xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x8c8-minmax-fp32-aarch64-neon-mlal-prfm.S (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-igemm/1x8c8-aarch64-neon-mlal.S.in
//   Generator: tools/xngen
//
// Copyright 2021 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.


#include <xnnpack/assembly.h>

# void xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm(
#     size_t mr,                 x0
#     size_t nc,                 x1
#     size_t kc,                 x2 / x0
#     size_t ks,                 x3 / x9
#     const int8_t**restrict a,  x4
#     const int8_t* restrict w,  x5
#     int8_t* restrict c,        x6
#     size_t cm_stride,          (x7)
#     size_t cn_stride,                  [sp] -> x10
#     size_t a_offset,                   [sp + 8] -> x8
#     const int8_t* zero,                [sp + 16] -> x12
#     const union xnn_qs8_minmax_params params [sp + 24] -> x11

# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.

# Register usage
# A0 x13  v0  v6
# B   x5  v4  v5  v2  v3
# C0  x6 v16 v18 v20 v22 v24 v26 v28 v30
# temp0  v17 v19 v21 v23


BEGIN_FUNCTION xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm

        # Clamp C pointers
        LDP     x10, x8, [sp]           // Load cn_stride, a_offset
        ADD     x2, x2, 7               // kc = (kc + 7) & ~7
        LDP     x12, x11, [sp, 16]      // Load zero, params pointer
        BIC     x2, x2, 7

        .p2align 3
0:
        # Load initial bias from w into accumulators
        LDP     s16, s18, [x5], 8
        LDP     s20, s22, [x5], 8
        LDP     s24, s26, [x5], 8
        LDP     s28, s30, [x5], 8
        MOV     x9, x3                  // p = ks

        .p2align 3
1:
        # Load next A pointer
        LDR     x13, [x4], 8
        CMP     x13, x12                // if a0 == zero
        ADD     x13, x13, x8            // a0 += a_offset
        CSEL    x13, x12, x13, EQ       //   a0 = zero, else += a0 + a_offset

        # Is there at least 16 bytes for epilogue?
        SUBS    x0, x2, 16              // k = kc - 16
        B.LO    5f

        # Prologue: load A0 and 4 B's
        LDP     d0, d6, [x13], 16       // Read A0
        LDP     d4, d5, [x5]            // Read B
        LDP     d2, d3, [x5, 64]        // Read B

        # Is there at least 16 bytes for main loop?
        SUBS    x0, x0, 16              // k = k - 16
        B.LO    3f

        # Main loop - 16 bytes of A
        # 4 groups of 2 mul/mla/adap = 6 cycles.
        # 2 load for A0, A1 = +4 cycle.  Total 36 cycles.

        .p2align 3
2:
        # BLOCK 0 - 4 cycles
        SMULL   v17.8h, v4.8b, v0.8b
        SMULL   v19.8h, v5.8b, v0.8b
        LDP     d4, d5, [x5, 16]
        SMLAL   v17.8h, v2.8b, v6.8b
        SMLAL   v19.8h, v3.8b, v6.8b
        LDP     d2, d3, [x5, 80]

        # BLOCK 1 - 6 cycles
        SMULL   v21.8h, v4.8b, v0.8b
        SMULL   v23.8h, v5.8b, v0.8b
        PRFM    PLDL1KEEP, [x5, 448]
        SADALP  v16.4s, v17.8h
        PRFM    PLDL1KEEP, [x5, 512]
        SADALP  v18.4s, v19.8h
        LDP     d4, d5, [x5, 32]
        SMLAL   v21.8h, v2.8b, v6.8b
        SMLAL   v23.8h, v3.8b, v6.8b
        LDP     d2, d3, [x5, 96]

        # BLOCK 2 - 6 cycles
        SMULL   v17.8h, v4.8b, v0.8b
        SMULL   v19.8h, v5.8b, v0.8b
        PRFM    PLDL1KEEP, [x13, 128]
        SADALP  v20.4s, v21.8h
        SADALP  v22.4s, v23.8h
        LDP     d4, d5, [x5, 48]
        SMLAL   v17.8h, v2.8b, v6.8b
        SMLAL   v19.8h, v3.8b, v6.8b
        LDP     d2, d3, [x5, 112]

        # BLOCK 3 - 14 cycles
        SMULL   v21.8h, v4.8b, v0.8b
        ADD     x5, x5, 128
        SMULL   v23.8h, v5.8b, v0.8b
        SADALP  v24.4s, v17.8h
        SUBS    x0, x0, 16
        SADALP  v26.4s, v19.8h
        LDP     d4, d5, [x5]            // Read B
        SMLAL   v21.8h, v2.8b, v6.8b
        SMLAL   v23.8h, v3.8b, v6.8b
        LDP     d0, d6, [x13], 16       // Read A0
        SADALP  v28.4s, v21.8h
        LDP     d2, d3, [x5, 64]        // Read B
        SADALP  v30.4s, v23.8h
        B.HS    2b

        # Epilogue
        # Same as main loop except no loads at end of loop

        .p2align 3
3:
       # BLOCK 0 - 4 cycles
        SMULL   v17.8h, v4.8b, v0.8b
        SMULL   v19.8h, v5.8b, v0.8b
        LDP     d4, d5, [x5, 16]
        SMLAL   v17.8h, v2.8b, v6.8b
        SMLAL   v19.8h, v3.8b, v6.8b
        LDP     d2, d3, [x5, 80]

        # BLOCK 1 - 6 cycles
        SMULL   v21.8h, v4.8b, v0.8b
        SMULL   v23.8h, v5.8b, v0.8b
        PRFM    PLDL1KEEP, [x5, 448]
        SADALP  v16.4s, v17.8h
        PRFM    PLDL1KEEP, [x5, 512]
        SADALP  v18.4s, v19.8h
        LDP     d4, d5, [x5, 32]
        SMLAL   v21.8h, v2.8b, v6.8b
        SMLAL   v23.8h, v3.8b, v6.8b
        LDP     d2, d3, [x5, 96]

        # BLOCK 2 - 6 cycles
        SMULL   v17.8h, v4.8b, v0.8b
        SMULL   v19.8h, v5.8b, v0.8b
        PRFM    PLDL1KEEP, [x13, 128]
        SADALP  v20.4s, v21.8h
        SADALP  v22.4s, v23.8h
        LDP     d4, d5, [x5, 48]
        SMLAL   v17.8h, v2.8b, v6.8b
        SMLAL   v19.8h, v3.8b, v6.8b
        LDP     d2, d3, [x5, 112]

        # BLOCK 3 - 8 cycles
        SMULL   v21.8h, v4.8b, v0.8b
        ADD     x5, x5, 128
        SMULL   v23.8h, v5.8b, v0.8b
        SADALP  v24.4s, v17.8h
        SUBS    x0, x0, 16
        SADALP  v26.4s, v19.8h
        SMLAL   v21.8h, v2.8b, v6.8b
        SMLAL   v23.8h, v3.8b, v6.8b
        SADALP  v28.4s, v21.8h
        SADALP  v30.4s, v23.8h

        # Is there a remainder?- 8 bytes of A
        TBNZ    x0, 3, 5f

        # ks loop
        SUBS    x9, x9, 8               // ks -= MR * sizeof(int8_t*)
        B.HI    1b

4:
        # Add columns
        ADDP    v16.4s, v16.4s, v18.4s
        ADDP    v20.4s, v20.4s, v22.4s
        ADDP    v24.4s, v24.4s, v26.4s
        ADDP    v28.4s, v28.4s, v30.4s
        ADDP    v0.4s, v16.4s, v20.4s
        ADDP    v1.4s, v24.4s, v28.4s

        # Load per channel scale values from weights
        SCVTF   v0.4s, v0.4s
        LDR     q4, [x5], 16
        SCVTF   v1.4s, v1.4s
        LDR     q5, [x5], 16
        FMUL    v0.4s, v0.4s, v4.4s
        FMUL    v1.4s, v1.4s, v5.4s

        FCVTNS  v0.4s, v0.4s
        FCVTNS  v1.4s, v1.4s

        LD1R    {v5.8h}, [x11], 2
        SQXTN   v0.4h, v0.4s
        SQXTN2  v0.8h, v1.4s
        SUBS    x1, x1, 8
        SQADD   v0.8h, v0.8h, v5.8h
        LD1R    {v1.16b}, [x11], 1
        SQXTN   v0.8b, v0.8h
        LD1R    {v17.16b}, [x11]
        SMAX    v0.8b, v0.8b, v1.8b
        SUB     x11, x11, 3          // rewind params pointer
        SMIN    v0.8b, v0.8b, v17.8b
        B.LO    6f

        # Store full 1 x 8
        ST1     {v0.8b}, [x6], x10
        SUB     x4, x4, x3              // a -= ks
        B.HI    0b
        RET

        # Remainder - 8 bytes of A
        .p2align 3
5:
        LDR     d0, [x13], 8
        LDP     d4, d5, [x5]
        LDP     d6, d7, [x5, 16]
        SMULL   v17.8h, v4.8b, v0.8b
        SMULL   v19.8h, v5.8b, v0.8b
        SMULL   v21.8h, v6.8b, v0.8b
        SMULL   v23.8h, v7.8b, v0.8b
        LDP     d4, d5, [x5, 32]
        LDP     d6, d7, [x5, 48]
        SADALP  v16.4s, v17.8h
        SADALP  v18.4s, v19.8h
        SADALP  v20.4s, v21.8h
        SADALP  v22.4s, v23.8h
        SMULL   v17.8h, v4.8b, v0.8b
        SMULL   v19.8h, v5.8b, v0.8b
        SMULL   v21.8h, v6.8b, v0.8b
        SMULL   v23.8h, v7.8b, v0.8b
        ADD     x5, x5, 64
        SADALP  v24.4s, v17.8h
        SADALP  v26.4s, v19.8h
        SADALP  v28.4s, v21.8h
        SADALP  v30.4s, v23.8h

        # ks loop
        SUBS    x9, x9, 8               // ks -= MR * sizeof(int8_t*)
        B.HI    1b
        B       4b

        # Store odd width
        .p2align 3
6:
        TBZ     x1, 2, 7f
        STR     s0, [x6], 4
        EXT     v0.16b, v0.16b, v0.16b, 4

7:
        TBZ     x1, 1, 8f
        STR     h0, [x6], 2
        EXT     v0.16b, v0.16b, v0.16b, 2
8:
        TBZ     x1, 0, 9f
        STR     b0, [x6]
9:
        RET

END_FUNCTION xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm

#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif