From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../generic.cpp | 582 ++++++++++----------- 1 file changed, 291 insertions(+), 291 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp index 1636225b31..612beb342a 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp @@ -73,96 +73,96 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( "ptrue p2.b\n" "mov x20, #0x6\n" "ldr x7, [%x[args], %[offsetof_Args_pad_top]]\n" - "ld1rh { z24.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_a_offset]]\n" + "ld1rh { z21.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_a_offset]]\n" "sub x20, x20, x6\n" ".inst 0x25207812 // ptrue pn10.b\n" "ldr x17, [%x[args], %[offsetof_Args_n_channels]]\n" "whilelt p1.s, XZR, x17\n" "whilelt p9.s, XZR, x20\n" - "ld1rw { z12.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n" + "ld1rw { z15.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n" "whilelt p8.s, XZR, x7\n" "addvl SP, SP, #-12\n" "ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n" - "neg z24.h, p2/M, z24.h\n" + "neg z21.h, p2/M, z21.h\n" "eor p8.b, p2/Z, p8.b, p9.b\n" - "ld1rw { z10.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_mul]]\n" - "ld1rw { z11.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z22.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n" - "ld1rw { z26.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n" + "ld1rw { z14.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z12.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z29.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n" + "ld1rw { z28.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n" "1:" // Channel loop "ldr x20, [%x[qp], %[offsetof_Requantize32_bias]]\n" - "mov z8.s, #0x0\n" + "mov z30.s, #0x0\n" "cbz x20, 2f\n" - "ld1w { z8.s }, p1/Z, [x20, x16, LSL #2]\n" + "ld1w { z30.s }, p1/Z, [x20, x16, LSL #2]\n" "2:" // Load bias: Done "ldr x22, [%x[args], %[offsetof_Args_weights]]\n" "mov x20, x22\n" - "ld1sb { z27.s }, p2/Z, [x20]\n" + "ld1sb { z10.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "ld1rh { z21.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n" - "mov z20.h, #0x0\n" - "sub z27.h, z27.h, z21.h\n" + "ld1rh { z31.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n" + "mov z7.h, #0x0\n" + "sub z10.h, z10.h, z31.h\n" "incw x22\n" - "ld1sb { z23.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #3\n" - "sub z23.h, z23.h, z21.h\n" - "trn1 z0.h, z20.h, z27.h\n" "ld1sb { z16.s }, p2/Z, [x20]\n" - "sub z16.h, z16.h, z21.h\n" + "incw x20, ALL, MUL #3\n" + "sub z16.h, z16.h, z31.h\n" + "trn1 z20.h, z7.h, z10.h\n" + "ld1sb { z11.s }, p2/Z, [x20]\n" + "sub z11.h, z11.h, z31.h\n" "mov x20, x22\n" - "trn1 z1.h, z27.h, z23.h\n" - "ld1sb { z27.s }, p2/Z, [x20]\n" + "trn1 z19.h, z10.h, z16.h\n" + "ld1sb { z24.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "trn1 z2.h, z23.h, z16.h\n" - "trn1 z3.h, z16.h, z20.h\n" - "ld1sb { z23.s }, p2/Z, [x20]\n" + "trn1 z26.h, z16.h, z11.h\n" + "trn1 z13.h, z11.h, z7.h\n" + "ld1sb { z11.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "sub z27.h, z27.h, z21.h\n" - "sub z23.h, z23.h, z21.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" - "sub z16.h, z16.h, z21.h\n" + "sub z24.h, z24.h, z31.h\n" + "sub z11.h, z11.h, z31.h\n" + "ld1sb { z2.s }, p2/Z, [x20]\n" + "sub z2.h, z2.h, z31.h\n" "addvl x21, SP, #12\n" "incw x22\n" "addvl x21, x21, #-4\n" "mov x20, x22\n" - "st1h { z0.h }, p2, [x21]\n" - "trn1 z0.h, z20.h, z27.h\n" - "st1h { z1.h }, p2, [x21, #1, MUL VL]\n" - "trn1 z1.h, z27.h, z23.h\n" - "ld1sb { z27.s }, p2/Z, [x20]\n" + "st1h { z20.h }, p2, [x21]\n" + "trn1 z22.h, z7.h, z24.h\n" + "st1h { z19.h }, p2, [x21, #1, MUL VL]\n" + "trn1 z1.h, z24.h, z11.h\n" + "ld1sb { z16.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "st1h { z2.h }, p2, [x21, #2, MUL VL]\n" - "trn1 z2.h, z23.h, z16.h\n" - "ld1sb { z23.s }, p2/Z, [x20]\n" + "st1h { z26.h }, p2, [x21, #2, MUL VL]\n" + "trn1 z3.h, z11.h, z2.h\n" + "ld1sb { z0.s }, p2/Z, [x20]\n" "incw x20, ALL, MUL #3\n" - "st1h { z3.h }, p2, [x21, #3, MUL VL]\n" - "trn1 z3.h, z16.h, z20.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" + "st1h { z13.h }, p2, [x21, #3, MUL VL]\n" + "trn1 z25.h, z2.h, z7.h\n" + "ld1sb { z4.s }, p2/Z, [x20]\n" "ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_muls]]\n" - "sub z27.h, z27.h, z21.h\n" - "sub z23.h, z23.h, z21.h\n" + "sub z16.h, z16.h, z31.h\n" + "sub z0.h, z0.h, z31.h\n" "addvl x21, x21, #-4\n" - "st1h { z0.h }, p2, [x21]\n" - "sub z16.h, z16.h, z21.h\n" + "st1h { z22.h }, p2, [x21]\n" + "sub z4.h, z4.h, z31.h\n" "st1h { z1.h }, p2, [x21, #1, MUL VL]\n" - "mov z9.d, z8.d\n" - "st1h { z2.h }, p2, [x21, #2, MUL VL]\n" - "trn1 z0.h, z20.h, z27.h\n" - "trn1 z1.h, z27.h, z23.h\n" - "st1h { z3.h }, p2, [x21, #3, MUL VL]\n" + "mov z31.d, z30.d\n" + "st1h { z3.h }, p2, [x21, #2, MUL VL]\n" + "trn1 z24.h, z7.h, z16.h\n" + "trn1 z18.h, z16.h, z0.h\n" + "st1h { z25.h }, p2, [x21, #3, MUL VL]\n" "addvl x21, x21, #-4\n" - "trn1 z2.h, z23.h, z16.h\n" - "trn1 z3.h, z16.h, z20.h\n" - "st1h { z0.h }, p2, [x21]\n" - "st1h { z1.h }, p2, [x21, #1, MUL VL]\n" - "st1h { z2.h }, p2, [x21, #2, MUL VL]\n" - "st1h { z3.h }, p2, [x21, #3, MUL VL]\n" + "trn1 z0.h, z0.h, z4.h\n" + "trn1 z1.h, z4.h, z7.h\n" + "st1h { z24.h }, p2, [x21]\n" + "st1h { z18.h }, p2, [x21, #1, MUL VL]\n" + "st1h { z0.h }, p2, [x21, #2, MUL VL]\n" + "st1h { z1.h }, p2, [x21, #3, MUL VL]\n" "cbz x20, 3f\n" - "ld1w { z10.s }, p1/Z, [x20, x16, LSL #2]\n" + "ld1w { z14.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 { z11.s }, p1/Z, [x20, x16, LSL #2]\n" + "ld1w { z12.s }, p1/Z, [x20, x16, LSL #2]\n" "4:" // Load right_shift: End "ldr x15, [%x[args], %[offsetof_Args_input_cols]]\n" "sub x20, x15, #0x1\n" @@ -182,21 +182,21 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( ".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n" "add x20, x20, %x[ld_in_col]\n" "bgt 5b\n" - "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x23, [%x[args], %[offsetof_Args_outptrs]]\n" "lsl x20, %x[ld_in_row], #0x0\n" "msub x14, x7, x20, x14\n" - ".inst 0xc0040900 // mova za.d[x8, #0], { z8.d-z9.d }\n" + ".inst 0xc0040bc0 // mova za.d[x8, #0], { z30.d-z31.d }\n" "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" - ".inst 0xc0040901 // mova za.d[x8, #1], { z8.d-z9.d }\n" + ".inst 0xc0040bc1 // mova za.d[x8, #1], { z30.d-z31.d }\n" "mov x22, #0x2\n" - "ldp x11, x10, [x25], #0x10\n" - ".inst 0xc0040902 // mova za.d[x8, #2], { z8.d-z9.d }\n" + "ldp x11, x10, [x23], #0x10\n" + ".inst 0xc0040bc2 // mova za.d[x8, #2], { z30.d-z31.d }\n" "ldp x9, x28, [x20], #0x10\n" - ".inst 0xc0040903 // mova za.d[x8, #3], { z8.d-z9.d }\n" + ".inst 0xc0040bc3 // mova za.d[x8, #3], { z30.d-z31.d }\n" "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" - ".inst 0xc0040904 // mova za.d[x8, #4], { z8.d-z9.d }\n" - "ldp x27, x26, [x25], #0x10\n" - ".inst 0xc0040905 // mova za.d[x8, #5], { z8.d-z9.d }\n" + ".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n" + "ldp x27, x26, [x23], #0x10\n" + ".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n" "ldp x25, x24, [x20], #0x10\n" "cbz x21, 7f\n" "cmp x21, x22\n" @@ -204,22 +204,22 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( "sub x21, x21, x20\n" "sub x22, x22, x20\n" "cbz x21, 7f\n" - ".inst 0xc0060804 // mova { z4.d-z5.d }, za.d[x8, #0]\n" + ".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n" "sub x13, x13, x21\n" - ".inst 0xc0060826 // mova { z6.d-z7.d }, za.d[x8, #1]\n" - ".inst 0xc1aaac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n" - ".inst 0xc1abaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z11.s\n" - ".inst 0xc1acab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n" - ".inst 0xc1bacec4 // sclamp { z4.s-z7.s }, z22.s, z26.s\n" + ".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n" + ".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n" + ".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n" + ".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n" + ".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n" "6:" // Left padding "subs x21, x21, #0x1\n" - "st1b { z4.s }, p1, [x11]\n" + "st1b { z24.s }, p1, [x11]\n" "add x11, x11, x9\n" - "st1b { z6.s }, p1, [x10]\n" + "st1b { z26.s }, p1, [x10]\n" "add x10, x10, x28\n" - "st1b { z5.s }, p1, [x27]\n" + "st1b { z25.s }, p1, [x27]\n" "add x27, x27, x25\n" - "st1b { z7.s }, p1, [x26]\n" + "st1b { z27.s }, p1, [x26]\n" "add x26, x26, x24\n" "bgt 6b\n" "7:" // Left padding: End @@ -231,148 +231,148 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( "beq 9f\n" "8:" // Unpadded: 2 priming loads "add x21, x14, %x[ld_in_row]\n" - "ld1b { z17.s }, p1/Z, [x14]\n" + "ld1b { z20.s }, p1/Z, [x14]\n" "addvl x20, SP, #8\n" "ld1b { z16.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z13.h, z17.h, z16.h\n" - "add z13.h, z13.h, z24.h\n" - "ld1b { z17.s }, p1/Z, [x21]\n" + "trn1 z4.h, z20.h, z16.h\n" + "add z4.h, z4.h, z21.h\n" + "ld1b { z23.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "add x14, x14, %x[ld_in_col]\n" - "ld1b { z16.s }, p1/Z, [x21]\n" + "ld1b { z22.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "trn1 z14.h, z17.h, z16.h\n" - "add z14.h, z14.h, z24.h\n" + "trn1 z5.h, z23.h, z22.h\n" + "add z5.h, z5.h, z21.h\n" "ld1b { z17.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "ld1b { z16.s }, p1/Z, [x21]\n" - "trn1 z15.h, z17.h, z16.h\n" - "add z15.h, z15.h, z24.h\n" - ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0412a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" + "trn1 z6.h, z17.h, z16.h\n" + "add z6.h, z6.h, z21.h\n" + ".inst 0xa1402a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20]\n" + ".inst 0xc16b1488 // sdot za.s[x8, 0], { z4.h-z5.h }, z11.h\n" + ".inst 0xc1631489 // sdot za.s[x8, 1], { z4.h-z5.h }, z3.h\n" + ".inst 0xa1412a80 // ld1h { z0.h, z8.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xc16814a8 // sdot za.s[x8, 0], { z5.h-z6.h }, z8.h\n" + ".inst 0xc16014a9 // sdot za.s[x8, 1], { z5.h-z6.h }, z0.h\n" "9:" // Unpadded: 1 priming loads "add x22, x14, %x[ld_in_row]\n" - "ld1b { z17.s }, p1/Z, [x14]\n" + "ld1b { z25.s }, p1/Z, [x14]\n" "addvl x21, SP, #4\n" - "ld1b { z16.s }, p1/Z, [x22]\n" + "ld1b { z6.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "trn1 z13.h, z17.h, z16.h\n" - "add z13.h, z13.h, z24.h\n" - "ld1b { z17.s }, p1/Z, [x22]\n" + "trn1 z3.h, z25.h, z6.h\n" + "add z3.h, z3.h, z21.h\n" + "ld1b { z18.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" "addvl x20, SP, #8\n" - "ld1b { z16.s }, p1/Z, [x22]\n" + "ld1b { z26.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "trn1 z14.h, z17.h, z16.h\n" - "add z14.h, z14.h, z24.h\n" - "ld1b { z17.s }, p1/Z, [x22]\n" + "trn1 z4.h, z18.h, z26.h\n" + "add z4.h, z4.h, z21.h\n" + "ld1b { z2.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" "add x14, x14, %x[ld_in_col]\n" - "ld1b { z16.s }, p1/Z, [x22]\n" - "trn1 z15.h, z17.h, z16.h\n" - "add z15.h, z15.h, z24.h\n" + "ld1b { z5.s }, p1/Z, [x22]\n" + "trn1 z5.h, z2.h, z5.h\n" + "add z5.h, z5.h, z21.h\n" ".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n" - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" - ".inst 0xa0412aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc16115aa // sdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n" - ".inst 0xc16015ab // sdot za.s[x8, 3], { z13.h-z14.h }, z0.h\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" - ".inst 0xa0412a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - ".inst 0xc16315ca // sdot za.s[x8, 2], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215cb // sdot za.s[x8, 3], { z14.h-z15.h }, z2.h\n" + ".inst 0xc1611468 // sdot za.s[x8, 0], { z3.h-z4.h }, z1.h\n" + ".inst 0xc1601469 // sdot za.s[x8, 1], { z3.h-z4.h }, z0.h\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa0412aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xc16a146a // sdot za.s[x8, 2], { z3.h-z4.h }, z10.h\n" + ".inst 0xc162146b // sdot za.s[x8, 3], { z3.h-z4.h }, z2.h\n" + ".inst 0xc1691488 // sdot za.s[x8, 0], { z4.h-z5.h }, z9.h\n" + ".inst 0xc1681489 // sdot za.s[x8, 1], { z4.h-z5.h }, z8.h\n" + ".inst 0xa1412a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xc16a148a // sdot za.s[x8, 2], { z4.h-z5.h }, z10.h\n" + ".inst 0xc162148b // sdot za.s[x8, 3], { z4.h-z5.h }, z2.h\n" "10:" // Unpadded: 0 priming loads - ".inst 0xa0402be0 // ld1h { z0.h-z1.h }, pn10.b/Z, [SP]\n" - ".inst 0xa0412be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" + ".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n" + ".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" "cbz x15, 18f\n" "add x20, x14, %x[ld_in_row]\n" "ld1b { z17.s }, p1/Z, [x14]\n" "sub x15, x15, #0x1\n" - "ld1b { z16.s }, p1/Z, [x20]\n" + "ld1b { z9.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z13.h, z17.h, z16.h\n" + "trn1 z6.h, z17.h, z9.h\n" "sub x13, x13, #0x1\n" "ld1b { z17.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" "cmp x15, x13\n" - "add z13.h, z13.h, z24.h\n" - "ld1b { z16.s }, p1/Z, [x20]\n" + "add z6.h, z6.h, z21.h\n" + "ld1b { z7.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z14.h, z17.h, z16.h\n" + "trn1 z7.h, z17.h, z7.h\n" "csel x23, x15, x13, LT\n" "ld1b { z17.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - "add z14.h, z14.h, z24.h\n" + "add z7.h, z7.h, z21.h\n" "add x14, x14, %x[ld_in_col]\n" - "ld1b { z16.s }, p1/Z, [x20]\n" - "trn1 z15.h, z17.h, z16.h\n" - "add z15.h, z15.h, z24.h\n" + "ld1b { z1.s }, p1/Z, [x20]\n" + "trn1 z8.h, z17.h, z1.h\n" + "add z8.h, z8.h, z21.h\n" "sub x13, x13, x23\n" "cbz x23, 17f\n" "11:" // Unpadded: Main loop - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" + ".inst 0xc16d14c8 // sdot za.s[x8, 0], { z6.h-z7.h }, z13.h\n" "addvl x22, SP, #4\n" "addvl x21, SP, #8\n" - "ld1b { z21.s }, p1/Z, [x14]\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0402ac0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x22]\n" + "ld1b { z2.s }, p1/Z, [x14]\n" + ".inst 0xc16514c9 // sdot za.s[x8, 1], { z6.h-z7.h }, z5.h\n" + ".inst 0xa1402ac5 // ld1h { z5.h, z13.h }, pn10.b/Z, [x22]\n" "add x20, x14, %x[ld_in_row]\n" "subs x23, x23, #0x1\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - "ld1b { z20.s }, p1/Z, [x20]\n" + ".inst 0xc16b14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z11.h\n" + "ld1b { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" "add x14, x14, %x[ld_in_col]\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" - ".inst 0xa0412ac2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" - ".inst 0xc0060804 // mova { z4.d-z5.d }, za.d[x8, #0]\n" - "ld1b { z19.s }, p1/Z, [x20]\n" + ".inst 0xc16a14e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z10.h\n" + ".inst 0xa1412ac3 // ld1h { z3.h, z11.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n" + ".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n" + "ld1b { z23.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - ".inst 0xc0060826 // mova { z6.d-z7.d }, za.d[x8, #1]\n" - ".inst 0xc1aaac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n" - ".inst 0xc16115aa // sdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n" + ".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n" + ".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n" + ".inst 0xc16d14ca // sdot za.s[x8, 2], { z6.h-z7.h }, z13.h\n" "ld1b { z18.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - ".inst 0xc16015ab // sdot za.s[x8, 3], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1abaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z11.s\n" - ".inst 0xc16115ac // sdot za.s[x8, 4], { z13.h-z14.h }, z1.h\n" + ".inst 0xc16514cb // sdot za.s[x8, 3], { z6.h-z7.h }, z5.h\n" + ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" + ".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n" + ".inst 0xc16914cc // sdot za.s[x8, 4], { z6.h-z7.h }, z9.h\n" "ld1b { z17.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row]\n" - ".inst 0xc16015ad // sdot za.s[x8, 5], { z13.h-z14.h }, z0.h\n" + ".inst 0xc16114cd // sdot za.s[x8, 5], { z6.h-z7.h }, z1.h\n" "ld1b { z16.s }, p1/Z, [x20]\n" - ".inst 0xc1acab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n" - ".inst 0xc16315ca // sdot za.s[x8, 2], { z14.h-z15.h }, z3.h\n" - "trn1 z13.h, z21.h, z20.h\n" - ".inst 0xa0402be0 // ld1h { z0.h-z1.h }, pn10.b/Z, [SP]\n" - ".inst 0xc16215cb // sdot za.s[x8, 3], { z14.h-z15.h }, z2.h\n" - ".inst 0xa0412aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc1bacec4 // sclamp { z4.s-z7.s }, z22.s, z26.s\n" - ".inst 0xc16315cc // sdot za.s[x8, 4], { z14.h-z15.h }, z3.h\n" - "st1b { z4.s }, p1, [x11]\n" + ".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n" + ".inst 0xc16b14ea // sdot za.s[x8, 2], { z7.h-z8.h }, z11.h\n" + "trn1 z6.h, z2.h, z19.h\n" + ".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n" + ".inst 0xc16314eb // sdot za.s[x8, 3], { z7.h-z8.h }, z3.h\n" + ".inst 0xa1412aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n" + ".inst 0xc16914ec // sdot za.s[x8, 4], { z7.h-z8.h }, z9.h\n" + "st1b { z24.s }, p1, [x11]\n" "add x11, x11, x9\n" - "add z13.h, z13.h, z24.h\n" - ".inst 0xc16215cd // sdot za.s[x8, 5], { z14.h-z15.h }, z2.h\n" - "trn1 z14.h, z19.h, z18.h\n" - "trn1 z15.h, z17.h, z16.h\n" + "add z6.h, z6.h, z21.h\n" + ".inst 0xc16114ed // sdot za.s[x8, 5], { z7.h-z8.h }, z1.h\n" + "trn1 z7.h, z23.h, z18.h\n" + "trn1 z8.h, z17.h, z16.h\n" "add x8, x8, #0x2\n" - ".inst 0xa0412be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" - "st1b { z6.s }, p1, [x10]\n" + ".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" + "st1b { z26.s }, p1, [x10]\n" "add x10, x10, x28\n" - ".inst 0xc0040904 // mova za.d[x8, #4], { z8.d-z9.d }\n" - "st1b { z5.s }, p1, [x27]\n" + ".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n" + "st1b { z25.s }, p1, [x27]\n" "add x27, x27, x25\n" - ".inst 0xc0040905 // mova za.d[x8, #5], { z8.d-z9.d }\n" - "add z14.h, z14.h, z24.h\n" - "st1b { z7.s }, p1, [x26]\n" + ".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n" + "add z7.h, z7.h, z21.h\n" + "st1b { z27.s }, p1, [x26]\n" "add x26, x26, x24\n" - "add z15.h, z15.h, z24.h\n" + "add z8.h, z8.h, z21.h\n" "bgt 11b\n" "b 17f\n" "12:" // Padded @@ -384,118 +384,118 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "ld1b { z19.s }, p0/Z, [x14]\n" - "add z19.h, p0/M, z19.h, z24.h\n" + "add z19.h, p0/M, z19.h, z21.h\n" "add x20, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1b { z18.s }, p0/Z, [x20]\n" - "add z18.h, p0/M, z18.h, z24.h\n" + "add z18.h, p0/M, z18.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "ld1b { z17.s }, p0/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "ld1b { z16.s }, p0/Z, [x20]\n" - "add z16.h, p0/M, z16.h, z24.h\n" + "add z16.h, p0/M, z16.h, z21.h\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z13.h, z19.h, z18.h\n" - "trn1 z14.h, z17.h, z16.h\n" + "trn1 z7.h, z19.h, z18.h\n" + "trn1 z8.h, z17.h, z16.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "ld1b { z17.s }, p0/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1b { z16.s }, p0/Z, [x20]\n" "addvl x20, SP, #8\n" - "add z16.h, p0/M, z16.h, z24.h\n" - ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" - "trn1 z15.h, z17.h, z16.h\n" - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" + "add z16.h, p0/M, z16.h, z21.h\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + "trn1 z9.h, z17.h, z16.h\n" + ".inst 0xc16a14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z10.h\n" "add x14, x14, %x[ld_in_col]\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0412a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" + ".inst 0xc16214e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z2.h\n" + ".inst 0xa1412a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xc16d1508 // sdot za.s[x8, 0], { z8.h-z9.h }, z13.h\n" + ".inst 0xc1651509 // sdot za.s[x8, 1], { z8.h-z9.h }, z5.h\n" "14:" // Padded: 1 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "ld1b { z19.s }, p0/Z, [x14]\n" - "add z19.h, p0/M, z19.h, z24.h\n" + "add z19.h, p0/M, z19.h, z21.h\n" "add x20, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1b { z18.s }, p0/Z, [x20]\n" - "add z18.h, p0/M, z18.h, z24.h\n" + "add z18.h, p0/M, z18.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "ld1b { z17.s }, p0/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "ld1b { z16.s }, p0/Z, [x20]\n" - "add z16.h, p0/M, z16.h, z24.h\n" + "add z16.h, p0/M, z16.h, z21.h\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z13.h, z19.h, z18.h\n" - "trn1 z14.h, z17.h, z16.h\n" + "trn1 z22.h, z19.h, z18.h\n" + "trn1 z23.h, z17.h, z16.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "ld1b { z17.s }, p0/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1b { z16.s }, p0/Z, [x20]\n" "addvl x21, SP, #4\n" - "add z16.h, p0/M, z16.h, z24.h\n" + "add z16.h, p0/M, z16.h, z21.h\n" ".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n" "addvl x20, SP, #8\n" - "trn1 z15.h, z17.h, z16.h\n" - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" + "trn1 z24.h, z17.h, z16.h\n" + ".inst 0xc16116c8 // sdot za.s[x8, 0], { z22.h-z23.h }, z1.h\n" + ".inst 0xc16016c9 // sdot za.s[x8, 1], { z22.h-z23.h }, z0.h\n" + ".inst 0xa1402a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20]\n" "add x14, x14, %x[ld_in_col]\n" - ".inst 0xa0412aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc16115aa // sdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n" - ".inst 0xc16015ab // sdot za.s[x8, 3], { z13.h-z14.h }, z0.h\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" - ".inst 0xa0412a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - ".inst 0xc16315ca // sdot za.s[x8, 2], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215cb // sdot za.s[x8, 3], { z14.h-z15.h }, z2.h\n" + ".inst 0xa0412aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xc16d16ca // sdot za.s[x8, 2], { z22.h-z23.h }, z13.h\n" + ".inst 0xc16516cb // sdot za.s[x8, 3], { z22.h-z23.h }, z5.h\n" + ".inst 0xc16116e8 // sdot za.s[x8, 0], { z23.h-z24.h }, z1.h\n" + ".inst 0xc16016e9 // sdot za.s[x8, 1], { z23.h-z24.h }, z0.h\n" + ".inst 0xa0412a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + ".inst 0xc16116ea // sdot za.s[x8, 2], { z23.h-z24.h }, z1.h\n" + ".inst 0xc16016eb // sdot za.s[x8, 3], { z23.h-z24.h }, z0.h\n" "15:" // Padded: 0 priming loads - ".inst 0xa0402be0 // ld1h { z0.h-z1.h }, pn10.b/Z, [SP]\n" - ".inst 0xa0412be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" + ".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n" + ".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" "cbz x15, 18f\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "ld1b { z19.s }, p0/Z, [x14]\n" - "add z19.h, p0/M, z19.h, z24.h\n" + "add z19.h, p0/M, z19.h, z21.h\n" "add x20, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1b { z18.s }, p0/Z, [x20]\n" - "add z18.h, p0/M, z18.h, z24.h\n" + "add z18.h, p0/M, z18.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "ld1b { z17.s }, p0/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "ld1b { z16.s }, p0/Z, [x20]\n" - "add z16.h, p0/M, z16.h, z24.h\n" + "add z16.h, p0/M, z16.h, z21.h\n" "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row]\n" - "trn1 z13.h, z19.h, z18.h\n" - "trn1 z14.h, z17.h, z16.h\n" + "trn1 z6.h, z19.h, z18.h\n" + "trn1 z7.h, z17.h, z16.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "ld1b { z17.s }, p0/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" "ld1b { z16.s }, p0/Z, [x20]\n" - "add z16.h, p0/M, z16.h, z24.h\n" + "add z16.h, p0/M, z16.h, z21.h\n" "sub x15, x15, #0x1\n" "sub x13, x13, #0x1\n" "cmp x15, x13\n" - "trn1 z15.h, z17.h, z16.h\n" + "trn1 z8.h, z17.h, z16.h\n" "csel x23, x15, x13, LT\n" "add x14, x14, %x[ld_in_col]\n" "sub x13, x13, x23\n" @@ -503,121 +503,121 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( "16:" // Padded: Main loop "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z21.s }, p0/Z, [x14]\n" - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" - "add z21.h, p0/M, z21.h, z24.h\n" + "ld1b { z9.s }, p0/Z, [x14]\n" + ".inst 0xc16d14c8 // sdot za.s[x8, 0], { z6.h-z7.h }, z13.h\n" + ".inst 0xc16514c9 // sdot za.s[x8, 1], { z6.h-z7.h }, z5.h\n" + "add z9.h, p0/M, z9.h, z21.h\n" "add x22, x14, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z20.s }, p0/Z, [x22]\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - "add z20.h, p0/M, z20.h, z24.h\n" + "ld1b { z19.s }, p0/Z, [x22]\n" + ".inst 0xc16b14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z11.h\n" + "add z19.h, p0/M, z19.h, z21.h\n" "add x22, x22, %x[ld_in_row]\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" + ".inst 0xc16a14e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z10.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z19.s }, p0/Z, [x22]\n" - "add z19.h, p0/M, z19.h, z24.h\n" - ".inst 0xc0060804 // mova { z4.d-z5.d }, za.d[x8, #0]\n" + "ld1b { z18.s }, p0/Z, [x22]\n" + "add z18.h, p0/M, z18.h, z21.h\n" + ".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n" "add x22, x22, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x22]\n" - ".inst 0xc0060826 // mova { z6.d-z7.d }, za.d[x8, #1]\n" + "ld1b { z16.s }, p0/Z, [x22]\n" + ".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n" "mov x12, #0x4\n" "addvl x21, SP, #4\n" - "add z18.h, p0/M, z18.h, z24.h\n" - ".inst 0xc1aaac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n" + "add z16.h, p0/M, z16.h, z21.h\n" + ".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n" "add x22, x22, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - ".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n" + ".inst 0xa1402aa3 // ld1h { z3.h, z11.h }, pn10.b/Z, [x21]\n" "addvl x20, SP, #8\n" - ".inst 0xc16115aa // sdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n" + ".inst 0xc16b14ca // sdot za.s[x8, 2], { z6.h-z7.h }, z11.h\n" "subs x23, x23, #0x1\n" "ld1b { z17.s }, p0/Z, [x22]\n" - ".inst 0xc16015ab // sdot za.s[x8, 3], { z13.h-z14.h }, z0.h\n" - ".inst 0xc1abaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z11.s\n" - ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" - "add z17.h, p0/M, z17.h, z24.h\n" + ".inst 0xc16314cb // sdot za.s[x8, 3], { z6.h-z7.h }, z3.h\n" + ".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n" + ".inst 0xa1402a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20]\n" + "add z17.h, p0/M, z17.h, z21.h\n" "add x22, x22, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - ".inst 0xa0412aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc16115ac // sdot za.s[x8, 4], { z13.h-z14.h }, z1.h\n" - ".inst 0xc1acab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n" - "ld1b { z16.s }, p0/Z, [x22]\n" - ".inst 0xc16015ad // sdot za.s[x8, 5], { z13.h-z14.h }, z0.h\n" - "add z16.h, p0/M, z16.h, z24.h\n" + ".inst 0xa0412aaa // ld1h { z10.h-z11.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xc16d14cc // sdot za.s[x8, 4], { z6.h-z7.h }, z13.h\n" + ".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n" + "ld1b { z2.s }, p0/Z, [x22]\n" + ".inst 0xc16514cd // sdot za.s[x8, 5], { z6.h-z7.h }, z5.h\n" + "add z2.h, p0/M, z2.h, z21.h\n" "add x14, x14, %x[ld_in_col]\n" - ".inst 0xc16315ca // sdot za.s[x8, 2], { z14.h-z15.h }, z3.h\n" - ".inst 0xa0402be0 // ld1h { z0.h-z1.h }, pn10.b/Z, [SP]\n" - ".inst 0xc1bacec4 // sclamp { z4.s-z7.s }, z22.s, z26.s\n" - ".inst 0xc16215cb // sdot za.s[x8, 3], { z14.h-z15.h }, z2.h\n" - ".inst 0xa0412a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - "st1b { z4.s }, p1, [x11]\n" + ".inst 0xc16b14ea // sdot za.s[x8, 2], { z7.h-z8.h }, z11.h\n" + ".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n" + ".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n" + ".inst 0xc16a14eb // sdot za.s[x8, 3], { z7.h-z8.h }, z10.h\n" + ".inst 0xa1412a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + "st1b { z24.s }, p1, [x11]\n" "add x11, x11, x9\n" - ".inst 0xc16315cc // sdot za.s[x8, 4], { z14.h-z15.h }, z3.h\n" - "st1b { z6.s }, p1, [x10]\n" + ".inst 0xc16b14ec // sdot za.s[x8, 4], { z7.h-z8.h }, z11.h\n" + "st1b { z26.s }, p1, [x10]\n" "add x10, x10, x28\n" - "trn1 z13.h, z21.h, z20.h\n" - ".inst 0xc16215cd // sdot za.s[x8, 5], { z14.h-z15.h }, z2.h\n" + "trn1 z6.h, z9.h, z19.h\n" + ".inst 0xc16314ed // sdot za.s[x8, 5], { z7.h-z8.h }, z3.h\n" "add x8, x8, #0x2\n" - ".inst 0xa0412be2 // ld1h { z2.h-z3.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" - "st1b { z5.s }, p1, [x27]\n" + ".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n" + "st1b { z25.s }, p1, [x27]\n" "add x27, x27, x25\n" - "st1b { z7.s }, p1, [x26]\n" + "st1b { z27.s }, p1, [x26]\n" "add x26, x26, x24\n" - ".inst 0xc0040904 // mova za.d[x8, #4], { z8.d-z9.d }\n" - ".inst 0xc0040905 // mova za.d[x8, #5], { z8.d-z9.d }\n" - "trn1 z14.h, z19.h, z18.h\n" - "trn1 z15.h, z17.h, z16.h\n" + ".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n" + ".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n" + "trn1 z7.h, z18.h, z16.h\n" + "trn1 z8.h, z17.h, z2.h\n" "bgt 16b\n" "17:" // Main loop tail - ".inst 0xc16115a8 // sdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n" + ".inst 0xc16d14c8 // sdot za.s[x8, 0], { z6.h-z7.h }, z13.h\n" "addvl x21, SP, #4\n" "addvl x20, SP, #8\n" - ".inst 0xc16015a9 // sdot za.s[x8, 1], { z13.h-z14.h }, z0.h\n" + ".inst 0xc16514c9 // sdot za.s[x8, 1], { z6.h-z7.h }, z5.h\n" ".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n" - ".inst 0xc16315c8 // sdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215c9 // sdot za.s[x8, 1], { z14.h-z15.h }, z2.h\n" + ".inst 0xc16b14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z11.h\n" + ".inst 0xc16a14e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z10.h\n" ".inst 0xa0412aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n" - ".inst 0xc0060804 // mova { z4.d-z5.d }, za.d[x8, #0]\n" - ".inst 0xc0060826 // mova { z6.d-z7.d }, za.d[x8, #1]\n" - ".inst 0xc1aaac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n" - ".inst 0xc16115aa // sdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n" - ".inst 0xc1abaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z11.s\n" - ".inst 0xc16015ab // sdot za.s[x8, 3], { z13.h-z14.h }, z0.h\n" - ".inst 0xa0402a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20]\n" - ".inst 0xc1acab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n" - ".inst 0xc16115ac // sdot za.s[x8, 4], { z13.h-z14.h }, z1.h\n" - ".inst 0xc1bacec4 // sclamp { z4.s-z7.s }, z22.s, z26.s\n" - "st1b { z4.s }, p1, [x11]\n" + ".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n" + ".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n" + ".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n" + ".inst 0xc16114ca // sdot za.s[x8, 2], { z6.h-z7.h }, z1.h\n" + ".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n" + ".inst 0xc16014cb // sdot za.s[x8, 3], { z6.h-z7.h }, z0.h\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n" + ".inst 0xc16914cc // sdot za.s[x8, 4], { z6.h-z7.h }, z9.h\n" + ".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n" + "st1b { z24.s }, p1, [x11]\n" "add x11, x11, x9\n" - ".inst 0xc16015ad // sdot za.s[x8, 5], { z13.h-z14.h }, z0.h\n" - "st1b { z6.s }, p1, [x10]\n" + ".inst 0xc16114cd // sdot za.s[x8, 5], { z6.h-z7.h }, z1.h\n" + "st1b { z26.s }, p1, [x10]\n" "add x10, x10, x28\n" - ".inst 0xc16315ca // sdot za.s[x8, 2], { z14.h-z15.h }, z3.h\n" - "st1b { z5.s }, p1, [x27]\n" + ".inst 0xc16314ea // sdot za.s[x8, 2], { z7.h-z8.h }, z3.h\n" + "st1b { z25.s }, p1, [x27]\n" "add x27, x27, x25\n" - ".inst 0xc16215cb // sdot za.s[x8, 3], { z14.h-z15.h }, z2.h\n" - ".inst 0xa0412a82 // ld1h { z2.h-z3.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" - "st1b { z7.s }, p1, [x26]\n" + ".inst 0xc16214eb // sdot za.s[x8, 3], { z7.h-z8.h }, z2.h\n" + ".inst 0xa0412a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n" + "st1b { z27.s }, p1, [x26]\n" "add x26, x26, x24\n" - ".inst 0xc16315cc // sdot za.s[x8, 4], { z14.h-z15.h }, z3.h\n" - ".inst 0xc16215cd // sdot za.s[x8, 5], { z14.h-z15.h }, z2.h\n" + ".inst 0xc16114ec // sdot za.s[x8, 4], { z7.h-z8.h }, z1.h\n" + ".inst 0xc16014ed // sdot za.s[x8, 5], { z7.h-z8.h }, z0.h\n" "add x8, x8, #0x2\n" - ".inst 0xc0040904 // mova za.d[x8, #4], { z8.d-z9.d }\n" - ".inst 0xc0040905 // mova za.d[x8, #5], { z8.d-z9.d }\n" + ".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n" + ".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n" "18:" // Main loop skip tail "cbz x13, 20f\n" "19:" // Right padding loop ".inst 0xc0060804 // mova { z4.d-z5.d }, za.d[x8, #0]\n" "subs x13, x13, #0x1\n" ".inst 0xc0060826 // mova { z6.d-z7.d }, za.d[x8, #1]\n" - ".inst 0xc1aaac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n" + ".inst 0xc1aeac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n" "add x8, x8, #0x2\n" - ".inst 0xc1abaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z11.s\n" - ".inst 0xc0040904 // mova za.d[x8, #4], { z8.d-z9.d }\n" - ".inst 0xc1acab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n" - ".inst 0xc0040905 // mova za.d[x8, #5], { z8.d-z9.d }\n" - ".inst 0xc1bacec4 // sclamp { z4.s-z7.s }, z22.s, z26.s\n" + ".inst 0xc1acaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n" + ".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n" + ".inst 0xc1afab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z15.s\n" + ".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n" + ".inst 0xc1bccfa4 // sclamp { z4.s-z7.s }, z29.s, z28.s\n" "st1b { z4.s }, p1, [x11]\n" "add x11, x11, x9\n" "st1b { z6.s }, p1, [x10]\n" @@ -628,15 +628,15 @@ void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl( "add x26, x26, x24\n" "bgt 19b\n" "20:" // End - "ldr x22, [%x[args], %[offsetof_Args_weights]]\n" - "incw x22, ALL, MUL #9\n" - "str x22, [%x[args], %[offsetof_Args_weights]]\n" + "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" + "incw x20, ALL, MUL #9\n" + "str x20, [%x[args], %[offsetof_Args_weights]]\n" "incw x16\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_in_vl]]\n" + "ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n" "whilelt p1.s, x16, x17\n" - "ldr x14, [%x[args], %[offsetof_Args_inptr]]\n" - "add x14, x14, x20\n" - "str x14, [%x[args], %[offsetof_Args_inptr]]\n" + "ldr x20, [%x[args], %[offsetof_Args_inptr]]\n" + "add x20, x20, x21\n" + "str x20, [%x[args], %[offsetof_Args_inptr]]\n" "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" "ldr x24, [%x[args], %[offsetof_Args_ld_out_vls]]\n" "ldp x23, x22, [x25, #0x0]\n" -- cgit v1.2.1