// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/4x8-aarch64-neonfma-cortex-a55.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_ukernel_4x8__aarch64_neonfma_cortex_a55(
#     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 static 1])  [sp + 8] -> x8

# d8-d15 need to be preserved if used.
# x19-30 need to be preserved if used.

# A pointers
#  x3 a0
#  x9 a1
# x10 a2
# x11 a3

# C pointers
#  x6 c0
# x16 c1
# x17 c2
# x18 c3

# x4 temporary vector shadow register

# Vector register usage
# A0  v0     v3
# A1  v0[1]  v3[1]
# A2  v1     v4
# A3  v1[1]  v4[1]

# B   v12 v13 v14 v15 second set of B
# B   v16 v17 v18 v19 first set
# C   v20 v21
# C   v22 v23
# C   v24 v25
# C   v26 v27
# Clamp v6 v7

# unused A   v8 v9 v10 v11
# x12 a4
# x13 c4
#  x7 c5
# A4  v2     v5
# A5  v2[1]  v5[1]
# C   v28 v29
# C   v30 v31

BEGIN_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a55

        # 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 x18, x17, x7         // c3 = c2 + cm_stride
        CSEL x11, x10, x11, LO   //   a3 = a2
        CSEL x18, x17, x18, LO   //   c3 = c2

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

        # Load min/max values
        LD2R {v6.4s, v7.4s}, [x8]

        # Load cn_stride
        LDR x14, [sp]

        // Save d12-d15 on stack
        STP d12, d13, [sp, -32]!
        STP d14, d15, [sp, 16]

0:
        # Load initial bias from w into accumulators
        LDP q20, q21, [x5], 32
        MOV v22.16b, v20.16b
        PRFM PLDL1KEEP,  [x3,  0]    // Prefetch A
        PRFM PLDL1KEEP,  [x3, 64]
        MOV v23.16b, v21.16b
        PRFM PLDL1KEEP,  [x9,  0]
        PRFM PLDL1KEEP,  [x9, 64]
        MOV v24.16b, v20.16b
        PRFM PLDL1KEEP, [x10,  0]
        PRFM PLDL1KEEP, [x10, 64]
        MOV v25.16b, v21.16b
        PRFM PLDL1KEEP, [x11,  0]
        PRFM PLDL1KEEP, [x11, 64]
        MOV v26.16b, v20.16b
        PRFM PLDL1KEEP, [x5,   0]  // Prefetch B
        MOV v27.16b, v21.16b
        PRFM PLDL1KEEP, [x5,  64]
        PRFM PLDL1KEEP, [x5, 128]
        PRFM PLDL1KEEP, [x5, 192]

        # Is there at least 4 floats (16 bytes) for prologue + epilogue?
        SUBS x0, x2, 16  // k = kc - 16
        B.LO 5f

        # Prologue - First group loads, no FMA
        LDR   d0, [x3], 8              // a0
        LDP q16, q17, [x5], 32         // b
        LDR   d1, [x10], 8             // a2
        LD1  {v0.d}[1],  [x9], 8       // a1
        LD1  {v1.d}[1], [x11], 8       // a3
        SUBS x0, x0, 16
        LDR  q18, [x5], 16
        LDR  d19, [x5], 8
        LDR  x4, [x5], 8   // ins is in BLOCK 0

        # Is there at least 4 floats (16 bytes) for main loop?
        B.LO 2f

        # Main loop - 4 floats of A (16 bytes)
        # 32 FMA + 8 LD64 A + 8 LDR B
1:
        # First group of 16 FMA, Second group loads
        // BLOCK 0
        FMLA v20.4s, v16.4s,  v0.s[0]
        LDR   d3, [x3], 8              // a0
        FMLA v22.4s, v16.4s,  v0.s[2]
        INS v19.d[1], x4               // b from second group
        FMLA v24.4s, v16.4s,  v1.s[0]
        LDR  x4, [x9], 8               // a1

        // BLOCK 1
        FMLA v26.4s, v16.4s,  v1.s[2]
        LDR  d12, [x5]
        FMLA v21.4s, v17.4s,  v0.s[0]
        INS v3.d[1], x4                // a1 ins
        FMLA v23.4s, v17.4s,  v0.s[2]
        LDR  x4, [x5, 8]   // b

        // BLOCK 2
        FMLA v25.4s, v17.4s,  v1.s[0]
        LDR   d4, [x10], 8             // a2
        FMLA v27.4s, v17.4s,  v1.s[2]
        INS v12.d[1], x4  // b  ins
        FMLA v20.4s, v18.4s,  v0.s[1]
        LDR  x4, [x11], 8              // a3

        // BLOCK 3
        FMLA v22.4s, v18.4s,  v0.s[3]
        LDR  d13, [x5, 16]
        FMLA v24.4s, v18.4s,  v1.s[1]
        INS v4.d[1], x4                // a3 ins
        FMLA v26.4s, v18.4s,  v1.s[3]
        LDR  x4, [x5, 24]

        // BLOCK 4
        FMLA v21.4s, v19.4s,  v0.s[1]
        LDR  d14, [x5, 32]
        FMLA v23.4s, v19.4s,  v0.s[3]
        INS v13.d[1], x4  // b
        FMLA v25.4s, v19.4s,  v1.s[1]
        LDR  x4, [x5, 40]

        // BLOCK 5
        // NOPs to ensure 4 cycle LDR lands on next LDR
        FMLA v27.4s, v19.4s,  v1.s[3]
        LDR  d15, [x5, 48]
        NOP
        INS v14.d[1], x4  // b from previous
        SUBS x0, x0, 16
        LDR x4, [x5, 56]

        # Second group of 16 FMA, First group of loads
        // BLOCK 0
        FMLA v20.4s, v12.4s,  v3.s[0]
        LDR   d0, [x3], 8              // a0
        FMLA v22.4s, v12.4s,  v3.s[2]
        INS v15.d[1], x4  // b from previous
        FMLA v24.4s, v12.4s,  v4.s[0]
        LDR  x4, [x9], 8               // a1

        // BLOCK 1
        FMLA v26.4s, v12.4s,  v4.s[2]
        LDR  d16, [x5, 64]
        FMLA v21.4s, v13.4s,  v3.s[0]
        INS v0.d[1], x4                // a1 ins
        FMLA v23.4s, v13.4s,  v3.s[2]
        LDR  x4, [x5, 72]  // b

        // BLOCK 2
        FMLA v25.4s, v13.4s,  v4.s[0]
        LDR   d1, [x10], 8             // a2
        FMLA v27.4s, v13.4s,  v4.s[2]
        INS v16.d[1], x4  // b
        FMLA v20.4s, v14.4s,  v3.s[1]
        LDR  x4, [x11], 8              // a3

        // BLOCK 3
        FMLA v22.4s, v14.4s,  v3.s[3]
        LDR  d17, [x5, 80]
        FMLA v24.4s, v14.4s,  v4.s[1]
        INS v1.d[1], x4                // a3 ins
        FMLA v26.4s, v14.4s,  v4.s[3]
        LDR  x4, [x5, 88]

        // BLOCK 4
        FMLA v21.4s, v15.4s,  v3.s[1]
        LDR  d18, [x5, 96]
        FMLA v23.4s, v15.4s,  v3.s[3]
        INS v17.d[1], x4  // b
        FMLA v25.4s, v15.4s,  v4.s[1]
        LDR  x4, [x5, 104]

        // BLOCK 5
        // NOTE that block needs to be 4 cycles for LDR not to stall
        FMLA v27.4s, v15.4s,  v4.s[3]
        LDR  d19, [x5, 112]
        INS v18.d[1], x4
        LDR  x4, [x5, 120]
        ADD x5, x5, 128
        B.HS 1b

        # Epilogue - 4 floats of A (16 bytes)
        # 32 FMA + 8 LD64 A + 8 LDR B
2:
        # First group of 16 FMA, Second group loads
        // BLOCK 0
        FMLA v20.4s, v16.4s,  v0.s[0]
        LDR   d3, [x3], 8              // a0
        FMLA v22.4s, v16.4s,  v0.s[2]
        INS v19.d[1], x4               // b from second group
        FMLA v24.4s, v16.4s,  v1.s[0]
        LDR  x4, [x9], 8               // a1

        // BLOCK 1
        FMLA v26.4s, v16.4s,  v1.s[2]
        LDR  d12, [x5]
        FMLA v21.4s, v17.4s,  v0.s[0]
        INS v3.d[1], x4                // a1 ins
        FMLA v23.4s, v17.4s,  v0.s[2]
        LDR  x4, [x5, 8]   // b

        // BLOCK 2
        FMLA v25.4s, v17.4s,  v1.s[0]
        LDR   d4, [x10], 8             // a2
        FMLA v27.4s, v17.4s,  v1.s[2]
        INS v12.d[1], x4  // b  ins
        FMLA v20.4s, v18.4s,  v0.s[1]
        LDR  x4, [x11], 8              // a3

        // BLOCK 3
        FMLA v22.4s, v18.4s,  v0.s[3]
        LDR  d13, [x5, 16]
        FMLA v24.4s, v18.4s,  v1.s[1]
        INS v4.d[1], x4                // a3 ins
        FMLA v26.4s, v18.4s,  v1.s[3]
        LDR  x4, [x5, 24]

        // BLOCK 4
        FMLA v21.4s, v19.4s,  v0.s[1]
        LDR  d14, [x5, 32]
        FMLA v23.4s, v19.4s,  v0.s[3]
        INS v13.d[1], x4  // b
        FMLA v25.4s, v19.4s,  v1.s[1]
        LDR  x4, [x5, 40]

        // BLOCK 5
        // NOPs to ensure 4 cycle LDR lands on next LDR
        FMLA v27.4s, v19.4s,  v1.s[3]
        LDR  d15, [x5, 48]
        NOP // fma
        INS v14.d[1], x4
        NOP
        LDR x4, [x5, 56]

        # Second group of 16 FMA, no loads
        // BLOCK 0
        FMLA v20.4s, v12.4s,  v3.s[0]
        FMLA v22.4s, v12.4s,  v3.s[2]
        INS v15.d[1], x4  // b from previous
        FMLA v24.4s, v12.4s,  v4.s[0]

        // BLOCK 1
        FMLA v26.4s, v12.4s,  v4.s[2]
        FMLA v21.4s, v13.4s,  v3.s[0]
        FMLA v23.4s, v13.4s,  v3.s[2]

        // BLOCK 2
        FMLA v25.4s, v13.4s,  v4.s[0]
        FMLA v27.4s, v13.4s,  v4.s[2]
        FMLA v20.4s, v14.4s,  v3.s[1]

        // BLOCK 3
        FMLA v22.4s, v14.4s,  v3.s[3]
        FMLA v24.4s, v14.4s,  v4.s[1]
        FMLA v26.4s, v14.4s,  v4.s[3]
        TST x0, 15

        // BLOCK 4
        FMLA v21.4s, v15.4s,  v3.s[1]
        FMLA v23.4s, v15.4s,  v3.s[3]
        FMLA v25.4s, v15.4s,  v4.s[1]
        ADD x5, x5, 64

        // BLOCK 5
        FMLA v27.4s, v15.4s,  v4.s[3]

        # Is there a remainder?- 2 floats of A (8 bytes) or less
        B.NE 5f

4:
        # Clamp
        FMAX v20.4s, v20.4s, v6.4s
        SUBS x1, x1, 8
        FMAX v21.4s, v21.4s, v6.4s
        FMAX v22.4s, v22.4s, v6.4s
        FMAX v23.4s, v23.4s, v6.4s
        FMAX v24.4s, v24.4s, v6.4s
        FMAX v25.4s, v25.4s, v6.4s
        FMAX v26.4s, v26.4s, v6.4s
        FMAX v27.4s, v27.4s, v6.4s
        FMIN v20.4s, v20.4s, v7.4s
        FMIN v21.4s, v21.4s, v7.4s
        FMIN v22.4s, v22.4s, v7.4s
        FMIN v23.4s, v23.4s, v7.4s
        FMIN v24.4s, v24.4s, v7.4s
        FMIN v25.4s, v25.4s, v7.4s
        FMIN v26.4s, v26.4s, v7.4s
        FMIN v27.4s, v27.4s, v7.4s

        # Store full 4 x 8
        B.LO 8f

        ST1 {v20.16b, v21.16b},  [x6], x14
        SUB  x3,  x3, x2 // a0 -= kc
        ST1 {v22.16b, v23.16b}, [x16], x14
        SUB  x9,  x9, x2 // a1 -= kc
        ST1 {v24.16b, v25.16b}, [x17], x14
        SUB x10, x10, x2 // a2 -= kc
        ST1 {v26.16b, v27.16b}, [x18], x14
        SUB x11, x11, x2 // a3 -= kc

        B.HI 0b

        // Restore d12-d15 from stack
        LDP d14, d15, [sp, 16]
        LDP d12, d13, [sp], 32
        RET

5:
        # Is there a remainder?- 2 floats of A (8 bytes)
        TBZ x0, 3, 6f

        # Remainder- 2 floats of A (8 bytes)
        LDR   d0,  [x3], 8
        LDR  q16, [x5], 16
        LD1   {v0.d}[1], [x9], 8
        LDR   d1, [x10], 8
        LD1   {v1.d}[1], [x11], 8
        LDR  q17, [x5], 16
        LDR  q18, [x5], 16
        LDR  q19, [x5], 16
        FMLA v20.4s, v16.4s,  v0.s[0]
        FMLA v22.4s, v16.4s,  v0.s[2]
        FMLA v24.4s, v16.4s,  v1.s[0]
        FMLA v26.4s, v16.4s,  v1.s[2]
        FMLA v21.4s, v17.4s,  v0.s[0]
        FMLA v23.4s, v17.4s,  v0.s[2]
        FMLA v25.4s, v17.4s,  v1.s[0]
        FMLA v27.4s, v17.4s,  v1.s[2]

        FMLA v20.4s, v18.4s,  v0.s[1]
        FMLA v22.4s, v18.4s,  v0.s[3]
        FMLA v24.4s, v18.4s,  v1.s[1]
        FMLA v26.4s, v18.4s,  v1.s[3]
        FMLA v21.4s, v19.4s,  v0.s[1]
        FMLA v23.4s, v19.4s,  v0.s[3]
        FMLA v25.4s, v19.4s,  v1.s[1]
        FMLA v27.4s, v19.4s,  v1.s[3]

        # Is there a remainder?- 1 floats of A (4 bytes)
        TBZ x0, 2, 4b

6:
        # Remainder- 1 floats of A (4 bytes)
        LDR   s0,  [x3], 4
        LDR  q16, [x5], 16
        LD1   {v0.s}[2], [x9], 4
        LDR   s1, [x10], 4
        LD1   {v1.s}[2], [x11], 4
        LDR  q17, [x5], 16

        FMLA v20.4s, v16.4s,  v0.s[0]
        FMLA v22.4s, v16.4s,  v0.s[2]
        FMLA v24.4s, v16.4s,  v1.s[0]
        FMLA v26.4s, v16.4s,  v1.s[2]
        FMLA v21.4s, v17.4s,  v0.s[0]
        FMLA v23.4s, v17.4s,  v0.s[2]
        FMLA v25.4s, v17.4s,  v1.s[0]
        FMLA v27.4s, v17.4s,  v1.s[2]
        B 4b

        # Store odd width
8:
        TBZ x1, 2, 9f
        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, [x18], 16
        MOV v26.16b, v27.16b

9:
        TBZ x1, 1, 10f
        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, [x18], 8
        DUP d26, v26.d[1]

10:
        TBZ x1, 0, 11f
        STR s20,  [x6]
        STR s22, [x16]
        STR s24, [x17]
        STR s26, [x18]
11:
        // Restore d12-d15 from stack
        LDP d14, d15, [sp, 16]
        LDP d12, d13, [sp], 32
        RET

END_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a55

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