diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp | 252 |
1 files changed, 126 insertions, 126 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp index 47de894306..ae14cd7d50 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,7 +48,6 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)), C(C), ldcb(ldc * sizeof(int32_t)), M(M), N(N), K(K), - n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2), bias(bias), accumulator_buffer(accumulator_buffer), @@ -69,10 +68,11 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in const long kstride_bytes; int32_t *const C; const long ldcb; - const long M, N, K, n_loops, n_tail_iters; + const long M, N, K; const int32_t *const bias; + int32_t *const accumulator_buffer; uint64_t flags; }; @@ -91,17 +91,17 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: Loop - ".inst 0xa040c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n" - ".inst 0xc0840500 // mova za0h.s[x12], { z8.s-z11.s }\n" - ".inst 0xa041c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840401 // mova za1h.s[x12], { z0.s-z3.s }\n" - ".inst 0xa042c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840582 // mova za2h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa043c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840403 // mova za3h.s[x12], { z0.s-z3.s }\n" + ".inst 0xa040c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n" + ".inst 0xa041c5fc // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xa042c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" + ".inst 0xa043c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840700 // mova za0h.s[x12], { z24.s-z27.s }\n" + "addvl x15, x15, #16\n" + ".inst 0xc0840781 // mova za1h.s[x12], { z28.s-z31.s }\n" + ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" + ".inst 0xc0840503 // mova za3h.s[x12], { z8.s-z11.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - "addvl x15, x15, #16\n" "blt 1b\n" "2:" // Initial accumulator load from buffer: End "ldr w13, [%x[args], %[offsetof_M]]\n" @@ -116,102 +116,102 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in "ldr x20, [%x[args], %[offsetof_bias]]\n" ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "cbz x20, 5f\n" - ".inst 0xa00a4295 // ldnt1w { z20.s-z21.s }, p8/Z, [x20, x10, LSL #2]\n" - ".inst 0xc0900280 // addha za0.s, p0/M, p0/M, z20.s\n" - ".inst 0xc09002a1 // addha za1.s, p0/M, p0/M, z21.s\n" - ".inst 0xc0900282 // addha za2.s, p0/M, p0/M, z20.s\n" - ".inst 0xc09002a3 // addha za3.s, p0/M, p0/M, z21.s\n" + ".inst 0xa10a4296 // ld1w { z22.s, z30.s }, p8/Z, [x20, x10, LSL #2]\n" + ".inst 0xc09002c0 // addha za0.s, p0/M, p0/M, z22.s\n" + ".inst 0xc09003c1 // addha za1.s, p0/M, p0/M, z30.s\n" + ".inst 0xc09002c2 // addha za2.s, p0/M, p0/M, z22.s\n" + ".inst 0xc09003c3 // addha za3.s, p0/M, p0/M, z30.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x10\n" "mov x21, x11\n" "incw x20, ALL, MUL #2\n" "incw x21, ALL, MUL #2\n" "cmp x20, x9\n" - "csel x21, x11, x21, LT\n" "mov x20, x16\n" + "csel x21, x11, x21, LT\n" "bfm x16, XZR, #0x0, #0x0 // bfc x16, #0x0, #0x1\n" "cmp x21, x13\n" "csel x16, x20, x16, LT\n" "5:" // Prepare accumulators: End "ldr x20, [%x[args], %[offsetof_K]]\n" + "ldr x23, [%x[args], %[offsetof_B]]\n" + "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n" "add x20, x20, #0x3\n" "lsr x20, x20, #0x2\n" - "ldr x23, [%x[args], %[offsetof_B]]\n" - "lsr x22, x20, #0x2\n" - "and x21, x20, #0x3\n" - "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n" - "madd x23, x10, x20, x23\n" // bptr = B + n * kstride_bytes - "cbz x22, 8f\n" - "subs x22, x22, #0x1\n" - ".inst 0xa040077c // ld1b { z28.b-z29.b }, pn9.b/Z, [x27]\n" - ".inst 0xa14006e8 // ldnt1b { z0.b, z8.b }, pn9.b/Z, [x23]\n" - ".inst 0xa0410762 // ld1b { z2.b-z3.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" - ".inst 0xa14106ff // ldnt1b { z23.b, z31.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xa042076e // ld1b { z14.b-z15.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa14206f8 // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa0430764 // ld1b { z4.b-z5.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" + "lsr x21, x20, #0x2\n" + "madd x23, x10, x22, x23\n" // bptr = B + n * kstride_bytes + "and x20, x20, #0x3\n" + "cbz x21, 8f\n" + "subs x21, x21, #0x1\n" + ".inst 0xa1400766 // ld1b { z6.b, z14.b }, pn9.b/Z, [x27]\n" + ".inst 0xa04006e2 // ld1b { z2.b-z3.b }, pn9.b/Z, [x23]\n" + ".inst 0xa041077a // ld1b { z26.b-z27.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" + ".inst 0xa04106f6 // ld1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa1420765 // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa14206f5 // ld1b { z21.b, z29.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa0430760 // ld1b { z0.b-z1.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" "addvl x27, x27, #8\n" - ".inst 0xa04306f5 // ldnt1b { z20.b-z21.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" + ".inst 0xa14306f1 // ld1b { z17.b, z25.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" "addvl x23, x23, #8\n" "ble 7f\n" "6:" // K loop - ".inst 0xa0800380 // smopa za0.s, p0/M, p0/M, z28.b, z0.b\n" - "subs x22, x22, #0x1\n" - ".inst 0xa0880381 // smopa za1.s, p0/M, p0/M, z28.b, z8.b\n" - ".inst 0xa08003a2 // smopa za2.s, p0/M, p0/M, z29.b, z0.b\n" - ".inst 0xa08803a3 // smopa za3.s, p0/M, p0/M, z29.b, z8.b\n" - ".inst 0xa040077c // ld1b { z28.b-z29.b }, pn9.b/Z, [x27]\n" - ".inst 0xa0970040 // smopa za0.s, p0/M, p0/M, z2.b, z23.b\n" - ".inst 0xa14006e8 // ldnt1b { z0.b, z8.b }, pn9.b/Z, [x23]\n" - ".inst 0xa09f0041 // smopa za1.s, p0/M, p0/M, z2.b, z31.b\n" - ".inst 0xa0970062 // smopa za2.s, p0/M, p0/M, z3.b, z23.b\n" - ".inst 0xa09f0063 // smopa za3.s, p0/M, p0/M, z3.b, z31.b\n" - ".inst 0xa0410762 // ld1b { z2.b-z3.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" - ".inst 0xa09001c0 // smopa za0.s, p0/M, p0/M, z14.b, z16.b\n" - ".inst 0xa14106ff // ldnt1b { z23.b, z31.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xa09801c1 // smopa za1.s, p0/M, p0/M, z14.b, z24.b\n" - ".inst 0xa09001e2 // smopa za2.s, p0/M, p0/M, z15.b, z16.b\n" - ".inst 0xa09801e3 // smopa za3.s, p0/M, p0/M, z15.b, z24.b\n" - ".inst 0xa042076e // ld1b { z14.b-z15.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa14206f8 // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa0940080 // smopa za0.s, p0/M, p0/M, z4.b, z20.b\n" - ".inst 0xa0950081 // smopa za1.s, p0/M, p0/M, z4.b, z21.b\n" - ".inst 0xa09400a2 // smopa za2.s, p0/M, p0/M, z5.b, z20.b\n" - ".inst 0xa09500a3 // smopa za3.s, p0/M, p0/M, z5.b, z21.b\n" - ".inst 0xa0430764 // ld1b { z4.b-z5.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" + ".inst 0xa08200c0 // smopa za0.s, p0/M, p0/M, z6.b, z2.b\n" + "subs x21, x21, #0x1\n" + ".inst 0xa08300c1 // smopa za1.s, p0/M, p0/M, z6.b, z3.b\n" + ".inst 0xa08201c2 // smopa za2.s, p0/M, p0/M, z14.b, z2.b\n" + ".inst 0xa08301c3 // smopa za3.s, p0/M, p0/M, z14.b, z3.b\n" + ".inst 0xa1400766 // ld1b { z6.b, z14.b }, pn9.b/Z, [x27]\n" + ".inst 0xa0960340 // smopa za0.s, p0/M, p0/M, z26.b, z22.b\n" + ".inst 0xa04006e2 // ld1b { z2.b-z3.b }, pn9.b/Z, [x23]\n" + ".inst 0xa0970341 // smopa za1.s, p0/M, p0/M, z26.b, z23.b\n" + ".inst 0xa0960362 // smopa za2.s, p0/M, p0/M, z27.b, z22.b\n" + ".inst 0xa0970363 // smopa za3.s, p0/M, p0/M, z27.b, z23.b\n" + ".inst 0xa041077a // ld1b { z26.b-z27.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" + ".inst 0xa09500a0 // smopa za0.s, p0/M, p0/M, z5.b, z21.b\n" + ".inst 0xa04106f6 // ld1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa09d00a1 // smopa za1.s, p0/M, p0/M, z5.b, z29.b\n" + ".inst 0xa09501a2 // smopa za2.s, p0/M, p0/M, z13.b, z21.b\n" + ".inst 0xa09d01a3 // smopa za3.s, p0/M, p0/M, z13.b, z29.b\n" + ".inst 0xa1420765 // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa14206f5 // ld1b { z21.b, z29.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa0910000 // smopa za0.s, p0/M, p0/M, z0.b, z17.b\n" + ".inst 0xa0990001 // smopa za1.s, p0/M, p0/M, z0.b, z25.b\n" + ".inst 0xa0910022 // smopa za2.s, p0/M, p0/M, z1.b, z17.b\n" + ".inst 0xa0990023 // smopa za3.s, p0/M, p0/M, z1.b, z25.b\n" + ".inst 0xa0430760 // ld1b { z0.b-z1.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" "addvl x27, x27, #8\n" - ".inst 0xa04306f5 // ldnt1b { z20.b-z21.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" + ".inst 0xa14306f1 // ld1b { z17.b, z25.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" "addvl x23, x23, #8\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0xa0800380 // smopa za0.s, p0/M, p0/M, z28.b, z0.b\n" - ".inst 0xa0880381 // smopa za1.s, p0/M, p0/M, z28.b, z8.b\n" - ".inst 0xa08003a2 // smopa za2.s, p0/M, p0/M, z29.b, z0.b\n" - ".inst 0xa08803a3 // smopa za3.s, p0/M, p0/M, z29.b, z8.b\n" - ".inst 0xa0970040 // smopa za0.s, p0/M, p0/M, z2.b, z23.b\n" - ".inst 0xa09f0041 // smopa za1.s, p0/M, p0/M, z2.b, z31.b\n" - ".inst 0xa0970062 // smopa za2.s, p0/M, p0/M, z3.b, z23.b\n" - ".inst 0xa09f0063 // smopa za3.s, p0/M, p0/M, z3.b, z31.b\n" - ".inst 0xa09001c0 // smopa za0.s, p0/M, p0/M, z14.b, z16.b\n" - ".inst 0xa09801c1 // smopa za1.s, p0/M, p0/M, z14.b, z24.b\n" - ".inst 0xa09001e2 // smopa za2.s, p0/M, p0/M, z15.b, z16.b\n" - ".inst 0xa09801e3 // smopa za3.s, p0/M, p0/M, z15.b, z24.b\n" - ".inst 0xa0940080 // smopa za0.s, p0/M, p0/M, z4.b, z20.b\n" - ".inst 0xa0950081 // smopa za1.s, p0/M, p0/M, z4.b, z21.b\n" - ".inst 0xa09400a2 // smopa za2.s, p0/M, p0/M, z5.b, z20.b\n" - ".inst 0xa09500a3 // smopa za3.s, p0/M, p0/M, z5.b, z21.b\n" + ".inst 0xa08200c0 // smopa za0.s, p0/M, p0/M, z6.b, z2.b\n" + ".inst 0xa08300c1 // smopa za1.s, p0/M, p0/M, z6.b, z3.b\n" + ".inst 0xa08201c2 // smopa za2.s, p0/M, p0/M, z14.b, z2.b\n" + ".inst 0xa08301c3 // smopa za3.s, p0/M, p0/M, z14.b, z3.b\n" + ".inst 0xa0960340 // smopa za0.s, p0/M, p0/M, z26.b, z22.b\n" + ".inst 0xa0970341 // smopa za1.s, p0/M, p0/M, z26.b, z23.b\n" + ".inst 0xa0960362 // smopa za2.s, p0/M, p0/M, z27.b, z22.b\n" + ".inst 0xa0970363 // smopa za3.s, p0/M, p0/M, z27.b, z23.b\n" + ".inst 0xa09500a0 // smopa za0.s, p0/M, p0/M, z5.b, z21.b\n" + ".inst 0xa09d00a1 // smopa za1.s, p0/M, p0/M, z5.b, z29.b\n" + ".inst 0xa09501a2 // smopa za2.s, p0/M, p0/M, z13.b, z21.b\n" + ".inst 0xa09d01a3 // smopa za3.s, p0/M, p0/M, z13.b, z29.b\n" + ".inst 0xa0910000 // smopa za0.s, p0/M, p0/M, z0.b, z17.b\n" + ".inst 0xa0990001 // smopa za1.s, p0/M, p0/M, z0.b, z25.b\n" + ".inst 0xa0910022 // smopa za2.s, p0/M, p0/M, z1.b, z17.b\n" + ".inst 0xa0990023 // smopa za3.s, p0/M, p0/M, z1.b, z25.b\n" "8:" // K oddments - "cbz x21, 10f\n" + "cbz x20, 10f\n" "9:" // K oddments: Loop - ".inst 0xa1400774 // ld1b { z20.b, z28.b }, pn9.b/Z, [x27]\n" - "subs x21, x21, #0x1\n" + ".inst 0xa1400777 // ld1b { z23.b, z31.b }, pn9.b/Z, [x27]\n" + "subs x20, x20, #0x1\n" "addvl x27, x27, #2\n" ".inst 0xa14006e7 // ld1b { z7.b, z15.b }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" - ".inst 0xa0870280 // smopa za0.s, p0/M, p0/M, z20.b, z7.b\n" - ".inst 0xa08f0281 // smopa za1.s, p0/M, p0/M, z20.b, z15.b\n" - ".inst 0xa0870382 // smopa za2.s, p0/M, p0/M, z28.b, z7.b\n" - ".inst 0xa08f0383 // smopa za3.s, p0/M, p0/M, z28.b, z15.b\n" + ".inst 0xa08702e0 // smopa za0.s, p0/M, p0/M, z23.b, z7.b\n" + ".inst 0xa08f02e1 // smopa za1.s, p0/M, p0/M, z23.b, z15.b\n" + ".inst 0xa08703e2 // smopa za2.s, p0/M, p0/M, z31.b, z7.b\n" + ".inst 0xa08f03e3 // smopa za3.s, p0/M, p0/M, z31.b, z15.b\n" "bgt 9b\n" "10:" // K oddments: End "tbz x16, #1, 14f\n" @@ -219,25 +219,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n" - ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" - ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" - ".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" - ".inst 0xa041c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n" - ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" - ".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n" - ".inst 0xa042c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840702 // mova za2h.s[x12], { z24.s-z27.s }\n" + ".inst 0xa040c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n" + ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" + ".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n" + ".inst 0xa041c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" + ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" + ".inst 0xa042c5f4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" ".inst 0xa043c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840700 // mova za0h.s[x12], { z24.s-z27.s }\n" + "addvl x15, x15, #16\n" + ".inst 0xc0840481 // mova za1h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa060c5dc // st1w { z28.s-z31.s }, pn9.b, [x14]\n" + ".inst 0xc0840682 // mova za2h.s[x12], { z20.s-z23.s }\n" + ".inst 0xa061c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n" ".inst 0xc0840603 // mova za3h.s[x12], { z16.s-z19.s }\n" "add x12, x12, #0x4\n" + ".inst 0xa062c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n" "cmp x12, x20\n" - ".inst 0xa060c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14]\n" - "addvl x15, x15, #16\n" - ".inst 0xa061c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x4, MUL VL]\n" - ".inst 0xa062c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa063c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 11b\n" "b 23f\n" @@ -245,16 +245,16 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in "mov x12, #0x0\n" "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop - ".inst 0xc086040c // mova { z12.s-z15.s }, za0h.s[x12]\n" + ".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n" ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" - ".inst 0xa060c5cc // st1w { z12.s-z15.s }, pn9.b, [x14]\n" - ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" - ".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" - ".inst 0xa061c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n" + ".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n" + ".inst 0xa060c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14]\n" "add x12, x12, #0x4\n" + ".inst 0xa061c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n" "cmp x12, x20\n" - ".inst 0xa062c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa062c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c5d8 // st1w { z24.s-z27.s }, pn9.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 13b\n" "b 23f\n" @@ -264,25 +264,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in "cntw x24\n" "ldr x23, [%x[args], %[offsetof_ldcb]]\n" "cmp x25, x24\n" + "mov x12, #0x0\n" "csel x22, x25, x24, LT\n" "add x26, x26, x10, LSL #2\n" // C += n "lsr x21, x22, #0x2\n" "madd x26, x11, x23, x26\n" // C += m * ldc - "mov x12, #0x0\n" "and x20, x22, #0x3\n" "cbz x21, 16f\n" "15:" // Store to output array: Accumulator row 0 loop - ".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n" - ".inst 0xc086042c // mova { z12.s-z15.s }, za1h.s[x12]\n" - ".inst 0xa1604344 // st1w { z4.s, z12.s }, p8, [x26]\n" - "add x26, x26, x23\n" - ".inst 0xa1604345 // st1w { z5.s, z13.s }, p8, [x26]\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n" + ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" "add x26, x26, x23\n" "add x12, x12, #0x4\n" - ".inst 0xa1604346 // st1w { z6.s, z14.s }, p8, [x26]\n" + ".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n" "add x26, x26, x23\n" "cmp x12, x21, LSL #2\n" - ".inst 0xa1604347 // st1w { z7.s, z15.s }, p8, [x26]\n" + ".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n" + "add x26, x26, x23\n" + ".inst 0xa1604353 // st1w { z19.s, z27.s }, p8, [x26]\n" "add x26, x26, x23\n" "blt 15b\n" "16:" // Store to output array: Accumulator row 0 oddments @@ -303,9 +303,9 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in "subs x25, x25, x22\n" "beq 21f\n" "cmp x25, x24\n" + "mov x12, #0x0\n" "csel x20, x25, x24, LT\n" "lsr x21, x20, #0x2\n" - "mov x12, #0x0\n" "and x20, x20, #0x3\n" "cbz x21, 19f\n" "18:" // Store to output array: Accumulator row 1 loop @@ -313,53 +313,53 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in ".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n" ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" "add x26, x26, x23\n" + "add x12, x12, #0x4\n" ".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n" "add x26, x26, x23\n" - "add x12, x12, #0x4\n" + "cmp x12, x21, LSL #2\n" ".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n" "add x26, x26, x23\n" - "cmp x12, x21, LSL #2\n" ".inst 0xa1604353 // st1w { z19.s, z27.s }, p8, [x26]\n" "add x26, x26, x23\n" "blt 18b\n" "19:" // Store to output array: Accumulator row 1 oddments "cbz x20, 20f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n" - ".inst 0xc0860468 // mova { z8.s-z11.s }, za3h.s[x12]\n" - ".inst 0xa1604340 // st1w { z0.s, z8.s }, p8, [x26]\n" + ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" + ".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n" + ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" "add x26, x26, x23\n" "beq 20f\n" "subs x20, x20, #0x1\n" - ".inst 0xa1604341 // st1w { z1.s, z9.s }, p8, [x26]\n" + ".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n" "add x26, x26, x23\n" "beq 20f\n" - ".inst 0xa1604342 // st1w { z2.s, z10.s }, p8, [x26]\n" + ".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n" "20:" // Store to output array: Accumulator row 1 oddments: End "21:" // Store to output array: End "tbz x16, #0, 23f\n" "mov x12, #0x0\n" "cntw x20\n" "22:" // Store to output array: Refill accumulators: Loop - ".inst 0xa040c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n" - ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa041c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840481 // mova za1h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa040c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n" + ".inst 0xa041c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" ".inst 0xa042c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840582 // mova za2h.s[x12], { z12.s-z15.s }\n" ".inst 0xa043c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840500 // mova za0h.s[x12], { z8.s-z11.s }\n" + "addvl x15, x15, #16\n" + ".inst 0xc0840401 // mova za1h.s[x12], { z0.s-z3.s }\n" + ".inst 0xc0840582 // mova za2h.s[x12], { z12.s-z15.s }\n" ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - "addvl x15, x15, #16\n" "blt 22b\n" "23:" // End block "incw x10, ALL, MUL #2\n" "cmp x10, x9\n" "blt 3b\n" "incw x11, ALL, MUL #2\n" - "cmp x11, x13\n" "mov x10, #0x0\n" + "cmp x11, x13\n" "mov x28, x27\n" "blt 3b\n" ".inst 0xd503467f // SMSTOP\n" |