xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld128.S (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qu8-igemm/4x16c4-aarch64-neondot-ld128.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_qu8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128(
#     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] -> (x0)
#     size_t a_offset,           [sp + 8] -> x8
#     const int8_t* zero,        [sp + 16] -> x12
#     const union xnn_qu8_conv_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
# A1  x14  v1
# A2  x15  v2
# A3  x10  v3
# B    x5  v4  v5  v6  v7
# C0   x6 v16 v20 v24 v28
# C1  x16 v17 v21 v25 v29
# C2  x17 v18 v22 v26 v30
# C3   x7 v19 v23 v27 v31
# zero_point v8 v12 v13 v14 v15
# unused v9 v10 v11

BEGIN_FUNCTION xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128

        # Clamp C pointers
        CMP     x0, 2                   // if mr < 2
        LDR     x8, [sp, 8]             // Load a_offset
        ADD     x16, x6, x7             // c1 = c0 + cm_stride
        CSEL    x16, x6,  x16, LO       //   c1 = c0
        LDP     x12, x11, [sp, 16]      // Load zero pointer, params
        ADD     x2, x2, 3               // kc = (kc + 3) & ~3
        ADD     x17, x16, x7            // c2 = c1 + cm_stride
                                        // if mr <= 2
        # Save d8,d12-d15 on stack
        STR     d8,  [sp, -48]!
        CSEL    x17, x16, x17, LS       //   c2 = c1
        BIC     x2, x2, 3
        STP     d12, d13, [sp, 16]
        CMP     x0, 4                   // if mr < 4
        ADD     x7,  x17, x7            // c3 = c2 + cm_stride
        STP     d14, d15, [sp, 32]
        CSEL    x7,  x17, x7, LO        //   c3 = c2
        LD1R    {v8.4s}, [x11], 4       // kernel_zero_point

        .p2align 3
0:
        # Load initial bias from w into accumulators
        LDP     q16, q20, [x5], 32

        MOVI    v12.4s, 0
        MOVI    v13.4s, 0
        MOVI    v14.4s, 0
        MOVI    v15.4s, 0

        MOV     v17.16b, v16.16b
        MOV     v18.16b, v16.16b
        LDP     q24, q28, [x5], 32
        MOV     v19.16b, v16.16b
        MOV     v21.16b, v20.16b
        MOV     v22.16b, v20.16b
        MOV     v23.16b, v20.16b
        MOV     v25.16b, v24.16b
        MOV     v26.16b, v24.16b
        MOV     v27.16b, v24.16b
        MOV     v29.16b, v28.16b
        MOV     v30.16b, v28.16b
        MOV     v31.16b, v28.16b

        MOV     x9, x3                  // p = ks

        .p2align 3
1:
        # Load next 4 A pointers
        LDP     x13, x14, [x4], 16
        LDP     x15, x10, [x4], 16

        CMP     x13, x12                // if a0 == zero
        ADD     x13, x13, x8            // a0 += a_offset
        CSEL    x13, x12, x13, EQ       //   a0 = zero, else += a0 + a_offset
        CMP     x14, x12                // if a1 == zero
        ADD     x14, x14, x8            // a1 += a_offset
        CSEL    x14, x12, x14, EQ       //   a1 = zero, else += a1 + a_offset
        CMP     x15, x12                // if a2 == zero
        ADD     x15, x15, x8            // a2 += a_offset
        CSEL    x15, x12, x15, EQ       //   a2 = zero, else += a2 + a_offset
        CMP     x10, x12                // if a3 == zero
        ADD     x10, x10, x8            // a3 += a_offset
        CSEL    x10, x12, x10, EQ       //   a3 = zero, else += a3 + a_offset

        # Is there at least 16 bytes for main loop?
        SUBS    x0, x2, 16              // k = kc - 16
        B.LO    4f

        # Main loop - 16 bytes of A
        .p2align 3
2:
        LDR     q0, [x13], 16
        LDR     q4,  [x5], 16
        LDR     q1, [x14], 16
        LDR     q2, [x15], 16
        LDR     q3, [x10], 16
        LDR     q5,  [x5], 16

        UDOT    v12.4s, v8.16b,  v0.16b  // update zero point
        UDOT    v13.4s, v8.16b,  v1.16b
        UDOT    v14.4s, v8.16b,  v2.16b
        UDOT    v15.4s, v8.16b,  v3.16b

        UDOT    v16.4s, v4.16b,  v0.4b[0]
        UDOT    v17.4s, v4.16b,  v1.4b[0]
        LDP     q6, q7, [x5], 32
        UDOT    v18.4s, v4.16b,  v2.4b[0]
        UDOT    v19.4s, v4.16b,  v3.4b[0]
        UDOT    v20.4s, v5.16b,  v0.4b[0]
        UDOT    v21.4s, v5.16b,  v1.4b[0]
        UDOT    v22.4s, v5.16b,  v2.4b[0]
        UDOT    v23.4s, v5.16b,  v3.4b[0]
        UDOT    v24.4s, v6.16b, v0.4b[0]
        UDOT    v25.4s, v6.16b, v1.4b[0]
        LDP     q4, q5, [x5], 32
        UDOT    v26.4s, v6.16b, v2.4b[0]
        UDOT    v27.4s, v6.16b, v3.4b[0]
        UDOT    v28.4s, v7.16b, v0.4b[0]
        UDOT    v29.4s, v7.16b, v1.4b[0]
        UDOT    v30.4s, v7.16b, v2.4b[0]
        UDOT    v31.4s, v7.16b, v3.4b[0]

        UDOT    v16.4s, v4.16b,  v0.4b[1]
        UDOT    v17.4s, v4.16b,  v1.4b[1]
        LDP     q6, q7, [x5], 32
        UDOT    v18.4s, v4.16b,  v2.4b[1]
        UDOT    v19.4s, v4.16b,  v3.4b[1]
        UDOT    v20.4s, v5.16b,  v0.4b[1]
        UDOT    v21.4s, v5.16b,  v1.4b[1]
        UDOT    v22.4s, v5.16b,  v2.4b[1]
        UDOT    v23.4s, v5.16b,  v3.4b[1]
        UDOT    v24.4s, v6.16b,  v0.4b[1]
        UDOT    v25.4s, v6.16b,  v1.4b[1]
        LDP     q4, q5, [x5], 32
        UDOT    v26.4s, v6.16b,  v2.4b[1]
        UDOT    v27.4s, v6.16b,  v3.4b[1]
        UDOT    v28.4s, v7.16b,  v0.4b[1]
        UDOT    v29.4s, v7.16b,  v1.4b[1]
        UDOT    v30.4s, v7.16b,  v2.4b[1]
        UDOT    v31.4s, v7.16b,  v3.4b[1]

        UDOT    v16.4s, v4.16b,  v0.4b[2]
        UDOT    v17.4s, v4.16b,  v1.4b[2]
        LDP     q6, q7, [x5], 32
        UDOT    v18.4s, v4.16b,  v2.4b[2]
        UDOT    v19.4s, v4.16b,  v3.4b[2]
        UDOT    v20.4s, v5.16b,  v0.4b[2]
        UDOT    v21.4s, v5.16b,  v1.4b[2]
        UDOT    v22.4s, v5.16b,  v2.4b[2]
        UDOT    v23.4s, v5.16b,  v3.4b[2]
        UDOT    v24.4s, v6.16b,  v0.4b[2]
        UDOT    v25.4s, v6.16b,  v1.4b[2]
        LDP     q4, q5, [x5], 32
        UDOT    v26.4s, v6.16b,  v2.4b[2]
        UDOT    v27.4s, v6.16b,  v3.4b[2]
        UDOT    v28.4s, v7.16b,  v0.4b[2]
        UDOT    v29.4s, v7.16b,  v1.4b[2]
        UDOT    v30.4s, v7.16b,  v2.4b[2]
        UDOT    v31.4s, v7.16b,  v3.4b[2]

        UDOT    v16.4s, v4.16b,  v0.4b[3]
        UDOT    v17.4s, v4.16b,  v1.4b[3]
        LDP     q6, q7, [x5], 32
        UDOT    v18.4s, v4.16b,  v2.4b[3]
        UDOT    v19.4s, v4.16b,  v3.4b[3]
        UDOT    v20.4s, v5.16b,  v0.4b[3]
        UDOT    v21.4s, v5.16b,  v1.4b[3]
        UDOT    v22.4s, v5.16b,  v2.4b[3]
        UDOT    v23.4s, v5.16b,  v3.4b[3]
        UDOT    v24.4s, v6.16b,  v0.4b[3]
        UDOT    v25.4s, v6.16b,  v1.4b[3]
        UDOT    v26.4s, v6.16b,  v2.4b[3]
        UDOT    v27.4s, v6.16b,  v3.4b[3]
        SUBS    x0, x0, 16
        UDOT    v28.4s, v7.16b,  v0.4b[3]
        UDOT    v29.4s, v7.16b,  v1.4b[3]
        UDOT    v30.4s, v7.16b,  v2.4b[3]
        UDOT    v31.4s, v7.16b,  v3.4b[3]
        B.HS    2b

        # Is there a remainder?- 4 to 12 bytes of A
        TST     x0, 15
        B.NE    4f

3:
        # ks loop
        SUBS    x9, x9, 32              // ks -= MR * sizeof(int8_t*)
        B.HI    1b

        ADDP    v0.4s, v12.4s, v12.4s
        ADDP    v1.4s, v13.4s, v13.4s
        ADDP    v2.4s, v14.4s, v14.4s
        ADDP    v3.4s, v15.4s, v15.4s
        ADDP    v12.4s, v0.4s, v0.4s
        ADDP    v13.4s, v1.4s, v1.4s
        ADDP    v14.4s, v2.4s, v2.4s
        ADDP    v15.4s, v3.4s, v3.4s

        # Subtract zero point from accumulators
        SUB     v16.4s, v16.4s, v12.4s
        SUB     v17.4s, v17.4s, v13.4s
        SUB     v18.4s, v18.4s, v14.4s
        SUB     v19.4s, v19.4s, v15.4s
        SUB     v20.4s, v20.4s, v12.4s
        SUB     v21.4s, v21.4s, v13.4s
        SUB     v22.4s, v22.4s, v14.4s
        SUB     v23.4s, v23.4s, v15.4s
        SUB     v24.4s, v24.4s, v12.4s
        SUB     v25.4s, v25.4s, v13.4s
        SUB     v26.4s, v26.4s, v14.4s
        SUB     v27.4s, v27.4s, v15.4s
        SUB     v28.4s, v28.4s, v12.4s
        SUB     v29.4s, v29.4s, v13.4s
        SUB     v30.4s, v30.4s, v14.4s
        SUB     v31.4s, v31.4s, v15.4s

        SCVTF   v16.4s, v16.4s
        SCVTF   v17.4s, v17.4s
        # Apply params - scale, bias and clamp
        LD1R    {v4.4s}, [x11], 4
        SCVTF   v18.4s, v18.4s
        SCVTF   v19.4s, v19.4s
        SCVTF   v20.4s, v20.4s
        SCVTF   v21.4s, v21.4s
        SCVTF   v22.4s, v22.4s
        SCVTF   v23.4s, v23.4s
        SCVTF   v24.4s, v24.4s
        SCVTF   v25.4s, v25.4s
        SCVTF   v26.4s, v26.4s
        SCVTF   v27.4s, v27.4s
        SCVTF   v28.4s, v28.4s
        SCVTF   v29.4s, v29.4s
        SCVTF   v30.4s, v30.4s
        SCVTF   v31.4s, v31.4s

        FMUL    v16.4s, v16.4s, v4.4s
        FMUL    v17.4s, v17.4s, v4.4s
        FMUL    v18.4s, v18.4s, v4.4s
        FMUL    v19.4s, v19.4s, v4.4s
        FMUL    v20.4s, v20.4s, v4.4s
        FMUL    v21.4s, v21.4s, v4.4s
        FMUL    v22.4s, v22.4s, v4.4s
        FMUL    v23.4s, v23.4s, v4.4s
        FMUL    v24.4s, v24.4s, v4.4s
        FMUL    v25.4s, v25.4s, v4.4s
        FMUL    v26.4s, v26.4s, v4.4s
        FMUL    v27.4s, v27.4s, v4.4s
        FMUL    v28.4s, v28.4s, v4.4s
        FMUL    v29.4s, v29.4s, v4.4s
        FMUL    v30.4s, v30.4s, v4.4s
        FMUL    v31.4s, v31.4s, v4.4s

        FCVTNS  v16.4s, v16.4s
        FCVTNS  v17.4s, v17.4s
        FCVTNS  v18.4s, v18.4s
        FCVTNS  v19.4s, v19.4s
        FCVTNS  v20.4s, v20.4s
        FCVTNS  v21.4s, v21.4s
        FCVTNS  v22.4s, v22.4s
        FCVTNS  v23.4s, v23.4s
        FCVTNS  v24.4s, v24.4s
        FCVTNS  v25.4s, v25.4s
        FCVTNS  v26.4s, v26.4s
        FCVTNS  v27.4s, v27.4s
        FCVTNS  v28.4s, v28.4s
        FCVTNS  v29.4s, v29.4s
        FCVTNS  v30.4s, v30.4s
        FCVTNS  v31.4s, v31.4s

        SQXTN   v16.4h, v16.4s
        SQXTN   v17.4h, v17.4s
        SQXTN   v18.4h, v18.4s
        SQXTN   v19.4h, v19.4s
        SQXTN   v24.4h, v24.4s
        SQXTN   v25.4h, v25.4s
        SQXTN   v26.4h, v26.4s
        SQXTN   v27.4h, v27.4s
        LD1R    {v6.8h}, [x11], 2        // add bias

        SQXTN2  v16.8h, v20.4s
        SQXTN2  v17.8h, v21.4s
        SQXTN2  v18.8h, v22.4s
        SQXTN2  v19.8h, v23.4s
        SQXTN2  v24.8h, v28.4s
        SQXTN2  v25.8h, v29.4s
        SQXTN2  v26.8h, v30.4s
        SQXTN2  v27.8h, v31.4s

        SQADD   v16.8h, v16.8h, v6.8h
        SQADD   v17.8h, v17.8h, v6.8h
        SQADD   v18.8h, v18.8h, v6.8h
        SQADD   v19.8h, v19.8h, v6.8h
        SQADD   v24.8h, v24.8h, v6.8h
        SQADD   v25.8h, v25.8h, v6.8h
        SQADD   v26.8h, v26.8h, v6.8h
        SQADD   v27.8h, v27.8h, v6.8h
        LD1R    {v4.16b}, [x11], 1      // clamp min value

        SQXTUN  v0.8b, v16.8h
        SQXTUN  v1.8b, v17.8h
        SQXTUN  v2.8b, v18.8h
        SQXTUN  v3.8b, v19.8h
        LD1R    {v5.16b}, [x11]         // clamp max value
        SQXTUN2 v0.16b, v24.8h
        SQXTUN2 v1.16b, v25.8h
        SQXTUN2 v2.16b, v26.8h
        SQXTUN2 v3.16b, v27.8h
        LDR     x0, [sp, 48]            // Load cn_stride

        UMAX    v0.16b, v0.16b, v4.16b
        UMAX    v1.16b, v1.16b, v4.16b
        SUB     x11, x11, 7             // rewind params pointer
        UMAX    v2.16b, v2.16b, v4.16b
        UMAX    v3.16b, v3.16b, v4.16b
        SUBS    x1, x1, 16
        UMIN    v0.16b, v0.16b, v5.16b
        UMIN    v1.16b, v1.16b, v5.16b
        UMIN    v2.16b, v2.16b, v5.16b
        UMIN    v3.16b, v3.16b, v5.16b
        B.LO    6f

        # Store full 4 x 16
        ST1     {v3.16b},  [x7], x0
        ST1     {v2.16b}, [x17], x0
        ST1     {v1.16b}, [x16], x0
        ST1     {v0.16b},  [x6], x0

        SUB     x4, x4, x3              // a -= ks

        # nc loop
        B.HI    0b

        # Restore d8,d12-d15 from stack
        LDP     d14, d15, [sp, 32]
        LDP     d12, d13, [sp, 16]
        LDR     d8,  [sp], 48
        RET

        # Remainder- 8 bytes of A
        .p2align 3
4:
        # Is there a remainder?- 8 bytes of A
        TBZ     x0, 3, 5f

        LDR     d0, [x13], 8
        LDR     q4,  [x5], 16
        LDR     d1, [x14], 8
        LDR     d2, [x15], 8
        LDR     d3, [x10], 8
        LDR     q5,  [x5], 16

        UDOT    v12.4s, v8.16b,  v0.16b  // update zero point
        UDOT    v13.4s, v8.16b,  v1.16b
        UDOT    v14.4s, v8.16b,  v2.16b
        UDOT    v15.4s, v8.16b,  v3.16b

        UDOT    v16.4s, v4.16b,  v0.4b[0]
        UDOT    v17.4s, v4.16b,  v1.4b[0]
        LDP     q6, q7, [x5], 32
        UDOT    v18.4s, v4.16b,  v2.4b[0]
        UDOT    v19.4s, v4.16b,  v3.4b[0]
        UDOT    v20.4s, v5.16b,  v0.4b[0]
        UDOT    v21.4s, v5.16b,  v1.4b[0]
        UDOT    v22.4s, v5.16b,  v2.4b[0]
        UDOT    v23.4s, v5.16b,  v3.4b[0]
        UDOT    v24.4s, v6.16b, v0.4b[0]
        UDOT    v25.4s, v6.16b, v1.4b[0]
        LDP     q4, q5, [x5], 32
        UDOT    v26.4s, v6.16b, v2.4b[0]
        UDOT    v27.4s, v6.16b, v3.4b[0]
        UDOT    v28.4s, v7.16b, v0.4b[0]
        UDOT    v29.4s, v7.16b, v1.4b[0]
        UDOT    v30.4s, v7.16b, v2.4b[0]
        UDOT    v31.4s, v7.16b, v3.4b[0]
        UDOT    v16.4s, v4.16b,  v0.4b[1]
        UDOT    v17.4s, v4.16b,  v1.4b[1]
        LDP     q6, q7, [x5], 32
        UDOT    v18.4s, v4.16b,  v2.4b[1]
        UDOT    v19.4s, v4.16b,  v3.4b[1]
        UDOT    v20.4s, v5.16b,  v0.4b[1]
        UDOT    v21.4s, v5.16b,  v1.4b[1]
        UDOT    v22.4s, v5.16b,  v2.4b[1]
        UDOT    v23.4s, v5.16b,  v3.4b[1]
        UDOT    v24.4s, v6.16b,  v0.4b[1]
        UDOT    v25.4s, v6.16b,  v1.4b[1]
        UDOT    v26.4s, v6.16b,  v2.4b[1]
        UDOT    v27.4s, v6.16b,  v3.4b[1]
        UDOT    v28.4s, v7.16b,  v0.4b[1]
        UDOT    v29.4s, v7.16b,  v1.4b[1]
        UDOT    v30.4s, v7.16b,  v2.4b[1]
        UDOT    v31.4s, v7.16b,  v3.4b[1]
        # Is there a remainder?- 4 bytes of A
        TBZ     x0, 2, 3b

        # Remainder- 4 bytes of A
5:
        LDR     s0, [x13], 4
        LDR     q4, [x5], 16
        LDR     s1, [x14], 4
        LDR     s2, [x15], 4
        LDR     s3, [x10], 4
        LDR     q5, [x5], 16

        UDOT    v12.4s, v8.16b,  v0.16b  // update zero point
        UDOT    v13.4s, v8.16b,  v1.16b
        UDOT    v14.4s, v8.16b,  v2.16b
        UDOT    v15.4s, v8.16b,  v3.16b

        UDOT    v16.4s, v4.16b,  v0.4b[0]
        UDOT    v17.4s, v4.16b,  v1.4b[0]
        UDOT    v18.4s, v4.16b,  v2.4b[0]
        UDOT    v19.4s, v4.16b,  v3.4b[0]
        LDP     q6, q7, [x5], 32
        UDOT    v20.4s, v5.16b,  v0.4b[0]
        UDOT    v21.4s, v5.16b,  v1.4b[0]
        UDOT    v22.4s, v5.16b,  v2.4b[0]
        UDOT    v23.4s, v5.16b,  v3.4b[0]
        UDOT    v24.4s, v6.16b, v0.4b[0]
        UDOT    v25.4s, v6.16b, v1.4b[0]
        UDOT    v26.4s, v6.16b, v2.4b[0]
        UDOT    v27.4s, v6.16b, v3.4b[0]
        UDOT    v28.4s, v7.16b, v0.4b[0]
        UDOT    v29.4s, v7.16b, v1.4b[0]
        UDOT    v30.4s, v7.16b, v2.4b[0]
        UDOT    v31.4s, v7.16b, v3.4b[0]
        B       3b

        # Store odd width
        .p2align 3
6:
        TBZ     x1, 3, 7f
        STR     d3, [x7], 8
        STR     d2, [x17], 8
        DUP     d3, v3.d[1]
        DUP     d2, v2.d[1]
        STR     d1, [x16], 8
        STR     d0, [x6], 8
        DUP     d1, v1.d[1]
        DUP     d0, v0.d[1]
7:
        TBZ     x1, 2, 8f
        STR     s3, [x7], 4
        STR     s2, [x17], 4
        DUP     s3, v3.s[1]
        DUP     s2, v2.s[1]
        STR     s1, [x16], 4
        STR     s0, [x6], 4
        DUP     s1, v1.s[1]
        DUP     s0, v0.s[1]
8:
        TBZ     x1, 1, 9f
        STR     h3, [x7], 2
        STR     h2, [x17], 2
        DUP     h3, v3.h[1]
        DUP     h2, v2.h[1]
        STR     h1, [x16], 2
        STR     h0, [x6], 2
        DUP     h1, v1.h[1]
        DUP     h0, v0.h[1]
9:
        TBZ     x1, 0, 10f
        STR     b3, [x7]
        STR     b2, [x17]
        STR     b1, [x16]
        STR     b0, [x6]
10:
        # Restore d8,d12-d15 from stack
        LDP     d14, d15, [sp, 32]
        LDP     d12, d13, [sp, 16]
        LDR     d8,  [sp], 48
        RET

END_FUNCTION xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128

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