// 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. $assert DATATYPE in ["F32", "QC8", "QC4"] #include $DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE] # void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}${"_goi" if GOI else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_ld128( # size_t mr, x0 # size_t nc, x1 # size_t kc, x2 / x0 # const float* a, x3 # size_t a_stride, x4 # const void* w, x5 # float* c, x6 # size_t cm_stride, x7 # size_t cn_stride, [sp] -> x14 $if INC: # const float* acc, [sp + 8] -> x15 # const xnn_f32_minmax_params* params) [sp + 16] -> (x8) $elif DATATYPE == "QC4": # const xnn_f32_qc4w_minmax_params* params) [sp + 8] -> (x8) $else: # const xnn_f32_minmax_params* params) [sp + 8] -> (x8) # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. # Register usage # A0 x3 v0 # A1 x11 v1 # A2 x12 v2 # A3 x4 v3 # B x5 v20 v24 v21 v25 v22 v26 v23 v27 # C0 x6 v16 v17 # C1 x9 v18 v19 # C2 x10 v28 v29 # C3 x7 v30 v31 # Clamp v4 v5 $if DATATYPE == "QC4": # zerop/mask v6 v7 $if GOI: # B stride x16, x17 BEGIN_FUNCTION xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}${"_goi" if GOI else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_ld128 $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] $if DATATYPE == "QC4": # Load min/max/zerop values LD3R {v4.4s, v5.4s, v6.4s}, [x8] MOVI v7.8b, 15 $else: # Load min/max values LD2R {v4.4s, v5.4s}, [x8] # 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 $if GOI: SUB x16, x2, x2, LSL #3 // B stride on 8th row to top (16 - KC * 7) ADD x16, x16, 16 LSL x17, x2, 3 // B stride for N loop SUB x17, x17, x2 0: $if INC: # Load initial accumulators LDP q16, q17, [x15], 32 LDP q18, q19, [x15], 32 LDP q28, q29, [x15], 32 LDP q30, q31, [x15], 32 $else: $if GOI: # Zero bias for GOI MOVI v16.16b, 0 MOVI v17.16b, 0 $else: # 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 4 floats (16 bytes)? SUBS x0, x2, 16 // k = kc - 16 B.LO 3f # Main loop - 4 floats of A (16 bytes) 1: LDR q0, [x3], 16 $if GOI: # Load 32 F32 weights from 8x4 and transpose to 4x8 LD4 {v20.s, v21.s, v22.s, v23.s}[0], [x5], x2 LD4 {v20.s, v21.s, v22.s, v23.s}[1], [x5], x2 LD4 {v20.s, v21.s, v22.s, v23.s}[2], [x5], x2 LD4 {v20.s, v21.s, v22.s, v23.s}[3], [x5], x2 LD4 {v24.s, v25.s, v26.s, v27.s}[0], [x5], x2 LD4 {v24.s, v25.s, v26.s, v27.s}[1], [x5], x2 LD4 {v24.s, v25.s, v26.s, v27.s}[2], [x5], x2 LD4 {v24.s, v25.s, v26.s, v27.s}[3], [x5], x16 $else: $if DATATYPE == "F32": LDP q20, q24, [x5], 32 // 8 F32 weights $else: LDP q20, q22, [x5], 32 // 32 QC8 weights SXTL v24.8h, v20.8b SXTL2 v25.8h, v20.16b SXTL v20.4s, v24.4h SXTL2 v24.4s, v24.8h SXTL v21.4s, v25.4h SXTL2 v25.4s, v25.8h SCVTF v20.4s, v20.4s SCVTF v24.4s, v24.4s SXTL v26.8h, v22.8b SXTL2 v27.8h, v22.16b LDR q1, [x11], 16 LDR q2, [x12], 16 LDR q3, [x4], 16 FMLA v16.4s, v20.4s, v0.s[0] FMLA v18.4s, v20.4s, v1.s[0] FMLA v28.4s, v20.4s, v2.s[0] FMLA v30.4s, v20.4s, v3.s[0] $if not GOI: $if DATATYPE == "F32": LDP q21, q25, [x5], 32 // 8 F32 weights $else: SCVTF v21.4s, v21.4s SCVTF v25.4s, v25.4s SXTL v22.4s, v26.4h SXTL2 v26.4s, v26.8h FMLA v17.4s, v24.4s, v0.s[0] FMLA v19.4s, v24.4s, v1.s[0] FMLA v29.4s, v24.4s, v2.s[0] FMLA v31.4s, v24.4s, v3.s[0] $if not GOI: $if DATATYPE == "F32": LDP q22, q26, [x5], 32 // 8 F32 weights $else: SCVTF v22.4s, v22.4s SCVTF v26.4s, v26.4s SXTL v23.4s, v27.4h SXTL2 v27.4s, v27.8h FMLA v16.4s, v21.4s, v0.s[1] FMLA v18.4s, v21.4s, v1.s[1] FMLA v28.4s, v21.4s, v2.s[1] FMLA v30.4s, v21.4s, v3.s[1] $if not GOI: $if DATATYPE == "F32": LDP q23, q27, [x5], 32 // 8 F32 weights $else: SCVTF v23.4s, v23.4s SCVTF v27.4s, v27.4s FMLA v17.4s, v25.4s, v0.s[1] FMLA v19.4s, v25.4s, v1.s[1] FMLA v29.4s, v25.4s, v2.s[1] FMLA v31.4s, v25.4s, v3.s[1] FMLA v16.4s, v22.4s, v0.s[2] FMLA v18.4s, v22.4s, v1.s[2] FMLA v28.4s, v22.4s, v2.s[2] FMLA v30.4s, v22.4s, v3.s[2] FMLA v17.4s, v26.4s, v0.s[2] FMLA v19.4s, v26.4s, v1.s[2] FMLA v29.4s, v26.4s, v2.s[2] FMLA v31.4s, v26.4s, v3.s[2] FMLA v16.4s, v23.4s, v0.s[3] FMLA v18.4s, v23.4s, v1.s[3] FMLA v28.4s, v23.4s, v2.s[3] FMLA v30.4s, v23.4s, v3.s[3] SUBS x0, x0, 16 FMLA v17.4s, v27.4s, v0.s[3] FMLA v19.4s, v27.4s, v1.s[3] FMLA v29.4s, v27.4s, v2.s[3] FMLA v31.4s, v27.4s, v3.s[3] B.HS 1b TST x0, 15 B.NE 3f 2: $if DATATYPE in ["QC8", "QC4"]: # Scale LDP q20, q24, [x5], 32 FMUL v16.4s, v16.4s, v20.4s FMUL v17.4s, v17.4s, v24.4s FMUL v18.4s, v18.4s, v20.4s FMUL v19.4s, v19.4s, v24.4s FMUL v28.4s, v28.4s, v20.4s FMUL v29.4s, v29.4s, v24.4s FMUL v30.4s, v30.4s, v20.4s FMUL v31.4s, v31.4s, v24.4s # 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 5f $if GOI: ADD x5, x5, x17 // advance to next row of weights $if INC: ST1 {v30.16b, v31.16b}, [x7], x14 SUB x3, x3, x2 // a0 -= kc ST1 {v28.16b, v29.16b}, [x10], x14 SUB x11, x11, x2 // a1 -= kc ST1 {v18.16b, v19.16b}, [x9], x14 SUB x12, x12, x2 // a2 -= kc ST1 {v16.16b, v17.16b}, [x6], x14 SUB x4, x4, x2 // a3 -= kc $else: ST1 {v16.16b, v17.16b}, [x6], x14 SUB x3, x3, x2 // a0 -= kc ST1 {v18.16b, v19.16b}, [x9], x14 SUB x11, x11, x2 // a1 -= kc ST1 {v28.16b, v29.16b}, [x10], x14 SUB x12, x12, x2 // a2 -= kc ST1 {v30.16b, v31.16b}, [x7], x14 SUB x4, x4, x2 // a3 -= kc B.HI 0b RET # Remainder- 2 floats of A (8 bytes) 3: # Is there a remainder?- 2 floats of A (8 bytes) TBZ x0, 3, 4f $if GOI: LD2 {v20.s, v21.s}[0], [x5], x2 LD2 {v20.s, v21.s}[1], [x5], x2 LD2 {v20.s, v21.s}[2], [x5], x2 LD2 {v20.s, v21.s}[3], [x5], x2 LD2 {v24.s, v25.s}[0], [x5], x2 LD2 {v24.s, v25.s}[1], [x5], x2 LD2 {v24.s, v25.s}[2], [x5], x2 LD2 {v24.s, v25.s}[3], [x5], x16 SUB x5, x5, 8 $else: # Remainder- 2 floats of A (8 bytes) $if DATATYPE == "F32": LDP q20, q24, [x5], 32 // 16 F32 weights LDP q21, q25, [x5], 32 $else: LDR q20, [x5], 16 // 16 QC8 weights SXTL v24.8h, v20.8b SXTL2 v25.8h, v20.16b SXTL v20.4s, v24.4h SXTL2 v24.4s, v24.8h SXTL v21.4s, v25.4h SXTL2 v25.4s, v25.8h SCVTF v20.4s, v20.4s SCVTF v24.4s, v24.4s SCVTF v21.4s, v21.4s SCVTF v25.4s, v25.4s LDR d0, [x3], 8 LDR d1, [x11], 8 LDR d2, [x12], 8 LDR d3, [x4], 8 FMLA v16.4s, v20.4s, v0.s[0] FMLA v18.4s, v20.4s, v1.s[0] FMLA v28.4s, v20.4s, v2.s[0] FMLA v30.4s, v20.4s, v3.s[0] FMLA v17.4s, v24.4s, v0.s[0] FMLA v19.4s, v24.4s, v1.s[0] FMLA v29.4s, v24.4s, v2.s[0] FMLA v31.4s, v24.4s, v3.s[0] FMLA v16.4s, v21.4s, v0.s[1] FMLA v18.4s, v21.4s, v1.s[1] FMLA v28.4s, v21.4s, v2.s[1] FMLA v30.4s, v21.4s, v3.s[1] FMLA v17.4s, v25.4s, v0.s[1] FMLA v19.4s, v25.4s, v1.s[1] FMLA v29.4s, v25.4s, v2.s[1] FMLA v31.4s, v25.4s, v3.s[1] # Is there a remainder?- 1 float of A (4 bytes) TBZ x0, 2, 2b # Remainder- 1 float of A (4 bytes) 4: $if GOI: LD1 {v20.s}[0], [x5], x2 LD1 {v20.s}[1], [x5], x2 LD1 {v20.s}[2], [x5], x2 LD1 {v20.s}[3], [x5], x2 LD1 {v24.s}[0], [x5], x2 LD1 {v24.s}[1], [x5], x2 LD1 {v24.s}[2], [x5], x2 LD1 {v24.s}[3], [x5], x16 SUB x5, x5, 12 $else: # Remainder- 2 floats of A (8 bytes) $if DATATYPE == "F32": LDP q20, q24, [x5], 32 // 8 F32 weights $else: LDR d20, [x5], 8 // 8 QC8 weights SXTL v24.8h, v20.8b SXTL v20.4s, v24.4h SXTL2 v24.4s, v24.8h SCVTF v20.4s, v20.4s SCVTF v24.4s, v24.4s LDR s0, [x3], 4 LDR s1, [x11], 4 LDR s2, [x12], 4 LDR s3, [x4], 4 FMLA v16.4s, v20.4s, v0.s[0] FMLA v18.4s, v20.4s, v1.s[0] FMLA v28.4s, v20.4s, v2.s[0] FMLA v30.4s, v20.4s, v3.s[0] FMLA v17.4s, v24.4s, v0.s[0] FMLA v19.4s, v24.4s, v1.s[0] FMLA v29.4s, v24.4s, v2.s[0] FMLA v31.4s, v24.4s, v3.s[0] B 2b # Store odd width 5: TBZ x1, 2, 6f $if INC: STR q30, [x7], 16 MOV v30.16b, v31.16b STR q28, [x10], 16 MOV v28.16b, v29.16b STR q18, [x9], 16 MOV v18.16b, v19.16b STR q16, [x6], 16 MOV v16.16b, v17.16b $else: 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 6: TBZ x1, 1, 7f $if INC: STR d30, [x7], 8 STR d28, [x10], 8 DUP d30, v30.d[1] DUP d28, v28.d[1] STR d18, [x9], 8 STR d16, [x6], 8 DUP d18, v18.d[1] DUP d16, v16.d[1] $else: STR d16, [x6], 8 STR d18, [x9], 8 DUP d16, v16.d[1] DUP d18, v18.d[1] STR d28, [x10], 8 STR d30, [x7], 8 DUP d28, v28.d[1] DUP d30, v30.d[1] 7: TBZ x1, 0, 8f $if INC: STR s30, [x7] STR s28, [x10] STR s18, [x9] STR s16, [x6] $else: STR s16, [x6] STR s18, [x9] STR s28, [x10] STR s30, [x7] 8: RET END_FUNCTION xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}${"_goi" if GOI else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_ld128 #ifdef __ELF__ .section ".note.GNU-stack","",%progbits #endif