diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp | 1063 |
1 files changed, 528 insertions, 535 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp index a3cfa94b03..fdebaea83b 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp @@ -70,124 +70,119 @@ void sme2_s8q_planar_3x3_s2_4rows_dot_za_impl( __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" "ldr x6, [%x[args], %[offsetof_Args_pad_bottom]]\n" - "mov x22, SP\n" - "mov x21, #0x9\n" "ptrue p2.b\n" + "mov x20, #0x9\n" "ldr x7, [%x[args], %[offsetof_Args_pad_top]]\n" - "sub x20, x22, #0x8\n" - "ldr x17, [%x[args], %[offsetof_Args_n_channels]]\n" - "ld1rh { z29.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_a_offset]]\n" - "and x20, x20, #-0x400\n" + "ld1rh { z11.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_a_offset]]\n" + "sub x20, x20, x6\n" ".inst 0x25207812 // ptrue pn10.b\n" - "sub x21, x21, x6\n" - "mov SP, x20\n" - "ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n" - "ld1rw { z8.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n" - "str x22, [SP]\n" - "addvl SP, SP, #-6\n" - "ld1rw { z1.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_mul]]\n" + "ldr x17, [%x[args], %[offsetof_Args_n_channels]]\n" "whilelt p1.s, XZR, x17\n" - "whilelt p9.s, XZR, x21\n" - "neg z29.h, p2/M, z29.h\n" - "ld1rw { z0.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "whilelt p9.s, XZR, x20\n" + "ld1rw { z13.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n" "whilelt p8.s, XZR, x7\n" - "ld1rw { z31.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n" + "addvl SP, SP, #-6\n" + "ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n" + "neg z11.h, p2/M, z11.h\n" "eor p8.b, p2/Z, p8.b, p9.b\n" - "ld1rw { z28.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n" + "ld1rw { z6.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z9.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z10.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n" + "ld1rw { z7.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n" "1:" // Channel loop "ldr x20, [%x[qp], %[offsetof_Requantize32_bias]]\n" - "mov z16.s, #0x0\n" + "mov z28.s, #0x0\n" "cbz x20, 2f\n" - "ld1w { z16.s }, p1/Z, [x20, x16, LSL #2]\n" + "ld1w { z28.s }, p1/Z, [x20, x16, LSL #2]\n" "2:" // Load bias: Done - "ldr x23, [%x[args], %[offsetof_Args_weights]]\n" - "ld1rh { z22.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n" - "mov z27.h, #0x0\n" - "addvl x22, SP, #6\n" - "addvl x22, x22, #-2\n" - "ldr x21, [%x[qp], %[offsetof_Requantize32_per_channel_muls]]\n" - "mov z17.d, z16.d\n" - "mov z18.d, z16.d\n" - "mov z19.d, z16.d\n" - "mov x20, x23\n" - "incw x23\n" - "ld1sb { z25.s }, p2/Z, [x20]\n" + "ldr x22, [%x[args], %[offsetof_Args_weights]]\n" + "mov x20, x22\n" + "ld1sb { z26.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "ld1sb { z15.s }, p2/Z, [x20]\n" + "ld1rh { z16.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n" + "sub z26.h, z26.h, z16.h\n" + "incw x22\n" + "mov z24.h, #0x0\n" + "ld1sb { z3.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "ld1sb { z9.s }, p2/Z, [x20]\n" - "mov x20, x23\n" - "incw x23\n" - "sub z25.h, z25.h, z22.h\n" - "sub z15.h, z15.h, z22.h\n" - "sub z9.h, z9.h, z22.h\n" - "trn1 z24.h, z25.h, z15.h\n" - "ld1sb { z12.s }, p2/Z, [x20]\n" + "sub z3.h, z3.h, z16.h\n" + "trn1 z31.h, z26.h, z3.h\n" + "ld1sb { z21.s }, p2/Z, [x20]\n" + "sub z21.h, z21.h, z16.h\n" + "mov x20, x22\n" + "trn1 z14.h, z21.h, z24.h\n" + "ld1sb { z2.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "ld1sb { z4.s }, p2/Z, [x20]\n" + "sub z2.h, z2.h, z16.h\n" + "addvl x21, SP, #6\n" + "ld1sb { z25.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "trn1 z11.h, z9.h, z27.h\n" - "ld1sb { z15.s }, p2/Z, [x20]\n" - "mov x20, x23\n" - "sub z12.h, z12.h, z22.h\n" - "sub z4.h, z4.h, z22.h\n" - "st1h { z24.h }, p2, [x22]\n" - "sub z15.h, z15.h, z22.h\n" - "st1h { z11.h }, p2, [x22, #1, MUL VL]\n" - "addvl x22, x22, #-2\n" - "trn1 z9.h, z12.h, z4.h\n" - "ld1sb { z14.s }, p2/Z, [x20]\n" + "sub z25.h, z25.h, z16.h\n" + "incw x22\n" + "ld1sb { z27.s }, p2/Z, [x20]\n" + "sub z27.h, z27.h, z16.h\n" + "addvl x21, x21, #-2\n" + "mov x20, x22\n" + "st1h { z31.h }, p2, [x21]\n" + "trn1 z4.h, z2.h, z25.h\n" + "ld1sb { z26.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "ld1sb { z10.s }, p2/Z, [x20]\n" + "ld1sb { z23.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "trn1 z21.h, z15.h, z27.h\n" - "ld1sb { z30.s }, p2/Z, [x20]\n" - "sub z14.h, z14.h, z22.h\n" - "sub z10.h, z10.h, z22.h\n" - "st1h { z9.h }, p2, [x22]\n" - "sub z30.h, z30.h, z22.h\n" - "st1h { z21.h }, p2, [x22, #1, MUL VL]\n" - "addvl x22, x22, #-2\n" - "trn1 z15.h, z14.h, z10.h\n" - "trn1 z25.h, z30.h, z27.h\n" - "st1h { z15.h }, p2, [x22]\n" - "st1h { z25.h }, p2, [x22, #1, MUL VL]\n" - "cbz x21, 3f\n" - "ld1w { z1.s }, p1/Z, [x21, x16, LSL #2]\n" + "st1h { z14.h }, p2, [x21, #1, MUL VL]\n" + "trn1 z12.h, z27.h, z24.h\n" + "ld1sb { z20.s }, p2/Z, [x20]\n" + "sub z26.h, z26.h, z16.h\n" + "sub z23.h, z23.h, z16.h\n" + "ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_muls]]\n" + "sub z20.h, z20.h, z16.h\n" + "addvl x21, x21, #-2\n" + "st1h { z4.h }, p2, [x21]\n" + "mov z29.d, z28.d\n" + "st1h { z12.h }, p2, [x21, #1, MUL VL]\n" + "addvl x21, x21, #-2\n" + "mov z30.d, z28.d\n" + "mov z31.d, z28.d\n" + "trn1 z25.h, z26.h, z23.h\n" + "st1h { z25.h }, p2, [x21]\n" + "trn1 z3.h, z20.h, z24.h\n" + "st1h { z3.h }, p2, [x21, #1, MUL VL]\n" + "cbz x20, 3f\n" + "ld1w { z6.s }, p1/Z, [x20, x16, LSL #2]\n" "3:" // Load mul: End "ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_right_shifts]]\n" "cbz x20, 4f\n" - "ld1w { z0.s }, p1/Z, [x20, x16, LSL #2]\n" + "ld1w { z9.s }, p1/Z, [x20, x16, LSL #2]\n" "4:" // Load right_shift: End "ldr x15, [%x[args], %[offsetof_Args_input_cols]]\n" - "mov x22, #0x9\n" - "add x20, x7, x6\n" + "sub x20, x15, #0x1\n" + "orr x23, x20, %x[ld_in_col], LSL #16\n" "ldr x14, [%x[args], %[offsetof_Args_inptr]]\n" - "lsl x21, %x[ld_in_row], #0x0\n" - "mov x8, #0x0\n" + "orr x23, x17, x23, LSL #22\n" + "mov x22, #0x9\n" + "add x21, x7, x6\n" + "lsl x20, %x[ld_in_row], #0x0\n" "ldr x13, [%x[args], %[offsetof_Args_output_cols]]\n" - "sub x22, x22, x20\n" - "sub x20, x15, #0x1\n" - "orr x20, x20, %x[ld_in_col], LSL #16\n" - "madd x21, x21, x7, x14\n" - "orr x20, x17, x20, LSL #22\n" - "lsl x20, x20, #0x0\n" + "mov x8, #0x0\n" + "lsl x23, x23, #0x0\n" + "sub x22, x22, x21\n" + "madd x20, x20, x7, x14\n" "5:" // Issue prefetches "subs x22, x22, #0x1\n" - ".inst 0xf8b44abc // rprfm pldstrm, x20, [x21]\n" - "add x21, x21, %x[ld_in_col]\n" + ".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n" + "add x20, x20, %x[ld_in_col]\n" "bgt 5b\n" "ldr x23, [%x[args], %[offsetof_Args_outptrs]]\n" - "lsl x21, %x[ld_in_row], #0x0\n" - ".inst 0xc0040e00 // mova za.d[x8, #0], { z16.d-z19.d }\n" - "mov x22, #0x2\n" + "lsl x20, %x[ld_in_row], #0x0\n" + "msub x14, x7, x20, x14\n" + ".inst 0xc0040f80 // mova za.d[x8, #0], { z28.d-z31.d }\n" "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" - "msub x14, x7, x21, x14\n" - ".inst 0xc0040e01 // mova za.d[x8, #1], { z16.d-z19.d }\n" - "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" - ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" + ".inst 0xc0040f81 // mova za.d[x8, #1], { z28.d-z31.d }\n" + "mov x22, #0x2\n" "ldp x11, x10, [x23], #0x10\n" + ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" "ldp x9, x28, [x20], #0x10\n" + "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" "ldp x27, x26, [x23], #0x10\n" "ldp x25, x24, [x20], #0x10\n" "cbz x21, 7f\n" @@ -196,24 +191,24 @@ void sme2_s8q_planar_3x3_s2_4rows_dot_za_impl( "sub x21, x21, x20\n" "sub x22, x22, x20\n" "cbz x21, 7f\n" - ".inst 0xc0060c0c // mova { z12.d-z15.d }, za.d[x8, #0]\n" + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + ".inst 0xc1a6ac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n" "and x22, x21, #0x1\n" + ".inst 0xc1a9aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z9.s\n" "add x21, x21, #0x1\n" "lsr x21, x21, #0x1\n" + ".inst 0xc1adab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z13.s\n" "sub x13, x13, x21\n" - ".inst 0xc1a1ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" - ".inst 0xc1a0aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n" - ".inst 0xc1a8ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z8.s\n" - ".inst 0xc1bccfec // sclamp { z12.s-z15.s }, z31.s, z28.s\n" + ".inst 0xc1a7cd58 // sclamp { z24.s-z27.s }, z10.s, z7.s\n" "6:" // Left padding "subs x21, x21, #0x1\n" - "st1b { z12.s }, p1, [x11]\n" + "st1b { z24.s }, p1, [x11]\n" "add x11, x11, x9\n" - "st1b { z13.s }, p1, [x10]\n" + "st1b { z25.s }, p1, [x10]\n" "add x10, x10, x28\n" - "st1b { z14.s }, p1, [x27]\n" + "st1b { z26.s }, p1, [x27]\n" "add x27, x27, x25\n" - "st1b { z15.s }, p1, [x26]\n" + "st1b { z27.s }, p1, [x26]\n" "add x26, x26, x24\n" "bgt 6b\n" "7:" // Left padding: End @@ -225,194 +220,194 @@ void sme2_s8q_planar_3x3_s2_4rows_dot_za_impl( "beq 9f\n" "8:" // Unpadded: 2 priming loads "add x21, x14, %x[ld_in_row]\n" - "ld1sb { z23.s }, p1/Z, [x14]\n" + "ld1sb { z1.s }, p1/Z, [x14]\n" "addvl x20, SP, #4\n" - "ld1sb { z4.s }, p1/Z, [x21]\n" + "ld1sb { z21.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "add x14, x14, %x[ld_in_col]\n" - "ld1sb { z24.s }, p1/Z, [x21]\n" + "trn1 z1.h, z1.h, z21.h\n" + "add z1.h, z1.h, z11.h\n" + "ld1sb { z2.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "ld1sb { z5.s }, p1/Z, [x21]\n" + "add x14, x14, %x[ld_in_col]\n" + "ld1sb { z15.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "ld1sb { z25.s }, p1/Z, [x21]\n" + "trn1 z2.h, z2.h, z15.h\n" + "add z2.h, z2.h, z11.h\n" + "ld1sb { z3.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z23.h, z23.h, z4.h\n" - "ld1sb { z6.s }, p1/Z, [x21]\n" + "ld1sb { z21.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "ld1sb { z26.s }, p1/Z, [x21]\n" + "trn1 z3.h, z3.h, z21.h\n" + "add z3.h, z3.h, z11.h\n" + "ld1sb { z4.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z24.h, z24.h, z5.h\n" - "ld1sb { z20.s }, p1/Z, [x21]\n" + "ld1sb { z19.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z25.h, z25.h, z6.h\n" - "ld1sb { z10.s }, p1/Z, [x21]\n" - "add z23.h, z23.h, z29.h\n" - ".inst 0xa1402a86 // ld1h { z6.h, z14.h }, pn10.b/Z, [x20]\n" - "trn1 z26.h, z26.h, z20.h\n" - "add z24.h, z24.h, z29.h\n" - "mov z27.d, z10.d\n" - "add z25.h, z25.h, z29.h\n" - "add z26.h, z26.h, z29.h\n" - "add z27.h, z27.h, z29.h\n" - ".inst 0xc17616e8 // sdot za.s[x8, 0], { z23.h-z26.h }, z6.h\n" - ".inst 0xc17e1708 // sdot za.s[x8, 0], { z24.h-z27.h }, z14.h\n" + "trn1 z4.h, z4.h, z19.h\n" + "add z4.h, z4.h, z11.h\n" + "ld1sb { z8.s }, p1/Z, [x21]\n" + "mov z5.d, z8.d\n" + "add z5.h, z5.h, z11.h\n" + ".inst 0xa1402a80 // ld1h { z0.h, z8.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1701428 // sdot za.s[x8, 0], { z1.h-z4.h }, z0.h\n" + ".inst 0xc1781448 // sdot za.s[x8, 0], { z2.h-z5.h }, z8.h\n" "9:" // Unpadded: 1 priming loads "add x21, x14, %x[ld_in_row]\n" - "ld1sb { z20.s }, p1/Z, [x14]\n" + "ld1sb { z1.s }, p1/Z, [x14]\n" "addvl x20, SP, #2\n" + "ld1sb { z21.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row]\n" + "trn1 z1.h, z1.h, z21.h\n" + "add z1.h, z1.h, z11.h\n" "ld1sb { z2.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "add x14, x14, %x[ld_in_col]\n" - "ld1sb { z21.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "ld1sb { z25.s }, p1/Z, [x21]\n" + "ld1sb { z12.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "ld1sb { z22.s }, p1/Z, [x21]\n" + "trn1 z2.h, z2.h, z12.h\n" + "add z2.h, z2.h, z11.h\n" + "ld1sb { z3.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z20.h, z20.h, z2.h\n" - "ld1sb { z24.s }, p1/Z, [x21]\n" + "ld1sb { z8.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "ld1sb { z23.s }, p1/Z, [x21]\n" + "trn1 z3.h, z3.h, z8.h\n" + "add z3.h, z3.h, z11.h\n" + "ld1sb { z4.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z21.h, z21.h, z25.h\n" - "ld1sb { z9.s }, p1/Z, [x21]\n" + "ld1sb { z5.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z22.h, z22.h, z24.h\n" - "ld1sb { z3.s }, p1/Z, [x21]\n" - "add z20.h, z20.h, z29.h\n" - ".inst 0xa0402a86 // ld1h { z6.h-z7.h }, pn10.b/Z, [x20]\n" - "trn1 z23.h, z23.h, z9.h\n" - "add z21.h, z21.h, z29.h\n" - "mov z24.d, z3.d\n" - "add z22.h, z22.h, z29.h\n" - "add z23.h, z23.h, z29.h\n" - "add z24.h, z24.h, z29.h\n" - ".inst 0xc1761688 // sdot za.s[x8, 0], { z20.h-z23.h }, z6.h\n" - ".inst 0xc17716a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z7.h\n" + "trn1 z4.h, z4.h, z5.h\n" + "add z4.h, z4.h, z11.h\n" + "ld1sb { z5.s }, p1/Z, [x21]\n" + "mov z5.d, z5.d\n" + "add z5.h, z5.h, z11.h\n" + ".inst 0xa1402a80 // ld1h { z0.h, z8.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1701428 // sdot za.s[x8, 0], { z1.h-z4.h }, z0.h\n" + ".inst 0xc1781448 // sdot za.s[x8, 0], { z2.h-z5.h }, z8.h\n" "10:" // Unpadded: 0 priming loads "cmp x15, #0x2\n" - ".inst 0xa0402be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP]\n" + ".inst 0xa0402bee // ld1h { z14.h-z15.h }, pn10.b/Z, [SP]\n" "blt 18f\n" "add x21, x14, %x[ld_in_row]\n" - "ld1sb { z10.s }, p1/Z, [x14]\n" + "ld1sb { z21.s }, p1/Z, [x14]\n" "sub x15, x15, #0x2\n" - "ld1sb { z25.s }, p1/Z, [x21]\n" + "ld1sb { z8.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" + "trn1 z21.h, z21.h, z8.h\n" "sub x13, x13, #0x1\n" - "ld1sb { z11.s }, p1/Z, [x21]\n" + "ld1sb { z22.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "lsr x20, x15, #0x1\n" - "ld1sb { z24.s }, p1/Z, [x21]\n" + "add z21.h, z21.h, z11.h\n" + "ld1sb { z25.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" + "trn1 z22.h, z22.h, z25.h\n" "cmp x20, x13\n" - "ld1sb { z12.s }, p1/Z, [x21]\n" + "ld1sb { z23.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z10.h, z10.h, z25.h\n" "csel x23, x20, x13, LT\n" - "ld1sb { z30.s }, p1/Z, [x21]\n" + "add z22.h, z22.h, z11.h\n" + "ld1sb { z18.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row]\n" + "trn1 z23.h, z23.h, z18.h\n" + "add z23.h, z23.h, z11.h\n" + "ld1sb { z24.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "add x14, x14, %x[ld_in_col]\n" - "ld1sb { z13.s }, p1/Z, [x21]\n" + "ld1sb { z19.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z24.h\n" + "trn1 z24.h, z24.h, z19.h\n" + "add z24.h, z24.h, z11.h\n" + "ld1sb { z8.s }, p1/Z, [x21]\n" + "mov z25.d, z8.d\n" + "add z25.h, z25.h, z11.h\n" "and x15, x15, #0x1\n" - "ld1sb { z20.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" "sub x13, x13, x23\n" - "ld1sb { z26.s }, p1/Z, [x21]\n" - "trn1 z12.h, z12.h, z30.h\n" - "add z10.h, z10.h, z29.h\n" - "trn1 z13.h, z13.h, z20.h\n" - "add z11.h, z11.h, z29.h\n" - "mov z14.d, z26.d\n" - "add z12.h, z12.h, z29.h\n" - "add z13.h, z13.h, z29.h\n" - "add z14.h, z14.h, z29.h\n" "cbz x23, 17f\n" "11:" // Unpadded: Main loop - ".inst 0xc1721548 // sdot za.s[x8, 0], { z10.h-z13.h }, z2.h\n" + ".inst 0xc17e16a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z14.h\n" "addvl x20, SP, #4\n" "add x22, x14, %x[ld_in_row]\n" + ".inst 0xc17f16c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z15.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" "addvl x21, SP, #2\n" "subs x23, x23, #0x1\n" - ".inst 0xc1731568 // sdot za.s[x8, 0], { z11.h-z14.h }, z3.h\n" - ".inst 0xa1402a87 // ld1h { z7.h, z15.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1771549 // sdot za.s[x8, 1], { z10.h-z13.h }, z7.h\n" - "ld1sb { z3.s }, p1/Z, [x14]\n" + ".inst 0xc17016a9 // sdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + "ld1sb { z21.s }, p1/Z, [x14]\n" "add x14, x14, %x[ld_in_col]\n" - "ld1sb { z9.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" "add x20, x14, %x[ld_in_row]\n" - ".inst 0xc17f1569 // sdot za.s[x8, 1], { z11.h-z14.h }, z15.h\n" - "ld1sb { z4.s }, p1/Z, [x22]\n" + "ld1sb { z18.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "trn1 z3.h, z3.h, z9.h\n" - "ld1sb { z15.s }, p1/Z, [x22]\n" + ".inst 0xc17116c9 // sdot za.s[x8, 1], { z22.h-z25.h }, z1.h\n" + "trn1 z21.h, z21.h, z18.h\n" + "ld1sb { z22.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "add x8, x8, #0x1\n" - "ld1sb { z5.s }, p1/Z, [x22]\n" + "add z21.h, z21.h, z11.h\n" + ".inst 0xc0060c00 // mova { z0.d-z3.d }, za.d[x8, #0]\n" + "ld1sb { z8.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - ".inst 0xc1a1ac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z1.s\n" - "ld1sb { z10.s }, p1/Z, [x22]\n" + "trn1 z22.h, z22.h, z8.h\n" + "add z22.h, z22.h, z11.h\n" + "ld1sb { z23.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" - "ld1sb { z6.s }, p1/Z, [x22]\n" + "add x8, x8, #0x1\n" + ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" + "ld1sb { z27.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "trn1 z4.h, z4.h, z15.h\n" - "add z3.h, z3.h, z29.h\n" - "ld1sb { z14.s }, p1/Z, [x22]\n" + "trn1 z23.h, z23.h, z27.h\n" + "add z23.h, z23.h, z11.h\n" + "ld1sb { z24.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "trn1 z5.h, z5.h, z10.h\n" - "ld1sb { z21.s }, p1/Z, [x22]\n" - ".inst 0xc1a0aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z0.s\n" - ".inst 0xa0402aaa // ld1h { z10.h-z11.h }, pn10.b/Z, [x21]\n" - "trn1 z6.h, z6.h, z14.h\n" - "add z4.h, z4.h, z29.h\n" - "mov z7.d, z21.d\n" - "add z5.h, z5.h, z29.h\n" - ".inst 0xc1a8ab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z8.s\n" - "add z6.h, z6.h, z29.h\n" - "add z7.h, z7.h, z29.h\n" - ".inst 0xc1bccff8 // sclamp { z24.s-z27.s }, z31.s, z28.s\n" - ".inst 0xc17a1468 // sdot za.s[x8, 0], { z3.h-z6.h }, z10.h\n" - "ld1sb { z10.s }, p1/Z, [x14]\n" - "add x14, x14, %x[ld_in_col]\n" + ".inst 0xc1a6ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" + "ld1sb { z8.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row]\n" + "trn1 z24.h, z24.h, z8.h\n" + "add z24.h, z24.h, z11.h\n" + "ld1sb { z4.s }, p1/Z, [x22]\n" + "mov z25.d, z4.d\n" + "add z25.h, z25.h, z11.h\n" + ".inst 0xa1402aa4 // ld1h { z4.h, z12.h }, pn10.b/Z, [x21]\n" + ".inst 0xc17416a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z4.h\n" + ".inst 0xc1a9aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n" + "ld1sb { z21.s }, p1/Z, [x14]\n" + ".inst 0xc17c16c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z12.h\n" + ".inst 0xc1adab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z13.s\n" + "ld1sb { z12.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z21.h, z21.h, z12.h\n" + ".inst 0xc1a7cd40 // sclamp { z0.s-z3.s }, z10.s, z7.s\n" "ld1sb { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "st1b { z24.s }, p1, [x11]\n" + "st1b { z0.s }, p1, [x11]\n" "add x11, x11, x9\n" - ".inst 0xc17b1488 // sdot za.s[x8, 0], { z4.h-z7.h }, z11.h\n" - "ld1sb { z11.s }, p1/Z, [x20]\n" + "ld1sb { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z10.h, z10.h, z22.h\n" - "ld1sb { z14.s }, p1/Z, [x20]\n" + "trn1 z22.h, z22.h, z20.h\n" + "st1b { z1.s }, p1, [x10]\n" + "ld1sb { z23.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "st1b { z25.s }, p1, [x10]\n" "add x10, x10, x28\n" - "ld1sb { z12.s }, p1/Z, [x20]\n" + "st1b { z2.s }, p1, [x27]\n" + "ld1sb { z24.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "st1b { z26.s }, p1, [x27]\n" + "trn1 z23.h, z23.h, z24.h\n" "add x27, x27, x25\n" - "ld1sb { z9.s }, p1/Z, [x20]\n" + "ld1sb { z24.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "st1b { z27.s }, p1, [x26]\n" + "st1b { z3.s }, p1, [x26]\n" "add x26, x26, x24\n" - "ld1sb { z13.s }, p1/Z, [x20]\n" + "ld1sb { z3.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z14.h\n" - "add z10.h, z10.h, z29.h\n" - "ld1sb { z6.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" - "ld1sb { z20.s }, p1/Z, [x20]\n" - "trn1 z12.h, z12.h, z9.h\n" - ".inst 0xa0402be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP]\n" - "trn1 z13.h, z13.h, z6.h\n" - "add z11.h, z11.h, z29.h\n" - "mov z14.d, z20.d\n" - "add z12.h, z12.h, z29.h\n" - "add z13.h, z13.h, z29.h\n" - "add z14.h, z14.h, z29.h\n" + "trn1 z24.h, z24.h, z3.h\n" + "add z21.h, z21.h, z11.h\n" + "ld1sb { z3.s }, p1/Z, [x20]\n" + "mov z25.d, z3.d\n" + "add z22.h, z22.h, z11.h\n" + "add x14, x14, %x[ld_in_col]\n" + ".inst 0xa0402bee // ld1h { z14.h-z15.h }, pn10.b/Z, [SP]\n" + "add z23.h, z23.h, z11.h\n" + "add z24.h, z24.h, z11.h\n" + "add z25.h, z25.h, z11.h\n" "bgt 11b\n" "b 17f\n" "12:" // Padded @@ -422,440 +417,440 @@ void sme2_s8q_planar_3x3_s2_4rows_dot_za_impl( "beq 14f\n" "13:" // Padded: 2 priming loads "mov x12, #0x0\n" - "add x21, x14, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "addvl x20, SP, #4\n" - ".inst 0xa1402a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20]\n" - "ld1sb { z20.s }, p0/Z, [x14]\n" - "add x14, x14, %x[ld_in_col]\n" - "add z20.h, p0/M, z20.h, z29.h\n" + "ld1sb { z22.s }, p0/Z, [x14]\n" + "add z22.h, p0/M, z22.h, z11.h\n" + "add x20, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z22.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z22.h, p0/M, z22.h, z29.h\n" + "ld1sb { z17.s }, p0/Z, [x20]\n" + "add z17.h, p0/M, z17.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z21.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z20.h, z20.h, z22.h\n" - "add z21.h, p0/M, z21.h, z29.h\n" + "ld1sb { z23.s }, p0/Z, [x20]\n" + "add z23.h, p0/M, z23.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z4.s }, p0/Z, [x20]\n" + "add z4.h, p0/M, z4.h, z11.h\n" "mov x12, #0x4\n" - "ld1sb { z23.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z23.h, p0/M, z23.h, z29.h\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z22.h, z22.h, z17.h\n" + "trn1 z23.h, z23.h, z4.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z22.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z21.h, z21.h, z23.h\n" - "add z22.h, p0/M, z22.h, z29.h\n" + "ld1sb { z24.s }, p0/Z, [x20]\n" + "add z24.h, p0/M, z24.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z24.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z24.h, p0/M, z24.h, z29.h\n" + "ld1sb { z18.s }, p0/Z, [x20]\n" + "add z18.h, p0/M, z18.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z23.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z22.h, z22.h, z24.h\n" - "add z23.h, p0/M, z23.h, z29.h\n" + "ld1sb { z25.s }, p0/Z, [x20]\n" + "add z25.h, p0/M, z25.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z17.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" - "ld1sb { z25.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z25.h, p0/M, z25.h, z29.h\n" + "add z17.h, p0/M, z17.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z24.s }, p0/Z, [x21]\n" - "trn1 z23.h, z23.h, z25.h\n" - "add z24.h, p0/M, z24.h, z29.h\n" - ".inst 0xc1731688 // sdot za.s[x8, 0], { z20.h-z23.h }, z3.h\n" - "mov z24.d, z24.d\n" - ".inst 0xc17b16a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z11.h\n" + "ld1sb { z1.s }, p0/Z, [x20]\n" + "add z1.h, p0/M, z1.h, z11.h\n" + "addvl x20, SP, #4\n" + "trn1 z24.h, z24.h, z18.h\n" + "trn1 z25.h, z25.h, z17.h\n" + ".inst 0xa1402a84 // ld1h { z4.h, z12.h }, pn10.b/Z, [x20]\n" + "mov z26.d, z1.d\n" + ".inst 0xc17416c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z4.h\n" + "add x14, x14, %x[ld_in_col]\n" + ".inst 0xc17c16e8 // sdot za.s[x8, 0], { z23.h-z26.h }, z12.h\n" "14:" // Padded: 1 priming loads "mov x12, #0x0\n" - "add x21, x14, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "addvl x20, SP, #2\n" - ".inst 0xa1402a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20]\n" - "ld1sb { z21.s }, p0/Z, [x14]\n" - "add x14, x14, %x[ld_in_col]\n" - "add z21.h, p0/M, z21.h, z29.h\n" + "ld1sb { z22.s }, p0/Z, [x14]\n" + "add z22.h, p0/M, z22.h, z11.h\n" + "add x20, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" + "ld1sb { z17.s }, p0/Z, [x20]\n" + "add z17.h, p0/M, z17.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z22.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z21.h, z21.h, z20.h\n" - "add z22.h, p0/M, z22.h, z29.h\n" + "ld1sb { z23.s }, p0/Z, [x20]\n" + "add z23.h, p0/M, z23.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z5.s }, p0/Z, [x20]\n" + "add z5.h, p0/M, z5.h, z11.h\n" "mov x12, #0x4\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z22.h, z22.h, z17.h\n" + "trn1 z23.h, z23.h, z5.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z23.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z22.h, z22.h, z20.h\n" - "add z23.h, p0/M, z23.h, z29.h\n" + "ld1sb { z24.s }, p0/Z, [x20]\n" + "add z24.h, p0/M, z24.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" + "ld1sb { z18.s }, p0/Z, [x20]\n" + "add z18.h, p0/M, z18.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z24.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z23.h, z23.h, z20.h\n" - "add z24.h, p0/M, z24.h, z29.h\n" + "ld1sb { z25.s }, p0/Z, [x20]\n" + "add z25.h, p0/M, z25.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z17.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" - "ld1sb { z25.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z25.h, p0/M, z25.h, z29.h\n" + "add z17.h, p0/M, z17.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" - "trn1 z24.h, z24.h, z25.h\n" - "add z20.h, p0/M, z20.h, z29.h\n" - ".inst 0xc17316a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z3.h\n" - "mov z25.d, z20.d\n" - ".inst 0xc17b16c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z11.h\n" + "ld1sb { z15.s }, p0/Z, [x20]\n" + "add z15.h, p0/M, z15.h, z11.h\n" + "addvl x20, SP, #2\n" + "trn1 z24.h, z24.h, z18.h\n" + "trn1 z25.h, z25.h, z17.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" + "mov z26.d, z15.d\n" + ".inst 0xc17016c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z0.h\n" + "add x14, x14, %x[ld_in_col]\n" + ".inst 0xc17116e8 // sdot za.s[x8, 0], { z23.h-z26.h }, z1.h\n" "15:" // Padded: 0 priming loads "cmp x15, #0x2\n" - ".inst 0xa0402be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP]\n" + ".inst 0xa0402bee // ld1h { z14.h-z15.h }, pn10.b/Z, [SP]\n" "blt 18f\n" "mov x12, #0x0\n" - "add x21, x14, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1sb { z21.s }, p0/Z, [x14]\n" + "add z21.h, p0/M, z21.h, z11.h\n" + "add x20, x14, %x[ld_in_row]\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1sb { z18.s }, p0/Z, [x20]\n" + "add z18.h, p0/M, z18.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1sb { z22.s }, p0/Z, [x20]\n" + "add z22.h, p0/M, z22.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z3.s }, p0/Z, [x20]\n" + "add z3.h, p0/M, z3.h, z11.h\n" + "mov x12, #0x4\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z21.h, z21.h, z18.h\n" + "trn1 z22.h, z22.h, z3.h\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1sb { z23.s }, p0/Z, [x20]\n" + "add z23.h, p0/M, z23.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1sb { z19.s }, p0/Z, [x20]\n" + "add z19.h, p0/M, z19.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1sb { z24.s }, p0/Z, [x20]\n" + "add z24.h, p0/M, z24.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z20.s }, p0/Z, [x20]\n" + "mov x12, #0x8\n" + "add z20.h, p0/M, z20.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1sb { z3.s }, p0/Z, [x20]\n" + "add z3.h, p0/M, z3.h, z11.h\n" "sub x15, x15, #0x2\n" "sub x13, x13, #0x1\n" + "trn1 z23.h, z23.h, z19.h\n" + "trn1 z24.h, z24.h, z20.h\n" "lsr x20, x15, #0x1\n" "cmp x20, x13\n" - "and x15, x15, #0x1\n" - "ld1sb { z10.s }, p0/Z, [x14]\n" - "csel x23, x20, x13, LT\n" + "mov z25.d, z3.d\n" + "csel x22, x20, x13, LT\n" "add x14, x14, %x[ld_in_col]\n" - "sub x13, x13, x23\n" - "add z10.h, p0/M, z10.h, z29.h\n" + "and x15, x15, #0x1\n" + "sub x13, x13, x22\n" + "cbz x22, 17f\n" + "16:" // Padded: Main loop + ".inst 0xc17e16a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z14.h\n" + "addvl x20, SP, #4\n" + "mov x12, #0x0\n" + ".inst 0xc17f16c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z15.h\n" + ".inst 0xa1402a84 // ld1h { z4.h, z12.h }, pn10.b/Z, [x20]\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "add x21, x14, %x[ld_in_row]\n" + ".inst 0xc17416a9 // sdot za.s[x8, 1], { z21.h-z24.h }, z4.h\n" + "ld1sb { z21.s }, p0/Z, [x14]\n" + "add z21.h, p0/M, z21.h, z11.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" + "ld1sb { z14.s }, p0/Z, [x21]\n" + "add z14.h, p0/M, z14.h, z11.h\n" "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z11.s }, p0/Z, [x21]\n" + ".inst 0xc17c16c9 // sdot za.s[x8, 1], { z22.h-z25.h }, z12.h\n" + "ld1sb { z22.s }, p0/Z, [x21]\n" + "add z22.h, p0/M, z22.h, z11.h\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z10.h, z10.h, z20.h\n" - "add z11.h, p0/M, z11.h, z29.h\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z15.s }, p0/Z, [x21]\n" "mov x12, #0x4\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" + "add z15.h, p0/M, z15.h, z11.h\n" "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z12.s }, p0/Z, [x21]\n" + "ld1sb { z23.s }, p0/Z, [x21]\n" + "add z23.h, p0/M, z23.h, z11.h\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z20.h\n" - "add z12.h, p0/M, z12.h, z29.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" + "ld1sb { z17.s }, p0/Z, [x21]\n" + "add z17.h, p0/M, z17.h, z11.h\n" "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z13.s }, p0/Z, [x21]\n" + "ld1sb { z24.s }, p0/Z, [x21]\n" + "add z24.h, p0/M, z24.h, z11.h\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z12.h, z12.h, z20.h\n" - "add z13.h, p0/M, z13.h, z29.h\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z4.s }, p0/Z, [x21]\n" + "add z4.h, p0/M, z4.h, z11.h\n" "mov x12, #0x8\n" - "ld1sb { z21.s }, p0/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "add z21.h, p0/M, z21.h, z29.h\n" + "trn1 z21.h, z21.h, z14.h\n" + "trn1 z22.h, z22.h, z15.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" - "trn1 z13.h, z13.h, z21.h\n" - "add z20.h, p0/M, z20.h, z29.h\n" - "mov z14.d, z20.d\n" - "cbz x23, 17f\n" - "16:" // Padded: Main loop - ".inst 0xc1721548 // sdot za.s[x8, 0], { z10.h-z13.h }, z2.h\n" - "addvl x20, SP, #4\n" + "addvl x20, SP, #2\n" + "ld1sb { z2.s }, p0/Z, [x21]\n" + "trn1 z23.h, z23.h, z17.h\n" + "trn1 z24.h, z24.h, z4.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" "mov x12, #0x0\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "add x22, x14, %x[ld_in_row]\n" - "addvl x21, SP, #2\n" - "subs x23, x23, #0x1\n" - ".inst 0xc1731568 // sdot za.s[x8, 0], { z11.h-z14.h }, z3.h\n" - ".inst 0xa0402a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1721549 // sdot za.s[x8, 1], { z10.h-z13.h }, z2.h\n" - "ld1sb { z10.s }, p0/Z, [x14]\n" + ".inst 0xc0060c10 // mova { z16.d-z19.d }, za.d[x8, #0]\n" + "add x8, x8, #0x1\n" + "add z2.h, p0/M, z2.h, z11.h\n" "add x14, x14, %x[ld_in_col]\n" - ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + ".inst 0xc17016a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z0.h\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1sb { z21.s }, p0/Z, [x14]\n" + "add z21.h, p0/M, z21.h, z11.h\n" "add x20, x14, %x[ld_in_row]\n" - "add z10.h, p0/M, z10.h, z29.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1731569 // sdot za.s[x8, 1], { z11.h-z14.h }, z3.h\n" - ".inst 0xa0402aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21]\n" - "add x8, x8, #0x1\n" - ".inst 0xc1a1ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" - "ld1sb { z26.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "add z26.h, p0/M, z26.h, z29.h\n" + "mov z25.d, z2.d\n" + "ld1sb { z20.s }, p0/Z, [x20]\n" + "add z20.h, p0/M, z20.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xc1a0aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" - "ld1sb { z11.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "trn1 z10.h, z10.h, z26.h\n" - ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" - "add z11.h, p0/M, z11.h, z29.h\n" + ".inst 0xc17116c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z1.h\n" + "ld1sb { z22.s }, p0/Z, [x20]\n" + "add z22.h, p0/M, z22.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z4.s }, p0/Z, [x20]\n" "mov x12, #0x4\n" - "ld1sb { z9.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - ".inst 0xc1bccfe4 // sclamp { z4.s-z7.s }, z31.s, z28.s\n" - "add z9.h, p0/M, z9.h, z29.h\n" + "add z4.h, p0/M, z4.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" + ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "st1b { z4.s }, p1, [x11]\n" - "add x11, x11, x9\n" - "st1b { z5.s }, p1, [x10]\n" - "add x10, x10, x28\n" - "ld1sb { z12.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "st1b { z6.s }, p1, [x27]\n" - "add x27, x27, x25\n" - "trn1 z11.h, z11.h, z9.h\n" - "st1b { z7.s }, p1, [x26]\n" - "add x26, x26, x24\n" - "add z12.h, p0/M, z12.h, z29.h\n" + "ld1sb { z23.s }, p0/Z, [x20]\n" + "add z23.h, p0/M, z23.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z9.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "add z9.h, p0/M, z9.h, z29.h\n" + "ld1sb { z27.s }, p0/Z, [x20]\n" + "add z27.h, p0/M, z27.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z13.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "trn1 z12.h, z12.h, z9.h\n" - "add z13.h, p0/M, z13.h, z29.h\n" + "ld1sb { z24.s }, p0/Z, [x20]\n" + ".inst 0xc1a6ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" + "add z24.h, p0/M, z24.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z12.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" - "ld1sb { z20.s }, p0/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" + ".inst 0xc1a9aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z9.s\n" + "add x20, x20, %x[ld_in_row]\n" + "add z12.h, p0/M, z12.h, z11.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1sb { z8.s }, p0/Z, [x20]\n" + "add z8.h, p0/M, z8.h, z11.h\n" + ".inst 0xc1adab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z13.s\n" + "subs x22, x22, #0x1\n" + ".inst 0xa0402bee // ld1h { z14.h-z15.h }, pn10.b/Z, [SP]\n" + ".inst 0xc1a7cd50 // sclamp { z16.s-z19.s }, z10.s, z7.s\n" + "st1b { z16.s }, p1, [x11]\n" + "add x11, x11, x9\n" + "trn1 z21.h, z21.h, z20.h\n" + "st1b { z17.s }, p1, [x10]\n" + "add x10, x10, x28\n" + "trn1 z22.h, z22.h, z4.h\n" + "trn1 z23.h, z23.h, z27.h\n" + "st1b { z18.s }, p1, [x27]\n" + "add x27, x27, x25\n" + "trn1 z24.h, z24.h, z12.h\n" + "mov z25.d, z8.d\n" + "st1b { z19.s }, p1, [x26]\n" + "add x26, x26, x24\n" + "add x14, x14, %x[ld_in_col]\n" + "bgt 16b\n" + "17:" // Main loop tail + ".inst 0xc17e16a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z14.h\n" + "addvl x20, SP, #4\n" "mov x12, #0x0\n" - "ld1sb { z9.s }, p0/Z, [x22]\n" - "trn1 z13.h, z13.h, z20.h\n" - "add z9.h, p0/M, z9.h, z29.h\n" + ".inst 0xc17f16c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z15.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xc1721548 // sdot za.s[x8, 0], { z10.h-z13.h }, z2.h\n" - "ld1sb { z10.s }, p0/Z, [x14]\n" - "add x14, x14, %x[ld_in_col]\n" - "mov z14.d, z9.d\n" - "add z10.h, p0/M, z10.h, z29.h\n" + "add x20, x14, %x[ld_in_row]\n" + ".inst 0xc17016a9 // sdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + "ld1sb { z0.s }, p0/Z, [x14]\n" + "add z0.h, p0/M, z0.h, z11.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1731568 // sdot za.s[x8, 0], { z11.h-z14.h }, z3.h\n" - ".inst 0xa0402be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP]\n" - "ld1sb { z25.s }, p0/Z, [x20]\n" + "ld1sb { z14.s }, p0/Z, [x20]\n" + "add z14.h, p0/M, z14.h, z11.h\n" "add x20, x20, %x[ld_in_row]\n" - "add z25.h, p0/M, z25.h, z29.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z11.s }, p0/Z, [x20]\n" + ".inst 0xc17116c9 // sdot za.s[x8, 1], { z22.h-z25.h }, z1.h\n" + "ld1sb { z1.s }, p0/Z, [x20]\n" + "add z1.h, p0/M, z1.h, z11.h\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z10.h, z10.h, z25.h\n" - "add z11.h, p0/M, z11.h, z29.h\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z12.s }, p0/Z, [x20]\n" "mov x12, #0x4\n" - "ld1sb { z15.s }, p0/Z, [x20]\n" + "add z12.h, p0/M, z12.h, z11.h\n" "add x20, x20, %x[ld_in_row]\n" - "add z15.h, p0/M, z15.h, z29.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z12.s }, p0/Z, [x20]\n" + "ld1sb { z2.s }, p0/Z, [x20]\n" + "add z2.h, p0/M, z2.h, z11.h\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z15.h\n" - "add z12.h, p0/M, z12.h, z29.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z4.s }, p0/Z, [x20]\n" + "ld1sb { z21.s }, p0/Z, [x20]\n" + "add z21.h, p0/M, z21.h, z11.h\n" "add x20, x20, %x[ld_in_row]\n" - "add z4.h, p0/M, z4.h, z29.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z13.s }, p0/Z, [x20]\n" + "ld1sb { z3.s }, p0/Z, [x20]\n" + ".inst 0xc0060c10 // mova { z16.d-z19.d }, za.d[x8, #0]\n" + "add z3.h, p0/M, z3.h, z11.h\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z12.h, z12.h, z4.h\n" - "add z13.h, p0/M, z13.h, z29.h\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z25.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" - "ld1sb { z4.s }, p0/Z, [x20]\n" + ".inst 0xc1a6ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" "add x20, x20, %x[ld_in_row]\n" - "add z4.h, p0/M, z4.h, z29.h\n" + "add z25.h, p0/M, z25.h, z11.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z26.s }, p0/Z, [x20]\n" - "trn1 z13.h, z13.h, z4.h\n" - "add z26.h, p0/M, z26.h, z29.h\n" - "mov z14.d, z26.d\n" - "bgt 16b\n" - "17:" // Main loop tail - ".inst 0xc1721548 // sdot za.s[x8, 0], { z10.h-z13.h }, z2.h\n" - "addvl x22, SP, #4\n" - "mov x12, #0x0\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "add x21, x14, %x[ld_in_row]\n" + "ld1sb { z27.s }, p0/Z, [x20]\n" "addvl x20, SP, #2\n" - ".inst 0xc1731568 // sdot za.s[x8, 0], { z11.h-z14.h }, z3.h\n" - ".inst 0xa0402ac2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x22]\n" - ".inst 0xc1721549 // sdot za.s[x8, 1], { z10.h-z13.h }, z2.h\n" - "ld1sb { z9.s }, p0/Z, [x14]\n" - "add x14, x14, %x[ld_in_col]\n" - ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" - "add z9.h, p0/M, z9.h, z29.h\n" - ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xc1731569 // sdot za.s[x8, 1], { z11.h-z14.h }, z3.h\n" - ".inst 0xa0402a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1a9aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z9.s\n" + "trn1 z0.h, z0.h, z14.h\n" "add x8, x8, #0x1\n" - ".inst 0xc1a1ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" - "ld1sb { z15.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z15.h, p0/M, z15.h, z29.h\n" - ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xc1a0aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" - "ld1sb { z10.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z9.h, z9.h, z15.h\n" - ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" - "add z10.h, p0/M, z10.h, z29.h\n" - ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "mov x12, #0x4\n" - "ld1sb { z15.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - ".inst 0xc1bccfe4 // sclamp { z4.s-z7.s }, z31.s, z28.s\n" - "add z15.h, p0/M, z15.h, z29.h\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "st1b { z4.s }, p1, [x11]\n" + "add z27.h, p0/M, z27.h, z11.h\n" + "trn1 z1.h, z1.h, z12.h\n" + "trn1 z2.h, z2.h, z21.h\n" + "add x14, x14, %x[ld_in_col]\n" + "trn1 z3.h, z3.h, z25.h\n" + ".inst 0xa0402a8e // ld1h { z14.h-z15.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1adab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z13.s\n" + "mov z4.d, z27.d\n" + ".inst 0xc17e1408 // sdot za.s[x8, 0], { z0.h-z3.h }, z14.h\n" + ".inst 0xc1a7cd50 // sclamp { z16.s-z19.s }, z10.s, z7.s\n" + "st1b { z16.s }, p1, [x11]\n" "add x11, x11, x9\n" - "st1b { z5.s }, p1, [x10]\n" + ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" + "st1b { z17.s }, p1, [x10]\n" "add x10, x10, x28\n" - "ld1sb { z11.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "st1b { z6.s }, p1, [x27]\n" + ".inst 0xc17f1428 // sdot za.s[x8, 0], { z1.h-z4.h }, z15.h\n" + ".inst 0xa0402bee // ld1h { z14.h-z15.h }, pn10.b/Z, [SP]\n" + "st1b { z18.s }, p1, [x27]\n" "add x27, x27, x25\n" - "trn1 z10.h, z10.h, z15.h\n" - "st1b { z7.s }, p1, [x26]\n" + "st1b { z19.s }, p1, [x26]\n" "add x26, x26, x24\n" - "add z11.h, p0/M, z11.h, z29.h\n" - ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z5.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z5.h, p0/M, z5.h, z29.h\n" - ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z12.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z5.h\n" - "add z12.h, p0/M, z12.h, z29.h\n" - ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "mov x12, #0x8\n" - "ld1sb { z15.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z15.h, p0/M, z15.h, z29.h\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z5.s }, p0/Z, [x21]\n" - "trn1 z12.h, z12.h, z15.h\n" - "add z5.h, p0/M, z5.h, z29.h\n" - ".inst 0xc1721528 // sdot za.s[x8, 0], { z9.h-z12.h }, z2.h\n" - "mov z13.d, z5.d\n" - ".inst 0xc1731548 // sdot za.s[x8, 0], { z10.h-z13.h }, z3.h\n" - ".inst 0xa0402be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP]\n" "18:" // Main loop skip tail "cbz x15, 19f\n" // Skip remainder inputs "mov x12, #0x0\n" - "add x21, x14, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "addvl x20, SP, #4\n" - "sub x13, x13, #0x1\n" "ld1sb { z21.s }, p0/Z, [x14]\n" - "add z21.h, p0/M, z21.h, z29.h\n" + "add z21.h, p0/M, z21.h, z11.h\n" + "add x20, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z15.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z15.h, p0/M, z15.h, z29.h\n" + "ld1sb { z17.s }, p0/Z, [x20]\n" + "add z17.h, p0/M, z17.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z22.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z21.h, z21.h, z15.h\n" - "add z22.h, p0/M, z22.h, z29.h\n" + "ld1sb { z22.s }, p0/Z, [x20]\n" + "add z22.h, p0/M, z22.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z0.s }, p0/Z, [x20]\n" + "add z0.h, p0/M, z0.h, z11.h\n" "mov x12, #0x4\n" - "ld1sb { z12.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z12.h, p0/M, z12.h, z29.h\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z21.h, z21.h, z17.h\n" + "trn1 z22.h, z22.h, z0.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z23.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z22.h, z22.h, z12.h\n" - "add z23.h, p0/M, z23.h, z29.h\n" + "ld1sb { z23.s }, p0/Z, [x20]\n" + "add z23.h, p0/M, z23.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1sb { z20.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z20.h, p0/M, z20.h, z29.h\n" + "ld1sb { z17.s }, p0/Z, [x20]\n" + "add z17.h, p0/M, z17.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1sb { z24.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z23.h, z23.h, z20.h\n" - "add z24.h, p0/M, z24.h, z29.h\n" + "ld1sb { z24.s }, p0/Z, [x20]\n" + "add z24.h, p0/M, z24.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1sb { z5.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" - "ld1sb { z30.s }, p0/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add z30.h, p0/M, z30.h, z29.h\n" + "add z5.h, p0/M, z5.h, z11.h\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1sb { z6.s }, p0/Z, [x21]\n" - "trn1 z24.h, z24.h, z30.h\n" - "add z6.h, p0/M, z6.h, z29.h\n" - ".inst 0xc17216a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z2.h\n" - "mov z25.d, z6.d\n" - ".inst 0xc17316c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z3.h\n" - ".inst 0xa1402a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20]\n" - ".inst 0xc17516a9 // sdot za.s[x8, 1], { z21.h-z24.h }, z5.h\n" - ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" - ".inst 0xc17d16c9 // sdot za.s[x8, 1], { z22.h-z25.h }, z13.h\n" + "ld1sb { z4.s }, p0/Z, [x20]\n" + "add z4.h, p0/M, z4.h, z11.h\n" + "trn1 z23.h, z23.h, z17.h\n" + "trn1 z24.h, z24.h, z5.h\n" + "mov z25.d, z4.d\n" + "addvl x20, SP, #4\n" + ".inst 0xc17e16a8 // sdot za.s[x8, 0], { z21.h-z24.h }, z14.h\n" + "sub x13, x13, #0x1\n" + ".inst 0xc17f16c8 // sdot za.s[x8, 0], { z22.h-z25.h }, z15.h\n" + ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" + ".inst 0xc0060c10 // mova { z16.d-z19.d }, za.d[x8, #0]\n" + ".inst 0xc1a6ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" + ".inst 0xc1a9aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z9.s\n" + ".inst 0xc17016a9 // sdot za.s[x8, 1], { z21.h-z24.h }, z0.h\n" + ".inst 0xc1adab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z13.s\n" + ".inst 0xc17116c9 // sdot za.s[x8, 1], { z22.h-z25.h }, z1.h\n" "add x8, x8, #0x1\n" - ".inst 0xc1a1ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" - ".inst 0xc1a0aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" - ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" - ".inst 0xc1bccfe4 // sclamp { z4.s-z7.s }, z31.s, z28.s\n" - "st1b { z4.s }, p1, [x11]\n" + ".inst 0xc1a7cd50 // sclamp { z16.s-z19.s }, z10.s, z7.s\n" + "st1b { z16.s }, p1, [x11]\n" "add x11, x11, x9\n" - "st1b { z5.s }, p1, [x10]\n" + ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" + "st1b { z17.s }, p1, [x10]\n" "add x10, x10, x28\n" - "st1b { z6.s }, p1, [x27]\n" + "st1b { z18.s }, p1, [x27]\n" "add x27, x27, x25\n" - "st1b { z7.s }, p1, [x26]\n" + "st1b { z19.s }, p1, [x26]\n" "add x26, x26, x24\n" "19:" // Tail input: End "cbz x13, 21f\n" "20:" // Right padding loop - ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + ".inst 0xc0060c00 // mova { z0.d-z3.d }, za.d[x8, #0]\n" + ".inst 0xc1a6ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" "add x8, x8, #0x1\n" + ".inst 0xc1a9aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n" "subs x13, x13, #0x1\n" - ".inst 0xc0040e02 // mova za.d[x8, #2], { z16.d-z19.d }\n" - ".inst 0xc1a1ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - ".inst 0xc1a0aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" - ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" - ".inst 0xc1bccfe4 // sclamp { z4.s-z7.s }, z31.s, z28.s\n" - "st1b { z4.s }, p1, [x11]\n" + ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" + ".inst 0xc1adab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z13.s\n" + ".inst 0xc1a7cd40 // sclamp { z0.s-z3.s }, z10.s, z7.s\n" + "st1b { z0.s }, p1, [x11]\n" "add x11, x11, x9\n" - "st1b { z5.s }, p1, [x10]\n" + "st1b { z1.s }, p1, [x10]\n" "add x10, x10, x28\n" - "st1b { z6.s }, p1, [x27]\n" + "st1b { z2.s }, p1, [x27]\n" "add x27, x27, x25\n" - "st1b { z7.s }, p1, [x26]\n" + "st1b { z3.s }, p1, [x26]\n" "add x26, x26, x24\n" "bgt 20b\n" "21:" // End "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" - "incw x16\n" - "whilelt p1.s, x16, x17\n" "incw x20, ALL, MUL #9\n" "str x20, [%x[args], %[offsetof_Args_weights]]\n" + "incw x16\n" "ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n" + "whilelt p1.s, x16, x17\n" "ldr x20, [%x[args], %[offsetof_Args_inptr]]\n" "add x20, x20, x21\n" "str x20, [%x[args], %[offsetof_Args_inptr]]\n" @@ -874,8 +869,6 @@ void sme2_s8q_planar_3x3_s2_4rows_dot_za_impl( "b.any 1b\n" "addvl SP, SP, #6\n" ".inst 0xd503467f // SMSTOP\n" - "ldr x20, [SP, #0x0]\n" - "mov SP, x20\n" : : [args] "r" (&args), [ld_in_col] "r" (ld_in_col), [ld_in_row] "r" (ld_in_row), [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)), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_bias] "I" (offsetof(arm_gemm::Requantize32, bias)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [offsetof_Requantize32_per_channel_muls] "I" (offsetof(arm_gemm::Requantize32, per_channel_muls)), [offsetof_Requantize32_per_channel_right_shifts] "I" (offsetof(arm_gemm::Requantize32, per_channel_right_shifts)), [offsetof_Requantize32_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_Requantize32_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [qp] "r" (&qp) : "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" |