Hi all,
This patch adds support for a new real_format for ARM Brain Floating
Point numbers to the middle end. This is to be used exclusively in the
ARM back-end.
The encode_arm_bfloat_half and decode_arm_bfloat_half functions are
provided to satisfy real_format struct requirements, but are never
intended to be called, which is why they are provided without an
explicit test.
Details on ARM Bfloat can be found here:
https://community.arm.com/developer/ip-products/processors/b/ml-ip-blog/posts/bfloat16-processing-for-neural-networks-on-armv8_2d00_a
Regtested on aarch64-none-elf for sanity.
Is this ok for trunk?
Also, I do not have commit rights, so could someone commit this on my
behalf?
Thank you!
Stam Markianos-Wright
2019-11-13 Stam Markianos-Wright <stam.markianos-wri...@arm.com>
* real.c (struct arm_bfloat_half_format,
encode_arm_bfloat_half, decode_arm_bfloat_half): New.
* real.h (arm_bfloat_half_format): New.
diff --git a/gcc/real.h b/gcc/real.h
index 0f660c9c671..2b337bb7f7d 100644
--- a/gcc/real.h
+++ b/gcc/real.h
@@ -368,6 +368,7 @@ extern const struct real_format decimal_double_format;
extern const struct real_format decimal_quad_format;
extern const struct real_format ieee_half_format;
extern const struct real_format arm_half_format;
+extern const struct real_format arm_bfloat_half_format;
/* ====================================================================== */
diff --git a/gcc/real.c b/gcc/real.c
index 90067f0087b..671a21241d8 100644
--- a/gcc/real.c
+++ b/gcc/real.c
@@ -4799,6 +4799,116 @@ decode_ieee_half (const struct real_format *fmt, REAL_VALUE_TYPE *r,
}
}
+/* Encode arm_bfloat types. */
+static void
+encode_arm_bfloat_half (const struct real_format *fmt, long *buf,
+ const REAL_VALUE_TYPE *r)
+{
+ unsigned long image, sig, exp;
+ unsigned long sign = r->sign;
+ bool denormal = (r->sig[SIGSZ-1] & SIG_MSB) == 0;
+
+ image = sign << 15;
+ sig = (r->sig[SIGSZ-1] >> (HOST_BITS_PER_LONG - 8)) & 0x7f;
+
+ switch (r->cl)
+ {
+ case rvc_zero:
+ break;
+
+ case rvc_inf:
+ if (fmt->has_inf)
+ image |= 255 << 7;
+ else
+ image |= 0x7fff;
+ break;
+
+ case rvc_nan:
+ if (fmt->has_nans)
+ {
+ if (r->canonical)
+ sig = (fmt->canonical_nan_lsbs_set ? (1 << 6) - 1 : 0);
+ if (r->signalling == fmt->qnan_msb_set)
+ sig &= ~(1 << 6);
+ else
+ sig |= 1 << 6;
+ if (sig == 0)
+ sig = 1 << 5;
+
+ image |= 255 << 7;
+ image |= sig;
+ }
+ else
+ image |= 0x7fff;
+ break;
+
+ case rvc_normal:
+ if (denormal)
+ exp = 0;
+ else
+ exp = REAL_EXP (r) + 127 - 1;
+ image |= exp << 7;
+ image |= sig;
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ buf[0] = image;
+}
+
+/* Decode arm_bfloat types. */
+static void
+decode_arm_bfloat_half (const struct real_format *fmt, REAL_VALUE_TYPE *r,
+ const long *buf)
+{
+ unsigned long image = buf[0] & 0xffff;
+ bool sign = (image >> 15) & 1;
+ int exp = (image >> 7) & 0xff;
+
+ memset (r, 0, sizeof (*r));
+ image <<= HOST_BITS_PER_LONG - 8;
+ image &= ~SIG_MSB;
+
+ if (exp == 0)
+ {
+ if (image && fmt->has_denorm)
+ {
+ r->cl = rvc_normal;
+ r->sign = sign;
+ SET_REAL_EXP (r, -126);
+ r->sig[SIGSZ-1] = image << 1;
+ normalize (r);
+ }
+ else if (fmt->has_signed_zero)
+ r->sign = sign;
+ }
+ else if (exp == 255 && (fmt->has_nans || fmt->has_inf))
+ {
+ if (image)
+ {
+ r->cl = rvc_nan;
+ r->sign = sign;
+ r->signalling = (((image >> (HOST_BITS_PER_LONG - 2)) & 1)
+ ^ fmt->qnan_msb_set);
+ r->sig[SIGSZ-1] = image;
+ }
+ else
+ {
+ r->cl = rvc_inf;
+ r->sign = sign;
+ }
+ }
+ else
+ {
+ r->cl = rvc_normal;
+ r->sign = sign;
+ SET_REAL_EXP (r, exp - 127 + 1);
+ r->sig[SIGSZ-1] = image | SIG_MSB;
+ }
+}
+
/* Half-precision format, as specified in IEEE 754R. */
const struct real_format ieee_half_format =
{
@@ -4848,6 +4958,33 @@ const struct real_format arm_half_format =
false,
"arm_half"
};
+
+/* ARM Bfloat half-precision format. This format resembles a truncated
+ (16-bit) version of the 32-bit IEEE 754 single-precision floating-point
+ format. */
+const struct real_format arm_bfloat_half_format =
+ {
+ encode_arm_bfloat_half,
+ decode_arm_bfloat_half,
+ 2,
+ 8,
+ 8,
+ -125,
+ 128,
+ 15,
+ 15,
+ 0,
+ false,
+ true,
+ true,
+ true,
+ true,
+ true,
+ true,
+ false,
+ "bfloat_half"
+ };
+
/* A synthetic "format" for internal arithmetic. It's the size of the
internal significand minus the two bits needed for proper rounding.
