diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp | 560 |
1 files changed, 280 insertions, 280 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp index 9f6b09ef88..26315101b4 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp @@ -69,69 +69,69 @@ void sme2_fp32_planar_3x3_s2_4rows_mla_za_impl( Args args = { inptr, ld_in_vl, pad_top, 9u - std::min(9u, 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 x6, [%x[args], %[offsetof_Args_pad_bottom]]\n" + "ldr x7, [%x[args], %[offsetof_Args_pad_bottom]]\n" "mov x20, #0x9\n" ".inst 0xd503477f // SMSTART ZA\n" - "sub x20, x20, x6\n" - "ldr x7, [%x[args], %[offsetof_Args_pad_top]]\n" + "sub x20, x20, x7\n" + "ldr x17, [%x[args], %[offsetof_Args_pad_top]]\n" "ptrue p2.b\n" ".inst 0x25207812 // ptrue pn10.b\n" - "ld1rw { z28.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" - "ldr x17, [%x[args], %[offsetof_Args_n_channels]]\n" - "whilelt p1.s, XZR, x17\n" + "ld1rw { z7.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" + "ldr x16, [%x[args], %[offsetof_Args_n_channels]]\n" + "whilelt p1.s, XZR, x16\n" "whilelt p9.s, XZR, x20\n" - "ld1rw { z19.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" - "whilelt p8.s, XZR, x7\n" + "ld1rw { z9.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" + "whilelt p8.s, XZR, x17\n" "eor p8.b, p2/Z, p8.b, p9.b\n" - "ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n" + "ldr x15, [%x[args], %[offsetof_Args_current_channel]]\n" "1:" // Channel loop "ldr x20, [%x[args], %[offsetof_Args_bias]]\n" - "fmov z24.s, #0x0\n" + "fmov z12.s, #0x0\n" "cbz x20, 2f\n" - "ld1w { z24.s }, p1/Z, [x20, x16, LSL #2]\n" + "ld1w { z12.s }, p1/Z, [x20, x15, LSL #2]\n" "2:" // Load bias: Done - "ldr x15, [%x[args], %[offsetof_Args_input_cols]]\n" - "sub x20, x15, #0x1\n" + "ldr x14, [%x[args], %[offsetof_Args_input_cols]]\n" + "sub x20, x14, #0x1\n" "orr x24, x20, %x[ld_in_col], LSL #18\n" - "mov z25.d, z24.d\n" + "mov z13.d, z12.d\n" "ldr x23, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xa0404ae2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x23]\n" - "orr x24, x17, x24, LSL #20\n" + ".inst 0xa1404ae2 // ld1w { z2.s, z10.s }, pn10.b/Z, [x23]\n" + "orr x24, x16, x24, LSL #20\n" "mov x22, #0x9\n" - "ldr x14, [%x[args], %[offsetof_Args_inptr]]\n" - "ld1w { z7.s }, p2/Z, [x23, #2, MUL VL]\n" + "ldr x13, [%x[args], %[offsetof_Args_inptr]]\n" + "ld1w { z8.s }, p2/Z, [x23, #2, MUL VL]\n" "addvl x23, x23, #3\n" - "add x21, x7, x6\n" - ".inst 0xa0404ae4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x23]\n" + "add x21, x17, x7\n" + ".inst 0xa0404ae0 // ld1w { z0.s-z1.s }, pn10.b/Z, [x23]\n" "lsl x20, %x[ld_in_row], #0x2\n" - "mov z26.d, z24.d\n" - "mov z27.d, z24.d\n" - "ld1w { z6.s }, p2/Z, [x23, #2, MUL VL]\n" + "mov z14.d, z12.d\n" + "mov z15.d, z12.d\n" + "ld1w { z5.s }, p2/Z, [x23, #2, MUL VL]\n" "addvl x23, x23, #3\n" "mov x8, #0x0\n" - "ldr x13, [%x[args], %[offsetof_Args_output_cols]]\n" - ".inst 0xa1404ae1 // ld1w { z1.s, z9.s }, pn10.b/Z, [x23]\n" + "ldr x11, [%x[args], %[offsetof_Args_output_cols]]\n" + ".inst 0xa1404ae3 // ld1w { z3.s, z11.s }, pn10.b/Z, [x23]\n" "lsl x24, x24, #0x2\n" "sub x22, x22, x21\n" - "ld1w { z8.s }, p2/Z, [x23, #2, MUL VL]\n" - "madd x20, x20, x7, x14\n" + "ld1w { z6.s }, p2/Z, [x23, #2, MUL VL]\n" + "madd x20, x20, x17, x13\n" "3:" // Issue prefetches "subs x22, x22, #0x1\n" ".inst 0xf8b84a9c // rprfm pldstrm, x24, [x20]\n" "add x20, x20, %x[ld_in_col], LSL #2\n" "bgt 3b\n" - "ldr x11, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x23, [%x[args], %[offsetof_Args_outptrs]]\n" "lsl x20, %x[ld_in_row], #0x2\n" - "msub x14, x7, x20, x14\n" - ".inst 0xc0040f00 // mova za.d[x8, #0], { z24.d-z27.d }\n" + "msub x13, x17, x20, x13\n" + ".inst 0xc0040d80 // mova za.d[x8, #0], { z12.d-z15.d }\n" "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" - ".inst 0xc0040f01 // mova za.d[x8, #1], { z24.d-z27.d }\n" + ".inst 0xc0040d81 // mova za.d[x8, #1], { z12.d-z15.d }\n" "mov x22, #0x2\n" - "ldp x10, x9, [x11], #0x10\n" - ".inst 0xc0040f02 // mova za.d[x8, #2], { z24.d-z27.d }\n" + "ldp x10, x9, [x23], #0x10\n" + ".inst 0xc0040d82 // mova za.d[x8, #2], { z12.d-z15.d }\n" "ldp x28, x27, [x20], #0x10\n" "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" - "ldp x26, x25, [x11], #0x10\n" + "ldp x26, x25, [x23], #0x10\n" "ldp x24, x23, [x20], #0x10\n" "cbz x21, 5f\n" "cmp x21, x22\n" @@ -142,9 +142,9 @@ void sme2_fp32_planar_3x3_s2_4rows_mla_za_impl( ".inst 0xc0060c14 // mova { z20.d-z23.d }, za.d[x8, #0]\n" "and x22, x21, #0x1\n" "add x21, x21, #0x1\n" - ".inst 0xc1b3cb94 // fclamp { z20.s-z23.s }, z28.s, z19.s\n" + ".inst 0xc1a9c8f4 // fclamp { z20.s-z23.s }, z7.s, z9.s\n" "lsr x21, x21, #0x1\n" - "sub x13, x13, x21\n" + "sub x11, x11, x21\n" "4:" // Left padding "subs x21, x21, #0x1\n" "st1w { z20.s }, p1, [x10]\n" @@ -157,490 +157,490 @@ void sme2_fp32_planar_3x3_s2_4rows_mla_za_impl( "add x25, x25, x23, LSL #2\n" "bgt 4b\n" "5:" // Left padding: End - "adds XZR, x7, x6\n" + "adds XZR, x17, x7\n" "bne 10f\n" "cbz x22, 8f\n" "cmp x22, #0x1\n" - "sub x15, x15, x22\n" + "sub x14, x14, x22\n" "beq 7f\n" "6:" // Unpadded: 2 priming loads - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z12.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z29.s }, p1/Z, [x20]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z19.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z24.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z13.s }, p1/Z, [x20]\n" + "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z30.s }, p1/Z, [x20]\n" + "ld1w { z25.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x20]\n" + "ld1w { z21.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z31.s }, p1/Z, [x20]\n" + "ld1w { z26.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" + "ld1w { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1321980 // fmla za.s[x8, 0], { z12.s-z15.s }, z2.s\n" - "ld1w { z0.s }, p1/Z, [x20]\n" + ".inst 0xc1321a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z2.s\n" + "ld1w { z27.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1341ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z4.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - ".inst 0xc13119a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z1.s\n" + ".inst 0xc1301b00 // fmla za.s[x8, 0], { z24.s-z27.s }, z0.s\n" + "ld1w { z23.s }, p1/Z, [x20]\n" + ".inst 0xc1331a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z3.s\n" "7:" // Unpadded: 1 priming loads - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z12.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z29.s }, p1/Z, [x20]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z26.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z13.s }, p1/Z, [x20]\n" + "ld1w { z27.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z30.s }, p1/Z, [x20]\n" + "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x20]\n" + "ld1w { z28.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z31.s }, p1/Z, [x20]\n" + "ld1w { z21.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" + "ld1w { z29.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1331980 // fmla za.s[x8, 0], { z12.s-z15.s }, z3.s\n" - "ld1w { z0.s }, p1/Z, [x20]\n" + ".inst 0xc13a1b40 // fmla za.s[x8, 0], { z26.s-z29.s }, z10.s\n" + "ld1w { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1351ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z5.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - ".inst 0xc13919a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z9.s\n" + ".inst 0xc1311a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z1.s\n" + "ld1w { z30.s }, p1/Z, [x20]\n" + ".inst 0xc13b1b60 // fmla za.s[x8, 0], { z27.s-z30.s }, z11.s\n" "8:" // Unpadded: 0 priming loads - "cmp x15, #0x2\n" + "cmp x14, #0x2\n" "blt 16f\n" - "add x21, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z12.s }, p1/Z, [x14]\n" - "sub x15, x15, #0x2\n" - "ld1w { z29.s }, p1/Z, [x21]\n" + "add x21, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z25.s }, p1/Z, [x13]\n" + "sub x14, x14, #0x2\n" + "ld1w { z19.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "sub x13, x13, #0x1\n" - "ld1w { z13.s }, p1/Z, [x21]\n" + "sub x11, x11, #0x1\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "lsr x20, x15, #0x1\n" - "ld1w { z30.s }, p1/Z, [x21]\n" + "lsr x20, x14, #0x1\n" + "ld1w { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "cmp x20, x13\n" - "ld1w { z14.s }, p1/Z, [x21]\n" + "cmp x20, x11\n" + "ld1w { z27.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "csel x22, x20, x13, LT\n" - "ld1w { z31.s }, p1/Z, [x21]\n" + "csel x22, x20, x11, LT\n" + "ld1w { z21.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x21]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z28.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "and x15, x15, #0x1\n" - "ld1w { z0.s }, p1/Z, [x21]\n" + "and x14, x14, #0x1\n" + "ld1w { z22.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "sub x13, x13, x22\n" - "ld1w { z16.s }, p1/Z, [x21]\n" + "sub x11, x11, x22\n" + "ld1w { z29.s }, p1/Z, [x21]\n" "cbz x22, 15f\n" "9:" // Unpadded: Main loop - ".inst 0xc1371980 // fmla za.s[x8, 0], { z12.s-z15.s }, z7.s\n" - "add x21, x14, %x[ld_in_row], LSL #2\n" + ".inst 0xc1381b20 // fmla za.s[x8, 0], { z25.s-z28.s }, z8.s\n" + "add x21, x13, %x[ld_in_row], LSL #2\n" "subs x22, x22, #0x1\n" - ".inst 0xc1321981 // fmla za.s[x8, 1], { z12.s-z15.s }, z2.s\n" - "ld1w { z12.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" - ".inst 0xc1361ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z6.s\n" - ".inst 0xc1341ba1 // fmla za.s[x8, 1], { z29.s-z0.s }, z4.s\n" - "ld1w { z29.s }, p1/Z, [x21]\n" + ".inst 0xc1321b21 // fmla za.s[x8, 1], { z25.s-z28.s }, z2.s\n" + "ld1w { z25.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + ".inst 0xc1351a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z5.s\n" + ".inst 0xc1301a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z0.s\n" + "ld1w { z18.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0xc13819a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z8.s\n" - ".inst 0xc13119a1 // fmla za.s[x8, 1], { z13.s-z16.s }, z1.s\n" - "ld1w { z13.s }, p1/Z, [x21]\n" + ".inst 0xc1361b40 // fmla za.s[x8, 0], { z26.s-z29.s }, z6.s\n" + ".inst 0xc1331b41 // fmla za.s[x8, 1], { z26.s-z29.s }, z3.s\n" + "ld1w { z26.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "ld1w { z30.s }, p1/Z, [x21]\n" + "ld1w { z19.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" ".inst 0xc0060c14 // mova { z20.d-z23.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" - "ld1w { z14.s }, p1/Z, [x21]\n" + "ld1w { z27.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0xc1b3cb94 // fclamp { z20.s-z23.s }, z28.s, z19.s\n" + ".inst 0xc1a9c8f4 // fclamp { z20.s-z23.s }, z7.s, z9.s\n" "st1w { z20.s }, p1, [x10]\n" - "ld1w { z31.s }, p1/Z, [x21]\n" + "ld1w { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" "add x10, x10, x28, LSL #2\n" "st1w { z21.s }, p1, [x9]\n" - "ld1w { z15.s }, p1/Z, [x21]\n" + "ld1w { z28.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0xc1331980 // fmla za.s[x8, 0], { z12.s-z15.s }, z3.s\n" + ".inst 0xc13a1b20 // fmla za.s[x8, 0], { z25.s-z28.s }, z10.s\n" "add x9, x9, x27, LSL #2\n" - "ld1w { z0.s }, p1/Z, [x21]\n" + "ld1w { z21.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0xc1351ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z5.s\n" + ".inst 0xc1311a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z1.s\n" "st1w { z22.s }, p1, [x26]\n" - "ld1w { z16.s }, p1/Z, [x21]\n" - ".inst 0xc13919a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z9.s\n" + "ld1w { z29.s }, p1/Z, [x21]\n" + ".inst 0xc13b1b40 // fmla za.s[x8, 0], { z26.s-z29.s }, z11.s\n" "add x26, x26, x24, LSL #2\n" "st1w { z23.s }, p1, [x25]\n" - "ld1w { z12.s }, p1/Z, [x14]\n" + "ld1w { z25.s }, p1/Z, [x13]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc0040f02 // mova za.d[x8, #2], { z24.d-z27.d }\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z29.s }, p1/Z, [x20]\n" + ".inst 0xc0040d82 // mova za.d[x8, #2], { z12.d-z15.d }\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z13.s }, p1/Z, [x20]\n" + "ld1w { z26.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z30.s }, p1/Z, [x20]\n" + "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x20]\n" + "ld1w { z27.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z31.s }, p1/Z, [x20]\n" + "ld1w { z21.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" + "ld1w { z28.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z0.s }, p1/Z, [x20]\n" + "ld1w { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z29.s }, p1/Z, [x20]\n" "bgt 9b\n" "b 15f\n" "10:" // Padded "cbz x22, 13f\n" "cmp x22, #0x1\n" - "sub x15, x15, x22\n" + "sub x14, x14, x22\n" "beq 12f\n" "11:" // Padded: 2 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z12.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z27.s }, p0/Z, [x13]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z29.s }, p0/Z, [x20]\n" + "ld1w { z23.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 { z13.s }, p0/Z, [x20]\n" + "ld1w { z28.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" - "ld1w { z30.s }, p0/Z, [x20]\n" + "ld1w { z24.s }, p0/Z, [x20]\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 { z14.s }, p0/Z, [x20]\n" + "ld1w { z29.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" - "ld1w { z31.s }, p0/Z, [x20]\n" + "ld1w { z25.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 { z15.s }, p0/Z, [x20]\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" "mov x12, #0x8\n" - ".inst 0xc1321980 // fmla za.s[x8, 0], { z12.s-z15.s }, z2.s\n" - "ld1w { z0.s }, p0/Z, [x20]\n" + ".inst 0xc1321b60 // fmla za.s[x8, 0], { z27.s-z30.s }, z2.s\n" + "ld1w { z26.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 0xc1341ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z4.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc13119a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z1.s\n" + ".inst 0xc1301ae0 // fmla za.s[x8, 0], { z23.s-z26.s }, z0.s\n" + "ld1w { z31.s }, p0/Z, [x20]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc1331b80 // fmla za.s[x8, 0], { z28.s-z31.s }, z3.s\n" "12:" // Padded: 1 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z12.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z22.s }, p0/Z, [x13]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z29.s }, p0/Z, [x20]\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" - "ld1w { z13.s }, p0/Z, [x20]\n" + "ld1w { z23.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" - "ld1w { z30.s }, p0/Z, [x20]\n" + "ld1w { z28.s }, p0/Z, [x20]\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 { z14.s }, p0/Z, [x20]\n" + "ld1w { z24.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" - "ld1w { z31.s }, p0/Z, [x20]\n" + "ld1w { z29.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 { z15.s }, p0/Z, [x20]\n" + "ld1w { z25.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" "mov x12, #0x8\n" - ".inst 0xc1331980 // fmla za.s[x8, 0], { z12.s-z15.s }, z3.s\n" - "ld1w { z0.s }, p0/Z, [x20]\n" + ".inst 0xc13a1ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z10.s\n" + "ld1w { z30.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 0xc1351ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z5.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc13919a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z9.s\n" + ".inst 0xc1311b60 // fmla za.s[x8, 0], { z27.s-z30.s }, z1.s\n" + "ld1w { z26.s }, p0/Z, [x20]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc13b1ae0 // fmla za.s[x8, 0], { z23.s-z26.s }, z11.s\n" "13:" // Padded: 0 priming loads - "cmp x15, #0x2\n" + "cmp x14, #0x2\n" "blt 16f\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z12.s }, p0/Z, [x14]\n" - "add x21, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z25.s }, p0/Z, [x13]\n" + "add x21, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z29.s }, p0/Z, [x21]\n" + "ld1w { z19.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" - "ld1w { z13.s }, p0/Z, [x21]\n" + "ld1w { z26.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" - "ld1w { z30.s }, p0/Z, [x21]\n" + "ld1w { z20.s }, p0/Z, [x21]\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 { z14.s }, p0/Z, [x21]\n" + "ld1w { z27.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" - "sub x15, x15, #0x2\n" - "ld1w { z31.s }, p0/Z, [x21]\n" + "sub x14, x14, #0x2\n" + "ld1w { z21.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" - "ld1w { z15.s }, p0/Z, [x21]\n" - "sub x13, x13, #0x1\n" - "lsr x20, x15, #0x1\n" + "ld1w { z28.s }, p0/Z, [x21]\n" + "sub x11, x11, #0x1\n" + "lsr x20, x14, #0x1\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 { z0.s }, p0/Z, [x21]\n" + "ld1w { z22.s }, p0/Z, [x21]\n" "mov x12, #0x8\n" - "cmp x20, x13\n" + "cmp x20, x11\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" - "csel x22, x20, x13, LT\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "and x15, x15, #0x1\n" - "sub x13, x13, x22\n" + "ld1w { z29.s }, p0/Z, [x21]\n" + "csel x22, x20, x11, LT\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "and x14, x14, #0x1\n" + "sub x11, x11, x22\n" "cbz x22, 15f\n" "14:" // Padded: Main loop - ".inst 0xc1371980 // fmla za.s[x8, 0], { z12.s-z15.s }, z7.s\n" + ".inst 0xc1381b20 // fmla za.s[x8, 0], { z25.s-z28.s }, z8.s\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1321981 // fmla za.s[x8, 1], { z12.s-z15.s }, z2.s\n" - "ld1w { z12.s }, p0/Z, [x14]\n" - "add x21, x14, %x[ld_in_row], LSL #2\n" + ".inst 0xc1321b21 // fmla za.s[x8, 1], { z25.s-z28.s }, z2.s\n" + "ld1w { z18.s }, p0/Z, [x13]\n" + "add x21, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1361ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z6.s\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" - ".inst 0xc1341ba1 // fmla za.s[x8, 1], { z29.s-z0.s }, z4.s\n" - "ld1w { z29.s }, p0/Z, [x21]\n" + ".inst 0xc1351a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z5.s\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + ".inst 0xc1301a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z0.s\n" + "ld1w { z25.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 0xc13819a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z8.s\n" + ".inst 0xc1361b40 // fmla za.s[x8, 0], { z26.s-z29.s }, z6.s\n" "subs x22, x22, #0x1\n" - ".inst 0xc13119a1 // fmla za.s[x8, 1], { z13.s-z16.s }, z1.s\n" - "ld1w { z13.s }, p0/Z, [x21]\n" + ".inst 0xc1331b41 // fmla za.s[x8, 1], { z26.s-z29.s }, z3.s\n" + "ld1w { z19.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" "mov x12, #0x4\n" - "ld1w { z30.s }, p0/Z, [x21]\n" + "ld1w { z26.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0xc0060c14 // mova { z20.d-z23.d }, za.d[x8, #0]\n" + ".inst 0xc0060c1c // mova { z28.d-z31.d }, za.d[x8, #0]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x21]\n" + "ld1w { z20.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - ".inst 0xc1b3cb94 // fclamp { z20.s-z23.s }, z28.s, z19.s\n" + ".inst 0xc1a9c8fc // fclamp { z28.s-z31.s }, z7.s, z9.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z31.s }, p0/Z, [x21]\n" + "ld1w { z27.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "st1w { z20.s }, p1, [x10]\n" + "st1w { z28.s }, p1, [x10]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z15.s }, p0/Z, [x21]\n" + "ld1w { z21.s }, p0/Z, [x21]\n" "add x8, x8, #0x1\n" - "st1w { z21.s }, p1, [x9]\n" + "st1w { z29.s }, p1, [x9]\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 { z0.s }, p0/Z, [x21]\n" - "st1w { z22.s }, p1, [x26]\n" + "ld1w { z28.s }, p0/Z, [x21]\n" + "st1w { z30.s }, p1, [x26]\n" "mov x12, #0x8\n" - ".inst 0xc1331980 // fmla za.s[x8, 0], { z12.s-z15.s }, z3.s\n" + ".inst 0xc13a1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z10.s\n" "add x21, x21, %x[ld_in_row], LSL #2\n" - "st1w { z23.s }, p1, [x25]\n" + "st1w { z31.s }, p1, [x25]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1351ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z5.s\n" + ".inst 0xc1311b20 // fmla za.s[x8, 0], { z25.s-z28.s }, z1.s\n" "mov x12, #0x0\n" - "ld1w { z16.s }, p0/Z, [x21]\n" + "ld1w { z22.s }, p0/Z, [x21]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z12.s }, p0/Z, [x14]\n" + "ld1w { z25.s }, p0/Z, [x13]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc13919a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z9.s\n" - "ld1w { z29.s }, p0/Z, [x20]\n" + ".inst 0xc13b1a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z11.s\n" + "ld1w { z19.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 0xc0040f02 // mova za.d[x8, #2], { z24.d-z27.d }\n" - "ld1w { z13.s }, p0/Z, [x20]\n" + ".inst 0xc0040d82 // mova za.d[x8, #2], { z12.d-z15.d }\n" + "ld1w { z26.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" "mov x12, #0x4\n" - "ld1w { z30.s }, p0/Z, [x20]\n" + "ld1w { z20.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" - "ld1w { z14.s }, p0/Z, [x20]\n" + "ld1w { z27.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" - "ld1w { z31.s }, p0/Z, [x20]\n" + "ld1w { z21.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 { z15.s }, p0/Z, [x20]\n" + "ld1w { z28.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" "mov x12, #0x8\n" - "ld1w { z0.s }, p0/Z, [x20]\n" + "ld1w { z22.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" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z29.s }, p0/Z, [x20]\n" "add x10, x10, x28, LSL #2\n" "add x9, x9, x27, LSL #2\n" "add x26, x26, x24, LSL #2\n" "add x25, x25, x23, LSL #2\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" "bgt 14b\n" "15:" // Main loop tail - ".inst 0xc1371980 // fmla za.s[x8, 0], { z12.s-z15.s }, z7.s\n" + ".inst 0xc1381b20 // fmla za.s[x8, 0], { z25.s-z28.s }, z8.s\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1321981 // fmla za.s[x8, 1], { z12.s-z15.s }, z2.s\n" - "ld1w { z12.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + ".inst 0xc1321b21 // fmla za.s[x8, 1], { z25.s-z28.s }, z2.s\n" + "ld1w { z18.s }, p0/Z, [x13]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1361ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z6.s\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc1341ba1 // fmla za.s[x8, 1], { z29.s-z0.s }, z4.s\n" - "ld1w { z29.s }, p0/Z, [x20]\n" + ".inst 0xc1351a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z5.s\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc1301a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z0.s\n" + "ld1w { z22.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 0xc13819a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z8.s\n" - ".inst 0xc13119a1 // fmla za.s[x8, 1], { z13.s-z16.s }, z1.s\n" - "ld1w { z13.s }, p0/Z, [x20]\n" + ".inst 0xc1361b40 // fmla za.s[x8, 0], { z26.s-z29.s }, z6.s\n" + ".inst 0xc1331b41 // fmla za.s[x8, 1], { z26.s-z29.s }, z3.s\n" + "ld1w { z19.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" "mov x12, #0x4\n" - "ld1w { z30.s }, p0/Z, [x20]\n" + "ld1w { z23.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc0060c14 // mova { z20.d-z23.d }, za.d[x8, #0]\n" + ".inst 0xc0060c1c // mova { z28.d-z31.d }, za.d[x8, #0]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x20]\n" + "ld1w { z20.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1b3cb94 // fclamp { z20.s-z23.s }, z28.s, z19.s\n" + ".inst 0xc1a9c8fc // fclamp { z28.s-z31.s }, z7.s, z9.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z31.s }, p0/Z, [x20]\n" + "ld1w { z24.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "st1w { z20.s }, p1, [x10]\n" + "st1w { z28.s }, p1, [x10]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z15.s }, p0/Z, [x20]\n" + "ld1w { z21.s }, p0/Z, [x20]\n" "add x8, x8, #0x1\n" - "st1w { z21.s }, p1, [x9]\n" + "st1w { z29.s }, p1, [x9]\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 { z0.s }, p0/Z, [x20]\n" - "st1w { z22.s }, p1, [x26]\n" + "ld1w { z25.s }, p0/Z, [x20]\n" + "st1w { z30.s }, p1, [x26]\n" "mov x12, #0x8\n" - ".inst 0xc1331980 // fmla za.s[x8, 0], { z12.s-z15.s }, z3.s\n" + ".inst 0xc13a1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z10.s\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "st1w { z23.s }, p1, [x25]\n" + "st1w { z31.s }, p1, [x25]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1351ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z5.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0xc1311ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z1.s\n" + "ld1w { z22.s }, p0/Z, [x20]\n" "add x10, x10, x28, LSL #2\n" "add x9, x9, x27, LSL #2\n" "add x26, x26, x24, LSL #2\n" - ".inst 0xc0040f02 // mova za.d[x8, #2], { z24.d-z27.d }\n" + ".inst 0xc0040d82 // mova za.d[x8, #2], { z12.d-z15.d }\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc13919a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z9.s\n" + ".inst 0xc13b1a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z11.s\n" "16:" // Main loop skip tail - "cbz x15, 17f\n" // Skip remainder inputs + "cbz x14, 17f\n" // Skip remainder inputs "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z12.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z21.s }, p0/Z, [x13]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z29.s }, p0/Z, [x20]\n" + "ld1w { z28.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 { z13.s }, p0/Z, [x20]\n" + "ld1w { z22.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" - "ld1w { z30.s }, p0/Z, [x20]\n" + "ld1w { z29.s }, p0/Z, [x20]\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 { z14.s }, p0/Z, [x20]\n" + "ld1w { z23.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" - "ld1w { z31.s }, p0/Z, [x20]\n" + "ld1w { z30.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 { z15.s }, p0/Z, [x20]\n" + "ld1w { z24.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" "mov x12, #0x8\n" - ".inst 0xc1371980 // fmla za.s[x8, 0], { z12.s-z15.s }, z7.s\n" - "ld1w { z0.s }, p0/Z, [x20]\n" + ".inst 0xc1381aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z8.s\n" + "ld1w { z31.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 0xc1361ba0 // fmla za.s[x8, 0], { z29.s-z0.s }, z6.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0xc1321981 // fmla za.s[x8, 1], { z12.s-z15.s }, z2.s\n" - "sub x13, x13, #0x1\n" - ".inst 0xc13819a0 // fmla za.s[x8, 0], { z13.s-z16.s }, z8.s\n" - ".inst 0xc1341ba1 // fmla za.s[x8, 1], { z29.s-z0.s }, z4.s\n" - ".inst 0xc0060c14 // mova { z20.d-z23.d }, za.d[x8, #0]\n" - ".inst 0xc1b3cb94 // fclamp { z20.s-z23.s }, z28.s, z19.s\n" - "st1w { z20.s }, p1, [x10]\n" + ".inst 0xc1351b80 // fmla za.s[x8, 0], { z28.s-z31.s }, z5.s\n" + "ld1w { z25.s }, p0/Z, [x20]\n" + ".inst 0xc1321aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z2.s\n" + "sub x11, x11, #0x1\n" + ".inst 0xc1361ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z6.s\n" + ".inst 0xc1301b81 // fmla za.s[x8, 1], { z28.s-z31.s }, z0.s\n" + ".inst 0xc0060c10 // mova { z16.d-z19.d }, za.d[x8, #0]\n" + ".inst 0xc1a9c8f0 // fclamp { z16.s-z19.s }, z7.s, z9.s\n" + "st1w { z16.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - ".inst 0xc13119a1 // fmla za.s[x8, 1], { z13.s-z16.s }, z1.s\n" + ".inst 0xc1331ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z3.s\n" "add x8, x8, #0x1\n" - "st1w { z21.s }, p1, [x9]\n" + "st1w { z17.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - "st1w { z22.s }, p1, [x26]\n" + "st1w { z18.s }, p1, [x26]\n" "add x26, x26, x24, LSL #2\n" - ".inst 0xc0040f02 // mova za.d[x8, #2], { z24.d-z27.d }\n" - "st1w { z23.s }, p1, [x25]\n" + ".inst 0xc0040d82 // mova za.d[x8, #2], { z12.d-z15.d }\n" + "st1w { z19.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" "17:" // Tail input: End - "cbz x13, 19f\n" + "cbz x11, 19f\n" "18:" // Right padding loop - ".inst 0xc0060c14 // mova { z20.d-z23.d }, za.d[x8, #0]\n" + ".inst 0xc0060c00 // mova { z0.d-z3.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" - "subs x13, x13, #0x1\n" - ".inst 0xc1b3cb94 // fclamp { z20.s-z23.s }, z28.s, z19.s\n" - "st1w { z20.s }, p1, [x10]\n" + "subs x11, x11, #0x1\n" + ".inst 0xc1a9c8e0 // fclamp { z0.s-z3.s }, z7.s, z9.s\n" + "st1w { z0.s }, p1, [x10]\n" "add x10, x10, x28, LSL #2\n" - ".inst 0xc0040f02 // mova za.d[x8, #2], { z24.d-z27.d }\n" - "st1w { z21.s }, p1, [x9]\n" + ".inst 0xc0040d82 // mova za.d[x8, #2], { z12.d-z15.d }\n" + "st1w { z1.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - "st1w { z22.s }, p1, [x26]\n" + "st1w { z2.s }, p1, [x26]\n" "add x26, x26, x24, LSL #2\n" - "st1w { z23.s }, p1, [x25]\n" + "st1w { z3.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" "bgt 18b\n" "19:" // End - "ldr x23, [%x[args], %[offsetof_Args_weights]]\n" - "incb x23, ALL, MUL #9\n" - "str x23, [%x[args], %[offsetof_Args_weights]]\n" - "incw x16\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_in_vl]]\n" - "whilelt p1.s, x16, x17\n" - "ldr x14, [%x[args], %[offsetof_Args_inptr]]\n" - "add x14, x14, x20, LSL #2\n" - "str x14, [%x[args], %[offsetof_Args_inptr]]\n" - "ldr x11, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" + "incb x20, ALL, MUL #9\n" + "str x20, [%x[args], %[offsetof_Args_weights]]\n" + "incw x15\n" + "ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n" + "whilelt p1.s, x15, x16\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, [x11, #0x0]\n" + "ldp x23, x22, [x25, #0x0]\n" "ldp x21, x20, [x24, #0x0]\n" "add x23, x23, x21, LSL #2\n" "add x22, x22, x20, LSL #2\n" - "stp x23, x22, [x11, #0x0]\n" - "ldp x23, x22, [x11, #0x10]\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 x22, x22, x20, LSL #2\n" - "stp x23, x22, [x11, #0x10]\n" + "stp x23, x22, [x25, #0x10]\n" "b.any 1b\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", "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", "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" ); } |