| // 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_igemm_ukernel_4x12__aarch64_neonfma_cortex_a53( |
| # size_t mr, x0 |
| # size_t nc, x1 |
| # size_t kc, x2 / x0 |
| # size_t ks, x3 / x9 |
| # const float**restrict a, x4 |
| # const float*restrict w, x5 |
| # float*restrict c, x6 |
| # size_t cm_stride, x7 |
| # size_t cn_stride, [sp] -> x10 |
| # size_t a_offset, [sp + 8] -> x11 |
| # const float* zero, [sp + 16] -> x12 |
| # const xnn_f32_output_params params [sp + 24] -> x8 |
| |
| # d8-d15 need to be preserved if used. |
| # x19-30 need to be preserved if used. |
| |
| # A pointers |
| # x22 a0 |
| # x23 a1 |
| # x24 a2 |
| # x25 a3 |
| |
| # C pointers |
| # x6 c0 |
| # x20 c1 |
| # x21 c2 |
| # x7 c3 / cm_stride |
| |
| # Vector register usage and GPR shadows |
| # A0 v0 first set of A |
| # A1 v0[1] x13 |
| # A2 v1 |
| # A3 v1[1] x8 |
| # A0 v2 second set of A |
| # A1 v2[1] x13 |
| # A2 v3 |
| # A3 v3[1] x8 |
| # B v4 v5 v6 x14 x15 x16 first set of B |
| # B v7 v8 v9 x17 x18 x19 |
| # B v14 v15 v16 x14 x15 x16 second set of B (same x as first set) |
| # B v17 v18 v19 x17 x18 x19 |
| # C v20 v21 v22 |
| # C v23 v24 v25 |
| # C v26 v27 v28 |
| # C v29 v30 v31 |
| # Clamp v10 v11 |
| # v12 to v13 unused. |
| |
| BEGIN_FUNCTION xnn_f32_igemm_ukernel_4x12__aarch64_neonfma_cortex_a53 |
| |
| # Load cn_stride, a_offset |
| LDP x10, x11, [sp] |
| |
| # Load zero, clamping params pointer |
| LDP x12, x8, [sp, 16] |
| |
| # Save x19-x25 on stack |
| STR x19, [sp, -112]! |
| STP x20, x21, [sp, 16] |
| STP x22, x23, [sp, 32] |
| STP x24, x25, [sp, 48] |
| |
| # Save d8-d11,d14,d15 on stack |
| STP d8, d9, [sp, 64] |
| STP d10, d11, [sp, 80] |
| STP d14, d15, [sp, 96] |
| |
| # Load clamping_params values |
| LD2R {v10.4s, v11.4s}, [x8] |
| |
| # Clamp C pointers |
| CMP x0, 2 // if mr < 2 |
| ADD x20, x6, x7 // c1 = c0 + cm_stride |
| CSEL x20, x6, x20, LO // c1 = c0 |
| |
| ADD x21, x20, x7 // c2 = c1 + cm_stride |
| // if mr <= 2 |
| |
| CSEL x21, x20, x21, LS // c2 = c1 |
| |
| CMP x0, 4 // if mr < 4 |
| ADD x7, x21, x7 // c3 = c2 + cm_stride |
| CSEL x7, x21, x7, LO // c3 = c2 |
| |
| 0: |
| # Load initial bias from w into accumulators |
| LD1 {v20.16b, v21.16b, v22.16b}, [x5], 48 |
| MOV v23.16b, v20.16b |
| PRFM PLDL1KEEP, [x5] |
| MOV v24.16b, v21.16b |
| PRFM PLDL1KEEP, [x5, 64] |
| MOV v25.16b, v22.16b |
| PRFM PLDL1KEEP, [x5, 128] |
| MOV v26.16b, v20.16b |
| PRFM PLDL1KEEP, [x5, 192] |
| MOV v27.16b, v21.16b |
| PRFM PLDL1KEEP, [x5, 256] |
| MOV v28.16b, v22.16b |
| PRFM PLDL1KEEP, [x5, 320] |
| MOV v29.16b, v20.16b |
| MOV v30.16b, v21.16b |
| MOV v31.16b, v22.16b |
| |
| MOV x9, x3 // p = ks |
| |
| 1: |
| # Load next 4 A pointers |
| LDP x22, x23, [x4], 16 |
| LDP x24, x25, [x4], 16 |
| |
| CMP x22, x12 // if a0 == zero |
| ADD x22, x22, x11 // a0 += a_offset |
| CSEL x22, x12, x22, EQ // a0 = zero, else += a0 + a_offset |
| CMP x23, x12 // if a1 == zero |
| ADD x23, x23, x11 // a1 += a_offset |
| CSEL x23, x12, x23, EQ // a1 = zero, else += a1 + a_offset |
| CMP x24, x12 // if a2 == zero |
| ADD x24, x24, x11 // a2 += a_offset |
| CSEL x24, x12, x24, EQ // a2 = zero, else += a2 + a_offset |
| CMP x25, x12 // if a3 == zero |
| ADD x25, x25, x11 // a3 += a_offset |
| CSEL x25, x12, x25, EQ // a3 = zero, else += a3 + a_offset |
| |
| # Is there at least 4 floats (16 bytes) for prologue + epilogue? |
| SUBS x0, x2, 16 // k = kc - 16 |
| B.LO 4f |
| |
| SUBS x0, x0, 16 // 4 floats for main loop |
| |
| # Prologue - loads for first group of 24 FMA |
| |
| # Read first block of 4 A. |
| LDR d0, [x22], 8 // a0 |
| LDR x13, [x23], 8 // a1 |
| LDR d1, [x24], 8 // a2 |
| LDR x8, [x25], 8 // a3 |
| |
| LDR d4, [x5] // vb0x0123 |
| LDR x14, [x5, 8] |
| |
| LDR d5, [x5, 16] // vb0x25567 |
| LDR x15, [x5, 24] |
| |
| LDR d6, [x5, 32] // vb0x89AB |
| LDR x16, [x5, 40] |
| |
| LDR d7, [x5, 48] // vb1x0123 |
| INS v0.d[1], x13 |
| LDR x17, [x5, 56] |
| |
| LDR d8, [x5, 64] // vb1x25567 |
| INS v1.d[1], x8 |
| LDR x18, [x5, 72] |
| |
| LDR d9, [x5, 80] // vb1x89AB |
| LDR x19, [x5, 88] |
| INS v4.d[1], x14 |
| ADD x5, x5, 96 |
| |
| # Is there at least 4 floats (16 bytes) for main loop? |
| B.LO 3f |
| |
| # Main loop - 4 floats of A (16 bytes) |
| 2: |
| # First group of 24 fma. 8 blocks of 4 cycles. LDR + 3 FMA |
| # A is loaded for 2nd group into v2/v3 |
| # INS is 4 blocks (16 cycles) after load |
| |
| # BLOCK 0 |
| LDR d2, [x22], 8 // a0 |
| INS v5.d[1], x15 |
| FMLA v20.4s, v4.4s, v0.s[0] |
| LDR x13, [x23], 8 // a1 |
| FMLA v23.4s, v4.4s, v0.s[2] |
| PRFM PLDL1KEEP, [x5, 192] |
| FMLA v26.4s, v4.4s, v1.s[0] |
| |
| # BLOCK 1 |
| LDR d3, [x24], 8 // a2 |
| INS v6.d[1], x16 |
| FMLA v29.4s, v4.4s, v1.s[2] |
| LDR x8, [x25], 8 // a3 |
| FMLA v21.4s, v5.4s, v0.s[0] |
| PRFM PLDL1KEEP, [x5, 256] |
| FMLA v24.4s, v5.4s, v0.s[2] |
| |
| # BLOCK 2 |
| LDR d14, [x5] // vb0x0123 |
| INS v7.d[1], x17 |
| FMLA v27.4s, v5.4s, v1.s[0] |
| LDR x14, [x5, 8] |
| FMLA v30.4s, v5.4s, v1.s[2] |
| PRFM PLDL1KEEP, [x5, 320] |
| FMLA v22.4s, v6.4s, v0.s[0] |
| |
| # BLOCK 3 |
| LDR d15, [x5, 16] // vb0x25567 |
| INS v8.d[1], x18 |
| FMLA v25.4s, v6.4s, v0.s[2] |
| LDR x15, [x5, 24] |
| FMLA v28.4s, v6.4s, v1.s[0] |
| FMLA v31.4s, v6.4s, v1.s[2] |
| |
| # BLOCK 4 |
| LDR d16, [x5, 32] // vb0x89AB |
| INS v9.d[1], x19 |
| FMLA v20.4s, v7.4s, v0.s[1] |
| LDR x16, [x5, 40] |
| FMLA v23.4s, v7.4s, v0.s[3] |
| FMLA v26.4s, v7.4s, v1.s[1] |
| |
| # BLOCK 5 |
| LDR d17, [x5, 48] // vb1x0123 |
| INS v2.d[1], x13 // a1 was loaded in block 0 |
| FMLA v29.4s, v7.4s, v1.s[3] |
| LDR x17, [x5, 56] |
| FMLA v21.4s, v8.4s, v0.s[1] |
| FMLA v24.4s, v8.4s, v0.s[3] |
| |
| # BLOCK 6 |
| LDR d18, [x5, 64] // vb1x25567 |
| INS v3.d[1], x8 // a3 was loaded in block 1 |
| FMLA v27.4s, v8.4s, v1.s[1] |
| LDR x18, [x5, 72] |
| FMLA v30.4s, v8.4s, v1.s[3] |
| FMLA v22.4s, v9.4s, v0.s[1] |
| |
| # BLOCK 7 |
| LDR d19, [x5, 80] // vb1x89AB |
| INS v14.d[1], x14 // v14 was loaded in block 2 |
| FMLA v25.4s, v9.4s, v0.s[3] |
| LDR x19, [x5, 88] |
| FMLA v28.4s, v9.4s, v1.s[1] |
| FMLA v31.4s, v9.4s, v1.s[3] |
| |
| # Second group of 24 fma. 8 blocks of 4 cycles. LDR + 3 FMA |
| # A is loaded for 1st group into v0/v1 |
| |
| # BLOCK 0 |
| LDR d0, [x22], 8 // a0 |
| INS v15.d[1], x15 |
| FMLA v20.4s, v14.4s, v2.s[0] |
| LDR x13, [x23], 8 // a1 |
| FMLA v23.4s, v14.4s, v2.s[2] |
| FMLA v26.4s, v14.4s, v3.s[0] |
| |
| # BLOCK 1 |
| LDR d1, [x24], 8 // a2 |
| INS v16.d[1], x16 |
| FMLA v29.4s, v14.4s, v3.s[2] |
| LDR x8, [x25], 8 // a3 |
| FMLA v21.4s, v15.4s, v2.s[0] |
| FMLA v24.4s, v15.4s, v2.s[2] |
| |
| # BLOCK 2 |
| LDR d4, [x5, 96] // vb0x0123 |
| INS v17.d[1], x17 |
| FMLA v27.4s, v15.4s, v3.s[0] |
| LDR x14, [x5, 104] |
| FMLA v30.4s, v15.4s, v3.s[2] |
| FMLA v22.4s, v16.4s, v2.s[0] |
| |
| # BLOCK 3 |
| LDR d5, [x5, 112] // vb0x25567 |
| INS v18.d[1], x18 |
| FMLA v25.4s, v16.4s, v2.s[2] |
| LDR x15, [x5, 120] |
| FMLA v28.4s, v16.4s, v3.s[0] |
| FMLA v31.4s, v16.4s, v3.s[2] |
| |
| # BLOCK 4 |
| LDR d6, [x5, 128] // vb0x89AB |
| INS v19.d[1], x19 |
| FMLA v20.4s, v17.4s, v2.s[1] |
| LDR x16, [x5, 136] |
| FMLA v23.4s, v17.4s, v2.s[3] |
| FMLA v26.4s, v17.4s, v3.s[1] |
| |
| # BLOCK 5 |
| LDR d7, [x5, 144] // vb1x0123 |
| INS v0.d[1], x13 // a1 |
| FMLA v29.4s, v17.4s, v3.s[3] |
| LDR x17, [x5, 152] |
| FMLA v21.4s, v18.4s, v2.s[1] |
| FMLA v24.4s, v18.4s, v2.s[3] |
| |
| # BLOCK 6 |
| LDR d8, [x5, 160] // vb1x25567 |
| INS v1.d[1], x8 // a3 |
| FMLA v27.4s, v18.4s, v3.s[1] |
| LDR x18, [x5, 168] |
| FMLA v30.4s, v18.4s, v3.s[3] |
| SUBS x0, x0, 16 |
| FMLA v22.4s, v19.4s, v2.s[1] |
| |
| # BLOCK 7 |
| LDR d9, [x5, 176] // vb1x89AB |
| INS v4.d[1], x14 |
| FMLA v25.4s, v19.4s, v2.s[3] |
| LDR x19, [x5, 184] |
| FMLA v28.4s, v19.4s, v3.s[1] |
| ADD x5, x5, 192 |
| FMLA v31.4s, v19.4s, v3.s[3] |
| B.HS 2b |
| |
| # Epilogue |
| # First block same as main loop. Second block has no loads. |
| 3: |
| # BLOCK 0 |
| LDR d2, [x22], 8 // a0 |
| INS v5.d[1], x15 |
| FMLA v20.4s, v4.4s, v0.s[0] |
| LDR x13, [x23], 8 // a1 |
| FMLA v23.4s, v4.4s, v0.s[2] |
| PRFM PLDL1KEEP, [x5, 192] |
| FMLA v26.4s, v4.4s, v1.s[0] |
| PRFM PSTL1KEEP, [x6] // Prefetch C0 |
| |
| # BLOCK 1 |
| LDR d3, [x24], 8 // a2 |
| INS v6.d[1], x16 |
| FMLA v29.4s, v4.4s, v1.s[2] |
| LDR x8, [x25], 8 // a3 |
| FMLA v21.4s, v5.4s, v0.s[0] |
| PRFM PLDL1KEEP, [x5, 256] |
| FMLA v24.4s, v5.4s, v0.s[2] |
| PRFM PSTL1KEEP, [x20] // Prefetch C1 |
| |
| # BLOCK 2 |
| LDR d14, [x5] // vb0x0123 |
| INS v7.d[1], x17 |
| FMLA v27.4s, v5.4s, v1.s[0] |
| LDR x14, [x5, 8] |
| FMLA v30.4s, v5.4s, v1.s[2] |
| PRFM PLDL1KEEP, [x5, 320] |
| FMLA v22.4s, v6.4s, v0.s[0] |
| PRFM PSTL1KEEP, [x21] // Prefetch C2 |
| |
| # BLOCK 3 |
| LDR d15, [x5, 16] // vb0x25567 |
| INS v8.d[1], x18 |
| FMLA v25.4s, v6.4s, v0.s[2] |
| LDR x15, [x5, 24] |
| FMLA v28.4s, v6.4s, v1.s[0] |
| FMLA v31.4s, v6.4s, v1.s[2] |
| PRFM PSTL1KEEP, [x7] // Prefetch C3 |
| |
| # BLOCK 4 |
| LDR d16, [x5, 32] // vb0x89AB |
| INS v9.d[1], x19 |
| FMLA v20.4s, v7.4s, v0.s[1] |
| LDR x16, [x5, 40] |
| FMLA v23.4s, v7.4s, v0.s[3] |
| FMLA v26.4s, v7.4s, v1.s[1] |
| |
| # BLOCK 5 |
| LDR d17, [x5, 48] // vb1x0123 |
| INS v2.d[1], x13 // a1 was loaded in block 0 |
| FMLA v29.4s, v7.4s, v1.s[3] |
| LDR x17, [x5, 56] |
| FMLA v21.4s, v8.4s, v0.s[1] |
| FMLA v24.4s, v8.4s, v0.s[3] |
| |
| # BLOCK 6 |
| LDR d18, [x5, 64] // vb1x25567 |
| INS v3.d[1], x8 // a3 was loaded in block 1 |
| FMLA v27.4s, v8.4s, v1.s[1] |
| LDR x18, [x5, 72] |
| FMLA v30.4s, v8.4s, v1.s[3] |
| FMLA v22.4s, v9.4s, v0.s[1] |
| |
| # BLOCK 7 |
| LDR d19, [x5, 80] // vb1x89AB |
| INS v14.d[1], x14 // v14 was loaded in block 2 |
| FMLA v25.4s, v9.4s, v0.s[3] |
| LDR x19, [x5, 88] |
| FMLA v28.4s, v9.4s, v1.s[1] |
| ADD x5, x5, 96 |
| FMLA v31.4s, v9.4s, v1.s[3] |
| |
| # Second group of 24 fma. 8 blocks of 4 cycles. |
| # Epilogue version does no loads |
| |
| # BLOCK 0 |
| INS v15.d[1], x15 |
| FMLA v20.4s, v14.4s, v2.s[0] |
| FMLA v23.4s, v14.4s, v2.s[2] |
| FMLA v26.4s, v14.4s, v3.s[0] |
| |
| # BLOCK 1 |
| INS v16.d[1], x16 |
| FMLA v29.4s, v14.4s, v3.s[2] |
| FMLA v21.4s, v15.4s, v2.s[0] |
| FMLA v24.4s, v15.4s, v2.s[2] |
| |
| # BLOCK 2 |
| INS v17.d[1], x17 |
| FMLA v27.4s, v15.4s, v3.s[0] |
| FMLA v30.4s, v15.4s, v3.s[2] |
| FMLA v22.4s, v16.4s, v2.s[0] |
| |
| # BLOCK 3 |
| INS v18.d[1], x18 |
| FMLA v25.4s, v16.4s, v2.s[2] |
| FMLA v28.4s, v16.4s, v3.s[0] |
| FMLA v31.4s, v16.4s, v3.s[2] |
| |
| # BLOCK 4 |
| INS v19.d[1], x19 |
| FMLA v20.4s, v17.4s, v2.s[1] |
| FMLA v23.4s, v17.4s, v2.s[3] |
| FMLA v26.4s, v17.4s, v3.s[1] |
| |
| # BLOCK 5 |
| FMLA v29.4s, v17.4s, v3.s[3] |
| FMLA v21.4s, v18.4s, v2.s[1] |
| FMLA v24.4s, v18.4s, v2.s[3] |
| |
| # BLOCK 6 |
| FMLA v27.4s, v18.4s, v3.s[1] |
| FMLA v30.4s, v18.4s, v3.s[3] |
| FMLA v22.4s, v19.4s, v2.s[1] |
| |
| # BLOCK 7 |
| FMLA v25.4s, v19.4s, v2.s[3] |
| FMLA v28.4s, v19.4s, v3.s[1] |
| FMLA v31.4s, v19.4s, v3.s[3] |
| |
| 4: |
| # Is there a remainder?- 2 floats of A (8 bytes) |
| TBNZ x0, 3, 6f |
| # Is there a remainder?- 1 floats of A (4 bytes) |
| TBNZ x0, 2, 7f |
| |
| 5: |
| # ks loop |
| SUBS x9, x9, 32 // ks -= MR * sizeof(void*) |
| B.NE 1b |
| |
| # Clamp |
| FMIN v20.4s, v20.4s, v10.4s |
| FMIN v21.4s, v21.4s, v10.4s |
| FMIN v22.4s, v22.4s, v10.4s |
| FMIN v23.4s, v23.4s, v10.4s |
| FMIN v24.4s, v24.4s, v10.4s |
| FMIN v25.4s, v25.4s, v10.4s |
| FMIN v26.4s, v26.4s, v10.4s |
| FMIN v27.4s, v27.4s, v10.4s |
| FMIN v28.4s, v28.4s, v10.4s |
| FMIN v29.4s, v29.4s, v10.4s |
| FMIN v30.4s, v30.4s, v10.4s |
| FMIN v31.4s, v31.4s, v10.4s |
| FMAX v20.4s, v20.4s, v11.4s |
| FMAX v21.4s, v21.4s, v11.4s |
| FMAX v22.4s, v22.4s, v11.4s |
| FMAX v23.4s, v23.4s, v11.4s |
| FMAX v24.4s, v24.4s, v11.4s |
| FMAX v25.4s, v25.4s, v11.4s |
| FMAX v26.4s, v26.4s, v11.4s |
| FMAX v27.4s, v27.4s, v11.4s |
| FMAX v28.4s, v28.4s, v11.4s |
| FMAX v29.4s, v29.4s, v11.4s |
| FMAX v30.4s, v30.4s, v11.4s |
| FMAX v31.4s, v31.4s, v11.4s |
| |
| # Store full 4 x 12 |
| CMP x1, 12 |
| B.LO 8f |
| |
| ST1 {v29.16b, v30.16b, v31.16b}, [x7], x10 |
| ST1 {v26.16b, v27.16b, v28.16b}, [x21], x10 |
| ST1 {v23.16b, v24.16b, v25.16b}, [x20], x10 |
| ST1 {v20.16b, v21.16b, v22.16b}, [x6], x10 |
| SUB x4, x4, x3 // a -= ks |
| |
| # nc loop |
| SUBS x1, x1, 12 |
| B.HI 0b |
| |
| # Restore d8-d11,d14,d15 from stack |
| LDP d14, d15, [sp, 96] |
| LDP d10, d11, [sp, 80] |
| LDP d8, d9, [sp, 64] |
| |
| # Restore x19-x25 from stack |
| LDP x24, x25, [sp, 48] |
| LDP x22, x23, [sp, 32] |
| LDP x20, x21, [sp, 16] |
| LDR x19, [sp], 112 |
| RET |
| |
| 6: |
| # Remainder - 2 floats of A (8 bytes) |
| # Read first block of 4 A. |
| LDR d0, [x22], 8 // a0 |
| LDR d1, [x23], 8 // a1 |
| LDR d2, [x24], 8 // a2 |
| LDR d3, [x25], 8 // a3 |
| LD1 {v4.16b, v5.16b, v6.16b}, [x5], 48 |
| LD1 {v7.16b, v8.16b, v9.16b}, [x5], 48 |
| |
| # First block of 3 B |
| FMLA v20.4s, v4.4s, v0.s[0] |
| FMLA v23.4s, v4.4s, v1.s[0] |
| FMLA v26.4s, v4.4s, v2.s[0] |
| FMLA v29.4s, v4.4s, v3.s[0] |
| FMLA v21.4s, v5.4s, v0.s[0] |
| FMLA v24.4s, v5.4s, v1.s[0] |
| FMLA v27.4s, v5.4s, v2.s[0] |
| FMLA v30.4s, v5.4s, v3.s[0] |
| FMLA v22.4s, v6.4s, v0.s[0] |
| FMLA v25.4s, v6.4s, v1.s[0] |
| FMLA v28.4s, v6.4s, v2.s[0] |
| FMLA v31.4s, v6.4s, v3.s[0] |
| |
| # Second block of 3 B |
| FMLA v20.4s, v7.4s, v0.s[1] |
| FMLA v23.4s, v7.4s, v1.s[1] |
| FMLA v26.4s, v7.4s, v2.s[1] |
| FMLA v29.4s, v7.4s, v3.s[1] |
| FMLA v21.4s, v8.4s, v0.s[1] |
| FMLA v24.4s, v8.4s, v1.s[1] |
| FMLA v27.4s, v8.4s, v2.s[1] |
| FMLA v30.4s, v8.4s, v3.s[1] |
| FMLA v22.4s, v9.4s, v0.s[1] |
| FMLA v25.4s, v9.4s, v1.s[1] |
| FMLA v28.4s, v9.4s, v2.s[1] |
| FMLA v31.4s, v9.4s, v3.s[1] |
| |
| TBZ x0, 2, 5b |
| 7: |
| # Remainder - 1 float of A (4 bytes) |
| LDR s0, [x22], 4 // a0 |
| LDR s1, [x23], 4 // a1 |
| LDR s2, [x24], 4 // a2 |
| LDR s3, [x25], 4 // a3 |
| LD1 {v4.16b, v5.16b, v6.16b}, [x5], 48 |
| |
| FMLA v20.4s, v4.4s, v0.s[0] |
| FMLA v23.4s, v4.4s, v1.s[0] |
| FMLA v26.4s, v4.4s, v2.s[0] |
| FMLA v29.4s, v4.4s, v3.s[0] |
| FMLA v21.4s, v5.4s, v0.s[0] |
| FMLA v24.4s, v5.4s, v1.s[0] |
| FMLA v27.4s, v5.4s, v2.s[0] |
| FMLA v30.4s, v5.4s, v3.s[0] |
| FMLA v22.4s, v6.4s, v0.s[0] |
| FMLA v25.4s, v6.4s, v1.s[0] |
| FMLA v28.4s, v6.4s, v2.s[0] |
| FMLA v31.4s, v6.4s, v3.s[0] |
| B 5b |
| |
| 8: |
| # Store odd channels |
| TBZ x1, 3, 9f |
| STP q29, q30, [x7], 32 |
| MOV v29.16b, v31.16b |
| STP q26, q27, [x21], 32 |
| MOV v26.16b, v28.16b |
| STP q23, q24, [x20], 32 |
| MOV v23.16b, v25.16b |
| STP q20, q21, [x6], 32 |
| MOV v20.16b, v22.16b |
| |
| 9: |
| TBZ x1, 2, 10f |
| STR q29, [x7], 16 |
| MOV v29.16b, v30.16b |
| STR q26, [x21], 16 |
| MOV v26.16b, v27.16b |
| STR q23, [x20], 16 |
| MOV v23.16b, v24.16b |
| STR q20, [x6], 16 |
| MOV v20.16b, v21.16b |
| |
| 10: |
| TBZ x1, 1, 11f |
| STR d29, [x7], 8 |
| DUP d29, v29.d[1] |
| STR d26, [x21], 8 |
| DUP d26, v26.d[1] |
| STR d23, [x20], 8 |
| DUP d23, v23.d[1] |
| STR d20, [x6], 8 |
| DUP d20, v20.d[1] |
| |
| 11: |
| TBZ x1, 0, 12f |
| STR s29, [x7] |
| STR s26, [x21] |
| STR s23, [x20] |
| STR s20, [x6] |
| 12: |
| # Restore d8-d11,d14,d15 from stack |
| LDP d14, d15, [sp, 96] |
| LDP d10, d11, [sp, 80] |
| LDP d8, d9, [sp, 64] |
| |
| # Restore x19-x25 from stack |
| LDP x24, x25, [sp, 48] |
| LDP x22, x23, [sp, 32] |
| LDP x20, x21, [sp, 16] |
| LDR x19, [sp], 112 |
| RET |
| |
| END_FUNCTION xnn_f32_igemm_ukernel_4x12__aarch64_neonfma_cortex_a53 |
| |
| #ifdef __ELF__ |
| .section ".note.GNU-stack","",%progbits |
| #endif |