diff options
author | Michael Tyler <michael.tyler@arm.com> | 2023-04-12 17:43:17 +0100 |
---|---|---|
committer | michael.tyler <michael.tyler@arm.com> | 2023-06-05 15:57:58 +0000 |
commit | 74921eee924625426429044decefe3673561b174 (patch) | |
tree | 654da1a95e3d42d6af8ad1ff27bb40d77b1fd8c5 /src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp | |
parent | df5d9878008be9b60586df97ebfff197abb5195e (diff) | |
download | ComputeLibrary-74921eee924625426429044decefe3673561b174.tar.gz |
Update CPU kernel implementations and guard directives
Resolves COMPMID-6023
Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp | 1288 |
1 files changed, 644 insertions, 644 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp index 01f689a0b4..3a56e69d26 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp @@ -76,134 +76,134 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( "ldr x4, [%x[args], %[offsetof_Args_pad_top]]\n" "ptrue p2.b\n" ".inst 0x25207812 // ptrue pn10.b\n" - "ld1rw { z30.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" + "ld1rw { z13.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" "ldr x5, [%x[args], %[offsetof_Args_n_channels]]\n" "whilelt p1.s, XZR, x5\n" "whilelt p9.s, XZR, x20\n" - "ld1rw { z22.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" + "ld1rw { z12.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" "whilelt p8.s, XZR, x4\n" "addvl SP, SP, #-15\n" "ldr x6, [%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" - "fmov z4.s, #0x0\n" + "fmov z16.s, #0x0\n" "cbz x20, 2f\n" - "ld1w { z4.s }, p1/Z, [x20, x6, LSL #2]\n" + "ld1w { z16.s }, p1/Z, [x20, x6, LSL #2]\n" "2:" // Load bias: Done - "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" - "mov x21, x20\n" - "ld1w { z31.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" - "incb x20\n" - "ld1w { z13.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + "ldr x21, [%x[args], %[offsetof_Args_weights]]\n" + "mov x20, x21\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "ld1w { z8.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aabef // bfcvt z15.h, p2/M, z31.s\n" + "incb x21\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aaa4e // bfcvt z14.h, p2/M, z18.s\n" "addvl x24, SP, #15\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x648aa90f // bfcvtnt z15.h, p2/M, z8.s\n" "addvl x24, x24, #-3\n" - "ld1w { z15.s }, p2/Z, [x21]\n" - "mov x21, x20\n" - "st1h { z1.h }, p2, [x24]\n" - ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" - "ld1w { z31.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" - "incb x20\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z9.h }, p2, [x24, #1, MUL VL]\n" - ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" - "ld1w { z13.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" - ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "st1h { z2.h }, p2, [x24, #2, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "mov x20, x21\n" + "st1h { z15.h }, p2, [x24]\n" + ".inst 0x648aaa2e // bfcvtnt z14.h, p2/M, z17.s\n" + "ld1w { z29.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aabb5 // bfcvt z21.h, p2/M, z29.s\n" + "incb x21\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z14.h }, p2, [x24, #1, MUL VL]\n" + ".inst 0x658aaa58 // bfcvt z24.h, p2/M, z18.s\n" + "ld1w { z26.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aab41 // bfcvt z1.h, p2/M, z26.s\n" + ".inst 0x648aaa35 // bfcvtnt z21.h, p2/M, z17.s\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z24.h }, p2, [x24, #2, MUL VL]\n" "addvl x24, x24, #-3\n" - "ld1w { z15.s }, p2/Z, [x21]\n" - "mov x21, x20\n" - "st1h { z1.h }, p2, [x24]\n" - ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" - "ld1w { z31.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - "incb x20\n" - ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" - "st1h { z9.h }, p2, [x24, #1, MUL VL]\n" - "ld1w { z13.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" - "st1h { z2.h }, p2, [x24, #2, MUL VL]\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + "ld1w { z9.s }, p2/Z, [x20]\n" + "mov x20, x21\n" + "st1h { z21.h }, p2, [x24]\n" + ".inst 0x648aaa21 // bfcvtnt z1.h, p2/M, z17.s\n" + "ld1w { z3.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "incb x21\n" + ".inst 0x658aa864 // bfcvt z4.h, p2/M, z3.s\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa92b // bfcvt z11.h, p2/M, z9.s\n" + "st1h { z1.h }, p2, [x24, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aaa46 // bfcvt z6.h, p2/M, z18.s\n" + "st1h { z11.h }, p2, [x24, #2, MUL VL]\n" + "ld1w { z5.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" "addvl x24, x24, #-3\n" - ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" - "ld1w { z15.s }, p2/Z, [x21]\n" - "mov x21, x20\n" - "st1h { z1.h }, p2, [x24]\n" - ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" - "ld1w { z31.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" - "incb x20\n" - "ld1w { z16.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" - "st1h { z9.h }, p2, [x24, #1, MUL VL]\n" - "ld1w { z13.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" - ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" - ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" - "ld1w { z18.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #5\n" + ".inst 0x648aabe4 // bfcvtnt z4.h, p2/M, z31.s\n" + "ld1w { z27.s }, p2/Z, [x20]\n" + "mov x20, x21\n" + "st1h { z4.h }, p2, [x24]\n" + ".inst 0x648aa8a6 // bfcvtnt z6.h, p2/M, z5.s\n" + "ld1w { z9.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa938 // bfcvt z24.h, p2/M, z9.s\n" + "incb x21\n" + "ld1w { z17.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aab75 // bfcvt z21.h, p2/M, z27.s\n" + "st1h { z6.h }, p2, [x24, #1, MUL VL]\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x648aaa38 // bfcvtnt z24.h, p2/M, z17.s\n" + ".inst 0x658aabf9 // bfcvt z25.h, p2/M, z31.s\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" "ldr x7, [%x[args], %[offsetof_Args_input_cols]]\n" - "st1h { z2.h }, p2, [x24, #2, MUL VL]\n" - "ld1w { z15.s }, p2/Z, [x21]\n" - "mov x21, x20\n" + "st1h { z21.h }, p2, [x24, #2, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x20]\n" + "mov x21, x21\n" "addvl x24, x24, #-3\n" - "st1h { z1.h }, p2, [x24]\n" - "ld1w { z31.s }, p2/Z, [x21]\n" + "st1h { z24.h }, p2, [x24]\n" + "ld1w { z17.s }, p2/Z, [x21]\n" "incb x21, ALL, MUL #5\n" - ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" - "st1h { z9.h }, p2, [x24, #1, MUL VL]\n" - "ld1w { z16.s }, p2/Z, [x21]\n" + ".inst 0x648aaa59 // bfcvtnt z25.h, p2/M, z18.s\n" + "st1h { z25.h }, p2, [x24, #1, MUL VL]\n" + "ld1w { z8.s }, p2/Z, [x21]\n" "incb x21, ALL, MUL #5\n" - ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" - ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" - "ld1w { z13.s }, p2/Z, [x21]\n" + ".inst 0x658aaa29 // bfcvt z9.h, p2/M, z17.s\n" + ".inst 0x658aa976 // bfcvt z22.h, p2/M, z11.s\n" + "ld1w { z28.s }, p2/Z, [x21]\n" "incb x21, ALL, MUL #5\n" - ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + ".inst 0x658aab85 // bfcvt z5.h, p2/M, z28.s\n" "ldr x17, [%x[args], %[offsetof_Args_inptr]]\n" - "ld1w { z18.s }, p2/Z, [x21]\n" + "ld1w { z25.s }, p2/Z, [x21]\n" "incb x21, ALL, MUL #5\n" "sub x20, x7, #0x1\n" - "st1h { z2.h }, p2, [x24, #2, MUL VL]\n" - "ld1w { z15.s }, p2/Z, [x21]\n" + "st1h { z22.h }, p2, [x24, #2, MUL VL]\n" + "ld1w { z11.s }, p2/Z, [x21]\n" "orr x23, x20, %x[ld_in_col], LSL #18\n" "addvl x24, x24, #-3\n" - "mov z5.d, z4.d\n" + "mov z17.d, z16.d\n" "orr x23, x5, x23, LSL #20\n" "mov x22, #0xb\n" - "mov z6.d, z4.d\n" - "mov z7.d, z4.d\n" + "mov z18.d, z16.d\n" + "mov z19.d, z16.d\n" "add x21, x4, x3\n" "lsl x20, %x[ld_in_row], #0x2\n" - ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" - "st1h { z1.h }, p2, [x24]\n" - ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" - "st1h { z9.h }, p2, [x24, #1, MUL VL]\n" - ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" + ".inst 0x648aa909 // bfcvtnt z9.h, p2/M, z8.s\n" + "st1h { z9.h }, p2, [x24]\n" + ".inst 0x648aab25 // bfcvtnt z5.h, p2/M, z25.s\n" + "st1h { z5.h }, p2, [x24, #1, MUL VL]\n" + ".inst 0x658aa97b // bfcvt z27.h, p2/M, z11.s\n" "mov x8, #0x0\n" - "st1h { z2.h }, p2, [x24, #2, MUL VL]\n" + "st1h { z27.h }, p2, [x24, #2, MUL VL]\n" "ldr x16, [%x[args], %[offsetof_Args_output_cols]]\n" "lsl x23, x23, #0x2\n" "sub x22, x22, x21\n" @@ -213,20 +213,20 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( ".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n" "add x20, x20, %x[ld_in_col], LSL #2\n" "bgt 3b\n" - "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x23, [%x[args], %[offsetof_Args_outptrs]]\n" "lsl x20, %x[ld_in_row], #0x2\n" "msub x17, x4, x20, x17\n" - ".inst 0xc0040c80 // mova za.d[x8, #0], { z4.d-z7.d }\n" + ".inst 0xc0040e00 // mova za.d[x8, #0], { z16.d-z19.d }\n" "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" - ".inst 0xc0040c81 // mova za.d[x8, #1], { z4.d-z7.d }\n" + ".inst 0xc0040e01 // mova za.d[x8, #1], { z16.d-z19.d }\n" "mov x22, #0x4\n" - "ldp x15, x14, [x25], #0x10\n" - ".inst 0xc0040c82 // mova za.d[x8, #2], { z4.d-z7.d }\n" + "ldp x15, x14, [x23], #0x10\n" + ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" "ldp x13, x11, [x20], #0x10\n" - ".inst 0xc0040c83 // mova za.d[x8, #3], { z4.d-z7.d }\n" + ".inst 0xc0040e03 // mova za.d[x8, #3], { z16.d-z19.d }\n" "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" - ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" - "ldp x10, x9, [x25], #0x10\n" + ".inst 0xc0040e04 // mova za.d[x8, #4], { z16.d-z19.d }\n" + "ldp x10, x9, [x23], #0x10\n" "ldp x28, x27, [x20], #0x10\n" "cbz x21, 5f\n" "cmp x21, x22\n" @@ -234,21 +234,21 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( "sub x21, x21, x20\n" "sub x22, x22, x20\n" "cbz x21, 5f\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" "and x22, x21, #0x1\n" "add x21, x21, #0x1\n" - ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + ".inst 0xc1acc9a4 // fclamp { z4.s-z7.s }, z13.s, z12.s\n" "lsr x21, x21, #0x1\n" "sub x16, x16, x21\n" "4:" // Left padding "subs x21, x21, #0x1\n" - "st1w { z24.s }, p1, [x15]\n" + "st1w { z4.s }, p1, [x15]\n" "add x15, x15, x13, LSL #2\n" - "st1w { z25.s }, p1, [x14]\n" + "st1w { z5.s }, p1, [x14]\n" "add x14, x14, x11, LSL #2\n" - "st1w { z26.s }, p1, [x10]\n" + "st1w { z6.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - "st1w { z27.s }, p1, [x9]\n" + "st1w { z7.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" "bgt 4b\n" "5:" // Left padding: End @@ -264,331 +264,331 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( "beq 7f\n" "6:" // Unpadded: 4 priming loads "add x21, x17, %x[ld_in_row], LSL #2\n" - "ld1w { z23.s }, p1/Z, [x17]\n" - ".inst 0x658aaaea // bfcvt z10.h, p2/M, z23.s\n" + "ld1w { z0.s }, p1/Z, [x17]\n" + ".inst 0x658aa816 // bfcvt z22.h, p2/M, z0.s\n" "addvl x20, SP, #12\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z9.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + ".inst 0x648aa936 // bfcvtnt z22.h, p2/M, z9.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z28.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aab97 // bfcvt z23.h, p2/M, z28.s\n" + "ld1w { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aaa97 // bfcvtnt z23.h, p2/M, z20.s\n" + "ld1w { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aaa98 // bfcvt z24.h, p2/M, z20.s\n" + "ld1w { z29.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" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aabb8 // bfcvtnt z24.h, p2/M, z29.s\n" + "ld1w { z30.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" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aabd9 // bfcvt z25.h, p2/M, z30.s\n" + "ld1w { z9.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" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aa939 // bfcvtnt z25.h, p2/M, z9.s\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aab5a // bfcvt z26.h, p2/M, z26.s\n" + ".inst 0xa1402a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20]\n" + ".inst 0xc13312d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z3.h\n" + "ld1w { z9.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - "ld1w { z16.s }, p1/Z, [x21]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + ".inst 0x648aa93a // bfcvtnt z26.h, p2/M, z9.s\n" + ".inst 0xc13b12f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z11.h\n" + "ld1w { z9.s }, p1/Z, [x21]\n" + ".inst 0x658aa93b // bfcvt z27.h, p2/M, z9.s\n" + "ld1h { z9.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1391310 // bfdot za.s[x8, 0], { z24.h-z27.h }, z9.h\n" "7:" // Unpadded: 3 priming loads "add x21, x17, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z27.s }, p1/Z, [x17]\n" + ".inst 0x658aab7d // bfcvt z29.h, p2/M, z27.s\n" "addvl x20, SP, #9\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + ".inst 0x648aab5d // bfcvtnt z29.h, p2/M, z26.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z9.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aa93e // bfcvt z30.h, p2/M, z9.s\n" + "ld1w { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aaa9e // bfcvtnt z30.h, p2/M, z20.s\n" + "ld1w { z25.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aab3f // bfcvt z31.h, p2/M, z25.s\n" + "ld1w { z26.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" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aab5f // bfcvtnt z31.h, p2/M, z26.s\n" + "ld1w { z27.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" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aab60 // bfcvt z0.h, p2/M, z27.s\n" + "ld1w { z9.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" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aa920 // bfcvtnt z0.h, p2/M, z9.s\n" + "ld1w { z23.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aaae1 // bfcvt z1.h, p2/M, z23.s\n" + ".inst 0xa0402a84 // ld1h { z4.h-z5.h }, pn10.b/Z, [x20]\n" + ".inst 0xc13413b0 // bfdot za.s[x8, 0], { z29.h-z0.h }, z4.h\n" + "ld1w { z9.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - "ld1w { z16.s }, p1/Z, [x21]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + ".inst 0x648aa921 // bfcvtnt z1.h, p2/M, z9.s\n" + ".inst 0xc13513d0 // bfdot za.s[x8, 0], { z30.h-z1.h }, z5.h\n" + "ld1w { z29.s }, p1/Z, [x21]\n" + ".inst 0x658aaba2 // bfcvt z2.h, p2/M, z29.s\n" + "ld1h { z9.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc13913f0 // bfdot za.s[x8, 0], { z31.h-z2.h }, z9.h\n" "8:" // Unpadded: 2 priming loads "add x22, x17, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z27.s }, p1/Z, [x17]\n" + ".inst 0x658aab7a // bfcvt z26.h, p2/M, z27.s\n" "addvl x21, SP, #6\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + "ld1w { z21.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + ".inst 0x648aaaba // bfcvtnt z26.h, p2/M, z21.s\n" "addvl x20, SP, #12\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + "ld1w { z25.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + ".inst 0x658aab3b // bfcvt z27.h, p2/M, z25.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + "ld1w { z4.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aa89b // bfcvtnt z27.h, p2/M, z4.s\n" + "ld1w { z10.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x658aa95c // bfcvt z28.h, p2/M, z10.s\n" + "ld1w { z4.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" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aa89c // bfcvtnt z28.h, p2/M, z4.s\n" + "ld1w { z5.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" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x658aa8bd // bfcvt z29.h, p2/M, z5.s\n" + "ld1w { z5.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" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aa8bd // bfcvtnt z29.h, p2/M, z5.s\n" + "ld1w { z5.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" - ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x22]\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0x658aa8be // bfcvt z30.h, p2/M, z5.s\n" + ".inst 0xa0402aae // ld1h { z14.h-z15.h }, pn10.b/Z, [x21]\n" + ".inst 0xc13e1350 // bfdot za.s[x8, 0], { z26.h-z29.h }, z14.h\n" + "ld1w { z5.s }, p1/Z, [x22]\n" + ".inst 0x648aa8be // bfcvtnt z30.h, p2/M, z5.s\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x22]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + ".inst 0xc13f1370 // bfdot za.s[x8, 0], { z27.h-z30.h }, z15.h\n" + ".inst 0xa0402a88 // ld1h { z8.h-z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1381351 // bfdot za.s[x8, 1], { z26.h-z29.h }, z8.h\n" + "ld1w { z23.s }, p1/Z, [x22]\n" + ".inst 0x658aaaff // bfcvt z31.h, p2/M, z23.s\n" + ".inst 0xc1391371 // bfdot za.s[x8, 1], { z27.h-z30.h }, z9.h\n" + "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1301390 // bfdot za.s[x8, 0], { z28.h-z31.h }, z0.h\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1301391 // bfdot za.s[x8, 1], { z28.h-z31.h }, z0.h\n" "9:" // Unpadded: 1 priming loads "add x22, x17, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z27.s }, p1/Z, [x17]\n" + ".inst 0x658aab77 // bfcvt z23.h, p2/M, z27.s\n" "addvl x21, SP, #3\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + "ld1w { z24.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + ".inst 0x648aab17 // bfcvtnt z23.h, p2/M, z24.s\n" "addvl x20, SP, #9\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + "ld1w { z31.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + ".inst 0x658aabf8 // bfcvt z24.h, p2/M, z31.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + "ld1w { z6.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aa8d8 // bfcvtnt z24.h, p2/M, z6.s\n" + "ld1w { z28.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x658aab99 // bfcvt z25.h, p2/M, z28.s\n" + "ld1w { z26.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" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aab59 // bfcvtnt z25.h, p2/M, z26.s\n" + "ld1w { z28.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" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x658aab9a // bfcvt z26.h, p2/M, z28.s\n" + "ld1w { z4.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" - "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aa89a // bfcvtnt z26.h, p2/M, z4.s\n" + "ld1w { z20.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" - ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x22]\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0x658aaa9b // bfcvt z27.h, p2/M, z20.s\n" + ".inst 0xa1402aa0 // ld1h { z0.h, z8.h }, pn10.b/Z, [x21]\n" + ".inst 0xc13012f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z0.h\n" + "ld1w { z20.s }, p1/Z, [x22]\n" + ".inst 0x648aaa9b // bfcvtnt z27.h, p2/M, z20.s\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x22]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + ".inst 0xc1381310 // bfdot za.s[x8, 0], { z24.h-z27.h }, z8.h\n" + ".inst 0xa0402a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20]\n" + ".inst 0xc13212f1 // bfdot za.s[x8, 1], { z23.h-z26.h }, z2.h\n" + "ld1w { z11.s }, p1/Z, [x22]\n" + ".inst 0x658aa97c // bfcvt z28.h, p2/M, z11.s\n" + ".inst 0xc1331311 // bfdot za.s[x8, 1], { z24.h-z27.h }, z3.h\n" + "ld1h { z4.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1341330 // bfdot za.s[x8, 0], { z25.h-z28.h }, z4.h\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1301331 // bfdot za.s[x8, 1], { z25.h-z28.h }, z0.h\n" "10:" // Unpadded: 0 priming loads "cmp x7, #0x2\n" - ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" - "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0xa1402be3 // ld1h { z3.h, z11.h }, pn10.b/Z, [SP]\n" + "ld1h { z7.h }, p2/Z, [SP, #2, MUL VL]\n" "blt 20f\n" "add x21, x17, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z27.s }, p1/Z, [x17]\n" + ".inst 0x658aab75 // bfcvt z21.h, p2/M, z27.s\n" "sub x7, x7, #0x2\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" "sub x16, x16, #0x1\n" - ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aab55 // bfcvtnt z21.h, p2/M, z26.s\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + ".inst 0x658aab56 // bfcvt z22.h, p2/M, z26.s\n" "lsr x20, x7, #0x1\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" "cmp x20, x16\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aab56 // bfcvtnt z22.h, p2/M, z26.s\n" + "ld1w { z8.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aa917 // bfcvt z23.h, p2/M, z8.s\n" "csel x26, x20, x16, LT\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z2.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" + ".inst 0x648aa857 // bfcvtnt z23.h, p2/M, z2.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z6.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" + ".inst 0x658aa8d8 // bfcvt z24.h, p2/M, z6.s\n" "and x7, x7, #0x1\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z15.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" + ".inst 0x648aa9f8 // bfcvtnt z24.h, p2/M, z15.s\n" "sub x16, x16, x26\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "ld1w { z27.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aab79 // bfcvt z25.h, p2/M, z27.s\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x21]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0x648aab59 // bfcvtnt z25.h, p2/M, z26.s\n" + "ld1w { z27.s }, p1/Z, [x21]\n" + ".inst 0x658aab7a // bfcvt z26.h, p2/M, z27.s\n" "cbz x26, 19f\n" "11:" // Unpadded: Main loop - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc13312b0 // bfdot za.s[x8, 0], { z21.h-z24.h }, z3.h\n" "addvl x25, SP, #6\n" "addvl x24, SP, #12\n" - "ld1w { z18.s }, p1/Z, [x17]\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402b21 // ld1h { z1.h, z9.h }, pn10.b/Z, [x25]\n" + "ld1w { z14.s }, p1/Z, [x17]\n" + ".inst 0xc13b12d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z11.h\n" + ".inst 0xa1402b20 // ld1h { z0.h, z8.h }, pn10.b/Z, [x25]\n" "add x23, x17, %x[ld_in_row], LSL #2\n" "addvl x22, SP, #3\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z17.s }, p1/Z, [x23]\n" + ".inst 0xc13012b1 // bfdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + "ld1w { z27.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402b01 // ld1h { z1.h, z9.h }, pn10.b/Z, [x24]\n" + ".inst 0xc13812d1 // bfdot za.s[x8, 1], { z22.h-z25.h }, z8.h\n" + ".inst 0xa1402b00 // ld1h { z0.h, z8.h }, pn10.b/Z, [x24]\n" "addvl x21, SP, #9\n" "add x20, x17, %x[ld_in_row], LSL #2\n" - ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p1/Z, [x23]\n" + ".inst 0xc13012b2 // bfdot za.s[x8, 2], { z21.h-z24.h }, z0.h\n" + "ld1w { z2.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa4a // bfcvt z10.h, p2/M, z18.s\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x25, #2, MUL VL]\n" - ".inst 0x648aaa2a // bfcvtnt z10.h, p2/M, z17.s\n" + ".inst 0x658aa9d5 // bfcvt z21.h, p2/M, z14.s\n" + ".inst 0xc13712f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z7.h\n" + "ld1h { z11.h }, p2/Z, [x25, #2, MUL VL]\n" + ".inst 0x648aab75 // bfcvtnt z21.h, p2/M, z27.s\n" "subs x26, x26, #0x1\n" - "ld1w { z17.s }, p1/Z, [x23]\n" + "ld1w { z14.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" - ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x23]\n" + ".inst 0xc13812d2 // bfdot za.s[x8, 2], { z22.h-z25.h }, z8.h\n" + ".inst 0x658aa856 // bfcvt z22.h, p2/M, z2.s\n" + "ld1w { z7.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - ".inst 0x648aaa2b // bfcvtnt z11.h, p2/M, z17.s\n" - "ld1w { z17.s }, p1/Z, [x23]\n" + ".inst 0xc13b12f1 // bfdot za.s[x8, 1], { z23.h-z26.h }, z11.h\n" + ".inst 0x648aa9d6 // bfcvtnt z22.h, p2/M, z14.s\n" + "ld1w { z31.s }, p1/Z, [x23]\n" "add x23, x23, %x[ld_in_row], LSL #2\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" - ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" - "ld1h { z2.h }, p2/Z, [x24, #2, MUL VL]\n" - ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0xc0060c08 // mova { z8.d-z11.d }, za.d[x8, #0]\n" + ".inst 0xc1acc9a8 // fclamp { z8.s-z11.s }, z13.s, z12.s\n" + "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n" + ".inst 0xc13012f2 // bfdot za.s[x8, 2], { z23.h-z26.h }, z0.h\n" + ".inst 0x658aa8f7 // bfcvt z23.h, p2/M, z7.s\n" "add x8, x8, #0x1\n" - "ld1w { z16.s }, p1/Z, [x23]\n" + "ld1w { z26.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" - ".inst 0x648aaa2c // bfcvtnt z12.h, p2/M, z17.s\n" - "ld1w { z16.s }, p1/Z, [x23]\n" + ".inst 0x658aab58 // bfcvt z24.h, p2/M, z26.s\n" + ".inst 0x648aabf7 // bfcvtnt z23.h, p2/M, z31.s\n" + "ld1w { z2.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" - "st1w { z24.s }, p1, [x15]\n" - "ld1w { z16.s }, p1/Z, [x23]\n" + ".inst 0x648aa858 // bfcvtnt z24.h, p2/M, z2.s\n" + "st1w { z8.s }, p1, [x15]\n" + "ld1w { z0.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" + ".inst 0x658aa819 // bfcvt z25.h, p2/M, z0.s\n" "add x15, x15, x13, LSL #2\n" - ".inst 0xa1402ac1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x22]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" - "st1w { z25.s }, p1, [x14]\n" + ".inst 0xa0402ac2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x22]\n" + ".inst 0xc13212b0 // bfdot za.s[x8, 0], { z21.h-z24.h }, z2.h\n" + "st1w { z9.s }, p1, [x14]\n" "add x14, x14, x11, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x23]\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "ld1w { z26.s }, p1/Z, [x23]\n" + ".inst 0x648aab59 // bfcvtnt z25.h, p2/M, z26.s\n" "add x23, x23, %x[ld_in_row], LSL #2\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xc13312d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z3.h\n" ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "st1w { z26.s }, p1, [x10]\n" + ".inst 0xc13112b1 // bfdot za.s[x8, 1], { z21.h-z24.h }, z1.h\n" + "st1w { z10.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x23]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - "ld1w { z16.s }, p1/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" - "st1w { z27.s }, p1, [x9]\n" + "ld1w { z26.s }, p1/Z, [x23]\n" + ".inst 0x658aab5a // bfcvt z26.h, p2/M, z26.s\n" + ".inst 0xc13912d1 // bfdot za.s[x8, 1], { z22.h-z25.h }, z9.h\n" + "ld1w { z31.s }, p1/Z, [x17]\n" + ".inst 0x658aabf5 // bfcvt z21.h, p2/M, z31.s\n" + "st1w { z11.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z30.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" - ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0xc0040e04 // mova za.d[x8, #4], { z16.d-z19.d }\n" + ".inst 0x648aabd5 // bfcvtnt z21.h, p2/M, z30.s\n" + "ld1w { z0.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + ".inst 0x658aa816 // bfcvt z22.h, p2/M, z0.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z1.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x648aa836 // bfcvtnt z22.h, p2/M, z1.s\n" + "ld1w { z11.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" "ld1h { z2.h }, p2/Z, [x22, #2, MUL VL]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1w { z19.s }, p1/Z, [x20]\n" + ".inst 0xc13212f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z2.h\n" + "ld1w { z28.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z17.s }, p1/Z, [x20]\n" + "ld1w { z14.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z18.s }, p1/Z, [x20]\n" + "ld1w { z27.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1h { z4.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc13412f1 // bfdot za.s[x8, 1], { z23.h-z26.h }, z4.h\n" + ".inst 0x658aa977 // bfcvt z23.h, p2/M, z11.s\n" + "ld1w { z29.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa2d // bfcvt z13.h, p2/M, z17.s\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - "ld1w { z17.s }, p1/Z, [x20]\n" + ".inst 0x658aa9d8 // bfcvt z24.h, p2/M, z14.s\n" + ".inst 0x658aabb9 // bfcvt z25.h, p2/M, z29.s\n" + "ld1w { z5.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa6c // bfcvtnt z12.h, p2/M, z19.s\n" - ".inst 0x648aaa4d // bfcvtnt z13.h, p2/M, z18.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - ".inst 0x648aaa2e // bfcvtnt z14.h, p2/M, z17.s\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" - "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0x648aab97 // bfcvtnt z23.h, p2/M, z28.s\n" + ".inst 0x648aab78 // bfcvtnt z24.h, p2/M, z27.s\n" + "ld1w { z11.s }, p1/Z, [x20]\n" + ".inst 0x648aa8b9 // bfcvtnt z25.h, p2/M, z5.s\n" + ".inst 0x658aa97a // bfcvt z26.h, p2/M, z11.s\n" + ".inst 0xa1402be3 // ld1h { z3.h, z11.h }, pn10.b/Z, [SP]\n" + "ld1h { z7.h }, p2/Z, [SP, #2, MUL VL]\n" "bgt 11b\n" "b 19f\n" "12:" // Padded @@ -603,282 +603,282 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( "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, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z1.s }, p0/Z, [x17]\n" + ".inst 0x658aa837 // bfcvt z23.h, p2/M, z1.s\n" "add x21, x17, %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 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z29.s }, p0/Z, [x21]\n" + ".inst 0x648aabb7 // bfcvtnt z23.h, p2/M, z29.s\n" "add x21, x21, %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, [x21]\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z30.s }, p0/Z, [x21]\n" + ".inst 0x658aabd8 // bfcvt z24.h, p2/M, z30.s\n" "add x21, x21, %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, [x21]\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z15.s }, p0/Z, [x21]\n" + ".inst 0x648aa9f8 // bfcvtnt z24.h, p2/M, z15.s\n" "mov x12, #0x4\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z27.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aab79 // bfcvt z25.h, p2/M, z27.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z20.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x648aaa99 // bfcvtnt z25.h, p2/M, z20.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z10.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aa95a // bfcvt z26.h, p2/M, z10.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x8\n" - "ld1w { z16.s }, p0/Z, [x21]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z8.s }, p0/Z, [x21]\n" + ".inst 0x648aa91a // bfcvtnt z26.h, p2/M, z8.s\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z28.s }, p0/Z, [x21]\n" + ".inst 0x658aab9b // bfcvt z27.h, p2/M, z28.s\n" "addvl x20, SP, #12\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc13112f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z1.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z28.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0x648aab9b // bfcvtnt z27.h, p2/M, z28.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z0.s }, p0/Z, [x21]\n" + ".inst 0x658aa81c // bfcvt z28.h, p2/M, z0.s\n" + ".inst 0xc1391310 // bfdot za.s[x8, 0], { z24.h-z27.h }, z9.h\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + ".inst 0xc1301330 // bfdot za.s[x8, 0], { z25.h-z28.h }, z0.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, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z21.s }, p0/Z, [x17]\n" + ".inst 0x658aaab4 // bfcvt z20.h, p2/M, z21.s\n" "add x21, x17, %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 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x21]\n" + ".inst 0x648aab74 // bfcvtnt z20.h, p2/M, z27.s\n" "add x21, x21, %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, [x21]\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x21]\n" + ".inst 0x658aab75 // bfcvt z21.h, p2/M, z27.s\n" "add x21, x21, %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, [x21]\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x21]\n" + ".inst 0x648aab75 // bfcvtnt z21.h, p2/M, z27.s\n" "mov x12, #0x4\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z29.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aabb6 // bfcvt z22.h, p2/M, z29.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z27.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x648aab76 // bfcvtnt z22.h, p2/M, z27.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z27.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aab77 // bfcvt z23.h, p2/M, z27.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x8\n" - "ld1w { z16.s }, p0/Z, [x21]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z8.s }, p0/Z, [x21]\n" + ".inst 0x648aa917 // bfcvtnt z23.h, p2/M, z8.s\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z28.s }, p0/Z, [x21]\n" + ".inst 0x658aab98 // bfcvt z24.h, p2/M, z28.s\n" "addvl x20, SP, #9\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc1311290 // bfdot za.s[x8, 0], { z20.h-z23.h }, z1.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z0.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0x648aa818 // bfcvtnt z24.h, p2/M, z0.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x21]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z1.s }, p0/Z, [x21]\n" + ".inst 0x658aa839 // bfcvt z25.h, p2/M, z1.s\n" + ".inst 0xc13912b0 // bfdot za.s[x8, 0], { z21.h-z24.h }, z9.h\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + ".inst 0xc13012d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z0.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, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z6.s }, p0/Z, [x17]\n" + ".inst 0x658aa8da // bfcvt z26.h, p2/M, z6.s\n" "add x22, x17, %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 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z29.s }, p0/Z, [x22]\n" + ".inst 0x648aabba // bfcvtnt z26.h, p2/M, z29.s\n" "add x22, x22, %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, [x22]\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z28.s }, p0/Z, [x22]\n" + ".inst 0x658aab9b // bfcvt z27.h, p2/M, z28.s\n" "add x22, x22, %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, [x22]\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z14.s }, p0/Z, [x22]\n" + ".inst 0x648aa9db // bfcvtnt z27.h, p2/M, z14.s\n" "mov x12, #0x4\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z24.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aab1c // bfcvt z28.h, p2/M, z24.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z1.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x648aa83c // bfcvtnt z28.h, p2/M, z1.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z3.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aa87d // bfcvt z29.h, p2/M, z3.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x8\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z0.s }, p0/Z, [x22]\n" + ".inst 0x648aa81d // bfcvtnt z29.h, p2/M, z0.s\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z24.s }, p0/Z, [x22]\n" + ".inst 0x658aab1e // bfcvt z30.h, p2/M, z24.s\n" "addvl x21, SP, #6\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc1311350 // bfdot za.s[x8, 0], { z26.h-z29.h }, z1.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "ld1w { z23.s }, p0/Z, [x22]\n" + ".inst 0x648aaafe // bfcvtnt z30.h, p2/M, z23.s\n" "addvl x20, SP, #12\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1391370 // bfdot za.s[x8, 0], { z27.h-z30.h }, z9.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" + "ld1w { z31.s }, p0/Z, [x22]\n" + ".inst 0xc1301351 // bfdot za.s[x8, 1], { z26.h-z29.h }, z0.h\n" + ".inst 0x658aabff // bfcvt z31.h, p2/M, z31.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1311371 // bfdot za.s[x8, 1], { z27.h-z30.h }, z1.h\n" + ".inst 0xc1301390 // bfdot za.s[x8, 0], { z28.h-z31.h }, z0.h\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1301391 // bfdot za.s[x8, 1], { z28.h-z31.h }, z0.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, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z22.s }, p0/Z, [x17]\n" + ".inst 0x658aaad5 // bfcvt z21.h, p2/M, z22.s\n" "add x22, x17, %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 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z3.s }, p0/Z, [x22]\n" + ".inst 0x648aa875 // bfcvtnt z21.h, p2/M, z3.s\n" "add x22, x22, %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, [x22]\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z20.s }, p0/Z, [x22]\n" + ".inst 0x658aaa96 // bfcvt z22.h, p2/M, z20.s\n" "add x22, x22, %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, [x22]\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z25.s }, p0/Z, [x22]\n" + ".inst 0x648aab36 // bfcvtnt z22.h, p2/M, z25.s\n" "mov x12, #0x4\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z24.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aab17 // bfcvt z23.h, p2/M, z24.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z0.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x648aa817 // bfcvtnt z23.h, p2/M, z0.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z7.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aa8f8 // bfcvt z24.h, p2/M, z7.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x8\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z28.s }, p0/Z, [x22]\n" + ".inst 0x648aab98 // bfcvtnt z24.h, p2/M, z28.s\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z6.s }, p0/Z, [x22]\n" + ".inst 0x658aa8d9 // bfcvt z25.h, p2/M, z6.s\n" "addvl x21, SP, #3\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc13112b0 // bfdot za.s[x8, 0], { z21.h-z24.h }, z1.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "ld1w { z6.s }, p0/Z, [x22]\n" + ".inst 0x648aa8d9 // bfcvtnt z25.h, p2/M, z6.s\n" "addvl x20, SP, #9\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc13912d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z9.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" + "ld1w { z3.s }, p0/Z, [x22]\n" + ".inst 0xc13012b1 // bfdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + ".inst 0x658aa87a // bfcvt z26.h, p2/M, z3.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc13112d1 // bfdot za.s[x8, 1], { z22.h-z25.h }, z1.h\n" + ".inst 0xc13012f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z0.h\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc13012f1 // bfdot za.s[x8, 1], { z23.h-z26.h }, z0.h\n" "17:" // Padded: 0 priming loads "cmp x7, #0x2\n" - ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" - "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0xa1402be3 // ld1h { z3.h, z11.h }, pn10.b/Z, [SP]\n" + "ld1h { z7.h }, p2/Z, [SP, #2, MUL VL]\n" "blt 20f\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z25.s }, p0/Z, [x17]\n" + ".inst 0x658aab35 // bfcvt z21.h, p2/M, z25.s\n" "add x20, x17, %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 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x20]\n" + ".inst 0x648aab75 // bfcvtnt z21.h, p2/M, z27.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 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x20]\n" + ".inst 0x658aab76 // bfcvt z22.h, p2/M, z27.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 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x20]\n" + ".inst 0x648aab76 // bfcvtnt z22.h, p2/M, z27.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" + "ld1w { z27.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aab77 // bfcvt z23.h, p2/M, z27.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z25.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x648aab37 // bfcvtnt z23.h, p2/M, z25.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z26.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aab58 // bfcvt z24.h, p2/M, z26.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x8\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x20]\n" + ".inst 0x648aab78 // bfcvtnt z24.h, p2/M, z27.s\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" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x20]\n" + ".inst 0x658aab79 // bfcvt z25.h, p2/M, z27.s\n" "add x20, x20, %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 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "ld1w { z26.s }, p0/Z, [x20]\n" + ".inst 0x648aab59 // bfcvtnt z25.h, p2/M, z26.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 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "ld1w { z27.s }, p0/Z, [x20]\n" + ".inst 0x658aab7a // bfcvt z26.h, p2/M, z27.s\n" "sub x7, x7, #0x2\n" "sub x16, x16, #0x1\n" "lsr x20, x7, #0x1\n" @@ -889,323 +889,323 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( "sub x16, x16, x24\n" "cbz x24, 19f\n" "18:" // Padded: Main loop - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc13312b0 // bfdot za.s[x8, 0], { z21.h-z24.h }, z3.h\n" "addvl x23, SP, #6\n" "addvl x21, SP, #12\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402ae1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x23]\n" + ".inst 0xc13b12d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z11.h\n" + ".inst 0xa0402ae0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x23]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p0/Z, [x17]\n" + ".inst 0xc13012b1 // bfdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + "ld1w { z9.s }, p0/Z, [x17]\n" "add x20, x17, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" + ".inst 0xc13112d1 // bfdot za.s[x8, 1], { z22.h-z25.h }, z1.h\n" + ".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n" "addvl x22, SP, #3\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" - "ld1w { z20.s }, p0/Z, [x20]\n" + ".inst 0xc13012b2 // bfdot za.s[x8, 2], { z21.h-z24.h }, z0.h\n" + "ld1w { z14.s }, p0/Z, [x20]\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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z27.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 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + ".inst 0xc13712f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z7.h\n" "mov x12, #0x4\n" - "ld1h { z2.h }, p2/Z, [x23, #2, MUL VL]\n" - ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" - "ld1w { z19.s }, p0/Z, [x20]\n" + "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" + ".inst 0xc13112d2 // bfdot za.s[x8, 2], { z22.h-z25.h }, z1.h\n" + ".inst 0x658aa921 // bfcvt z1.h, p2/M, z9.s\n" + "ld1w { z15.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0xc13012f1 // bfdot za.s[x8, 1], { z23.h-z26.h }, z0.h\n" + "ld1w { z9.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0x658aaa2b // bfcvt z11.h, p2/M, z17.s\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + ".inst 0x658aab62 // bfcvt z2.h, p2/M, z27.s\n" + "ld1w { z27.s }, p0/Z, [x20]\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 0x648aaa8a // bfcvtnt z10.h, p2/M, z20.s\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x648aa9c1 // bfcvtnt z1.h, p2/M, z14.s\n" + "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc13012f2 // bfdot za.s[x8, 2], { z23.h-z26.h }, z0.h\n" + ".inst 0x658aa923 // bfcvt z3.h, p2/M, z9.s\n" "addvl x21, SP, #9\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z9.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 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aa924 // bfcvt z4.h, p2/M, z9.s\n" "mov x12, #0x8\n" - "ld1w { z17.s }, p0/Z, [x20]\n" + "ld1w { z24.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa6b // bfcvtnt z11.h, p2/M, z19.s\n" + ".inst 0x648aa9e2 // bfcvtnt z2.h, p2/M, z15.s\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z9.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa4c // bfcvtnt z12.h, p2/M, z18.s\n" + ".inst 0x648aab63 // bfcvtnt z3.h, p2/M, z27.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0x648aaa2d // bfcvtnt z13.h, p2/M, z17.s\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - ".inst 0xa1402ac1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x22]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aab04 // bfcvtnt z4.h, p2/M, z24.s\n" + ".inst 0x658aa925 // bfcvt z5.h, p2/M, z9.s\n" + ".inst 0xa1402ac0 // ld1h { z0.h, z8.h }, pn10.b/Z, [x22]\n" + "ld1w { z30.s }, p0/Z, [x20]\n" ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0x648aabc5 // bfcvtnt z5.h, p2/M, z30.s\n" + ".inst 0xc1301030 // bfdot za.s[x8, 0], { z1.h-z4.h }, z0.h\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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z29.s }, p0/Z, [x20]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - "ld1w { z16.s }, p0/Z, [x17]\n" + ".inst 0xc1381050 // bfdot za.s[x8, 0], { z2.h-z5.h }, z8.h\n" + ".inst 0xa1402aa6 // ld1h { z6.h, z14.h }, pn10.b/Z, [x21]\n" + "ld1w { z0.s }, p0/Z, [x17]\n" "add x20, x17, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z21.s }, p0/Z, [x20]\n" + ".inst 0xc1361031 // bfdot za.s[x8, 1], { z1.h-z4.h }, z6.h\n" + "ld1w { z10.s }, p0/Z, [x20]\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 0x658aaa2f // bfcvt z15.h, p2/M, z17.s\n" - "ld1w { z17.s }, p0/Z, [x20]\n" + ".inst 0x658aaba6 // bfcvt z6.h, p2/M, z29.s\n" + "ld1w { z9.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 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xc13e1051 // bfdot za.s[x8, 1], { z2.h-z5.h }, z14.h\n" "mov x12, #0x4\n" - "ld1w { z20.s }, p0/Z, [x20]\n" + "ld1w { z29.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + ".inst 0x658aa815 // bfcvt z21.h, p2/M, z0.s\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z31.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa2b // bfcvt z11.h, p2/M, z17.s\n" + ".inst 0x658aa936 // bfcvt z22.h, p2/M, z9.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1h { z2.h }, p2/Z, [x22, #2, MUL VL]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z0.h }, p2/Z, [x22, #2, MUL VL]\n" + ".inst 0xc1301070 // bfdot za.s[x8, 0], { z3.h-z6.h }, z0.h\n" "subs x24, x24, #0x1\n" - "ld1w { z19.s }, p0/Z, [x20]\n" + "ld1w { z15.s }, p0/Z, [x20]\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 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" - "ld1w { z17.s }, p0/Z, [x20]\n" + ".inst 0xc1acc9b8 // fclamp { z24.s-z27.s }, z13.s, z12.s\n" + "ld1w { z30.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" "st1w { z24.s }, p1, [x15]\n" "mov x12, #0x8\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + "ld1w { z14.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" "st1w { z25.s }, p1, [x14]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1301071 // bfdot za.s[x8, 1], { z3.h-z6.h }, z0.h\n" + ".inst 0x658aabf7 // bfcvt z23.h, p2/M, z31.s\n" + "ld1w { z8.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0x658aaa2d // bfcvt z13.h, p2/M, z17.s\n" - "ld1w { z17.s }, p0/Z, [x20]\n" + ".inst 0x658aabd8 // bfcvt z24.h, p2/M, z30.s\n" + "ld1w { z4.s }, p0/Z, [x20]\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 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aa919 // bfcvt z25.h, p2/M, z8.s\n" + "ld1w { z5.s }, p0/Z, [x20]\n" "add x15, x15, x13, LSL #2\n" "add x14, x14, x11, LSL #2\n" "st1w { z26.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" "st1w { z27.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" - ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" - ".inst 0x648aaaaa // bfcvtnt z10.h, p2/M, z21.s\n" - ".inst 0x648aaa8b // bfcvtnt z11.h, p2/M, z20.s\n" + ".inst 0xc0040e04 // mova za.d[x8, #4], { z16.d-z19.d }\n" + ".inst 0xa1402be3 // ld1h { z3.h, z11.h }, pn10.b/Z, [SP]\n" + ".inst 0x648aa955 // bfcvtnt z21.h, p2/M, z10.s\n" + ".inst 0x648aabb6 // bfcvtnt z22.h, p2/M, z29.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" - ".inst 0x648aaa6c // bfcvtnt z12.h, p2/M, z19.s\n" - ".inst 0x648aaa4d // bfcvtnt z13.h, p2/M, z18.s\n" - ".inst 0x648aaa2e // bfcvtnt z14.h, p2/M, z17.s\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "ld1h { z7.h }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0x648aa9f7 // bfcvtnt z23.h, p2/M, z15.s\n" + ".inst 0x648aa9d8 // bfcvtnt z24.h, p2/M, z14.s\n" + ".inst 0x648aa899 // bfcvtnt z25.h, p2/M, z4.s\n" + ".inst 0x658aa8ba // bfcvt z26.h, p2/M, z5.s\n" "bgt 18b\n" "19:" // Main loop tail - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0xc13312b0 // bfdot za.s[x8, 0], { z21.h-z24.h }, z3.h\n" "addvl x24, SP, #6\n" "addvl x23, SP, #12\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402b01 // ld1h { z1.h, z9.h }, pn10.b/Z, [x24]\n" + ".inst 0xc13b12d0 // bfdot za.s[x8, 0], { z22.h-z25.h }, z11.h\n" + ".inst 0xa0402b00 // ld1h { z0.h-z1.h }, pn10.b/Z, [x24]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p0/Z, [x17]\n" + ".inst 0xc13012b1 // bfdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + "ld1w { z5.s }, p0/Z, [x17]\n" "add x22, x17, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402ae1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x23]\n" + ".inst 0xc13112d1 // bfdot za.s[x8, 1], { z22.h-z25.h }, z1.h\n" + ".inst 0xa0402ae0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x23]\n" "addvl x21, SP, #3\n" "addvl x20, SP, #9\n" - ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" - "ld1w { z20.s }, p0/Z, [x22]\n" + ".inst 0xc13012b2 // bfdot za.s[x8, 2], { z21.h-z24.h }, z0.h\n" + "ld1w { z29.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z17.s }, p0/Z, [x22]\n" + "ld1w { z2.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + ".inst 0xc13712f0 // bfdot za.s[x8, 0], { z23.h-z26.h }, z7.h\n" "mov x12, #0x4\n" - "ld1h { z2.h }, p2/Z, [x24, #2, MUL VL]\n" - ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" - "ld1w { z19.s }, p0/Z, [x22]\n" + "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n" + ".inst 0xc13112d2 // bfdot za.s[x8, 2], { z22.h-z25.h }, z1.h\n" + ".inst 0x658aa8bb // bfcvt z27.h, p2/M, z5.s\n" + "ld1w { z20.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + ".inst 0xc13012f1 // bfdot za.s[x8, 1], { z23.h-z26.h }, z0.h\n" + "ld1w { z1.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0x658aaa2b // bfcvt z11.h, p2/M, z17.s\n" - "ld1w { z18.s }, p0/Z, [x22]\n" + ".inst 0x658aa85c // bfcvt z28.h, p2/M, z2.s\n" + "ld1w { z14.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0x648aaa8a // bfcvtnt z10.h, p2/M, z20.s\n" - "ld1h { z2.h }, p2/Z, [x23, #2, MUL VL]\n" - ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x648aabbb // bfcvtnt z27.h, p2/M, z29.s\n" + "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" + ".inst 0xc13012f2 // bfdot za.s[x8, 2], { z23.h-z26.h }, z0.h\n" + ".inst 0x658aa83d // bfcvt z29.h, p2/M, z1.s\n" "add x17, x17, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z1.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aa83e // bfcvt z30.h, p2/M, z1.s\n" "mov x12, #0x8\n" - "ld1w { z17.s }, p0/Z, [x22]\n" + "ld1w { z31.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa6b // bfcvtnt z11.h, p2/M, z19.s\n" + ".inst 0x648aaa9c // bfcvtnt z28.h, p2/M, z20.s\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z26.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa4c // bfcvtnt z12.h, p2/M, z18.s\n" + ".inst 0x648aa9dd // bfcvtnt z29.h, p2/M, z14.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0x648aaa2d // bfcvtnt z13.h, p2/M, z17.s\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" - ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + ".inst 0x648aabfe // bfcvtnt z30.h, p2/M, z31.s\n" + ".inst 0x658aab5f // bfcvt z31.h, p2/M, z26.s\n" + ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" + "ld1w { z9.s }, p0/Z, [x22]\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" - ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0x648aa93f // bfcvtnt z31.h, p2/M, z9.s\n" + ".inst 0xc1321370 // bfdot za.s[x8, 0], { z27.h-z30.h }, z2.h\n" "add x22, x22, %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, [x22]\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - "st1w { z24.s }, p1, [x15]\n" + "ld1w { z26.s }, p0/Z, [x22]\n" + ".inst 0xc13a1390 // bfdot za.s[x8, 0], { z28.h-z31.h }, z10.h\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0x658aab40 // bfcvt z0.h, p2/M, z26.s\n" + ".inst 0xc1321371 // bfdot za.s[x8, 1], { z27.h-z30.h }, z2.h\n" + "ld1h { z9.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1acc9a4 // fclamp { z4.s-z7.s }, z13.s, z12.s\n" + ".inst 0xc13a1391 // bfdot za.s[x8, 1], { z28.h-z31.h }, z10.h\n" + "st1w { z4.s }, p1, [x15]\n" "add x15, x15, x13, LSL #2\n" - ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - "st1w { z25.s }, p1, [x14]\n" + ".inst 0xa1402be3 // ld1h { z3.h, z11.h }, pn10.b/Z, [SP]\n" + ".inst 0xc13913b0 // bfdot za.s[x8, 0], { z29.h-z0.h }, z9.h\n" + "ld1h { z9.h }, p2/Z, [x20, #2, MUL VL]\n" + "st1w { z5.s }, p1, [x14]\n" "add x14, x14, x11, LSL #2\n" - "st1w { z26.s }, p1, [x10]\n" + "st1w { z6.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" - "st1w { z27.s }, p1, [x9]\n" + ".inst 0xc0040e04 // mova za.d[x8, #4], { z16.d-z19.d }\n" + "st1w { z7.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0xc13913b1 // bfdot za.s[x8, 1], { z29.h-z0.h }, z9.h\n" + "ld1h { z7.h }, p2/Z, [SP, #2, MUL VL]\n" "20:" // Main loop skip tail "cbz x7, 21f\n" // Skip remainder inputs "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x17]\n" - ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z25.s }, p0/Z, [x17]\n" + ".inst 0x658aab3d // bfcvt z29.h, p2/M, z25.s\n" "add x22, x17, %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 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z26.s }, p0/Z, [x22]\n" + ".inst 0x648aab5d // bfcvtnt z29.h, p2/M, z26.s\n" "add x22, x22, %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, [x22]\n" - ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z25.s }, p0/Z, [x22]\n" + ".inst 0x658aab3e // bfcvt z30.h, p2/M, z25.s\n" "add x22, x22, %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, [x22]\n" - ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z24.s }, p0/Z, [x22]\n" + ".inst 0x648aab1e // bfcvtnt z30.h, p2/M, z24.s\n" "mov x12, #0x4\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z26.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x658aab5f // bfcvt z31.h, p2/M, z26.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z9.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x648aa93f // bfcvtnt z31.h, p2/M, z9.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" + "ld1w { z9.s }, p0/Z, [x22]\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x658aa920 // bfcvt z0.h, p2/M, z9.s\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x8\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z24.s }, p0/Z, [x22]\n" + ".inst 0x648aab00 // bfcvtnt z0.h, p2/M, z24.s\n" "add x22, x22, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z9.s }, p0/Z, [x22]\n" + ".inst 0x658aa921 // bfcvt z1.h, p2/M, z9.s\n" "add x22, x22, %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 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" - ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "ld1w { z25.s }, p0/Z, [x22]\n" + ".inst 0x648aab21 // bfcvtnt z1.h, p2/M, z25.s\n" + ".inst 0xc13313b0 // bfdot za.s[x8, 0], { z29.h-z0.h }, z3.h\n" "addvl x21, SP, #6\n" "add x22, x22, %x[ld_in_row], LSL #2\n" - ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" + ".inst 0xc13b13d0 // bfdot za.s[x8, 0], { z30.h-z1.h }, z11.h\n" + ".inst 0xa0402aae // ld1h { z14.h-z15.h }, pn10.b/Z, [x21]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "addvl x20, SP, #12\n" - ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" - "ld1w { z16.s }, p0/Z, [x22]\n" - ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc13e13b1 // bfdot za.s[x8, 1], { z29.h-z0.h }, z14.h\n" + "ld1w { z25.s }, p0/Z, [x22]\n" + ".inst 0x658aab22 // bfcvt z2.h, p2/M, z25.s\n" "sub x16, x16, #0x1\n" - ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" - ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" - ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" - ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" - ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" - "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" - ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" - "st1w { z24.s }, p1, [x15]\n" + ".inst 0xc13f13d1 // bfdot za.s[x8, 1], { z30.h-z1.h }, z15.h\n" + ".inst 0xa0402a8e // ld1h { z14.h-z15.h }, pn10.b/Z, [x20]\n" + ".inst 0xc13e13b2 // bfdot za.s[x8, 2], { z29.h-z0.h }, z14.h\n" + ".inst 0xc13713f0 // bfdot za.s[x8, 0], { z31.h-z2.h }, z7.h\n" + "ld1h { z4.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc13f13d2 // bfdot za.s[x8, 2], { z30.h-z1.h }, z15.h\n" + ".inst 0xc13413f1 // bfdot za.s[x8, 1], { z31.h-z2.h }, z4.h\n" + "ld1h { z9.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + ".inst 0xc1acc9a4 // fclamp { z4.s-z7.s }, z13.s, z12.s\n" + "st1w { z4.s }, p1, [x15]\n" "add x15, x15, x13, LSL #2\n" - ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" + ".inst 0xc13913f2 // bfdot za.s[x8, 2], { z31.h-z2.h }, z9.h\n" "add x8, x8, #0x1\n" - "st1w { z25.s }, p1, [x14]\n" + "st1w { z5.s }, p1, [x14]\n" "add x14, x14, x11, LSL #2\n" - "st1w { z26.s }, p1, [x10]\n" + "st1w { z6.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" - "st1w { z27.s }, p1, [x9]\n" + ".inst 0xc0040e04 // mova za.d[x8, #4], { z16.d-z19.d }\n" + "st1w { z7.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" "21:" // Tail input: End "cbz x16, 23f\n" "22:" // Right padding loop - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" "subs x16, x16, #0x1\n" - ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" - "st1w { z24.s }, p1, [x15]\n" + ".inst 0xc1acc9a4 // fclamp { z4.s-z7.s }, z13.s, z12.s\n" + "st1w { z4.s }, p1, [x15]\n" "add x15, x15, x13, LSL #2\n" - ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" - "st1w { z25.s }, p1, [x14]\n" + ".inst 0xc0040e04 // mova za.d[x8, #4], { z16.d-z19.d }\n" + "st1w { z5.s }, p1, [x14]\n" "add x14, x14, x11, LSL #2\n" - "st1w { z26.s }, p1, [x10]\n" + "st1w { z6.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - "st1w { z27.s }, p1, [x9]\n" + "st1w { z7.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" "bgt 22b\n" "23:" // End @@ -1213,12 +1213,12 @@ void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( "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" + "ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n" "incw x6\n" "whilelt p1.s, x6, x5\n" - "ldr x17, [%x[args], %[offsetof_Args_inptr]]\n" - "add x17, x17, x20, LSL #2\n" - "str x17, [%x[args], %[offsetof_Args_inptr]]\n" + "ldr x20, [%x[args], %[offsetof_Args_inptr]]\n" + "add x20, x20, x21, LSL #2\n" + "str x20, [%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" |