From ba209750abc1ac7e42bab9fef5db284384d70fb3 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Thu, 15 Dec 2022 12:39:29 +0000 Subject: Update CPU kernels to remove x19 Resolves: COMPMID-5805 Signed-off-by: Michael Tyler Change-Id: I250f64531e209625e4ff176dd5a552c1c34bc484 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8909 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Gunes Bayir Reviewed-by: Viet-Hoa Do Benchmark: Arm Jenkins --- .../sme2_gemv_bf16fp32_dot_16VL/generic.cpp | 558 ++++++++++----------- 1 file changed, 279 insertions(+), 279 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp index 26861fb931..520eeedfec 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -10,16 +10,16 @@ * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * - * The above copyright notice and this permission notice shall be included in - * all copies or substantial portions of the Software. + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING - * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS - * IN THE SOFTWARE. + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. */ #ifdef __ARM_FEATURE_SVE #ifdef ARM_COMPUTE_ENABLE_SME2 @@ -64,487 +64,487 @@ void sme2_gemv_bf16fp32_dot_16VL ( __asm__ __volatile__( "ptrue p1.b\n" ".inst 0xd503477f // SMSTART ZA\n" - "cntw x27, ALL, MUL #4\n" - "add x26, %x[N], x27\n" - "sub x26, x26, #0x1\n" - "udiv x26, x26, x27\n" - "add x21, x26, #0x3\n" - "and x21, x21, #0xfffffffffffffffc\n" - "mul x21, x21, x27\n" - "mul x21, x21, %x[K]\n" + "cntw x28, ALL, MUL #4\n" + "add x27, %x[N], x28\n" + "sub x27, x27, #0x1\n" + "udiv x27, x27, x28\n" + "add x22, x27, #0x3\n" + "and x22, x22, #0xfffffffffffffffc\n" + "mul x22, x22, x28\n" + "mul x22, x22, %x[K]\n" "mov x9, #0x0\n" - "mov x25, %x[B_ptr]\n" - "mov x24, %x[output_ptr]\n" + "mov x26, %x[B_ptr]\n" + "mov x25, %x[output_ptr]\n" "ptrue p1.b\n" ".inst 0x25207811 // ptrue pn9.b\n" - "lsl x21, x21, #0x1\n" - "mov x20, #0x1\n" + "lsl x22, x22, #0x1\n" + "mov x21, #0x1\n" "1:" // RHS size check loop - "cmp x21, #0x200000\n" + "cmp x22, #0x200000\n" "blt 2f\n" - "tbnz x21, #0, 3f\n" - "lsr x21, x21, #0x1\n" - "lsl x20, x20, #0x1\n" + "tbnz x22, #0, 3f\n" + "lsr x22, x22, #0x1\n" + "lsl x21, x21, #0x1\n" "b 1b\n" "2:" // RHS do prefetch - "lsl x19, x21, #0x26\n" - "sub x20, x20, #0x1\n" - "lsl x20, x20, #0x16\n" - "orr x21, x21, x19\n" - "orr x21, x21, x20\n" - ".inst 0xf8b54b3a // rprfm pldonce, x21, [x25]\n" + "lsl x20, x22, #0x26\n" + "sub x21, x21, #0x1\n" + "lsl x21, x21, #0x16\n" + "orr x22, x22, x20\n" + "orr x22, x22, x21\n" + ".inst 0xf8b64b5a // rprfm pldonce, x22, [x26]\n" "3:" // RHS prefetch exit - "mov x23, %x[bias]\n" + "mov x24, %x[bias]\n" "4:" // Column loop - "cmp x26, #0x4\n" + "cmp x27, #0x4\n" "bge 28f\n" - "cmp x26, #0x2\n" + "cmp x27, #0x2\n" "bgt 20f\n" "beq 12f\n" - "mov x22, %x[A_ptr]\n" - "lsl x21, %x[K], #0x1\n" - "mov x19, %x[N]\n" - "mov x20, %x[K]\n" - ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n" - ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n" - "cbz x23, 5f\n" - ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n" + "mov x23, %x[A_ptr]\n" + "lsl x22, %x[K], #0x1\n" + "mov x20, %x[N]\n" + "mov x21, %x[K]\n" + ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" + "cbz x24, 5f\n" + ".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n" ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" "b 6f\n" "5:" // Width 1: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "6:" // Width 1: setup done - "cmp x20, #0x8\n" + "cmp x21, #0x8\n" "ble 8f\n" "7:" // Width 1: Multiply loop: Main loop head - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "sub x20, x20, #0x8\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "sub x21, x21, #0x8\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" - "cmp x20, #0x8\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + "cmp x21, #0x8\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" - "add x22, x22, #0x10\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + "add x23, x23, #0x10\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "bgt 7b\n" "8:" // Width 1: Multiply loop: Single iteration only - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "subs x20, x20, #0x2\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" - "add x22, x22, #0x10\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "subs x21, x21, #0x2\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + "add x23, x23, #0x10\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 9f\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 9f\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 9f\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "9:" // Width 1: Multiply loop: multiply skip "tbz %x[flags], #1, 10f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x19, %x[args_ptr], %[offset_max]\n" + "add x21, %x[args_ptr], %[offset_min]\n" + "add x20, %x[args_ptr], %[offset_max]\n" ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - "ld1rw { z0.s }, p1/Z, [x20]\n" - "ld1rw { z6.s }, p1/Z, [x19]\n" + "ld1rw { z0.s }, p1/Z, [x21]\n" + "ld1rw { z6.s }, p1/Z, [x20]\n" ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n" - ".inst 0xa060c308 // st1w { z8.s-z11.s }, p8, [x24]\n" - "addvl x24, x24, #4\n" + ".inst 0xa060c328 // st1w { z8.s-z11.s }, p8, [x25]\n" + "addvl x25, x25, #4\n" "b 11f\n" "10:" // Width 1: No activation ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - ".inst 0xa060c308 // st1w { z8.s-z11.s }, p8, [x24]\n" - "addvl x24, x24, #4\n" + ".inst 0xa060c328 // st1w { z8.s-z11.s }, p8, [x25]\n" + "addvl x25, x25, #4\n" "11:" // Width 1: Output done "b 36f\n" "12:" // Width 2 - "mov x22, %x[A_ptr]\n" - "lsl x21, %x[K], #0x1\n" - "sub x19, %x[N], x27\n" - "mov x20, %x[K]\n" - ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n" - ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n" - "cbz x23, 13f\n" - ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n" + "mov x23, %x[A_ptr]\n" + "lsl x22, %x[K], #0x1\n" + "sub x20, %x[N], x28\n" + "mov x21, %x[K]\n" + ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" + "cbz x24, 13f\n" + ".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n" ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" - ".inst 0xa041c6e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa041c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" ".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n" "b 14f\n" "13:" // Width 2: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "14:" // Width 2: setup done - "cmp x20, #0x8\n" + "cmp x21, #0x8\n" "ble 16f\n" "15:" // Width 2: Multiply loop: Main loop head - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "sub x20, x20, #0x8\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "sub x21, x21, #0x8\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - "cmp x20, #0x8\n" - "add x22, x22, #0x10\n" - ".inst 0xa041a725 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + "cmp x21, #0x8\n" + "add x23, x23, #0x10\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15aba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa041a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15abf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "bgt 15b\n" "16:" // Width 2: Multiply loop: Single iteration only - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "subs x20, x20, #0x2\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" - "add x22, x22, #0x10\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "subs x21, x21, #0x2\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + "add x23, x23, #0x10\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - ".inst 0xa041a725 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 17f\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 17f\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15aba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 17f\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa041a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15abf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "17:" // Width 2: Multiply loop: multiply skip "tbz %x[flags], #1, 18f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x19, %x[args_ptr], %[offset_max]\n" + "add x21, %x[args_ptr], %[offset_min]\n" + "add x20, %x[args_ptr], %[offset_max]\n" ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - "ld1rw { z0.s }, p1/Z, [x20]\n" + "ld1rw { z0.s }, p1/Z, [x21]\n" ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n" - "ld1rw { z6.s }, p1/Z, [x19]\n" + "ld1rw { z6.s }, p1/Z, [x20]\n" ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n" - ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" ".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n" - ".inst 0xa061c314 // st1w { z20.s-z23.s }, p8, [x24, #0x4, MUL VL]\n" - "addvl x24, x24, #8\n" + ".inst 0xa061c334 // st1w { z20.s-z23.s }, p8, [x25, #0x4, MUL VL]\n" + "addvl x25, x25, #8\n" "b 19f\n" "18:" // Width 2: No activation ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n" - ".inst 0xa061c314 // st1w { z20.s-z23.s }, p8, [x24, #0x4, MUL VL]\n" - "addvl x24, x24, #8\n" + ".inst 0xa061c334 // st1w { z20.s-z23.s }, p8, [x25, #0x4, MUL VL]\n" + "addvl x25, x25, #8\n" "19:" // Width 2: Output done "b 36f\n" "20:" // Width 3 - "mov x19, #0x2\n" - "mov x22, %x[A_ptr]\n" - "lsl x21, %x[K], #0x1\n" - "msub x19, x27, x19, %x[N]\n" - "mov x20, %x[K]\n" - ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n" - ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n" - "cbz x23, 21f\n" - ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n" + "mov x20, #0x2\n" + "mov x23, %x[A_ptr]\n" + "lsl x22, %x[K], #0x1\n" + "msub x20, x28, x20, %x[N]\n" + "mov x21, %x[K]\n" + ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" + "cbz x24, 21f\n" + ".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n" ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" - ".inst 0xa041c6e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa041c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" ".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n" - ".inst 0xa042c6e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x23, #0x8, MUL VL]\n" + ".inst 0xa042c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n" "b 22f\n" "21:" // Width 3: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "22:" // Width 3: setup done - "cmp x20, #0x8\n" + "cmp x21, #0x8\n" "ble 24f\n" "23:" // Width 3: Multiply loop: Main loop head - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "sub x20, x20, #0x8\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "sub x21, x21, #0x8\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - "cmp x20, #0x8\n" - "add x22, x22, #0x10\n" - ".inst 0xa041a725 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + "cmp x21, #0x8\n" + "add x23, x23, #0x10\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z10.h[0]\n" - ".inst 0xa042a735 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab29a // bfdot za.s[x9, 2], { z20.h-z23.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[1]\n" - ".inst 0xa042a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab71a // bfdot za.s[x9, 2], { z24.h-z27.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15aba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[2]\n" - ".inst 0xa042a73d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abb9a // bfdot za.s[x9, 2], { z28.h-z31.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa041a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15abf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z10.h[3]\n" - ".inst 0xa042a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "bgt 23b\n" "24:" // Width 3: Multiply loop: Single iteration only - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "subs x20, x20, #0x2\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" - "add x22, x22, #0x10\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "subs x21, x21, #0x2\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + "add x23, x23, #0x10\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - ".inst 0xa041a725 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z10.h[0]\n" - ".inst 0xa042a735 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab29a // bfdot za.s[x9, 2], { z20.h-z23.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 25f\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[1]\n" - ".inst 0xa042a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab71a // bfdot za.s[x9, 2], { z24.h-z27.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 25f\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15aba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[2]\n" - ".inst 0xa042a73d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abb9a // bfdot za.s[x9, 2], { z28.h-z31.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 25f\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa041a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15abf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z10.h[3]\n" - ".inst 0xa042a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "25:" // Width 3: Multiply loop: multiply skip "tbz %x[flags], #1, 26f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x19, %x[args_ptr], %[offset_max]\n" + "add x21, %x[args_ptr], %[offset_min]\n" + "add x20, %x[args_ptr], %[offset_max]\n" ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - "ld1rw { z0.s }, p1/Z, [x20]\n" + "ld1rw { z0.s }, p1/Z, [x21]\n" ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n" - "ld1rw { z6.s }, p1/Z, [x19]\n" + "ld1rw { z6.s }, p1/Z, [x20]\n" ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n" ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" - ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" ".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n" - ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n" + ".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n" ".inst 0xc1a6c810 // fclamp { z16.s-z19.s }, z0.s, z6.s\n" - ".inst 0xa062c310 // st1w { z16.s-z19.s }, p8, [x24, #0x8, MUL VL]\n" - "addvl x24, x24, #12\n" + ".inst 0xa062c330 // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n" + "addvl x25, x25, #12\n" "b 27f\n" "26:" // Width 3: No activation ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n" - ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n" + ".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n" ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" - ".inst 0xa062c310 // st1w { z16.s-z19.s }, p8, [x24, #0x8, MUL VL]\n" - "addvl x24, x24, #12\n" + ".inst 0xa062c330 // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n" + "addvl x25, x25, #12\n" "27:" // Width 3: Output done "b 36f\n" "28:" // Width 4 - "mov x19, #0x3\n" - "mov x22, %x[A_ptr]\n" - "lsl x21, %x[K], #0x1\n" - "msub x19, x27, x19, %x[N]\n" - "mov x20, %x[K]\n" - ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n" - ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n" - "cbz x23, 29f\n" - ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n" + "mov x20, #0x3\n" + "mov x23, %x[A_ptr]\n" + "lsl x22, %x[K], #0x1\n" + "msub x20, x28, x20, %x[N]\n" + "mov x21, %x[K]\n" + ".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n" + ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" + "cbz x24, 29f\n" + ".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n" ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" - ".inst 0xa041c6e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa041c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" ".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n" - ".inst 0xa042c6e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x23, #0x8, MUL VL]\n" + ".inst 0xa042c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n" - ".inst 0xa043c6f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x23, #0xc, MUL VL]\n" + ".inst 0xa043c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n" ".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n" - "addvl x23, x23, #16\n" + "addvl x24, x24, #16\n" "b 30f\n" "29:" // Width 4: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "30:" // Width 4: setup done - "cmp x20, #0x8\n" + "cmp x21, #0x8\n" "ble 32f\n" "31:" // Width 4: Multiply loop: Main loop head - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "sub x20, x20, #0x8\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "sub x21, x21, #0x8\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - "cmp x20, #0x8\n" - "add x22, x22, #0x10\n" - ".inst 0xa041a725 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + "cmp x21, #0x8\n" + "add x23, x23, #0x10\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z10.h[0]\n" - ".inst 0xa042a735 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab29a // bfdot za.s[x9, 2], { z20.h-z23.h }, z10.h[0]\n" - ".inst 0xa043a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15ab21b // bfdot za.s[x9, 3], { z16.h-z19.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[1]\n" - ".inst 0xa042a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab71a // bfdot za.s[x9, 2], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa043a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15ab61b // bfdot za.s[x9, 3], { z16.h-z19.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15aba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[2]\n" - ".inst 0xa042a73d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abb9a // bfdot za.s[x9, 2], { z28.h-z31.h }, z10.h[2]\n" - ".inst 0xa043a735 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15aba9b // bfdot za.s[x9, 3], { z20.h-z23.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + "addvl x26, x26, #16\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa041a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15abf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z10.h[3]\n" - ".inst 0xa042a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa043a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15abe1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "bgt 31b\n" "32:" // Width 4: Multiply loop: Single iteration only - "whilelt p0.h, XZR, x20\n" - "ld1rqh { z10.h }, p0/Z, [x22]\n" - "subs x20, x20, #0x2\n" - ".inst 0xa040a721 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x25]\n" - "add x22, x22, #0x10\n" + "whilelt p0.h, XZR, x21\n" + "ld1rqh { z10.h }, p0/Z, [x23]\n" + "subs x21, x21, #0x2\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + "add x23, x23, #0x10\n" ".inst 0xc15ab018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[0]\n" - ".inst 0xa041a725 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z10.h[0]\n" - ".inst 0xa042a735 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab29a // bfdot za.s[x9, 2], { z20.h-z23.h }, z10.h[0]\n" - ".inst 0xa043a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15ab21b // bfdot za.s[x9, 3], { z16.h-z19.h }, z10.h[0]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 33f\n" - ".inst 0xa040a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15ab619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[1]\n" - ".inst 0xa042a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15ab71a // bfdot za.s[x9, 2], { z24.h-z27.h }, z10.h[1]\n" - ".inst 0xa043a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15ab61b // bfdot za.s[x9, 3], { z16.h-z19.h }, z10.h[1]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 33f\n" - ".inst 0xa040a72d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x25]\n" - "subs x20, x20, #0x2\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + "subs x21, x21, #0x2\n" ".inst 0xc15ab998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[2]\n" - ".inst 0xa041a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15aba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z10.h[2]\n" - ".inst 0xa042a73d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abb9a // bfdot za.s[x9, 2], { z28.h-z31.h }, z10.h[2]\n" - ".inst 0xa043a735 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15aba9b // bfdot za.s[x9, 3], { z20.h-z23.h }, z10.h[2]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "ble 33f\n" - ".inst 0xa040a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" ".inst 0xc15abe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa041a739 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x25, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" ".inst 0xc15abf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z10.h[3]\n" - ".inst 0xa042a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0x8, MUL VL]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0xc15abe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z10.h[3]\n" - ".inst 0xa043a731 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x25, #0xc, MUL VL]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" ".inst 0xc15abe1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z10.h[3]\n" - "addvl x25, x25, #16\n" + "addvl x26, x26, #16\n" "33:" // Width 4: Multiply loop: multiply skip "tbz %x[flags], #1, 34f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x19, %x[args_ptr], %[offset_max]\n" + "add x21, %x[args_ptr], %[offset_min]\n" + "add x20, %x[args_ptr], %[offset_max]\n" ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - "ld1rw { z0.s }, p1/Z, [x20]\n" + "ld1rw { z0.s }, p1/Z, [x21]\n" ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n" - "ld1rw { z6.s }, p1/Z, [x19]\n" + "ld1rw { z6.s }, p1/Z, [x20]\n" ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n" ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" - ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" ".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n" ".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n" - ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n" + ".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n" ".inst 0xc1a6c810 // fclamp { z16.s-z19.s }, z0.s, z6.s\n" - ".inst 0xa062c710 // st1w { z16.s-z19.s }, pn9.b, [x24, #0x8, MUL VL]\n" + ".inst 0xa062c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xc1a6c818 // fclamp { z24.s-z27.s }, z0.s, z6.s\n" - ".inst 0xa063c318 // st1w { z24.s-z27.s }, p8, [x24, #0xc, MUL VL]\n" - "addvl x24, x24, #16\n" + ".inst 0xa063c338 // st1w { z24.s-z27.s }, p8, [x25, #0xc, MUL VL]\n" + "addvl x25, x25, #16\n" "b 35f\n" "34:" // Width 4: No activation ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n" - ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n" + ".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n" ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" - ".inst 0xa062c710 // st1w { z16.s-z19.s }, pn9.b, [x24, #0x8, MUL VL]\n" + ".inst 0xa062c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n" - ".inst 0xa063c318 // st1w { z24.s-z27.s }, p8, [x24, #0xc, MUL VL]\n" - "addvl x24, x24, #16\n" + ".inst 0xa063c338 // st1w { z24.s-z27.s }, p8, [x25, #0xc, MUL VL]\n" + "addvl x25, x25, #16\n" "35:" // Width 4: Output done - "subs x26, x26, #0x4\n" - "sub %x[N], %x[N], x27, LSL #2\n" + "subs x27, x27, #0x4\n" + "sub %x[N], %x[N], x28, LSL #2\n" "bgt 4b\n" "36:" // Exit ".inst 0xd503467f // SMSTOP\n" "ptrue p1.b\n" : [N] "+&r" (N) : [A_ptr] "r" (A_ptr), [B_ptr] "r" (B_ptr), [K] "r" (K), [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [output_ptr] "r" (output_ptr) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1