diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp | 248 |
1 files changed, 123 insertions, 125 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp index dd99387c5e..97be758bd6 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp @@ -21,7 +21,6 @@ * 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 #include "arm_gemm.hpp" @@ -113,12 +112,12 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: Loop - ".inst 0xa040c5cc // ld1w { z12.s-z15.s }, pn9.b/Z, [x14]\n" - ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa041c5d0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n" - ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa042c5c4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n" - ".inst 0xc0840482 // mova za2h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa040c5c4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x14]\n" + ".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa041c5d8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n" + ".inst 0xc0840701 // mova za1h.s[x12], { z24.s-z27.s }\n" + ".inst 0xa042c5d4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n" + ".inst 0xc0840682 // mova za2h.s[x12], { z20.s-z23.s }\n" ".inst 0xa043c5c4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n" ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" "add x12, x12, #0x4\n" @@ -138,12 +137,12 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(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 z21.s, #1.0\n" + "fmov z6.s, #1.0\n" ".inst 0xa009c29d // ldnt1w { z28.s-z31.s }, p8/Z, [x20, x9, LSL #2]\n" - ".inst 0x809c02a0 // fmopa za0.s, p0/M, p0/M, z21.s, z28.s\n" - ".inst 0x809d02a1 // fmopa za1.s, p0/M, p0/M, z21.s, z29.s\n" - ".inst 0x809e02a2 // fmopa za2.s, p0/M, p0/M, z21.s, z30.s\n" - ".inst 0x809f02a3 // fmopa za3.s, p0/M, p0/M, z21.s, z31.s\n" + ".inst 0x809c00c0 // fmopa za0.s, p0/M, p0/M, z6.s, z28.s\n" + ".inst 0x809d00c1 // fmopa za1.s, p0/M, p0/M, z6.s, z29.s\n" + ".inst 0x809e00c2 // fmopa za2.s, p0/M, p0/M, z6.s, z30.s\n" + ".inst 0x809f00c3 // fmopa za3.s, p0/M, p0/M, z6.s, z31.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x9\n" "mov x21, x10\n" @@ -164,75 +163,75 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa "madd x21, x9, x20, x21\n" // bptr = B + n * kstride_bytes "cbz x23, 8f\n" "subs x23, x23, #0x1\n" - "ld1w { z0.s }, p0/Z, [x26]\n" - ".inst 0xa140c6bb // ldnt1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21]\n" - "ld1w { z13.s }, p0/Z, [x26, #1, MUL VL]\n" - ".inst 0xa141c6aa // ldnt1w { z2.s, z6.s, z10.s, z14.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n" - "ld1w { z12.s }, p0/Z, [x26, #2, MUL VL]\n" - ".inst 0xa142c6ab // ldnt1w { z3.s, z7.s, z11.s, z15.s }, pn9.b/Z, [x21, #0x8, MUL VL]\n" - "ld1w { z26.s }, p0/Z, [x26, #3, MUL VL]\n" + "ld1w { z28.s }, p0/Z, [x26]\n" + ".inst 0xa040c6a9 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x21]\n" + "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n" + ".inst 0xa041c6ad // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n" + "ld1w { z30.s }, p0/Z, [x26, #2, MUL VL]\n" + ".inst 0xa042c6a5 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x21, #0x8, MUL VL]\n" + "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n" "addvl x26, x26, #4\n" - ".inst 0xa143c6b8 // ldnt1w { z16.s, z20.s, z24.s, z28.s }, pn9.b/Z, [x21, #0xc, MUL VL]\n" + ".inst 0xa143c6bb // ldnt1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21, #0xc, MUL VL]\n" "addvl x21, x21, #16\n" "ble 7f\n" "6:" // K loop - ".inst 0x80930000 // fmopa za0.s, p0/M, p0/M, z0.s, z19.s\n" + ".inst 0x80880380 // fmopa za0.s, p0/M, p0/M, z28.s, z8.s\n" "subs x23, x23, #0x1\n" - ".inst 0x80970001 // fmopa za1.s, p0/M, p0/M, z0.s, z23.s\n" - ".inst 0x809b0002 // fmopa za2.s, p0/M, p0/M, z0.s, z27.s\n" - ".inst 0x809f0003 // fmopa za3.s, p0/M, p0/M, z0.s, z31.s\n" - "ld1w { z0.s }, p0/Z, [x26]\n" - ".inst 0x808201a0 // fmopa za0.s, p0/M, p0/M, z13.s, z2.s\n" - ".inst 0xa140c6bb // ldnt1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21]\n" - ".inst 0x808601a1 // fmopa za1.s, p0/M, p0/M, z13.s, z6.s\n" - ".inst 0x808a01a2 // fmopa za2.s, p0/M, p0/M, z13.s, z10.s\n" - ".inst 0x808e01a3 // fmopa za3.s, p0/M, p0/M, z13.s, z14.s\n" - "ld1w { z13.s }, p0/Z, [x26, #1, MUL VL]\n" - ".inst 0x80830180 // fmopa za0.s, p0/M, p0/M, z12.s, z3.s\n" - ".inst 0xa141c6aa // ldnt1w { z2.s, z6.s, z10.s, z14.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n" - ".inst 0x80870181 // fmopa za1.s, p0/M, p0/M, z12.s, z7.s\n" - ".inst 0x808b0182 // fmopa za2.s, p0/M, p0/M, z12.s, z11.s\n" - ".inst 0x808f0183 // fmopa za3.s, p0/M, p0/M, z12.s, z15.s\n" - "ld1w { z12.s }, p0/Z, [x26, #2, MUL VL]\n" - ".inst 0xa142c6ab // ldnt1w { z3.s, z7.s, z11.s, z15.s }, pn9.b/Z, [x21, #0x8, MUL VL]\n" - ".inst 0x80900340 // fmopa za0.s, p0/M, p0/M, z26.s, z16.s\n" - ".inst 0x80940341 // fmopa za1.s, p0/M, p0/M, z26.s, z20.s\n" - ".inst 0x80980342 // fmopa za2.s, p0/M, p0/M, z26.s, z24.s\n" - ".inst 0x809c0343 // fmopa za3.s, p0/M, p0/M, z26.s, z28.s\n" - "ld1w { z26.s }, p0/Z, [x26, #3, MUL VL]\n" + ".inst 0x80890381 // fmopa za1.s, p0/M, p0/M, z28.s, z9.s\n" + ".inst 0x808a0382 // fmopa za2.s, p0/M, p0/M, z28.s, z10.s\n" + ".inst 0x808b0383 // fmopa za3.s, p0/M, p0/M, z28.s, z11.s\n" + "ld1w { z28.s }, p0/Z, [x26]\n" + ".inst 0x808c02c0 // fmopa za0.s, p0/M, p0/M, z22.s, z12.s\n" + ".inst 0xa040c6a9 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x21]\n" + ".inst 0x808d02c1 // fmopa za1.s, p0/M, p0/M, z22.s, z13.s\n" + ".inst 0x808e02c2 // fmopa za2.s, p0/M, p0/M, z22.s, z14.s\n" + ".inst 0x808f02c3 // fmopa za3.s, p0/M, p0/M, z22.s, z15.s\n" + "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n" + ".inst 0x808403c0 // fmopa za0.s, p0/M, p0/M, z30.s, z4.s\n" + ".inst 0xa041c6ad // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n" + ".inst 0x808503c1 // fmopa za1.s, p0/M, p0/M, z30.s, z5.s\n" + ".inst 0x808603c2 // fmopa za2.s, p0/M, p0/M, z30.s, z6.s\n" + ".inst 0x808703c3 // fmopa za3.s, p0/M, p0/M, z30.s, z7.s\n" + "ld1w { z30.s }, p0/Z, [x26, #2, MUL VL]\n" + ".inst 0xa042c6a5 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x21, #0x8, MUL VL]\n" + ".inst 0x80930280 // fmopa za0.s, p0/M, p0/M, z20.s, z19.s\n" + ".inst 0x80970281 // fmopa za1.s, p0/M, p0/M, z20.s, z23.s\n" + ".inst 0x809b0282 // fmopa za2.s, p0/M, p0/M, z20.s, z27.s\n" + ".inst 0x809f0283 // fmopa za3.s, p0/M, p0/M, z20.s, z31.s\n" + "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n" "addvl x26, x26, #4\n" - ".inst 0xa143c6b8 // ldnt1w { z16.s, z20.s, z24.s, z28.s }, pn9.b/Z, [x21, #0xc, MUL VL]\n" + ".inst 0xa143c6bb // ldnt1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21, #0xc, MUL VL]\n" "addvl x21, x21, #16\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0x80930000 // fmopa za0.s, p0/M, p0/M, z0.s, z19.s\n" - ".inst 0x80970001 // fmopa za1.s, p0/M, p0/M, z0.s, z23.s\n" - ".inst 0x809b0002 // fmopa za2.s, p0/M, p0/M, z0.s, z27.s\n" - ".inst 0x809f0003 // fmopa za3.s, p0/M, p0/M, z0.s, z31.s\n" - ".inst 0x808201a0 // fmopa za0.s, p0/M, p0/M, z13.s, z2.s\n" - ".inst 0x808601a1 // fmopa za1.s, p0/M, p0/M, z13.s, z6.s\n" - ".inst 0x808a01a2 // fmopa za2.s, p0/M, p0/M, z13.s, z10.s\n" - ".inst 0x808e01a3 // fmopa za3.s, p0/M, p0/M, z13.s, z14.s\n" - ".inst 0x80830180 // fmopa za0.s, p0/M, p0/M, z12.s, z3.s\n" - ".inst 0x80870181 // fmopa za1.s, p0/M, p0/M, z12.s, z7.s\n" - ".inst 0x808b0182 // fmopa za2.s, p0/M, p0/M, z12.s, z11.s\n" - ".inst 0x808f0183 // fmopa za3.s, p0/M, p0/M, z12.s, z15.s\n" - ".inst 0x80900340 // fmopa za0.s, p0/M, p0/M, z26.s, z16.s\n" - ".inst 0x80940341 // fmopa za1.s, p0/M, p0/M, z26.s, z20.s\n" - ".inst 0x80980342 // fmopa za2.s, p0/M, p0/M, z26.s, z24.s\n" - ".inst 0x809c0343 // fmopa za3.s, p0/M, p0/M, z26.s, z28.s\n" + ".inst 0x80880380 // fmopa za0.s, p0/M, p0/M, z28.s, z8.s\n" + ".inst 0x80890381 // fmopa za1.s, p0/M, p0/M, z28.s, z9.s\n" + ".inst 0x808a0382 // fmopa za2.s, p0/M, p0/M, z28.s, z10.s\n" + ".inst 0x808b0383 // fmopa za3.s, p0/M, p0/M, z28.s, z11.s\n" + ".inst 0x808c02c0 // fmopa za0.s, p0/M, p0/M, z22.s, z12.s\n" + ".inst 0x808d02c1 // fmopa za1.s, p0/M, p0/M, z22.s, z13.s\n" + ".inst 0x808e02c2 // fmopa za2.s, p0/M, p0/M, z22.s, z14.s\n" + ".inst 0x808f02c3 // fmopa za3.s, p0/M, p0/M, z22.s, z15.s\n" + ".inst 0x808403c0 // fmopa za0.s, p0/M, p0/M, z30.s, z4.s\n" + ".inst 0x808503c1 // fmopa za1.s, p0/M, p0/M, z30.s, z5.s\n" + ".inst 0x808603c2 // fmopa za2.s, p0/M, p0/M, z30.s, z6.s\n" + ".inst 0x808703c3 // fmopa za3.s, p0/M, p0/M, z30.s, z7.s\n" + ".inst 0x80930280 // fmopa za0.s, p0/M, p0/M, z20.s, z19.s\n" + ".inst 0x80970281 // fmopa za1.s, p0/M, p0/M, z20.s, z23.s\n" + ".inst 0x809b0282 // fmopa za2.s, p0/M, p0/M, z20.s, z27.s\n" + ".inst 0x809f0283 // fmopa za3.s, p0/M, p0/M, z20.s, z31.s\n" "8:" // K oddments "cbz x22, 10f\n" "9:" // K oddments: Loop - "ld1w { z0.s }, p0/Z, [x26]\n" + "ld1w { z8.s }, p0/Z, [x26]\n" "subs x22, x22, #0x1\n" "addvl x26, x26, #1\n" - ".inst 0xa140c6b3 // ld1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21]\n" + ".inst 0xa140c6a3 // ld1w { z3.s, z7.s, z11.s, z15.s }, pn9.b/Z, [x21]\n" "addvl x21, x21, #4\n" - ".inst 0x80930000 // fmopa za0.s, p0/M, p0/M, z0.s, z19.s\n" - ".inst 0x80970001 // fmopa za1.s, p0/M, p0/M, z0.s, z23.s\n" - ".inst 0x809b0002 // fmopa za2.s, p0/M, p0/M, z0.s, z27.s\n" - ".inst 0x809f0003 // fmopa za3.s, p0/M, p0/M, z0.s, z31.s\n" + ".inst 0x80830100 // fmopa za0.s, p0/M, p0/M, z8.s, z3.s\n" + ".inst 0x80870101 // fmopa za1.s, p0/M, p0/M, z8.s, z7.s\n" + ".inst 0x808b0102 // fmopa za2.s, p0/M, p0/M, z8.s, z11.s\n" + ".inst 0x808f0103 // fmopa za3.s, p0/M, p0/M, z8.s, z15.s\n" "bgt 9b\n" "10:" // K oddments: End "tbz x15, #1, 14f\n" @@ -240,25 +239,25 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c5cc // ld1w { z12.s-z15.s }, pn9.b/Z, [x14]\n" - ".inst 0xc0860418 // mova { z24.s-z27.s }, za0h.s[x12]\n" - ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" - ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" - ".inst 0xa041c5dc // ld1w { z28.s-z31.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n" - ".inst 0xc0840781 // mova za1h.s[x12], { z28.s-z31.s }\n" - ".inst 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n" - ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" - ".inst 0xa042c5cc // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n" - ".inst 0xc0840582 // mova za2h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa043c5cc // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n" - ".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa040c5d4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x14]\n" + ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" + ".inst 0xc0840680 // mova za0h.s[x12], { z20.s-z23.s }\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xa041c5cc // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #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 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" + ".inst 0xa042c5d0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n" + ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa043c5d0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n" + ".inst 0xc0840603 // mova za3h.s[x12], { z16.s-z19.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa060c5b8 // st1w { z24.s-z27.s }, pn9.b, [x13]\n" + ".inst 0xa060c5a8 // st1w { z8.s-z11.s }, pn9.b, [x13]\n" "addvl x14, x14, #16\n" - ".inst 0xa061c5b4 // st1w { z20.s-z23.s }, pn9.b, [x13, #0x4, MUL VL]\n" - ".inst 0xa062c5bc // st1w { z28.s-z31.s }, pn9.b, [x13, #0x8, MUL VL]\n" - ".inst 0xa063c5b0 // st1w { z16.s-z19.s }, pn9.b, [x13, #0xc, MUL VL]\n" + ".inst 0xa061c5a4 // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n" + ".inst 0xa062c5ac // st1w { z12.s-z15.s }, pn9.b, [x13, #0x8, MUL VL]\n" + ".inst 0xa063c5a0 // st1w { z0.s-z3.s }, pn9.b, [x13, #0xc, MUL VL]\n" "addvl x13, x13, #16\n" "blt 11b\n" "b 24f\n" @@ -266,15 +265,15 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa "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 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" - ".inst 0xa060c5ac // st1w { z12.s-z15.s }, pn9.b, [x13]\n" - ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n" + ".inst 0xa060c5b0 // st1w { z16.s-z19.s }, pn9.b, [x13]\n" + ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" ".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n" - ".inst 0xa061c5bc // st1w { z28.s-z31.s }, pn9.b, [x13, #0x4, MUL VL]\n" + ".inst 0xa061c5b8 // st1w { z24.s-z27.s }, pn9.b, [x13, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa062c5b0 // st1w { z16.s-z19.s }, pn9.b, [x13, #0x8, MUL VL]\n" + ".inst 0xa062c5a8 // st1w { z8.s-z11.s }, pn9.b, [x13, #0x8, MUL VL]\n" ".inst 0xa063c5a4 // st1w { z4.s-z7.s }, pn9.b, [x13, #0xc, MUL VL]\n" "addvl x13, x13, #16\n" "blt 13b\n" @@ -312,18 +311,18 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa "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 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" - ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xa160c320 // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" + ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" + ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xa160c330 // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n" "add x25, x25, x23\n" "beq 17f\n" "subs x20, x20, #0x1\n" - ".inst 0xa160c321 // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n" + ".inst 0xa160c331 // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n" "add x25, x25, x23\n" "beq 17f\n" - ".inst 0xa160c322 // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n" + ".inst 0xa160c332 // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n" "add x25, x25, x23\n" "17:" // Store to output array: Skip activation: Accumulator row 0 oddments: End "subs x24, x24, x22\n" @@ -332,66 +331,66 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa "18:" // Store to output array: Skip activation: End "cntw x20\n" "cmp x24, x20\n" - "ld1rw { z23.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n" + "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n" "csel x20, x24, x20, LT\n" "lsr x21, x20, #0x2\n" - "ld1rw { z16.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n" + "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n" "mov x12, #0x0\n" "and x20, x20, #0x3\n" "cbz x21, 20f\n" "19:" // Store to output array: Accumulator row 0 loop - ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" - ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" - ".inst 0xc1b0cae0 // fclamp { z0.s-z3.s }, z23.s, z16.s\n" - ".inst 0xc1b0cae4 // fclamp { z4.s-z7.s }, z23.s, z16.s\n" - ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xc1b0cae8 // fclamp { z8.s-z11.s }, z23.s, z16.s\n" - ".inst 0xc1b0caec // fclamp { z12.s-z15.s }, z23.s, z16.s\n" - ".inst 0xa160c320 // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" + ".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n" + ".inst 0xc1a0c834 // fclamp { z20.s-z23.s }, z1.s, z0.s\n" + ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" + ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n" + ".inst 0xc1a0c83c // fclamp { z28.s-z31.s }, z1.s, z0.s\n" + ".inst 0xa160c330 // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n" "add x25, x25, x23\n" "add x12, x12, #0x4\n" - ".inst 0xa160c321 // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n" + ".inst 0xa160c331 // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n" "add x25, x25, x23\n" "cmp x12, x21, LSL #2\n" - ".inst 0xa160c322 // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n" + ".inst 0xa160c332 // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n" "add x25, x25, x23\n" - ".inst 0xa160c323 // st1w { z3.s, z7.s, z11.s, z15.s }, p8, [x25]\n" + ".inst 0xa160c333 // st1w { z19.s, z23.s, z27.s, z31.s }, p8, [x25]\n" "add x25, x25, x23\n" "blt 19b\n" "20:" // Store to output array: Accumulator row 0 oddments "cbz x20, 21f\n" - ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" - ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" - ".inst 0xc1b0cae0 // fclamp { z0.s-z3.s }, z23.s, z16.s\n" - ".inst 0xc1b0cae4 // fclamp { z4.s-z7.s }, z23.s, z16.s\n" - ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xc1b0cae8 // fclamp { z8.s-z11.s }, z23.s, z16.s\n" - ".inst 0xc1b0caec // fclamp { z12.s-z15.s }, z23.s, z16.s\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" + ".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n" + ".inst 0xc1a0c834 // fclamp { z20.s-z23.s }, z1.s, z0.s\n" + ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" + ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n" + ".inst 0xc1a0c83c // fclamp { z28.s-z31.s }, z1.s, z0.s\n" "subs x20, x20, #0x1\n" - ".inst 0xa160c320 // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n" + ".inst 0xa160c330 // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n" "add x25, x25, x23\n" "beq 21f\n" "subs x20, x20, #0x1\n" - ".inst 0xa160c321 // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n" + ".inst 0xa160c331 // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n" "add x25, x25, x23\n" "beq 21f\n" - ".inst 0xa160c322 // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n" + ".inst 0xa160c332 // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n" "21:" // Store to output array: Accumulator row 0 oddments: End "22:" // Store to output array: End "tbz x15, #0, 24f\n" "mov x12, #0x0\n" "cntw x20\n" "23:" // Store to output array: Refill accumulators: Loop - ".inst 0xa040c5d0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x14]\n" - ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa040c5cc // ld1w { z12.s-z15.s }, pn9.b/Z, [x14]\n" + ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" ".inst 0xa041c5d0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n" ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" ".inst 0xa042c5d0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n" ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa043c5c8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n" - ".inst 0xc0840503 // mova za3h.s[x12], { z8.s-z11.s }\n" + ".inst 0xa043c5c4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n" + ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x14, x14, #16\n" @@ -415,4 +414,3 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |