diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp | 296 |
1 files changed, 147 insertions, 149 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp index 0b642818e2..7033de5fe3 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/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_2VLx2VL(const int8_t *const A, const int8 "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: Loop - ".inst 0xa040c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n" - ".inst 0xc0840400 // mova za0h.s[x12], { z0.s-z3.s }\n" - ".inst 0xa041c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n" - ".inst 0xa042c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840402 // mova za2h.s[x12], { z0.s-z3.s }\n" - ".inst 0xa043c5fc // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840783 // mova za3h.s[x12], { z28.s-z31.s }\n" + ".inst 0xa040c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n" + ".inst 0xc0840500 // mova za0h.s[x12], { z8.s-z11.s }\n" + ".inst 0xa041c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840401 // mova za1h.s[x12], { z0.s-z3.s }\n" + ".inst 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 0xa043c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #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 x15, x15, #16\n" @@ -125,11 +124,11 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(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" - ".inst 0xa00a4295 // ldnt1w { z20.s-z21.s }, p8/Z, [x20, x10, LSL #2]\n" - ".inst 0xc0902680 // addha za0.s, p1/M, p1/M, z20.s\n" - ".inst 0xc09026a1 // addha za1.s, p1/M, p1/M, z21.s\n" - ".inst 0xc0902682 // addha za2.s, p1/M, p1/M, z20.s\n" - ".inst 0xc09026a3 // addha za3.s, p1/M, p1/M, z21.s\n" + ".inst 0xa00a4299 // ldnt1w { z24.s-z25.s }, p8/Z, [x20, x10, 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 0xc0902702 // addha za2.s, p1/M, p1/M, z24.s\n" + ".inst 0xc0902723 // addha za3.s, p1/M, p1/M, z25.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x10\n" "mov x21, x11\n" @@ -152,75 +151,75 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(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 0xa040077e // ld1b { z30.b-z31.b }, pn9.b/Z, [x27]\n" - ".inst 0xa04006f1 // ldnt1b { z16.b-z17.b }, pn9.b/Z, [x23]\n" - ".inst 0xa041076e // ld1b { z14.b-z15.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" - ".inst 0xa04106e9 // ldnt1b { z8.b-z9.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xa0420760 // ld1b { z0.b-z1.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa14206fc // ldnt1b { z20.b, z28.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa0430764 // ld1b { z4.b-z5.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" + ".inst 0xa1400763 // ld1b { z3.b, z11.b }, pn9.b/Z, [x27]\n" + ".inst 0xa14006f9 // ldnt1b { z17.b, z25.b }, pn9.b/Z, [x23]\n" + ".inst 0xa1410774 // ld1b { z20.b, z28.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" + ".inst 0xa04106f7 // ldnt1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa1420775 // ld1b { z21.b, z29.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa14206f8 // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa1430765 // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" "addvl x27, x27, #8\n" - ".inst 0xa14306ea // ldnt1b { z2.b, z10.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" + ".inst 0xa14306ef // ldnt1b { z7.b, z15.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" "addvl x23, x23, #8\n" "ble 7f\n" "6:" // K loop - ".inst 0xa09027c0 // smopa za0.s, p1/M, p1/M, z30.b, z16.b\n" + ".inst 0xa0912460 // smopa za0.s, p1/M, p1/M, z3.b, z17.b\n" "subs x22, x22, #0x1\n" - ".inst 0xa09127c1 // smopa za1.s, p1/M, p1/M, z30.b, z17.b\n" - ".inst 0xa09027e2 // smopa za2.s, p1/M, p1/M, z31.b, z16.b\n" - ".inst 0xa09127e3 // smopa za3.s, p1/M, p1/M, z31.b, z17.b\n" - ".inst 0xa040077e // ld1b { z30.b-z31.b }, pn9.b/Z, [x27]\n" - ".inst 0xa08825c0 // smopa za0.s, p1/M, p1/M, z14.b, z8.b\n" - ".inst 0xa04006f1 // ldnt1b { z16.b-z17.b }, pn9.b/Z, [x23]\n" - ".inst 0xa08925c1 // smopa za1.s, p1/M, p1/M, z14.b, z9.b\n" - ".inst 0xa08825e2 // smopa za2.s, p1/M, p1/M, z15.b, z8.b\n" - ".inst 0xa08925e3 // smopa za3.s, p1/M, p1/M, z15.b, z9.b\n" - ".inst 0xa041076e // ld1b { z14.b-z15.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" - ".inst 0xa0942400 // smopa za0.s, p1/M, p1/M, z0.b, z20.b\n" - ".inst 0xa04106e9 // ldnt1b { z8.b-z9.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" - ".inst 0xa09c2401 // smopa za1.s, p1/M, p1/M, z0.b, z28.b\n" - ".inst 0xa0942422 // smopa za2.s, p1/M, p1/M, z1.b, z20.b\n" - ".inst 0xa09c2423 // smopa za3.s, p1/M, p1/M, z1.b, z28.b\n" - ".inst 0xa0420760 // ld1b { z0.b-z1.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa14206fc // ldnt1b { z20.b, z28.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa0822480 // smopa za0.s, p1/M, p1/M, z4.b, z2.b\n" - ".inst 0xa08a2481 // smopa za1.s, p1/M, p1/M, z4.b, z10.b\n" - ".inst 0xa08224a2 // smopa za2.s, p1/M, p1/M, z5.b, z2.b\n" - ".inst 0xa08a24a3 // smopa za3.s, p1/M, p1/M, z5.b, z10.b\n" - ".inst 0xa0430764 // ld1b { z4.b-z5.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" + ".inst 0xa0992461 // smopa za1.s, p1/M, p1/M, z3.b, z25.b\n" + ".inst 0xa0912562 // smopa za2.s, p1/M, p1/M, z11.b, z17.b\n" + ".inst 0xa0992563 // smopa za3.s, p1/M, p1/M, z11.b, z25.b\n" + ".inst 0xa1400763 // ld1b { z3.b, z11.b }, pn9.b/Z, [x27]\n" + ".inst 0xa0962680 // smopa za0.s, p1/M, p1/M, z20.b, z22.b\n" + ".inst 0xa14006f9 // ldnt1b { z17.b, z25.b }, pn9.b/Z, [x23]\n" + ".inst 0xa0972681 // smopa za1.s, p1/M, p1/M, z20.b, z23.b\n" + ".inst 0xa0962782 // smopa za2.s, p1/M, p1/M, z28.b, z22.b\n" + ".inst 0xa0972783 // smopa za3.s, p1/M, p1/M, z28.b, z23.b\n" + ".inst 0xa1410774 // ld1b { z20.b, z28.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n" + ".inst 0xa09026a0 // smopa za0.s, p1/M, p1/M, z21.b, z16.b\n" + ".inst 0xa04106f7 // ldnt1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n" + ".inst 0xa09826a1 // smopa za1.s, p1/M, p1/M, z21.b, z24.b\n" + ".inst 0xa09027a2 // smopa za2.s, p1/M, p1/M, z29.b, z16.b\n" + ".inst 0xa09827a3 // smopa za3.s, p1/M, p1/M, z29.b, z24.b\n" + ".inst 0xa1420775 // ld1b { z21.b, z29.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa14206f8 // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa08724a0 // smopa za0.s, p1/M, p1/M, z5.b, z7.b\n" + ".inst 0xa08f24a1 // smopa za1.s, p1/M, p1/M, z5.b, z15.b\n" + ".inst 0xa08725a2 // smopa za2.s, p1/M, p1/M, z13.b, z7.b\n" + ".inst 0xa08f25a3 // smopa za3.s, p1/M, p1/M, z13.b, z15.b\n" + ".inst 0xa1430765 // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n" "addvl x27, x27, #8\n" - ".inst 0xa14306ea // ldnt1b { z2.b, z10.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" + ".inst 0xa14306ef // ldnt1b { z7.b, z15.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n" "addvl x23, x23, #8\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0xa09027c0 // smopa za0.s, p1/M, p1/M, z30.b, z16.b\n" - ".inst 0xa09127c1 // smopa za1.s, p1/M, p1/M, z30.b, z17.b\n" - ".inst 0xa09027e2 // smopa za2.s, p1/M, p1/M, z31.b, z16.b\n" - ".inst 0xa09127e3 // smopa za3.s, p1/M, p1/M, z31.b, z17.b\n" - ".inst 0xa08825c0 // smopa za0.s, p1/M, p1/M, z14.b, z8.b\n" - ".inst 0xa08925c1 // smopa za1.s, p1/M, p1/M, z14.b, z9.b\n" - ".inst 0xa08825e2 // smopa za2.s, p1/M, p1/M, z15.b, z8.b\n" - ".inst 0xa08925e3 // smopa za3.s, p1/M, p1/M, z15.b, z9.b\n" - ".inst 0xa0942400 // smopa za0.s, p1/M, p1/M, z0.b, z20.b\n" - ".inst 0xa09c2401 // smopa za1.s, p1/M, p1/M, z0.b, z28.b\n" - ".inst 0xa0942422 // smopa za2.s, p1/M, p1/M, z1.b, z20.b\n" - ".inst 0xa09c2423 // smopa za3.s, p1/M, p1/M, z1.b, z28.b\n" - ".inst 0xa0822480 // smopa za0.s, p1/M, p1/M, z4.b, z2.b\n" - ".inst 0xa08a2481 // smopa za1.s, p1/M, p1/M, z4.b, z10.b\n" - ".inst 0xa08224a2 // smopa za2.s, p1/M, p1/M, z5.b, z2.b\n" - ".inst 0xa08a24a3 // smopa za3.s, p1/M, p1/M, z5.b, z10.b\n" + ".inst 0xa0912460 // smopa za0.s, p1/M, p1/M, z3.b, z17.b\n" + ".inst 0xa0992461 // smopa za1.s, p1/M, p1/M, z3.b, z25.b\n" + ".inst 0xa0912562 // smopa za2.s, p1/M, p1/M, z11.b, z17.b\n" + ".inst 0xa0992563 // smopa za3.s, p1/M, p1/M, z11.b, z25.b\n" + ".inst 0xa0962680 // smopa za0.s, p1/M, p1/M, z20.b, z22.b\n" + ".inst 0xa0972681 // smopa za1.s, p1/M, p1/M, z20.b, z23.b\n" + ".inst 0xa0962782 // smopa za2.s, p1/M, p1/M, z28.b, z22.b\n" + ".inst 0xa0972783 // smopa za3.s, p1/M, p1/M, z28.b, z23.b\n" + ".inst 0xa09026a0 // smopa za0.s, p1/M, p1/M, z21.b, z16.b\n" + ".inst 0xa09826a1 // smopa za1.s, p1/M, p1/M, z21.b, z24.b\n" + ".inst 0xa09027a2 // smopa za2.s, p1/M, p1/M, z29.b, z16.b\n" + ".inst 0xa09827a3 // smopa za3.s, p1/M, p1/M, z29.b, z24.b\n" + ".inst 0xa08724a0 // smopa za0.s, p1/M, p1/M, z5.b, z7.b\n" + ".inst 0xa08f24a1 // smopa za1.s, p1/M, p1/M, z5.b, z15.b\n" + ".inst 0xa08725a2 // smopa za2.s, p1/M, p1/M, z13.b, z7.b\n" + ".inst 0xa08f25a3 // smopa za3.s, p1/M, p1/M, z13.b, z15.b\n" "8:" // K oddments "cbz x21, 10f\n" "9:" // K oddments: Loop - ".inst 0xa040077e // ld1b { z30.b-z31.b }, pn9.b/Z, [x27]\n" + ".inst 0xa1400773 // ld1b { z19.b, z27.b }, pn9.b/Z, [x27]\n" "subs x21, x21, #0x1\n" "addvl x27, x27, #2\n" ".inst 0xa04006f0 // ld1b { z16.b-z17.b }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" - ".inst 0xa09027c0 // smopa za0.s, p1/M, p1/M, z30.b, z16.b\n" - ".inst 0xa09127c1 // smopa za1.s, p1/M, p1/M, z30.b, z17.b\n" - ".inst 0xa09027e2 // smopa za2.s, p1/M, p1/M, z31.b, z16.b\n" - ".inst 0xa09127e3 // smopa za3.s, p1/M, p1/M, z31.b, z17.b\n" + ".inst 0xa0902660 // smopa za0.s, p1/M, p1/M, z19.b, z16.b\n" + ".inst 0xa0912661 // smopa za1.s, p1/M, p1/M, z19.b, z17.b\n" + ".inst 0xa0902762 // smopa za2.s, p1/M, p1/M, z27.b, z16.b\n" + ".inst 0xa0912763 // smopa za3.s, p1/M, p1/M, z27.b, z17.b\n" "bgt 9b\n" "10:" // K oddments: End ".inst 0xa040476e // ld1w { z14.s-z15.s }, pn9.b/Z, [x27]\n" @@ -234,25 +233,25 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8 "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c5fc // ld1w { z28.s-z31.s }, pn9.b/Z, [x15]\n" - ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" - ".inst 0xc0840780 // mova za0h.s[x12], { z28.s-z31.s }\n" - ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" - ".inst 0xa041c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" - ".inst 0xc0840701 // mova za1h.s[x12], { z24.s-z27.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 0xa042c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840702 // mova za2h.s[x12], { z24.s-z27.s }\n" - ".inst 0xa043c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa040c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15]\n" + ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" + ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n" + ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" + ".inst 0xa041c5fc // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840781 // mova za1h.s[x12], { z28.s-z31.s }\n" + ".inst 0xc0860444 // mova { z4.s-z7.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 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa043c5fc // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840783 // mova za3h.s[x12], { z28.s-z31.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa060c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14]\n" + ".inst 0xa060c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14]\n" "addvl x15, x15, #16\n" - ".inst 0xa061c5d4 // st1w { z20.s-z23.s }, pn9.b, [x14, #0x4, MUL VL]\n" - ".inst 0xa062c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa061c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c5c4 // st1w { z4.s-z7.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" "b 24f\n" @@ -260,16 +259,16 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(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 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" - ".inst 0xa060c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14]\n" - ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xa061c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xc086040c // mova { z12.s-z15.s }, za0h.s[x12]\n" + ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" + ".inst 0xa060c5cc // st1w { z12.s-z15.s }, pn9.b, [x14]\n" + ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" + ".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\n" + ".inst 0xa061c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa062c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa062c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c5d4 // st1w { z20.s-z23.s }, pn9.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 13b\n" "b 24f\n" @@ -277,13 +276,13 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(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 { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z0.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 { z3.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" - "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" - "ld1rw { z11.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n" + "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n" + "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z3.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n" + "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n" "ld1rw { z25.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n" "ld1rw { z24.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n" "tbz x16, #2, 15f\n" @@ -291,10 +290,10 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8 "add x21, x21, x10\n" "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n" "add x20, x20, x21, LSL #2\n" - ".inst 0xa0404282 // ld1w { z2.s-z3.s }, p8/Z, [x20]\n" + ".inst 0xa0404280 // ld1w { z0.s-z1.s }, p8/Z, [x20]\n" "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n" "add x20, x20, x21, LSL #2\n" - ".inst 0xa0404280 // ld1w { z0.s-z1.s }, p8/Z, [x20]\n" + ".inst 0xa0404282 // ld1w { z2.s-z3.s }, p8/Z, [x20]\n" "15:" // Store to output array: Load per-channel parameters: End "cntw x23\n" "whilelt p0.h, x10, x9\n" @@ -305,26 +304,26 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(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 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" - ".inst 0xc1a2ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" - ".inst 0xc1a3ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z3.s\n" + ".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n" + ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" + ".inst 0xc1a0ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" + ".inst 0xc1a1ac08 // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n" "add x12, x12, #0x4\n" "cmp x12, x21, LSL #2\n" - ".inst 0xc1a0aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n" - ".inst 0xc1a1aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n" - ".inst 0xc1abab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z11.s\n" - ".inst 0xc1abab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z11.s\n" - ".inst 0xc1b8cf2c // sclamp { z12.s-z15.s }, z25.s, z24.s\n" - ".inst 0xc1b8cf3c // sclamp { z28.s-z31.s }, z25.s, z24.s\n" - "uzp1 z16.h, z12.h, z28.h\n" + ".inst 0xc1a2aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" + ".inst 0xc1a3aa28 // srshl { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n" + ".inst 0xc1aeab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n" + ".inst 0xc1aeab08 // add { z8.s-z11.s }, { z8.s-z11.s }, z14.s\n" + ".inst 0xc1b8cf24 // sclamp { z4.s-z7.s }, z25.s, z24.s\n" + ".inst 0xc1b8cf28 // sclamp { z8.s-z11.s }, z25.s, z24.s\n" + "uzp1 z16.h, z4.h, z8.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" - "uzp1 z16.h, z13.h, z29.h\n" - "uzp1 z17.h, z14.h, z30.h\n" + "uzp1 z16.h, z5.h, z9.h\n" + "uzp1 z17.h, z6.h, z10.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" - "uzp1 z16.h, z15.h, z31.h\n" + "uzp1 z16.h, z7.h, z11.h\n" "st1b { z17.h }, p0, [x26]\n" "add x26, x26, x24\n" "st1b { z16.h }, p0, [x26]\n" @@ -332,27 +331,27 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8 "blt 16b\n" "17:" // Store to output array: Accumulator row 0 oddments "cbz x20, 18f\n" - ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" - ".inst 0xc086042c // mova { z12.s-z15.s }, za1h.s[x12]\n" - ".inst 0xc1a2ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n" - ".inst 0xc1a3ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n" + ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xc1a0ac08 // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z0.s\n" + ".inst 0xc1a1ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1a0aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n" - ".inst 0xc1a1aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" - ".inst 0xc1abab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z11.s\n" - ".inst 0xc1abab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z11.s\n" - ".inst 0xc1b8cf3c // sclamp { z28.s-z31.s }, z25.s, z24.s\n" - ".inst 0xc1b8cf2c // sclamp { z12.s-z15.s }, z25.s, z24.s\n" - "uzp1 z16.h, z28.h, z12.h\n" + ".inst 0xc1a2aa28 // srshl { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n" + ".inst 0xc1a3aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" + ".inst 0xc1aeab08 // add { z8.s-z11.s }, { z8.s-z11.s }, z14.s\n" + ".inst 0xc1aeab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n" + ".inst 0xc1b8cf28 // sclamp { z8.s-z11.s }, z25.s, z24.s\n" + ".inst 0xc1b8cf24 // sclamp { z4.s-z7.s }, z25.s, z24.s\n" + "uzp1 z16.h, z8.h, z4.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" "beq 18f\n" "subs x20, x20, #0x1\n" - "uzp1 z16.h, z29.h, z13.h\n" + "uzp1 z16.h, z9.h, z5.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" "beq 18f\n" - "uzp1 z16.h, z30.h, z14.h\n" + "uzp1 z16.h, z10.h, z6.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" "18:" // Store to output array: Accumulator row 0 oddments: End @@ -367,25 +366,25 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8 "cbz x21, 20f\n" "19:" // Store to output array: Accumulator row 1 loop ".inst 0xc0860444 // mova { z4.s-z7.s }, za2h.s[x12]\n" - ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" - ".inst 0xc1a2ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" - ".inst 0xc1a3ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z3.s\n" + ".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\n" + ".inst 0xc1a0ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" + ".inst 0xc1a1ac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z1.s\n" "add x12, x12, #0x4\n" "cmp x12, x21, LSL #2\n" - ".inst 0xc1a0aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" - ".inst 0xc1a1aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n" - ".inst 0xc1abab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z11.s\n" - ".inst 0xc1abab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n" + ".inst 0xc1a2aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" + ".inst 0xc1a3aa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z3.s\n" + ".inst 0xc1aeab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n" + ".inst 0xc1aeab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z14.s\n" ".inst 0xc1b8cf24 // sclamp { z4.s-z7.s }, z25.s, z24.s\n" - ".inst 0xc1b8cf30 // sclamp { z16.s-z19.s }, z25.s, z24.s\n" - "uzp1 z16.h, z4.h, z16.h\n" + ".inst 0xc1b8cf34 // sclamp { z20.s-z23.s }, z25.s, z24.s\n" + "uzp1 z16.h, z4.h, z20.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" - "uzp1 z16.h, z5.h, z17.h\n" - "uzp1 z17.h, z6.h, z18.h\n" + "uzp1 z16.h, z5.h, z21.h\n" + "uzp1 z17.h, z6.h, z22.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" - "uzp1 z16.h, z7.h, z19.h\n" + "uzp1 z16.h, z7.h, z23.h\n" "st1b { z17.h }, p0, [x26]\n" "add x26, x26, x24\n" "st1b { z16.h }, p0, [x26]\n" @@ -393,27 +392,27 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8 "blt 19b\n" "20:" // Store to output array: Accumulator row 1 oddments "cbz x20, 21f\n" - ".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n" + ".inst 0xc0860444 // mova { z4.s-z7.s }, za2h.s[x12]\n" ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" - ".inst 0xc1a2ac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z2.s\n" - ".inst 0xc1a3ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z3.s\n" + ".inst 0xc1a0ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n" + ".inst 0xc1a1ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n" "subs x20, x20, #0x1\n" - ".inst 0xc1a0aa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z0.s\n" - ".inst 0xc1a1aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n" - ".inst 0xc1abab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z11.s\n" - ".inst 0xc1abab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n" - ".inst 0xc1b8cf34 // sclamp { z20.s-z23.s }, z25.s, z24.s\n" + ".inst 0xc1a2aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" + ".inst 0xc1a3aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z3.s\n" + ".inst 0xc1aeab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n" + ".inst 0xc1aeab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z14.s\n" + ".inst 0xc1b8cf24 // sclamp { z4.s-z7.s }, z25.s, z24.s\n" ".inst 0xc1b8cf30 // sclamp { z16.s-z19.s }, z25.s, z24.s\n" - "uzp1 z16.h, z20.h, z16.h\n" + "uzp1 z16.h, z4.h, z16.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" "beq 21f\n" "subs x20, x20, #0x1\n" - "uzp1 z16.h, z21.h, z17.h\n" + "uzp1 z16.h, z5.h, z17.h\n" "st1b { z16.h }, p0, [x26]\n" "add x26, x26, x24\n" "beq 21f\n" - "uzp1 z16.h, z22.h, z18.h\n" + "uzp1 z16.h, z6.h, z18.h\n" "st1b { z16.h }, p0, [x26]\n" "21:" // Store to output array: Accumulator row 1 oddments: End "22:" // Store to output array: End @@ -452,4 +451,3 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8 } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |