// Auto-generated file. Do not edit! // Template: src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S.in // Generator: tools/xngen // // Copyright 2020 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_qd8_f32_qc8w_gemm_minmax__ukernel_4x16c4__asm_aarch64_neondot_ld64( # size_t mr, x0 # size_t nc, x1 # size_t kc, x2 / x0 # const int8_t* restrict a, x3 # size_t a_stride, x4 # const void* restrict w, x5 # int8_t* restrict c, x6 # size_t cm_stride, x7 # size_t cn_stride, [sp] -> x12 # const union xnn_f32_minmax_params *params, [sp + 8] -> x11 # const struct xnn_qd8_quantization_params *quantization_params) [sp + 16] -> x16 # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. // Register usage // A0 x3 v0 // A1 x15 v1 // A2 x13 v2 // A3 x4 v3 // B x5 v4 v5 v6 v7 // C0 x6 v16 v20 v24 v28 // C1 x8 v17 v21 v25 v29 // C2 x9 v18 v22 v26 v30 // C3 x7 v19 v23 v27 v31 // unused v14 v15 BEGIN_FUNCTION xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x16c4__asm_aarch64_neondot_ld64 # Clamp A and C pointers CMP x0, 2 // if mr < 2 ADD x2, x2, 3 // kc = (kc + 3) & ~3 ADD x15, x3, x4 // a1 = a0 + a_stride ADD x8, x6, x7 // c1 = c0 + cm_stride CSEL x15, x3, x15, LO // a1 = a0 CSEL x8, x6, x8, LO // c1 = c0 BIC x2, x2, 3 ADD x13, x15, x4 // a2 = a1 + a_stride ADD x9, x8, x7 // c2 = c1 + cm_stride // if mr <= 2 CSEL x13, x15, x13, LS // a2 = a1 CSEL x9, x8, x9, LS // c2 = c1 LDP x12, x11, [sp] // cn_stride, params LDR x16, [sp, 16] // &quantization_params[0].zero_point STP d8, d9, [sp, -48]! STP d10, d11, [sp, 16] STP d12, d13, [sp, 32] LDP q12, q13, [x16] // v12 & v13 interleaved zero_point & scale CMP x0, 4 // if mr < 4 ADD x4, x13, x4 // a3 = a2 + a_stride ADD x7, x9, x7 // c3 = c2 + cm_stride CSEL x4, x13, x4, LO // a3 = a2 CSEL x7, x9, x7, LO // c3 = c2 .p2align 3 0: # Load initial bias from w into accumulators SUBS x0, x2, 8 // k = kc - 8 LDP q0, q1, [x5], 32 MUL v16.4s, v0.4s, v12.s[0] MUL v17.4s, v0.4s, v12.s[2] MUL v18.4s, v0.4s, v13.s[0] LDP q2, q3, [x5], 32 MUL v19.4s, v0.4s, v13.s[2] MUL v20.4s, v1.4s, v12.s[0] MUL v21.4s, v1.4s, v12.s[2] MUL v22.4s, v1.4s, v13.s[0] MUL v23.4s, v1.4s, v13.s[2] MUL v24.4s, v2.4s, v12.s[0] MUL v25.4s, v2.4s, v12.s[2] MUL v26.4s, v2.4s, v13.s[0] MUL v27.4s, v2.4s, v13.s[2] MUL v28.4s, v3.4s, v12.s[0] MUL v29.4s, v3.4s, v12.s[2] MUL v30.4s, v3.4s, v13.s[0] MUL v31.4s, v3.4s, v13.s[2] # Is there at least 8 bytes? B.LO 3f # Main loop - 8 bytes of A .p2align 3 1: LDR d0, [x3], 8 LDR q4, [x5], 16 LDR d1, [x15], 8 LDR d2, [x13], 8 LDR d3, [x4], 8 LDR q5, [x5], 16 SDOT v16.4s, v4.16b, v0.4b[0] SDOT v17.4s, v4.16b, v1.4b[0] LDP q6, q7, [x5], 32 SDOT v18.4s, v4.16b, v2.4b[0] SDOT v19.4s, v4.16b, v3.4b[0] SDOT v20.4s, v5.16b, v0.4b[0] SDOT v21.4s, v5.16b, v1.4b[0] SDOT v22.4s, v5.16b, v2.4b[0] SDOT v23.4s, v5.16b, v3.4b[0] SDOT v24.4s, v6.16b, v0.4b[0] SDOT v25.4s, v6.16b, v1.4b[0] LDP q4, q5, [x5], 32 SDOT v26.4s, v6.16b, v2.4b[0] SDOT v27.4s, v6.16b, v3.4b[0] SDOT v28.4s, v7.16b, v0.4b[0] SDOT v29.4s, v7.16b, v1.4b[0] SDOT v30.4s, v7.16b, v2.4b[0] SDOT v31.4s, v7.16b, v3.4b[0] SDOT v16.4s, v4.16b, v0.4b[1] SDOT v17.4s, v4.16b, v1.4b[1] LDP q6, q7, [x5], 32 SDOT v18.4s, v4.16b, v2.4b[1] SDOT v19.4s, v4.16b, v3.4b[1] SDOT v20.4s, v5.16b, v0.4b[1] SDOT v21.4s, v5.16b, v1.4b[1] SDOT v22.4s, v5.16b, v2.4b[1] SDOT v23.4s, v5.16b, v3.4b[1] SDOT v24.4s, v6.16b, v0.4b[1] SDOT v25.4s, v6.16b, v1.4b[1] SDOT v26.4s, v6.16b, v2.4b[1] SDOT v27.4s, v6.16b, v3.4b[1] SDOT v28.4s, v7.16b, v0.4b[1] SDOT v29.4s, v7.16b, v1.4b[1] SDOT v30.4s, v7.16b, v2.4b[1] SUBS x0, x0, 8 SDOT v31.4s, v7.16b, v3.4b[1] B.HS 1b # Is there a remainder?- 4 bytes of A TBNZ x0, 2, 3f 2: LDP q0, q1, [x5], 32 // kernel_scale LDP q2, q3, [x5], 32 SCVTF v19.4s, v19.4s SCVTF v23.4s, v23.4s SCVTF v27.4s, v27.4s SCVTF v31.4s, v31.4s SCVTF v18.4s, v18.4s SCVTF v22.4s, v22.4s SCVTF v26.4s, v26.4s FMUL v8.4s, v0.4s, v13.s[3] // kernel_scale * scale FMUL v9.4s, v1.4s, v13.s[3] FMUL v10.4s, v2.4s, v13.s[3] FMUL v11.4s, v3.4s, v13.s[3] FMUL v4.4s, v0.4s, v13.s[1] FMUL v5.4s, v1.4s, v13.s[1] FMUL v6.4s, v2.4s, v13.s[1] FMUL v7.4s, v3.4s, v13.s[1] SCVTF v30.4s, v30.4s SCVTF v17.4s, v17.4s SCVTF v21.4s, v21.4s SCVTF v25.4s, v25.4s SCVTF v29.4s, v29.4s SCVTF v16.4s, v16.4s SCVTF v20.4s, v20.4s SCVTF v24.4s, v24.4s SCVTF v28.4s, v28.4s FMUL v19.4s, v19.4s, v8.4s FMUL v8.4s, v0.4s, v12.s[3] FMUL v23.4s, v23.4s, v9.4s FMUL v9.4s, v1.4s, v12.s[3] FMUL v27.4s, v27.4s, v10.4s FMUL v10.4s, v2.4s, v12.s[3] FMUL v31.4s, v31.4s, v11.4s FMUL v11.4s, v3.4s, v12.s[3] FMUL v18.4s, v18.4s, v4.4s FMUL v4.4s, v0.4s, v12.s[1] FMUL v22.4s, v22.4s, v5.4s FMUL v5.4s, v1.4s, v12.s[1] FMUL v26.4s, v26.4s, v6.4s FMUL v6.4s, v2.4s, v12.s[1] FMUL v30.4s, v30.4s, v7.4s FMUL v7.4s, v3.4s, v12.s[1] LDP q0, q1, [x5], 32 // bias LDP q2, q3, [x5], 32 FMUL v17.4s, v17.4s, v8.4s FMUL v21.4s, v21.4s, v9.4s FMUL v25.4s, v25.4s, v10.4s FMUL v29.4s, v29.4s, v11.4s FMUL v16.4s, v16.4s, v4.4s FMUL v20.4s, v20.4s, v5.4s FMUL v24.4s, v24.4s, v6.4s FMUL v28.4s, v28.4s, v7.4s LD2R {v4.4s, v5.4s}, [x11] // min max FADD v19.4s, v19.4s, v0.4s FADD v23.4s, v23.4s, v1.4s FADD v27.4s, v27.4s, v2.4s FADD v31.4s, v31.4s, v3.4s FADD v18.4s, v18.4s, v0.4s FADD v22.4s, v22.4s, v1.4s FADD v26.4s, v26.4s, v2.4s FADD v30.4s, v30.4s, v3.4s FADD v17.4s, v17.4s, v0.4s FADD v21.4s, v21.4s, v1.4s FADD v25.4s, v25.4s, v2.4s FADD v29.4s, v29.4s, v3.4s FADD v16.4s, v16.4s, v0.4s FADD v20.4s, v20.4s, v1.4s FADD v24.4s, v24.4s, v2.4s FADD v28.4s, v28.4s, v3.4s FMAX v19.4s, v19.4s, v4.4s FMAX v23.4s, v23.4s, v4.4s FMAX v27.4s, v27.4s, v4.4s FMAX v31.4s, v31.4s, v4.4s FMAX v18.4s, v18.4s, v4.4s FMAX v22.4s, v22.4s, v4.4s FMAX v26.4s, v26.4s, v4.4s FMAX v30.4s, v30.4s, v4.4s FMAX v17.4s, v17.4s, v4.4s FMAX v21.4s, v21.4s, v4.4s FMAX v25.4s, v25.4s, v4.4s FMAX v29.4s, v29.4s, v4.4s FMAX v16.4s, v16.4s, v4.4s FMAX v20.4s, v20.4s, v4.4s FMAX v24.4s, v24.4s, v4.4s FMAX v28.4s, v28.4s, v4.4s FMIN v19.4s, v19.4s, v5.4s FMIN v23.4s, v23.4s, v5.4s FMIN v27.4s, v27.4s, v5.4s FMIN v31.4s, v31.4s, v5.4s FMIN v18.4s, v18.4s, v5.4s FMIN v22.4s, v22.4s, v5.4s FMIN v26.4s, v26.4s, v5.4s FMIN v30.4s, v30.4s, v5.4s FMIN v17.4s, v17.4s, v5.4s FMIN v21.4s, v21.4s, v5.4s FMIN v25.4s, v25.4s, v5.4s FMIN v29.4s, v29.4s, v5.4s FMIN v16.4s, v16.4s, v5.4s FMIN v20.4s, v20.4s, v5.4s FMIN v24.4s, v24.4s, v5.4s FMIN v28.4s, v28.4s, v5.4s SUBS x1, x1, 16 B.LO 4f STP q19, q23, [x7] STP q27, q31, [x7, #32] ADD x7, x7, x12 STP q18, q22, [x9] STP q26, q30, [x9, #32] ADD x9, x9, x12 STP q17, q21, [x8] STP q25, q29, [x8, #32] ADD x8, x8, x12 STP q16, q20, [x6] STP q24, q28, [x6, #32] ADD x6, x6, x12 SUB x3, x3, x2 // a0 -= kc SUB x15, x15, x2 // a1 -= kc SUB x13, x13, x2 // a2 -= kc SUB x4, x4, x2 // a3 -= kc B.NE 0b # Restore d8-d13 from stack LDP d12, d13, [sp, 32] LDP d10, d11, [sp, 16] LDP d8, d9, [sp], 48 RET # Remainder- 4 bytes of A .p2align 3 3: LDR s0, [x3], 4 LDR q4, [x5], 16 LDR s1, [x15], 4 LDR s2, [x13], 4 LDR s3, [x4], 4 SDOT v16.4s, v4.16b, v0.4b[0] LDR q5, [x5], 16 SDOT v17.4s, v4.16b, v1.4b[0] SDOT v18.4s, v4.16b, v2.4b[0] SDOT v19.4s, v4.16b, v3.4b[0] SDOT v20.4s, v5.16b, v0.4b[0] LDP q6, q7, [x5], 32 SDOT v21.4s, v5.16b, v1.4b[0] SDOT v22.4s, v5.16b, v2.4b[0] SDOT v23.4s, v5.16b, v3.4b[0] SDOT v24.4s, v6.16b, v0.4b[0] SDOT v25.4s, v6.16b, v1.4b[0] SDOT v26.4s, v6.16b, v2.4b[0] SDOT v27.4s, v6.16b, v3.4b[0] SDOT v28.4s, v7.16b, v0.4b[0] SDOT v29.4s, v7.16b, v1.4b[0] SDOT v30.4s, v7.16b, v2.4b[0] SDOT v31.4s, v7.16b, v3.4b[0] B 2b # Store odd width .p2align 3 4: TBZ x1, 3, 5f STP q19, q23, [x7] STP q18, q22, [x9] MOV v19.16b, v27.16b MOV v23.16b, v31.16b MOV v18.16b, v26.16b MOV v22.16b, v30.16b STP q17, q21, [x8] STP q16, q20, [x6] MOV v17.16b, v25.16b MOV v21.16b, v29.16b MOV v16.16b, v24.16b MOV v20.16b, v28.16b ADD x6, x6, #32 ADD x7, x7, #32 ADD x8, x8, #32 ADD x9, x9, #32 5: TBZ x1, 2, 6f STR q19, [x7] STR q18, [x9] MOV v19.16b, v23.16b MOV v18.16b, v22.16b STR q17, [x8] STR q16, [x6] MOV v17.16b, v21.16b MOV v16.16b, v20.16b ADD x6, x6, #16 ADD x7, x7, #16 ADD x8, x8, #16 ADD x9, x9, #16 6: TBZ x1, 1, 7f STR d19, [x7], 8 STR d18, [x9], 8 DUP d19, v19.d[1] DUP d18, v18.d[1] STR d17, [x8], 8 STR d16, [x6], 8 DUP d17, v17.d[1] DUP d16, v16.d[1] 7: TBZ x1, 0, 8f STR s19, [x7] STR s18, [x9] STR s17, [x8] STR s16, [x6] 8: # Restore d8-d13 from stack LDP d12, d13, [sp, 32] LDP d10, d11, [sp, 16] LDP d8, d9, [sp], 48 RET END_FUNCTION xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x16c4__asm_aarch64_neondot_ld64 #ifdef __ELF__ .section ".note.GNU-stack","",%progbits #endif