Mihail Ionescu <mihail.ione...@foss.arm.com> writes:
> Hi Richard,
>
> On 02/18/2020 05:06 PM, Richard Sandiford wrote:
>> 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
>>
>
> Thanks for pointing this out.
> I've updated the patch with the fixes.
> The ChangeLog is the same.
LGTM thanks, pushed.
Richard
