diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp | 144 |
1 files changed, 57 insertions, 87 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp index 0bfc28776f..be6d5b901d 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp @@ -158,7 +158,6 @@ void sve_hybrid_u8qa_dot_4x4VL ( "tbnz %x[flags], #31, 8f\n" "udot z11.s, z0.b, z15.b\n" "8:" // Height 1: Multiply loop: unique 1: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" "sub x24, x24, #0x10\n" "cmp x24, #0x10\n" "bgt 7b\n" @@ -170,7 +169,6 @@ void sve_hybrid_u8qa_dot_4x4VL ( "ld1rqb { z0.b }, p0/Z, [x23]\n" "udot z16.s, z4.b, z0.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "add x23, x23, #0x10\n" "udot z17.s, z5.b, z0.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "addvl x28, x28, #4\n" @@ -212,9 +210,8 @@ void sve_hybrid_u8qa_dot_4x4VL ( "tbnz %x[flags], #31, 11f\n" "udot z11.s, z0.b, z15.b\n" "11:" // Height 1: Multiply loop: unique 2: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" - "add x25, x25, #0x1\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x25, x25, #0x1\n" "cmp x25, x19\n" "bne 4b\n" "tbnz %x[flags], #31, 12f\n" @@ -251,16 +248,16 @@ void sve_hybrid_u8qa_dot_4x4VL ( ".inst 0x04a47673 // sqrdmulh z19.s, z19.s, z4.s\n" "tbz %x[flags], #5, 13f\n" "and z4.d, z16.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z17.d, z0.d\n" "and z6.d, z18.d, z0.d\n" - "asr z5.s, z5.s, #0x1f\n" "and z7.d, z19.d, z0.d\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z16.s, z16.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z17.s, z17.s, z5.s\n" "sqadd z18.s, z18.s, z6.s\n" + "asr z7.s, z7.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" "13:" // Height 1: no shift correction ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" @@ -396,9 +393,7 @@ void sve_hybrid_u8qa_dot_4x4VL ( "udot z11.s, z0.b, z15.b\n" "udot z12.s, z1.b, z15.b\n" "22:" // Height 2: Multiply loop: unique 3: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" "sub x24, x24, #0x10\n" - "prfm pldl1keep, [x22, #0x80]\n" "cmp x24, #0x10\n" "bgt 21b\n" "23:" // Height 2: Multiply loop: Single iteration only @@ -409,12 +404,10 @@ void sve_hybrid_u8qa_dot_4x4VL ( "ld1rqb { z0.b }, p0/Z, [x23]\n" "udot z16.s, z4.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" "udot z17.s, z5.b, z0.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "add x22, x22, #0x10\n" - "udot z20.s, z4.b, z1.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" + "udot z20.s, z4.b, z1.b[0]\n" "addvl x28, x28, #4\n" "udot z21.s, z5.b, z1.b[0]\n" "udot z18.s, z6.b, z0.b[0]\n" @@ -470,10 +463,8 @@ void sve_hybrid_u8qa_dot_4x4VL ( "udot z11.s, z0.b, z15.b\n" "udot z12.s, z1.b, z15.b\n" "25:" // Height 2: Multiply loop: unique 4: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" - "add x25, x25, #0x1\n" - "prfm pldl1keep, [x22, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x25, x25, #0x1\n" "cmp x25, x19\n" "bne 18b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -527,27 +518,27 @@ void sve_hybrid_u8qa_dot_4x4VL ( ".inst 0x04a476f7 // sqrdmulh z23.s, z23.s, z4.s\n" "tbz %x[flags], #5, 27f\n" "and z4.d, z16.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z17.d, z0.d\n" "and z6.d, z18.d, z0.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" + "sqadd z16.s, z16.s, z4.s\n" + "sqadd z17.s, z17.s, z5.s\n" + "sqadd z18.s, z18.s, z6.s\n" "and z7.d, z19.d, z0.d\n" "and z8.d, z20.d, z0.d\n" - "asr z6.s, z6.s, #0x1f\n" "and z9.d, z21.d, z0.d\n" "asr z7.s, z7.s, #0x1f\n" - "sqadd z16.s, z16.s, z4.s\n" - "and z10.d, z22.d, z0.d\n" "asr z8.s, z8.s, #0x1f\n" - "and z4.d, z23.d, z0.d\n" "asr z9.s, z9.s, #0x1f\n" - "sqadd z17.s, z17.s, z5.s\n" - "asr z10.s, z10.s, #0x1f\n" - "sqadd z18.s, z18.s, z6.s\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" "sqadd z20.s, z20.s, z8.s\n" "sqadd z21.s, z21.s, z9.s\n" + "and z10.d, z22.d, z0.d\n" + "and z4.d, z23.d, z0.d\n" + "asr z10.s, z10.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "sqadd z22.s, z22.s, z10.s\n" "sqadd z23.s, z23.s, z4.s\n" "27:" // Height 2: no shift correction @@ -731,11 +722,8 @@ void sve_hybrid_u8qa_dot_4x4VL ( "udot z12.s, z1.b, z15.b\n" "udot z13.s, z2.b, z15.b\n" "36:" // Height 3: Multiply loop: unique 5: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" "sub x24, x24, #0x10\n" - "prfm pldl1keep, [x22, #0x80]\n" "cmp x24, #0x10\n" - "prfm pldl1keep, [x21, #0x80]\n" "bgt 35b\n" "37:" // Height 3: Multiply loop: Single iteration only "ld1b { z4.b }, p2/Z, [x28]\n" @@ -745,16 +733,13 @@ void sve_hybrid_u8qa_dot_4x4VL ( "ld1rqb { z0.b }, p0/Z, [x23]\n" "udot z16.s, z4.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" "udot z17.s, z5.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" - "udot z20.s, z4.b, z1.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "add x21, x21, #0x10\n" - "udot z24.s, z4.b, z2.b[0]\n" + "udot z20.s, z4.b, z1.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "addvl x28, x28, #4\n" + "udot z24.s, z4.b, z2.b[0]\n" "udot z21.s, z5.b, z1.b[0]\n" "udot z25.s, z5.b, z2.b[0]\n" "udot z18.s, z6.b, z0.b[0]\n" @@ -825,11 +810,8 @@ void sve_hybrid_u8qa_dot_4x4VL ( "udot z12.s, z1.b, z15.b\n" "udot z13.s, z2.b, z15.b\n" "39:" // Height 3: Multiply loop: unique 6: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" - "add x25, x25, #0x1\n" - "prfm pldl1keep, [x22, #0x80]\n" - "prfm pldl1keep, [x21, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x25, x25, #0x1\n" "cmp x25, x19\n" "bne 32b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -899,39 +881,39 @@ void sve_hybrid_u8qa_dot_4x4VL ( ".inst 0x04a4777b // sqrdmulh z27.s, z27.s, z4.s\n" "tbz %x[flags], #5, 41f\n" "and z4.d, z16.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z17.d, z0.d\n" "and z6.d, z18.d, z0.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" + "sqadd z16.s, z16.s, z4.s\n" + "sqadd z17.s, z17.s, z5.s\n" + "sqadd z18.s, z18.s, z6.s\n" "and z7.d, z19.d, z0.d\n" "and z8.d, z20.d, z0.d\n" - "asr z6.s, z6.s, #0x1f\n" "and z9.d, z21.d, z0.d\n" "asr z7.s, z7.s, #0x1f\n" - "sqadd z16.s, z16.s, z4.s\n" - "and z10.d, z22.d, z0.d\n" "asr z8.s, z8.s, #0x1f\n" - "and z4.d, z23.d, z0.d\n" "asr z9.s, z9.s, #0x1f\n" - "sqadd z17.s, z17.s, z5.s\n" - "asr z10.s, z10.s, #0x1f\n" - "sqadd z18.s, z18.s, z6.s\n" - "asr z4.s, z4.s, #0x1f\n" - "and z5.d, z24.d, z0.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" "sqadd z20.s, z20.s, z8.s\n" "sqadd z21.s, z21.s, z9.s\n" + "and z10.d, z22.d, z0.d\n" + "and z4.d, z23.d, z0.d\n" + "and z5.d, z24.d, z0.d\n" + "asr z10.s, z10.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "sqadd z22.s, z22.s, z10.s\n" "sqadd z23.s, z23.s, z4.s\n" - "and z6.d, z25.d, z0.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z24.s, z24.s, z5.s\n" + "and z6.d, z25.d, z0.d\n" "and z7.d, z26.d, z0.d\n" - "asr z7.s, z7.s, #0x1f\n" "and z8.d, z27.d, z0.d\n" - "sqadd z25.s, z25.s, z6.s\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z8.s, z8.s, #0x1f\n" + "sqadd z25.s, z25.s, z6.s\n" "sqadd z26.s, z26.s, z7.s\n" "sqadd z27.s, z27.s, z8.s\n" "41:" // Height 3: no shift correction @@ -1165,12 +1147,8 @@ void sve_hybrid_u8qa_dot_4x4VL ( "udot z13.s, z2.b, z15.b\n" "udot z14.s, z3.b, z15.b\n" "50:" // Height 4: Multiply loop: unique 7: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" "sub x24, x24, #0x10\n" - "prfm pldl1keep, [x22, #0x80]\n" "cmp x24, #0x10\n" - "prfm pldl1keep, [x21, #0x80]\n" - "prfm pldl1keep, [x20, #0x80]\n" "bgt 49b\n" "51:" // Height 4: Multiply loop: Single iteration only "ld1b { z4.b }, p2/Z, [x28]\n" @@ -1180,19 +1158,15 @@ void sve_hybrid_u8qa_dot_4x4VL ( "ld1rqb { z0.b }, p0/Z, [x23]\n" "udot z16.s, z4.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" "udot z17.s, z5.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" - "udot z20.s, z4.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x20]\n" - "add x21, x21, #0x10\n" - "udot z24.s, z4.b, z2.b[0]\n" + "udot z20.s, z4.b, z1.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "add x20, x20, #0x10\n" "udot z21.s, z5.b, z1.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "addvl x28, x28, #4\n" + "udot z24.s, z4.b, z2.b[0]\n" "udot z28.s, z4.b, z3.b[0]\n" "udot z25.s, z5.b, z2.b[0]\n" "udot z29.s, z5.b, z3.b[0]\n" @@ -1279,12 +1253,8 @@ void sve_hybrid_u8qa_dot_4x4VL ( "udot z13.s, z2.b, z15.b\n" "udot z14.s, z3.b, z15.b\n" "53:" // Height 4: Multiply loop: unique 8: skip row sum - "prfm pldl1keep, [x23, #0x80]\n" - "add x25, x25, #0x1\n" - "prfm pldl1keep, [x22, #0x80]\n" - "prfm pldl1keep, [x21, #0x80]\n" - "prfm pldl1keep, [x20, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x25, x25, #0x1\n" "cmp x25, x19\n" "bne 46b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1370,52 +1340,52 @@ void sve_hybrid_u8qa_dot_4x4VL ( ".inst 0x04a477ff // sqrdmulh z31.s, z31.s, z4.s\n" "tbz %x[flags], #5, 55f\n" "and z4.d, z16.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z17.d, z0.d\n" "and z6.d, z18.d, z0.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" + "sqadd z16.s, z16.s, z4.s\n" + "sqadd z17.s, z17.s, z5.s\n" + "sqadd z18.s, z18.s, z6.s\n" "and z7.d, z19.d, z0.d\n" "and z8.d, z20.d, z0.d\n" - "asr z6.s, z6.s, #0x1f\n" "and z9.d, z21.d, z0.d\n" "asr z7.s, z7.s, #0x1f\n" - "sqadd z16.s, z16.s, z4.s\n" - "and z10.d, z22.d, z0.d\n" "asr z8.s, z8.s, #0x1f\n" - "and z4.d, z23.d, z0.d\n" "asr z9.s, z9.s, #0x1f\n" - "sqadd z17.s, z17.s, z5.s\n" - "asr z10.s, z10.s, #0x1f\n" - "sqadd z18.s, z18.s, z6.s\n" - "asr z4.s, z4.s, #0x1f\n" - "and z5.d, z24.d, z0.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" "sqadd z20.s, z20.s, z8.s\n" "sqadd z21.s, z21.s, z9.s\n" + "and z10.d, z22.d, z0.d\n" + "and z4.d, z23.d, z0.d\n" + "and z5.d, z24.d, z0.d\n" + "asr z10.s, z10.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "sqadd z22.s, z22.s, z10.s\n" "sqadd z23.s, z23.s, z4.s\n" - "and z6.d, z25.d, z0.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z24.s, z24.s, z5.s\n" + "and z6.d, z25.d, z0.d\n" "and z7.d, z26.d, z0.d\n" - "asr z7.s, z7.s, #0x1f\n" "and z8.d, z27.d, z0.d\n" - "and z9.d, z28.d, z0.d\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z8.s, z8.s, #0x1f\n" "sqadd z25.s, z25.s, z6.s\n" + "sqadd z26.s, z26.s, z7.s\n" + "sqadd z27.s, z27.s, z8.s\n" + "and z9.d, z28.d, z0.d\n" "and z10.d, z29.d, z0.d\n" - "asr z9.s, z9.s, #0x1f\n" "and z4.d, z30.d, z0.d\n" + "asr z9.s, z9.s, #0x1f\n" "asr z10.s, z10.s, #0x1f\n" - "sqadd z26.s, z26.s, z7.s\n" - "and z5.d, z31.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" - "sqadd z27.s, z27.s, z8.s\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z28.s, z28.s, z9.s\n" "sqadd z29.s, z29.s, z10.s\n" "sqadd z30.s, z30.s, z4.s\n" + "and z5.d, z31.d, z0.d\n" + "asr z5.s, z5.s, #0x1f\n" "sqadd z31.s, z31.s, z5.s\n" "55:" // Height 4: no shift correction ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" @@ -1529,4 +1499,4 @@ void sve_hybrid_u8qa_dot_4x4VL ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // __ARM_FEATURE_SVE |