diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp | 554 |
1 files changed, 277 insertions, 277 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp index be82e04613..412d786d8a 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -72,99 +72,99 @@ void sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za_impl( "ldr x7, [%x[args], %[offsetof_Args_pad_bottom]]\n" "mov x20, #0x6\n" ".inst 0xd503477f // SMSTART ZA\n" - "sub x20, x20, x7\n" "ldr x17, [%x[args], %[offsetof_Args_pad_top]]\n" "ptrue p2.b\n" - "ld1rw { z25.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" "ldr x16, [%x[args], %[offsetof_Args_n_channels]]\n" + "ld1rw { z23.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" + "sub x20, x20, x7\n" + "ldr x15, [%x[args], %[offsetof_Args_current_channel]]\n" + "ld1rw { z13.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" "whilelt p1.s, XZR, x16\n" "whilelt p9.s, XZR, x20\n" - "ld1rw { z13.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" "whilelt p8.s, XZR, x17\n" "eor p8.b, p2/Z, p8.b, p9.b\n" - "ldr x15, [%x[args], %[offsetof_Args_current_channel]]\n" "1:" // Channel loop "ldr x20, [%x[args], %[offsetof_Args_bias]]\n" - "fmov z26.s, #0x0\n" + "fmov z4.s, #0x0\n" "cbz x20, 2f\n" - "ld1w { z26.s }, p1/Z, [x20, x15, LSL #2]\n" + "ld1w { z4.s }, p1/Z, [x20, x15, LSL #2]\n" "2:" // Load bias: Done - "ldr x21, [%x[args], %[offsetof_Args_weights]]\n" - "mov x20, x21\n" - "fmov z6.s, #0x0\n" - "ld1w { z15.s }, p2/Z, [x20]\n" + "ldr x25, [%x[args], %[offsetof_Args_weights]]\n" + "fmov z1.s, #0x0\n" + "fmov z12.s, #0x0\n" + "mov x24, #0x6\n" + "ldr x14, [%x[args], %[offsetof_Args_input_cols]]\n" + "fmov z15.s, #0x0\n" + "add x20, x17, x7\n" + "lsl x23, %x[ld_in_row], #0x2\n" + "ldr x13, [%x[args], %[offsetof_Args_inptr]]\n" + "mov z5.d, z4.d\n" + "mov x8, #0x0\n" + "sub x24, x24, x20\n" + "mov x22, x25\n" + "incb x25\n" + "ldr x11, [%x[args], %[offsetof_Args_output_cols]]\n" + "ld1w { z31.s }, p2/Z, [x22]\n" + "incb x22, ALL, MUL #3\n" + "sub x20, x14, #0x1\n" + "ld1w { z21.s }, p2/Z, [x22]\n" + "incb x22, ALL, MUL #3\n" + "orr x21, x20, %x[ld_in_col], LSL #18\n" + "ld1w { z10.s }, p2/Z, [x22]\n" + "mov x20, x25\n" + "incb x25\n" + ".inst 0x648aabe1 // bfcvtnt z1.h, p2/M, z31.s\n" + ".inst 0x658aabe2 // bfcvt z2.h, p2/M, z31.s\n" + "ld1w { z8.s }, p2/Z, [x20]\n" "incb x20, ALL, MUL #3\n" - "incb x21\n" + ".inst 0x658aaaa3 // bfcvt z3.h, p2/M, z21.s\n" + "orr x21, x16, x21, LSL #20\n" + "madd x23, x23, x17, x13\n" + ".inst 0x658aa94e // bfcvt z14.h, p2/M, z10.s\n" + "lsl x21, x21, #0x2\n" + ".inst 0x658aa907 // bfcvt z7.h, p2/M, z8.s\n" + ".inst 0x648aa90c // bfcvtnt z12.h, p2/M, z8.s\n" + ".inst 0x648aaaa2 // bfcvtnt z2.h, p2/M, z21.s\n" "ld1w { z29.s }, p2/Z, [x20]\n" - ".inst 0x648aa9e6 // bfcvtnt z6.h, p2/M, z15.s\n" "incb x20, ALL, MUL #3\n" - "ld1w { z30.s }, p2/Z, [x20]\n" - "mov x20, x21\n" - ".inst 0x658aa9e5 // bfcvt z5.h, p2/M, z15.s\n" - "ld1w { z14.s }, p2/Z, [x20]\n" - ".inst 0x658aaba8 // bfcvt z8.h, p2/M, z29.s\n" - "fmov z11.s, #0x0\n" + ".inst 0x648aa943 // bfcvtnt z3.h, p2/M, z10.s\n" + "ld1w { z20.s }, p2/Z, [x20]\n" + "mov x20, x25\n" + "ld1w { z17.s }, p2/Z, [x20]\n" "incb x20, ALL, MUL #3\n" - ".inst 0x658aa9ca // bfcvt z10.h, p2/M, z14.s\n" - ".inst 0x648aaba5 // bfcvtnt z5.h, p2/M, z29.s\n" - "incb x21\n" - "ld1w { z24.s }, p2/Z, [x20]\n" - "incb x20, ALL, MUL #3\n" - ".inst 0x648aabc8 // bfcvtnt z8.h, p2/M, z30.s\n" - ".inst 0x658aabcc // bfcvt z12.h, p2/M, z30.s\n" - "ld1w { z28.s }, p2/Z, [x20]\n" - "mov x21, x21\n" - ".inst 0x648aa9cb // bfcvtnt z11.h, p2/M, z14.s\n" - "ld1w { z20.s }, p2/Z, [x21]\n" - "incb x21, ALL, MUL #3\n" - ".inst 0x648aab0a // bfcvtnt z10.h, p2/M, z24.s\n" - ".inst 0x658aab09 // bfcvt z9.h, p2/M, z24.s\n" - "ld1w { z15.s }, p2/Z, [x21]\n" - "ldr x14, [%x[args], %[offsetof_Args_input_cols]]\n" - "incb x21, ALL, MUL #3\n" - "fmov z14.s, #0x0\n" - ".inst 0x658aaa81 // bfcvt z1.h, p2/M, z20.s\n" - "ldr x13, [%x[args], %[offsetof_Args_inptr]]\n" - ".inst 0x658aa9e7 // bfcvt z7.h, p2/M, z15.s\n" - ".inst 0x648aab89 // bfcvtnt z9.h, p2/M, z28.s\n" - "sub x20, x14, #0x1\n" - "orr x23, x20, %x[ld_in_col], LSL #18\n" - ".inst 0x658aab84 // bfcvt z4.h, p2/M, z28.s\n" - "ld1w { z29.s }, p2/Z, [x21]\n" - "orr x23, x16, x23, LSL #20\n" - "mov x22, #0x6\n" - "add x21, x17, x7\n" - "lsl x20, %x[ld_in_row], #0x2\n" - "mov z27.d, z26.d\n" - ".inst 0x648aaa8e // bfcvtnt z14.h, p2/M, z20.s\n" - ".inst 0x648aa9e1 // bfcvtnt z1.h, p2/M, z15.s\n" ".inst 0x648aaba7 // bfcvtnt z7.h, p2/M, z29.s\n" - "mov x8, #0x0\n" - "ldr x11, [%x[args], %[offsetof_Args_output_cols]]\n" - ".inst 0x658aaba2 // bfcvt z2.h, p2/M, z29.s\n" - "lsl x23, x23, #0x2\n" - "sub x22, x22, x21\n" - "madd x20, x20, x17, x13\n" + ".inst 0x658aaba0 // bfcvt z0.h, p2/M, z29.s\n" + "ld1w { z25.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #3\n" + ".inst 0x658aaa86 // bfcvt z6.h, p2/M, z20.s\n" + ".inst 0x658aaa29 // bfcvt z9.h, p2/M, z17.s\n" + ".inst 0x648aaa2f // bfcvtnt z15.h, p2/M, z17.s\n" + ".inst 0x658aab28 // bfcvt z8.h, p2/M, z25.s\n" + ".inst 0x648aaa80 // bfcvtnt z0.h, p2/M, z20.s\n" + "ld1w { z10.s }, p2/Z, [x20]\n" + ".inst 0x648aab29 // bfcvtnt z9.h, p2/M, z25.s\n" + ".inst 0x648aa948 // bfcvtnt z8.h, p2/M, z10.s\n" + ".inst 0x658aa94a // bfcvt z10.h, p2/M, z10.s\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" + "subs x24, x24, #0x1\n" + ".inst 0xf8b54afc // rprfm pldstrm, x21, [x23]\n" + "add x23, x23, %x[ld_in_col], LSL #2\n" "bgt 3b\n" "ldr x22, [%x[args], %[offsetof_Args_outptrs]]\n" - "lsl x20, %x[ld_in_row], #0x2\n" - "msub x13, x17, x20, x13\n" - ".inst 0xc0040b40 // mova za.d[x8, #0], { z26.d-z27.d }\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" - ".inst 0xc0040b41 // mova za.d[x8, #1], { z26.d-z27.d }\n" + "lsl x21, %x[ld_in_row], #0x2\n" + ".inst 0xc0040880 // mova za.d[x8, #0], { z4.d-z5.d }\n" "mov x10, #0x2\n" + "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" + "msub x13, x17, x21, x13\n" + ".inst 0xc0040881 // mova za.d[x8, #1], { z4.d-z5.d }\n" + "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" + ".inst 0xc0040882 // mova za.d[x8, #2], { z4.d-z5.d }\n" "ldp x9, x28, [x22], #0x10\n" - ".inst 0xc0040b42 // mova za.d[x8, #2], { z26.d-z27.d }\n" + ".inst 0xc0040883 // mova za.d[x8, #3], { z4.d-z5.d }\n" "ldp x27, x26, [x20], #0x10\n" - ".inst 0xc0040b43 // mova za.d[x8, #3], { z26.d-z27.d }\n" - "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" - ".inst 0xc0040b44 // mova za.d[x8, #4], { z26.d-z27.d }\n" + ".inst 0xc0040884 // mova za.d[x8, #4], { z4.d-z5.d }\n" + ".inst 0xc0040885 // mova za.d[x8, #5], { z4.d-z5.d }\n" "ldp x25, x24, [x22], #0x10\n" - ".inst 0xc0040b45 // mova za.d[x8, #5], { z26.d-z27.d }\n" "ldp x23, x22, [x20], #0x10\n" "cbz x21, 5f\n" "cmp x21, x10\n" @@ -172,19 +172,19 @@ void sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za_impl( "sub x21, x21, x20\n" "sub x10, x10, x20\n" "cbz x21, 5f\n" - ".inst 0xc0060814 // mova { z20.d-z21.d }, za.d[x8, #0]\n" + ".inst 0xc006081c // mova { z28.d-z29.d }, za.d[x8, #0]\n" "sub x11, x11, x21\n" - ".inst 0xc0060836 // mova { z22.d-z23.d }, za.d[x8, #1]\n" - ".inst 0xc1adcb34 // fclamp { z20.s-z23.s }, z25.s, z13.s\n" + ".inst 0xc006083e // mova { z30.d-z31.d }, za.d[x8, #1]\n" + ".inst 0xc1adcafc // fclamp { z28.s-z31.s }, z23.s, z13.s\n" "4:" // Left padding "subs x21, x21, #0x1\n" - "st1w { z20.s }, p1, [x9]\n" + "st1w { z28.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - "st1w { z22.s }, p1, [x28]\n" + "st1w { z30.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - "st1w { z21.s }, p1, [x25]\n" + "st1w { z29.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - "st1w { z23.s }, p1, [x24]\n" + "st1w { z31.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" "bgt 4b\n" "5:" // Left padding: End @@ -196,124 +196,124 @@ void sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za_impl( "beq 7f\n" "6:" // Unpadded: 2 priming loads "add x20, x13, %x[ld_in_row], LSL #2\n" - "ld1w { z17.s }, p1/Z, [x13]\n" - ".inst 0x658aaa3e // bfcvt z30.h, p2/M, z17.s\n" + "ld1w { z21.s }, p1/Z, [x13]\n" "add x13, x13, %x[ld_in_col], LSL #2\n" - "ld1w { z28.s }, p1/Z, [x20]\n" + "ld1w { z19.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aab9e // bfcvtnt z30.h, p2/M, z28.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z24.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa1f // bfcvt z31.h, p2/M, z16.s\n" - "ld1w { z15.s }, p1/Z, [x20]\n" + "ld1w { z17.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aa9ff // bfcvtnt z31.h, p2/M, z15.s\n" - ".inst 0xc12513d0 // bfdot za.s[x8, 0], { z30.h-z31.h }, z5.h\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x658aaabb // bfcvt z27.h, p2/M, z21.s\n" + "ld1w { z22.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa00 // bfcvt z0.h, p2/M, z16.s\n" - ".inst 0xc12613d1 // bfdot za.s[x8, 1], { z30.h-z31.h }, z6.h\n" - "ld1w { z15.s }, p1/Z, [x20]\n" - ".inst 0x648aa9e0 // bfcvtnt z0.h, p2/M, z15.s\n" - ".inst 0xc12c13f0 // bfdot za.s[x8, 0], { z31.h-z0.h }, z12.h\n" - ".inst 0xc12813f1 // bfdot za.s[x8, 1], { z31.h-z0.h }, z8.h\n" + ".inst 0x658aab1c // bfcvt z28.h, p2/M, z24.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x648aaa7b // bfcvtnt z27.h, p2/M, z19.s\n" + ".inst 0x658aaadd // bfcvt z29.h, p2/M, z22.s\n" + ".inst 0x648aaa3c // bfcvtnt z28.h, p2/M, z17.s\n" + ".inst 0x648aaa1d // bfcvtnt z29.h, p2/M, z16.s\n" + ".inst 0xc1221370 // bfdot za.s[x8, 0], { z27.h-z28.h }, z2.h\n" + ".inst 0xc1211371 // bfdot za.s[x8, 1], { z27.h-z28.h }, z1.h\n" + ".inst 0xc12e1390 // bfdot za.s[x8, 0], { z28.h-z29.h }, z14.h\n" + ".inst 0xc1231391 // bfdot za.s[x8, 1], { z28.h-z29.h }, z3.h\n" "7:" // Unpadded: 1 priming loads "add x20, x13, %x[ld_in_row], LSL #2\n" - "ld1w { z31.s }, p1/Z, [x13]\n" - ".inst 0x658aabef // bfcvt z15.h, p2/M, z31.s\n" + "ld1w { z24.s }, p1/Z, [x13]\n" "add x13, x13, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z30.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" + "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" - "ld1w { z17.s }, p1/Z, [x20]\n" + "ld1w { z18.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x648aaa30 // bfcvtnt z16.h, p2/M, z17.s\n" - ".inst 0xc12a11f0 // bfdot za.s[x8, 0], { z15.h-z16.h }, z10.h\n" - "ld1w { z22.s }, p1/Z, [x20]\n" + ".inst 0x658aab13 // bfcvt z19.h, p2/M, z24.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaad1 // bfcvt z17.h, p2/M, z22.s\n" - ".inst 0xc12b11f1 // bfdot za.s[x8, 1], { z15.h-z16.h }, z11.h\n" - "ld1w { z18.s }, p1/Z, [x20]\n" - ".inst 0x648aaa51 // bfcvtnt z17.h, p2/M, z18.s\n" - ".inst 0xc12511f2 // bfdot za.s[x8, 2], { z15.h-z16.h }, z5.h\n" - ".inst 0xc12611f3 // bfdot za.s[x8, 3], { z15.h-z16.h }, z6.h\n" - ".inst 0xc1241210 // bfdot za.s[x8, 0], { z16.h-z17.h }, z4.h\n" - ".inst 0xc1291211 // bfdot za.s[x8, 1], { z16.h-z17.h }, z9.h\n" - ".inst 0xc12c1212 // bfdot za.s[x8, 2], { z16.h-z17.h }, z12.h\n" - ".inst 0xc1281213 // bfdot za.s[x8, 3], { z16.h-z17.h }, z8.h\n" + ".inst 0x658aaa94 // bfcvt z20.h, p2/M, z20.s\n" + "ld1w { z17.s }, p1/Z, [x20]\n" + ".inst 0x648aabd3 // bfcvtnt z19.h, p2/M, z30.s\n" + ".inst 0x658aaa15 // bfcvt z21.h, p2/M, z16.s\n" + ".inst 0x648aaa54 // bfcvtnt z20.h, p2/M, z18.s\n" + ".inst 0x648aaa35 // bfcvtnt z21.h, p2/M, z17.s\n" + ".inst 0xc1271270 // bfdot za.s[x8, 0], { z19.h-z20.h }, z7.h\n" + ".inst 0xc12c1271 // bfdot za.s[x8, 1], { z19.h-z20.h }, z12.h\n" + ".inst 0xc1221272 // bfdot za.s[x8, 2], { z19.h-z20.h }, z2.h\n" + ".inst 0xc1211273 // bfdot za.s[x8, 3], { z19.h-z20.h }, z1.h\n" + ".inst 0xc1261290 // bfdot za.s[x8, 0], { z20.h-z21.h }, z6.h\n" + ".inst 0xc1201291 // bfdot za.s[x8, 1], { z20.h-z21.h }, z0.h\n" + ".inst 0xc12e1292 // bfdot za.s[x8, 2], { z20.h-z21.h }, z14.h\n" + ".inst 0xc1231293 // bfdot za.s[x8, 3], { z20.h-z21.h }, z3.h\n" "8:" // Unpadded: 0 priming loads "cbz x14, 16f\n" "add x20, x13, %x[ld_in_row], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x13]\n" - ".inst 0x658aaa16 // bfcvt z22.h, p2/M, z16.s\n" + "ld1w { z17.s }, p1/Z, [x13]\n" "sub x14, x14, #0x1\n" "ld1w { z16.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" "sub x11, x11, #0x1\n" - ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n" - "ld1w { z0.s }, p1/Z, [x20]\n" + "ld1w { z21.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aa817 // bfcvt z23.h, p2/M, z0.s\n" "cmp x14, x11\n" - "ld1w { z24.s }, p1/Z, [x20]\n" + "ld1w { z27.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa3c // bfcvt z28.h, p2/M, z17.s\n" "csel x21, x14, x11, LT\n" - ".inst 0x648aab17 // bfcvtnt z23.h, p2/M, z24.s\n" - "ld1w { z0.s }, p1/Z, [x20]\n" + "ld1w { z20.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aa818 // bfcvt z24.h, p2/M, z0.s\n" "add x13, x13, %x[ld_in_col], LSL #2\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - ".inst 0x648aaa18 // bfcvtnt z24.h, p2/M, z16.s\n" + ".inst 0x658aaabd // bfcvt z29.h, p2/M, z21.s\n" + "ld1w { z17.s }, p1/Z, [x20]\n" "sub x11, x11, x21\n" + ".inst 0x658aaa9e // bfcvt z30.h, p2/M, z20.s\n" + ".inst 0x648aaa1c // bfcvtnt z28.h, p2/M, z16.s\n" + ".inst 0x648aab7d // bfcvtnt z29.h, p2/M, z27.s\n" + ".inst 0x648aaa3e // bfcvtnt z30.h, p2/M, z17.s\n" "cbz x21, 15f\n" "9:" // Unpadded: Main loop - ".inst 0xc12112d0 // bfdot za.s[x8, 0], { z22.h-z23.h }, z1.h\n" + ".inst 0xc1291390 // bfdot za.s[x8, 0], { z28.h-z29.h }, z9.h\n" "add x20, x13, %x[ld_in_row], LSL #2\n" - "ld1w { z0.s }, p1/Z, [x13]\n" + "ld1w { z19.s }, p1/Z, [x13]\n" "subs x21, x21, #0x1\n" - ".inst 0xc12e12d1 // bfdot za.s[x8, 1], { z22.h-z23.h }, z14.h\n" - "ld1w { z20.s }, p1/Z, [x20]\n" + ".inst 0xc12f1391 // bfdot za.s[x8, 1], { z28.h-z29.h }, z15.h\n" + "ld1w { z18.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" "add x13, x13, %x[ld_in_col], LSL #2\n" - "ld1w { z19.s }, p1/Z, [x20]\n" + "ld1w { z17.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc12212f0 // bfdot za.s[x8, 0], { z23.h-z24.h }, z2.h\n" - ".inst 0xc12712f1 // bfdot za.s[x8, 1], { z23.h-z24.h }, z7.h\n" - "ld1w { z18.s }, p1/Z, [x20]\n" + ".inst 0xc1271392 // bfdot za.s[x8, 2], { z28.h-z29.h }, z7.h\n" + ".inst 0xc12c1393 // bfdot za.s[x8, 3], { z28.h-z29.h }, z12.h\n" + "ld1w { z21.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc12a12d2 // bfdot za.s[x8, 2], { z22.h-z23.h }, z10.h\n" - "ld1w { z17.s }, p1/Z, [x20]\n" + ".inst 0xc12a13b0 // bfdot za.s[x8, 0], { z29.h-z30.h }, z10.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc12b12d3 // bfdot za.s[x8, 3], { z22.h-z23.h }, z11.h\n" - "ld1w { z28.s }, p1/Z, [x20]\n" - ".inst 0xc12512d4 // bfdot za.s[x8, 4], { z22.h-z23.h }, z5.h\n" - ".inst 0xc12612d5 // bfdot za.s[x8, 5], { z22.h-z23.h }, z6.h\n" - ".inst 0x658aa816 // bfcvt z22.h, p2/M, z0.s\n" - ".inst 0x648aaa96 // bfcvtnt z22.h, p2/M, z20.s\n" - ".inst 0xc12412f2 // bfdot za.s[x8, 2], { z23.h-z24.h }, z4.h\n" - ".inst 0xc12912f3 // bfdot za.s[x8, 3], { z23.h-z24.h }, z9.h\n" - ".inst 0xc12c12f4 // bfdot za.s[x8, 4], { z23.h-z24.h }, z12.h\n" - ".inst 0xc12812f5 // bfdot za.s[x8, 5], { z23.h-z24.h }, z8.h\n" - ".inst 0x658aaa77 // bfcvt z23.h, p2/M, z19.s\n" - ".inst 0x658aaa38 // bfcvt z24.h, p2/M, z17.s\n" + ".inst 0xc12813b1 // bfdot za.s[x8, 1], { z29.h-z30.h }, z8.h\n" + "ld1w { z20.s }, p1/Z, [x20]\n" + ".inst 0xc1221394 // bfdot za.s[x8, 4], { z28.h-z29.h }, z2.h\n" + ".inst 0xc1211395 // bfdot za.s[x8, 5], { z28.h-z29.h }, z1.h\n" + ".inst 0x658aaa7c // bfcvt z28.h, p2/M, z19.s\n" + ".inst 0xc12613b2 // bfdot za.s[x8, 2], { z29.h-z30.h }, z6.h\n" + ".inst 0xc12013b3 // bfdot za.s[x8, 3], { z29.h-z30.h }, z0.h\n" + ".inst 0xc12e13b4 // bfdot za.s[x8, 4], { z29.h-z30.h }, z14.h\n" + ".inst 0xc12313b5 // bfdot za.s[x8, 5], { z29.h-z30.h }, z3.h\n" + ".inst 0x658aaa3d // bfcvt z29.h, p2/M, z17.s\n" + ".inst 0x658aaa1e // bfcvt z30.h, p2/M, z16.s\n" ".inst 0xc0060810 // mova { z16.d-z17.d }, za.d[x8, #0]\n" - ".inst 0x648aaa57 // bfcvtnt z23.h, p2/M, z18.s\n" - ".inst 0x648aab98 // bfcvtnt z24.h, p2/M, z28.s\n" + ".inst 0x648aaa5c // bfcvtnt z28.h, p2/M, z18.s\n" ".inst 0xc0060832 // mova { z18.d-z19.d }, za.d[x8, #1]\n" "add x8, x8, #0x2\n" - ".inst 0xc1adcb30 // fclamp { z16.s-z19.s }, z25.s, z13.s\n" + ".inst 0xc0040884 // mova za.d[x8, #4], { z4.d-z5.d }\n" + ".inst 0xc0040885 // mova za.d[x8, #5], { z4.d-z5.d }\n" + ".inst 0x648aaabd // bfcvtnt z29.h, p2/M, z21.s\n" + ".inst 0x648aaa9e // bfcvtnt z30.h, p2/M, z20.s\n" + ".inst 0xc1adcaf0 // fclamp { z16.s-z19.s }, z23.s, z13.s\n" "st1w { z16.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" "st1w { z18.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc0040b44 // mova za.d[x8, #4], { z26.d-z27.d }\n" "st1w { z17.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc0040b45 // mova za.d[x8, #5], { z26.d-z27.d }\n" "st1w { z19.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" "bgt 9b\n" @@ -325,186 +325,186 @@ void sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za_impl( "beq 12f\n" "11:" // Padded: 2 priming loads "mov x12, #0x0\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x13]\n" - ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n" "add x20, x13, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z17.s }, p0/Z, [x13]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x658aaa15 // bfcvt z21.h, p2/M, z16.s\n" "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa33 // bfcvt z19.h, p2/M, z17.s\n" + "ld1w { z17.s }, p0/Z, [x20]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n" - "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc1251290 // bfdot za.s[x8, 0], { z20.h-z21.h }, z5.h\n" + "mov x12, #0x4\n" + ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x20]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z23.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaaf6 // bfcvt z22.h, p2/M, z23.s\n" + ".inst 0x658aaa34 // bfcvt z20.h, p2/M, z17.s\n" + "ld1w { z31.s }, p0/Z, [x20]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" + ".inst 0x658aabf5 // bfcvt z21.h, p2/M, z31.s\n" "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n" - ".inst 0xc1261291 // bfdot za.s[x8, 1], { z20.h-z21.h }, z6.h\n" - "add x13, x13, %x[ld_in_col], LSL #2\n" - ".inst 0xc12c12b0 // bfdot za.s[x8, 0], { z21.h-z22.h }, z12.h\n" - ".inst 0xc12812b1 // bfdot za.s[x8, 1], { z21.h-z22.h }, z8.h\n" + ".inst 0xc1221270 // bfdot za.s[x8, 0], { z19.h-z20.h }, z2.h\n" + ".inst 0xc1211271 // bfdot za.s[x8, 1], { z19.h-z20.h }, z1.h\n" + ".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n" + ".inst 0xc12e1290 // bfdot za.s[x8, 0], { z20.h-z21.h }, z14.h\n" + ".inst 0xc1231291 // bfdot za.s[x8, 1], { z20.h-z21.h }, z3.h\n" "12:" // Padded: 1 priming loads "mov x12, #0x0\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x13]\n" - ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n" "add x20, x13, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z17.s }, p0/Z, [x13]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n" "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa34 // bfcvt z20.h, p2/M, z17.s\n" + "ld1w { z17.s }, p0/Z, [x20]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" - "mov x12, #0x4\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc12a1270 // bfdot za.s[x8, 0], { z19.h-z20.h }, z10.h\n" + "mov x12, #0x4\n" + ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x20]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z15.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aa9f5 // bfcvt z21.h, p2/M, z15.s\n" + ".inst 0x658aaa35 // bfcvt z21.h, p2/M, z17.s\n" + "ld1w { z25.s }, p0/Z, [x20]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" ".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n" - ".inst 0xc12b1271 // bfdot za.s[x8, 1], { z19.h-z20.h }, z11.h\n" - ".inst 0xc1251272 // bfdot za.s[x8, 2], { z19.h-z20.h }, z5.h\n" - "add x13, x13, %x[ld_in_col], LSL #2\n" - ".inst 0xc1261273 // bfdot za.s[x8, 3], { z19.h-z20.h }, z6.h\n" - ".inst 0xc1241290 // bfdot za.s[x8, 0], { z20.h-z21.h }, z4.h\n" - ".inst 0xc1291291 // bfdot za.s[x8, 1], { z20.h-z21.h }, z9.h\n" - ".inst 0xc12c1292 // bfdot za.s[x8, 2], { z20.h-z21.h }, z12.h\n" - ".inst 0xc1281293 // bfdot za.s[x8, 3], { z20.h-z21.h }, z8.h\n" + ".inst 0x658aab36 // bfcvt z22.h, p2/M, z25.s\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0xc1271290 // bfdot za.s[x8, 0], { z20.h-z21.h }, z7.h\n" + ".inst 0xc12c1291 // bfdot za.s[x8, 1], { z20.h-z21.h }, z12.h\n" + ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n" + ".inst 0xc1221292 // bfdot za.s[x8, 2], { z20.h-z21.h }, z2.h\n" + ".inst 0xc1211293 // bfdot za.s[x8, 3], { z20.h-z21.h }, z1.h\n" + ".inst 0xc12612b0 // bfdot za.s[x8, 0], { z21.h-z22.h }, z6.h\n" + ".inst 0xc12012b1 // bfdot za.s[x8, 1], { z21.h-z22.h }, z0.h\n" + ".inst 0xc12e12b2 // bfdot za.s[x8, 2], { z21.h-z22.h }, z14.h\n" + ".inst 0xc12312b3 // bfdot za.s[x8, 3], { z21.h-z22.h }, z3.h\n" "13:" // Padded: 0 priming loads "cbz x14, 16f\n" "mov x12, #0x0\n" + "add x20, x13, %x[ld_in_row], LSL #2\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "sub x14, x14, #0x1\n" + "sub x11, x11, #0x1\n" + "cmp x14, x11\n" "ld1w { z16.s }, p0/Z, [x13]\n" - ".inst 0x658aaa16 // bfcvt z22.h, p2/M, z16.s\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 { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + "csel x21, x14, x11, LT\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + "sub x11, x11, x21\n" + "ld1w { z17.s }, p0/Z, [x20]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x658aaa17 // bfcvt z23.h, p2/M, z16.s\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0x658aaa1c // bfcvt z28.h, p2/M, z16.s\n" "ld1w { z16.s }, p0/Z, [x20]\n" - ".inst 0x648aaa17 // bfcvtnt z23.h, p2/M, z16.s\n" - "mov x12, #0x4\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" + "mov x12, #0x4\n" + ".inst 0x648aaa3c // bfcvtnt z28.h, p2/M, z17.s\n" + "ld1w { z17.s }, p0/Z, [x20]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0x658aaa18 // bfcvt z24.h, p2/M, z16.s\n" + ".inst 0x658aaa1d // bfcvt z29.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x20]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa3d // bfcvtnt z29.h, p2/M, z17.s\n" + ".inst 0x658aaa1e // bfcvt z30.h, p2/M, z16.s\n" "ld1w { z16.s }, p0/Z, [x20]\n" - "sub x14, x14, #0x1\n" - ".inst 0x648aaa18 // bfcvtnt z24.h, p2/M, z16.s\n" - "sub x11, x11, #0x1\n" - "cmp x14, x11\n" - "csel x21, x14, x11, LT\n" - "add x13, x13, %x[ld_in_col], LSL #2\n" - "sub x11, x11, x21\n" + ".inst 0x648aaa1e // bfcvtnt z30.h, p2/M, z16.s\n" "cbz x21, 15f\n" "14:" // Padded: Main loop "mov x12, #0x0\n" - ".inst 0xc12112d0 // bfdot za.s[x8, 0], { z22.h-z23.h }, z1.h\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1w { z20.s }, p0/Z, [x13]\n" - ".inst 0xc12e12d1 // bfdot za.s[x8, 1], { z22.h-z23.h }, z14.h\n" "add x20, x13, %x[ld_in_row], LSL #2\n" + ".inst 0xc1291390 // bfdot za.s[x8, 0], { z28.h-z29.h }, z9.h\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc12f1391 // bfdot za.s[x8, 1], { z28.h-z29.h }, z15.h\n" + "subs x21, x21, #0x1\n" + ".inst 0xc1271392 // bfdot za.s[x8, 2], { z28.h-z29.h }, z7.h\n" + ".inst 0xc12c1393 // bfdot za.s[x8, 3], { z28.h-z29.h }, z12.h\n" + "ld1w { z17.s }, p0/Z, [x13]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z19.s }, p0/Z, [x20]\n" - "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc12a13b0 // bfdot za.s[x8, 0], { z29.h-z30.h }, z10.h\n" + "add x13, x13, %x[ld_in_col], LSL #2\n" + ".inst 0xc12813b1 // bfdot za.s[x8, 1], { z29.h-z30.h }, z8.h\n" + ".inst 0xc1221394 // bfdot za.s[x8, 4], { z28.h-z29.h }, z2.h\n" + "ld1w { z16.s }, p0/Z, [x20]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1w { z17.s }, p0/Z, [x20]\n" - ".inst 0xc12212f0 // bfdot za.s[x8, 0], { z23.h-z24.h }, z2.h\n" "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc1211395 // bfdot za.s[x8, 5], { z28.h-z29.h }, z1.h\n" + ".inst 0xc12613b2 // bfdot za.s[x8, 2], { z29.h-z30.h }, z6.h\n" + ".inst 0x658aaa3c // bfcvt z28.h, p2/M, z17.s\n" + ".inst 0xc12013b3 // bfdot za.s[x8, 3], { z29.h-z30.h }, z0.h\n" + "ld1w { z17.s }, p0/Z, [x20]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - ".inst 0xc12712f1 // bfdot za.s[x8, 1], { z23.h-z24.h }, z7.h\n" - "ld1w { z18.s }, p0/Z, [x20]\n" - "mov x12, #0x4\n" - ".inst 0xc12a12d2 // bfdot za.s[x8, 2], { z22.h-z23.h }, z10.h\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc12b12d3 // bfdot za.s[x8, 3], { z22.h-z23.h }, z11.h\n" - ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc12e13b4 // bfdot za.s[x8, 4], { z29.h-z30.h }, z14.h\n" + "mov x12, #0x4\n" + ".inst 0xc12313b5 // bfdot za.s[x8, 5], { z29.h-z30.h }, z3.h\n" + ".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n" + ".inst 0x648aaa1c // bfcvtnt z28.h, p2/M, z16.s\n" "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" "add x20, x20, %x[ld_in_row], LSL #2\n" - ".inst 0xc12512d4 // bfdot za.s[x8, 4], { z22.h-z23.h }, z5.h\n" - ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1w { z15.s }, p0/Z, [x20]\n" - "subs x21, x21, #0x1\n" - ".inst 0xc12612d5 // bfdot za.s[x8, 5], { z22.h-z23.h }, z6.h\n" - ".inst 0x658aaa96 // bfcvt z22.h, p2/M, z20.s\n" - ".inst 0x648aaa76 // bfcvtnt z22.h, p2/M, z19.s\n" - "add x13, x13, %x[ld_in_col], LSL #2\n" - ".inst 0xc12412f2 // bfdot za.s[x8, 2], { z23.h-z24.h }, z4.h\n" - ".inst 0xc12912f3 // bfdot za.s[x8, 3], { z23.h-z24.h }, z9.h\n" - ".inst 0xc12c12f4 // bfdot za.s[x8, 4], { z23.h-z24.h }, z12.h\n" - ".inst 0xc12812f5 // bfdot za.s[x8, 5], { z23.h-z24.h }, z8.h\n" - ".inst 0x658aaa37 // bfcvt z23.h, p2/M, z17.s\n" - ".inst 0x658aaa18 // bfcvt z24.h, p2/M, z16.s\n" - ".inst 0xc0060810 // mova { z16.d-z17.d }, za.d[x8, #0]\n" - ".inst 0x648aaa57 // bfcvtnt z23.h, p2/M, z18.s\n" - ".inst 0x648aa9f8 // bfcvtnt z24.h, p2/M, z15.s\n" - ".inst 0xc0060832 // mova { z18.d-z19.d }, za.d[x8, #1]\n" + ".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n" + ".inst 0x658aaa3d // bfcvt z29.h, p2/M, z17.s\n" "add x8, x8, #0x2\n" - ".inst 0xc1adcb30 // fclamp { z16.s-z19.s }, z25.s, z13.s\n" - "st1w { z16.s }, p1, [x9]\n" + ".inst 0xc0040884 // mova za.d[x8, #4], { z4.d-z5.d }\n" + "ld1w { z31.s }, p0/Z, [x20]\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xc0040885 // mova za.d[x8, #5], { z4.d-z5.d }\n" + ".inst 0xc1adcaf8 // fclamp { z24.s-z27.s }, z23.s, z13.s\n" + ".inst 0x648aaa1d // bfcvtnt z29.h, p2/M, z16.s\n" + ".inst 0x658aabfe // bfcvt z30.h, p2/M, z31.s\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "st1w { z24.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - "st1w { z18.s }, p1, [x28]\n" + "st1w { z26.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc0040b44 // mova za.d[x8, #4], { z26.d-z27.d }\n" - "st1w { z17.s }, p1, [x25]\n" + "st1w { z25.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc0040b45 // mova za.d[x8, #5], { z26.d-z27.d }\n" - "st1w { z19.s }, p1, [x24]\n" + ".inst 0x648aaa1e // bfcvtnt z30.h, p2/M, z16.s\n" + "st1w { z27.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" "bgt 14b\n" "15:" // Main loop tail - ".inst 0xc12112d0 // bfdot za.s[x8, 0], { z22.h-z23.h }, z1.h\n" - ".inst 0xc12e12d1 // bfdot za.s[x8, 1], { z22.h-z23.h }, z14.h\n" - ".inst 0xc12212f0 // bfdot za.s[x8, 0], { z23.h-z24.h }, z2.h\n" - ".inst 0xc12712f1 // bfdot za.s[x8, 1], { z23.h-z24.h }, z7.h\n" - ".inst 0xc12a12d2 // bfdot za.s[x8, 2], { z22.h-z23.h }, z10.h\n" - ".inst 0xc12b12d3 // bfdot za.s[x8, 3], { z22.h-z23.h }, z11.h\n" - ".inst 0xc12512d4 // bfdot za.s[x8, 4], { z22.h-z23.h }, z5.h\n" - ".inst 0xc12612d5 // bfdot za.s[x8, 5], { z22.h-z23.h }, z6.h\n" - ".inst 0xc0060810 // mova { z16.d-z17.d }, za.d[x8, #0]\n" - ".inst 0xc0060832 // mova { z18.d-z19.d }, za.d[x8, #1]\n" - ".inst 0xc1adcb30 // fclamp { z16.s-z19.s }, z25.s, z13.s\n" - "st1w { z16.s }, p1, [x9]\n" + ".inst 0xc1291390 // bfdot za.s[x8, 0], { z28.h-z29.h }, z9.h\n" + ".inst 0xc12f1391 // bfdot za.s[x8, 1], { z28.h-z29.h }, z15.h\n" + ".inst 0xc1271392 // bfdot za.s[x8, 2], { z28.h-z29.h }, z7.h\n" + ".inst 0xc12c1393 // bfdot za.s[x8, 3], { z28.h-z29.h }, z12.h\n" + ".inst 0xc12a13b0 // bfdot za.s[x8, 0], { z29.h-z30.h }, z10.h\n" + ".inst 0xc12813b1 // bfdot za.s[x8, 1], { z29.h-z30.h }, z8.h\n" + ".inst 0xc1221394 // bfdot za.s[x8, 4], { z28.h-z29.h }, z2.h\n" + ".inst 0xc1211395 // bfdot za.s[x8, 5], { z28.h-z29.h }, z1.h\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 0xc12613b2 // bfdot za.s[x8, 2], { z29.h-z30.h }, z6.h\n" + ".inst 0xc12013b3 // bfdot za.s[x8, 3], { z29.h-z30.h }, z0.h\n" + ".inst 0xc12e13b4 // bfdot za.s[x8, 4], { z29.h-z30.h }, z14.h\n" + ".inst 0xc12313b5 // bfdot za.s[x8, 5], { z29.h-z30.h }, z3.h\n" + "add x8, x8, #0x2\n" + ".inst 0xc1adcaf8 // fclamp { z24.s-z27.s }, z23.s, z13.s\n" + ".inst 0xc0040884 // mova za.d[x8, #4], { z4.d-z5.d }\n" + ".inst 0xc0040885 // mova za.d[x8, #5], { z4.d-z5.d }\n" + "st1w { z24.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" - ".inst 0xc12412f2 // bfdot za.s[x8, 2], { z23.h-z24.h }, z4.h\n" - "st1w { z18.s }, p1, [x28]\n" + "st1w { z26.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc12912f3 // bfdot za.s[x8, 3], { z23.h-z24.h }, z9.h\n" - "st1w { z17.s }, p1, [x25]\n" + "st1w { z25.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc12c12f4 // bfdot za.s[x8, 4], { z23.h-z24.h }, z12.h\n" - "st1w { z19.s }, p1, [x24]\n" + "st1w { z27.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" - ".inst 0xc12812f5 // bfdot za.s[x8, 5], { z23.h-z24.h }, z8.h\n" - "add x8, x8, #0x2\n" - ".inst 0xc0040b44 // mova za.d[x8, #4], { z26.d-z27.d }\n" - ".inst 0xc0040b45 // mova za.d[x8, #5], { z26.d-z27.d }\n" "16:" // Main loop skip tail "cbz x11, 18f\n" "17:" // Right padding loop @@ -512,25 +512,25 @@ void sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za_impl( "subs x11, x11, #0x1\n" ".inst 0xc006083e // mova { z30.d-z31.d }, za.d[x8, #1]\n" "add x8, x8, #0x2\n" - ".inst 0xc1adcb3c // fclamp { z28.s-z31.s }, z25.s, z13.s\n" + ".inst 0xc0040884 // mova za.d[x8, #4], { z4.d-z5.d }\n" + ".inst 0xc0040885 // mova za.d[x8, #5], { z4.d-z5.d }\n" + ".inst 0xc1adcafc // fclamp { z28.s-z31.s }, z23.s, z13.s\n" "st1w { z28.s }, p1, [x9]\n" "add x9, x9, x27, LSL #2\n" "st1w { z30.s }, p1, [x28]\n" "add x28, x28, x26, LSL #2\n" - ".inst 0xc0040b44 // mova za.d[x8, #4], { z26.d-z27.d }\n" "st1w { z29.s }, p1, [x25]\n" "add x25, x25, x23, LSL #2\n" - ".inst 0xc0040b45 // mova za.d[x8, #5], { z26.d-z27.d }\n" "st1w { z31.s }, p1, [x24]\n" "add x24, x24, x22, LSL #2\n" "bgt 17b\n" "18:" // End "ldr x20, [%x[args], %[offsetof_Args_weights]]\n" + "incw x15\n" + "whilelt p1.s, x15, x16\n" "incb x20, ALL, MUL #9\n" "str x20, [%x[args], %[offsetof_Args_weights]]\n" - "incw x15\n" "ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n" - "whilelt p1.s, x15, x16\n" "ldr x20, [%x[args], %[offsetof_Args_inptr]]\n" "add x20, x20, x21, LSL #2\n" "str x20, [%x[args], %[offsetof_Args_inptr]]\n" |