diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp | 298 |
1 files changed, 148 insertions, 150 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp index d868ed2b67..96247d2db5 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_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" @@ -100,14 +99,14 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: Loop - ".inst 0xa040c5ac // ld1w { z12.s-z15.s }, pn9.b/Z, [x13]\n" - ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa041c5bc // ld1w { z28.s-z31.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n" - ".inst 0xc0840781 // mova za1h.s[x12], { z28.s-z31.s }\n" - ".inst 0xa042c5bc // ld1w { z28.s-z31.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n" - ".inst 0xc0840782 // mova za2h.s[x12], { z28.s-z31.s }\n" - ".inst 0xa043c5a4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n" - ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa040c5a4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x13]\n" + ".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa041c5a8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n" + ".inst 0xc0840501 // mova za1h.s[x12], { z8.s-z11.s }\n" + ".inst 0xa042c5a8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n" + ".inst 0xc0840502 // mova za2h.s[x12], { z8.s-z11.s }\n" + ".inst 0xa043c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n" + ".inst 0xc0840603 // mova za3h.s[x12], { z16.s-z19.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x13, x13, #16\n" @@ -125,11 +124,11 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "ldr x20, [%x[args], %[offsetof_bias]]\n" ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "cbz x20, 5f\n" - ".inst 0xa01cc299 // ldnt1w { z24.s-z27.s }, p8/Z, [x20, x28, LSL #2]\n" - ".inst 0xc0902700 // addha za0.s, p1/M, p1/M, z24.s\n" - ".inst 0xc0902721 // addha za1.s, p1/M, p1/M, z25.s\n" - ".inst 0xc0902742 // addha za2.s, p1/M, p1/M, z26.s\n" - ".inst 0xc0902763 // addha za3.s, p1/M, p1/M, z27.s\n" + ".inst 0xa11cc289 // ldnt1w { z1.s, z5.s, z9.s, z13.s }, p8/Z, [x20, x28, LSL #2]\n" + ".inst 0xc0902420 // addha za0.s, p1/M, p1/M, z1.s\n" + ".inst 0xc09024a1 // addha za1.s, p1/M, p1/M, z5.s\n" + ".inst 0xc0902522 // addha za2.s, p1/M, p1/M, z9.s\n" + ".inst 0xc09025a3 // addha za3.s, p1/M, p1/M, z13.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x28\n" "mov x21, x9\n" @@ -152,107 +151,107 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "madd x23, x28, x20, x23\n" // bptr = B + n * kstride_bytes "cbz x22, 8f\n" "subs x22, x22, #0x1\n" - "ld1b { z10.b }, p1/Z, [x25]\n" - ".inst 0xa04086fd // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23]\n" - "ld1b { z16.b }, p1/Z, [x25, #1, MUL VL]\n" - ".inst 0xa04186ed // ldnt1b { z12.b-z15.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - "ld1b { z21.b }, p1/Z, [x25, #2, MUL VL]\n" - ".inst 0xa04286f9 // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" - "ld1b { z19.b }, p1/Z, [x25, #3, MUL VL]\n" + "ld1b { z20.b }, p1/Z, [x25]\n" + ".inst 0xa04086e5 // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23]\n" + "ld1b { z11.b }, p1/Z, [x25, #1, MUL VL]\n" + ".inst 0xa04186f9 // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + "ld1b { z2.b }, p1/Z, [x25, #2, MUL VL]\n" + ".inst 0xa04286fd // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" + "ld1b { z14.b }, p1/Z, [x25, #3, MUL VL]\n" "addvl x25, x25, #4\n" - ".inst 0xa04386e1 // ldnt1b { z0.b-z3.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n" + ".inst 0xa04386f1 // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n" "addvl x23, x23, #16\n" "ble 7f\n" "6:" // K loop - ".inst 0xa1bc2540 // umopa za0.s, p1/M, p1/M, z10.b, z28.b\n" + ".inst 0xa1a42680 // umopa za0.s, p1/M, p1/M, z20.b, z4.b\n" "subs x22, x22, #0x1\n" - ".inst 0xa1bd2541 // umopa za1.s, p1/M, p1/M, z10.b, z29.b\n" - ".inst 0xa1be2542 // umopa za2.s, p1/M, p1/M, z10.b, z30.b\n" - ".inst 0xa1bf2543 // umopa za3.s, p1/M, p1/M, z10.b, z31.b\n" - "ld1b { z10.b }, p1/Z, [x25]\n" - ".inst 0xa1ac2600 // umopa za0.s, p1/M, p1/M, z16.b, z12.b\n" - ".inst 0xa04086fd // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23]\n" - ".inst 0xa1ad2601 // umopa za1.s, p1/M, p1/M, z16.b, z13.b\n" - ".inst 0xa1ae2602 // umopa za2.s, p1/M, p1/M, z16.b, z14.b\n" - ".inst 0xa1af2603 // umopa za3.s, p1/M, p1/M, z16.b, z15.b\n" - "ld1b { z16.b }, p1/Z, [x25, #1, MUL VL]\n" - ".inst 0xa1b826a0 // umopa za0.s, p1/M, p1/M, z21.b, z24.b\n" - ".inst 0xa04186ed // ldnt1b { z12.b-z15.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa1b926a1 // umopa za1.s, p1/M, p1/M, z21.b, z25.b\n" - ".inst 0xa1ba26a2 // umopa za2.s, p1/M, p1/M, z21.b, z26.b\n" - ".inst 0xa1bb26a3 // umopa za3.s, p1/M, p1/M, z21.b, z27.b\n" - "ld1b { z21.b }, p1/Z, [x25, #2, MUL VL]\n" - ".inst 0xa04286f9 // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" - ".inst 0xa1a02660 // umopa za0.s, p1/M, p1/M, z19.b, z0.b\n" - ".inst 0xa1a12661 // umopa za1.s, p1/M, p1/M, z19.b, z1.b\n" - ".inst 0xa1a22662 // umopa za2.s, p1/M, p1/M, z19.b, z2.b\n" - ".inst 0xa1a32663 // umopa za3.s, p1/M, p1/M, z19.b, z3.b\n" - "ld1b { z19.b }, p1/Z, [x25, #3, MUL VL]\n" + ".inst 0xa1a52681 // umopa za1.s, p1/M, p1/M, z20.b, z5.b\n" + ".inst 0xa1a62682 // umopa za2.s, p1/M, p1/M, z20.b, z6.b\n" + ".inst 0xa1a72683 // umopa za3.s, p1/M, p1/M, z20.b, z7.b\n" + "ld1b { z20.b }, p1/Z, [x25]\n" + ".inst 0xa1b82560 // umopa za0.s, p1/M, p1/M, z11.b, z24.b\n" + ".inst 0xa04086e5 // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23]\n" + ".inst 0xa1b92561 // umopa za1.s, p1/M, p1/M, z11.b, z25.b\n" + ".inst 0xa1ba2562 // umopa za2.s, p1/M, p1/M, z11.b, z26.b\n" + ".inst 0xa1bb2563 // umopa za3.s, p1/M, p1/M, z11.b, z27.b\n" + "ld1b { z11.b }, p1/Z, [x25, #1, MUL VL]\n" + ".inst 0xa1bc2440 // umopa za0.s, p1/M, p1/M, z2.b, z28.b\n" + ".inst 0xa04186f9 // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa1bd2441 // umopa za1.s, p1/M, p1/M, z2.b, z29.b\n" + ".inst 0xa1be2442 // umopa za2.s, p1/M, p1/M, z2.b, z30.b\n" + ".inst 0xa1bf2443 // umopa za3.s, p1/M, p1/M, z2.b, z31.b\n" + "ld1b { z2.b }, p1/Z, [x25, #2, MUL VL]\n" + ".inst 0xa04286fd // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" + ".inst 0xa1b025c0 // umopa za0.s, p1/M, p1/M, z14.b, z16.b\n" + ".inst 0xa1b125c1 // umopa za1.s, p1/M, p1/M, z14.b, z17.b\n" + ".inst 0xa1b225c2 // umopa za2.s, p1/M, p1/M, z14.b, z18.b\n" + ".inst 0xa1b325c3 // umopa za3.s, p1/M, p1/M, z14.b, z19.b\n" + "ld1b { z14.b }, p1/Z, [x25, #3, MUL VL]\n" "addvl x25, x25, #4\n" - ".inst 0xa04386e1 // ldnt1b { z0.b-z3.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n" + ".inst 0xa04386f1 // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n" "addvl x23, x23, #16\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0xa1bc2540 // umopa za0.s, p1/M, p1/M, z10.b, z28.b\n" - ".inst 0xa1bd2541 // umopa za1.s, p1/M, p1/M, z10.b, z29.b\n" - ".inst 0xa1be2542 // umopa za2.s, p1/M, p1/M, z10.b, z30.b\n" - ".inst 0xa1bf2543 // umopa za3.s, p1/M, p1/M, z10.b, z31.b\n" - ".inst 0xa1ac2600 // umopa za0.s, p1/M, p1/M, z16.b, z12.b\n" - ".inst 0xa1ad2601 // umopa za1.s, p1/M, p1/M, z16.b, z13.b\n" - ".inst 0xa1ae2602 // umopa za2.s, p1/M, p1/M, z16.b, z14.b\n" - ".inst 0xa1af2603 // umopa za3.s, p1/M, p1/M, z16.b, z15.b\n" - ".inst 0xa1b826a0 // umopa za0.s, p1/M, p1/M, z21.b, z24.b\n" - ".inst 0xa1b926a1 // umopa za1.s, p1/M, p1/M, z21.b, z25.b\n" - ".inst 0xa1ba26a2 // umopa za2.s, p1/M, p1/M, z21.b, z26.b\n" - ".inst 0xa1bb26a3 // umopa za3.s, p1/M, p1/M, z21.b, z27.b\n" - ".inst 0xa1a02660 // umopa za0.s, p1/M, p1/M, z19.b, z0.b\n" - ".inst 0xa1a12661 // umopa za1.s, p1/M, p1/M, z19.b, z1.b\n" - ".inst 0xa1a22662 // umopa za2.s, p1/M, p1/M, z19.b, z2.b\n" - ".inst 0xa1a32663 // umopa za3.s, p1/M, p1/M, z19.b, z3.b\n" + ".inst 0xa1a42680 // umopa za0.s, p1/M, p1/M, z20.b, z4.b\n" + ".inst 0xa1a52681 // umopa za1.s, p1/M, p1/M, z20.b, z5.b\n" + ".inst 0xa1a62682 // umopa za2.s, p1/M, p1/M, z20.b, z6.b\n" + ".inst 0xa1a72683 // umopa za3.s, p1/M, p1/M, z20.b, z7.b\n" + ".inst 0xa1b82560 // umopa za0.s, p1/M, p1/M, z11.b, z24.b\n" + ".inst 0xa1b92561 // umopa za1.s, p1/M, p1/M, z11.b, z25.b\n" + ".inst 0xa1ba2562 // umopa za2.s, p1/M, p1/M, z11.b, z26.b\n" + ".inst 0xa1bb2563 // umopa za3.s, p1/M, p1/M, z11.b, z27.b\n" + ".inst 0xa1bc2440 // umopa za0.s, p1/M, p1/M, z2.b, z28.b\n" + ".inst 0xa1bd2441 // umopa za1.s, p1/M, p1/M, z2.b, z29.b\n" + ".inst 0xa1be2442 // umopa za2.s, p1/M, p1/M, z2.b, z30.b\n" + ".inst 0xa1bf2443 // umopa za3.s, p1/M, p1/M, z2.b, z31.b\n" + ".inst 0xa1b025c0 // umopa za0.s, p1/M, p1/M, z14.b, z16.b\n" + ".inst 0xa1b125c1 // umopa za1.s, p1/M, p1/M, z14.b, z17.b\n" + ".inst 0xa1b225c2 // umopa za2.s, p1/M, p1/M, z14.b, z18.b\n" + ".inst 0xa1b325c3 // umopa za3.s, p1/M, p1/M, z14.b, z19.b\n" "8:" // K oddments "cbz x21, 10f\n" "9:" // K oddments: Loop - "ld1b { z10.b }, p1/Z, [x25]\n" + "ld1b { z16.b }, p1/Z, [x25]\n" "subs x21, x21, #0x1\n" "addvl x25, x25, #1\n" - ".inst 0xa04086fc // ld1b { z28.b-z31.b }, pn9.b/Z, [x23]\n" + ".inst 0xa04086e4 // ld1b { z4.b-z7.b }, pn9.b/Z, [x23]\n" "addvl x23, x23, #4\n" - ".inst 0xa1bc2540 // umopa za0.s, p1/M, p1/M, z10.b, z28.b\n" - ".inst 0xa1bd2541 // umopa za1.s, p1/M, p1/M, z10.b, z29.b\n" - ".inst 0xa1be2542 // umopa za2.s, p1/M, p1/M, z10.b, z30.b\n" - ".inst 0xa1bf2543 // umopa za3.s, p1/M, p1/M, z10.b, z31.b\n" + ".inst 0xa1a42600 // umopa za0.s, p1/M, p1/M, z16.b, z4.b\n" + ".inst 0xa1a52601 // umopa za1.s, p1/M, p1/M, z16.b, z5.b\n" + ".inst 0xa1a62602 // umopa za2.s, p1/M, p1/M, z16.b, z6.b\n" + ".inst 0xa1a72603 // umopa za3.s, p1/M, p1/M, z16.b, z7.b\n" "bgt 9b\n" "10:" // K oddments: End - "ld1w { z14.s }, p1/Z, [x25]\n" + "ld1w { z15.s }, p1/Z, [x25]\n" "addvl x25, x25, #1\n" - ".inst 0xc09125c0 // addva za0.s, p1/M, p1/M, z14.s\n" - ".inst 0xc09125c1 // addva za1.s, p1/M, p1/M, z14.s\n" - ".inst 0xc09125c2 // addva za2.s, p1/M, p1/M, z14.s\n" - ".inst 0xc09125c3 // addva za3.s, p1/M, p1/M, z14.s\n" + ".inst 0xc09125e0 // addva za0.s, p1/M, p1/M, z15.s\n" + ".inst 0xc09125e1 // addva za1.s, p1/M, p1/M, z15.s\n" + ".inst 0xc09125e2 // addva za2.s, p1/M, p1/M, z15.s\n" + ".inst 0xc09125e3 // addva za3.s, p1/M, p1/M, z15.s\n" "tbz x14, #1, 14f\n" "tbz x14, #0, 12f\n" "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c5b8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x13]\n" - ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" - ".inst 0xc0840700 // mova za0h.s[x12], { z24.s-z27.s }\n" - ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" - ".inst 0xa041c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n" - ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" - ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xa042c5a4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n" - ".inst 0xc0840482 // mova za2h.s[x12], { z4.s-z7.s }\n" - ".inst 0xa043c5a4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n" - ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa040c5a0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x13]\n" + ".inst 0xc0860418 // mova { z24.s-z27.s }, za0h.s[x12]\n" + ".inst 0xc0840400 // mova za0h.s[x12], { z0.s-z3.s }\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xa041c5a0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n" + ".inst 0xc0840401 // mova za1h.s[x12], { z0.s-z3.s }\n" + ".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n" + ".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" + ".inst 0xa042c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n" + ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa043c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #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 0xa060c57c // st1w { z28.s-z31.s }, pn9.b, [x11]\n" + ".inst 0xa060c578 // st1w { z24.s-z27.s }, pn9.b, [x11]\n" "addvl x13, x13, #16\n" - ".inst 0xa061c568 // st1w { z8.s-z11.s }, pn9.b, [x11, #0x4, MUL VL]\n" - ".inst 0xa062c578 // st1w { z24.s-z27.s }, pn9.b, [x11, #0x8, MUL VL]\n" - ".inst 0xa063c56c // st1w { z12.s-z15.s }, pn9.b, [x11, #0xc, MUL VL]\n" + ".inst 0xa061c564 // st1w { z4.s-z7.s }, pn9.b, [x11, #0x4, MUL VL]\n" + ".inst 0xa062c574 // st1w { z20.s-z23.s }, pn9.b, [x11, #0x8, MUL VL]\n" + ".inst 0xa063c560 // st1w { z0.s-z3.s }, pn9.b, [x11, #0xc, MUL VL]\n" "addvl x11, x11, #16\n" "blt 11b\n" "b 21f\n" @@ -260,16 +259,16 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "mov x12, #0x0\n" "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop - ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" - ".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n" - ".inst 0xa060c57c // st1w { z28.s-z31.s }, pn9.b, [x11]\n" - ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" - ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" - ".inst 0xa061c560 // st1w { z0.s-z3.s }, pn9.b, [x11, #0x4, MUL VL]\n" + ".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n" + ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" + ".inst 0xa060c564 // st1w { z4.s-z7.s }, pn9.b, [x11]\n" + ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" + ".inst 0xc0860468 // mova { z8.s-z11.s }, za3h.s[x12]\n" + ".inst 0xa061c574 // st1w { z20.s-z23.s }, pn9.b, [x11, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa062c568 // st1w { z8.s-z11.s }, pn9.b, [x11, #0x8, MUL VL]\n" - ".inst 0xa063c570 // st1w { z16.s-z19.s }, pn9.b, [x11, #0xc, MUL VL]\n" + ".inst 0xa062c56c // st1w { z12.s-z15.s }, pn9.b, [x11, #0x8, MUL VL]\n" + ".inst 0xa063c568 // st1w { z8.s-z11.s }, pn9.b, [x11, #0xc, MUL VL]\n" "addvl x11, x11, #16\n" "blt 13b\n" "b 21f\n" @@ -277,17 +276,17 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "ldr x24, [%x[args], %[offsetof_C]]\n" "add x24, x24, x28\n" // C += n "sub x23, x10, x9\n" - "ld1rw { z12.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z4.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" "ldr x22, [%x[args], %[offsetof_ldcb]]\n" "madd x24, x9, x22, x24\n" // C += m * ldc - "ld1rw { z13.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" - "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" - "ld1rw { z15.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" - "ld1rw { z4.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z5.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z6.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z7.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n" + "ld1rw { z5.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z6.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z7.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z12.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z13.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z15.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n" "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n" "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n" "tbz x14, #2, 15f\n" @@ -295,10 +294,10 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "add x21, x21, x28\n" "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n" "add x20, x20, x21, LSL #2\n" - ".inst 0xa040c28c // ld1w { z12.s-z15.s }, p8/Z, [x20]\n" + ".inst 0xa040c284 // ld1w { z4.s-z7.s }, p8/Z, [x20]\n" "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n" "add x20, x20, x21, LSL #2\n" - ".inst 0xa040c284 // ld1w { z4.s-z7.s }, p8/Z, [x20]\n" + ".inst 0xa040c28c // ld1w { z12.s-z15.s }, p8/Z, [x20]\n" "15:" // Store to output array: Load per-channel parameters: End "cntw x20\n" "whilelt p0.b, x28, x27\n" @@ -311,22 +310,22 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "16:" // Store to output array: Accumulator row 0 loop ".inst 0xc086001a // mova { z26.s-z27.s }, za0h.s[x12, 0:1]\n" ".inst 0xc086005c // mova { z28.s-z29.s }, za1h.s[x12, 0:1]\n" - ".inst 0xc1aca41a // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z12.s\n" + ".inst 0xc1a4a41a // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z4.s\n" ".inst 0xc0860096 // mova { z22.s-z23.s }, za2h.s[x12, 0:1]\n" ".inst 0xc08600d0 // mova { z16.s-z17.s }, za3h.s[x12, 0:1]\n" - ".inst 0xc1ada41c // sqdmulh { z28.s-z29.s }, { z28.s-z29.s }, z13.s\n" - ".inst 0xc1aea416 // sqdmulh { z22.s-z23.s }, { z22.s-z23.s }, z14.s\n" + ".inst 0xc1a5a41c // sqdmulh { z28.s-z29.s }, { z28.s-z29.s }, z5.s\n" + ".inst 0xc1a6a416 // sqdmulh { z22.s-z23.s }, { z22.s-z23.s }, z6.s\n" "add x12, x12, #0x2\n" "cmp x12, x21, LSL #1\n" - ".inst 0xc1afa410 // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z15.s\n" - ".inst 0xc1a4a23a // srshl { z26.s-z27.s }, { z26.s-z27.s }, z4.s\n" - ".inst 0xc1a5a23c // srshl { z28.s-z29.s }, { z28.s-z29.s }, z5.s\n" - ".inst 0xc1a6a236 // srshl { z22.s-z23.s }, { z22.s-z23.s }, z6.s\n" - ".inst 0xc1a7a230 // srshl { z16.s-z17.s }, { z16.s-z17.s }, z7.s\n" - ".inst 0xc1a1a31a // add { z26.s-z27.s }, { z26.s-z27.s }, z1.s\n" - ".inst 0xc1a1a31c // add { z28.s-z29.s }, { z28.s-z29.s }, z1.s\n" - ".inst 0xc1a1a316 // add { z22.s-z23.s }, { z22.s-z23.s }, z1.s\n" - ".inst 0xc1a1a310 // add { z16.s-z17.s }, { z16.s-z17.s }, z1.s\n" + ".inst 0xc1a7a410 // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z7.s\n" + ".inst 0xc1aca23a // srshl { z26.s-z27.s }, { z26.s-z27.s }, z12.s\n" + ".inst 0xc1ada23c // srshl { z28.s-z29.s }, { z28.s-z29.s }, z13.s\n" + ".inst 0xc1aea236 // srshl { z22.s-z23.s }, { z22.s-z23.s }, z14.s\n" + ".inst 0xc1afa230 // srshl { z16.s-z17.s }, { z16.s-z17.s }, z15.s\n" + ".inst 0xc1a0a31a // add { z26.s-z27.s }, { z26.s-z27.s }, z0.s\n" + ".inst 0xc1a0a31c // add { z28.s-z29.s }, { z28.s-z29.s }, z0.s\n" + ".inst 0xc1a0a316 // add { z22.s-z23.s }, { z22.s-z23.s }, z0.s\n" + ".inst 0xc1a0a310 // add { z16.s-z17.s }, { z16.s-z17.s }, z0.s\n" ".inst 0xc1b4c6ba // sclamp { z26.s-z27.s }, z21.s, z20.s\n" ".inst 0xc1b4c6bc // sclamp { z28.s-z29.s }, z21.s, z20.s\n" "uzp1 z19.b, z26.b, z28.b\n" @@ -344,29 +343,29 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "blt 16b\n" "17:" // Store to output array: Accumulator row 0 oddments "cbz x20, 18f\n" - ".inst 0xc0860002 // mova { z2.s-z3.s }, za0h.s[x12, 0:1]\n" + ".inst 0xc086000a // mova { z10.s-z11.s }, za0h.s[x12, 0:1]\n" ".inst 0xc0860058 // mova { z24.s-z25.s }, za1h.s[x12, 0:1]\n" - ".inst 0xc1aca402 // sqdmulh { z2.s-z3.s }, { z2.s-z3.s }, z12.s\n" - ".inst 0xc0860090 // mova { z16.s-z17.s }, za2h.s[x12, 0:1]\n" - ".inst 0xc08600ca // mova { z10.s-z11.s }, za3h.s[x12, 0:1]\n" - ".inst 0xc1ada418 // sqdmulh { z24.s-z25.s }, { z24.s-z25.s }, z13.s\n" - ".inst 0xc1aea410 // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z14.s\n" - ".inst 0xc1afa40a // sqdmulh { z10.s-z11.s }, { z10.s-z11.s }, z15.s\n" - ".inst 0xc1a4a222 // srshl { z2.s-z3.s }, { z2.s-z3.s }, z4.s\n" - ".inst 0xc1a5a238 // srshl { z24.s-z25.s }, { z24.s-z25.s }, z5.s\n" - ".inst 0xc1a6a230 // srshl { z16.s-z17.s }, { z16.s-z17.s }, z6.s\n" - ".inst 0xc1a7a22a // srshl { z10.s-z11.s }, { z10.s-z11.s }, z7.s\n" - ".inst 0xc1a1a302 // add { z2.s-z3.s }, { z2.s-z3.s }, z1.s\n" - ".inst 0xc1a1a318 // add { z24.s-z25.s }, { z24.s-z25.s }, z1.s\n" - ".inst 0xc1a1a310 // add { z16.s-z17.s }, { z16.s-z17.s }, z1.s\n" - ".inst 0xc1a1a30a // add { z10.s-z11.s }, { z10.s-z11.s }, z1.s\n" - ".inst 0xc1b4c6a2 // sclamp { z2.s-z3.s }, z21.s, z20.s\n" - ".inst 0xc1b4c6b8 // sclamp { z24.s-z25.s }, z21.s, z20.s\n" - "uzp1 z23.b, z2.b, z24.b\n" - ".inst 0xc1b4c6b0 // sclamp { z16.s-z17.s }, z21.s, z20.s\n" + ".inst 0xc1a4a40a // sqdmulh { z10.s-z11.s }, { z10.s-z11.s }, z4.s\n" + ".inst 0xc086009a // mova { z26.s-z27.s }, za2h.s[x12, 0:1]\n" + ".inst 0xc08600de // mova { z30.s-z31.s }, za3h.s[x12, 0:1]\n" + ".inst 0xc1a5a418 // sqdmulh { z24.s-z25.s }, { z24.s-z25.s }, z5.s\n" + ".inst 0xc1a6a41a // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z6.s\n" + ".inst 0xc1a7a41e // sqdmulh { z30.s-z31.s }, { z30.s-z31.s }, z7.s\n" + ".inst 0xc1aca22a // srshl { z10.s-z11.s }, { z10.s-z11.s }, z12.s\n" + ".inst 0xc1ada238 // srshl { z24.s-z25.s }, { z24.s-z25.s }, z13.s\n" + ".inst 0xc1aea23a // srshl { z26.s-z27.s }, { z26.s-z27.s }, z14.s\n" + ".inst 0xc1afa23e // srshl { z30.s-z31.s }, { z30.s-z31.s }, z15.s\n" + ".inst 0xc1a0a30a // add { z10.s-z11.s }, { z10.s-z11.s }, z0.s\n" + ".inst 0xc1a0a318 // add { z24.s-z25.s }, { z24.s-z25.s }, z0.s\n" + ".inst 0xc1a0a31a // add { z26.s-z27.s }, { z26.s-z27.s }, z0.s\n" + ".inst 0xc1a0a31e // add { z30.s-z31.s }, { z30.s-z31.s }, z0.s\n" ".inst 0xc1b4c6aa // sclamp { z10.s-z11.s }, z21.s, z20.s\n" - "uzp1 z16.b, z16.b, z10.b\n" - "uzp1 z16.b, z23.b, z16.b\n" + ".inst 0xc1b4c6b8 // sclamp { z24.s-z25.s }, z21.s, z20.s\n" + "uzp1 z17.b, z10.b, z24.b\n" + ".inst 0xc1b4c6ba // sclamp { z26.s-z27.s }, z21.s, z20.s\n" + ".inst 0xc1b4c6be // sclamp { z30.s-z31.s }, z21.s, z20.s\n" + "uzp1 z16.b, z26.b, z30.b\n" + "uzp1 z16.b, z17.b, z16.b\n" "st1b { z16.b }, p0, [x24]\n" "18:" // Store to output array: Accumulator row 0 oddments: End "19:" // Store to output array: End @@ -374,14 +373,14 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin "mov x12, #0x0\n" "cntw x20\n" "20:" // Store to output array: Refill accumulators: Loop - ".inst 0xa040c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13]\n" - ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa041c5ac // ld1w { z12.s-z15.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n" - ".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa042c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n" - ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa043c5ac // ld1w { z12.s-z15.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n" - ".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa040c5bc // ld1w { z28.s-z31.s }, pn9.b/Z, [x13]\n" + ".inst 0xc0840780 // mova za0h.s[x12], { z28.s-z31.s }\n" + ".inst 0xa041c5b0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n" + ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa042c5ac // ld1w { z12.s-z15.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n" + ".inst 0xc0840582 // mova za2h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa043c5a0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n" + ".inst 0xc0840403 // mova za3h.s[x12], { z0.s-z3.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x13, x13, #16\n" @@ -405,4 +404,3 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |