// Auto-generated file. Do not edit!
//   Template: src/f16-gemm/6x16-aarch64-neonfp16arith-ld32.S.in
//   Generator: tools/xngen
//
// Copyright 2020 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_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32(
#     size_t mr,                x0
#     size_t nc,                x1
#     size_t kc,                x2 / x0
#     const uint8_t*restrict a, x3
#     size_t a_stride,          x4
#     const void*restrict w,    x5
#     uint8_t*restrict c,       x6
#     size_t cm_stride,         x7
#     size_t cn_stride,         [sp] -> (x0)
#     const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 8] -> x8

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

# A pointers
#  x3 a0
#  x9 a1
# x10 a2
# x11 a3
# x12 a4
#  x4 a5

# C pointers
#  x6 c0
# x16 c1
# x17 c2
# x14 c3
# x13 c4
#  x7 c5

# Vector register usage
# A0   v0
# A1   v1
# A2   v2
# A3   v3
# A4   v4
# A5   v5
# B   v16 v17 v18 v19
# C   v20 v21
# C   v22 v23
# C   v24 v25
# C   v26 v27
# C   v28 v29
# C   v30 v31
# Clamp v6, (v4), (v5)
# unused A   v8 v9 v10 v11
# unused B   v12 v13 v14 v15


BEGIN_FUNCTION xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32

        # Load params pointer
        LDR x8, [sp, 8]

        # Clamp A and C pointers
        CMP x0, 2                // if mr < 2
        ADD x9, x3, x4           // a1 = a0 + a_stride
        ADD x16, x6, x7          // c1 = c0 + cm_stride
        CSEL x9, x3, x9, LO      //   a1 = a0
        CSEL x16, x6, x16, LO    //   c1 = c0

        ADD x10, x9, x4          // a2 = a1 + a_stride
        ADD x17, x16, x7         // c2 = c1 + cm_stride
                                 // if mr <= 2
        CSEL x10, x9, x10, LS    //   a2 = a1
        CSEL x17, x16, x17, LS   //   c2 = c1

        CMP x0, 4                // if mr < 4
        ADD x11, x10, x4         // a3 = a2 + a_stride
        ADD x14, x17, x7         // c3 = c2 + cm_stride
        CSEL x11, x10, x11, LO   //   a3 = a2
        CSEL x14, x17, x14, LO   //   c3 = c2

        ADD x12, x11, x4         // a4 = a3 + a_stride
        ADD x13, x14, x7         // c4 = c3 + cm_stride
                                 // if mr <= 4
        CSEL x12, x11, x12, LS   //   a4 = a3
        CSEL x13, x14, x13, LS   //   c4 = c3

        CMP x0, 6                // if mr < 6
        ADD x4, x12, x4          // a5 = a4 + a_stride
        ADD x7, x13, x7          // c5 = c4 + cm_stride
        CSEL x4, x12, x4, LO     //   a5 = a4
        CSEL x7, x13, x7, LO     //   c5 = c4

        # Load params scale value
        LD1R {v6.8h}, [x8]
        ADD x8, x8, 2

0:
        # Load initial bias from w into accumulators
        LDP q20, q21, [x5], 32
        MOV v22.16b, v20.16b
        MOV v23.16b, v21.16b
        MOV v24.16b, v20.16b
        MOV v25.16b, v21.16b
        MOV v26.16b, v20.16b
        MOV v27.16b, v21.16b
        MOV v28.16b, v20.16b
        MOV v29.16b, v21.16b
        MOV v30.16b, v20.16b
        MOV v31.16b, v21.16b

         # Is there at least 2 halffloats (4 bytes)?
        SUBS x0, x2, 4  // k = kc - 4
        B.LO 3f

        # Main loop - 2 halffloats of A (4 bytes)
        # 24 FMA + 6 ld32 A + 4 LDR B
1:
        LDR   s0,  [x3], 4
        LDR  q16, [x5], 16
        LDR  q17, [x5], 16
        LDR   s1,  [x9], 4
        LDR   s2, [x10], 4
        LDR   s3, [x11], 4
        LDR   s4, [x12], 4
        LDR   s5,  [x4], 4
        SUBS x0, x0, 4
        FMLA v20.8h, v16.8h,  v0.h[0]
        FMLA v22.8h, v16.8h,  v1.h[0]
        FMLA v24.8h, v16.8h,  v2.h[0]
        FMLA v26.8h, v16.8h,  v3.h[0]
        LDR  q18, [x5], 16
        LDR  q19, [x5], 16
        FMLA v28.8h, v16.8h,  v4.h[0]
        FMLA v30.8h, v16.8h,  v5.h[0]
        FMLA v21.8h, v17.8h,  v0.h[0]
        FMLA v23.8h, v17.8h,  v1.h[0]
        FMLA v25.8h, v17.8h,  v2.h[0]
        FMLA v27.8h, v17.8h,  v3.h[0]
        FMLA v29.8h, v17.8h,  v4.h[0]
        FMLA v31.8h, v17.8h,  v5.h[0]

        FMLA v20.8h, v18.8h,  v0.h[1]
        FMLA v22.8h, v18.8h,  v1.h[1]
        FMLA v24.8h, v18.8h,  v2.h[1]
        FMLA v26.8h, v18.8h,  v3.h[1]
        FMLA v28.8h, v18.8h,  v4.h[1]
        FMLA v30.8h, v18.8h,  v5.h[1]
        FMLA v21.8h, v19.8h,  v0.h[1]
        FMLA v23.8h, v19.8h,  v1.h[1]
        FMLA v25.8h, v19.8h,  v2.h[1]
        FMLA v27.8h, v19.8h,  v3.h[1]
        FMLA v29.8h, v19.8h,  v4.h[1]
        FMLA v31.8h, v19.8h,  v5.h[1]
        B.HS 1b

        # Is there a remainder?- 1 halffloat of A (2 bytes)
        TBNZ x0, 1, 3f
2:
        # Scale and Clamp
        FMUL v20.8h, v20.8h, v6.8h
        # Load params values
        LD2R {v4.8h, v5.8h}, [x8]
        FMUL v21.8h, v21.8h, v6.8h
        FMUL v22.8h, v22.8h, v6.8h
        FMUL v23.8h, v23.8h, v6.8h
        FMUL v24.8h, v24.8h, v6.8h
        FMUL v25.8h, v25.8h, v6.8h
        FMUL v26.8h, v26.8h, v6.8h
        FMUL v27.8h, v27.8h, v6.8h
        FMUL v28.8h, v28.8h, v6.8h
        FMUL v29.8h, v29.8h, v6.8h
        FMUL v30.8h, v30.8h, v6.8h
        FMUL v31.8h, v31.8h, v6.8h
        # Load cn_stride
        LDR x0, [sp, 0]
        FMAX v20.8h, v20.8h, v4.8h
        FMAX v21.8h, v21.8h, v4.8h
        FMAX v22.8h, v22.8h, v4.8h
        FMAX v23.8h, v23.8h, v4.8h
        FMAX v24.8h, v24.8h, v4.8h
        FMAX v25.8h, v25.8h, v4.8h
        FMAX v26.8h, v26.8h, v4.8h
        FMAX v27.8h, v27.8h, v4.8h
        FMAX v28.8h, v28.8h, v4.8h
        FMAX v29.8h, v29.8h, v4.8h
        FMAX v30.8h, v30.8h, v4.8h
        FMAX v31.8h, v31.8h, v4.8h
        SUBS x1, x1, 16
        FMIN v20.8h, v20.8h, v5.8h
        FMIN v21.8h, v21.8h, v5.8h
        FMIN v22.8h, v22.8h, v5.8h
        FMIN v23.8h, v23.8h, v5.8h
        FMIN v24.8h, v24.8h, v5.8h
        FMIN v25.8h, v25.8h, v5.8h
        FMIN v26.8h, v26.8h, v5.8h
        FMIN v27.8h, v27.8h, v5.8h
        FMIN v28.8h, v28.8h, v5.8h
        FMIN v29.8h, v29.8h, v5.8h
        FMIN v30.8h, v30.8h, v5.8h
        FMIN v31.8h, v31.8h, v5.8h

        # Store full 6 x 16
        B.LO 4f

        ST1 {v20.16b, v21.16b},  [x6], x0
        SUB  x3,  x3, x2 // a0 -= kc
        ST1 {v22.16b, v23.16b}, [x16], x0
        SUB  x9,  x9, x2 // a1 -= kc
        ST1 {v24.16b, v25.16b}, [x17], x0
        SUB x10, x10, x2 // a2 -= kc
        ST1 {v26.16b, v27.16b}, [x14], x0
        SUB x11, x11, x2 // a3 -= kc
        ST1 {v28.16b, v29.16b}, [x13], x0
        SUB x12, x12, x2 // a4 -= kc
        ST1 {v30.16b, v31.16b},  [x7], x0
        SUB  x4,  x4, x2 // a5 -= kc

        B.HI 0b
        RET

3:
        # Remainder- 1 halffloat of A (2 bytes)
        LDR   h0,  [x3], 2
        LDR  q16, [x5], 16
        LDR  q17, [x5], 16
        LDR   h1,  [x9], 2
        LDR   h2, [x10], 2
        LDR   h3, [x11], 2
        LDR   h4, [x12], 2
        LDR   h5,  [x4], 2
        FMLA v20.8h, v16.8h,  v0.h[0]
        FMLA v22.8h, v16.8h,  v1.h[0]
        FMLA v24.8h, v16.8h,  v2.h[0]
        FMLA v26.8h, v16.8h,  v3.h[0]
        FMLA v28.8h, v16.8h,  v4.h[0]
        FMLA v30.8h, v16.8h,  v5.h[0]
        FMLA v21.8h, v17.8h,  v0.h[0]
        FMLA v23.8h, v17.8h,  v1.h[0]
        FMLA v25.8h, v17.8h,  v2.h[0]
        FMLA v27.8h, v17.8h,  v3.h[0]
        FMLA v29.8h, v17.8h,  v4.h[0]
        FMLA v31.8h, v17.8h,  v5.h[0]
        B 2b

        # Store odd width
4:
        TBZ x1, 3, 5f
        STR q20,  [x6], 16
        MOV v20.16b, v21.16b
        STR q22, [x16], 16
        MOV v22.16b, v23.16b
        STR q24, [x17], 16
        MOV v24.16b, v25.16b
        STR q26, [x14], 16
        MOV v26.16b, v27.16b
        STR q28, [x13], 16
        MOV v28.16b, v29.16b
        STR q30,  [x7], 16
        MOV v30.16b, v31.16b

5:
        TBZ x1, 2, 6f
        STR d20,  [x6], 8
        DUP d20, v20.d[1]
        STR d22, [x16], 8
        DUP d22, v22.d[1]
        STR d24, [x17], 8
        DUP d24, v24.d[1]
        STR d26, [x14], 8
        DUP d26, v26.d[1]
        STR d28, [x13], 8
        DUP d28, v28.d[1]
        STR d30,  [x7], 8
        DUP d30, v30.d[1]

6:
        TBZ x1, 1, 7f
        STR s20,  [x6], 4
        DUP s20, v20.s[1]
        STR s22, [x16], 4
        DUP s22, v22.s[1]
        STR s24, [x17], 4
        DUP s24, v24.s[1]
        STR s26, [x14], 4
        DUP s26, v26.s[1]
        STR s28, [x13], 4
        DUP s28, v28.s[1]
        STR s30,  [x7], 4
        DUP s30, v30.s[1]

7:
        TBZ x1, 0, 8f
        STR h20,  [x6]
        STR h22, [x16]
        STR h24, [x17]
        STR h26, [x14]
        STR h28, [x13]
        STR h30,  [x7]
8:
        RET

END_FUNCTION xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32

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