diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp | 318 |
1 files changed, 159 insertions, 159 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp index 45fcc7a860..1e46aee27a 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp @@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa B(B), kstride_bytes(K * sizeof(float)), C(C), ldcb(ldc * sizeof(float)), M(M), N(N), K(K), + n_loops((K - 1) / 2), n_tail_iters((K - 1) % 2), min(-std::numeric_limits<float>::infinity()), max(std::numeric_limits<float>::infinity()), bias(bias), @@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa const long kstride_bytes; float *const C; const long ldcb; - const long M, N, K; + const long M, N, K, n_loops, n_tail_iters; float min = -std::numeric_limits<float>::infinity(); float max = std::numeric_limits<float>::infinity(); const float *const bias; - float *const accumulator_buffer; uint64_t flags; }; @@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: Loop - ".inst 0xa040c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n" - ".inst 0xa041c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xa042c5e0 // ld1w { z0.s-z3.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 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n" - ".inst 0xc0840402 // mova za2h.s[x12], { z0.s-z3.s }\n" - ".inst 0xc0840603 // mova za3h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa040c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15]\n" + ".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa041c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840501 // mova za1h.s[x12], { z8.s-z11.s }\n" + ".inst 0xa042c5f4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" + ".inst 0xc0840682 // mova za2h.s[x12], { z20.s-z23.s }\n" + ".inst 0xa043c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\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" @@ -137,101 +137,101 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "ldr x20, [%x[args], %[offsetof_bias]]\n" ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "cbz x20, 5f\n" - "fmov z17.s, #1.0\n" - ".inst 0xa00a428a // ld1w { z10.s-z11.s }, p8/Z, [x20, x10, LSL #2]\n" - ".inst 0x808a0220 // fmopa za0.s, p0/M, p0/M, z17.s, z10.s\n" - ".inst 0x808b0221 // fmopa za1.s, p0/M, p0/M, z17.s, z11.s\n" - ".inst 0x808a0222 // fmopa za2.s, p0/M, p0/M, z17.s, z10.s\n" - ".inst 0x808b0223 // fmopa za3.s, p0/M, p0/M, z17.s, z11.s\n" + "fmov z12.s, #1.0\n" + ".inst 0xa10a4289 // ldnt1w { z1.s, z9.s }, p8/Z, [x20, x10, LSL #2]\n" + ".inst 0x80810180 // fmopa za0.s, p0/M, p0/M, z12.s, z1.s\n" + ".inst 0x80890181 // fmopa za1.s, p0/M, p0/M, z12.s, z9.s\n" + ".inst 0x80810182 // fmopa za2.s, p0/M, p0/M, z12.s, z1.s\n" + ".inst 0x80890183 // fmopa za3.s, p0/M, p0/M, z12.s, z9.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" - "mov x20, x16\n" "csel x21, x11, x21, LT\n" + "mov x20, x16\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" - "lsr x21, x20, #0x2\n" - "and x20, x20, #0x3\n" - "madd x23, x10, x22, x23\n" // bptr = B + n * kstride_bytes - "cbz x21, 8f\n" - "subs x21, x21, #0x1\n" - ".inst 0xa0404776 // ld1w { z22.s-z23.s }, pn9.b/Z, [x27]\n" - ".inst 0xa14046e7 // ld1w { z7.s, z15.s }, pn9.b/Z, [x23]\n" - ".inst 0xa1414766 // ld1w { z6.s, z14.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n" - ".inst 0xa04146f4 // ld1w { z20.s-z21.s }, pn9.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xa1424762 // ld1w { z2.s, z10.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa14246e3 // ld1w { z3.s, z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa1434761 // ld1w { z1.s, z9.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n" + "lsr x23, x20, #0x2\n" + "and x22, x20, #0x3\n" + "ldr x21, [%x[args], %[offsetof_B]]\n" + "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n" + "madd x21, x10, x20, x21\n" // bptr = B + n * kstride_bytes + "cbz x23, 8f\n" + "subs x23, x23, #0x1\n" + ".inst 0xa0404772 // ld1w { z18.s-z19.s }, pn9.b/Z, [x27]\n" + ".inst 0xa04046a3 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x21]\n" + ".inst 0xa0414764 // ld1w { z4.s-z5.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n" + ".inst 0xa04146bb // ldnt1w { z26.s-z27.s }, pn9.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0xa042476a // ld1w { z10.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa04246b5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0xa0434766 // ld1w { z6.s-z7.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n" "addvl x27, x27, #8\n" - ".inst 0xa04346e4 // ld1w { z4.s-z5.s }, pn9.b/Z, [x23, #0x6, MUL VL]\n" - "addvl x23, x23, #8\n" + ".inst 0xa04346a9 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x21, #0x6, MUL VL]\n" + "addvl x21, x21, #8\n" "ble 7f\n" "6:" // K loop - ".inst 0x808702c0 // fmopa za0.s, p0/M, p0/M, z22.s, z7.s\n" - "subs x21, x21, #0x1\n" - ".inst 0x808f02c1 // fmopa za1.s, p0/M, p0/M, z22.s, z15.s\n" - ".inst 0x808702e2 // fmopa za2.s, p0/M, p0/M, z23.s, z7.s\n" - ".inst 0x808f02e3 // fmopa za3.s, p0/M, p0/M, z23.s, z15.s\n" - ".inst 0xa0404776 // ld1w { z22.s-z23.s }, pn9.b/Z, [x27]\n" - ".inst 0x809400c0 // fmopa za0.s, p0/M, p0/M, z6.s, z20.s\n" - ".inst 0xa14046e7 // ld1w { z7.s, z15.s }, pn9.b/Z, [x23]\n" - ".inst 0x809500c1 // fmopa za1.s, p0/M, p0/M, z6.s, z21.s\n" - ".inst 0x809401c2 // fmopa za2.s, p0/M, p0/M, z14.s, z20.s\n" - ".inst 0x809501c3 // fmopa za3.s, p0/M, p0/M, z14.s, z21.s\n" - ".inst 0xa1414766 // ld1w { z6.s, z14.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n" - ".inst 0x80830040 // fmopa za0.s, p0/M, p0/M, z2.s, z3.s\n" - ".inst 0xa04146f4 // ld1w { z20.s-z21.s }, pn9.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0x808b0041 // fmopa za1.s, p0/M, p0/M, z2.s, z11.s\n" - ".inst 0x80830142 // fmopa za2.s, p0/M, p0/M, z10.s, z3.s\n" - ".inst 0x808b0143 // fmopa za3.s, p0/M, p0/M, z10.s, z11.s\n" - ".inst 0xa1424762 // ld1w { z2.s, z10.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa14246e3 // ld1w { z3.s, z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0x80840020 // fmopa za0.s, p0/M, p0/M, z1.s, z4.s\n" - ".inst 0x80850021 // fmopa za1.s, p0/M, p0/M, z1.s, z5.s\n" - ".inst 0x80840122 // fmopa za2.s, p0/M, p0/M, z9.s, z4.s\n" - ".inst 0x80850123 // fmopa za3.s, p0/M, p0/M, z9.s, z5.s\n" - ".inst 0xa1434761 // ld1w { z1.s, z9.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n" + ".inst 0x80820240 // fmopa za0.s, p0/M, p0/M, z18.s, z2.s\n" + "subs x23, x23, #0x1\n" + ".inst 0x80830241 // fmopa za1.s, p0/M, p0/M, z18.s, z3.s\n" + ".inst 0x80820262 // fmopa za2.s, p0/M, p0/M, z19.s, z2.s\n" + ".inst 0x80830263 // fmopa za3.s, p0/M, p0/M, z19.s, z3.s\n" + ".inst 0xa0404772 // ld1w { z18.s-z19.s }, pn9.b/Z, [x27]\n" + ".inst 0x809a0080 // fmopa za0.s, p0/M, p0/M, z4.s, z26.s\n" + ".inst 0xa04046a3 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x21]\n" + ".inst 0x809b0081 // fmopa za1.s, p0/M, p0/M, z4.s, z27.s\n" + ".inst 0x809a00a2 // fmopa za2.s, p0/M, p0/M, z5.s, z26.s\n" + ".inst 0x809b00a3 // fmopa za3.s, p0/M, p0/M, z5.s, z27.s\n" + ".inst 0xa0414764 // ld1w { z4.s-z5.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n" + ".inst 0x80940140 // fmopa za0.s, p0/M, p0/M, z10.s, z20.s\n" + ".inst 0xa04146bb // ldnt1w { z26.s-z27.s }, pn9.b/Z, [x21, #0x2, MUL VL]\n" + ".inst 0x80950141 // fmopa za1.s, p0/M, p0/M, z10.s, z21.s\n" + ".inst 0x80940162 // fmopa za2.s, p0/M, p0/M, z11.s, z20.s\n" + ".inst 0x80950163 // fmopa za3.s, p0/M, p0/M, z11.s, z21.s\n" + ".inst 0xa042476a // ld1w { z10.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa04246b5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0x808800c0 // fmopa za0.s, p0/M, p0/M, z6.s, z8.s\n" + ".inst 0x808900c1 // fmopa za1.s, p0/M, p0/M, z6.s, z9.s\n" + ".inst 0x808800e2 // fmopa za2.s, p0/M, p0/M, z7.s, z8.s\n" + ".inst 0x808900e3 // fmopa za3.s, p0/M, p0/M, z7.s, z9.s\n" + ".inst 0xa0434766 // ld1w { z6.s-z7.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n" "addvl x27, x27, #8\n" - ".inst 0xa04346e4 // ld1w { z4.s-z5.s }, pn9.b/Z, [x23, #0x6, MUL VL]\n" - "addvl x23, x23, #8\n" + ".inst 0xa04346a9 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x21, #0x6, MUL VL]\n" + "addvl x21, x21, #8\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0x808702c0 // fmopa za0.s, p0/M, p0/M, z22.s, z7.s\n" - ".inst 0x808f02c1 // fmopa za1.s, p0/M, p0/M, z22.s, z15.s\n" - ".inst 0x808702e2 // fmopa za2.s, p0/M, p0/M, z23.s, z7.s\n" - ".inst 0x808f02e3 // fmopa za3.s, p0/M, p0/M, z23.s, z15.s\n" - ".inst 0x809400c0 // fmopa za0.s, p0/M, p0/M, z6.s, z20.s\n" - ".inst 0x809500c1 // fmopa za1.s, p0/M, p0/M, z6.s, z21.s\n" - ".inst 0x809401c2 // fmopa za2.s, p0/M, p0/M, z14.s, z20.s\n" - ".inst 0x809501c3 // fmopa za3.s, p0/M, p0/M, z14.s, z21.s\n" - ".inst 0x80830040 // fmopa za0.s, p0/M, p0/M, z2.s, z3.s\n" - ".inst 0x808b0041 // fmopa za1.s, p0/M, p0/M, z2.s, z11.s\n" - ".inst 0x80830142 // fmopa za2.s, p0/M, p0/M, z10.s, z3.s\n" - ".inst 0x808b0143 // fmopa za3.s, p0/M, p0/M, z10.s, z11.s\n" - ".inst 0x80840020 // fmopa za0.s, p0/M, p0/M, z1.s, z4.s\n" - ".inst 0x80850021 // fmopa za1.s, p0/M, p0/M, z1.s, z5.s\n" - ".inst 0x80840122 // fmopa za2.s, p0/M, p0/M, z9.s, z4.s\n" - ".inst 0x80850123 // fmopa za3.s, p0/M, p0/M, z9.s, z5.s\n" + ".inst 0x80820240 // fmopa za0.s, p0/M, p0/M, z18.s, z2.s\n" + ".inst 0x80830241 // fmopa za1.s, p0/M, p0/M, z18.s, z3.s\n" + ".inst 0x80820262 // fmopa za2.s, p0/M, p0/M, z19.s, z2.s\n" + ".inst 0x80830263 // fmopa za3.s, p0/M, p0/M, z19.s, z3.s\n" + ".inst 0x809a0080 // fmopa za0.s, p0/M, p0/M, z4.s, z26.s\n" + ".inst 0x809b0081 // fmopa za1.s, p0/M, p0/M, z4.s, z27.s\n" + ".inst 0x809a00a2 // fmopa za2.s, p0/M, p0/M, z5.s, z26.s\n" + ".inst 0x809b00a3 // fmopa za3.s, p0/M, p0/M, z5.s, z27.s\n" + ".inst 0x80940140 // fmopa za0.s, p0/M, p0/M, z10.s, z20.s\n" + ".inst 0x80950141 // fmopa za1.s, p0/M, p0/M, z10.s, z21.s\n" + ".inst 0x80940162 // fmopa za2.s, p0/M, p0/M, z11.s, z20.s\n" + ".inst 0x80950163 // fmopa za3.s, p0/M, p0/M, z11.s, z21.s\n" + ".inst 0x808800c0 // fmopa za0.s, p0/M, p0/M, z6.s, z8.s\n" + ".inst 0x808900c1 // fmopa za1.s, p0/M, p0/M, z6.s, z9.s\n" + ".inst 0x808800e2 // fmopa za2.s, p0/M, p0/M, z7.s, z8.s\n" + ".inst 0x808900e3 // fmopa za3.s, p0/M, p0/M, z7.s, z9.s\n" "8:" // K oddments - "cbz x20, 10f\n" + "cbz x22, 10f\n" "9:" // K oddments: Loop - ".inst 0xa040476a // ld1w { z10.s-z11.s }, pn9.b/Z, [x27]\n" - "subs x20, x20, #0x1\n" + ".inst 0xa040477e // ld1w { z30.s-z31.s }, pn9.b/Z, [x27]\n" + "subs x22, x22, #0x1\n" "addvl x27, x27, #2\n" - ".inst 0xa04046ee // ld1w { z14.s-z15.s }, pn9.b/Z, [x23]\n" - "addvl x23, x23, #2\n" - ".inst 0x808e0140 // fmopa za0.s, p0/M, p0/M, z10.s, z14.s\n" - ".inst 0x808f0141 // fmopa za1.s, p0/M, p0/M, z10.s, z15.s\n" - ".inst 0x808e0162 // fmopa za2.s, p0/M, p0/M, z11.s, z14.s\n" - ".inst 0x808f0163 // fmopa za3.s, p0/M, p0/M, z11.s, z15.s\n" + ".inst 0xa14046a5 // ld1w { z5.s, z13.s }, pn9.b/Z, [x21]\n" + "addvl x21, x21, #2\n" + ".inst 0x808503c0 // fmopa za0.s, p0/M, p0/M, z30.s, z5.s\n" + ".inst 0x808d03c1 // fmopa za1.s, p0/M, p0/M, z30.s, z13.s\n" + ".inst 0x808503e2 // fmopa za2.s, p0/M, p0/M, z31.s, z5.s\n" + ".inst 0x808d03e3 // fmopa za3.s, p0/M, p0/M, z31.s, z13.s\n" "bgt 9b\n" "10:" // K oddments: End "tbz x16, #1, 14f\n" @@ -239,24 +239,24 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n" - ".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n" - ".inst 0xc086043c // mova { z28.s-z31.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 0xa040c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n" + ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" + ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xa041c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" + ".inst 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n" ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" ".inst 0xa042c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xa043c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840400 // mova za0h.s[x12], { z0.s-z3.s }\n" - "addvl x15, x15, #16\n" - ".inst 0xc0840481 // mova za1h.s[x12], { z4.s-z7.s }\n" - ".inst 0xa060c5d4 // st1w { z20.s-z23.s }, pn9.b, [x14]\n" ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa061c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x4, MUL VL]\n" - ".inst 0xc0840703 // mova za3h.s[x12], { z24.s-z27.s }\n" + ".inst 0xa043c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840503 // mova za3h.s[x12], { z8.s-z11.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 0xa060c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14]\n" + "addvl x15, x15, #16\n" + ".inst 0xa061c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, 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" @@ -265,31 +265,31 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "mov x12, #0x0\n" "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop - ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" - ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" - ".inst 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n" - ".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\n" - ".inst 0xa060c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14]\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n" + ".inst 0xa060c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14]\n" + ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" + ".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n" + ".inst 0xa061c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" - ".inst 0xa061c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14, #0x4, MUL VL]\n" "cmp x12, x20\n" - ".inst 0xa062c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c5d4 // st1w { z20.s-z23.s }, pn9.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa062c5d8 // st1w { z24.s-z27.s }, pn9.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 13b\n" "b 30f\n" "14:" // Store to output array "ldr x26, [%x[args], %[offsetof_C]]\n" + "add x26, x26, x10, LSL #2\n" // C += n "sub x25, x13, x11\n" "ldr x24, [%x[args], %[offsetof_ldcb]]\n" - "add x26, x26, x10, LSL #2\n" // C += n "madd x26, x11, x24, x26\n" // C += m * ldc "tbz x16, #2, 21f\n" "cntw x23\n" - "mov x12, #0x0\n" "cmp x25, x23\n" "csel x22, x25, x23, LT\n" "lsr x21, x22, #0x2\n" + "mov x12, #0x0\n" "and x20, x22, #0x3\n" "cbz x21, 16f\n" "15:" // Store to output array: Skip activation: Accumulator row 0 loop @@ -297,36 +297,36 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa ".inst 0xc086042c // mova { z12.s-z15.s }, za1h.s[x12]\n" ".inst 0xa1604344 // st1w { z4.s, z12.s }, p8, [x26]\n" "add x26, x26, x24\n" - "add x12, x12, #0x4\n" ".inst 0xa1604345 // st1w { z5.s, z13.s }, p8, [x26]\n" "add x26, x26, x24\n" - "cmp x12, x21, LSL #2\n" + "add x12, x12, #0x4\n" ".inst 0xa1604346 // st1w { z6.s, z14.s }, p8, [x26]\n" "add x26, x26, x24\n" + "cmp x12, x21, LSL #2\n" ".inst 0xa1604347 // st1w { z7.s, z15.s }, p8, [x26]\n" "add x26, x26, x24\n" "blt 15b\n" "16:" // Store to output array: Skip activation: Accumulator row 0 oddments "cbz x20, 17f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" - ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" - ".inst 0xa1604340 // st1w { z0.s, z8.s }, p8, [x26]\n" + ".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n" + ".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" + ".inst 0xa1604354 // st1w { z20.s, z28.s }, p8, [x26]\n" "add x26, x26, x24\n" "beq 17f\n" "subs x20, x20, #0x1\n" - ".inst 0xa1604341 // st1w { z1.s, z9.s }, p8, [x26]\n" + ".inst 0xa1604355 // st1w { z21.s, z29.s }, p8, [x26]\n" "add x26, x26, x24\n" "beq 17f\n" - ".inst 0xa1604342 // st1w { z2.s, z10.s }, p8, [x26]\n" + ".inst 0xa1604356 // st1w { z22.s, z30.s }, p8, [x26]\n" "add x26, x26, x24\n" "17:" // Store to output array: Skip activation: Accumulator row 0 oddments: End "subs x25, x25, x22\n" "beq 21f\n" "cmp x25, x23\n" - "mov x12, #0x0\n" "csel x22, x25, x23, LT\n" "lsr x21, x22, #0x2\n" + "mov x12, #0x0\n" "and x20, x22, #0x3\n" "cbz x21, 19f\n" "18:" // Store to output array: Skip activation: Accumulator row 1 loop @@ -334,28 +334,28 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" ".inst 0xa1604344 // st1w { z4.s, z12.s }, p8, [x26]\n" "add x26, x26, x24\n" - "add x12, x12, #0x4\n" ".inst 0xa1604345 // st1w { z5.s, z13.s }, p8, [x26]\n" "add x26, x26, x24\n" - "cmp x12, x21, LSL #2\n" + "add x12, x12, #0x4\n" ".inst 0xa1604346 // st1w { z6.s, z14.s }, p8, [x26]\n" "add x26, x26, x24\n" + "cmp x12, x21, LSL #2\n" ".inst 0xa1604347 // st1w { z7.s, z15.s }, p8, [x26]\n" "add x26, x26, x24\n" "blt 18b\n" "19:" // Store to output array: Skip activation: Accumulator row 1 oddments "cbz x20, 20f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860444 // mova { z4.s-z7.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xa1604344 // st1w { z4.s, z12.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, x24\n" "beq 20f\n" "subs x20, x20, #0x1\n" - ".inst 0xa1604345 // st1w { z5.s, z13.s }, p8, [x26]\n" + ".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n" "add x26, x26, x24\n" "beq 20f\n" - ".inst 0xa1604346 // st1w { z6.s, z14.s }, p8, [x26]\n" + ".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n" "add x26, x26, x24\n" "20:" // Store to output array: Skip activation: Accumulator row 1 oddments: End "subs x25, x25, x22\n" @@ -363,37 +363,37 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "b 28f\n" "21:" // Store to output array: Skip activation: End "cntw x23\n" - "ld1rw { z21.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n" - "mov x12, #0x0\n" "cmp x25, x23\n" - "ld1rw { z20.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n" + "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n" "csel x22, x25, x23, LT\n" "lsr x21, x22, #0x2\n" + "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n" + "mov x12, #0x0\n" "and x20, x22, #0x3\n" "cbz x21, 23f\n" "22:" // Store to output array: Accumulator row 0 loop - ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" - ".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n" - ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" - ".inst 0xc1b4cab8 // fclamp { z24.s-z27.s }, z21.s, z20.s\n" - "add x12, x12, #0x4\n" - "cmp x12, x21, LSL #2\n" - ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" + ".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n" + ".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" + ".inst 0xc1a0c834 // fclamp { z20.s-z23.s }, z1.s, z0.s\n" + ".inst 0xc1a0c83c // fclamp { z28.s-z31.s }, z1.s, z0.s\n" + ".inst 0xa1604354 // st1w { z20.s, z28.s }, p8, [x26]\n" "add x26, x26, x24\n" - ".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n" + "add x12, x12, #0x4\n" + ".inst 0xa1604355 // st1w { z21.s, z29.s }, p8, [x26]\n" "add x26, x26, x24\n" - ".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n" + "cmp x12, x21, LSL #2\n" + ".inst 0xa1604356 // st1w { z22.s, z30.s }, p8, [x26]\n" "add x26, x26, x24\n" - ".inst 0xa1604353 // st1w { z19.s, z27.s }, p8, [x26]\n" + ".inst 0xa1604357 // st1w { z23.s, z31.s }, p8, [x26]\n" "add x26, x26, x24\n" "blt 22b\n" "23:" // Store to output array: Accumulator row 0 oddments "cbz x20, 24f\n" ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" ".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n" + ".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n" + ".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" - ".inst 0xc1b4cab8 // fclamp { z24.s-z27.s }, z21.s, z20.s\n" ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" "add x26, x26, x24\n" "beq 24f\n" @@ -407,34 +407,34 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "subs x25, x25, x22\n" "beq 28f\n" "cmp x25, x23\n" - "mov x12, #0x0\n" "csel x20, x25, x23, LT\n" "lsr x21, x20, #0x2\n" + "mov x12, #0x0\n" "and x20, x20, #0x3\n" "cbz x21, 26f\n" "25:" // Store to output array: Accumulator row 1 loop - ".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n" - ".inst 0xc0860468 // mova { z8.s-z11.s }, za3h.s[x12]\n" - ".inst 0xc1b4caa0 // fclamp { z0.s-z3.s }, z21.s, z20.s\n" - ".inst 0xc1b4caa8 // fclamp { z8.s-z11.s }, z21.s, z20.s\n" - "add x12, x12, #0x4\n" - "cmp x12, x21, LSL #2\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 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n" + ".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n" + ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" "add x26, x26, x24\n" - ".inst 0xa1604341 // st1w { z1.s, z9.s }, p8, [x26]\n" + "add x12, x12, #0x4\n" + ".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n" "add x26, x26, x24\n" - ".inst 0xa1604342 // st1w { z2.s, z10.s }, p8, [x26]\n" + "cmp x12, x21, LSL #2\n" + ".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n" "add x26, x26, x24\n" - ".inst 0xa1604343 // st1w { z3.s, z11.s }, p8, [x26]\n" + ".inst 0xa1604353 // st1w { z19.s, z27.s }, p8, [x26]\n" "add x26, x26, x24\n" "blt 25b\n" "26:" // Store to output array: Accumulator row 1 oddments "cbz x20, 27f\n" ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" ".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n" + ".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n" + ".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" - ".inst 0xc1b4cab8 // fclamp { z24.s-z27.s }, z21.s, z20.s\n" ".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n" "add x26, x26, x24\n" "beq 27f\n" @@ -449,25 +449,25 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa "mov x12, #0x0\n" "cntw x20\n" "29:" // Store to output array: Refill accumulators: Loop - ".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 0xa042c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" + ".inst 0xa040c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n" + ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa041c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa042c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" + ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" ".inst 0xa043c5ec // ld1w { z12.s-z15.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 0xc0840482 // mova za2h.s[x12], { z4.s-z7.s }\n" ".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" + "addvl x15, x15, #16\n" "blt 29b\n" "30:" // End block "incw x10, ALL, MUL #2\n" "cmp x10, x9\n" "blt 3b\n" "incw x11, ALL, MUL #2\n" - "mov x10, #0x0\n" "cmp x11, x13\n" + "mov x10, #0x0\n" "mov x28, x27\n" "blt 3b\n" ".inst 0xd503467f // SMSTOP\n" |