[PATCH 2/2] Add bfloat16 support for AVX512 register view.
Kevin Buettner
kevinb@redhat.com
Wed Jul 22 22:19:44 GMT 2020
Hi Felix,
See my comments inline below.
Assuming you get approval on the include and libiberty portions, I
think this should be pushed (without another round trip back to this
list) after addressing my concerns.
On Tue, 21 Jul 2020 17:52:55 +0200
Felix Willgerodt via Gdb-patches <gdb-patches@sourceware.org> wrote:
> This adds support for the bfloat16 datatype, which can be seen as a short
> version of FP32, skipping the least significant 16 bits of the mantissa.
> Since the datatype is currently only supported by the AVX512 registers,
> the printing of bfloat16 values is only supported for xmm, ymm and zmm
> registers.
>
> gdb/ChangeLog:
> 2020-07-07 Moritz Riesterer <moritz.riesterer@intel.com>
> Felix Willgerodt <Felix.Willgerodt@intel.com>
>
> * gdbarch.sh: Added bfloat16 type.
> * gdbarch.c: Regenerated.
> * gdbarch.h: Regenerated.
> * gdbtypes.c (floatformats_bfloat16): New struct.
> (gdbtypes_post_init): Add builtin_bfloat16.
> * gdbtypes.h (struct builtin_type) <builtin_bfloat16>: New member.
> (floatformats_bfloat16): New struct.
> * i386-tdep.c (i386_zmm_type): Add field "v32_bfloat16"
> (i386_ymm_type): Add field "v16_bfloat16"
> (i386_gdbarch_init): Add set_gdbarch_bfloat16_format.
> * target-descriptions.c (make_gdb_type): Add case TDESC_TYPE_BFLOAT16.
> * gdbsupport/tdesc.cc (tdesc_predefined_types): New member bfloat16.
> * gdbsupport/tdesc.h (tdesc_type_kind): New member TDESC_TYPE_BFLOAT16.
>
> gdb/features/ChangeLog:
I think this is just gdb/ChangeLog, not gdb/features/ChangeLog.
> 2020-07-07 Moritz Riesterer <moritz.riesterer@intel.com>
> Felix Willgerodt <Felix.Willgerodt@intel.com>
>
> * features/64bit-avx512.c
> (create_feature_i386_64bit_avx512): Add vector v8bf.
> (create_feature_i386_64bit_avx512): Add field v8_bfloat16.
> * features/64bit-avx512.xml: Likewise.
> * features/64bit-sse.c
> (create_feature_i386_64bit_sse): Add vector v8bf.
> (create_feature_i386_64bit_sse): Add field v8_bfloat16.
> * features/64bit-sse.xml: Likewise.
These files are in features/i386.
Also, the .c files were (re)generated from the .xml files, right?
> gdb/testsuite/ChangeLog:
> 2020-07-07 Moritz Riesterer <moritz.riesterer@intel.com>
> Felix Willgerodt <Felix.Willgerodt@intel.com>
>
> * x86-avx512bf16.c: New file.
> * x86-avx512bf16.exp: Likewise.
> * lib/gdb.exp (skip_avx512bf16_tests): New function.
>
> include/ChangeLog:
> 2020-07-07 Moritz Riesterer <moritz.riesterer@intel.com>
>
> * floatformat.h: Add struct floatformat_intel_bfloat16_big.
> Add struct floatformat_intel_bfloat16_little.
>
> libiberty/ChangeLog:
> 2020-07-07 Moritz Riesterer <moritz.riesterer@intel.com>
>
> * floatformat.c: Add struct floatformat_intel_bfloat16_big.
> Add struct floatformat_intel_bfloat16_little.
>
> 2020-07-21 Felix Willgerodt <felix.willgerodt@intel.com>
> ---
> gdb/features/i386/64bit-avx512.c | 5 +
> gdb/features/i386/64bit-avx512.xml | 4 +-
> gdb/features/i386/64bit-sse.c | 5 +
> gdb/features/i386/64bit-sse.xml | 2 +
> gdb/gdbarch.c | 45 ++++++
> gdb/gdbarch.h | 8 +-
> gdb/gdbarch.sh | 4 +-
> gdb/gdbtypes.c | 7 +
> gdb/gdbtypes.h | 3 +-
> gdb/i386-tdep.c | 9 ++
> gdb/target-descriptions.c | 5 +
> gdb/testsuite/gdb.arch/x86-avx512bf16.c | 166 ++++++++++++++++++++++
> gdb/testsuite/gdb.arch/x86-avx512bf16.exp | 74 ++++++++++
> gdb/testsuite/lib/gdb.exp | 51 +++++++
> gdbsupport/tdesc.cc | 3 +-
> gdbsupport/tdesc.h | 1 +
> include/floatformat.h | 3 +
> libiberty/floatformat.c | 19 ++-
> 18 files changed, 408 insertions(+), 6 deletions(-)
> create mode 100644 gdb/testsuite/gdb.arch/x86-avx512bf16.c
> create mode 100644 gdb/testsuite/gdb.arch/x86-avx512bf16.exp
>
> diff --git a/gdb/features/i386/64bit-avx512.c b/gdb/features/i386/64bit-avx512.c
> index d12234c41ce..1bd49dcd9d5 100644
> --- a/gdb/features/i386/64bit-avx512.c
> +++ b/gdb/features/i386/64bit-avx512.c
> @@ -10,6 +10,9 @@ create_feature_i386_64bit_avx512 (struct target_desc *result, long regnum)
>
> feature = tdesc_create_feature (result, "org.gnu.gdb.i386.avx512");
> tdesc_type *element_type;
> + element_type = tdesc_named_type (feature, "bfloat16");
> + tdesc_create_vector (feature, "v8bf16", element_type, 8);
> +
> element_type = tdesc_named_type (feature, "ieee_single");
> tdesc_create_vector (feature, "v4f", element_type, 4);
>
> @@ -31,6 +34,8 @@ create_feature_i386_64bit_avx512 (struct target_desc *result, long regnum)
> tdesc_type_with_fields *type_with_fields;
> type_with_fields = tdesc_create_union (feature, "vec128");
> tdesc_type *field_type;
> + field_type = tdesc_named_type (feature, "v8bf16");
> + tdesc_add_field (type_with_fields, "v8_bfloat16", field_type);
> field_type = tdesc_named_type (feature, "v4f");
> tdesc_add_field (type_with_fields, "v4_float", field_type);
> field_type = tdesc_named_type (feature, "v2d");
> diff --git a/gdb/features/i386/64bit-avx512.xml b/gdb/features/i386/64bit-avx512.xml
> index 4f02136533d..3636121e847 100644
> --- a/gdb/features/i386/64bit-avx512.xml
> +++ b/gdb/features/i386/64bit-avx512.xml
> @@ -7,13 +7,15 @@
>
> <!DOCTYPE feature SYSTEM "gdb-target.dtd">
> <feature name="org.gnu.gdb.i386.avx512">
> -<vector id="v4f" type="ieee_single" count="4"/>
> + <vector id="v8bf16" type="bfloat16" count="8"/>
> + <vector id="v4f" type="ieee_single" count="4"/>
> <vector id="v2d" type="ieee_double" count="2"/>
> <vector id="v16i8" type="int8" count="16"/>
> <vector id="v8i16" type="int16" count="8"/>
> <vector id="v4i32" type="int32" count="4"/>
> <vector id="v2i64" type="int64" count="2"/>
> <union id="vec128">
> + <field name="v8_bfloat16" type="v8bf16"/>
> <field name="v4_float" type="v4f"/>
> <field name="v2_double" type="v2d"/>
> <field name="v16_int8" type="v16i8"/>
> diff --git a/gdb/features/i386/64bit-sse.c b/gdb/features/i386/64bit-sse.c
> index 48b5aa09793..645e3141337 100644
> --- a/gdb/features/i386/64bit-sse.c
> +++ b/gdb/features/i386/64bit-sse.c
> @@ -10,6 +10,9 @@ create_feature_i386_64bit_sse (struct target_desc *result, long regnum)
>
> feature = tdesc_create_feature (result, "org.gnu.gdb.i386.sse");
> tdesc_type *element_type;
> + element_type = tdesc_named_type (feature, "bfloat16");
> + tdesc_create_vector (feature, "v8bf16", element_type, 8);
> +
> element_type = tdesc_named_type (feature, "ieee_single");
> tdesc_create_vector (feature, "v4f", element_type, 4);
>
> @@ -31,6 +34,8 @@ create_feature_i386_64bit_sse (struct target_desc *result, long regnum)
> tdesc_type_with_fields *type_with_fields;
> type_with_fields = tdesc_create_union (feature, "vec128");
> tdesc_type *field_type;
> + field_type = tdesc_named_type (feature, "v8bf16");
> + tdesc_add_field (type_with_fields, "v8_bfloat16", field_type);
> field_type = tdesc_named_type (feature, "v4f");
> tdesc_add_field (type_with_fields, "v4_float", field_type);
> field_type = tdesc_named_type (feature, "v2d");
> diff --git a/gdb/features/i386/64bit-sse.xml b/gdb/features/i386/64bit-sse.xml
> index 4ec1e7c1e36..e3dec544cf4 100644
> --- a/gdb/features/i386/64bit-sse.xml
> +++ b/gdb/features/i386/64bit-sse.xml
> @@ -7,6 +7,7 @@
>
> <!DOCTYPE feature SYSTEM "gdb-target.dtd">
> <feature name="org.gnu.gdb.i386.sse">
> + <vector id="v8bf16" type="bfloat16" count="8"/>
> <vector id="v4f" type="ieee_single" count="4"/>
> <vector id="v2d" type="ieee_double" count="2"/>
> <vector id="v16i8" type="int8" count="16"/>
> @@ -14,6 +15,7 @@
> <vector id="v4i32" type="int32" count="4"/>
> <vector id="v2i64" type="int64" count="2"/>
> <union id="vec128">
> + <field name="v8_bfloat16" type="v8bf16"/>
> <field name="v4_float" type="v4f"/>
> <field name="v2_double" type="v2d"/>
> <field name="v16_int8" type="v16i8"/>
> diff --git a/gdb/gdbarch.c b/gdb/gdbarch.c
> index 21ee840e88f..d9e5dcfbaf5 100644
> --- a/gdb/gdbarch.c
> +++ b/gdb/gdbarch.c
> @@ -166,6 +166,8 @@ struct gdbarch
> int int_bit;
> int long_bit;
> int long_long_bit;
> + int bfloat16_bit;
> + const struct floatformat ** bfloat16_format;
> int half_bit;
> const struct floatformat ** half_format;
> int float_bit;
> @@ -382,6 +384,7 @@ gdbarch_alloc (const struct gdbarch_info *info,
> gdbarch->int_bit = 4*TARGET_CHAR_BIT;
> gdbarch->long_bit = 4*TARGET_CHAR_BIT;
> gdbarch->long_long_bit = 2*gdbarch->long_bit;
> + gdbarch->bfloat16_bit = 2*TARGET_CHAR_BIT;
> gdbarch->half_bit = 2*TARGET_CHAR_BIT;
> gdbarch->float_bit = 4*TARGET_CHAR_BIT;
> gdbarch->double_bit = 8*TARGET_CHAR_BIT;
> @@ -521,6 +524,9 @@ verify_gdbarch (struct gdbarch *gdbarch)
> /* Skip verify of int_bit, invalid_p == 0 */
> /* Skip verify of long_bit, invalid_p == 0 */
> /* Skip verify of long_long_bit, invalid_p == 0 */
> + /* Skip verify of bfloat16_bit, invalid_p == 0 */
> + if (gdbarch->bfloat16_format == 0)
> + gdbarch->bfloat16_format = floatformats_bfloat16;
> /* Skip verify of half_bit, invalid_p == 0 */
> if (gdbarch->half_format == 0)
> gdbarch->half_format = floatformats_ieee_half;
> @@ -804,6 +810,12 @@ gdbarch_dump (struct gdbarch *gdbarch, struct ui_file *file)
> fprintf_unfiltered (file,
> "gdbarch_dump: bfd_arch_info = %s\n",
> gdbarch_bfd_arch_info (gdbarch)->printable_name);
> + fprintf_unfiltered (file,
> + "gdbarch_dump: bfloat16_bit = %s\n",
> + plongest (gdbarch->bfloat16_bit));
> + fprintf_unfiltered (file,
> + "gdbarch_dump: bfloat16_format = %s\n",
> + pformat (gdbarch->bfloat16_format));
> fprintf_unfiltered (file,
> "gdbarch_dump: breakpoint_from_pc = <%s>\n",
> host_address_to_string (gdbarch->breakpoint_from_pc));
> @@ -1614,6 +1626,39 @@ set_gdbarch_long_long_bit (struct gdbarch *gdbarch,
> gdbarch->long_long_bit = long_long_bit;
> }
>
> +int
> +gdbarch_bfloat16_bit (struct gdbarch *gdbarch)
> +{
> + gdb_assert (gdbarch != NULL);
> + /* Skip verify of bfloat16_bit, invalid_p == 0 */
> + if (gdbarch_debug >= 2)
> + fprintf_unfiltered (gdb_stdlog, "gdbarch_bfloat16_bit called\n");
> + return gdbarch->bfloat16_bit;
> +}
> +
> +void
> +set_gdbarch_bfloat16_bit (struct gdbarch *gdbarch,
> + int bfloat16_bit)
> +{
> + gdbarch->bfloat16_bit = bfloat16_bit;
> +}
> +
> +const struct floatformat **
> +gdbarch_bfloat16_format (struct gdbarch *gdbarch)
> +{
> + gdb_assert (gdbarch != NULL);
> + if (gdbarch_debug >= 2)
> + fprintf_unfiltered (gdb_stdlog, "gdbarch_bfloat16_format called\n");
> + return gdbarch->bfloat16_format;
> +}
> +
> +void
> +set_gdbarch_bfloat16_format (struct gdbarch *gdbarch,
> + const struct floatformat ** bfloat16_format)
> +{
> + gdbarch->bfloat16_format = bfloat16_format;
> +}
> +
> int
> gdbarch_half_bit (struct gdbarch *gdbarch)
> {
> diff --git a/gdb/gdbarch.h b/gdb/gdbarch.h
> index 0940156aeb8..c555fc2bc42 100644
> --- a/gdb/gdbarch.h
> +++ b/gdb/gdbarch.h
> @@ -158,12 +158,18 @@ extern void set_gdbarch_long_bit (struct gdbarch *gdbarch, int long_bit);
> extern int gdbarch_long_long_bit (struct gdbarch *gdbarch);
> extern void set_gdbarch_long_long_bit (struct gdbarch *gdbarch, int long_long_bit);
>
> -/* The ABI default bit-size and format for "half", "float", "double", and
> +/* The ABI default bit-size and format for "bfloat16", "half", "float", "double", and
> "long double". These bit/format pairs should eventually be combined
> into a single object. For the moment, just initialize them as a pair.
> Each format describes both the big and little endian layouts (if
> useful). */
>
> +extern int gdbarch_bfloat16_bit (struct gdbarch *gdbarch);
> +extern void set_gdbarch_bfloat16_bit (struct gdbarch *gdbarch, int bfloat16_bit);
> +
> +extern const struct floatformat ** gdbarch_bfloat16_format (struct gdbarch *gdbarch);
> +extern void set_gdbarch_bfloat16_format (struct gdbarch *gdbarch, const struct floatformat ** bfloat16_format);
> +
> extern int gdbarch_half_bit (struct gdbarch *gdbarch);
> extern void set_gdbarch_half_bit (struct gdbarch *gdbarch, int half_bit);
>
> diff --git a/gdb/gdbarch.sh b/gdb/gdbarch.sh
> index 41e7b8d5cc3..f2f2b76d954 100755
> --- a/gdb/gdbarch.sh
> +++ b/gdb/gdbarch.sh
> @@ -325,12 +325,14 @@ v;int;long_bit;;;8 * sizeof (long);4*TARGET_CHAR_BIT;;0
> # machine.
> v;int;long_long_bit;;;8 * sizeof (LONGEST);2*gdbarch->long_bit;;0
>
> -# The ABI default bit-size and format for "half", "float", "double", and
> +# The ABI default bit-size and format for "bfloat16", "half", "float", "double", and
> # "long double". These bit/format pairs should eventually be combined
> # into a single object. For the moment, just initialize them as a pair.
> # Each format describes both the big and little endian layouts (if
> # useful).
>
> +v;int;bfloat16_bit;;;16;2*TARGET_CHAR_BIT;;0
> +v;const struct floatformat **;bfloat16_format;;;;;floatformats_bfloat16;;pformat (gdbarch->bfloat16_format)
> v;int;half_bit;;;16;2*TARGET_CHAR_BIT;;0
> v;const struct floatformat **;half_format;;;;;floatformats_ieee_half;;pformat (gdbarch->half_format)
> v;int;float_bit;;;8 * sizeof (float);4*TARGET_CHAR_BIT;;0
> diff --git a/gdb/gdbtypes.c b/gdb/gdbtypes.c
> index e87648813ec..a9ce0e6c34c 100644
> --- a/gdb/gdbtypes.c
> +++ b/gdb/gdbtypes.c
> @@ -115,6 +115,10 @@ const struct floatformat *floatformats_ibm_long_double[BFD_ENDIAN_UNKNOWN] = {
> &floatformat_ibm_long_double_big,
> &floatformat_ibm_long_double_little
> };
> +const struct floatformat *floatformats_bfloat16[BFD_ENDIAN_UNKNOWN] = {
> + &floatformat_bfloat16_big,
> + &floatformat_bfloat16_little
> +};
>
> /* Should opaque types be resolved? */
>
> @@ -5739,6 +5743,9 @@ gdbtypes_post_init (struct gdbarch *gdbarch)
> builtin_type->builtin_float
> = arch_float_type (gdbarch, gdbarch_float_bit (gdbarch),
> "float", gdbarch_float_format (gdbarch));
> + builtin_type->builtin_bfloat16
> + = arch_float_type (gdbarch, gdbarch_bfloat16_bit (gdbarch),
> + "bfloat16", gdbarch_bfloat16_format (gdbarch));
> builtin_type->builtin_double
> = arch_float_type (gdbarch, gdbarch_double_bit (gdbarch),
> "double", gdbarch_double_format (gdbarch));
> diff --git a/gdb/gdbtypes.h b/gdb/gdbtypes.h
> index eaa4cff608d..961b87d43d9 100644
> --- a/gdb/gdbtypes.h
> +++ b/gdb/gdbtypes.h
> @@ -1857,6 +1857,7 @@ struct builtin_type
> struct type *builtin_unsigned_short;
> struct type *builtin_unsigned_int;
> struct type *builtin_unsigned_long;
> + struct type *builtin_bfloat16;
> struct type *builtin_half;
> struct type *builtin_float;
> struct type *builtin_double;
> @@ -1991,7 +1992,7 @@ extern const struct floatformat *floatformats_ia64_quad[BFD_ENDIAN_UNKNOWN];
> extern const struct floatformat *floatformats_vax_f[BFD_ENDIAN_UNKNOWN];
> extern const struct floatformat *floatformats_vax_d[BFD_ENDIAN_UNKNOWN];
> extern const struct floatformat *floatformats_ibm_long_double[BFD_ENDIAN_UNKNOWN];
> -
> +extern const struct floatformat *floatformats_bfloat16[BFD_ENDIAN_UNKNOWN];
>
> /* Allocate space for storing data associated with a particular
> type. We ensure that the space is allocated using the same
> diff --git a/gdb/i386-tdep.c b/gdb/i386-tdep.c
> index 648974455f6..51c5a644d5d 100644
> --- a/gdb/i386-tdep.c
> +++ b/gdb/i386-tdep.c
> @@ -3110,6 +3110,7 @@ i386_zmm_type (struct gdbarch *gdbarch)
> int8_t v64_int8[64];
> double v8_double[8];
> float v16_float[16];
> + bfloat16_t v32_bfloat16[32];
> };
> #endif
>
> @@ -3117,6 +3118,8 @@ i386_zmm_type (struct gdbarch *gdbarch)
>
> t = arch_composite_type (gdbarch,
> "__gdb_builtin_type_vec512i", TYPE_CODE_UNION);
> + append_composite_type_field (t, "v32_bfloat16",
> + init_vector_type (bt->builtin_bfloat16, 32));
> append_composite_type_field (t, "v16_float",
> init_vector_type (bt->builtin_float, 16));
> append_composite_type_field (t, "v8_double",
> @@ -3163,6 +3166,7 @@ i386_ymm_type (struct gdbarch *gdbarch)
> int8_t v32_int8[32];
> double v4_double[4];
> float v8_float[8];
> + bfloat16_t v16_bfloat16[16];
> };
> #endif
>
> @@ -3170,6 +3174,8 @@ i386_ymm_type (struct gdbarch *gdbarch)
>
> t = arch_composite_type (gdbarch,
> "__gdb_builtin_type_vec256i", TYPE_CODE_UNION);
> + append_composite_type_field (t, "v16_bfloat16",
> + init_vector_type (bt->builtin_bfloat16, 16));
> append_composite_type_field (t, "v8_float",
> init_vector_type (bt->builtin_float, 8));
> append_composite_type_field (t, "v4_double",
> @@ -8483,6 +8489,9 @@ i386_gdbarch_init (struct gdbarch_info info, struct gdbarch_list *arches)
> alignment. */
> set_gdbarch_long_double_bit (gdbarch, 96);
>
> + /* Support of bfloat16 format. */
> + set_gdbarch_bfloat16_format (gdbarch, floatformats_bfloat16);
> +
> /* Support for floating-point data type variants. */
> set_gdbarch_floatformat_for_type (gdbarch, i386_floatformat_for_type);
>
> diff --git a/gdb/target-descriptions.c b/gdb/target-descriptions.c
> index 20d624c0c65..6778b93400b 100644
> --- a/gdb/target-descriptions.c
> +++ b/gdb/target-descriptions.c
> @@ -141,6 +141,11 @@ make_gdb_type (struct gdbarch *gdbarch, struct tdesc_type *ttype)
> m_type = arch_float_type (m_gdbarch, -1, "builtin_type_i387_ext",
> floatformats_i387_ext);
> return;
> +
> + case TDESC_TYPE_BFLOAT16:
> + m_type = arch_float_type (m_gdbarch, -1, "builtin_type_bfloat16",
> + floatformats_bfloat16);
> + return;
> }
>
> internal_error (__FILE__, __LINE__,
> diff --git a/gdb/testsuite/gdb.arch/x86-avx512bf16.c b/gdb/testsuite/gdb.arch/x86-avx512bf16.c
> new file mode 100644
> index 00000000000..fdc378818ac
> --- /dev/null
> +++ b/gdb/testsuite/gdb.arch/x86-avx512bf16.c
> @@ -0,0 +1,166 @@
> +/* Test program for bfloat16 of AVX 512 registers.
> +
> + Copyright 2020 Free Software Foundation, Inc.
> +
> + This file is part of GDB.
> +
> + This program is free software; you can redistribute it and/or modify
> + it under the terms of the GNU General Public License as published by
> + the Free Software Foundation; either version 3 of the License, or
> + (at your option) any later version.
> +
> + This program is distributed in the hope that it will be useful,
> + but WITHOUT ANY WARRANTY; without even the implied warranty of
> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> + GNU General Public License for more details.
> +
> + You should have received a copy of the GNU General Public License
> + along with this program. If not, see <http://www.gnu.org/licenses/>. */
> +
> +#include "x86-cpuid.h"
Is x86-cpuid.h needed for this test?
> +
> +typedef struct
> +{
> + float f[4];
> +} v4sd_t;
> +
> +typedef struct
> +{
> + float f[8];
> +} v8sd_t;
> +
> +typedef struct
> +{
> + float f[16];
> +} v16sd_t;
> +
> +v4sd_t xmm_data[] =
> +{
> + { { 0.0, 0.125, 0.25, 0.375 } },
> + { { 0.5, 0.625, 0.75, 0.875 } },
> + { { 1.0, 1.125, 1.25, 1.375 } },
> + { { 1.5, 1.625, 1.75, 1.875 } },
> + { { 2.0, 2.125, 2.25, 2.375 } },
> + { { 2.5, 2.625, 2.75, 2.875 } },
> + { { 3.0, 3.125, 3.25, 3.375 } },
> + { { 3.5, 3.625, 3.75, 3.875 } },
> +};
> +
> +v8sd_t ymm_data[] =
> +{
> + { { 8.0, 8.25, 8.5, 8.75, 9.0, 9.25, 9.5, 9.75 } },
> + { { 10.0, 10.25, 10.5, 10.75, 11.0, 11.25, 11.5, 11.75 } },
> + { { 12.0, 12.25, 12.5, 12.75, 13.0, 13.25, 13.5, 13.75 } },
> + { { 14.0, 14.25, 14.5, 14.75, 15.0, 15.25, 15.5, 15.75 } },
> + { { 16.0, 16.25, 16.5, 16.75, 17.0, 17.25, 17.5, 17.75 } },
> + { { 18.0, 18.25, 18.5, 18.75, 19.0, 19.25, 19.5, 19.75 } },
> + { { 20.0, 20.25, 20.5, 20.75, 21.0, 21.25, 21.5, 21.75 } },
> + { { 22.0, 22.25, 22.5, 22.75, 23.0, 23.25, 23.5, 23.75 } },
> +};
> +
> +v16sd_t zmm_data[] =
> +{
> + { { 20.0, 20.5, 21.0, 21.5, 22.0, 22.5, 23.0, 23.5, 24.0, 24.5,
> + 25.0, 25.5, 26.0, 26.5, 27.0, 27.5 } },
> + { { 28.0, 28.5, 29.0, 29.5, 30.0, 30.5, 31.0, 31.5, 32.0, 32.5,
> + 33.0, 33.5, 34.0, 34.5, 35.0, 35.5 } },
> + { { 36.0, 36.5, 37.0, 37.5, 38.0, 38.5, 39.0, 39.5, 40.0, 40.5,
> + 41.0, 41.5, 42.0, 42.5, 43.0, 43.5 } },
> + { { 44.0, 44.5, 45.0, 45.5, 46.0, 46.5, 47.0, 47.5, 48.0, 48.5,
> + 49.0, 49.5, 50.0, 50.5, 51.0, 51.5 } },
> + { { 52.0, 52.5, 53.0, 53.5, 54.0, 54.5, 55.0, 55.5, 56.0, 56.5,
> + 57.0, 57.5, 58.0, 58.5, 59.0, 59.5 } },
> + { { 60.0, 60.5, 61.0, 61.5, 62.0, 62.5, 63.0, 63.5, 64.0, 64.5,
> + 65.0, 65.5, 66.0, 66.5, 67.0, 67.5 } },
> + { { 68.0, 68.5, 69.0, 69.5, 70.0, 70.5, 71.0, 71.5, 72.0, 72.5,
> + 73.0, 73.5, 74.0, 74.5, 75.0, 75.5 } },
> + { { 76.0, 76.5, 77.0, 77.5, 78.0, 78.5, 79.0, 79.5, 80.0, 80.5,
> + 81.0, 81.5, 82.0, 82.5, 83.0, 83.5 } },
> +};
> +
> +void
> +move_data_to_xmm_reg (void)
> +{
> + asm ("vmovups 0(%0), %%xmm0 \n\t"
> + "vmovups 16(%0), %%xmm1 \n\t"
> + "vmovups 32(%0), %%xmm2 \n\t"
> + "vmovups 48(%0), %%xmm3 \n\t"
> + "vmovups 64(%0), %%xmm4 \n\t"
> + "vmovups 80(%0), %%xmm5 \n\t"
> + "vmovups 96(%0), %%xmm6 \n\t"
> + "vmovups 112(%0), %%xmm7 \n\t"
> + : /* no output operands */
> + : "r" (xmm_data));
> +}
> +
> +void
> +move_data_to_ymm_reg (void)
> +{
> + asm ("vmovups 0(%0), %%ymm0 \n\t"
> + "vmovups 32(%0), %%ymm1 \n\t"
> + "vmovups 64(%0), %%ymm2 \n\t"
> + "vmovups 96(%0), %%ymm3 \n\t"
> + "vmovups 128(%0), %%ymm4 \n\t"
> + "vmovups 160(%0), %%ymm5 \n\t"
> + "vmovups 192(%0), %%ymm6 \n\t"
> + "vmovups 224(%0), %%ymm7 \n\t"
> + : /* no output operands */
> + : "r" (ymm_data));
> +}
> +
> +void
> +move_data_to_zmm_reg (void)
> +{
> + asm ("vmovups 0(%0), %%zmm0 \n\t"
> + "vmovups 64(%0), %%zmm1 \n\t"
> + "vmovups 128(%0), %%zmm2 \n\t"
> + "vmovups 192(%0), %%zmm3 \n\t"
> + "vmovups 256(%0), %%zmm4 \n\t"
> + "vmovups 320(%0), %%zmm5 \n\t"
> + "vmovups 384(%0), %%zmm6 \n\t"
> + "vmovups 448(%0), %%zmm7 \n\t"
> + : /* no output operands */
> + : "r" (zmm_data));
> +}
> +
> +void
> +convert_xmm_from_float_to_bfloat16 (void)
> +{
> + asm("vcvtne2ps2bf16 %xmm0, %xmm1, %xmm0");
> + asm("vcvtne2ps2bf16 %xmm6, %xmm7, %xmm6");
> +}
> +
> +void
> +convert_ymm_from_float_to_bfloat16 (void)
> +{
> + asm("vcvtne2ps2bf16 %ymm0, %ymm1, %ymm0");
> + asm("vcvtne2ps2bf16 %ymm6, %ymm7, %ymm6");
> +}
> +
> +void
> +convert_zmm_from_float_to_bfloat16 (void)
> +{
> + asm("vcvtne2ps2bf16 %zmm0, %zmm1, %zmm0");
> + asm("vcvtne2ps2bf16 %zmm6, %zmm7, %zmm6");
> +}
> +
> +int
> +main (int argc, char **argv)
> +{
> + /* Move initial values from array to registers and read from XMM regs. */
> + move_data_to_xmm_reg ();
> + convert_xmm_from_float_to_bfloat16 ();
> + asm ("nop"); /* first breakpoint here */
> +
> + /* Move initial values from array to registers and read from YMM regs. */
> + move_data_to_ymm_reg ();
> + convert_ymm_from_float_to_bfloat16 ();
> + asm ("nop"); /* second breakpoint here */
> +
> + /* Move initial values from array to registers and read from ZMM regs. */
> + move_data_to_zmm_reg ();
> + convert_zmm_from_float_to_bfloat16 ();
> + asm ("nop"); /* third breakpoint here */
> +
> + return 0;
> +}
> diff --git a/gdb/testsuite/gdb.arch/x86-avx512bf16.exp b/gdb/testsuite/gdb.arch/x86-avx512bf16.exp
> new file mode 100644
> index 00000000000..5f90d8aca28
> --- /dev/null
> +++ b/gdb/testsuite/gdb.arch/x86-avx512bf16.exp
> @@ -0,0 +1,74 @@
> +# Copyright 2020 Free Software Foundation, Inc.
> +
> +# This program is free software; you can redistribute it and/or modify
> +# it under the terms of the GNU General Public License as published by
> +# the Free Software Foundation; either version 3 of the License, or
> +# (at your option) any later version.
> +#
> +# This program is distributed in the hope that it will be useful,
> +# but WITHOUT ANY WARRANTY; without even the implied warranty of
> +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> +# GNU General Public License for more details.
> +#
> +# You should have received a copy of the GNU General Public License
> +# along with this program. If not, see <http://www.gnu.org/licenses/>.
> +
> +# Please email any bugs, comments, and/or additions to this file to:
> +# bug-gdb@gnu.org
> +
> +# Test bfloat16 support in AVX512 registers
> +
> +if { [skip_avx512bf16_tests] } {
> + unsupported "target does not support AVX512BF16"
> + return -1
> +}
> +
> +standard_testfile
> +
> +if { [prepare_for_testing "failed to prepare" ${testfile} ${srcfile}] } {
> + return -1
> +}
> +
> +if { ![runto_main] } {
> + unsupported "could not run to main"
> + return -1
> +}
> +
> +# Test xmm
> +set line1 [gdb_get_line_number "first breakpoint here"]
> +gdb_breakpoint $line1
> +gdb_continue_to_breakpoint "line1" ".*$srcfile:$line1.*"
> +
> +gdb_test "print \$xmm0.v8_bfloat16" \
> + "= \\{0, 0.125, 0.25, 0.375, 0.5, 0.625, 0.75, 0.875\\}"
> +
> +gdb_test "print \$xmm6.v8_bfloat16" \
> + "= \\{3, 3.125, 3.25, 3.375, 3.5, 3.625, 3.75, 3.875\\}"
> +
> +# Test ymm
> +set line2 [gdb_get_line_number "second breakpoint here"]
> +gdb_breakpoint $line2
> +gdb_continue_to_breakpoint "line2" ".*$srcfile:$line2.*"
> +
> +gdb_test "print \$ymm0.v16_bfloat16\[1\]" "= 8.25"
> +gdb_test "print \$ymm6.v16_bfloat16\[1\]" "= 20.25"
> +
> +# Test zmm
> +set line3 [gdb_get_line_number "third breakpoint here"]
> +gdb_breakpoint $line3
> +gdb_continue_to_breakpoint "line3" ".*$srcfile:$line3.*"
> +
> +gdb_test "print \$zmm0.v32_bfloat16\[1\]" "= 20.5"
> +gdb_test "print \$zmm6.v32_bfloat16\[1\]" "= 68.5"
> +
> +# Test setting of bfloat values
> +gdb_test_no_output "set var \$xmm0.v8_bfloat16\[0\] = 32.25" \
> + "set %xmm0.v8_bfloat16\[0\]"
> +gdb_test_no_output "set var \$ymm8.v16_bfloat16\[1\] = 33.5" \
> + "set %ymm8.v16_bfloat16\[1\]"
> +gdb_test_no_output "set var \$zmm16.v32_bfloat16\[2\] = 22.75" \
> + "set %zmm16.v32_bfloat16\[2\]"
> +
> +gdb_test "p \$xmm0.v8_bfloat16\[0\]" "= 32.25"
> +gdb_test "p \$ymm8.v16_bfloat16\[1\]" "= 33.5"
> +gdb_test "p \$zmm16.v32_bfloat16\[2\]" "= 22.75"
> diff --git a/gdb/testsuite/lib/gdb.exp b/gdb/testsuite/lib/gdb.exp
> index 78dae17b0c7..fd8b5936166 100644
> --- a/gdb/testsuite/lib/gdb.exp
> +++ b/gdb/testsuite/lib/gdb.exp
> @@ -3061,6 +3061,57 @@ gdb_caching_proc skip_tsx_tests {
> return $skip_tsx_tests
> }
>
> +# Run a test on the target to see if it supports avx512bf16. Return 0 if so,
> +# 1 if it does not. Based on 'check_vmx_hw_available' from the GCC testsuite.
> +
> +gdb_caching_proc skip_avx512bf16_tests {
> + global srcdir subdir gdb_prompt inferior_exited_re
> +
> + set me "skip_avx512bf16_tests"
> + if { ![istarget "i?86-*-*"] && ![istarget "x86_64-*-*"] } {
> + verbose "$me: target does not support avx512bf16, returning 1" 2
> + return 1
> + }
> +
> + # Compile a test program.
> + set src {
> + int main() {
> + asm volatile ("vcvtne2ps2bf16 %xmm0, %xmm1, %xmm0");
> + return 0;
> + }
> + }
> + if {![gdb_simple_compile $me $src executable]} {
> + return 1
> + }
> +
> + # No error message, compilation succeeded so now run it via gdb.
> +
> + gdb_exit
> + gdb_start
> + gdb_reinitialize_dir $srcdir/$subdir
> + gdb_load "$obj"
> + gdb_run_cmd
> + gdb_expect {
> + -re ".*Illegal instruction.*${gdb_prompt} $" {
> + verbose -log "$me: avx512bf16 hardware not detected."
> + set skip_avx512bf16_tests 1
> + }
> + -re ".*$inferior_exited_re normally.*${gdb_prompt} $" {
> + verbose -log "$me: avx512bf16 hardware detected."
> + set skip_avx512bf16_tests 0
> + }
> + default {
> + warning "\n$me: default case taken."
> + set skip_avx512bf16_tests 1
> + }
> + }
> + gdb_exit
> + remote_file build delete $obj
> +
> + verbose "$me: returning $skip_avx512bf16_tests" 2
> + return $skip_avx512bf16_tests
> +}
> +
> # Run a test on the target to see if it supports btrace hardware. Return 0 if so,
> # 1 if it does not. Based on 'check_vmx_hw_available' from the GCC testsuite.
>
> diff --git a/gdbsupport/tdesc.cc b/gdbsupport/tdesc.cc
> index 624588b6563..c2a3af700e3 100644
> --- a/gdbsupport/tdesc.cc
> +++ b/gdbsupport/tdesc.cc
> @@ -54,7 +54,8 @@ static tdesc_type_builtin tdesc_predefined_types[] =
> { "ieee_single", TDESC_TYPE_IEEE_SINGLE },
> { "ieee_double", TDESC_TYPE_IEEE_DOUBLE },
> { "arm_fpa_ext", TDESC_TYPE_ARM_FPA_EXT },
> - { "i387_ext", TDESC_TYPE_I387_EXT }
> + { "i387_ext", TDESC_TYPE_I387_EXT },
> + { "bfloat16", TDESC_TYPE_BFLOAT16 }
> };
>
> void tdesc_feature::accept (tdesc_element_visitor &v) const
> diff --git a/gdbsupport/tdesc.h b/gdbsupport/tdesc.h
> index fdc2a6a3708..14b5b5fc9b3 100644
> --- a/gdbsupport/tdesc.h
> +++ b/gdbsupport/tdesc.h
> @@ -173,6 +173,7 @@ enum tdesc_type_kind
> TDESC_TYPE_IEEE_DOUBLE,
> TDESC_TYPE_ARM_FPA_EXT,
> TDESC_TYPE_I387_EXT,
> + TDESC_TYPE_BFLOAT16,
>
> /* Types defined by a target feature. */
> TDESC_TYPE_VECTOR,
> diff --git a/include/floatformat.h b/include/floatformat.h
> index ce8d6d4add8..b3338eb63ad 100644
> --- a/include/floatformat.h
> +++ b/include/floatformat.h
> @@ -133,6 +133,9 @@ extern const struct floatformat floatformat_ia64_quad_little;
> /* IBM long double (double+double). */
> extern const struct floatformat floatformat_ibm_long_double_big;
> extern const struct floatformat floatformat_ibm_long_double_little;
> +/* Intel bfloat16. */
> +extern const struct floatformat floatformat_bfloat16_big;
> +extern const struct floatformat floatformat_bfloat16_little;
>
> /* Convert from FMT to a double.
> FROM is the address of the extended float.
See below; I think the lib/floatformat.h change needs approval from a GCC
maintainer.
> diff --git a/libiberty/floatformat.c b/libiberty/floatformat.c
> index 2fd5e688ec4..6b9b03288e2 100644
> --- a/libiberty/floatformat.c
> +++ b/libiberty/floatformat.c
> @@ -389,7 +389,24 @@ const struct floatformat floatformat_ibm_long_double_little =
> floatformat_ibm_long_double_is_valid,
> &floatformat_ieee_double_little
> };
> -
> +
> +const struct floatformat floatformat_bfloat16_big =
> +{
> + floatformat_big, 16, 0, 1, 8, 127, 255, 9, 7,
> + floatformat_intbit_no,
> + "floatformat_bfloat16_big",
> + floatformat_always_valid,
> + NULL
> +};
> +
> +const struct floatformat floatformat_bfloat16_little =
> +{
> + floatformat_little, 16, 0, 1, 8, 127, 255, 9, 7,
> + floatformat_intbit_no,
> + "floatformat_bfloat16_little",
> + floatformat_always_valid,
> + NULL
> +};
The definitions of floatformat_bfloat16_{bit,little} look correct to
me, but I think it needs approval from someone else. libiberty/README
says to send fixes to gcc-patches@gcc.gnu.org.
More information about the Gdb-patches
mailing list