Thanks. When trying a bootstrap locally I get:
include/arm_neon.h:34709:38: error: cannot convert ‘const __bf16*’ to ‘const
__fp16*’
34709 | __o = __builtin_aarch64_ld1x2v4hf ((const __builtin_aarch64_simd_bf
*) __a);
|
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
| |
| const __bf16*
I think the affected lines are:
Mihail Ionescu <mihail.ione...@foss.arm.com> writes:
> +__extension__ extern __inline bfloat16x4x2_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vld1_bf16_x2 (const bfloat16_t *__a)
> +{
> + bfloat16x4x2_t ret;
> + __builtin_aarch64_simd_oi __o;
> + __o = __builtin_aarch64_ld1x2v4hf ((const __builtin_aarch64_simd_bf *)
> __a);
bf rather than hf here (the error above).
> + ret.val[0] = (bfloat16x4_t) __builtin_aarch64_get_dregoiv4bf (__o, 0);
> + ret.val[1] = (bfloat16x4_t) __builtin_aarch64_get_dregoiv4bf (__o, 1);
> + return ret;
> +}
> [...]
> +__extension__ extern __inline bfloat16x4x2_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vld2_dup_bf16 (const bfloat16_t * __a)
> +{
> + bfloat16x4x2_t ret;
> + __builtin_aarch64_simd_oi __o;
> + __o = __builtin_aarch64_ld2rv4bf ((const __builtin_aarch64_simd_bf *) __a);
> + ret.val[0] = (bfloat16x4_t) __builtin_aarch64_get_dregoiv4hf (__o, 0);
Here too (although the choice is cosmetic).
> + ret.val[1] = (bfloat16x4_t) __builtin_aarch64_get_dregoiv4bf (__o, 1);
> + return ret;
> +}
> +
> +__extension__ extern __inline bfloat16x8x2_t
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vld2q_dup_bf16 (const bfloat16_t * __a)
> +{
> + bfloat16x8x2_t ret;
> + __builtin_aarch64_simd_oi __o;
> + __o = __builtin_aarch64_ld2rv8bf ((const __builtin_aarch64_simd_bf *) __a);
> + ret.val[0] = (bfloat16x8_t) __builtin_aarch64_get_qregoiv8hf (__o, 0);
Same here (again cosmetic).
> + ret.val[1] = (bfloat16x8_t) __builtin_aarch64_get_qregoiv8bf (__o, 1);
> + return ret;
> +}
> [...]
> +__extension__ extern __inline void
> +__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> +vst1_bf16_x3 (bfloat16_t * __a, bfloat16x4x3_t __val)
> +{
> + __builtin_aarch64_simd_ci __o;
> + bfloat16x8x3_t __temp;
> + __temp.val[0] = vcombine_bf16 (__val.val[0], vcreate_bf16
> (__AARCH64_UINT64_C (0)));
> + __temp.val[1] = vcombine_bf16 (__val.val[1], vcreate_bf16
> (__AARCH64_UINT64_C (0)));
> + __temp.val[2] = vcombine_bf16 (__val.val[2], vcreate_bf16
> (__AARCH64_UINT64_C (0)));
> + __o = __builtin_aarch64_set_qregciv8bf (__o, (bfloat16x8_t) __temp.val[0],
> 0);
> + __o = __builtin_aarch64_set_qregciv8bf (__o, (bfloat16x8_t) __temp.val[1],
> 1);
> + __o = __builtin_aarch64_set_qregciv8bf (__o, (bfloat16x8_t) __temp.val[2],
> 2);
> + __builtin_aarch64_st1x3v4hf ((__builtin_aarch64_simd_bf *) __a, __o);
Here too, to avoid:
include/arm_neon.h:35000:32: error: cannot convert ‘__bf16*’ to ‘__fp16*’
35000 | __builtin_aarch64_st1x3v4hf ((__builtin_aarch64_simd_bf *) __a, __o);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
| |
| __bf16*
Looks good otherwise. I guess this shows we should be running the
intrinsics tests for C++ as well as C (in general, not just for this patch).
Richard
