// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/4x2-aarch64-neonfma-ld64.S.in
//   Generator: tools/xngen
//
// Copyright 2019 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_minmax_ukernel_4x2__aarch64_neonfma_ld64(
#     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] -> x14
#     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 pointers
# x3  a0
# x11 a1
# x12 a2
# x4  a3 / a_stride

# C pointers
# x6  c0
# x9  c1
# x10 c2
# x7  c3 / cm_stride

# Vector register usage
# A0  v0
# A1  v1
# A2  v2
# A3  v3
# B  v20 v21
# C  v24 v25
# C  v26 v27
# C  v28 v29
# C  v30 v31
# Clamp v4 v5

BEGIN_FUNCTION xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_ld64

        # Load cn_stride, params pointer
        LDP     x14, x8, [sp]

        # Clamp A and C pointers
        CMP     x0, 2                   // if mr < 2
        ADD     x11, x3, x4             // a1 = a0 + a_stride
        ADD     x9, x6, x7              // c1 = c0 + cm_stride
        CSEL    x11, x3, x11, LO        //   a1 = a0
        CSEL    x9, x6, x9, LO          //   c1 = c0

        # Load min/max values
        LD2R    {v4.2s, v5.2s}, [x8]

        ADD     x12, x11, x4            // a2 = a1 + a_stride
        ADD     x10, x9, x7             // c2 = c1 + cm_stride
                                        // if mr <= 2
        CSEL    x12, x11, x12, LS       //   a2 = a1
        CSEL    x10, x9, x10, LS        //   c2 = c1

        CMP     x0, 4                   // if mr < 4
        ADD     x4, x12, x4             // a3 = a2 + a_stride
        ADD     x7, x10, x7             // c3 = c2 + cm_stride
        CSEL    x4, x12, x4, LO         //   a3 = a2
        CSEL    x7, x10, x7, LO         //   c3 = c2

0:
        # Load initial bias from w into accumulators
        LDR     d24, [x5], 8
        MOV     v26.8b, v24.8b
        MOV     v28.8b, v24.8b
        MOV     v30.8b, v24.8b
        MOVI    v25.2s, 0
        MOVI    v27.2s, 0
        MOVI    v29.2s, 0
        MOVI    v31.2s, 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     d20, d21, [x5], 16
        LDR     d1, [x11], 8
        LDR     d2, [x12], 8
        LDR     d3,  [x4], 8
        SUBS    x0, x0, 8
        FMLA    v24.2s, v20.2s, v0.s[0]
        FMLA    v26.2s, v20.2s, v1.s[0]
        FMLA    v28.2s, v20.2s, v2.s[0]
        FMLA    v30.2s, v20.2s, v3.s[0]
        FMLA    v25.2s, v21.2s, v0.s[1]
        FMLA    v27.2s, v21.2s, v1.s[1]
        FMLA    v29.2s, v21.2s, v2.s[1]
        FMLA    v31.2s, v21.2s, v3.s[1]
        B.HS    1b

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

2:
        FADD    v24.2s, v24.2s, v25.2s
        FADD    v26.2s, v26.2s, v27.2s
        FADD    v28.2s, v28.2s, v29.2s
        FADD    v30.2s, v30.2s, v31.2s

        # Clamp
        FMAX    v24.2s, v24.2s, v4.2s
        SUBS    x1, x1, 2
        FMAX    v26.2s, v26.2s, v4.2s
        FMAX    v28.2s, v28.2s, v4.2s
        FMAX    v30.2s, v30.2s, v4.2s
        FMIN    v24.2s, v24.2s, v5.2s
        FMIN    v26.2s, v26.2s, v5.2s
        FMIN    v28.2s, v28.2s, v5.2s
        FMIN    v30.2s, v30.2s, v5.2s

        # Store full 4 x 2
        B.LO    4f

        ST1     {v24.8b},  [x6], x14
        SUB     x3,  x3, x2             // a0 -= kc
        ST1     {v26.8b},  [x9], x14
        SUB     x11, x11, x2            // a1 -= kc
        ST1     {v28.8b}, [x10], x14
        SUB     x12, x12, x2            // a2 -= kc
        ST1     {v30.8b},  [x7], x14
        SUB     x4,  x4, x2             // a3 -= kc

        B.HI    0b

        RET

        # Remainder- 1 float of A (4 bytes)
3:
        LDR     s0,  [x3], 4
        LDR     d20, [x5], 8
        LDR     s1, [x11], 4
        LDR     s2, [x12], 4
        LDR     s3,  [x4], 4
        SUBS    x0, x0, 4
        FMLA    v24.2s, v20.2s, v0.s[0]
        FMLA    v26.2s, v20.2s, v1.s[0]
        FMLA    v28.2s, v20.2s, v2.s[0]
        FMLA    v30.2s, v20.2s, v3.s[0]
        B       2b

        # Store odd width
4:
        STR     s24,  [x6]
        STR     s26,  [x9]
        STR     s28, [x10]
        STR     s30,  [x7]
7:
        RET

END_FUNCTION xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_ld64

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