T__conv_short__Float16); +#endif /* From unsigned short to floats */ status |= @@ -1243,23 +1362,39 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "ushort_dbl", native_ushort, native_double, H5T__conv_ushort_double); status |= H5T__register_int(H5T_PERS_HARD, "ushort_ldbl", native_ushort, native_ldouble, H5T__conv_ushort_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "ushort_flt16", native_ushort, native_float16, + H5T__conv_ushort__Float16); +#endif /* From int to floats */ status |= H5T__register_int(H5T_PERS_HARD, "int_flt", native_int, native_float, H5T__conv_int_float); status |= H5T__register_int(H5T_PERS_HARD, "int_dbl", native_int, native_double, H5T__conv_int_double); status |= H5T__register_int(H5T_PERS_HARD, "int_ldbl", native_int, native_ldouble, H5T__conv_int_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "int_flt16", native_int, native_float16, H5T__conv_int__Float16); +#endif /* From unsigned int to floats */ status |= H5T__register_int(H5T_PERS_HARD, "uint_flt", native_uint, native_float, H5T__conv_uint_float); status |= H5T__register_int(H5T_PERS_HARD, "uint_dbl", native_uint, native_double, H5T__conv_uint_double); status |= H5T__register_int(H5T_PERS_HARD, "uint_ldbl", native_uint, native_ldouble, H5T__conv_uint_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "uint_flt16", native_uint, native_float16, H5T__conv_uint__Float16); +#endif /* From long to floats */ status |= H5T__register_int(H5T_PERS_HARD, "long_flt", native_long, native_float, H5T__conv_long_float); status |= H5T__register_int(H5T_PERS_HARD, "long_dbl", native_long, native_double, H5T__conv_long_double); status |= H5T__register_int(H5T_PERS_HARD, "long_ldbl", native_long, native_ldouble, H5T__conv_long_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "long_flt16", native_long, native_float16, H5T__conv_long__Float16); +#endif /* From unsigned long to floats */ status |= @@ -1268,6 +1403,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "ulong_dbl", native_ulong, native_double, H5T__conv_ulong_double); status |= H5T__register_int(H5T_PERS_HARD, "ulong_ldbl", native_ulong, native_ldouble, H5T__conv_ulong_ldouble); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "ulong_flt16", native_ulong, native_float16, + H5T__conv_ulong__Float16); +#endif /* From long long to floats */ status |= @@ -1278,6 +1417,10 @@ H5T_init(void) status |= H5T__register_int(H5T_PERS_HARD, "llong_ldbl", native_llong, native_ldouble, H5T__conv_llong_ldouble); #endif /* H5T_CONV_INTERNAL_LLONG_LDOUBLE */ +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "llong_flt16", native_llong, native_float16, + H5T__conv_llong__Float16); +#endif /* From unsigned long long to floats */ status |= @@ -1288,6 +1431,10 @@ H5T_init(void) status |= H5T__register_int(H5T_PERS_HARD, "ullong_ldbl", native_ullong, native_ldouble, H5T__conv_ullong_ldouble); #endif /* H5T_CONV_INTERNAL_ULLONG_LDOUBLE */ +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "ullong_flt16", native_ullong, native_float16, + H5T__conv_ullong__Float16); +#endif /* From floats to char */ status |= @@ -1296,6 +1443,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "dbl_schar", native_double, native_schar, H5T__conv_double_schar); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_schar", native_ldouble, native_schar, H5T__conv_ldouble_schar); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_schar", native_float16, native_schar, + H5T__conv__Float16_schar); +#endif /* From floats to unsigned char */ status |= @@ -1304,6 +1455,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "dbl_uchar", native_double, native_uchar, H5T__conv_double_uchar); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_uchar", native_ldouble, native_uchar, H5T__conv_ldouble_uchar); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_uchar", native_float16, native_uchar, + H5T__conv__Float16_uchar); +#endif /* From floats to short */ status |= @@ -1312,6 +1467,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "dbl_short", native_double, native_short, H5T__conv_double_short); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_short", native_ldouble, native_short, H5T__conv_ldouble_short); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_short", native_float16, native_short, + H5T__conv__Float16_short); +#endif /* From floats to unsigned short */ status |= @@ -1320,23 +1479,39 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "dbl_ushort", native_double, native_ushort, H5T__conv_double_ushort); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_ushort", native_ldouble, native_ushort, H5T__conv_ldouble_ushort); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_ushort", native_float16, native_ushort, + H5T__conv__Float16_ushort); +#endif /* From floats to int */ status |= H5T__register_int(H5T_PERS_HARD, "flt_int", native_float, native_int, H5T__conv_float_int); status |= H5T__register_int(H5T_PERS_HARD, "dbl_int", native_double, native_int, H5T__conv_double_int); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_int", native_ldouble, native_int, H5T__conv_ldouble_int); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "flt16_int", native_float16, native_int, H5T__conv__Float16_int); +#endif /* From floats to unsigned int */ status |= H5T__register_int(H5T_PERS_HARD, "flt_uint", native_float, native_uint, H5T__conv_float_uint); status |= H5T__register_int(H5T_PERS_HARD, "dbl_uint", native_double, native_uint, H5T__conv_double_uint); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_uint", native_ldouble, native_uint, H5T__conv_ldouble_uint); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "flt16_uint", native_float16, native_uint, H5T__conv__Float16_uint); +#endif /* From floats to long */ status |= H5T__register_int(H5T_PERS_HARD, "flt_long", native_float, native_long, H5T__conv_float_long); status |= H5T__register_int(H5T_PERS_HARD, "dbl_long", native_double, native_long, H5T__conv_double_long); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_long", native_ldouble, native_long, H5T__conv_ldouble_long); +#ifdef H5_HAVE__FLOAT16 + status |= + H5T__register_int(H5T_PERS_HARD, "flt16_long", native_float16, native_long, H5T__conv__Float16_long); +#endif /* From floats to unsigned long */ status |= @@ -1345,6 +1520,10 @@ H5T_init(void) H5T__register_int(H5T_PERS_HARD, "dbl_ulong", native_double, native_ulong, H5T__conv_double_ulong); status |= H5T__register_int(H5T_PERS_HARD, "ldbl_ulong", native_ldouble, native_ulong, H5T__conv_ldouble_ulong); +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_ulong", native_float16, native_ulong, + H5T__conv__Float16_ulong); +#endif /* From floats to long long */ status |= @@ -1355,6 +1534,10 @@ H5T_init(void) status |= H5T__register_int(H5T_PERS_HARD, "ldbl_llong", native_ldouble, native_llong, H5T__conv_ldouble_llong); #endif /* H5T_CONV_INTERNAL_LDOUBLE_LLONG */ +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_llong", native_float16, native_llong, + H5T__conv__Float16_llong); +#endif /* From floats to unsigned long long */ status |= @@ -1365,6 +1548,10 @@ H5T_init(void) status |= H5T__register_int(H5T_PERS_HARD, "ldbl_ullong", native_ldouble, native_ullong, H5T__conv_ldouble_ullong); #endif /* H5T_CONV_INTERNAL_LDOUBLE_ULLONG */ +#ifdef H5_HAVE__FLOAT16 + status |= H5T__register_int(H5T_PERS_HARD, "flt16_ullong", native_float16, native_ullong, + H5T__conv__Float16_ullong); +#endif /* * The special no-op conversion is the fastest, so we list it last. The @@ -1546,6 +1733,8 @@ H5T_top_term_package(void) /* Reset all the datatype IDs */ if (H5T_IEEE_F32BE_g > 0) { + H5T_IEEE_F16BE_g = FAIL; + H5T_IEEE_F16LE_g = FAIL; H5T_IEEE_F32BE_g = FAIL; H5T_IEEE_F32LE_g = FAIL; H5T_IEEE_F64BE_g = FAIL; @@ -1598,6 +1787,7 @@ H5T_top_term_package(void) H5T_NATIVE_ULONG_g = FAIL; H5T_NATIVE_LLONG_g = FAIL; H5T_NATIVE_ULLONG_g = FAIL; + H5T_NATIVE_FLOAT16_g = FAIL; H5T_NATIVE_FLOAT_g = FAIL; H5T_NATIVE_DOUBLE_g = FAIL; H5T_NATIVE_LDOUBLE_g = FAIL; diff --git a/src/H5Tconv.c b/src/H5Tconv.c index 72debe8..33852a9 100644 --- a/src/H5Tconv.c +++ b/src/H5Tconv.c @@ -113,6 +113,14 @@ * at least as wide as the destination. Overflow can occur * when the source magnitude is too large for the destination. * + * fX: Floating-point values to integers where the destination is at least + * as wide as the source. This case cannot generate overflows. + * + * Xf: Integers to floating-point values where the source is at least as + * wide as the destination. Overflows can occur when the destination is + * narrower than the source. + * + * * The macros take a subset of these arguments in the order listed here: * * CDATA: A pointer to the H5T_cdata_t structure that was passed to the @@ -690,6 +698,72 @@ H5T_CONV(H5T_CONV_Fx, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, Y) \ } while (0) +#define H5T_CONV_fX(STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ + do { \ + HDcompile_assert(sizeof(ST) <= sizeof(DT)); \ + H5T_CONV(H5T_CONV_xX, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, N) \ + } while (0) + +#define H5T_CONV_Xf_CORE(STYPE, DTYPE, S, D, ST, DT, D_MIN, D_MAX) \ + { \ + if (*(S) > (ST)(D_MAX) || (sprec < dprec && *(S) == (ST)(D_MAX))) { \ + H5T_conv_ret_t except_ret = \ + (cb_struct.func)(H5T_CONV_EXCEPT_RANGE_HI, src_id, dst_id, S, D, cb_struct.user_data); \ + if (except_ret == H5T_CONV_UNHANDLED) \ + /* Let compiler convert if case is ignored by user handler*/ \ + *(D) = H5_GLUE3(H5T_NATIVE_, DTYPE, _POS_INF_g); \ + else if (except_ret == H5T_CONV_ABORT) \ + HGOTO_ERROR(H5E_DATATYPE, H5E_CANTCONVERT, FAIL, "can't handle conversion exception"); \ + /* if(except_ret==H5T_CONV_HANDLED): Fall through, user handled it */ \ + } \ + else if (*(S) < (ST)(D_MIN)) { \ + H5T_conv_ret_t except_ret = \ + (cb_struct.func)(H5T_CONV_EXCEPT_RANGE_LOW, src_id, dst_id, S, D, cb_struct.user_data); \ + if (except_ret == H5T_CONV_UNHANDLED) \ + /* Let compiler convert if case is ignored by user handler*/ \ + *(D) = H5_GLUE3(H5T_NATIVE_, DTYPE, _NEG_INF_g); \ + else if (except_ret == H5T_CONV_ABORT) \ + HGOTO_ERROR(H5E_DATATYPE, H5E_CANTCONVERT, FAIL, "can't handle conversion exception"); \ + /* if(except_ret==H5T_CONV_HANDLED): Fall through, user handled it */ \ + } \ + else if (sprec > dprec) { \ + unsigned low_bit_pos, high_bit_pos; \ + \ + /* Detect high & low bits set in source */ \ + H5T_HI_LO_BIT_SET(ST, *(S), low_bit_pos, high_bit_pos) \ + \ + /* Check for more bits of precision in src than available in dst */ \ + if ((high_bit_pos - low_bit_pos) >= dprec) { \ + H5T_conv_ret_t except_ret = \ + (cb_struct.func)(H5T_CONV_EXCEPT_PRECISION, src_id, dst_id, S, D, cb_struct.user_data); \ + if (except_ret == H5T_CONV_UNHANDLED) \ + /* Let compiler convert if case is ignored by user handler*/ \ + *(D) = (DT)(*(S)); \ + else if (except_ret == H5T_CONV_ABORT) \ + HGOTO_ERROR(H5E_DATATYPE, H5E_CANTCONVERT, FAIL, "can't handle conversion exception"); \ + /* if(except_ret==H5T_CONV_HANDLED): Fall through, user handled it */ \ + } \ + else \ + *(D) = (DT)(*(S)); \ + } \ + else \ + *(D) = (DT)(*(S)); \ + } +#define H5T_CONV_Xf_NOEX_CORE(STYPE, DTYPE, S, D, ST, DT, D_MIN, D_MAX) \ + { \ + if (*(S) > (ST)(D_MAX)) \ + *(D) = H5_GLUE3(H5T_NATIVE_, DTYPE, _POS_INF_g); \ + else if (*(S) < (ST)(D_MIN)) \ + *(D) = H5_GLUE3(H5T_NATIVE_, DTYPE, _NEG_INF_g); \ + else \ + *(D) = (DT)(*(S)); \ + } + +#define H5T_CONV_Xf(STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ + do { \ + H5T_CONV(H5T_CONV_Xf, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, Y) \ + } while (0) + /* Since all "no exception" cores do the same thing (assign the value in the * source location to the destination location, using casting), use one "core" * to do them all. @@ -7874,6 +7948,227 @@ H5T__conv_ldouble_ullong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t } #endif /*H5T_CONV_INTERNAL_LDOUBLE_ULLONG*/ +/* Conversions for _Float16 type */ +#ifdef H5_HAVE__FLOAT16 +herr_t +H5T__conv_schar__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_xF(SCHAR, FLOAT16, signed char, H5__Float16, -, -); +} + +herr_t +H5T__conv_uchar__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_xF(UCHAR, FLOAT16, unsigned char, H5__Float16, -, -); +} + +herr_t +H5T__conv_short__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_xF(SHORT, FLOAT16, short, H5__Float16, -, -); +} + +herr_t +H5T__conv_ushort__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(USHORT, FLOAT16, unsigned short, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_int__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(INT, FLOAT16, int, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_uint__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(UINT, FLOAT16, unsigned int, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_long__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(LONG, FLOAT16, long, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_ulong__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(ULONG, FLOAT16, unsigned long, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_llong__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(LLONG, FLOAT16, long long, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_ullong__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Xf(ULLONG, FLOAT16, unsigned long long, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_float__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Ff(FLOAT, FLOAT16, float, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_double__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Ff(DOUBLE, FLOAT16, double, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv_ldouble__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + /* Suppress warning about non-standard floating-point literal suffix */ + H5_GCC_CLANG_DIAG_OFF("pedantic") + H5T_CONV_Ff(LDOUBLE, FLOAT16, long double, H5__Float16, -FLT16_MAX, FLT16_MAX); + H5_GCC_CLANG_DIAG_ON("pedantic") +} + +herr_t +H5T__conv__Float16_schar(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5_GCC_CLANG_DIAG_OFF("float-equal") + H5T_CONV_Fx(FLOAT16, SCHAR, H5__Float16, signed char, SCHAR_MIN, SCHAR_MAX); + H5_GCC_CLANG_DIAG_ON("float-equal") +} + +herr_t +H5T__conv__Float16_uchar(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5_GCC_CLANG_DIAG_OFF("float-equal") + H5T_CONV_Fx(FLOAT16, UCHAR, H5__Float16, unsigned char, 0, UCHAR_MAX); + H5_GCC_CLANG_DIAG_ON("float-equal") +} + +herr_t +H5T__conv__Float16_short(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5_GCC_CLANG_DIAG_OFF("float-equal") + H5T_CONV_Fx(FLOAT16, SHORT, H5__Float16, short, SHRT_MIN, SHRT_MAX); + H5_GCC_CLANG_DIAG_ON("float-equal") +} + +herr_t +H5T__conv__Float16_ushort(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, USHORT, H5__Float16, unsigned short, 0, USHRT_MAX); +} + +herr_t +H5T__conv__Float16_int(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, INT, H5__Float16, int, INT_MIN, INT_MAX); +} + +herr_t +H5T__conv__Float16_uint(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, UINT, H5__Float16, unsigned int, 0, UINT_MAX); +} + +herr_t +H5T__conv__Float16_long(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, LONG, H5__Float16, long, LONG_MIN, LONG_MAX); +} + +herr_t +H5T__conv__Float16_ulong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, ULONG, H5__Float16, unsigned long, 0, ULONG_MAX); +} + +herr_t +H5T__conv__Float16_llong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, LLONG, H5__Float16, long long, LLONG_MIN, LLONG_MAX); +} + +herr_t +H5T__conv__Float16_ullong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fX(FLOAT16, ULLONG, H5__Float16, unsigned long long, 0, ULLONG_MAX); +} + +herr_t +H5T__conv__Float16_float(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fF(FLOAT16, FLOAT, H5__Float16, float, -, -); +} + +herr_t +H5T__conv__Float16_double(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fF(FLOAT16, DOUBLE, H5__Float16, double, -, -); +} + +herr_t +H5T__conv__Float16_ldouble(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, + size_t H5_ATTR_UNUSED bkg_stride, void *buf, void H5_ATTR_UNUSED *bkg) +{ + H5T_CONV_fF(FLOAT16, LDOUBLE, H5__Float16, long double, -, -); +} +#endif + /*------------------------------------------------------------------------- * Function: H5T__conv_f_i * @@ -7910,8 +8205,9 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz uint8_t *int_buf = NULL; /*buffer for temporary value */ size_t buf_size; /*buffer size for temporary value */ size_t i; /*miscellaneous counters */ - size_t first; /*first bit(MSB) in an integer */ - ssize_t sfirst; /*a signed version of `first' */ + ssize_t msb_pos_s; /*first bit(MSB) in an integer */ + ssize_t new_msb_pos; /*MSB position after shifting mantissa by exponent */ + hssize_t shift_val; /*shift value when shifting mantissa by exponent */ H5T_conv_cb_t cb_struct = {NULL, NULL}; /*conversion callback structure */ bool truncated; /*if fraction value is dropped */ bool reverse; /*if reverse order of destination at the end */ @@ -7976,8 +8272,11 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz /* Allocate enough space for the buffer holding temporary * converted value */ - buf_size = (size_t)(pow(2.0, (double)src.u.f.esize) / 8 + 1); - int_buf = (uint8_t *)H5MM_calloc(buf_size); + if (dst.prec / 8 > src_p->shared->size) + buf_size = (dst.prec + 7) / 8; + else + buf_size = src_p->shared->size; + int_buf = (uint8_t *)H5MM_calloc(buf_size); /* Get conversion exception callback property */ if (H5CX_get_dt_conv_cb(&cb_struct) < 0) @@ -8234,35 +8533,46 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz H5T__bit_inc(int_buf, src.u.f.msize, 8 * buf_size - src.u.f.msize); /* + * What is the bit position for the most significant bit(MSB) of S + * which is set? This is checked before shifting and before possibly + * converting to a negative integer. Note that later use of this value + * assumes that H5T__bit_shift will always shift in 0 during a right + * shift. + */ + msb_pos_s = H5T__bit_find(int_buf, (size_t)0, src.prec, H5T_BIT_MSB, true); + + /* + * The temporary buffer has no bits set and must therefore be + * zero; nothing to do. + */ + if (msb_pos_s < 0) + goto padding; + + /* * Shift mantissa part by exponent minus mantissa size(right shift), * or by mantissa size minus exponent(left shift). Example: Sequence * 10...010111, expo=20, expo-msize=-3. Right-shift the sequence, we get * 00010...10. The last three bits were dropped. */ - H5T__bit_shift(int_buf, expo - (ssize_t)src.u.f.msize, (size_t)0, buf_size * 8); + shift_val = expo - (ssize_t)src.u.f.msize; + H5T__bit_shift(int_buf, shift_val, (size_t)0, buf_size * 8); + + /* Calculate the new position of the MSB after shifting and + * skip to the padding section if we shifted exactly to 0 + * (MSB position is -1) + */ + new_msb_pos = msb_pos_s + shift_val; + if (new_msb_pos == -1) + goto padding; /* - * If expo is less than mantissa size, the frantional value is dropped off + * If expo is less than mantissa size, the fractional value is dropped off * during conversion. Set exception type to be "truncate" */ if ((size_t)expo < src.u.f.msize && cb_struct.func) truncated = true; - /* - * What is the bit position for the most significant bit(MSB) of S - * which is set? This is checked before converted to negative - * integer. - */ - sfirst = H5T__bit_find(int_buf, (size_t)0, 8 * buf_size, H5T_BIT_MSB, true); - first = (size_t)sfirst; - - if (sfirst < 0) { - /* - * The source has no bits set and must therefore be zero. - * Set the destination to zero - nothing to do. - */ - } - else if (H5T_SGN_NONE == dst.u.i.sign) { /*destination is unsigned*/ + if (H5T_SGN_NONE == dst.u.i.sign) { /*destination is unsigned*/ /* * Destination is unsigned. Library's default way: If the source value * is greater than the maximal destination value then it overflows, the @@ -8289,7 +8599,7 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz } } else { /*source is positive*/ - if (first >= dst.prec) { + if (new_msb_pos >= (ssize_t)dst.prec) { /*overflow*/ if (cb_struct.func) { /*If user's exception handler is present, use it*/ /*reverse order first*/ @@ -8310,7 +8620,7 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz HGOTO_ERROR(H5E_DATATYPE, H5E_CANTCONVERT, FAIL, "can't handle conversion exception"); } - else if (first < dst.prec) { + else { if (truncated && cb_struct.func) { /*If user's exception handler is present, use it*/ /*reverse order first*/ @@ -8320,9 +8630,11 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz src_rev, d, cb_struct.user_data); } - if (except_ret == H5T_CONV_UNHANDLED) + if (except_ret == H5T_CONV_UNHANDLED) { /*copy source value into it if case is ignored by user handler*/ - H5T__bit_copy(d, dst.offset, int_buf, (size_t)0, first + 1); + if (new_msb_pos >= 0) + H5T__bit_copy(d, dst.offset, int_buf, (size_t)0, (size_t)new_msb_pos + 1); + } else if (except_ret == H5T_CONV_HANDLED) { /*No need to reverse the order of destination because user handles it*/ reverse = false; @@ -8336,7 +8648,7 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz } else if (H5T_SGN_2 == dst.u.i.sign) { /*Destination is signed*/ if (sign) { /*source is negative*/ - if (first < dst.prec - 1) { + if ((new_msb_pos >= 0) && ((size_t)new_msb_pos < dst.prec - 1)) { if (truncated && cb_struct.func) { /*If user's exception handler is present, use it*/ /*reverse order first*/ @@ -8348,8 +8660,8 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz if (except_ret == H5T_CONV_UNHANDLED) { /*If this case ignored by user handler*/ /*Convert to integer representation. Equivalent to ~(value - 1).*/ - H5T__bit_dec(int_buf, (size_t)0, 8 * buf_size); - H5T__bit_neg(int_buf, (size_t)0, 8 * buf_size); + H5T__bit_dec(int_buf, (size_t)0, dst.prec); + H5T__bit_neg(int_buf, (size_t)0, dst.prec); /*copy source value into destination*/ H5T__bit_copy(d, dst.offset, int_buf, (size_t)0, dst.prec - 1); @@ -8389,7 +8701,7 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz } } else { /*source is positive*/ - if (first >= dst.prec - 1) { + if (new_msb_pos >= (ssize_t)dst.prec - 1) { /*overflow*/ if (cb_struct.func) { /*If user's exception handler is present, use it*/ /*reverse order first*/ @@ -8410,7 +8722,7 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz goto next; } } - else if (first < dst.prec - 1) { + else if (new_msb_pos < (ssize_t)dst.prec - 1) { if (truncated && cb_struct.func) { /*If user's exception handler is present, use it*/ /*reverse order first*/ @@ -8422,7 +8734,8 @@ H5T__conv_f_i(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz if (except_ret == H5T_CONV_UNHANDLED) { /*copy source value into it if case is ignored by user handler*/ - H5T__bit_copy(d, dst.offset, int_buf, (size_t)0, first + 1); + if (new_msb_pos >= 0) + H5T__bit_copy(d, dst.offset, int_buf, (size_t)0, (size_t)new_msb_pos + 1); } else if (except_ret == H5T_CONV_ABORT) HGOTO_ERROR(H5E_DATATYPE, H5E_CANTCONVERT, FAIL, @@ -8601,7 +8914,7 @@ H5T__conv_i_f(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, siz /* Allocate enough space for the buffer holding temporary * converted value */ - buf_size = (src.prec > dst.u.f.msize ? src.prec : dst.u.f.msize) / 8 + 1; + buf_size = ((src.prec > dst.u.f.msize ? src.prec : dst.u.f.msize) + 7) / 8; int_buf = (uint8_t *)H5MM_calloc(buf_size); /* Get conversion exception callback property */ diff --git a/src/H5Tinit_float.c b/src/H5Tinit_float.c index 3213f00..7ddd65b 100644 --- a/src/H5Tinit_float.c +++ b/src/H5Tinit_float.c @@ -53,8 +53,8 @@ * Purpose: This macro takes a floating point type like `double' and * and detects byte order, mantissa location, exponent location, * sign bit location, presence or absence of implicit mantissa - * bit, and exponent bias and initializes a detected_t structure - * with those properties. + * bit, and exponent bias and initializes a H5T_fpoint_det_t + * structure with those properties. * * Note that these operations can raise floating-point * exceptions and building with some compiler options @@ -306,14 +306,17 @@ H5T__fix_order(int n, int last, int *perm, H5T_order_t *order) if (last <= 0) HGOTO_ERROR(H5E_DATATYPE, H5E_CANTINIT, FAIL, "failed to detect byte order"); - /* We have at least three points to consider */ - if (perm[last] < perm[last - 1] && perm[last - 1] < perm[last - 2]) { + if (perm[last] < perm[last - 1] && + /* Only check perm[last - 2] if we have more than 2 points to consider */ + ((last < 2) || (perm[last - 1] < perm[last - 2]))) { /* Little endian */ *order = H5T_ORDER_LE; for (int i = 0; i < n; i++) perm[i] = i; } - else if (perm[last] > perm[last - 1] && perm[last - 1] > perm[last - 2]) { + else if (perm[last] > perm[last - 1] && + /* Only check perm[last - 2] if we have more than 2 points to consider */ + ((last < 2) || (perm[last - 1] > perm[last - 2]))) { /* Big endian */ *order = H5T_ORDER_BE; for (int i = 0; i < n; i++) @@ -358,7 +361,7 @@ done: * Return: imp_bit will be set to 1 if the most significant bit * of the mantissa is discarded (ie, the mantissa has an * implicit `one' as the most significant bit). Otherwise, - * imp_bit will be set to zero zero. + * imp_bit will be set to zero. * * SUCCEED/FAIL *------------------------------------------------------------------------- @@ -570,6 +573,39 @@ H5T__init_native_float_types(void) */ H5T_native_order_g = det.order; +#ifdef H5_HAVE__FLOAT16 + /* H5T_NATIVE_FLOAT16 */ + + /* Get the type's characteristics */ + memset(&det, 0, sizeof(H5T_fpoint_det_t)); + DETECT_F(H5__Float16, det); + + /* Allocate and fill type structure */ + if (NULL == (dt = H5T__alloc())) + HGOTO_ERROR(H5E_DATATYPE, H5E_NOSPACE, FAIL, "datatype allocation failed"); + dt->shared->state = H5T_STATE_IMMUTABLE; + dt->shared->type = H5T_FLOAT; + dt->shared->size = det.size; + dt->shared->u.atomic.order = det.order; + dt->shared->u.atomic.offset = det.offset; + dt->shared->u.atomic.prec = det.prec; + dt->shared->u.atomic.lsb_pad = H5T_PAD_ZERO; + dt->shared->u.atomic.msb_pad = H5T_PAD_ZERO; + dt->shared->u.atomic.u.f.sign = det.sign; + dt->shared->u.atomic.u.f.epos = det.epos; + dt->shared->u.atomic.u.f.esize = det.esize; + dt->shared->u.atomic.u.f.ebias = det.ebias; + dt->shared->u.atomic.u.f.mpos = det.mpos; + dt->shared->u.atomic.u.f.msize = det.msize; + dt->shared->u.atomic.u.f.norm = det.norm; + dt->shared->u.atomic.u.f.pad = H5T_PAD_ZERO; + + /* Register the type and set global variables */ + if ((H5T_NATIVE_FLOAT16_g = H5I_register(H5I_DATATYPE, dt, false)) < 0) + HGOTO_ERROR(H5E_DATATYPE, H5E_CANTINIT, FAIL, "can't register ID for built-in datatype"); + H5T_NATIVE_FLOAT16_ALIGN_g = det.comp_align; +#endif + done: /* Clear any FE_INVALID exceptions from NaN handling */ if (feclearexcept(FE_INVALID) != 0) diff --git a/src/H5Tmodule.h b/src/H5Tmodule.h index f1b7b17..3654025 100644 --- a/src/H5Tmodule.h +++ b/src/H5Tmodule.h @@ -712,6 +712,14 @@ * </tr> * <tr> * <td> + * #H5T_NATIVE_FLOAT16 + * </td> + * <td span='3'> + * _Float16 + * </td> + * </tr> + * <tr> + * <td> * #H5T_NATIVE_FLOAT * </td> * <td span='3'> @@ -3753,10 +3761,11 @@ filled according to the value of this property. The padding can be: * H5T_NATIVE_INT | H5T_NATIVE_UINT | * H5T_NATIVE_LONG | H5T_NATIVE_ULONG | * H5T_NATIVE_LLONG | H5T_NATIVE_ULLONG - * <float> ::= H5T_IEEE_F32BE | H5T_IEEE_F32LE | + * <float> ::= H5T_IEEE_F16BE | H5T_IEEE_F16LE | + * H5T_IEEE_F32BE | H5T_IEEE_F32LE | * H5T_IEEE_F64BE | H5T_IEEE_F64LE | - * H5T_NATIVE_FLOAT | H5T_NATIVE_DOUBLE | - * H5T_NATIVE_LDOUBLE + * H5T_NATIVE_FLOAT16 | H5T_NATIVE_FLOAT | + * H5T_NATIVE_DOUBLE | H5T_NATIVE_LDOUBLE * <time> ::= H5T_TIME: not yet implemented * <string> ::= H5T_STRING { * STRSIZE <strsize> ; diff --git a/src/H5Tnative.c b/src/H5Tnative.c index f83e9c3..8136f96 100644 --- a/src/H5Tnative.c +++ b/src/H5Tnative.c @@ -50,6 +50,7 @@ static herr_t H5T__cmp_offset(size_t *comp_size, size_t *offset, size_t elem_siz * H5T_NATIVE_LONG H5T_NATIVE_ULONG * H5T_NATIVE_LLONG H5T_NATIVE_ULLONG * + * H5T_NATIVE_FLOAT16 (if available) * H5T_NATIVE_FLOAT * H5T_NATIVE_DOUBLE * H5T_NATIVE_LDOUBLE @@ -703,6 +704,7 @@ H5T__get_native_float(size_t size, H5T_direction_t direction, size_t *struct_ali size_t align = 0; /* Alignment necessary for native datatype */ size_t native_size = 0; /* Datatype size of the native type */ enum match_type { /* The different kinds of floating point types we can match */ + H5T_NATIVE_FLOAT_MATCH_FLOAT16, H5T_NATIVE_FLOAT_MATCH_FLOAT, H5T_NATIVE_FLOAT_MATCH_DOUBLE, H5T_NATIVE_FLOAT_MATCH_LDOUBLE, @@ -715,7 +717,14 @@ H5T__get_native_float(size_t size, H5T_direction_t direction, size_t *struct_ali assert(size > 0); if (direction == H5T_DIR_DEFAULT || direction == H5T_DIR_ASCEND) { - if (size <= sizeof(float)) { +#ifdef H5_HAVE__FLOAT16 + if (size <= sizeof(H5__Float16)) { + match = H5T_NATIVE_FLOAT_MATCH_FLOAT16; + native_size = sizeof(H5__Float16); + } + else +#endif + if (size <= sizeof(float)) { match = H5T_NATIVE_FLOAT_MATCH_FLOAT; native_size = sizeof(float); } @@ -741,14 +750,29 @@ H5T__get_native_float(size_t size, H5T_direction_t direction, size_t *struct_ali match = H5T_NATIVE_FLOAT_MATCH_DOUBLE; native_size = sizeof(double); } - else { + else +#ifdef H5_HAVE__FLOAT16 + if (size > sizeof(H5__Float16)) +#endif + { match = H5T_NATIVE_FLOAT_MATCH_FLOAT; native_size = sizeof(float); } +#ifdef H5_HAVE__FLOAT16 + else { + match = H5T_NATIVE_FLOAT_MATCH_FLOAT16; + native_size = sizeof(H5__Float16); + } +#endif } /* Set the appropriate native floating point information */ switch (match) { + case H5T_NATIVE_FLOAT_MATCH_FLOAT16: + tid = H5T_NATIVE_FLOAT16; + align = H5T_NATIVE_FLOAT16_ALIGN_g; + break; + case H5T_NATIVE_FLOAT_MATCH_FLOAT: tid = H5T_NATIVE_FLOAT; align = H5T_NATIVE_FLOAT_ALIGN_g; diff --git a/src/H5Tpkg.h b/src/H5Tpkg.h index ef5ba36..8f063b9 100644 --- a/src/H5Tpkg.h +++ b/src/H5Tpkg.h @@ -416,6 +416,7 @@ H5_DLLVAR size_t H5T_NATIVE_LONG_ALIGN_g; H5_DLLVAR size_t H5T_NATIVE_ULONG_ALIGN_g; H5_DLLVAR size_t H5T_NATIVE_LLONG_ALIGN_g; H5_DLLVAR size_t H5T_NATIVE_ULLONG_ALIGN_g; +H5_DLLVAR size_t H5T_NATIVE_FLOAT16_ALIGN_g; H5_DLLVAR size_t H5T_NATIVE_FLOAT_ALIGN_g; H5_DLLVAR size_t H5T_NATIVE_DOUBLE_ALIGN_g; H5_DLLVAR size_t H5T_NATIVE_LDOUBLE_ALIGN_g; @@ -451,6 +452,10 @@ H5_DLLVAR size_t H5T_NATIVE_UINT_FAST64_ALIGN_g; /* Useful floating-point values for conversion routines */ /* (+/- Inf for all floating-point types) */ +#ifdef H5_HAVE__FLOAT16 +H5_DLLVAR H5__Float16 H5T_NATIVE_FLOAT16_POS_INF_g; +H5_DLLVAR H5__Float16 H5T_NATIVE_FLOAT16_NEG_INF_g; +#endif H5_DLLVAR float H5T_NATIVE_FLOAT_POS_INF_g; H5_DLLVAR float H5T_NATIVE_FLOAT_NEG_INF_g; H5_DLLVAR double H5T_NATIVE_DOUBLE_POS_INF_g; @@ -831,6 +836,61 @@ H5_DLL herr_t H5T__conv_ldouble_llong(hid_t src_id, hid_t dst_id, H5T_cdata_t *c H5_DLL herr_t H5T__conv_ldouble_ullong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +#ifdef H5_HAVE__FLOAT16 +H5_DLL herr_t H5T__conv_schar__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_uchar__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_short__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_ushort__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_int__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_uint__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_long__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_ulong__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_llong__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_ullong__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_float__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_double__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv_ldouble__Float16(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_schar(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_uchar(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_short(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_ushort(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_int(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_uint(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_long(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_ulong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_llong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_ullong(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_float(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_double(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +H5_DLL herr_t H5T__conv__Float16_ldouble(hid_t src_id, hid_t dst_id, H5T_cdata_t *cdata, size_t nelmts, + size_t buf_stride, size_t bkg_stride, void *buf, void *bkg); +#endif + /* Bit twiddling functions */ H5_DLL void H5T__bit_copy(uint8_t *dst, size_t dst_offset, const uint8_t *src, size_t src_offset, size_t size); diff --git a/src/H5Tpublic.h b/src/H5Tpublic.h index a117075..cfeb384 100644 --- a/src/H5Tpublic.h +++ b/src/H5Tpublic.h @@ -251,6 +251,16 @@ typedef H5T_conv_ret_t (*H5T_conv_except_func_t)(H5T_conv_except_t except_type, */ /** * \ingroup PDTIEEE + * 16-bit big-endian IEEE floating-point numbers + */ +#define H5T_IEEE_F16BE (H5OPEN H5T_IEEE_F16BE_g) +/** + * \ingroup PDTIEEE + * 16-bit little-endian IEEE floating-point numbers + */ +#define H5T_IEEE_F16LE (H5OPEN H5T_IEEE_F16LE_g) +/** + * \ingroup PDTIEEE * 32-bit big-endian IEEE floating-point numbers */ #define H5T_IEEE_F32BE (H5OPEN H5T_IEEE_F32BE_g) @@ -269,6 +279,8 @@ typedef H5T_conv_ret_t (*H5T_conv_except_func_t)(H5T_conv_except_t except_type, * 64-bit little-endian IEEE floating-point numbers */ #define H5T_IEEE_F64LE (H5OPEN H5T_IEEE_F64LE_g) +H5_DLLVAR hid_t H5T_IEEE_F16BE_g; +H5_DLLVAR hid_t H5T_IEEE_F16LE_g; H5_DLLVAR hid_t H5T_IEEE_F32BE_g; H5_DLLVAR hid_t H5T_IEEE_F32LE_g; H5_DLLVAR hid_t H5T_IEEE_F64BE_g; @@ -792,6 +804,11 @@ H5_DLLVAR hid_t H5T_VAX_F64_g; #define H5T_NATIVE_ULLONG (H5OPEN H5T_NATIVE_ULLONG_g) /** * \ingroup PDTNAT + * C-style \Code{_Float16} + */ +#define H5T_NATIVE_FLOAT16 (H5OPEN H5T_NATIVE_FLOAT16_g) +/** + * \ingroup PDTNAT * C-style \Code{float} */ #define H5T_NATIVE_FLOAT (H5OPEN H5T_NATIVE_FLOAT_g) @@ -865,6 +882,7 @@ H5_DLLVAR hid_t H5T_NATIVE_LONG_g; H5_DLLVAR hid_t H5T_NATIVE_ULONG_g; H5_DLLVAR hid_t H5T_NATIVE_LLONG_g; H5_DLLVAR hid_t H5T_NATIVE_ULLONG_g; +H5_DLLVAR hid_t H5T_NATIVE_FLOAT16_g; H5_DLLVAR hid_t H5T_NATIVE_FLOAT_g; H5_DLLVAR hid_t H5T_NATIVE_DOUBLE_g; H5_DLLVAR hid_t H5T_NATIVE_LDOUBLE_g; @@ -2319,6 +2337,7 @@ H5_DLL htri_t H5Tis_variable_str(hid_t type_id); * \li #H5T_NATIVE_ULONG * \li #H5T_NATIVE_ULLONG * + * \li #H5T_NATIVE_FLOAT16 (if available) * \li #H5T_NATIVE_FLOAT * \li #H5T_NATIVE_DOUBLE * \li #H5T_NATIVE_LDOUBLE diff --git a/src/H5private.h b/src/H5private.h index 5663c00..0cbda2e 100644 --- a/src/H5private.h +++ b/src/H5private.h @@ -27,7 +27,6 @@ #include <errno.h> #include <fcntl.h> #include <fenv.h> -#include <float.h> #include <math.h> #include <setjmp.h> #include <signal.h> @@ -36,6 +35,10 @@ #include <string.h> #include <time.h> +/* Define __STDC_WANT_IEC_60559_TYPES_EXT__ for _FloatN support, if available */ +#define __STDC_WANT_IEC_60559_TYPES_EXT__ +#include <float.h> + /* POSIX headers */ #ifdef H5_HAVE_SYS_TIME_H #include <sys/time.h> @@ -540,6 +543,21 @@ #define H5_GCC_CLANG_DIAG_ON(x) #endif +/* Create a typedef for library usage of the _Float16 type + * to avoid issues when compiling the library with the + * -pedantic flag or similar where we get warnings about + * _Float16 not being an ISO C type. + */ +#ifdef H5_HAVE__FLOAT16 +#if defined(__GNUC__) +__extension__ typedef _Float16 H5__Float16; +#elif defined(__clang__) +/* TODO */ +#else +typedef _Float16 H5__Float16; +#endif +#endif + /* Function pointer typedef for qsort */ typedef int (*H5_sort_func_cb_t)(const void *, const void *); diff --git a/src/H5trace.c b/src/H5trace.c index 7657356..37f45fc 100644 --- a/src/H5trace.c +++ b/src/H5trace.c @@ -1625,12 +1625,20 @@ H5_trace_args(H5RS_str_t *rs, const char *type, va_list ap) H5RS_acat(rs, "H5T_NATIVE_LLONG"); else if (obj == H5T_NATIVE_ULLONG_g) H5RS_acat(rs, "H5T_NATIVE_ULLONG"); +#ifdef H5_HAVE__FLOAT16 + else if (obj == H5T_NATIVE_FLOAT16_g) + H5RS_acat(rs, "H5T_NATIVE_FLOAT16"); +#endif else if (obj == H5T_NATIVE_FLOAT_g) H5RS_acat(rs, "H5T_NATIVE_FLOAT"); else if (obj == H5T_NATIVE_DOUBLE_g) H5RS_acat(rs, "H5T_NATIVE_DOUBLE"); else if (obj == H5T_NATIVE_LDOUBLE_g) H5RS_acat(rs, "H5T_NATIVE_LDOUBLE"); + else if (obj == H5T_IEEE_F16BE_g) + H5RS_acat(rs, "H5T_IEEE_F16BE"); + else if (obj == H5T_IEEE_F16LE_g) + H5RS_acat(rs, "H5T_IEEE_F16LE"); else if (obj == H5T_IEEE_F32BE_g) H5RS_acat(rs, "H5T_IEEE_F32BE"); else if (obj == H5T_IEEE_F32LE_g) |