diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp | 1132 |
1 files changed, 566 insertions, 566 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp index bf12b42ddc..3741b973b4 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp @@ -69,71 +69,71 @@ void sme2_fp32_planar_5x5_s1_4rows_mla_za_impl( Args args = { inptr, ld_in_vl, pad_top, 8u - std::min(8u, pad_top + valid_input_rows), pad_left, weights, bias, valid_input_cols, output_cols, outptrs, outlds, outvllds, start_channel, valid_channels, act_min, act_max }; __asm__ __volatile__( - "ldr x5, [%x[args], %[offsetof_Args_pad_bottom]]\n" + "ldr x6, [%x[args], %[offsetof_Args_pad_bottom]]\n" "mov x20, #0x8\n" ".inst 0xd503477f // SMSTART ZA\n" - "sub x20, x20, x5\n" - "ldr x6, [%x[args], %[offsetof_Args_pad_top]]\n" + "sub x20, x20, x6\n" + "ldr x7, [%x[args], %[offsetof_Args_pad_top]]\n" "ptrue p2.b\n" ".inst 0x25207812 // ptrue pn10.b\n" - "ld1rw { z22.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" - "ldr x7, [%x[args], %[offsetof_Args_n_channels]]\n" - "whilelt p1.s, XZR, x7\n" + "ld1rw { z16.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" "whilelt p9.s, XZR, x20\n" - "ld1rw { z11.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" - "whilelt p8.s, XZR, x6\n" + "ld1rw { z17.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" + "whilelt p8.s, XZR, x7\n" "eor p8.b, p2/Z, p8.b, p9.b\n" - "ldr x17, [%x[args], %[offsetof_Args_current_channel]]\n" + "ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n" "1:" // Channel loop "ldr x20, [%x[args], %[offsetof_Args_bias]]\n" "fmov z28.s, #0x0\n" "cbz x20, 2f\n" - "ld1w { z28.s }, p1/Z, [x20, x17, LSL #2]\n" + "ld1w { z28.s }, p1/Z, [x20, x16, LSL #2]\n" "2:" // Load bias: Done - "ldr x16, [%x[args], %[offsetof_Args_input_cols]]\n" - "sub x20, x16, #0x1\n" + "ldr x15, [%x[args], %[offsetof_Args_input_cols]]\n" + "sub x20, x15, #0x1\n" "orr x23, x20, %x[ld_in_col], LSL #18\n" "mov z29.d, z28.d\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "orr x23, x7, x23, LSL #20\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "orr x23, x17, x23, LSL #20\n" "mov x22, #0x8\n" - "ldr x14, [%x[args], %[offsetof_Args_inptr]]\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "add x21, x6, x5\n" + "ldr x13, [%x[args], %[offsetof_Args_inptr]]\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "add x21, x7, x6\n" "lsl x20, %x[ld_in_row], #0x2\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - "addvl x15, x15, #5\n" + "ld1w { z2.s }, p2/Z, [x14, #4, MUL VL]\n" + "addvl x14, x14, #5\n" "mov z30.d, z28.d\n" "mov z31.d, z28.d\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" "mov x8, #0x0\n" - "ldr x13, [%x[args], %[offsetof_Args_output_cols]]\n" + "ldr x11, [%x[args], %[offsetof_Args_output_cols]]\n" "lsl x23, x23, #0x2\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" "sub x22, x22, x21\n" - "madd x20, x20, x6, x14\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - "addvl x15, x15, #5\n" + "madd x20, x20, x7, x13\n" + "ld1w { z3.s }, p2/Z, [x14, #4, MUL VL]\n" + "addvl x14, x14, #5\n" "3:" // Issue prefetches "subs x22, x22, #0x1\n" ".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n" "add x20, x20, %x[ld_in_col], LSL #2\n" "bgt 3b\n" - "ldr x11, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x22, [%x[args], %[offsetof_Args_outptrs]]\n" "lsl x20, %x[ld_in_row], #0x2\n" - "msub x14, x6, x20, x14\n" + "msub x13, x7, x20, x13\n" ".inst 0xc0040f80 // mova za.d[x8, #0], { z28.d-z31.d }\n" "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" ".inst 0xc0040f81 // mova za.d[x8, #1], { z28.d-z31.d }\n" "mov x10, #0x4\n" - "ldp x9, x28, [x11], #0x10\n" + "ldp x9, x28, [x22], #0x10\n" ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" "ldp x27, x26, [x20], #0x10\n" ".inst 0xc0040f83 // mova za.d[x8, #3], { z28.d-z31.d }\n" "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" - "ldp x25, x24, [x11], #0x10\n" + "ldp x25, x24, [x22], #0x10\n" "ldp x23, x22, [x20], #0x10\n" "cbz x21, 5f\n" "cmp x21, x10\n" @@ -141,308 +141,308 @@ void sme2_fp32_planar_5x5_s1_4rows_mla_za_impl( "sub x21, x21, x20\n" "sub x10, x10, x20\n" "cbz x21, 5f\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" - "sub x13, x13, x21\n" - ".inst 0xc1abcad8 // fclamp { z24.s-z27.s }, z22.s, z11.s\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + "sub x11, x11, x21\n" + ".inst 0xc1b1ca04 // fclamp { z4.s-z7.s }, z16.s, z17.s\n" "4:" // Left padding "subs x21, x21, #0x1\n" - "st1w { z24.s }, p1, [x9]\n" + "st1w { z4.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - "st1w { z25.s }, p1, [x28]\n" + "st1w { z5.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - "st1w { z26.s }, p1, [x25]\n" + "st1w { z6.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - "st1w { z27.s }, p1, [x24]\n" + "st1w { z7.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" "bgt 4b\n" "5:" // Left padding: End - "adds XZR, x6, x5\n" + "adds XZR, x7, x6\n" "bne 12f\n" "cbz x10, 10f\n" "cmp x10, #0x1\n" - "sub x16, x16, x10\n" + "sub x15, x15, x10\n" "beq 9f\n" "cmp x10, #0x2\n" "beq 8f\n" "cmp x10, #0x3\n" "beq 7f\n" "6:" // Unpadded: 4 priming loads - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13419c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z4.s\n" - "ld1w { z18.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13019e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z0.s\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z18.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" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1341a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z4.s\n" "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1301a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z0.s\n" "ld1w { z21.s }, p1/Z, [x20]\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1341a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - "7:" // Unpadded: 3 priming loads - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13519c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z5.s\n" - ".inst 0xc13419c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z4.s\n" - "ld1w { z18.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13819e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z8.s\n" - "ld1w { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13019e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z0.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1351a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z5.s\n" - "ld1w { z20.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1341a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z4.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1381a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z8.s\n" + ".inst 0xc13e1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z14.s\n" + "ld1w { z22.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13a1a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z10.s\n" + "ld1w { z23.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xa14049c5 // ld1w { z5.s, z13.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1351a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z5.s\n" + "ld1w { z24.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xa04049cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13c1aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z12.s\n" + "ld1w { z25.s }, p1/Z, [x20]\n" + ".inst 0xa14049c1 // ld1w { z1.s, z9.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc1311ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z1.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + "7:" // Unpadded: 3 priming loads + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z20.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" "ld1w { z21.s }, p1/Z, [x20]\n" - ".inst 0xc1301a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z0.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1351a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z5.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1341a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" - "8:" // Unpadded: 2 priming loads - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13219c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z2.s\n" - ".inst 0xc13519c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z5.s\n" - "ld1w { z18.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13419c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z4.s\n" - "ld1w { z19.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13619e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z6.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - ".inst 0xc13819e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z8.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13019e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z0.s\n" - "ld1w { z20.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1321a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z2.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xc1351a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z5.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1341a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z4.s\n" - "ld1w { z21.s }, p1/Z, [x20]\n" - ".inst 0xc1361a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z6.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - ".inst 0xc1381a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z8.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1301a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z0.s\n" - ".inst 0xc1321a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z5.s\n" - ".inst 0xc1341a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" + "ld1w { z22.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z23.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13f1a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z15.s\n" + ".inst 0xc13e1a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z14.s\n" + "ld1w { z24.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13b1aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z11.s\n" + "ld1w { z25.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13a1aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z10.s\n" + ".inst 0xa04049c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1371ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z7.s\n" + "ld1w { z26.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc1361ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z6.s\n" + ".inst 0xa14049c5 // ld1w { z5.s, z13.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13d1ae0 // fmla za.s[x8, 0], { z23.s-z26.s }, z13.s\n" + "ld1w { z27.s }, p1/Z, [x20]\n" + ".inst 0xc1351ae1 // fmla za.s[x8, 1], { z23.s-z26.s }, z5.s\n" + ".inst 0xa04049c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc1371b00 // fmla za.s[x8, 0], { z24.s-z27.s }, z7.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1361b01 // fmla za.s[x8, 1], { z24.s-z27.s }, z6.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" + "8:" // Unpadded: 2 priming loads + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z1.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z2.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z3.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z4.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13c1820 // fmla za.s[x8, 0], { z1.s-z4.s }, z12.s\n" + ".inst 0xc13f1821 // fmla za.s[x8, 1], { z1.s-z4.s }, z15.s\n" + "ld1w { z5.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13e1822 // fmla za.s[x8, 2], { z1.s-z4.s }, z14.s\n" + "ld1w { z6.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc1381840 // fmla za.s[x8, 0], { z2.s-z5.s }, z8.s\n" + ".inst 0xa04049cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14]\n" + ".inst 0xc13b1841 // fmla za.s[x8, 1], { z2.s-z5.s }, z11.s\n" + ".inst 0xa04149ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13a1842 // fmla za.s[x8, 2], { z2.s-z5.s }, z10.s\n" + "ld1w { z7.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13e1860 // fmla za.s[x8, 0], { z3.s-z6.s }, z14.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + ".inst 0xc13d1861 // fmla za.s[x8, 1], { z3.s-z6.s }, z13.s\n" + ".inst 0xa14149c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13c1862 // fmla za.s[x8, 2], { z3.s-z6.s }, z12.s\n" + "ld1w { z8.s }, p1/Z, [x20]\n" + ".inst 0xc1301880 // fmla za.s[x8, 0], { z4.s-z7.s }, z0.s\n" + ".inst 0xa04049c0 // ld1w { z0.s-z1.s }, pn10.b/Z, [x14]\n" + ".inst 0xc13f1881 // fmla za.s[x8, 1], { z4.s-z7.s }, z15.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc13e1882 // fmla za.s[x8, 2], { z4.s-z7.s }, z14.s\n" + ".inst 0xc13c18a0 // fmla za.s[x8, 0], { z5.s-z8.s }, z12.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13118a1 // fmla za.s[x8, 1], { z5.s-z8.s }, z1.s\n" + ".inst 0xc13018a2 // fmla za.s[x8, 2], { z5.s-z8.s }, z0.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" "9:" // Unpadded: 1 priming loads - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "ld1w { z15.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ld1w { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13319c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z3.s\n" - ".inst 0xc13219c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z2.s\n" - "ld1w { z18.s }, p1/Z, [x20]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z20.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z21.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13519c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z5.s\n" + "ld1w { z22.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "ld1w { z23.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13d1a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z13.s\n" + ".inst 0xc13c1a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z12.s\n" + "ld1w { z24.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13f1a82 // fmla za.s[x8, 2], { z20.s-z23.s }, z15.s\n" + "ld1w { z25.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13e1a83 // fmla za.s[x8, 3], { z20.s-z23.s }, z14.s\n" + ".inst 0xa04049c4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x14]\n" + ".inst 0xc1391aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z9.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1381aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z8.s\n" + "ld1w { z26.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13b1aa2 // fmla za.s[x8, 2], { z21.s-z24.s }, z11.s\n" + ".inst 0xa14149c6 // ld1w { z6.s, z14.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13a1aa3 // fmla za.s[x8, 3], { z21.s-z24.s }, z10.s\n" + ".inst 0xa14049c1 // ld1w { z1.s, z9.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13d1ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z13.s\n" + "ld1w { z27.s }, p1/Z, [x20]\n" + ".inst 0xc13c1ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z12.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1351ac2 // fmla za.s[x8, 2], { z22.s-z25.s }, z5.s\n" + ".inst 0xc1341ac3 // fmla za.s[x8, 3], { z22.s-z25.s }, z4.s\n" + ".inst 0xa04049c4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc13e1ae0 // fmla za.s[x8, 0], { z23.s-z26.s }, z14.s\n" + "ld1w { z2.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1361ae1 // fmla za.s[x8, 1], { z23.s-z26.s }, z6.s\n" + ".inst 0xc1391ae2 // fmla za.s[x8, 2], { z23.s-z26.s }, z9.s\n" + ".inst 0xc1311ae3 // fmla za.s[x8, 3], { z23.s-z26.s }, z1.s\n" + ".inst 0xc13d1b00 // fmla za.s[x8, 0], { z24.s-z27.s }, z13.s\n" + ".inst 0xc13c1b01 // fmla za.s[x8, 1], { z24.s-z27.s }, z12.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1351b02 // fmla za.s[x8, 2], { z24.s-z27.s }, z5.s\n" + ".inst 0xc1341b03 // fmla za.s[x8, 3], { z24.s-z27.s }, z4.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x14, #4, MUL VL]\n" + "addvl x14, x14, #5\n" + "10:" // Unpadded: 0 priming loads + "cbz x15, 20f\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + "ld1w { z18.s }, p1/Z, [x13]\n" + "sub x15, x15, #0x1\n" "ld1w { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13419c3 // fmla za.s[x8, 3], { z14.s-z17.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - ".inst 0xc13719e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z7.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13619e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z6.s\n" + "sub x11, x11, #0x1\n" "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13819e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z8.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13019e3 // fmla za.s[x8, 3], { z15.s-z18.s }, z0.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1331a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z3.s\n" + "cmp x15, x11\n" "ld1w { z21.s }, p1/Z, [x20]\n" - ".inst 0xc1321a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z5.s\n" - ".inst 0xc1341a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1371a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z7.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1361a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z6.s\n" - ".inst 0xc1381a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z8.s\n" - ".inst 0xc1301a23 // fmla za.s[x8, 3], { z17.s-z20.s }, z0.s\n" - ".inst 0xc1331a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z3.s\n" - ".inst 0xc1321a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z5.s\n" - ".inst 0xc1341a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - "addvl x15, x15, #5\n" - "10:" // Unpadded: 0 priming loads - "cbz x16, 20f\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" - "ld1w { z14.s }, p1/Z, [x14]\n" - "sub x16, x16, #0x1\n" - "ld1w { z15.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "sub x13, x13, #0x1\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "csel x21, x15, x11, LT\n" + "ld1w { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "cmp x16, x13\n" - "ld1w { z17.s }, p1/Z, [x20]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "ld1w { z23.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "csel x21, x16, x13, LT\n" - "ld1w { z18.s }, p1/Z, [x20]\n" + "sub x11, x11, x21\n" + "ld1w { z24.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" + "ld1w { z25.s }, p1/Z, [x20]\n" + "cbz x21, 19f\n" + "11:" // Unpadded: Main loop + ".inst 0xc1321a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z2.s\n" + "ld1w { z6.s }, p2/Z, [x14, #4, MUL VL]\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" + "subs x21, x21, #0x1\n" + ".inst 0xc1331a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z3.s\n" + ".inst 0xc13d1a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z13.s\n" + ".inst 0xc13c1a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z12.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13f1a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z15.s\n" + ".inst 0xc13e1a44 // fmla za.s[x8, 4], { z18.s-z21.s }, z14.s\n" + ".inst 0xa04049c4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1361a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z6.s\n" + "ld1w { z6.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1391a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z9.s\n" + "ld1w { z18.s }, p1/Z, [x13]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc1381a62 // fmla za.s[x8, 2], { z19.s-z22.s }, z8.s\n" + ".inst 0xa04149ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13b1a63 // fmla za.s[x8, 3], { z19.s-z22.s }, z11.s\n" + ".inst 0xc13a1a64 // fmla za.s[x8, 4], { z19.s-z22.s }, z10.s\n" + ".inst 0xa14049c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1361aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z6.s\n" + "ld1w { z2.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc13d1a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z13.s\n" "ld1w { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - "sub x13, x13, x21\n" + ".inst 0xc13c1a82 // fmla za.s[x8, 2], { z20.s-z23.s }, z12.s\n" + ".inst 0xa04149c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1351a83 // fmla za.s[x8, 3], { z20.s-z23.s }, z5.s\n" + ".inst 0xc1341a84 // fmla za.s[x8, 4], { z20.s-z23.s }, z4.s\n" + ".inst 0xa04049c4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc1321ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z2.s\n" "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc13f1aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z15.s\n" + "ld1w { z2.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc13e1aa2 // fmla za.s[x8, 2], { z21.s-z24.s }, z14.s\n" + ".inst 0xc1381aa3 // fmla za.s[x8, 3], { z21.s-z24.s }, z8.s\n" + ".inst 0xc1301aa4 // fmla za.s[x8, 4], { z21.s-z24.s }, z0.s\n" "ld1w { z21.s }, p1/Z, [x20]\n" - "cbz x21, 19f\n" - "11:" // Unpadded: Main loop - ".inst 0xc13a19c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z10.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" - "subs x21, x21, #0x1\n" - ".inst 0xc13119e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z1.s\n" - ".inst 0xc13319c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z3.s\n" - ".inst 0xc13219c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13519c3 // fmla za.s[x8, 3], { z14.s-z17.s }, z5.s\n" - ".inst 0xc13419c4 // fmla za.s[x8, 4], { z14.s-z17.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13a1a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z10.s\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc13719e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z7.s\n" - "ld1w { z14.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc13619e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z6.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13819e3 // fmla za.s[x8, 3], { z15.s-z18.s }, z8.s\n" - ".inst 0xc13019e4 // fmla za.s[x8, 4], { z15.s-z18.s }, z0.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1311a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z1.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1331a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z3.s\n" - "ld1w { z15.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1321a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z5.s\n" - ".inst 0xc1341a04 // fmla za.s[x8, 4], { z16.s-z19.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc13a1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z10.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1371a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z7.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1361a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z6.s\n" - ".inst 0xc1381a23 // fmla za.s[x8, 3], { z17.s-z20.s }, z8.s\n" - ".inst 0xc1301a24 // fmla za.s[x8, 4], { z17.s-z20.s }, z0.s\n" - "ld1w { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" - ".inst 0xc1abcad8 // fclamp { z24.s-z27.s }, z22.s, z11.s\n" - "st1w { z24.s }, p1, [x9]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc0060c0c // mova { z12.d-z15.d }, za.d[x8, #0]\n" + ".inst 0xc1b1ca0c // fclamp { z12.s-z15.s }, z16.s, z17.s\n" + "st1w { z12.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - ".inst 0xc1331a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z3.s\n" - "st1w { z25.s }, p1, [x28]\n" + ".inst 0xc1371ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z7.s\n" + "st1w { z13.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc1321a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "st1w { z26.s }, p1, [x25]\n" + ".inst 0xc1361ac2 // fmla za.s[x8, 2], { z22.s-z25.s }, z6.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "st1w { z14.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc1351a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z5.s\n" - "st1w { z27.s }, p1, [x24]\n" + ".inst 0xc1351ac3 // fmla za.s[x8, 3], { z22.s-z25.s }, z5.s\n" + "st1w { z15.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" - ".inst 0xc1341a44 // fmla za.s[x8, 4], { z18.s-z21.s }, z4.s\n" - "ld1w { z18.s }, p1/Z, [x20]\n" + ".inst 0xc1341ac4 // fmla za.s[x8, 4], { z22.s-z25.s }, z4.s\n" + "ld1w { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" "add x8, x8, #0x1\n" - "ld1w { z19.s }, p1/Z, [x20]\n" + "ld1w { z23.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - "ld1w { z20.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - "addvl x15, x15, #5\n" - "ld1w { z21.s }, p1/Z, [x20]\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + "ld1w { z24.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x14, #4, MUL VL]\n" + "addvl x14, x14, #5\n" + "ld1w { z25.s }, p1/Z, [x20]\n" "bgt 11b\n" "b 19f\n" "12:" // Padded "cbz x10, 17f\n" "cmp x10, #0x1\n" - "sub x16, x16, x10\n" + "sub x15, x15, x10\n" "beq 16f\n" "cmp x10, #0x2\n" "beq 15f\n" @@ -451,429 +451,429 @@ void sme2_fp32_planar_5x5_s1_4rows_mla_za_impl( "13:" // Padded: 4 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z19.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 { z15.s }, p0/Z, [x20]\n" + "ld1w { z20.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 { z16.s }, p0/Z, [x20]\n" + "ld1w { z21.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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z22.s }, p0/Z, [x20]\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13419c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" + ".inst 0xc13e1a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z14.s\n" + ".inst 0xa14049c1 // ld1w { z1.s, z9.s }, pn10.b/Z, [x14]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + "ld1w { z23.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13019e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z0.s\n" + ".inst 0xc13a1a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z10.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z19.s }, p0/Z, [x20]\n" + "ld1w { z24.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1341a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z4.s\n" + ".inst 0xc1311aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z1.s\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "addvl x15, x15, #5\n" - "ld1w { z20.s }, p0/Z, [x20]\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" + "addvl x14, x14, #5\n" + "ld1w { z25.s }, p0/Z, [x20]\n" + ".inst 0xa04049c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14]\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 0xc1301a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z0.s\n" - "addvl x15, x15, #5\n" - "ld1w { z21.s }, p0/Z, [x20]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1341a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" + ".inst 0xc1361ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z6.s\n" + "addvl x14, x14, #5\n" + "ld1w { z26.s }, p0/Z, [x20]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xa14049c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc1301ae0 // fmla za.s[x8, 0], { z23.s-z26.s }, z0.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" "14:" // Padded: 3 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z0.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 { z15.s }, p0/Z, [x20]\n" + "ld1w { z1.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 { z16.s }, p0/Z, [x20]\n" + "ld1w { z2.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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z3.s }, p0/Z, [x20]\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13519c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z5.s\n" + ".inst 0xc13f1800 // fmla za.s[x8, 0], { z0.s-z3.s }, z15.s\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc13419c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z4.s\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + ".inst 0xc13e1801 // fmla za.s[x8, 1], { z0.s-z3.s }, z14.s\n" + "ld1w { z4.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 0xc13819e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z8.s\n" - "ld1w { z19.s }, p0/Z, [x20]\n" + ".inst 0xc13b1820 // fmla za.s[x8, 0], { z1.s-z4.s }, z11.s\n" + "ld1w { z5.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13019e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z0.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" + ".inst 0xc13a1821 // fmla za.s[x8, 1], { z1.s-z4.s }, z10.s\n" + ".inst 0xa04049c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1351a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z5.s\n" - "ld1w { z20.s }, p0/Z, [x20]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1391840 // fmla za.s[x8, 0], { z2.s-z5.s }, z9.s\n" + "ld1w { z6.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 0xc1341a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z4.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc1381a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z8.s\n" - "ld1w { z21.s }, p0/Z, [x20]\n" - ".inst 0xc1301a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z0.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1351a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z5.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1341a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" + ".inst 0xc1381841 // fmla za.s[x8, 1], { z2.s-z5.s }, z8.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc13f1860 // fmla za.s[x8, 0], { z3.s-z6.s }, z15.s\n" + "ld1w { z7.s }, p0/Z, [x20]\n" + ".inst 0xc13e1861 // fmla za.s[x8, 1], { z3.s-z6.s }, z14.s\n" + ".inst 0xa14049c3 // ld1w { z3.s, z11.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc13b1880 // fmla za.s[x8, 0], { z4.s-z7.s }, z11.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1331881 // fmla za.s[x8, 1], { z4.s-z7.s }, z3.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" "15:" // Padded: 2 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + "ld1w { z19.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 { z15.s }, p0/Z, [x20]\n" + "ld1w { z20.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 { z16.s }, p0/Z, [x20]\n" + "ld1w { z21.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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z22.s }, p0/Z, [x20]\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13219c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" + ".inst 0xc13c1a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z12.s\n" + ".inst 0xa04149c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc13519c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z5.s\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + ".inst 0xc13f1a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z15.s\n" + "ld1w { z23.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13419c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z4.s\n" + ".inst 0xc13e1a62 // fmla za.s[x8, 2], { z19.s-z22.s }, z14.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z19.s }, p0/Z, [x20]\n" + "ld1w { z24.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13619e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z6.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" + ".inst 0xc1381a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z8.s\n" + ".inst 0xa14049c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13819e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z8.s\n" - "ld1w { z20.s }, p0/Z, [x20]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13b1a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z11.s\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" - ".inst 0xc13019e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z0.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc1321a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z2.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1351a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z5.s\n" - "ld1w { z21.s }, p0/Z, [x20]\n" - ".inst 0xc1341a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - ".inst 0xc1361a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z6.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1381a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z8.s\n" - ".inst 0xc1301a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z0.s\n" - ".inst 0xc1321a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z5.s\n" - ".inst 0xc1341a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" + ".inst 0xc13a1a82 // fmla za.s[x8, 2], { z20.s-z23.s }, z10.s\n" + ".inst 0xa14049c2 // ld1w { z2.s, z10.s }, pn10.b/Z, [x14]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc1361aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z6.s\n" + ".inst 0xa14149c4 // ld1w { z4.s, z12.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1381aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z8.s\n" + "ld1w { z26.s }, p0/Z, [x20]\n" + ".inst 0xc1301aa2 // fmla za.s[x8, 2], { z21.s-z24.s }, z0.s\n" + ".inst 0xa04049c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14]\n" + ".inst 0xc1341ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z4.s\n" + ".inst 0xa14149c3 // ld1w { z3.s, z11.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc13a1ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z10.s\n" + ".inst 0xc1321ac2 // fmla za.s[x8, 2], { z22.s-z25.s }, z2.s\n" + ".inst 0xc1331ae0 // fmla za.s[x8, 0], { z23.s-z26.s }, z3.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1371ae1 // fmla za.s[x8, 1], { z23.s-z26.s }, z7.s\n" + ".inst 0xc1361ae2 // fmla za.s[x8, 2], { z23.s-z26.s }, z6.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" "16:" // Padded: 1 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\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" - "ld1w { z15.s }, p0/Z, [x20]\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" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z20.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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z21.s }, p0/Z, [x20]\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13319c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z3.s\n" + ".inst 0xc13d1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z13.s\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc13219c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z2.s\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + ".inst 0xc13c1a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z12.s\n" + "ld1w { z22.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13519c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z5.s\n" + ".inst 0xc13f1a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z15.s\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z19.s }, p0/Z, [x20]\n" + "ld1w { z23.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc13419c3 // fmla za.s[x8, 3], { z14.s-z17.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" + ".inst 0xc13e1a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z14.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc13719e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z7.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13619e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z6.s\n" - "ld1w { z20.s }, p0/Z, [x20]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc1391a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z9.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1381a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z8.s\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" - ".inst 0xc13819e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z8.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13019e3 // fmla za.s[x8, 3], { z15.s-z18.s }, z0.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1331a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z3.s\n" - "ld1w { z21.s }, p0/Z, [x20]\n" - ".inst 0xc1321a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z5.s\n" - ".inst 0xc1341a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1371a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z7.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1361a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z6.s\n" - ".inst 0xc1381a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z8.s\n" - ".inst 0xc1301a23 // fmla za.s[x8, 3], { z17.s-z20.s }, z0.s\n" - ".inst 0xc1331a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z3.s\n" - ".inst 0xc1321a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z5.s\n" - ".inst 0xc1341a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - "addvl x15, x15, #5\n" + ".inst 0xc13b1a62 // fmla za.s[x8, 2], { z19.s-z22.s }, z11.s\n" + ".inst 0xa14149c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13a1a63 // fmla za.s[x8, 3], { z19.s-z22.s }, z10.s\n" + ".inst 0xa14049c1 // ld1w { z1.s, z9.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13d1a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z13.s\n" + "ld1w { z25.s }, p0/Z, [x20]\n" + ".inst 0xc13c1a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z12.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13f1a82 // fmla za.s[x8, 2], { z20.s-z23.s }, z15.s\n" + ".inst 0xc13e1a83 // fmla za.s[x8, 3], { z20.s-z23.s }, z14.s\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc1381aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z8.s\n" + "ld1w { z2.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1301aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z0.s\n" + ".inst 0xc1391aa2 // fmla za.s[x8, 2], { z21.s-z24.s }, z9.s\n" + ".inst 0xc1311aa3 // fmla za.s[x8, 3], { z21.s-z24.s }, z1.s\n" + ".inst 0xc13d1ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z13.s\n" + ".inst 0xc13c1ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z12.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13b1ac2 // fmla za.s[x8, 2], { z22.s-z25.s }, z11.s\n" + ".inst 0xc13a1ac3 // fmla za.s[x8, 3], { z22.s-z25.s }, z10.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x14, #4, MUL VL]\n" + "addvl x14, x14, #5\n" "17:" // Padded: 0 priming loads - "cbz x16, 20f\n" + "cbz x15, 20f\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z14.s }, p0/Z, [x14]\n" - "add x20, x14, %x[ld_in_row], LSL #2\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" - "ld1w { z15.s }, p0/Z, [x20]\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" - "ld1w { z16.s }, p0/Z, [x20]\n" + "ld1w { z20.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 { z17.s }, p0/Z, [x20]\n" + "ld1w { z21.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 { z18.s }, p0/Z, [x20]\n" + "ld1w { z22.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" - "sub x16, x16, #0x1\n" - "ld1w { z19.s }, p0/Z, [x20]\n" - "sub x13, x13, #0x1\n" + "sub x15, x15, #0x1\n" + "ld1w { z23.s }, p0/Z, [x20]\n" + "sub x11, x11, #0x1\n" "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "cmp x16, x13\n" - "ld1w { z20.s }, p0/Z, [x20]\n" + "cmp x15, x11\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" - "ld1w { z21.s }, p0/Z, [x20]\n" - "csel x21, x16, x13, LT\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - "sub x13, x13, x21\n" + "ld1w { z25.s }, p0/Z, [x20]\n" + "csel x21, x15, x11, LT\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "sub x11, x11, x21\n" "cbz x21, 19f\n" "18:" // Padded: Main loop - ".inst 0xc13a19c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z10.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" + ".inst 0xc1321a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z2.s\n" + "ld1w { z0.s }, p2/Z, [x14, #4, MUL VL]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc13119e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z1.s\n" - "add x20, x14, %x[ld_in_row], LSL #2\n" + ".inst 0xc1331a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z3.s\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" "subs x21, x21, #0x1\n" - ".inst 0xc13319c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z3.s\n" - ".inst 0xc13219c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13519c3 // fmla za.s[x8, 3], { z14.s-z17.s }, z5.s\n" - ".inst 0xc13419c4 // fmla za.s[x8, 4], { z14.s-z17.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13a1a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z10.s\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc13719e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z7.s\n" - "ld1w { z14.s }, p0/Z, [x14]\n" + ".inst 0xc13d1a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z13.s\n" + ".inst 0xc13c1a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z12.s\n" + ".inst 0xa04149c2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13f1a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z15.s\n" + ".inst 0xc13e1a44 // fmla za.s[x8, 4], { z18.s-z21.s }, z14.s\n" + ".inst 0xa04049c4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1301a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z0.s\n" + "ld1w { z12.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1391a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z9.s\n" + "ld1w { z18.s }, p0/Z, [x13]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "add x14, x14, %x[ld_in_col], LSL #2\n" - ".inst 0xc13619e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z6.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13819e3 // fmla za.s[x8, 3], { z15.s-z18.s }, z8.s\n" - ".inst 0xc13019e4 // fmla za.s[x8, 4], { z15.s-z18.s }, z0.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1311a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z1.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1331a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z3.s\n" - "ld1w { z15.s }, p0/Z, [x20]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc1381a62 // fmla za.s[x8, 2], { z19.s-z22.s }, z8.s\n" + ".inst 0xa14149c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13b1a63 // fmla za.s[x8, 3], { z19.s-z22.s }, z11.s\n" + ".inst 0xc13a1a64 // fmla za.s[x8, 4], { z19.s-z22.s }, z10.s\n" + ".inst 0xa04049c6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc13c1aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z12.s\n" + "ld1w { z12.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1331a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z3.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 0xc1321a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z5.s\n" - ".inst 0xc1341a04 // fmla za.s[x8, 4], { z16.s-z19.s }, z4.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0xc1321a82 // fmla za.s[x8, 2], { z20.s-z23.s }, z2.s\n" + ".inst 0xa14149c3 // ld1w { z3.s, z11.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1351a83 // fmla za.s[x8, 3], { z20.s-z23.s }, z5.s\n" + ".inst 0xc1341a84 // fmla za.s[x8, 4], { z20.s-z23.s }, z4.s\n" + "ld1w { z20.s }, p0/Z, [x20]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x4\n" - ".inst 0xc13a1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z10.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - ".inst 0xc1371a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z7.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1361a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z6.s\n" - ".inst 0xc1381a23 // fmla za.s[x8, 3], { z17.s-z20.s }, z8.s\n" - ".inst 0xc1301a24 // fmla za.s[x8, 4], { z17.s-z20.s }, z0.s\n" - "ld1w { z17.s }, p0/Z, [x20]\n" + ".inst 0xc13c1ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z12.s\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "ldr x14, [%x[args], %[offsetof_Args_weights]]\n" + ".inst 0xc1381aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z8.s\n" + "ld1w { z2.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1301aa2 // fmla za.s[x8, 2], { z21.s-z24.s }, z0.s\n" + ".inst 0xc1371aa3 // fmla za.s[x8, 3], { z21.s-z24.s }, z7.s\n" + ".inst 0xc1361aa4 // fmla za.s[x8, 4], { z21.s-z24.s }, z6.s\n" + "ld1w { z21.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 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" - ".inst 0xc1abcad8 // fclamp { z24.s-z27.s }, z22.s, z11.s\n" - "st1w { z24.s }, p1, [x9]\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + ".inst 0xc1b1ca04 // fclamp { z4.s-z7.s }, z16.s, z17.s\n" + "st1w { z4.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - ".inst 0xc1331a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z3.s\n" - "st1w { z25.s }, p1, [x28]\n" + ".inst 0xc13b1ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z11.s\n" + "st1w { z5.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc1321a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "st1w { z26.s }, p1, [x25]\n" + ".inst 0xc1331ac2 // fmla za.s[x8, 2], { z22.s-z25.s }, z3.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "st1w { z6.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc1351a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z5.s\n" - "st1w { z27.s }, p1, [x24]\n" + ".inst 0xc13f1ac3 // fmla za.s[x8, 3], { z22.s-z25.s }, z15.s\n" + "st1w { z7.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" - ".inst 0xc1341a44 // fmla za.s[x8, 4], { z18.s-z21.s }, z4.s\n" - "ld1w { z18.s }, p0/Z, [x20]\n" + ".inst 0xc13e1ac4 // fmla za.s[x8, 4], { z22.s-z25.s }, z14.s\n" + "ld1w { z22.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 { z19.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" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" + ".inst 0xa04049ce // ld1w { z14.s-z15.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" "add x8, x8, #0x1\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" - "ld1w { z20.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" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - "addvl x15, x15, #5\n" - "ld1w { z21.s }, p0/Z, [x20]\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + ".inst 0xa04149c8 // ld1w { z8.s-z9.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + "ld1w { z3.s }, p2/Z, [x14, #4, MUL VL]\n" + "addvl x14, x14, #5\n" + "ld1w { z25.s }, p0/Z, [x20]\n" "bgt 18b\n" "19:" // Main loop tail - ".inst 0xc13a19c0 // fmla za.s[x8, 0], { z14.s-z17.s }, z10.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc13119e0 // fmla za.s[x8, 0], { z15.s-z18.s }, z1.s\n" - ".inst 0xc13319c1 // fmla za.s[x8, 1], { z14.s-z17.s }, z3.s\n" - ".inst 0xc13219c2 // fmla za.s[x8, 2], { z14.s-z17.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13519c3 // fmla za.s[x8, 3], { z14.s-z17.s }, z5.s\n" - ".inst 0xc13419c4 // fmla za.s[x8, 4], { z14.s-z17.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc13a1a00 // fmla za.s[x8, 0], { z16.s-z19.s }, z10.s\n" - "ld1w { z1.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc13719e1 // fmla za.s[x8, 1], { z15.s-z18.s }, z7.s\n" - ".inst 0xc13619e2 // fmla za.s[x8, 2], { z15.s-z18.s }, z6.s\n" - ".inst 0xa04149e6 // ld1w { z6.s-z7.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc13819e3 // fmla za.s[x8, 3], { z15.s-z18.s }, z8.s\n" - ".inst 0xc13019e4 // fmla za.s[x8, 4], { z15.s-z18.s }, z0.s\n" - ".inst 0xa14049e0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x15]\n" - "addvl x15, x15, #5\n" - ".inst 0xc1311a20 // fmla za.s[x8, 0], { z17.s-z20.s }, z1.s\n" - "ld1w { z10.s }, p2/Z, [x15, #4, MUL VL]\n" - ".inst 0xc1331a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z3.s\n" - ".inst 0xc1321a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z2.s\n" - ".inst 0xa04149e2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x15, #0x2, MUL VL]\n" - ".inst 0xc1351a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z5.s\n" - ".inst 0xc1341a04 // fmla za.s[x8, 4], { z16.s-z19.s }, z4.s\n" - ".inst 0xa04049e4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x15]\n" - ".inst 0xc13a1a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z10.s\n" - ".inst 0xc1371a21 // fmla za.s[x8, 1], { z17.s-z20.s }, z7.s\n" - ".inst 0xc1361a22 // fmla za.s[x8, 2], { z17.s-z20.s }, z6.s\n" - ".inst 0xc1381a23 // fmla za.s[x8, 3], { z17.s-z20.s }, z8.s\n" - ".inst 0xc1301a24 // fmla za.s[x8, 4], { z17.s-z20.s }, z0.s\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" - ".inst 0xc1abcad8 // fclamp { z24.s-z27.s }, z22.s, z11.s\n" - "st1w { z24.s }, p1, [x9]\n" + ".inst 0xc1321a40 // fmla za.s[x8, 0], { z18.s-z21.s }, z2.s\n" + "ld1w { z6.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1331a60 // fmla za.s[x8, 0], { z19.s-z22.s }, z3.s\n" + ".inst 0xc13d1a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z13.s\n" + ".inst 0xc13c1a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z12.s\n" + ".inst 0xa04149c4 // ld1w { z4.s-z5.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13f1a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z15.s\n" + ".inst 0xc13e1a44 // fmla za.s[x8, 4], { z18.s-z21.s }, z14.s\n" + ".inst 0xa04049c2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1361a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z6.s\n" + "ld1w { z7.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1391a61 // fmla za.s[x8, 1], { z19.s-z22.s }, z9.s\n" + ".inst 0xc1381a62 // fmla za.s[x8, 2], { z19.s-z22.s }, z8.s\n" + ".inst 0xa14149c0 // ld1w { z0.s, z8.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc13b1a63 // fmla za.s[x8, 3], { z19.s-z22.s }, z11.s\n" + ".inst 0xc13a1a64 // fmla za.s[x8, 4], { z19.s-z22.s }, z10.s\n" + ".inst 0xa04049ca // ld1w { z10.s-z11.s }, pn10.b/Z, [x14]\n" + "addvl x14, x14, #5\n" + ".inst 0xc1371aa0 // fmla za.s[x8, 0], { z21.s-z24.s }, z7.s\n" + "ld1w { z1.s }, p2/Z, [x14, #4, MUL VL]\n" + ".inst 0xc1351a81 // fmla za.s[x8, 1], { z20.s-z23.s }, z5.s\n" + ".inst 0xc1341a82 // fmla za.s[x8, 2], { z20.s-z23.s }, z4.s\n" + ".inst 0xa04149cc // ld1w { z12.s-z13.s }, pn10.b/Z, [x14, #0x2, MUL VL]\n" + ".inst 0xc1331a83 // fmla za.s[x8, 3], { z20.s-z23.s }, z3.s\n" + ".inst 0xc1321a84 // fmla za.s[x8, 4], { z20.s-z23.s }, z2.s\n" + ".inst 0xa04049c2 // ld1w { z2.s-z3.s }, pn10.b/Z, [x14]\n" + ".inst 0xc1311ac0 // fmla za.s[x8, 0], { z22.s-z25.s }, z1.s\n" + ".inst 0xc1381aa1 // fmla za.s[x8, 1], { z21.s-z24.s }, z8.s\n" + ".inst 0xc1301aa2 // fmla za.s[x8, 2], { z21.s-z24.s }, z0.s\n" + ".inst 0xc13b1aa3 // fmla za.s[x8, 3], { z21.s-z24.s }, z11.s\n" + ".inst 0xc13a1aa4 // fmla za.s[x8, 4], { z21.s-z24.s }, z10.s\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + ".inst 0xc1b1ca04 // fclamp { z4.s-z7.s }, z16.s, z17.s\n" + "st1w { z4.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - ".inst 0xc1331a41 // fmla za.s[x8, 1], { z18.s-z21.s }, z3.s\n" - "st1w { z25.s }, p1, [x28]\n" + ".inst 0xc13d1ac1 // fmla za.s[x8, 1], { z22.s-z25.s }, z13.s\n" + "st1w { z5.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc1321a42 // fmla za.s[x8, 2], { z18.s-z21.s }, z2.s\n" - "st1w { z26.s }, p1, [x25]\n" + ".inst 0xc13c1ac2 // fmla za.s[x8, 2], { z22.s-z25.s }, z12.s\n" + "st1w { z6.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc1351a43 // fmla za.s[x8, 3], { z18.s-z21.s }, z5.s\n" - "st1w { z27.s }, p1, [x24]\n" + ".inst 0xc1331ac3 // fmla za.s[x8, 3], { z22.s-z25.s }, z3.s\n" + "st1w { z7.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" - ".inst 0xc1341a44 // fmla za.s[x8, 4], { z18.s-z21.s }, z4.s\n" + ".inst 0xc1321ac4 // fmla za.s[x8, 4], { z22.s-z25.s }, z2.s\n" "add x8, x8, #0x1\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" "20:" // Main loop skip tail - "cbz x13, 22f\n" + "cbz x11, 22f\n" "21:" // Right padding loop - ".inst 0xc0060c18 // mova { z24.d-z27.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 0xc1abcad8 // fclamp { z24.s-z27.s }, z22.s, z11.s\n" - "st1w { z24.s }, p1, [x9]\n" + "subs x11, x11, #0x1\n" + ".inst 0xc1b1ca00 // fclamp { z0.s-z3.s }, z16.s, z17.s\n" + "st1w { z0.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" - "st1w { z25.s }, p1, [x28]\n" + "st1w { z1.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - "st1w { z26.s }, p1, [x25]\n" + "st1w { z2.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - "st1w { z27.s }, p1, [x24]\n" + "st1w { z3.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" "bgt 21b\n" "22:" // End - "ldr x15, [%x[args], %[offsetof_Args_weights]]\n" - "incb x15, ALL, MUL #16\n" - "incb x15, ALL, MUL #9\n" - "str x15, [%x[args], %[offsetof_Args_weights]]\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_in_vl]]\n" - "incw x17\n" - "whilelt p1.s, x17, x7\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 #16\n" + "incb x20, ALL, MUL #9\n" + "str x20, [%x[args], %[offsetof_Args_weights]]\n" + "ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n" + "incw x16\n" + "whilelt p1.s, x16, x17\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", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "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" ); } |