diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp | 998 |
1 files changed, 499 insertions, 499 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp index 2b3a247686..17f2455469 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -69,258 +69,294 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( Args args = { inptr, ld_in_vl, pad_top, 8u - std::min(8u, pad_top + valid_input_rows), pad_left, weights, bias, valid_input_cols, output_cols, outptrs, outlds, outvllds, start_channel, valid_channels, act_min, act_max }; __asm__ __volatile__( - "ldr x4, [%x[args], %[offsetof_Args_pad_bottom]]\n" - "mov x20, #0x8\n" + "ldr x6, [%x[args], %[offsetof_Args_pad_bottom]]\n" + "mov x19, #0x8\n" ".inst 0xd503477f // SMSTART ZA\n" - "sub x20, x20, x4\n" + "sub x19, x19, x6\n" "ldr x5, [%x[args], %[offsetof_Args_pad_top]]\n" "ptrue p2.b\n" ".inst 0x25207812 // ptrue pn10.b\n" "ld1rw { z26.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" - "ldr x6, [%x[args], %[offsetof_Args_n_channels]]\n" - "whilelt p1.s, XZR, x6\n" - "whilelt p9.s, XZR, x20\n" + "ldr x7, [%x[args], %[offsetof_Args_n_channels]]\n" + "whilelt p1.s, XZR, x7\n" + "whilelt p9.s, XZR, x19\n" "ld1rw { z31.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" "whilelt p8.s, XZR, x5\n" "addvl SP, SP, #-30\n" - "ldr x7, [%x[args], %[offsetof_Args_current_channel]]\n" + "ldr x17, [%x[args], %[offsetof_Args_current_channel]]\n" "eor p8.b, p2/Z, p8.b, p9.b\n" "1:" // Channel loop - "ldr x20, [%x[args], %[offsetof_Args_bias]]\n" + "ldr x19, [%x[args], %[offsetof_Args_bias]]\n" "fmov z24.s, #0x0\n" - "cbz x20, 2f\n" - "ld1w { z24.s }, p1/Z, [x20, x7, LSL #2]\n" + "cbz x19, 2f\n" + "ld1w { z24.s }, p1/Z, [x19, x17, LSL #2]\n" "2:" // Load bias: Done - "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" - "mov x21, x20\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "ld1w { z11.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ldr x19, [%x[args], %[offsetof_Args_weights]]\n" + "mov x20, x19\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "ld1w { z11.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" "fmov z4.s, #0x0\n" - "incb x20\n" - "ld1w { z3.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "incb x19\n" + "ld1w { z3.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x658aaa45 // bfcvt z5.h, p2/M, z18.s\n" ".inst 0x658aa966 // bfcvt z6.h, p2/M, z11.s\n" - "ld1w { z17.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "addvl x24, SP, #30\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "addvl x23, SP, #30\n" ".inst 0x648aaa44 // bfcvtnt z4.h, p2/M, z18.s\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "mov x21, x20\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "mov x20, x19\n" ".inst 0x658aa867 // bfcvt z7.h, p2/M, z3.s\n" - "addvl x24, x24, #-6\n" - "ld1w { z18.s }, p2/Z, [x21]\n" + "addvl x23, x23, #-6\n" + "ld1w { z18.s }, p2/Z, [x20]\n" ".inst 0x658aaa28 // bfcvt z8.h, p2/M, z17.s\n" - "incb x21, ALL, MUL #5\n" - "st1h { z4.h }, p2, [x24]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z4.h }, p2, [x23]\n" ".inst 0x648aa965 // bfcvtnt z5.h, p2/M, z11.s\n" - "ld1w { z11.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ld1w { z11.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" "fmov z4.s, #0x0\n" - "st1h { z5.h }, p2, [x24, #1, MUL VL]\n" + "st1h { z5.h }, p2, [x23, #1, MUL VL]\n" ".inst 0x648aa866 // bfcvtnt z6.h, p2/M, z3.s\n" - "ld1w { z3.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ld1w { z3.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x658aaa45 // bfcvt z5.h, p2/M, z18.s\n" ".inst 0x648aaa27 // bfcvtnt z7.h, p2/M, z17.s\n" - "incb x20\n" - "ld1w { z17.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z6.h }, p2, [x24, #2, MUL VL]\n" + "incb x19\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z6.h }, p2, [x23, #2, MUL VL]\n" ".inst 0x648aaa08 // bfcvtnt z8.h, p2/M, z16.s\n" ".inst 0x658aaa09 // bfcvt z9.h, p2/M, z16.s\n" - "ld1w { z16.s }, p2/Z, [x21]\n" + "ld1w { z16.s }, p2/Z, [x20]\n" ".inst 0x658aa966 // bfcvt z6.h, p2/M, z11.s\n" - "mov x21, x20\n" - "st1h { z7.h }, p2, [x24, #3, MUL VL]\n" + "mov x20, x19\n" + "st1h { z7.h }, p2, [x23, #3, MUL VL]\n" ".inst 0x648aaa44 // bfcvtnt z4.h, p2/M, z18.s\n" ".inst 0x658aa867 // bfcvt z7.h, p2/M, z3.s\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z8.h }, p2, [x24, #4, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z8.h }, p2, [x23, #4, MUL VL]\n" ".inst 0x648aa965 // bfcvtnt z5.h, p2/M, z11.s\n" ".inst 0x658aaa28 // bfcvt z8.h, p2/M, z17.s\n" - "incb x20\n" - "st1h { z9.h }, p2, [x24, #5, MUL VL]\n" - "addvl x24, x24, #-6\n" - "ld1w { z11.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z4.h }, p2, [x24]\n" + "incb x19\n" + "st1h { z9.h }, p2, [x23, #5, MUL VL]\n" + "addvl x23, x23, #-6\n" + "ld1w { z11.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z4.h }, p2, [x23]\n" "fmov z4.s, #0x0\n" ".inst 0x648aa866 // bfcvtnt z6.h, p2/M, z3.s\n" - "ldr x17, [%x[args], %[offsetof_Args_input_cols]]\n" - "st1h { z5.h }, p2, [x24, #1, MUL VL]\n" - "ld1w { z3.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ldr x16, [%x[args], %[offsetof_Args_input_cols]]\n" + "st1h { z5.h }, p2, [x23, #1, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x658aaa45 // bfcvt z5.h, p2/M, z18.s\n" - "st1h { z6.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z6.h }, p2, [x23, #2, MUL VL]\n" ".inst 0x648aaa27 // bfcvtnt z7.h, p2/M, z17.s\n" - "ld1w { z17.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x658aa966 // bfcvt z6.h, p2/M, z11.s\n" ".inst 0x648aaa08 // bfcvtnt z8.h, p2/M, z16.s\n" - "st1h { z7.h }, p2, [x24, #3, MUL VL]\n" - "ldr x16, [%x[args], %[offsetof_Args_inptr]]\n" + "st1h { z7.h }, p2, [x23, #3, MUL VL]\n" + "ldr x15, [%x[args], %[offsetof_Args_inptr]]\n" ".inst 0x658aaa09 // bfcvt z9.h, p2/M, z16.s\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "mov x21, x20\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "mov x20, x19\n" ".inst 0x648aaa44 // bfcvtnt z4.h, p2/M, z18.s\n" ".inst 0x658aa867 // bfcvt z7.h, p2/M, z3.s\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z8.h }, p2, [x24, #4, MUL VL]\n" - "st1h { z9.h }, p2, [x24, #5, MUL VL]\n" - "addvl x24, x24, #-6\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z8.h }, p2, [x23, #4, MUL VL]\n" + "st1h { z9.h }, p2, [x23, #5, MUL VL]\n" + "addvl x23, x23, #-6\n" ".inst 0x648aa965 // bfcvtnt z5.h, p2/M, z11.s\n" ".inst 0x658aaa28 // bfcvt z8.h, p2/M, z17.s\n" - "ld1w { z11.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z4.h }, p2, [x24]\n" + "ld1w { z11.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z4.h }, p2, [x23]\n" ".inst 0x648aa866 // bfcvtnt z6.h, p2/M, z3.s\n" - "ld1w { z3.s }, p2/Z, [x21]\n" + "ld1w { z3.s }, p2/Z, [x20]\n" "fmov z4.s, #0x0\n" - "st1h { z5.h }, p2, [x24, #1, MUL VL]\n" - "incb x21, ALL, MUL #5\n" + "st1h { z5.h }, p2, [x23, #1, MUL VL]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x658aaa45 // bfcvt z5.h, p2/M, z18.s\n" - "st1h { z6.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z6.h }, p2, [x23, #2, MUL VL]\n" ".inst 0x648aaa27 // bfcvtnt z7.h, p2/M, z17.s\n" - "incb x20\n" - "ld1w { z17.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "incb x19\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x658aa966 // bfcvt z6.h, p2/M, z11.s\n" - "st1h { z7.h }, p2, [x24, #3, MUL VL]\n" + "st1h { z7.h }, p2, [x23, #3, MUL VL]\n" ".inst 0x648aaa08 // bfcvtnt z8.h, p2/M, z16.s\n" ".inst 0x658aaa09 // bfcvt z9.h, p2/M, z16.s\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "mov x21, x20\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "mov x20, x19\n" ".inst 0x658aa867 // bfcvt z7.h, p2/M, z3.s\n" ".inst 0x648aaa44 // bfcvtnt z4.h, p2/M, z18.s\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x648aa965 // bfcvtnt z5.h, p2/M, z11.s\n" - "ld1w { z11.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z8.h }, p2, [x24, #4, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z8.h }, p2, [x23, #4, MUL VL]\n" ".inst 0x648aa866 // bfcvtnt z6.h, p2/M, z3.s\n" ".inst 0x658aaa28 // bfcvt z8.h, p2/M, z17.s\n" - "ld1w { z3.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ld1w { z3.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" ".inst 0x648aaa27 // bfcvtnt z7.h, p2/M, z17.s\n" - "ld1w { z17.s }, p2/Z, [x21]\n" - "st1h { z9.h }, p2, [x24, #5, MUL VL]\n" - "addvl x24, x24, #-6\n" - "st1h { z4.h }, p2, [x24]\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "st1h { z9.h }, p2, [x23, #5, MUL VL]\n" + "addvl x23, x23, #-6\n" + "st1h { z4.h }, p2, [x23]\n" ".inst 0x648aaa08 // bfcvtnt z8.h, p2/M, z16.s\n" - "incb x21, ALL, MUL #5\n" + "incb x20, ALL, MUL #5\n" "fmov z4.s, #0x0\n" - "st1h { z5.h }, p2, [x24, #1, MUL VL]\n" + "st1h { z5.h }, p2, [x23, #1, MUL VL]\n" ".inst 0x658aaa45 // bfcvt z5.h, p2/M, z18.s\n" ".inst 0x658aaa09 // bfcvt z9.h, p2/M, z16.s\n" - "sub x20, x17, #0x1\n" - "st1h { z6.h }, p2, [x24, #2, MUL VL]\n" + "sub x19, x16, #0x1\n" + "st1h { z6.h }, p2, [x23, #2, MUL VL]\n" ".inst 0x658aa966 // bfcvt z6.h, p2/M, z11.s\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "orr x23, x20, %x[ld_in_col], LSL #18\n" - "st1h { z7.h }, p2, [x24, #3, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "orr x22, x19, %x[ld_in_col], LSL #18\n" + "st1h { z7.h }, p2, [x23, #3, MUL VL]\n" ".inst 0x658aa867 // bfcvt z7.h, p2/M, z3.s\n" - "orr x23, x6, x23, LSL #20\n" - "mov x22, #0x8\n" - "st1h { z8.h }, p2, [x24, #4, MUL VL]\n" + "orr x22, x7, x22, LSL #20\n" + "mov x21, #0x8\n" + "st1h { z8.h }, p2, [x23, #4, MUL VL]\n" ".inst 0x658aaa28 // bfcvt z8.h, p2/M, z17.s\n" - "add x21, x5, x4\n" - "lsl x20, %x[ld_in_row], #0x2\n" - "st1h { z9.h }, p2, [x24, #5, MUL VL]\n" - "addvl x24, x24, #-6\n" + "add x20, x5, x6\n" + "lsl x19, %x[ld_in_row], #0x2\n" + "st1h { z9.h }, p2, [x23, #5, MUL VL]\n" + "addvl x23, x23, #-6\n" "mov z25.d, z24.d\n" ".inst 0x648aaa44 // bfcvtnt z4.h, p2/M, z18.s\n" - "st1h { z4.h }, p2, [x24]\n" + "st1h { z4.h }, p2, [x23]\n" ".inst 0x648aa965 // bfcvtnt z5.h, p2/M, z11.s\n" ".inst 0x648aa866 // bfcvtnt z6.h, p2/M, z3.s\n" "mov x11, #0x0\n" - "st1h { z5.h }, p2, [x24, #1, MUL VL]\n" + "st1h { z5.h }, p2, [x23, #1, MUL VL]\n" ".inst 0x648aaa27 // bfcvtnt z7.h, p2/M, z17.s\n" ".inst 0x648aaa08 // bfcvtnt z8.h, p2/M, z16.s\n" "mov x8, #0x8\n" - "st1h { z6.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z6.h }, p2, [x23, #2, MUL VL]\n" ".inst 0x658aaa09 // bfcvt z9.h, p2/M, z16.s\n" - "ldr x15, [%x[args], %[offsetof_Args_output_cols]]\n" - "lsl x23, x23, #0x2\n" - "st1h { z7.h }, p2, [x24, #3, MUL VL]\n" - "sub x22, x22, x21\n" - "madd x20, x20, x5, x16\n" - "st1h { z8.h }, p2, [x24, #4, MUL VL]\n" - "st1h { z9.h }, p2, [x24, #5, MUL VL]\n" + "ldr x14, [%x[args], %[offsetof_Args_output_cols]]\n" + "lsl x22, x22, #0x2\n" + "st1h { z7.h }, p2, [x23, #3, MUL VL]\n" + "sub x21, x21, x20\n" + "madd x19, x19, x5, x15\n" + "st1h { z8.h }, p2, [x23, #4, MUL VL]\n" + "st1h { z9.h }, p2, [x23, #5, MUL VL]\n" "3:" // Issue prefetches - "subs x22, x22, #0x1\n" - ".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n" - "add x20, x20, %x[ld_in_col], LSL #2\n" + "subs x21, x21, #0x1\n" + ".inst 0xf8b64a7c // rprfm pldstrm, x22, [x19]\n" + "add x19, x19, %x[ld_in_col], LSL #2\n" "bgt 3b\n" - "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" - "lsl x20, %x[ld_in_row], #0x2\n" - "msub x16, x5, x20, x16\n" + "ldr x24, [%x[args], %[offsetof_Args_outptrs]]\n" + "lsl x19, %x[ld_in_row], #0x2\n" + "msub x15, x5, x19, x15\n" ".inst 0xc0046b00 // mova za.d[x11, #0], { z24.d-z25.d }\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" + "ldr x19, [%x[args], %[offsetof_Args_ld_out_cols]]\n" ".inst 0xc0046b01 // mova za.d[x11, #1], { z24.d-z25.d }\n" - "mov x22, #0x4\n" - "ldp x14, x13, [x25], #0x10\n" + "mov x21, #0x4\n" + "ldp x13, x0, [x24], #0x10\n" ".inst 0xc0046b02 // mova za.d[x11, #2], { z24.d-z25.d }\n" - "ldp x0, x10, [x20], #0x10\n" + "ldp x10, x9, [x19], #0x10\n" ".inst 0xc0046b03 // mova za.d[x11, #3], { z24.d-z25.d }\n" - "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" + "ldr x20, [%x[args], %[offsetof_Args_pad_left]]\n" ".inst 0xc0046b04 // mova za.d[x11, #4], { z24.d-z25.d }\n" - "ldp x9, x28, [x25], #0x10\n" + "ldp x28, x27, [x24], #0x10\n" ".inst 0xc0046b05 // mova za.d[x11, #5], { z24.d-z25.d }\n" - "ldp x27, x26, [x20], #0x10\n" + "ldp x26, x25, [x19], #0x10\n" ".inst 0xc0046b06 // mova za.d[x11, #6], { z24.d-z25.d }\n" ".inst 0xc0046b07 // mova za.d[x11, #7], { z24.d-z25.d }\n" ".inst 0xc0040b00 // mova za.d[x8, #0], { z24.d-z25.d }\n" ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" - "cbz x21, 5f\n" - "cmp x21, x22\n" - "csel x20, x21, x22, LT\n" - "sub x21, x21, x20\n" - "sub x22, x22, x20\n" - "cbz x21, 5f\n" + "cbz x20, 5f\n" + "cmp x20, x21\n" + "csel x19, x20, x21, LT\n" + "sub x20, x20, x19\n" + "sub x21, x21, x19\n" + "cbz x20, 5f\n" ".inst 0xc0066800 // mova { z0.d-z1.d }, za.d[x11, #0]\n" - "sub x15, x15, x21\n" + "sub x14, x14, x20\n" ".inst 0xc0066822 // mova { z2.d-z3.d }, za.d[x11, #1]\n" ".inst 0xc1bfcb40 // fclamp { z0.s-z3.s }, z26.s, z31.s\n" "4:" // Left padding - "subs x21, x21, #0x1\n" - "st1w { z0.s }, p1, [x14]\n" - "add x14, x14, x0, LSL #2\n" - "st1w { z2.s }, p1, [x13]\n" + "subs x20, x20, #0x1\n" + "st1w { z0.s }, p1, [x13]\n" "add x13, x13, x10, LSL #2\n" - "st1w { z1.s }, p1, [x9]\n" - "add x9, x9, x27, LSL #2\n" - "st1w { z3.s }, p1, [x28]\n" + "st1w { z2.s }, p1, [x0]\n" + "add x0, x0, x9, LSL #2\n" + "st1w { z1.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" + "st1w { z3.s }, p1, [x27]\n" + "add x27, x27, x25, LSL #2\n" "bgt 4b\n" "5:" // Left padding: End - "adds XZR, x5, x4\n" + "adds XZR, x5, x6\n" "bne 12f\n" - "cbz x22, 10f\n" - "cmp x22, #0x1\n" - "sub x17, x17, x22\n" + "cbz x21, 10f\n" + "cmp x21, #0x1\n" + "sub x16, x16, x21\n" "beq 9f\n" - "cmp x22, #0x2\n" + "cmp x21, #0x2\n" "beq 8f\n" - "cmp x22, #0x3\n" + "cmp x21, #0x3\n" "beq 7f\n" "6:" // Unpadded: 4 priming loads - "add x21, x16, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x16]\n" + "add x20, x15, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "addvl x20, SP, #24\n" + "addvl x19, SP, #24\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" + ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" + ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" + ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" + ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" + "7:" // Unpadded: 3 priming loads + "add x21, x15, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x15]\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "addvl x20, SP, #18\n" "ld1w { z16.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + "addvl x19, SP, #24\n" "ld1w { z16.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" @@ -336,30 +372,40 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" + ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" "ld1w { z16.s }, p1/Z, [x21]\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" + ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" + ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - "7:" // Unpadded: 3 priming loads - "add x22, x16, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" + ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" + ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" + "8:" // Unpadded: 2 priming loads + "add x22, x15, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "addvl x21, SP, #18\n" + "addvl x21, SP, #12\n" "ld1w { z16.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "addvl x20, SP, #24\n" + "addvl x20, SP, #18\n" "ld1w { z16.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + "addvl x19, SP, #24\n" "ld1w { z16.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" @@ -379,36 +425,46 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( "ld1w { z16.s }, p1/Z, [x22]\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" + ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - "8:" // Unpadded: 2 priming loads - "add x23, x16, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" + ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" + ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" + "9:" // Unpadded: 1 priming loads + "add x23, x15, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "addvl x22, SP, #12\n" + "addvl x22, SP, #6\n" "ld1w { z16.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "addvl x21, SP, #18\n" + "addvl x21, SP, #12\n" "ld1w { z16.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" - "addvl x20, SP, #24\n" + "addvl x20, SP, #18\n" "ld1w { z16.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + "addvl x19, SP, #24\n" "ld1w { z16.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" @@ -431,177 +487,121 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" - ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" - ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" - ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" - ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" - ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - "9:" // Unpadded: 1 priming loads - "add x24, x16, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x16]\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "addvl x23, SP, #6\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "addvl x22, SP, #12\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" - "addvl x21, SP, #18\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" - "addvl x20, SP, #24\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" - ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" - ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" - ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" - ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" - ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" - ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" - ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" - ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" - ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" - ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc1257196 // bfdot za.s[x11, 6], { z12.h-z13.h }, z5.h\n" ".inst 0xc1247197 // bfdot za.s[x11, 7], { z12.h-z13.h }, z4.h\n" ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" ".inst 0xc12771b6 // bfdot za.s[x11, 6], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b7 // bfdot za.s[x11, 7], { z13.h-z14.h }, z6.h\n" ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" ".inst 0xc12971d6 // bfdot za.s[x11, 6], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d7 // bfdot za.s[x11, 7], { z14.h-z15.h }, z8.h\n" "10:" // Unpadded: 0 priming loads ".inst 0xa0402be4 // ld1h { z4.h-z5.h }, pn10.b/Z, [SP]\n" ".inst 0xa0412be6 // ld1h { z6.h-z7.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" ".inst 0xa0422be8 // ld1h { z8.h-z9.h }, pn10.b/Z, [SP, #0x4, MUL VL]\n" - "cbz x17, 20f\n" - "add x20, x16, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x16]\n" + "cbz x16, 20f\n" + "add x19, x15, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "sub x17, x17, #0x1\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "sub x15, x15, #0x1\n" + "sub x16, x16, #0x1\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "sub x14, x14, #0x1\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" - "cmp x17, x15\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "csel x25, x17, x15, LT\n" + "cmp x16, x14\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "csel x24, x16, x14, LT\n" ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - "sub x15, x15, x25\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "sub x14, x14, x24\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z16.s }, p1/Z, [x19]\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - "cbz x25, 19f\n" + "cbz x24, 19f\n" "11:" // Unpadded: Main loop - "addvl x24, SP, #6\n" + "addvl x23, SP, #6\n" ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" - "addvl x23, SP, #12\n" - "ld1w { z23.s }, p1/Z, [x16]\n" + "addvl x22, SP, #12\n" + "ld1w { z23.s }, p1/Z, [x15]\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402b04 // ld1h { z4.h-z5.h }, pn10.b/Z, [x24]\n" - "addvl x22, SP, #18\n" - "addvl x21, SP, #24\n" + ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" + "addvl x21, SP, #18\n" + "addvl x20, SP, #24\n" ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" - "add x20, x16, %x[ld_in_row], LSL #2\n" - "ld1w { z22.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "add x19, x15, %x[ld_in_row], LSL #2\n" + "ld1w { z22.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" - "subs x25, x25, #0x1\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" + "subs x24, x24, #0x1\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" - "ld1w { z21.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z21.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412b06 // ld1h { z6.h-z7.h }, pn10.b/Z, [x24, #0x2, MUL VL]\n" + ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" - "ld1w { z20.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z20.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" + ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" - "ld1w { z19.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z19.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" - "ld1w { z18.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z18.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422b08 // ld1h { z8.h-z9.h }, pn10.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" ".inst 0xc1257196 // bfdot za.s[x11, 6], { z12.h-z13.h }, z5.h\n" - "ld1w { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z17.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0xc1247197 // bfdot za.s[x11, 7], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" + ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z16.s }, p1/Z, [x19]\n" ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" + ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc12771b6 // bfdot za.s[x11, 6], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b7 // bfdot za.s[x11, 7], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" + ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" ".inst 0xc12971d6 // bfdot za.s[x11, 6], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d7 // bfdot za.s[x11, 7], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" ".inst 0xc1251190 // bfdot za.s[x8, 0], { z12.h-z13.h }, z5.h\n" ".inst 0xc1241191 // bfdot za.s[x8, 1], { z12.h-z13.h }, z4.h\n" ".inst 0x658aaaec // bfcvt z12.h, p2/M, z23.s\n" @@ -622,35 +622,80 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0x648aaa4e // bfcvtnt z14.h, p2/M, z18.s\n" ".inst 0xc0066822 // mova { z2.d-z3.d }, za.d[x11, #1]\n" ".inst 0xc1bfcb40 // fclamp { z0.s-z3.s }, z26.s, z31.s\n" - "st1w { z0.s }, p1, [x14]\n" - "add x14, x14, x0, LSL #2\n" - "st1w { z2.s }, p1, [x13]\n" + "st1w { z0.s }, p1, [x13]\n" "add x13, x13, x10, LSL #2\n" + "st1w { z2.s }, p1, [x0]\n" + "add x0, x0, x9, LSL #2\n" "add x11, x11, #0x2\n" ".inst 0xc0040b00 // mova za.d[x8, #0], { z24.d-z25.d }\n" - "st1w { z1.s }, p1, [x9]\n" - "add x9, x9, x27, LSL #2\n" + "st1w { z1.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - "st1w { z3.s }, p1, [x28]\n" - "add x28, x28, x26, LSL #2\n" + "st1w { z3.s }, p1, [x27]\n" + "add x27, x27, x25, LSL #2\n" "bgt 11b\n" "b 19f\n" "12:" // Padded - "cbz x22, 17f\n" - "cmp x22, #0x1\n" - "sub x17, x17, x22\n" + "cbz x21, 17f\n" + "cmp x21, #0x1\n" + "sub x16, x16, x21\n" "beq 16f\n" - "cmp x22, #0x2\n" + "cmp x21, #0x2\n" "beq 15f\n" - "cmp x22, #0x3\n" + "cmp x21, #0x3\n" "beq 14f\n" "13:" // Padded: 4 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x16]\n" + "ld1w { z16.s }, p0/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "add x21, x16, %x[ld_in_row], LSL #2\n" + "add x20, x15, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "addvl x19, SP, #24\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" + ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" + ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" + ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" + ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" + "14:" // Padded: 3 priming loads + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x15]\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "add x21, x15, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1w { z16.s }, p0/Z, [x21]\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" @@ -669,7 +714,7 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "addvl x20, SP, #24\n" + "addvl x20, SP, #18\n" "ld1w { z16.s }, p0/Z, [x21]\n" ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" "add x21, x21, %x[ld_in_row], LSL #2\n" @@ -677,25 +722,35 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" "ld1w { z16.s }, p0/Z, [x21]\n" + "addvl x19, SP, #24\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" - ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" + ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" + ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - "14:" // Padded: 3 priming loads + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" + ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" + ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" + "15:" // Padded: 2 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x16]\n" + "ld1w { z16.s }, p0/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "add x22, x16, %x[ld_in_row], LSL #2\n" + "add x22, x15, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1w { z16.s }, p0/Z, [x22]\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" @@ -714,7 +769,7 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "addvl x21, SP, #18\n" + "addvl x21, SP, #12\n" "ld1w { z16.s }, p0/Z, [x22]\n" ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" "add x22, x22, %x[ld_in_row], LSL #2\n" @@ -722,35 +777,45 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" "ld1w { z16.s }, p0/Z, [x22]\n" - "addvl x20, SP, #24\n" + "addvl x20, SP, #18\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + "addvl x19, SP, #24\n" ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p0/Z, [x22]\n" ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" - ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" + ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" + ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - "15:" // Padded: 2 priming loads + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" + ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" + ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" + "16:" // Padded: 1 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x16]\n" + "ld1w { z16.s }, p0/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "add x23, x16, %x[ld_in_row], LSL #2\n" + "add x23, x15, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1w { z16.s }, p0/Z, [x23]\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" @@ -769,7 +834,7 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "addvl x22, SP, #12\n" + "addvl x22, SP, #6\n" "ld1w { z16.s }, p0/Z, [x23]\n" ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" "add x23, x23, %x[ld_in_row], LSL #2\n" @@ -777,16 +842,17 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" "ld1w { z16.s }, p0/Z, [x23]\n" - "addvl x21, SP, #18\n" + "addvl x21, SP, #12\n" "add x23, x23, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" - "addvl x20, SP, #24\n" + "addvl x20, SP, #18\n" ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + "addvl x19, SP, #24\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p0/Z, [x23]\n" ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" @@ -797,207 +863,141 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" - ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" - ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" - ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" - ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" - ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - "16:" // Padded: 1 priming loads - "mov x12, #0x0\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x16]\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "add x24, x16, %x[ld_in_row], LSL #2\n" - ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" - "mov x12, #0x4\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "addvl x23, SP, #6\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" - ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - "addvl x22, SP, #12\n" - "add x24, x24, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" - "addvl x21, SP, #18\n" - ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" - "addvl x20, SP, #24\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" - ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" - ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" - ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" - ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" - ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" - ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" - ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" - ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc1257196 // bfdot za.s[x11, 6], { z12.h-z13.h }, z5.h\n" ".inst 0xc1247197 // bfdot za.s[x11, 7], { z12.h-z13.h }, z4.h\n" ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" ".inst 0xc12771b6 // bfdot za.s[x11, 6], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b7 // bfdot za.s[x11, 7], { z13.h-z14.h }, z6.h\n" ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" ".inst 0xc12971d6 // bfdot za.s[x11, 6], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d7 // bfdot za.s[x11, 7], { z14.h-z15.h }, z8.h\n" "17:" // Padded: 0 priming loads ".inst 0xa0402be4 // ld1h { z4.h-z5.h }, pn10.b/Z, [SP]\n" ".inst 0xa0412be6 // ld1h { z6.h-z7.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" ".inst 0xa0422be8 // ld1h { z8.h-z9.h }, pn10.b/Z, [SP, #0x4, MUL VL]\n" - "cbz x17, 20f\n" + "cbz x16, 20f\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x16]\n" + "ld1w { z16.s }, p0/Z, [x15]\n" ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "add x20, x16, %x[ld_in_row], LSL #2\n" + "add x19, x15, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" "mov x12, #0x4\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "sub x17, x17, #0x1\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "sub x16, x16, #0x1\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - "sub x15, x15, #0x1\n" - "cmp x17, x15\n" - "csel x25, x17, x15, LT\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" - "sub x15, x15, x25\n" - "cbz x25, 19f\n" + "sub x14, x14, #0x1\n" + "cmp x16, x14\n" + "csel x24, x16, x14, LT\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" + "sub x14, x14, x24\n" + "cbz x24, 19f\n" "18:" // Padded: Main loop - "addvl x24, SP, #6\n" + "addvl x23, SP, #6\n" ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" - "addvl x23, SP, #12\n" + "addvl x22, SP, #12\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402b04 // ld1h { z4.h-z5.h }, pn10.b/Z, [x24]\n" + ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" - "ld1w { z23.s }, p0/Z, [x16]\n" - "add x22, x16, %x[ld_in_row], LSL #2\n" + "ld1w { z23.s }, p0/Z, [x15]\n" + "add x21, x15, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" - "addvl x21, SP, #18\n" - "addvl x20, SP, #24\n" + ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" + "addvl x20, SP, #18\n" + "addvl x19, SP, #24\n" ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" - "ld1w { z22.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row], LSL #2\n" + "ld1w { z22.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412b06 // ld1h { z6.h-z7.h }, pn10.b/Z, [x24, #0x2, MUL VL]\n" - "subs x25, x25, #0x1\n" - "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" + "subs x24, x24, #0x1\n" + "add x15, x15, %x[ld_in_col], LSL #2\n" ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" - "ld1w { z21.s }, p0/Z, [x22]\n" + "ld1w { z21.s }, p0/Z, [x21]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x4\n" ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" - "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" - "ld1w { z20.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row], LSL #2\n" + "ld1w { z20.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" - "ld1w { z19.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row], LSL #2\n" + "ld1w { z19.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422b08 // ld1h { z8.h-z9.h }, pn10.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" ".inst 0xc1257196 // bfdot za.s[x11, 6], { z12.h-z13.h }, z5.h\n" - "ld1w { z18.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row], LSL #2\n" + "ld1w { z18.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" ".inst 0xc1247197 // bfdot za.s[x11, 7], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" - "ld1w { z17.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row], LSL #2\n" + "ld1w { z17.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc12771b6 // bfdot za.s[x11, 6], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b7 // bfdot za.s[x11, 7], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" ".inst 0xc12971d6 // bfdot za.s[x11, 6], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d7 // bfdot za.s[x11, 7], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" ".inst 0xc1251190 // bfdot za.s[x8, 0], { z12.h-z13.h }, z5.h\n" ".inst 0xc1241191 // bfdot za.s[x8, 1], { z12.h-z13.h }, z4.h\n" ".inst 0x658aaaec // bfcvt z12.h, p2/M, z23.s\n" @@ -1018,60 +1018,60 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0x648aaa4e // bfcvtnt z14.h, p2/M, z18.s\n" ".inst 0xc0066822 // mova { z2.d-z3.d }, za.d[x11, #1]\n" ".inst 0xc1bfcb40 // fclamp { z0.s-z3.s }, z26.s, z31.s\n" - "st1w { z0.s }, p1, [x14]\n" - "add x14, x14, x0, LSL #2\n" - "st1w { z2.s }, p1, [x13]\n" + "st1w { z0.s }, p1, [x13]\n" "add x13, x13, x10, LSL #2\n" + "st1w { z2.s }, p1, [x0]\n" + "add x0, x0, x9, LSL #2\n" "add x11, x11, #0x2\n" ".inst 0xc0040b00 // mova za.d[x8, #0], { z24.d-z25.d }\n" - "st1w { z1.s }, p1, [x9]\n" - "add x9, x9, x27, LSL #2\n" + "st1w { z1.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - "st1w { z3.s }, p1, [x28]\n" - "add x28, x28, x26, LSL #2\n" + "st1w { z3.s }, p1, [x27]\n" + "add x27, x27, x25, LSL #2\n" "bgt 18b\n" "19:" // Main loop tail - "addvl x23, SP, #6\n" + "addvl x22, SP, #6\n" ".inst 0xc1257190 // bfdot za.s[x11, 0], { z12.h-z13.h }, z5.h\n" - "addvl x22, SP, #12\n" + "addvl x21, SP, #12\n" ".inst 0xc1247191 // bfdot za.s[x11, 1], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ae4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x23]\n" - "addvl x21, SP, #18\n" - "addvl x20, SP, #24\n" + ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" + "addvl x20, SP, #18\n" + "addvl x19, SP, #24\n" ".inst 0xc1257192 // bfdot za.s[x11, 2], { z12.h-z13.h }, z5.h\n" ".inst 0xc1247193 // bfdot za.s[x11, 3], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402ac4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x22]\n" + ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" ".inst 0xc12771b0 // bfdot za.s[x11, 0], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b1 // bfdot za.s[x11, 1], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ae6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" ".inst 0xc1257194 // bfdot za.s[x11, 4], { z12.h-z13.h }, z5.h\n" ".inst 0xc1247195 // bfdot za.s[x11, 5], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402aa4 // ld1h { z4.h-z5.h }, pn10.b/Z, [x21]\n" + ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" ".inst 0xc12771b2 // bfdot za.s[x11, 2], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b3 // bfdot za.s[x11, 3], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412ac6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" + ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" ".inst 0xc12971d0 // bfdot za.s[x11, 0], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d1 // bfdot za.s[x11, 1], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ae8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" ".inst 0xc1257196 // bfdot za.s[x11, 6], { z12.h-z13.h }, z5.h\n" ".inst 0xc1247197 // bfdot za.s[x11, 7], { z12.h-z13.h }, z4.h\n" - ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" + ".inst 0xa0402a64 // ld1h { z4.h-z5.h }, pn10.b/Z, [x19]\n" ".inst 0xc12771b4 // bfdot za.s[x11, 4], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b5 // bfdot za.s[x11, 5], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412aa6 // ld1h { z6.h-z7.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" ".inst 0xc12971d2 // bfdot za.s[x11, 2], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d3 // bfdot za.s[x11, 3], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422ac8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x22, #0x4, MUL VL]\n" + ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" ".inst 0xc12771b6 // bfdot za.s[x11, 6], { z13.h-z14.h }, z7.h\n" ".inst 0xc12671b7 // bfdot za.s[x11, 7], { z13.h-z14.h }, z6.h\n" - ".inst 0xa0412a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xa0412a66 // ld1h { z6.h-z7.h }, pn10.b/Z, [x19, #0x2, MUL VL]\n" ".inst 0xc12971d4 // bfdot za.s[x11, 4], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d5 // bfdot za.s[x11, 5], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" ".inst 0xc12971d6 // bfdot za.s[x11, 6], { z14.h-z15.h }, z9.h\n" ".inst 0xc12871d7 // bfdot za.s[x11, 7], { z14.h-z15.h }, z8.h\n" - ".inst 0xa0422a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20, #0x4, MUL VL]\n" + ".inst 0xa0422a68 // ld1h { z8.h-z9.h }, pn10.b/Z, [x19, #0x4, MUL VL]\n" ".inst 0xc1251190 // bfdot za.s[x8, 0], { z12.h-z13.h }, z5.h\n" ".inst 0xc1241191 // bfdot za.s[x8, 1], { z12.h-z13.h }, z4.h\n" ".inst 0xc12711b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z7.h\n" @@ -1082,66 +1082,66 @@ void sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za_impl( ".inst 0xc0066800 // mova { z0.d-z1.d }, za.d[x11, #0]\n" ".inst 0xc0066822 // mova { z2.d-z3.d }, za.d[x11, #1]\n" ".inst 0xc1bfcb40 // fclamp { z0.s-z3.s }, z26.s, z31.s\n" - "st1w { z0.s }, p1, [x14]\n" - "add x14, x14, x0, LSL #2\n" - "st1w { z2.s }, p1, [x13]\n" + "st1w { z0.s }, p1, [x13]\n" "add x13, x13, x10, LSL #2\n" + "st1w { z2.s }, p1, [x0]\n" + "add x0, x0, x9, LSL #2\n" "add x11, x11, #0x2\n" ".inst 0xc0040b00 // mova za.d[x8, #0], { z24.d-z25.d }\n" - "st1w { z1.s }, p1, [x9]\n" - "add x9, x9, x27, LSL #2\n" - ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" - "st1w { z3.s }, p1, [x28]\n" + "st1w { z1.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" + ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" + "st1w { z3.s }, p1, [x27]\n" + "add x27, x27, x25, LSL #2\n" "20:" // Main loop skip tail - "cbz x15, 22f\n" + "cbz x14, 22f\n" "21:" // Right padding loop ".inst 0xc0066800 // mova { z0.d-z1.d }, za.d[x11, #0]\n" "add x8, x8, #0x2\n" - "subs x15, x15, #0x1\n" + "subs x14, x14, #0x1\n" ".inst 0xc0066822 // mova { z2.d-z3.d }, za.d[x11, #1]\n" ".inst 0xc1bfcb40 // fclamp { z0.s-z3.s }, z26.s, z31.s\n" - "st1w { z0.s }, p1, [x14]\n" - "add x14, x14, x0, LSL #2\n" - "st1w { z2.s }, p1, [x13]\n" + "st1w { z0.s }, p1, [x13]\n" "add x13, x13, x10, LSL #2\n" + "st1w { z2.s }, p1, [x0]\n" + "add x0, x0, x9, LSL #2\n" "add x11, x11, #0x2\n" ".inst 0xc0040b00 // mova za.d[x8, #0], { z24.d-z25.d }\n" - "st1w { z1.s }, p1, [x9]\n" - "add x9, x9, x27, LSL #2\n" - ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" - "st1w { z3.s }, p1, [x28]\n" + "st1w { z1.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" + ".inst 0xc0040b01 // mova za.d[x8, #1], { z24.d-z25.d }\n" + "st1w { z3.s }, p1, [x27]\n" + "add x27, x27, x25, LSL #2\n" "bgt 21b\n" "22:" // End - "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" - "incb x20, ALL, MUL #16\n" - "incb x20, ALL, MUL #9\n" - "str x20, [%x[args], %[offsetof_Args_weights]]\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_in_vl]]\n" - "incw x7\n" - "whilelt p1.s, x7, x6\n" - "ldr x16, [%x[args], %[offsetof_Args_inptr]]\n" - "add x16, x16, x20, LSL #2\n" - "str x16, [%x[args], %[offsetof_Args_inptr]]\n" - "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" - "ldr x24, [%x[args], %[offsetof_Args_ld_out_vls]]\n" - "ldp x23, x22, [x25, #0x0]\n" - "ldp x21, x20, [x24, #0x0]\n" - "add x23, x23, x21, LSL #2\n" + "ldr x19, [%x[args], %[offsetof_Args_weights]]\n" + "incb x19, ALL, MUL #16\n" + "incb x19, ALL, MUL #9\n" + "str x19, [%x[args], %[offsetof_Args_weights]]\n" + "ldr x19, [%x[args], %[offsetof_Args_ld_in_vl]]\n" + "incw x17\n" + "whilelt p1.s, x17, x7\n" + "ldr x15, [%x[args], %[offsetof_Args_inptr]]\n" + "add x15, x15, x19, LSL #2\n" + "str x15, [%x[args], %[offsetof_Args_inptr]]\n" + "ldr x24, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x23, [%x[args], %[offsetof_Args_ld_out_vls]]\n" + "ldp x22, x21, [x24, #0x0]\n" + "ldp x20, x19, [x23, #0x0]\n" "add x22, x22, x20, LSL #2\n" - "stp x23, x22, [x25, #0x0]\n" - "ldp x23, x22, [x25, #0x10]\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x23, x23, x21, LSL #2\n" + "add x21, x21, x19, LSL #2\n" + "stp x22, x21, [x24, #0x0]\n" + "ldp x22, x21, [x24, #0x10]\n" + "ldp x20, x19, [x23, #0x10]\n" "add x22, x22, x20, LSL #2\n" - "stp x23, x22, [x25, #0x10]\n" + "add x21, x21, x19, LSL #2\n" + "stp x22, x21, [x24, #0x10]\n" "b.any 1b\n" "addvl SP, SP, #30\n" ".inst 0xd503467f // SMSTOP\n" : : [args] "r" (&args), [ld_in_col] "r" (ld_in_col), [ld_in_row] "r" (ld_in_row), [offsetof_Args_bias] "I" (offsetof(Args, bias)), [offsetof_Args_clamp_max] "I" (offsetof(Args, clamp_max)), [offsetof_Args_clamp_min] "I" (offsetof(Args, clamp_min)), [offsetof_Args_current_channel] "I" (offsetof(Args, current_channel)), [offsetof_Args_inptr] "I" (offsetof(Args, inptr)), [offsetof_Args_input_cols] "I" (offsetof(Args, input_cols)), [offsetof_Args_ld_in_vl] "I" (offsetof(Args, ld_in_vl)), [offsetof_Args_ld_out_cols] "I" (offsetof(Args, ld_out_cols)), [offsetof_Args_ld_out_vls] "I" (offsetof(Args, ld_out_vls)), [offsetof_Args_n_channels] "I" (offsetof(Args, n_channels)), [offsetof_Args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_Args_output_cols] "I" (offsetof(Args, output_cols)), [offsetof_Args_pad_bottom] "I" (offsetof(Args, pad_bottom)), [offsetof_Args_pad_left] "I" (offsetof(Args, pad_left)), [offsetof_Args_pad_top] "I" (offsetof(Args, pad_top)), [offsetof_Args_weights] "I" (offsetof(Args, weights)) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x0", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x0", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |