// 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${"inc" if INC else ""}_minmax_ukernel_1x8__asm_aarch64_neonfp16arith_ld64(
#     size_t mr,                (x0) - unused.  mr = 1
#     size_t nc,                x1
#     size_t kc,                x2 / x0
#     const void* restrict a,    x3
#     size_t a_stride,          (x4) - unused
#     const void* restrict w,    x5
#     void* restrict c,          x6
#     size_t cm_stride,         (x7) - unused
#     size_t cn_stride,         [sp] -> x14
$if INC:
  #     const float* restrict acc,  [sp + 8] -> x15
  #     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 16] -> (x8)
$else:
  #     const union xnn_f16_minmax_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.

# Register usage
# A0  x3  v0 v1
# B   x5 v20 v21 v22 v23
# C0  x6 v24 v26
# Clamp v4, v5


BEGIN_FUNCTION xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_1x8__asm_aarch64_neonfp16arith_ld64

        $if INC:
          # Load cn_stride, acc
          LDP         x14, x15, [sp]
          # Load params pointer
          LDR         x8, [sp, 16]
        $else:
          # Load cn_stride, params pointer
          LDP         x14, x8, [sp]

        # Load params values
        LD2R        {v4.8h, v5.8h}, [x8]
0:
        $if INC:
          # Load initial accumulators
          LDR         q24, [x15], 16
        $else:
          # Load initial bias from w into accumulators
          LDR         q24, [x5], 16

        MOVI        v26.8h, 0               // second set of C for pipelining FMLA

        # Is there at least 4 halffloats (8 bytes)
        SUBS        x0, x2, 8               // k = kc - 8

        B.LO        3f

        # Main loop - 4 halffloats of A (8 bytes)
1:
        LDR         d0, [x3], 8
        LDR         q20, [x5, 0]
        LDR         q21, [x5, 16]
        LDR         q22, [x5, 32]
        LDR         q23, [x5, 48]
        SUBS        x0, x0, 8
        FMLA        v24.8h, v20.8h, v0.h[0]
        FMLA        v26.8h, v21.8h, v0.h[1]
        FMLA        v24.8h, v22.8h, v0.h[2]
        FMLA        v26.8h, v23.8h, v0.h[3]
        ADD         x5, x5, 64
        B.HS        1b

        # Is there a remainder?- 2 halffloats of A (4 bytes)
        TBNZ        x0, 2, 4f
        # Is there a remainder?- 1 halffloat of A (2 bytes)
        TBNZ        x0, 1, 5f

2:
        FADD        v24.8h, v24.8h, v26.8h
        SUBS        x1, x1, 8

        # Clamp
        FMAX        v24.8h, v24.8h, v4.8h
        FMIN        v24.8h, v24.8h, v5.8h

        # Store full 1 x 8
        B.LO        6f

        ST1         {v24.16b},  [x6], x14
        SUB         x3,  x3, x2             // a0 -= kc

        B.HI        0b
        RET

3:
        TBZ         x0, 2, 5f
4:
        # Remainder- 2 halffloats of A (4 bytes)
        LDR         s0, [x3], 4
        LDR         q20, [x5], 16
        LDR         q21, [x5], 16
        FMLA        v24.8h, v20.8h, v0.h[0]
        FMLA        v26.8h, v21.8h, v0.h[1]
        TBZ         x0, 1, 2b

5:
        # Remainder- 1 halffloat of A (2 bytes)
        LDR         h0,  [x3], 2
        LDR         q20, [x5], 16
        FMLA        v24.8h, v20.8h, v0.h[0]
        B           2b

        # Store odd channels
6:
        TBZ         x1, 2, 7f
        STR         d24, [x6], 8
        DUP         d24, v24.d[1]

7:
        TBZ         x1, 1, 8f
        STR         s24, [x6], 4
        DUP         s24, v24.s[1]

8:
        TBZ         x1, 0, 9f
        STR         h24, [x6]
9:
        RET

END_FUNCTION xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_1x8__asm_aarch64_neonfp16arith_ld64

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