// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.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_a75(
#     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
# 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  v4
# A1  v1  v5
# A2  v2  v6
# A3  v3  v7
# B   v8  v9 v10 v11
# B  v12 v13 v14 v15
# B  v20 v21 v22 v23
# B  v24 v25 v26 v27
# C  v16 v17
# C  v18 v19
# C  v28 v29
# C  v30 v31
# Clamp v4 v5

BEGIN_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a75

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

        # Load min/max values
        LD2R {v4.4s, v5.4s}, [x8]

        # Save d8-d15 on stack
        STP  d8,  d9, [sp, -64]!
        STP d10, d11, [sp, 16]
        STP d12, d13, [sp, 32]
        STP d14, d15, [sp, 48]

        # 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

        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
        LDP q16, q17, [x5], 32
        MOV v18.16b, v16.16b
        MOV v19.16b, v17.16b
        MOV v28.16b, v16.16b
        MOV v29.16b, v17.16b
        MOV v30.16b, v16.16b
        MOV v31.16b, v17.16b

        # Is there at least 8 floats (32 bytes) for prologue + epilogue?
        SUBS x0, x2, 32  // k = kc - 32
        B.LO 3f

        # 16 prologue
        # Read first block of 4 A and B.
        LDR q0,  [x3], 16
        LDP q20, q21, [x5], 32
        LDR q1, [x11], 16
        LDR q2, [x12], 16
        LDR q3,  [x4], 16
        LDP q22, q23, [x5], 32
        LDP q24, q25, [x5], 32
        LDP q26, q27, [x5], 32

        # Is there at least 32.  yes do main loop
        SUBS x0, x0, 32
        B.LO 2f

        # Main loop - 8 floats of A (32 bytes)
1:
        # First block of 4.  FMA for first 4, loads for 2nd block of 4.
        FMLA v16.4s, v20.4s, v0.s[0]
        LDP q8, q9, [x5], 32
        FMLA v17.4s, v21.4s, v0.s[0]
        FMLA v18.4s, v20.4s, v1.s[0]
        LDP q10, q11, [x5], 32
        FMLA v19.4s, v21.4s, v1.s[0]
        FMLA v28.4s, v20.4s, v2.s[0]
        LDP q12, q13, [x5], 32
        FMLA v29.4s, v21.4s, v2.s[0]
        FMLA v30.4s, v20.4s, v3.s[0]
        LDP q14, q15, [x5], 32
        FMLA v31.4s, v21.4s, v3.s[0]
        FMLA v16.4s, v22.4s, v0.s[1]
        LDR q4, [x3], 16
        FMLA v17.4s, v23.4s, v0.s[1]
        FMLA v18.4s, v22.4s, v1.s[1]
        LDR q5, [x11], 16
        FMLA v19.4s, v23.4s, v1.s[1]
        FMLA v28.4s, v22.4s, v2.s[1]
        LDR q6, [x12], 16
        FMLA v29.4s, v23.4s, v2.s[1]
        FMLA v30.4s, v22.4s, v3.s[1]
        LDR q7, [x4], 16
        FMLA v31.4s, v23.4s, v3.s[1]
        FMLA v16.4s, v24.4s, v0.s[2]
        PRFM PLDL1KEEP, [x5, 128]
        FMLA v17.4s, v25.4s, v0.s[2]
        FMLA v18.4s, v24.4s, v1.s[2]
        PRFM PLDL1KEEP, [x5, 192]
        FMLA v19.4s, v25.4s, v1.s[2]
        FMLA v28.4s, v24.4s, v2.s[2]
        PRFM PLDL1KEEP, [x5, 256]
        FMLA v29.4s, v25.4s, v2.s[2]
        FMLA v30.4s, v24.4s, v3.s[2]
        PRFM PLDL1KEEP, [x5, 320]
        FMLA v31.4s, v25.4s, v3.s[2]
        FMLA v16.4s, v26.4s, v0.s[3]
        FMLA v17.4s, v27.4s, v0.s[3]
        FMLA v18.4s, v26.4s, v1.s[3]
        FMLA v19.4s, v27.4s, v1.s[3]
        FMLA v28.4s, v26.4s, v2.s[3]
        FMLA v29.4s, v27.4s, v2.s[3]
        FMLA v30.4s, v26.4s, v3.s[3]
        FMLA v31.4s, v27.4s, v3.s[3]

        # Second block of 4.  FMA for second 4, loads for 1nd block of 4.
        FMLA v16.4s, v8.4s, v4.s[0]
        LDP q20, q21, [x5], 32
        FMLA v17.4s, v9.4s, v4.s[0]
        FMLA v18.4s, v8.4s, v5.s[0]
        LDP q22, q23, [x5], 32
        FMLA v19.4s, v9.4s, v5.s[0]
        FMLA v28.4s, v8.4s, v6.s[0]
        LDP q24, q25, [x5], 32
        FMLA v29.4s, v9.4s, v6.s[0]
        FMLA v30.4s, v8.4s, v7.s[0]
        LDP q26, q27, [x5], 32
        FMLA v31.4s, v9.4s, v7.s[0]
        FMLA v16.4s, v10.4s, v4.s[1]
        LDR q0, [x3], 16
        FMLA v17.4s, v11.4s, v4.s[1]
        FMLA v18.4s, v10.4s, v5.s[1]
        LDR q1, [x11], 16
        FMLA v19.4s, v11.4s, v5.s[1]
        FMLA v28.4s, v10.4s, v6.s[1]
        LDR q2, [x12], 16
        FMLA v29.4s, v11.4s, v6.s[1]
        FMLA v30.4s, v10.4s, v7.s[1]
        LDR q3, [x4], 16
        FMLA v31.4s, v11.4s, v7.s[1]
        FMLA v16.4s, v12.4s, v4.s[2]
        FMLA v17.4s, v13.4s, v4.s[2]
        FMLA v18.4s, v12.4s, v5.s[2]
        FMLA v19.4s, v13.4s, v5.s[2]
        FMLA v28.4s, v12.4s, v6.s[2]
        FMLA v29.4s, v13.4s, v6.s[2]
        FMLA v30.4s, v12.4s, v7.s[2]
        FMLA v31.4s, v13.4s, v7.s[2]
        FMLA v16.4s, v14.4s, v4.s[3]
        FMLA v17.4s, v15.4s, v4.s[3]
        FMLA v18.4s, v14.4s, v5.s[3]
        FMLA v19.4s, v15.4s, v5.s[3]
        FMLA v28.4s, v14.4s, v6.s[3]
        FMLA v29.4s, v15.4s, v6.s[3]
        SUBS x0, x0, 32
        FMLA v30.4s, v14.4s, v7.s[3]
        FMLA v31.4s, v15.4s, v7.s[3]
        B.HS 1b

2:
        # Epilogue
        # First block of 4.  FMA for first 4, loads for 2nd block of 4.
        FMLA v16.4s, v20.4s, v0.s[0]
        LDP q8, q9, [x5], 32
        FMLA v17.4s, v21.4s, v0.s[0]
        FMLA v18.4s, v20.4s, v1.s[0]
        LDP q10, q11, [x5], 32
        FMLA v19.4s, v21.4s, v1.s[0]
        FMLA v28.4s, v20.4s, v2.s[0]
        LDP q12, q13, [x5], 32
        FMLA v29.4s, v21.4s, v2.s[0]
        FMLA v30.4s, v20.4s, v3.s[0]
        LDP q14, q15, [x5], 32
        FMLA v31.4s, v21.4s, v3.s[0]
        FMLA v16.4s, v22.4s, v0.s[1]
        LDR q4, [x3], 16
        FMLA v17.4s, v23.4s, v0.s[1]
        FMLA v18.4s, v22.4s, v1.s[1]
        LDR q5, [x11], 16
        FMLA v19.4s, v23.4s, v1.s[1]
        FMLA v28.4s, v22.4s, v2.s[1]
        LDR q6, [x12], 16
        FMLA v29.4s, v23.4s, v2.s[1]
        FMLA v30.4s, v22.4s, v3.s[1]
        LDR q7, [x4], 16
        FMLA v31.4s, v23.4s, v3.s[1]
        FMLA v16.4s, v24.4s, v0.s[2]
        FMLA v17.4s, v25.4s, v0.s[2]
        FMLA v18.4s, v24.4s, v1.s[2]
        FMLA v19.4s, v25.4s, v1.s[2]
        FMLA v28.4s, v24.4s, v2.s[2]
        FMLA v29.4s, v25.4s, v2.s[2]
        FMLA v30.4s, v24.4s, v3.s[2]
        FMLA v31.4s, v25.4s, v3.s[2]
        FMLA v16.4s, v26.4s, v0.s[3]
        FMLA v17.4s, v27.4s, v0.s[3]
        FMLA v18.4s, v26.4s, v1.s[3]
        FMLA v19.4s, v27.4s, v1.s[3]
        FMLA v28.4s, v26.4s, v2.s[3]
        FMLA v29.4s, v27.4s, v2.s[3]
        FMLA v30.4s, v26.4s, v3.s[3]
        FMLA v31.4s, v27.4s, v3.s[3]

        # Second block of 4.  FMA for second 4, noloads
        FMLA v16.4s, v8.4s, v4.s[0]
        FMLA v17.4s, v9.4s, v4.s[0]
        FMLA v18.4s, v8.4s, v5.s[0]
        FMLA v19.4s, v9.4s, v5.s[0]
        FMLA v28.4s, v8.4s, v6.s[0]
        FMLA v29.4s, v9.4s, v6.s[0]
        FMLA v30.4s, v8.4s, v7.s[0]
        FMLA v31.4s, v9.4s, v7.s[0]

        FMLA v16.4s, v10.4s, v4.s[1]
        FMLA v17.4s, v11.4s, v4.s[1]
        FMLA v18.4s, v10.4s, v5.s[1]
        FMLA v19.4s, v11.4s, v5.s[1]
        FMLA v28.4s, v10.4s, v6.s[1]
        FMLA v29.4s, v11.4s, v6.s[1]
        FMLA v30.4s, v10.4s, v7.s[1]
        FMLA v31.4s, v11.4s, v7.s[1]

        FMLA v16.4s, v12.4s, v4.s[2]
        FMLA v17.4s, v13.4s, v4.s[2]
        FMLA v18.4s, v12.4s, v5.s[2]
        FMLA v19.4s, v13.4s, v5.s[2]
        FMLA v28.4s, v12.4s, v6.s[2]
        FMLA v29.4s, v13.4s, v6.s[2]
        FMLA v30.4s, v12.4s, v7.s[2]
        FMLA v31.4s, v13.4s, v7.s[2]

        FMLA v16.4s, v14.4s, v4.s[3]
        FMLA v17.4s, v15.4s, v4.s[3]
        FMLA v18.4s, v14.4s, v5.s[3]
        FMLA v19.4s, v15.4s, v5.s[3]

        # Load min/max values
        LD2R {v4.4s, v5.4s}, [x8]

        FMLA v28.4s, v14.4s, v6.s[3]
        FMLA v29.4s, v15.4s, v6.s[3]
        FMLA v30.4s, v14.4s, v7.s[3]
        FMLA v31.4s, v15.4s, v7.s[3]

3:
        # Remainder- 4 floats of A (16 bytes)
        TBZ x0, 4, 4f

        LDR q0,  [x3], 16
        LDP q20, q21, [x5], 32
        LDR q1, [x11], 16
        LDR q2, [x12], 16
        LDR q3,  [x4], 16
        FMLA v16.4s, v20.4s, v0.s[0]
        FMLA v17.4s, v21.4s, v0.s[0]
        LDP q22, q23, [x5], 32
        FMLA v18.4s, v20.4s, v1.s[0]
        FMLA v19.4s, v21.4s, v1.s[0]
        LDP q24, q25, [x5], 32
        FMLA v28.4s, v20.4s, v2.s[0]
        FMLA v29.4s, v21.4s, v2.s[0]
        LDP q26, q27, [x5], 32
        FMLA v30.4s, v20.4s, v3.s[0]
        FMLA v31.4s, v21.4s, v3.s[0]
        FMLA v16.4s, v22.4s, v0.s[1]
        FMLA v17.4s, v23.4s, v0.s[1]
        FMLA v18.4s, v22.4s, v1.s[1]
        FMLA v19.4s, v23.4s, v1.s[1]
        FMLA v28.4s, v22.4s, v2.s[1]
        FMLA v29.4s, v23.4s, v2.s[1]
        FMLA v30.4s, v22.4s, v3.s[1]
        FMLA v31.4s, v23.4s, v3.s[1]
        FMLA v16.4s, v24.4s, v0.s[2]
        FMLA v17.4s, v25.4s, v0.s[2]
        FMLA v18.4s, v24.4s, v1.s[2]
        FMLA v19.4s, v25.4s, v1.s[2]
        FMLA v28.4s, v24.4s, v2.s[2]
        FMLA v29.4s, v25.4s, v2.s[2]
        FMLA v30.4s, v24.4s, v3.s[2]
        FMLA v31.4s, v25.4s, v3.s[2]
        FMLA v16.4s, v26.4s, v0.s[3]
        FMLA v17.4s, v27.4s, v0.s[3]
        FMLA v18.4s, v26.4s, v1.s[3]
        FMLA v19.4s, v27.4s, v1.s[3]
        FMLA v28.4s, v26.4s, v2.s[3]
        FMLA v29.4s, v27.4s, v2.s[3]
        FMLA v30.4s, v26.4s, v3.s[3]
        FMLA v31.4s, v27.4s, v3.s[3]

4:
        # Remainder- 2 floats of A (8 bytes)
        TBZ x0, 3, 5f

        LDR d0,  [x3], 8
        LDP q20, q21, [x5], 32
        LDR d1, [x11], 8
        LDR d2, [x12], 8
        LDR d3,  [x4], 8
        FMLA v16.4s, v20.4s, v0.s[0]
        FMLA v17.4s, v21.4s, v0.s[0]
        LDP q22, q23, [x5], 32
        FMLA v18.4s, v20.4s, v1.s[0]
        FMLA v19.4s, v21.4s, v1.s[0]
        FMLA v28.4s, v20.4s, v2.s[0]
        FMLA v29.4s, v21.4s, v2.s[0]
        FMLA v30.4s, v20.4s, v3.s[0]
        FMLA v31.4s, v21.4s, v3.s[0]
        FMLA v16.4s, v22.4s, v0.s[1]
        FMLA v17.4s, v23.4s, v0.s[1]
        FMLA v18.4s, v22.4s, v1.s[1]
        FMLA v19.4s, v23.4s, v1.s[1]
        FMLA v28.4s, v22.4s, v2.s[1]
        FMLA v29.4s, v23.4s, v2.s[1]
        FMLA v30.4s, v22.4s, v3.s[1]
        FMLA v31.4s, v23.4s, v3.s[1]

5:
        # Remainder- 1 float of A (4 bytes)
        TBZ x0, 2, 6f

        LDR s0,  [x3], 4
        LDP q20, q21, [x5], 32
        LDR s1, [x11], 4
        LDR s2, [x12], 4
        LDR s3,  [x4], 4
        FMLA v16.4s, v20.4s, v0.s[0]
        FMLA v17.4s, v21.4s, v0.s[0]
        FMLA v18.4s, v20.4s, v1.s[0]
        FMLA v19.4s, v21.4s, v1.s[0]
        FMLA v28.4s, v20.4s, v2.s[0]
        FMLA v29.4s, v21.4s, v2.s[0]
        FMLA v30.4s, v20.4s, v3.s[0]
        FMLA v31.4s, v21.4s, v3.s[0]

6:
        # Clamp
        FMAX v16.4s, v16.4s, v4.4s
        SUBS x1, x1, 8
        FMAX v17.4s, v17.4s, v4.4s
        FMAX v18.4s, v18.4s, v4.4s
        FMAX v19.4s, v19.4s, v4.4s
        FMAX v28.4s, v28.4s, v4.4s
        FMAX v29.4s, v29.4s, v4.4s
        FMAX v30.4s, v30.4s, v4.4s
        FMAX v31.4s, v31.4s, v4.4s
        FMIN v16.4s, v16.4s, v5.4s
        FMIN v17.4s, v17.4s, v5.4s
        FMIN v18.4s, v18.4s, v5.4s
        FMIN v19.4s, v19.4s, v5.4s
        FMIN v28.4s, v28.4s, v5.4s
        FMIN v29.4s, v29.4s, v5.4s
        FMIN v30.4s, v30.4s, v5.4s
        FMIN v31.4s, v31.4s, v5.4s

        # Store full 4 x 8
        B.LO 7f

        STP q16, q17,  [x6]
        SUB  x3,  x3, x2 // a0 -= kc
        ADD  x6,  x6, x14
        STP q18, q19,  [x9]
        SUB x11, x11, x2 // a1 -= kc
        ADD  x9,  x9, x14
        STP q28, q29, [x10]
        SUB x12, x12, x2 // a2 -= kc
        ADD x10, x10, x14
        STP q30, q31,  [x7]
        SUB  x4,  x4, x2 // a3 -= kc
        ADD  x7,  x7, x14

        B.HI 0b

        # Restore d8-d15 from stack
        LDP d14, d15, [sp, 48]
        LDP d12, d13, [sp, 32]
        LDP d10, d11, [sp, 16]
        LDP  d8,  d9, [sp], 64
        RET

        # Store odd width
7:
        TBZ x1, 2, 8f
        STR q16, [x6], 16
        MOV v16.16b, v17.16b
        STR q18, [x9], 16
        MOV v18.16b, v19.16b
        STR q28, [x10], 16
        MOV v28.16b, v29.16b
        STR q30, [x7], 16
        MOV v30.16b, v31.16b

8:
        TBZ x1, 1, 9f
        STR d16, [x6], 8
        DUP d16, v16.d[1]
        STR d18, [x9], 8
        DUP d18, v18.d[1]
        STR d28, [x10], 8
        DUP d28, v28.d[1]
        STR d30, [x7], 8
        DUP d30, v30.d[1]

9:
        TBZ x1, 0, 10f
        STR s16,  [x6]
        STR s18,  [x9]
        STR s28, [x10]
        STR s30,  [x7]
10:
        # Restore d8-d15 from stack
        LDP d14, d15, [sp, 48]
        LDP d12, d13, [sp, 32]
        LDP d10, d11, [sp, 16]
        LDP  d8,  d9, [sp], 64
        RET


END_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a75

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