[BUGFIX] Fix neon2rvv failing tests #309
Conversation
neon2rvv.h
Outdated
| @@ -3846,127 +3845,127 @@ FORCE_INLINE uint64x2_t vrsraq_n_u64(uint64x2_t __a, uint64x2_t __b, const int _ | |||
| } | |||
|
|
|||
| FORCE_INLINE int8x8_t vsri_n_s8(int8x8_t __a, int8x8_t __b, const int __c) { | |||
| uint8_t mask = UINT8_MAX >> __c; | |||
| uint8_t mask = (__c == 8) ? 0 : UINT8_MAX >> __c; | |||
There was a problem hiding this comment.
how about using
+uint8_t mask = (uint64_t) UINT8_MAX >> __c;
-uint8_t mask = (__c == 8) ? 0 : UINT8_MAX >> __c;I think in this way we may avoid branching?
There was a problem hiding this comment.
Hmm except the cast won't work for the 64-bit data types, better to keep things uniform between all implementations?
There was a problem hiding this comment.
Umm I feel ok for this part, because there are already some exception happened in 64 bits cases. Therefore if the feedback is good enough I am willing to change this.
How do you think about it? If you think this is unnecessary then I will directly merge this PR
There was a problem hiding this comment.
Okay, made the change for 8/16/32 versions.
|
Thank you so much for the contribution, and helping me to resolve my confusion! Except the question I raised above, all the rest of changes are amazing |
Hello, made the following changes to fix failing tests:
vcnt_u8- correctly load entire look-up tablevsri- right shifting an n-bit value by n bits is undefined behaviour, in such cases mask should be set to 0 to copy all bits from input __avst1_lane- strided store causes errors,vmv_x_scan be used to copy first vector element to scalar register insteadvdup- Noticed possible confusion, the use ofvmv_s_xis to modify the first element in the vector regiser,vmv_v_xshould be used to broadcast copy to all vector elements. Also fixed return-type in forward declaration forvdupq, and corrected usage ofvdupqtovdupinside 64-bit vector intrinsics.Test suite executed with QEMU and Spike.
Using GCC(d6479050ecef10fd5e67b4da989229e4cfac53ee) all tests passing on optimization level 0 to 3.
Using Clang(6008cd40b7bbdc66555550c2e38648d5ce99cc78) all tests passing on optimization level 0 and 1. However on -O2 Clang initially crashed - raised an issue. After fix compilation no longer crashes, but almost half of the tests fail. Further investigation needed.