diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp | 202 |
1 files changed, 99 insertions, 103 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp index 12e714a471..d11faa634d 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_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" @@ -32,10 +31,8 @@ namespace arm_gemm { -void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const int8_t *const B, int32_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Activation act, bool accumulate, int32_t *const accumulator_buffer) +void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const int8_t *const B, int32_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Activation, bool accumulate, int32_t *const accumulator_buffer) { - ARM_COMPUTE_UNUSED(act); - struct KernelArgs { KernelArgs( @@ -96,12 +93,12 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in "1:" // Initial accumulator load from buffer: Loop ".inst 0xa040c564 // ld1w { z4.s-z7.s }, pn9.b/Z, [x11]\n" ".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n" - ".inst 0xa041c57c // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n" - ".inst 0xc0840781 // mova za1h.s[x12], { z28.s-z31.s }\n" - ".inst 0xa042c570 // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n" - ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa043c560 // ld1w { z0.s-z3.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n" - ".inst 0xc0840403 // mova za3h.s[x12], { z0.s-z3.s }\n" + ".inst 0xa041c56c // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n" + ".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa042c57c // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n" + ".inst 0xc0840782 // mova za2h.s[x12], { z28.s-z31.s }\n" + ".inst 0xa043c574 // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #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 x11, x11, #16\n" @@ -119,11 +116,11 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in "ldr x20, [%x[args], %[offsetof_bias]]\n" ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "cbz x20, 5f\n" - ".inst 0xa11bc28a // ldnt1w { z2.s, z6.s, z10.s, z14.s }, p8/Z, [x20, x27, LSL #2]\n" - ".inst 0xc0900040 // addha za0.s, p0/M, p0/M, z2.s\n" - ".inst 0xc09000c1 // addha za1.s, p0/M, p0/M, z6.s\n" - ".inst 0xc0900142 // addha za2.s, p0/M, p0/M, z10.s\n" - ".inst 0xc09001c3 // addha za3.s, p0/M, p0/M, z14.s\n" + ".inst 0xa11bc29b // ldnt1w { z19.s, z23.s, z27.s, z31.s }, p8/Z, [x20, x27, LSL #2]\n" + ".inst 0xc0900260 // addha za0.s, p0/M, p0/M, z19.s\n" + ".inst 0xc09002e1 // addha za1.s, p0/M, p0/M, z23.s\n" + ".inst 0xc0900362 // addha za2.s, p0/M, p0/M, z27.s\n" + ".inst 0xc09003e3 // addha za3.s, p0/M, p0/M, z31.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x27\n" "mov x21, x28\n" @@ -146,75 +143,75 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in "madd x23, x27, x20, x23\n" // bptr = B + n * kstride_bytes "cbz x22, 8f\n" "subs x22, x22, #0x1\n" - "ld1b { z20.b }, p0/Z, [x24]\n" - ".inst 0xa14086e9 // ldnt1b { z1.b, z5.b, z9.b, z13.b }, pn9.b/Z, [x23]\n" - "ld1b { z10.b }, p0/Z, [x24, #1, MUL VL]\n" - ".inst 0xa14186fa // ldnt1b { z18.b, z22.b, z26.b, z30.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - "ld1b { z16.b }, p0/Z, [x24, #2, MUL VL]\n" - ".inst 0xa14286eb // ldnt1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" - "ld1b { z25.b }, p0/Z, [x24, #3, MUL VL]\n" + "ld1b { z30.b }, p0/Z, [x24]\n" + ".inst 0xa04086e1 // ldnt1b { z0.b-z3.b }, pn9.b/Z, [x23]\n" + "ld1b { z21.b }, p0/Z, [x24, #1, MUL VL]\n" + ".inst 0xa04186f9 // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + "ld1b { z28.b }, p0/Z, [x24, #2, MUL VL]\n" + ".inst 0xa04286e5 // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" + "ld1b { z11.b }, p0/Z, [x24, #3, MUL VL]\n" "addvl x24, x24, #4\n" - ".inst 0xa14386e8 // ldnt1b { z0.b, z4.b, z8.b, z12.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 0xa0810280 // smopa za0.s, p0/M, p0/M, z20.b, z1.b\n" + ".inst 0xa08003c0 // smopa za0.s, p0/M, p0/M, z30.b, z0.b\n" "subs x22, x22, #0x1\n" - ".inst 0xa0850281 // smopa za1.s, p0/M, p0/M, z20.b, z5.b\n" - ".inst 0xa0890282 // smopa za2.s, p0/M, p0/M, z20.b, z9.b\n" - ".inst 0xa08d0283 // smopa za3.s, p0/M, p0/M, z20.b, z13.b\n" - "ld1b { z20.b }, p0/Z, [x24]\n" - ".inst 0xa0920140 // smopa za0.s, p0/M, p0/M, z10.b, z18.b\n" - ".inst 0xa14086e9 // ldnt1b { z1.b, z5.b, z9.b, z13.b }, pn9.b/Z, [x23]\n" - ".inst 0xa0960141 // smopa za1.s, p0/M, p0/M, z10.b, z22.b\n" - ".inst 0xa09a0142 // smopa za2.s, p0/M, p0/M, z10.b, z26.b\n" - ".inst 0xa09e0143 // smopa za3.s, p0/M, p0/M, z10.b, z30.b\n" - "ld1b { z10.b }, p0/Z, [x24, #1, MUL VL]\n" - ".inst 0xa0830200 // smopa za0.s, p0/M, p0/M, z16.b, z3.b\n" - ".inst 0xa14186fa // ldnt1b { z18.b, z22.b, z26.b, z30.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" - ".inst 0xa0870201 // smopa za1.s, p0/M, p0/M, z16.b, z7.b\n" - ".inst 0xa08b0202 // smopa za2.s, p0/M, p0/M, z16.b, z11.b\n" - ".inst 0xa08f0203 // smopa za3.s, p0/M, p0/M, z16.b, z15.b\n" - "ld1b { z16.b }, p0/Z, [x24, #2, MUL VL]\n" - ".inst 0xa14286eb // ldnt1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" - ".inst 0xa0800320 // smopa za0.s, p0/M, p0/M, z25.b, z0.b\n" - ".inst 0xa0840321 // smopa za1.s, p0/M, p0/M, z25.b, z4.b\n" - ".inst 0xa0880322 // smopa za2.s, p0/M, p0/M, z25.b, z8.b\n" - ".inst 0xa08c0323 // smopa za3.s, p0/M, p0/M, z25.b, z12.b\n" - "ld1b { z25.b }, p0/Z, [x24, #3, MUL VL]\n" + ".inst 0xa08103c1 // smopa za1.s, p0/M, p0/M, z30.b, z1.b\n" + ".inst 0xa08203c2 // smopa za2.s, p0/M, p0/M, z30.b, z2.b\n" + ".inst 0xa08303c3 // smopa za3.s, p0/M, p0/M, z30.b, z3.b\n" + "ld1b { z30.b }, p0/Z, [x24]\n" + ".inst 0xa09802a0 // smopa za0.s, p0/M, p0/M, z21.b, z24.b\n" + ".inst 0xa04086e1 // ldnt1b { z0.b-z3.b }, pn9.b/Z, [x23]\n" + ".inst 0xa09902a1 // smopa za1.s, p0/M, p0/M, z21.b, z25.b\n" + ".inst 0xa09a02a2 // smopa za2.s, p0/M, p0/M, z21.b, z26.b\n" + ".inst 0xa09b02a3 // smopa za3.s, p0/M, p0/M, z21.b, z27.b\n" + "ld1b { z21.b }, p0/Z, [x24, #1, MUL VL]\n" + ".inst 0xa0840380 // smopa za0.s, p0/M, p0/M, z28.b, z4.b\n" + ".inst 0xa04186f9 // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n" + ".inst 0xa0850381 // smopa za1.s, p0/M, p0/M, z28.b, z5.b\n" + ".inst 0xa0860382 // smopa za2.s, p0/M, p0/M, z28.b, z6.b\n" + ".inst 0xa0870383 // smopa za3.s, p0/M, p0/M, z28.b, z7.b\n" + "ld1b { z28.b }, p0/Z, [x24, #2, MUL VL]\n" + ".inst 0xa04286e5 // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n" + ".inst 0xa0900160 // smopa za0.s, p0/M, p0/M, z11.b, z16.b\n" + ".inst 0xa0910161 // smopa za1.s, p0/M, p0/M, z11.b, z17.b\n" + ".inst 0xa0920162 // smopa za2.s, p0/M, p0/M, z11.b, z18.b\n" + ".inst 0xa0930163 // smopa za3.s, p0/M, p0/M, z11.b, z19.b\n" + "ld1b { z11.b }, p0/Z, [x24, #3, MUL VL]\n" "addvl x24, x24, #4\n" - ".inst 0xa14386e8 // ldnt1b { z0.b, z4.b, z8.b, z12.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 0xa0810280 // smopa za0.s, p0/M, p0/M, z20.b, z1.b\n" - ".inst 0xa0850281 // smopa za1.s, p0/M, p0/M, z20.b, z5.b\n" - ".inst 0xa0890282 // smopa za2.s, p0/M, p0/M, z20.b, z9.b\n" - ".inst 0xa08d0283 // smopa za3.s, p0/M, p0/M, z20.b, z13.b\n" - ".inst 0xa0920140 // smopa za0.s, p0/M, p0/M, z10.b, z18.b\n" - ".inst 0xa0960141 // smopa za1.s, p0/M, p0/M, z10.b, z22.b\n" - ".inst 0xa09a0142 // smopa za2.s, p0/M, p0/M, z10.b, z26.b\n" - ".inst 0xa09e0143 // smopa za3.s, p0/M, p0/M, z10.b, z30.b\n" - ".inst 0xa0830200 // smopa za0.s, p0/M, p0/M, z16.b, z3.b\n" - ".inst 0xa0870201 // smopa za1.s, p0/M, p0/M, z16.b, z7.b\n" - ".inst 0xa08b0202 // smopa za2.s, p0/M, p0/M, z16.b, z11.b\n" - ".inst 0xa08f0203 // smopa za3.s, p0/M, p0/M, z16.b, z15.b\n" - ".inst 0xa0800320 // smopa za0.s, p0/M, p0/M, z25.b, z0.b\n" - ".inst 0xa0840321 // smopa za1.s, p0/M, p0/M, z25.b, z4.b\n" - ".inst 0xa0880322 // smopa za2.s, p0/M, p0/M, z25.b, z8.b\n" - ".inst 0xa08c0323 // smopa za3.s, p0/M, p0/M, z25.b, z12.b\n" + ".inst 0xa08003c0 // smopa za0.s, p0/M, p0/M, z30.b, z0.b\n" + ".inst 0xa08103c1 // smopa za1.s, p0/M, p0/M, z30.b, z1.b\n" + ".inst 0xa08203c2 // smopa za2.s, p0/M, p0/M, z30.b, z2.b\n" + ".inst 0xa08303c3 // smopa za3.s, p0/M, p0/M, z30.b, z3.b\n" + ".inst 0xa09802a0 // smopa za0.s, p0/M, p0/M, z21.b, z24.b\n" + ".inst 0xa09902a1 // smopa za1.s, p0/M, p0/M, z21.b, z25.b\n" + ".inst 0xa09a02a2 // smopa za2.s, p0/M, p0/M, z21.b, z26.b\n" + ".inst 0xa09b02a3 // smopa za3.s, p0/M, p0/M, z21.b, z27.b\n" + ".inst 0xa0840380 // smopa za0.s, p0/M, p0/M, z28.b, z4.b\n" + ".inst 0xa0850381 // smopa za1.s, p0/M, p0/M, z28.b, z5.b\n" + ".inst 0xa0860382 // smopa za2.s, p0/M, p0/M, z28.b, z6.b\n" + ".inst 0xa0870383 // smopa za3.s, p0/M, p0/M, z28.b, z7.b\n" + ".inst 0xa0900160 // smopa za0.s, p0/M, p0/M, z11.b, z16.b\n" + ".inst 0xa0910161 // smopa za1.s, p0/M, p0/M, z11.b, z17.b\n" + ".inst 0xa0920162 // smopa za2.s, p0/M, p0/M, z11.b, z18.b\n" + ".inst 0xa0930163 // smopa za3.s, p0/M, p0/M, z11.b, z19.b\n" "8:" // K oddments "cbz x21, 10f\n" "9:" // K oddments: Loop - "ld1b { z20.b }, p0/Z, [x24]\n" + "ld1b { z22.b }, p0/Z, [x24]\n" "subs x21, x21, #0x1\n" "addvl x24, x24, #1\n" - ".inst 0xa14086e1 // ld1b { z1.b, z5.b, z9.b, z13.b }, pn9.b/Z, [x23]\n" + ".inst 0xa14086f1 // ld1b { z17.b, z21.b, z25.b, z29.b }, pn9.b/Z, [x23]\n" "addvl x23, x23, #4\n" - ".inst 0xa0810280 // smopa za0.s, p0/M, p0/M, z20.b, z1.b\n" - ".inst 0xa0850281 // smopa za1.s, p0/M, p0/M, z20.b, z5.b\n" - ".inst 0xa0890282 // smopa za2.s, p0/M, p0/M, z20.b, z9.b\n" - ".inst 0xa08d0283 // smopa za3.s, p0/M, p0/M, z20.b, z13.b\n" + ".inst 0xa09102c0 // smopa za0.s, p0/M, p0/M, z22.b, z17.b\n" + ".inst 0xa09502c1 // smopa za1.s, p0/M, p0/M, z22.b, z21.b\n" + ".inst 0xa09902c2 // smopa za2.s, p0/M, p0/M, z22.b, z25.b\n" + ".inst 0xa09d02c3 // smopa za3.s, p0/M, p0/M, z22.b, z29.b\n" "bgt 9b\n" "10:" // K oddments: End "tbz x13, #1, 14f\n" @@ -222,25 +219,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c57c // ld1w { z28.s-z31.s }, pn9.b/Z, [x11]\n" + ".inst 0xa040c570 // ld1w { z16.s-z19.s }, pn9.b/Z, [x11]\n" ".inst 0xc0860404 // mova { z4.s-z7.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 0xa041c560 // ld1w { z0.s-z3.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n" - ".inst 0xc0840401 // mova za1h.s[x12], { z0.s-z3.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 0xa042c570 // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n" - ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa043c56c // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n" - ".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n" + ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n" + ".inst 0xc086042c // mova { z12.s-z15.s }, za1h.s[x12]\n" + ".inst 0xa041c570 // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n" + ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" + ".inst 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n" + ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" + ".inst 0xa042c560 // ld1w { z0.s-z3.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n" + ".inst 0xc0840402 // mova za2h.s[x12], { z0.s-z3.s }\n" + ".inst 0xa043c574 // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n" + ".inst 0xc0840683 // mova za3h.s[x12], { z20.s-z23.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" ".inst 0xa060c544 // st1w { z4.s-z7.s }, pn9.b, [x10]\n" "addvl x11, x11, #16\n" - ".inst 0xa061c554 // st1w { z20.s-z23.s }, pn9.b, [x10, #0x4, MUL VL]\n" - ".inst 0xa062c558 // st1w { z24.s-z27.s }, pn9.b, [x10, #0x8, MUL VL]\n" - ".inst 0xa063c55c // st1w { z28.s-z31.s }, pn9.b, [x10, #0xc, MUL VL]\n" + ".inst 0xa061c54c // st1w { z12.s-z15.s }, pn9.b, [x10, #0x4, MUL VL]\n" + ".inst 0xa062c55c // st1w { z28.s-z31.s }, pn9.b, [x10, #0x8, MUL VL]\n" + ".inst 0xa063c550 // st1w { z16.s-z19.s }, pn9.b, [x10, #0xc, MUL VL]\n" "addvl x10, x10, #16\n" "blt 11b\n" "b 20f\n" @@ -248,16 +245,16 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in "mov x12, #0x0\n" "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop - ".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n" - ".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n" - ".inst 0xa060c554 // st1w { z20.s-z23.s }, pn9.b, [x10]\n" - ".inst 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - ".inst 0xa061c540 // st1w { z0.s-z3.s }, pn9.b, [x10, #0x4, MUL VL]\n" + ".inst 0xc086040c // mova { z12.s-z15.s }, za0h.s[x12]\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xa060c54c // st1w { z12.s-z15.s }, pn9.b, [x10]\n" + ".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n" + ".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n" + ".inst 0xa061c544 // st1w { z4.s-z7.s }, pn9.b, [x10, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa062c548 // st1w { z8.s-z11.s }, pn9.b, [x10, #0x8, MUL VL]\n" - ".inst 0xa063c54c // st1w { z12.s-z15.s }, pn9.b, [x10, #0xc, MUL VL]\n" + ".inst 0xa062c540 // st1w { z0.s-z3.s }, pn9.b, [x10, #0x8, MUL VL]\n" + ".inst 0xa063c558 // st1w { z24.s-z27.s }, pn9.b, [x10, #0xc, MUL VL]\n" "addvl x10, x10, #16\n" "blt 13b\n" "b 20f\n" @@ -293,32 +290,32 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in "16:" // Store to output array: 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 0xa160c2e0 // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x23]\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 0xa160c2f0 // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x23]\n" "add x23, x23, x22\n" "beq 17f\n" "subs x20, x20, #0x1\n" - ".inst 0xa160c2e1 // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x23]\n" + ".inst 0xa160c2f1 // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x23]\n" "add x23, x23, x22\n" "beq 17f\n" - ".inst 0xa160c2e2 // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x23]\n" + ".inst 0xa160c2f2 // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x23]\n" "17:" // Store to output array: Accumulator row 0 oddments: End "18:" // Store to output array: End "tbz x13, #0, 20f\n" "mov x12, #0x0\n" "cntw x20\n" "19:" // Store to output array: Refill accumulators: Loop - ".inst 0xa040c56c // ld1w { z12.s-z15.s }, pn9.b/Z, [x11]\n" - ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n" + ".inst 0xa040c568 // ld1w { z8.s-z11.s }, pn9.b/Z, [x11]\n" + ".inst 0xc0840500 // mova za0h.s[x12], { z8.s-z11.s }\n" ".inst 0xa041c570 // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n" ".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n" ".inst 0xa042c570 // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n" ".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n" - ".inst 0xa043c564 // ld1w { z4.s-z7.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n" - ".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n" + ".inst 0xa043c568 // ld1w { z8.s-z11.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n" + ".inst 0xc0840503 // mova za3h.s[x12], { z8.s-z11.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x11, x11, #16\n" @@ -342,4 +339,3 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |