diff options
author | Michael Tyler <michael.tyler@arm.com> | 2023-04-12 17:43:17 +0100 |
---|---|---|
committer | michael.tyler <michael.tyler@arm.com> | 2023-06-05 15:57:58 +0000 |
commit | 74921eee924625426429044decefe3673561b174 (patch) | |
tree | 654da1a95e3d42d6af8ad1ff27bb40d77b1fd8c5 /src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL | |
parent | df5d9878008be9b60586df97ebfff197abb5195e (diff) | |
download | ComputeLibrary-74921eee924625426429044decefe3673561b174.tar.gz |
Update CPU kernel implementations and guard directives
Resolves COMPMID-6023
Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp | 326 |
1 files changed, 162 insertions, 164 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp index 0d0e3da224..4601f05501 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/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_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: Loop - ".inst 0xa040c1fc // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n" - ".inst 0xc0840780 // mova za0h.s[x12], { z28.s-z31.s }\n" - ".inst 0xa041c1ec // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa042c1f4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840682 // mova za2h.s[x12], { z20.s-z23.s }\n" - ".inst 0xa043c1f8 // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840703 // mova za3h.s[x12], { z24.s-z27.s }\n" + ".inst 0xa040c1f0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n" + ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa041c1f8 // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840701 // mova za1h.s[x12], { z24.s-z27.s }\n" + ".inst 0xa042c1fc // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" + ".inst 0xc0840782 // mova za2h.s[x12], { z28.s-z31.s }\n" + ".inst 0xa043c1f4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840683 // mova za3h.s[x12], { z20.s-z23.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x15, x15, #16\n" @@ -125,11 +124,11 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "ldr x20, [%x[args], %[offsetof_bias]]\n" ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "cbz x20, 5f\n" - "ldnt1w { z15.s }, p0/Z, [x20, x10, LSL #2]\n" - ".inst 0xc09025e0 // addha za0.s, p1/M, p1/M, z15.s\n" - ".inst 0xc09025e1 // addha za1.s, p1/M, p1/M, z15.s\n" - ".inst 0xc09025e2 // addha za2.s, p1/M, p1/M, z15.s\n" - ".inst 0xc09025e3 // addha za3.s, p1/M, p1/M, z15.s\n" + "ldnt1w { z8.s }, p0/Z, [x20, x10, LSL #2]\n" + ".inst 0xc0902500 // addha za0.s, p1/M, p1/M, z8.s\n" + ".inst 0xc0902501 // addha za1.s, p1/M, p1/M, z8.s\n" + ".inst 0xc0902502 // addha za2.s, p1/M, p1/M, z8.s\n" + ".inst 0xc0902503 // addha za3.s, p1/M, p1/M, z8.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x10\n" "mov x21, x11\n" @@ -152,107 +151,107 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "madd x23, x10, x20, x23\n" // bptr = B + n * kstride_bytes "cbz x22, 8f\n" "subs x22, x22, #0x1\n" - ".inst 0xa1408372 // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n" - "ldnt1b { z0.b }, p1/Z, [x23]\n" - ".inst 0xa1418373 // ld1b { z19.b, z23.b, z27.b, z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" - "ldnt1b { z9.b }, p1/Z, [x23, #1, MUL VL]\n" - ".inst 0xa1428370 // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" - "ldnt1b { z21.b }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0xa1438362 // ld1b { z2.b, z6.b, z10.b, z14.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xa0408364 // ld1b { z4.b-z7.b }, pn8.b/Z, [x27]\n" + "ldnt1b { z14.b }, p1/Z, [x23]\n" + ".inst 0xa0418374 // ld1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "ldnt1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n" + ".inst 0xa0428378 // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "ldnt1b { z13.b }, p1/Z, [x23, #2, MUL VL]\n" + ".inst 0xa0438368 // ld1b { z8.b-z11.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" "addvl x27, x27, #16\n" - "ldnt1b { z12.b }, p1/Z, [x23, #3, MUL VL]\n" + "ldnt1b { z29.b }, p1/Z, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "ble 7f\n" "6:" // K loop - ".inst 0xa0802640 // smopa za0.s, p1/M, p1/M, z18.b, z0.b\n" + ".inst 0xa08e2480 // smopa za0.s, p1/M, p1/M, z4.b, z14.b\n" "subs x22, x22, #0x1\n" - ".inst 0xa08026c1 // smopa za1.s, p1/M, p1/M, z22.b, z0.b\n" - ".inst 0xa0802742 // smopa za2.s, p1/M, p1/M, z26.b, z0.b\n" - ".inst 0xa08027c3 // smopa za3.s, p1/M, p1/M, z30.b, z0.b\n" - ".inst 0xa1408372 // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n" - ".inst 0xa0892660 // smopa za0.s, p1/M, p1/M, z19.b, z9.b\n" - "ldnt1b { z0.b }, p1/Z, [x23]\n" - ".inst 0xa08926e1 // smopa za1.s, p1/M, p1/M, z23.b, z9.b\n" - ".inst 0xa0892762 // smopa za2.s, p1/M, p1/M, z27.b, z9.b\n" - ".inst 0xa08927e3 // smopa za3.s, p1/M, p1/M, z31.b, z9.b\n" - ".inst 0xa1418373 // ld1b { z19.b, z23.b, z27.b, z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa0952600 // smopa za0.s, p1/M, p1/M, z16.b, z21.b\n" - "ldnt1b { z9.b }, p1/Z, [x23, #1, MUL VL]\n" - ".inst 0xa0952681 // smopa za1.s, p1/M, p1/M, z20.b, z21.b\n" - ".inst 0xa0952702 // smopa za2.s, p1/M, p1/M, z24.b, z21.b\n" - ".inst 0xa0952783 // smopa za3.s, p1/M, p1/M, z28.b, z21.b\n" - ".inst 0xa1428370 // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" - "ldnt1b { z21.b }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0xa08c2440 // smopa za0.s, p1/M, p1/M, z2.b, z12.b\n" - ".inst 0xa08c24c1 // smopa za1.s, p1/M, p1/M, z6.b, z12.b\n" - ".inst 0xa08c2542 // smopa za2.s, p1/M, p1/M, z10.b, z12.b\n" - ".inst 0xa08c25c3 // smopa za3.s, p1/M, p1/M, z14.b, z12.b\n" - ".inst 0xa1438362 // ld1b { z2.b, z6.b, z10.b, z14.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xa08e24a1 // smopa za1.s, p1/M, p1/M, z5.b, z14.b\n" + ".inst 0xa08e24c2 // smopa za2.s, p1/M, p1/M, z6.b, z14.b\n" + ".inst 0xa08e24e3 // smopa za3.s, p1/M, p1/M, z7.b, z14.b\n" + ".inst 0xa0408364 // ld1b { z4.b-z7.b }, pn8.b/Z, [x27]\n" + ".inst 0xa09f2680 // smopa za0.s, p1/M, p1/M, z20.b, z31.b\n" + "ldnt1b { z14.b }, p1/Z, [x23]\n" + ".inst 0xa09f26a1 // smopa za1.s, p1/M, p1/M, z21.b, z31.b\n" + ".inst 0xa09f26c2 // smopa za2.s, p1/M, p1/M, z22.b, z31.b\n" + ".inst 0xa09f26e3 // smopa za3.s, p1/M, p1/M, z23.b, z31.b\n" + ".inst 0xa0418374 // ld1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa08d2700 // smopa za0.s, p1/M, p1/M, z24.b, z13.b\n" + "ldnt1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n" + ".inst 0xa08d2721 // smopa za1.s, p1/M, p1/M, z25.b, z13.b\n" + ".inst 0xa08d2742 // smopa za2.s, p1/M, p1/M, z26.b, z13.b\n" + ".inst 0xa08d2763 // smopa za3.s, p1/M, p1/M, z27.b, z13.b\n" + ".inst 0xa0428378 // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "ldnt1b { z13.b }, p1/Z, [x23, #2, MUL VL]\n" + ".inst 0xa09d2500 // smopa za0.s, p1/M, p1/M, z8.b, z29.b\n" + ".inst 0xa09d2521 // smopa za1.s, p1/M, p1/M, z9.b, z29.b\n" + ".inst 0xa09d2542 // smopa za2.s, p1/M, p1/M, z10.b, z29.b\n" + ".inst 0xa09d2563 // smopa za3.s, p1/M, p1/M, z11.b, z29.b\n" + ".inst 0xa0438368 // ld1b { z8.b-z11.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" "addvl x27, x27, #16\n" - "ldnt1b { z12.b }, p1/Z, [x23, #3, MUL VL]\n" + "ldnt1b { z29.b }, p1/Z, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0xa0802640 // smopa za0.s, p1/M, p1/M, z18.b, z0.b\n" - ".inst 0xa08026c1 // smopa za1.s, p1/M, p1/M, z22.b, z0.b\n" - ".inst 0xa0802742 // smopa za2.s, p1/M, p1/M, z26.b, z0.b\n" - ".inst 0xa08027c3 // smopa za3.s, p1/M, p1/M, z30.b, z0.b\n" - ".inst 0xa0892660 // smopa za0.s, p1/M, p1/M, z19.b, z9.b\n" - ".inst 0xa08926e1 // smopa za1.s, p1/M, p1/M, z23.b, z9.b\n" - ".inst 0xa0892762 // smopa za2.s, p1/M, p1/M, z27.b, z9.b\n" - ".inst 0xa08927e3 // smopa za3.s, p1/M, p1/M, z31.b, z9.b\n" - ".inst 0xa0952600 // smopa za0.s, p1/M, p1/M, z16.b, z21.b\n" - ".inst 0xa0952681 // smopa za1.s, p1/M, p1/M, z20.b, z21.b\n" - ".inst 0xa0952702 // smopa za2.s, p1/M, p1/M, z24.b, z21.b\n" - ".inst 0xa0952783 // smopa za3.s, p1/M, p1/M, z28.b, z21.b\n" - ".inst 0xa08c2440 // smopa za0.s, p1/M, p1/M, z2.b, z12.b\n" - ".inst 0xa08c24c1 // smopa za1.s, p1/M, p1/M, z6.b, z12.b\n" - ".inst 0xa08c2542 // smopa za2.s, p1/M, p1/M, z10.b, z12.b\n" - ".inst 0xa08c25c3 // smopa za3.s, p1/M, p1/M, z14.b, z12.b\n" + ".inst 0xa08e2480 // smopa za0.s, p1/M, p1/M, z4.b, z14.b\n" + ".inst 0xa08e24a1 // smopa za1.s, p1/M, p1/M, z5.b, z14.b\n" + ".inst 0xa08e24c2 // smopa za2.s, p1/M, p1/M, z6.b, z14.b\n" + ".inst 0xa08e24e3 // smopa za3.s, p1/M, p1/M, z7.b, z14.b\n" + ".inst 0xa09f2680 // smopa za0.s, p1/M, p1/M, z20.b, z31.b\n" + ".inst 0xa09f26a1 // smopa za1.s, p1/M, p1/M, z21.b, z31.b\n" + ".inst 0xa09f26c2 // smopa za2.s, p1/M, p1/M, z22.b, z31.b\n" + ".inst 0xa09f26e3 // smopa za3.s, p1/M, p1/M, z23.b, z31.b\n" + ".inst 0xa08d2700 // smopa za0.s, p1/M, p1/M, z24.b, z13.b\n" + ".inst 0xa08d2721 // smopa za1.s, p1/M, p1/M, z25.b, z13.b\n" + ".inst 0xa08d2742 // smopa za2.s, p1/M, p1/M, z26.b, z13.b\n" + ".inst 0xa08d2763 // smopa za3.s, p1/M, p1/M, z27.b, z13.b\n" + ".inst 0xa09d2500 // smopa za0.s, p1/M, p1/M, z8.b, z29.b\n" + ".inst 0xa09d2521 // smopa za1.s, p1/M, p1/M, z9.b, z29.b\n" + ".inst 0xa09d2542 // smopa za2.s, p1/M, p1/M, z10.b, z29.b\n" + ".inst 0xa09d2563 // smopa za3.s, p1/M, p1/M, z11.b, z29.b\n" "8:" // K oddments "cbz x21, 10f\n" "9:" // K oddments: Loop ".inst 0xa1408372 // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n" "subs x21, x21, #0x1\n" "addvl x27, x27, #4\n" - "ld1b { z0.b }, p1/Z, [x23]\n" + "ld1b { z15.b }, p1/Z, [x23]\n" "addvl x23, x23, #1\n" - ".inst 0xa0802640 // smopa za0.s, p1/M, p1/M, z18.b, z0.b\n" - ".inst 0xa08026c1 // smopa za1.s, p1/M, p1/M, z22.b, z0.b\n" - ".inst 0xa0802742 // smopa za2.s, p1/M, p1/M, z26.b, z0.b\n" - ".inst 0xa08027c3 // smopa za3.s, p1/M, p1/M, z30.b, z0.b\n" + ".inst 0xa08f2640 // smopa za0.s, p1/M, p1/M, z18.b, z15.b\n" + ".inst 0xa08f26c1 // smopa za1.s, p1/M, p1/M, z22.b, z15.b\n" + ".inst 0xa08f2742 // smopa za2.s, p1/M, p1/M, z26.b, z15.b\n" + ".inst 0xa08f27c3 // smopa za3.s, p1/M, p1/M, z30.b, z15.b\n" "bgt 9b\n" "10:" // K oddments: End - ".inst 0xa040c360 // ld1w { z0.s-z3.s }, pn8.b/Z, [x27]\n" + ".inst 0xa140c363 // ld1w { z3.s, z7.s, z11.s, z15.s }, pn8.b/Z, [x27]\n" "addvl x27, x27, #4\n" - ".inst 0xc0912400 // addva za0.s, p1/M, p1/M, z0.s\n" - ".inst 0xc0912421 // addva za1.s, p1/M, p1/M, z1.s\n" - ".inst 0xc0912442 // addva za2.s, p1/M, p1/M, z2.s\n" - ".inst 0xc0912463 // addva za3.s, p1/M, p1/M, z3.s\n" + ".inst 0xc0912460 // addva za0.s, p1/M, p1/M, z3.s\n" + ".inst 0xc09124e1 // addva za1.s, p1/M, p1/M, z7.s\n" + ".inst 0xc0912562 // addva za2.s, p1/M, p1/M, z11.s\n" + ".inst 0xc09125e3 // addva za3.s, p1/M, p1/M, z15.s\n" "tbz x16, #1, 14f\n" "tbz x16, #0, 12f\n" "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c1f4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n" - ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" - ".inst 0xc0840680 // mova za0h.s[x12], { z20.s-z23.s }\n" - ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" - ".inst 0xa041c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840481 // mova za1h.s[x12], { z4.s-z7.s }\n" - ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" - ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xa040c1ec // ld1w { z12.s-z15.s }, pn8.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 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" + ".inst 0xa041c1f0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" + ".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n" + ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" ".inst 0xa042c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" ".inst 0xc0840482 // mova za2h.s[x12], { z4.s-z7.s }\n" - ".inst 0xa043c1f4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840683 // mova za3h.s[x12], { z20.s-z23.s }\n" + ".inst 0xa043c1e8 // ld1w { z8.s-z11.s }, pn8.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" - ".inst 0xa060c1d0 // st1w { z16.s-z19.s }, pn8.b, [x14]\n" + ".inst 0xa060c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14]\n" "addvl x15, x15, #16\n" - ".inst 0xa061c1c8 // st1w { z8.s-z11.s }, pn8.b, [x14, #0x4, MUL VL]\n" - ".inst 0xa062c1cc // st1w { z12.s-z15.s }, pn8.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c1dc // st1w { z28.s-z31.s }, pn8.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa061c1dc // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c1d4 // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c1cc // st1w { z12.s-z15.s }, pn8.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 11b\n" "b 30f\n" @@ -260,16 +259,16 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "mov x12, #0x0\n" "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop - ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" ".inst 0xc086042c // mova { z12.s-z15.s }, za1h.s[x12]\n" - ".inst 0xa060c1d0 // st1w { z16.s-z19.s }, pn8.b, [x14]\n" - ".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n" - ".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n" + ".inst 0xa060c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14]\n" + ".inst 0xc0860444 // mova { z4.s-z7.s }, za2h.s[x12]\n" + ".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" ".inst 0xa061c1cc // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa062c1d4 // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c1d8 // st1w { z24.s-z27.s }, pn8.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa062c1c4 // st1w { z4.s-z7.s }, pn8.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 13b\n" "b 30f\n" @@ -277,22 +276,22 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "ldr x26, [%x[args], %[offsetof_C]]\n" "add x26, x26, x10\n" // C += n "sub x25, x13, x11\n" - "ld1rw { z8.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" "ldr x24, [%x[args], %[offsetof_ldcb]]\n" "madd x26, x11, x24, x26\n" // C += m * ldc - "ld1rw { z7.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z6.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n" - "ld1rw { z5.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n" - "ld1rw { z4.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n" + "ld1rw { z1.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 x16, #2, 15f\n" "ldr w21, [%x[args], %[offsetof_n_0]]\n" "add x21, x21, x10\n" "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n" "add x20, x20, x21, LSL #2\n" - "ld1w { z8.s }, p0/Z, [x20]\n" + "ld1w { z2.s }, p0/Z, [x20]\n" "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n" "add x20, x20, x21, LSL #2\n" - "ld1w { z7.s }, p0/Z, [x20]\n" + "ld1w { z1.s }, p0/Z, [x20]\n" "15:" // Store to output array: Load per-channel parameters: End "cntw x23\n" "whilelt p0.s, x10, x9\n" @@ -303,30 +302,30 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "and x20, x22, #0x3\n" "cbz x21, 17f\n" "16:" // Store to output array: Accumulator row 0 loop - ".inst 0xc086040c // mova { z12.s-z15.s }, za0h.s[x12]\n" - ".inst 0xc1a8ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z8.s\n" + ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" + ".inst 0xc1a2ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n" "add x12, x12, #0x4\n" - ".inst 0xc1a7aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" + ".inst 0xc1a1aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n" "cmp x12, x21, LSL #2\n" - ".inst 0xc1a6ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" - ".inst 0xc1a4ccac // sclamp { z12.s-z15.s }, z5.s, z4.s\n" - "st1b { z12.s }, p0, [x26]\n" + ".inst 0xc1a0ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n" + ".inst 0xc1b4ceb0 // sclamp { z16.s-z19.s }, z21.s, z20.s\n" + "st1b { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z13.s }, p0, [x26]\n" + "st1b { z17.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z14.s }, p0, [x26]\n" + "st1b { z18.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z15.s }, p0, [x26]\n" + "st1b { z19.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 16b\n" "17:" // Store to output array: Accumulator row 0 oddments "cbz x20, 18f\n" ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" - ".inst 0xc1a8ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z8.s\n" + ".inst 0xc1a2ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1a7aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z7.s\n" - ".inst 0xc1a6ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" - ".inst 0xc1a4ccb0 // sclamp { z16.s-z19.s }, z5.s, z4.s\n" + ".inst 0xc1a1aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n" + ".inst 0xc1a0ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n" + ".inst 0xc1b4ceb0 // sclamp { z16.s-z19.s }, z21.s, z20.s\n" "st1b { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 18f\n" @@ -347,38 +346,38 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "and x20, x22, #0x3\n" "cbz x21, 20f\n" "19:" // Store to output array: Accumulator row 1 loop - ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" - ".inst 0xc1a8ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z8.s\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xc1a2ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" "add x12, x12, #0x4\n" - ".inst 0xc1a7aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z7.s\n" + ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" "cmp x12, x21, LSL #2\n" - ".inst 0xc1a6ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" - ".inst 0xc1a4ccb0 // sclamp { z16.s-z19.s }, z5.s, z4.s\n" - "st1b { z16.s }, p0, [x26]\n" + ".inst 0xc1a0ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" + ".inst 0xc1b4cea4 // sclamp { z4.s-z7.s }, z21.s, z20.s\n" + "st1b { z4.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z17.s }, p0, [x26]\n" + "st1b { z5.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z18.s }, p0, [x26]\n" + "st1b { z6.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z19.s }, p0, [x26]\n" + "st1b { z7.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 19b\n" "20:" // Store to output array: Accumulator row 1 oddments "cbz x20, 21f\n" - ".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" - ".inst 0xc1a8ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z8.s\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xc1a2ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1a7aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z7.s\n" - ".inst 0xc1a6ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n" - ".inst 0xc1a4ccbc // sclamp { z28.s-z31.s }, z5.s, z4.s\n" - "st1b { z28.s }, p0, [x26]\n" + ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" + ".inst 0xc1a0ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" + ".inst 0xc1b4cea4 // sclamp { z4.s-z7.s }, z21.s, z20.s\n" + "st1b { z4.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 21f\n" "subs x20, x20, #0x1\n" - "st1b { z29.s }, p0, [x26]\n" + "st1b { z5.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 21f\n" - "st1b { z30.s }, p0, [x26]\n" + "st1b { z6.s }, p0, [x26]\n" "add x26, x26, x24\n" "21:" // Store to output array: Accumulator row 1 oddments: End "subs x25, x25, x22\n" @@ -391,30 +390,30 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "and x20, x22, #0x3\n" "cbz x21, 23f\n" "22:" // Store to output array: Accumulator row 2 loop - ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" - ".inst 0xc1a8ac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z8.s\n" + ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" + ".inst 0xc1a2ac08 // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n" "add x12, x12, #0x4\n" - ".inst 0xc1a7aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z7.s\n" + ".inst 0xc1a1aa28 // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n" "cmp x12, x21, LSL #2\n" - ".inst 0xc1a6ab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n" - ".inst 0xc1a4ccb8 // sclamp { z24.s-z27.s }, z5.s, z4.s\n" - "st1b { z24.s }, p0, [x26]\n" + ".inst 0xc1a0ab08 // add { z8.s-z11.s }, { z8.s-z11.s }, z0.s\n" + ".inst 0xc1b4cea8 // sclamp { z8.s-z11.s }, z21.s, z20.s\n" + "st1b { z8.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z25.s }, p0, [x26]\n" + "st1b { z9.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z26.s }, p0, [x26]\n" + "st1b { z10.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z27.s }, p0, [x26]\n" + "st1b { z11.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 22b\n" "23:" // Store to output array: Accumulator row 2 oddments "cbz x20, 24f\n" ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" - ".inst 0xc1a8ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z8.s\n" + ".inst 0xc1a2ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1a7aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" - ".inst 0xc1a6ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" - ".inst 0xc1a4ccac // sclamp { z12.s-z15.s }, z5.s, z4.s\n" + ".inst 0xc1a1aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" + ".inst 0xc1a0ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n" + ".inst 0xc1b4ceac // sclamp { z12.s-z15.s }, z21.s, z20.s\n" "st1b { z12.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 24f\n" @@ -435,52 +434,52 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 "and x20, x20, #0x3\n" "cbz x21, 26f\n" "25:" // Store to output array: Accumulator row 3 loop - ".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\n" - ".inst 0xc1a8ac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z8.s\n" + ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xc1a2ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n" "add x12, x12, #0x4\n" - ".inst 0xc1a7aa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z7.s\n" + ".inst 0xc1a1aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n" "cmp x12, x21, LSL #2\n" - ".inst 0xc1a6ab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z6.s\n" - ".inst 0xc1a4ccb4 // sclamp { z20.s-z23.s }, z5.s, z4.s\n" - "st1b { z20.s }, p0, [x26]\n" + ".inst 0xc1a0ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n" + ".inst 0xc1b4cebc // sclamp { z28.s-z31.s }, z21.s, z20.s\n" + "st1b { z28.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z21.s }, p0, [x26]\n" + "st1b { z29.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z22.s }, p0, [x26]\n" + "st1b { z30.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1b { z23.s }, p0, [x26]\n" + "st1b { z31.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 25b\n" "26:" // Store to output array: Accumulator row 3 oddments "cbz x20, 27f\n" - ".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" - ".inst 0xc1a8ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z8.s\n" + ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xc1a2ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1a7aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z7.s\n" - ".inst 0xc1a6ab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" - ".inst 0xc1a4cca0 // sclamp { z0.s-z3.s }, z5.s, z4.s\n" - "st1b { z0.s }, p0, [x26]\n" + ".inst 0xc1a1aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n" + ".inst 0xc1a0ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n" + ".inst 0xc1b4cebc // sclamp { z28.s-z31.s }, z21.s, z20.s\n" + "st1b { z28.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 27f\n" "subs x20, x20, #0x1\n" - "st1b { z1.s }, p0, [x26]\n" + "st1b { z29.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 27f\n" - "st1b { z2.s }, p0, [x26]\n" + "st1b { z30.s }, p0, [x26]\n" "27:" // Store to output array: Accumulator row 3 oddments: End "28:" // Store to output array: End "tbz x16, #0, 30f\n" "mov x12, #0x0\n" "cntw x20\n" "29:" // Store to output array: Refill accumulators: Loop - ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n" - ".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa040c1fc // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n" + ".inst 0xc0840780 // mova za0h.s[x12], { z28.s-z31.s }\n" ".inst 0xa041c1f0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" ".inst 0xa042c1f0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa043c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa043c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #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 x15, x15, #16\n" @@ -504,4 +503,3 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8 } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |