diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp | 343 |
1 files changed, 140 insertions, 203 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp index 0328c107e2..6b08d2834b 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -135,13 +135,12 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" - "cmp x26, #0x10\n" + "add x25, x25, #0x10\n" "sdot z10.s, z6.b, z0.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" - "prfm pldl1keep, [x25, #0x80]\n" "sdot z11.s, z7.b, z0.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "sdot z8.s, z6.b, z0.b[1]\n" @@ -176,7 +175,6 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "addvl x28, x28, #4\n" @@ -215,9 +213,8 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z10.s, z6.b, z0.b[3]\n" "sdot z11.s, z7.b, z0.b[3]\n" "9:" // Height 1: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 4b\n" "ld1w { z0.s }, p2/Z, [x11]\n" @@ -259,16 +256,16 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n" "tbz %x[flags], #5, 12f\n" "and z4.d, z8.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" "and z6.d, z10.d, z2.d\n" - "asr z5.s, z5.s, #0x1f\n" "and z7.d, z11.d, z3.d\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z8.s, z8.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z9.s, z9.s, z5.s\n" "sqadd z10.s, z10.s, z6.s\n" + "asr z7.s, z7.s, #0x1f\n" "sqadd z11.s, z11.s, z7.s\n" "12:" // Height 1: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" @@ -351,16 +348,14 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "cmp x26, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" - "prfm pldl1keep, [x24, #0x80]\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #4, MUL VL]\n" @@ -411,9 +406,7 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" - "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" "sdot z13.s, z7.b, z1.b[0]\n" @@ -468,10 +461,8 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z11.s, z7.b, z0.b[3]\n" "sdot z15.s, z7.b, z1.b[3]\n" "22:" // Height 2: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 17b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -523,27 +514,27 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a775ef // sqrdmulh z15.s, z15.s, z7.s\n" "tbz %x[flags], #5, 25f\n" "and z4.d, z8.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" "and z6.d, z10.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" - "and z7.d, z11.d, z3.d\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z8.s, z8.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" - "and z4.d, z12.d, z0.d\n" "sqadd z9.s, z9.s, z5.s\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z10.s, z10.s, z6.s\n" + "and z7.d, z11.d, z3.d\n" + "and z4.d, z12.d, z0.d\n" "and z5.d, z13.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z11.s, z11.s, z7.s\n" - "and z6.d, z14.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z12.s, z12.s, z4.s\n" + "sqadd z13.s, z13.s, z5.s\n" + "and z6.d, z14.d, z2.d\n" "and z7.d, z15.d, z3.d\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" - "sqadd z13.s, z13.s, z5.s\n" "sqadd z14.s, z14.s, z6.s\n" "sqadd z15.s, z15.s, z7.s\n" "25:" // Height 2: no shift correction @@ -654,21 +645,18 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" + "sdot z13.s, z7.b, z1.b[0]\n" "add x23, x23, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "cmp x26, #0x10\n" - "sdot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" "sdot z17.s, z7.b, z2.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" - "prfm pldl1keep, [x23, #0x80]\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" @@ -733,12 +721,9 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" - "add x23, x23, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "sdot z16.s, z6.b, z2.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" @@ -808,11 +793,8 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z15.s, z7.b, z1.b[3]\n" "sdot z19.s, z7.b, z2.b[3]\n" "35:" // Height 3: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 30b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -873,37 +855,37 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a77673 // sqrdmulh z19.s, z19.s, z7.s\n" "tbz %x[flags], #5, 38f\n" "and z4.d, z8.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" "and z6.d, z10.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" - "and z7.d, z11.d, z3.d\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z8.s, z8.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" - "and z4.d, z12.d, z0.d\n" "sqadd z9.s, z9.s, z5.s\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z10.s, z10.s, z6.s\n" + "and z7.d, z11.d, z3.d\n" + "and z4.d, z12.d, z0.d\n" "and z5.d, z13.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z11.s, z11.s, z7.s\n" - "and z6.d, z14.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z12.s, z12.s, z4.s\n" - "and z7.d, z15.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z13.s, z13.s, z5.s\n" + "and z6.d, z14.d, z2.d\n" + "and z7.d, z15.d, z3.d\n" "and z4.d, z16.d, z0.d\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z4.s, z4.s, #0x1f\n" "sqadd z14.s, z14.s, z6.s\n" - "and z5.d, z17.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z15.s, z15.s, z7.s\n" - "and z6.d, z18.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z16.s, z16.s, z4.s\n" + "and z5.d, z17.d, z1.d\n" + "and z6.d, z18.d, z2.d\n" "and z7.d, z19.d, z3.d\n" + "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" "sqadd z17.s, z17.s, z5.s\n" "sqadd z18.s, z18.s, z6.s\n" @@ -1043,26 +1025,22 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" + "add x24, x24, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x10\n" + "add x22, x22, #0x10\n" + "sdot z17.s, z7.b, z2.b[0]\n" "sdot z20.s, z6.b, z3.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "sdot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "sdot z21.s, z7.b, z3.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" - "prfm pldl1keep, [x22, #0x80]\n" "sdot z10.s, z6.b, z0.b[0]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" @@ -1141,19 +1119,15 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "sdot z16.s, z6.b, z2.b[0]\n" - "add x22, x22, #0x10\n" + "sdot z12.s, z6.b, z1.b[0]\n" "sdot z13.s, z7.b, z1.b[0]\n" - "sdot z17.s, z7.b, z2.b[0]\n" + "sdot z16.s, z6.b, z2.b[0]\n" "sdot z20.s, z6.b, z3.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "sdot z17.s, z7.b, z2.b[0]\n" "sdot z21.s, z7.b, z3.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "addvl x28, x28, #4\n" @@ -1234,12 +1208,8 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z19.s, z7.b, z2.b[3]\n" "sdot z23.s, z7.b, z3.b[3]\n" "48:" // Height 4: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 43b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1309,52 +1279,52 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a776f7 // sqrdmulh z23.s, z23.s, z7.s\n" "tbz %x[flags], #5, 51f\n" "and z4.d, z8.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" "and z6.d, z10.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" - "and z7.d, z11.d, z3.d\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z8.s, z8.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" - "and z4.d, z12.d, z0.d\n" "sqadd z9.s, z9.s, z5.s\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z10.s, z10.s, z6.s\n" + "and z7.d, z11.d, z3.d\n" + "and z4.d, z12.d, z0.d\n" "and z5.d, z13.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z11.s, z11.s, z7.s\n" - "and z6.d, z14.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z12.s, z12.s, z4.s\n" - "and z7.d, z15.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z13.s, z13.s, z5.s\n" + "and z6.d, z14.d, z2.d\n" + "and z7.d, z15.d, z3.d\n" "and z4.d, z16.d, z0.d\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z4.s, z4.s, #0x1f\n" "sqadd z14.s, z14.s, z6.s\n" - "and z5.d, z17.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z15.s, z15.s, z7.s\n" - "and z6.d, z18.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z16.s, z16.s, z4.s\n" + "and z5.d, z17.d, z1.d\n" + "and z6.d, z18.d, z2.d\n" "and z7.d, z19.d, z3.d\n" + "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" "sqadd z17.s, z17.s, z5.s\n" - "and z4.d, z20.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z18.s, z18.s, z6.s\n" - "and z5.d, z21.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" + "and z4.d, z20.d, z0.d\n" + "and z5.d, z21.d, z1.d\n" "and z6.d, z22.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z20.s, z20.s, z4.s\n" - "and z7.d, z23.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z21.s, z21.s, z5.s\n" "sqadd z22.s, z22.s, z6.s\n" + "and z7.d, z23.d, z3.d\n" + "asr z7.s, z7.s, #0x1f\n" "sqadd z23.s, z23.s, z7.s\n" "51:" // Height 4: no shift correction ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n" @@ -1518,32 +1488,27 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" + "add x24, x24, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" "ld1rqb { z4.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x22, x22, #0x10\n" + "sdot z17.s, z7.b, z2.b[0]\n" "add x21, x21, #0x10\n" "sdot z20.s, z6.b, z3.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x10\n" "sdot z24.s, z6.b, z4.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "sdot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "sdot z21.s, z7.b, z3.b[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" "sdot z25.s, z7.b, z4.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "sdot z10.s, z6.b, z0.b[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" "sdot z22.s, z6.b, z3.b[0]\n" @@ -1635,22 +1600,17 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "sdot z16.s, z6.b, z2.b[0]\n" + "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z4.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" - "add x21, x21, #0x10\n" - "sdot z17.s, z7.b, z2.b[0]\n" + "sdot z16.s, z6.b, z2.b[0]\n" "sdot z20.s, z6.b, z3.b[0]\n" "sdot z24.s, z6.b, z4.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "sdot z17.s, z7.b, z2.b[0]\n" "sdot z21.s, z7.b, z3.b[0]\n" "sdot z25.s, z7.b, z4.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" @@ -1746,13 +1706,8 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z23.s, z7.b, z3.b[3]\n" "sdot z27.s, z7.b, z4.b[3]\n" "61:" // Height 5: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "prfm pldl1keep, [x21, #0x80]\n" "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x27, x27, #0x1\n" "cmp x27, x19\n" "bne 56b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1831,63 +1786,63 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a7777b // sqrdmulh z27.s, z27.s, z7.s\n" "tbz %x[flags], #5, 64f\n" "and z4.d, z8.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" "and z6.d, z10.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" - "and z7.d, z11.d, z3.d\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z8.s, z8.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" - "and z4.d, z12.d, z0.d\n" "sqadd z9.s, z9.s, z5.s\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z10.s, z10.s, z6.s\n" + "and z7.d, z11.d, z3.d\n" + "and z4.d, z12.d, z0.d\n" "and z5.d, z13.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z11.s, z11.s, z7.s\n" - "and z6.d, z14.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z12.s, z12.s, z4.s\n" - "and z7.d, z15.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z13.s, z13.s, z5.s\n" + "and z6.d, z14.d, z2.d\n" + "and z7.d, z15.d, z3.d\n" "and z4.d, z16.d, z0.d\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z4.s, z4.s, #0x1f\n" "sqadd z14.s, z14.s, z6.s\n" - "and z5.d, z17.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z15.s, z15.s, z7.s\n" - "and z6.d, z18.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z16.s, z16.s, z4.s\n" + "and z5.d, z17.d, z1.d\n" + "and z6.d, z18.d, z2.d\n" "and z7.d, z19.d, z3.d\n" + "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" "sqadd z17.s, z17.s, z5.s\n" - "and z4.d, z20.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z18.s, z18.s, z6.s\n" - "and z5.d, z21.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" + "and z4.d, z20.d, z0.d\n" + "and z5.d, z21.d, z1.d\n" "and z6.d, z22.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z20.s, z20.s, z4.s\n" - "and z7.d, z23.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z21.s, z21.s, z5.s\n" - "and z4.d, z24.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z22.s, z22.s, z6.s\n" + "and z7.d, z23.d, z3.d\n" + "and z4.d, z24.d, z0.d\n" "and z5.d, z25.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z23.s, z23.s, z7.s\n" - "and z6.d, z26.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z24.s, z24.s, z4.s\n" + "sqadd z25.s, z25.s, z5.s\n" + "and z6.d, z26.d, z2.d\n" "and z7.d, z27.d, z3.d\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" - "sqadd z25.s, z25.s, z5.s\n" "sqadd z26.s, z26.s, z6.s\n" "sqadd z27.s, z27.s, z7.s\n" "64:" // Height 5: no shift correction @@ -2082,37 +2037,31 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" + "add x25, x25, #0x10\n" "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" + "add x24, x24, #0x10\n" "sdot z16.s, z6.b, z2.b[0]\n" "ld1rqb { z4.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "ld1rqb { z5.b }, p0/Z, [x20]\n" - "add x21, x21, #0x10\n" + "add x22, x22, #0x10\n" "sdot z20.s, z6.b, z3.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x21, x21, #0x10\n" + "sdot z17.s, z7.b, z2.b[0]\n" "add x20, x20, #0x10\n" "sdot z24.s, z6.b, z4.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x10\n" "sdot z28.s, z6.b, z5.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" - "sdot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "sdot z21.s, z7.b, z3.b[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" "sdot z25.s, z7.b, z4.b[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" "sdot z29.s, z7.b, z5.b[0]\n" "ld1b { z7.b }, p2/Z, [x28, #3, MUL VL]\n" "sdot z10.s, z6.b, z0.b[0]\n" - "prfm pldl1keep, [x20, #0x80]\n" "sdot z14.s, z6.b, z1.b[0]\n" "sdot z18.s, z6.b, z2.b[0]\n" "sdot z22.s, z6.b, z3.b[0]\n" @@ -2218,25 +2167,19 @@ void sve_hybrid_s8qs_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "sdot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "sdot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "sdot z16.s, z6.b, z2.b[0]\n" + "sdot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z4.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" "sdot z13.s, z7.b, z1.b[0]\n" "ld1rqb { z5.b }, p0/Z, [x20]\n" - "add x21, x21, #0x10\n" + "sdot z16.s, z6.b, z2.b[0]\n" "sdot z20.s, z6.b, z3.b[0]\n" - "add x20, x20, #0x10\n" - "sdot z17.s, z7.b, z2.b[0]\n" "sdot z24.s, z6.b, z4.b[0]\n" "sdot z28.s, z6.b, z5.b[0]\n" "ld1b { z6.b }, p2/Z, [x28, #2, MUL VL]\n" + "sdot z17.s, z7.b, z2.b[0]\n" "sdot z21.s, z7.b, z3.b[0]\n" "sdot z25.s, z7.b, z4.b[0]\n" "sdot z29.s, z7.b, z5.b[0]\n" @@ -2347,14 +2290,8 @@ void sve_hybrid_s8qs_dot_6x4VL ( "sdot z27.s, z7.b, z4.b[3]\n" "sdot z31.s, z7.b, z5.b[3]\n" "74:" // Height 6: Multiply loop: multiply skip - "prfm pldl1keep, [x25, #0x80]\n" - "add x27, x27, #0x1\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\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 x27, x27, #0x1\n" "cmp x27, x19\n" "bne 69b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -2442,73 +2379,73 @@ void sve_hybrid_s8qs_dot_6x4VL ( ".inst 0x04a777ff // sqrdmulh z31.s, z31.s, z7.s\n" "tbz %x[flags], #5, 77f\n" "and z4.d, z8.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "and z5.d, z9.d, z1.d\n" "and z6.d, z10.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" - "and z7.d, z11.d, z3.d\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z8.s, z8.s, z4.s\n" - "asr z7.s, z7.s, #0x1f\n" - "and z4.d, z12.d, z0.d\n" "sqadd z9.s, z9.s, z5.s\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z10.s, z10.s, z6.s\n" + "and z7.d, z11.d, z3.d\n" + "and z4.d, z12.d, z0.d\n" "and z5.d, z13.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z11.s, z11.s, z7.s\n" - "and z6.d, z14.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z12.s, z12.s, z4.s\n" - "and z7.d, z15.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z13.s, z13.s, z5.s\n" + "and z6.d, z14.d, z2.d\n" + "and z7.d, z15.d, z3.d\n" "and z4.d, z16.d, z0.d\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z4.s, z4.s, #0x1f\n" "sqadd z14.s, z14.s, z6.s\n" - "and z5.d, z17.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z15.s, z15.s, z7.s\n" - "and z6.d, z18.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z16.s, z16.s, z4.s\n" + "and z5.d, z17.d, z1.d\n" + "and z6.d, z18.d, z2.d\n" "and z7.d, z19.d, z3.d\n" + "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" "sqadd z17.s, z17.s, z5.s\n" - "and z4.d, z20.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z18.s, z18.s, z6.s\n" - "and z5.d, z21.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z19.s, z19.s, z7.s\n" + "and z4.d, z20.d, z0.d\n" + "and z5.d, z21.d, z1.d\n" "and z6.d, z22.d, z2.d\n" + "asr z4.s, z4.s, #0x1f\n" + "asr z5.s, z5.s, #0x1f\n" "asr z6.s, z6.s, #0x1f\n" "sqadd z20.s, z20.s, z4.s\n" - "and z7.d, z23.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z21.s, z21.s, z5.s\n" - "and z4.d, z24.d, z0.d\n" - "asr z4.s, z4.s, #0x1f\n" "sqadd z22.s, z22.s, z6.s\n" + "and z7.d, z23.d, z3.d\n" + "and z4.d, z24.d, z0.d\n" "and z5.d, z25.d, z1.d\n" + "asr z7.s, z7.s, #0x1f\n" + "asr z4.s, z4.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" "sqadd z23.s, z23.s, z7.s\n" - "and z6.d, z26.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z24.s, z24.s, z4.s\n" - "and z7.d, z27.d, z3.d\n" - "asr z7.s, z7.s, #0x1f\n" "sqadd z25.s, z25.s, z5.s\n" + "and z6.d, z26.d, z2.d\n" + "and z7.d, z27.d, z3.d\n" "and z4.d, z28.d, z0.d\n" + "asr z6.s, z6.s, #0x1f\n" + "asr z7.s, z7.s, #0x1f\n" "asr z4.s, z4.s, #0x1f\n" "sqadd z26.s, z26.s, z6.s\n" - "and z5.d, z29.d, z1.d\n" - "asr z5.s, z5.s, #0x1f\n" "sqadd z27.s, z27.s, z7.s\n" - "and z6.d, z30.d, z2.d\n" - "asr z6.s, z6.s, #0x1f\n" "sqadd z28.s, z28.s, z4.s\n" + "and z5.d, z29.d, z1.d\n" + "and z6.d, z30.d, z2.d\n" "and z7.d, z31.d, z3.d\n" + "asr z5.s, z5.s, #0x1f\n" + "asr z6.s, z6.s, #0x1f\n" "asr z7.s, z7.s, #0x1f\n" "sqadd z29.s, z29.s, z5.s\n" "sqadd z30.s, z30.s, z6.s\n" @@ -2665,4 +2602,4 @@ void sve_hybrid_s8qs_dot_6x4VL ( } } // namespace arm_gemm -#endif // ARM_COMPUTE_ENABLE_SVE +#endif // __ARM_FEATURE_SVE |