// 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_f32_gemm${"inc" if INC else ""}_minmax_ukernel_1x8__aarch64_neonfma_ld64(
#     size_t mr,                (x0) - unused.  mr = 1
#     size_t nc,                x1
#     size_t kc,                x2 / x0
#     const uint8_t*restrict a, x3
#     size_t a_stride,          (x4) - unused
#     const void*restrict w,    x5
#     uint8_t*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_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])  [sp + 16] -> x8
$else:
  #     const union xnn_f32_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.

# A pointer
# x3  a0

# C pointer
# x6  c0

# Clamp v4 v5

BEGIN_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_1x8__aarch64_neonfma_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 min/max values
        LD2R {v4.4s, v5.4s}, [x8]
0:
        $if INC:
          # Load initial accumulators
          LDP q16, q17, [x15], 32
        $else:
          # Load initial bias from w into accumulators
          LDP q16, q17, [x5], 32

        MOVI v18.4s, 0  // second set of C for pipelining FMLA
        MOVI v19.4s, 0

        # Is there at least 2 floats (8 bytes)
        SUBS x0, x2, 8  // k = kc - 8

        B.LO 3f

        # Main loop - 2 floats of A (8 bytes)
1:
        LDR d0, [x3], 8
        LDP q20, q21, [x5], 32
        LDP q22, q23, [x5], 32
        FMLA v16.4s, v20.4s, v0.s[0]
        FMLA v17.4s, v21.4s, v0.s[0]
        SUBS x0, x0, 8
        FMLA v18.4s, v22.4s, v0.s[1]
        FMLA v19.4s, v23.4s, v0.s[1]
        B.HS 1b

        # Is there a remainder?- 1 floats of A (4 bytes)
        TBNZ x0, 2, 3f

2:
        FADD v16.4s, v16.4s, v18.4s
        FADD v17.4s, v17.4s, v19.4s
        SUBS x1, x1, 8

        # Clamp
        FMAX v16.4s, v16.4s, v4.4s
        FMAX v17.4s, v17.4s, v4.4s
        FMIN v16.4s, v16.4s, v5.4s
        FMIN v17.4s, v17.4s, v5.4s

        # Store full 1 x 8
        B.LO 4f

        STP q16, q17, [x6]
        ADD x6, x6, x14

        SUB  x3,  x3, x2 // a0 -= kc

        B.HI 0b

        RET

3:
        # Remainder- 1 float of A (4 bytes)
        LDP q20, q21, [x5], 32
        LDR s0, [x3], 4
        FMLA v16.4s, v20.4s, v0.s[0]
        FMLA v17.4s, v21.4s, v0.s[0]
        B 2b

        # Store odd channels
4:
        TBZ x1, 2, 5f
        STR q16, [x6], 16
        MOV v16.16b, v17.16b

5:
        TBZ x1, 1, 6f
        STR d16, [x6], 8
        DUP d16, v16.d[1]

6:
        TBZ x1, 0, 7f
        STR s16, [x6]
7:
        RET

END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_1x8__aarch64_neonfma_ld64

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