diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp | 149 |
1 files changed, 43 insertions, 106 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp index 413bc65288..fc8bdb50a9 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -137,13 +137,12 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" - "cmp x26, #0x10\n" + "add x25, x25, #0x10\n" "udot z10.s, z6.b, z0.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #4, MUL VL]\n" - "prfm pldl1keep, [x25, #0x80]\n" "udot z11.s, z7.b, z0.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #5, MUL VL]\n" "udot z8.s, z6.b, z0.b[1]\n" @@ -178,7 +177,6 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "add x25, x25, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" @@ -217,9 +215,8 @@ void sve_hybrid_u8u32_dot_6x4VL ( "udot z10.s, z6.b, z0.b[3]\n" "udot z11.s, z7.b, z0.b[3]\n" "10:" // 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 5b\n" "st1w { z8.s }, p4, [x28]\n" @@ -296,16 +293,14 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" "udot z12.s, z6.b, z1.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "cmp x26, #0x10\n" "udot z13.s, z7.b, z1.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" - "prfm pldl1keep, [x24, #0x80]\n" "udot z10.s, z6.b, z0.b[0]\n" "udot z14.s, z6.b, z1.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #4, MUL VL]\n" @@ -356,9 +351,7 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" - "add x24, x24, #0x10\n" "udot z12.s, z6.b, z1.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" "udot z13.s, z7.b, z1.b[0]\n" @@ -413,10 +406,8 @@ void sve_hybrid_u8u32_dot_6x4VL ( "udot z11.s, z7.b, z0.b[3]\n" "udot z15.s, z7.b, z1.b[3]\n" "21:" // 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 16b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -511,21 +502,18 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "udot 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" "udot z12.s, z6.b, z1.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" "add x23, x23, #0x10\n" "udot z16.s, z6.b, z2.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "cmp x26, #0x10\n" - "udot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" "udot z17.s, z7.b, z2.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" - "prfm pldl1keep, [x23, #0x80]\n" "udot z10.s, z6.b, z0.b[0]\n" "udot z14.s, z6.b, z1.b[0]\n" "udot z18.s, z6.b, z2.b[0]\n" @@ -590,12 +578,9 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" "udot z12.s, z6.b, z1.b[0]\n" - "add x23, x23, #0x10\n" "udot z13.s, z7.b, z1.b[0]\n" "udot z16.s, z6.b, z2.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" @@ -665,11 +650,8 @@ void sve_hybrid_u8u32_dot_6x4VL ( "udot z15.s, z7.b, z1.b[3]\n" "udot z19.s, z7.b, z2.b[3]\n" "32:" // 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 27b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -781,26 +763,22 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "udot 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" "udot 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" "udot z16.s, z6.b, z2.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "add x22, x22, #0x10\n" + "add x23, x23, #0x10\n" "udot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x10\n" + "add x22, x22, #0x10\n" + "udot z17.s, z7.b, z2.b[0]\n" "udot z20.s, z6.b, z3.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "udot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "udot z21.s, z7.b, z3.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" - "prfm pldl1keep, [x22, #0x80]\n" "udot z10.s, z6.b, z0.b[0]\n" "udot z14.s, z6.b, z1.b[0]\n" "udot z18.s, z6.b, z2.b[0]\n" @@ -879,19 +857,15 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "udot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "udot z16.s, z6.b, z2.b[0]\n" - "add x22, x22, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" "udot z13.s, z7.b, z1.b[0]\n" - "udot z17.s, z7.b, z2.b[0]\n" + "udot z16.s, z6.b, z2.b[0]\n" "udot z20.s, z6.b, z3.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" "udot z21.s, z7.b, z3.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" @@ -972,12 +946,8 @@ void sve_hybrid_u8u32_dot_6x4VL ( "udot z19.s, z7.b, z2.b[3]\n" "udot z23.s, z7.b, z3.b[3]\n" "43:" // 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 38b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1106,32 +1076,27 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "udot 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" "udot 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" "udot 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" "udot z13.s, z7.b, z1.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x22, x22, #0x10\n" + "udot z17.s, z7.b, z2.b[0]\n" "add x21, x21, #0x10\n" "udot z20.s, z6.b, z3.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x10\n" "udot z24.s, z6.b, z4.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "udot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "udot z21.s, z7.b, z3.b[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" "udot z25.s, z7.b, z4.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" "udot z10.s, z6.b, z0.b[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" "udot z14.s, z6.b, z1.b[0]\n" "udot z18.s, z6.b, z2.b[0]\n" "udot z22.s, z6.b, z3.b[0]\n" @@ -1223,22 +1188,17 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "udot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "udot z16.s, z6.b, z2.b[0]\n" + "udot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z4.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" "udot z13.s, z7.b, z1.b[0]\n" - "add x21, x21, #0x10\n" - "udot z17.s, z7.b, z2.b[0]\n" + "udot z16.s, z6.b, z2.b[0]\n" "udot z20.s, z6.b, z3.b[0]\n" "udot z24.s, z6.b, z4.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" "udot z21.s, z7.b, z3.b[0]\n" "udot z25.s, z7.b, z4.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" @@ -1334,13 +1294,8 @@ void sve_hybrid_u8u32_dot_6x4VL ( "udot z23.s, z7.b, z3.b[3]\n" "udot z27.s, z7.b, z4.b[3]\n" "54:" // 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 49b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -1489,37 +1444,31 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" + "cmp x26, #0x10\n" "udot 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" "udot 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" "udot 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" "udot 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" "udot z20.s, z6.b, z3.b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x21, x21, #0x10\n" + "udot z17.s, z7.b, z2.b[0]\n" "add x20, x20, #0x10\n" "udot z24.s, z6.b, z4.b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "cmp x26, #0x10\n" "udot z28.s, z6.b, z5.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" - "udot z17.s, z7.b, z2.b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" "udot z21.s, z7.b, z3.b[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" "udot z25.s, z7.b, z4.b[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" "udot z29.s, z7.b, z5.b[0]\n" "ld1b { z7.b }, p5/Z, [x9, #3, MUL VL]\n" "udot z10.s, z6.b, z0.b[0]\n" - "prfm pldl1keep, [x20, #0x80]\n" "udot z14.s, z6.b, z1.b[0]\n" "udot z18.s, z6.b, z2.b[0]\n" "udot z22.s, z6.b, z3.b[0]\n" @@ -1625,25 +1574,19 @@ void sve_hybrid_u8u32_dot_6x4VL ( "ld1rqb { z0.b }, p0/Z, [x25]\n" "udot z8.s, z6.b, z0.b[0]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "add x25, x25, #0x10\n" "udot z9.s, z7.b, z0.b[0]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "udot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x23, x23, #0x10\n" - "udot z16.s, z6.b, z2.b[0]\n" + "udot z12.s, z6.b, z1.b[0]\n" "ld1rqb { z4.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" "udot z13.s, z7.b, z1.b[0]\n" "ld1rqb { z5.b }, p0/Z, [x20]\n" - "add x21, x21, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" "udot z20.s, z6.b, z3.b[0]\n" - "add x20, x20, #0x10\n" - "udot z17.s, z7.b, z2.b[0]\n" "udot z24.s, z6.b, z4.b[0]\n" "udot z28.s, z6.b, z5.b[0]\n" "ld1b { z6.b }, p5/Z, [x9, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" "udot z21.s, z7.b, z3.b[0]\n" "udot z25.s, z7.b, z4.b[0]\n" "udot z29.s, z7.b, z5.b[0]\n" @@ -1754,14 +1697,8 @@ void sve_hybrid_u8u32_dot_6x4VL ( "udot z27.s, z7.b, z4.b[3]\n" "udot z31.s, z7.b, z5.b[3]\n" "65:" // 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 60b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" |