diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp | 14 |
1 files changed, 6 insertions, 8 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp index 618570de08..76c1d053cd 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp @@ -22,16 +22,14 @@ * SOFTWARE. */ -#if defined(__ARM_FEATURE_SVE) +#if defined(ARM_COMPUTE_ENABLE_SME) template <> void interleave_block<1, 4, VLType::SME, false>( uint8_t * &out, const uint8_t * const *in, - size_t width, size_t height, size_t row_offset, bool first + size_t width, size_t height, size_t row_offset, bool ) { - ARM_COMPUTE_UNUSED(first); - __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" "cntb x21\n" @@ -179,11 +177,11 @@ void interleave_block<1, 4, VLType::SME, false>( "9:" // K loop: Tails: Even: First ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" - "ldr x25, [x26, #0x0]\n" + "ldr x20, [x26, #0x0]\n" "add x12, x12, #0x1\n" ".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" "cmp x12, x9\n" - ".inst 0xe0162322 // ld1b { za0h.b[x13, #2] }, p0/Z, [x25, x22]\n" + ".inst 0xe0162282 // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x22]\n" "add x26, x26, #0x8\n" "addvl x21, x21, #1\n" "add x13, x13, #0x4\n" @@ -200,7 +198,7 @@ void interleave_block<1, 4, VLType::SME, false>( "addvl x21, x21, #1\n" "add x20, x20, #0x4\n" "blt 10b\n" - "whilelt p9.b, x27, %x[width]\n" + "whilelt p8.b, x27, %x[width]\n" "b 13f\n" "11:" // K loop: Tails: Odd "mov x12, #0x0\n" @@ -220,4 +218,4 @@ void interleave_block<1, 4, VLType::SME, false>( ); } -#endif // defined(__ARM_FEATURE_SVE) +#endif // defined(ARM_COMPUTE_ENABLE_SME) |