diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp | 264 |
1 files changed, 130 insertions, 134 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp index d863b6c72a..a23c44b7da 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_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" @@ -32,10 +31,8 @@ namespace arm_gemm { -void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(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_4VLx1VL(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( @@ -94,14 +91,14 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "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 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 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 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 0xa040c1f8 // ld1w { z24.s-z27.s }, pn8.b/Z, [x15]\n" + ".inst 0xc0840700 // mova za0h.s[x12], { z24.s-z27.s }\n" + ".inst 0xa041c1fc // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840781 // mova za1h.s[x12], { z28.s-z31.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 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" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x15, x15, #16\n" @@ -119,11 +116,11 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(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" - "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 { z17.s }, p0/Z, [x20, x10, LSL #2]\n" + ".inst 0xc0902620 // addha za0.s, p1/M, p1/M, z17.s\n" + ".inst 0xc0902621 // addha za1.s, p1/M, p1/M, z17.s\n" + ".inst 0xc0902622 // addha za2.s, p1/M, p1/M, z17.s\n" + ".inst 0xc0902623 // addha za3.s, p1/M, p1/M, z17.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x10\n" "mov x21, x11\n" @@ -146,75 +143,75 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "madd x23, x10, x20, x23\n" // bptr = B + n * kstride_bytes "cbz x22, 8f\n" "subs x22, x22, #0x1\n" - ".inst 0xa0408370 // ld1b { z16.b-z19.b }, pn8.b/Z, [x27]\n" - "ldnt1b { z7.b }, p1/Z, [x23]\n" - ".inst 0xa041837c // ld1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" - "ldnt1b { z13.b }, p1/Z, [x23, #1, MUL VL]\n" - ".inst 0xa0428360 // ld1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" - "ldnt1b { z12.b }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0xa0438378 // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xa1408372 // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n" + "ldnt1b { z12.b }, p1/Z, [x23]\n" + ".inst 0xa1418370 // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "ldnt1b { z5.b }, p1/Z, [x23, #1, MUL VL]\n" + ".inst 0xa1428363 // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "ldnt1b { z4.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" "addvl x27, x27, #16\n" - "ldnt1b { z23.b }, p1/Z, [x23, #3, MUL VL]\n" + "ldnt1b { z19.b }, p1/Z, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "ble 7f\n" "6:" // K loop - ".inst 0xa0872600 // smopa za0.s, p1/M, p1/M, z16.b, z7.b\n" + ".inst 0xa08c2640 // smopa za0.s, p1/M, p1/M, z18.b, z12.b\n" "subs x22, x22, #0x1\n" - ".inst 0xa0872621 // smopa za1.s, p1/M, p1/M, z17.b, z7.b\n" - ".inst 0xa0872642 // smopa za2.s, p1/M, p1/M, z18.b, z7.b\n" - ".inst 0xa0872663 // smopa za3.s, p1/M, p1/M, z19.b, z7.b\n" - ".inst 0xa0408370 // ld1b { z16.b-z19.b }, pn8.b/Z, [x27]\n" - ".inst 0xa08d2780 // smopa za0.s, p1/M, p1/M, z28.b, z13.b\n" - "ldnt1b { z7.b }, p1/Z, [x23]\n" - ".inst 0xa08d27a1 // smopa za1.s, p1/M, p1/M, z29.b, z13.b\n" - ".inst 0xa08d27c2 // smopa za2.s, p1/M, p1/M, z30.b, z13.b\n" - ".inst 0xa08d27e3 // smopa za3.s, p1/M, p1/M, z31.b, z13.b\n" - ".inst 0xa041837c // ld1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0xa08c2400 // smopa za0.s, p1/M, p1/M, z0.b, z12.b\n" - "ldnt1b { z13.b }, p1/Z, [x23, #1, MUL VL]\n" - ".inst 0xa08c2421 // smopa za1.s, p1/M, p1/M, z1.b, z12.b\n" - ".inst 0xa08c2442 // smopa za2.s, p1/M, p1/M, z2.b, z12.b\n" - ".inst 0xa08c2463 // smopa za3.s, p1/M, p1/M, z3.b, z12.b\n" - ".inst 0xa0428360 // ld1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" - "ldnt1b { z12.b }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0xa0972700 // smopa za0.s, p1/M, p1/M, z24.b, z23.b\n" - ".inst 0xa0972721 // smopa za1.s, p1/M, p1/M, z25.b, z23.b\n" - ".inst 0xa0972742 // smopa za2.s, p1/M, p1/M, z26.b, z23.b\n" - ".inst 0xa0972763 // smopa za3.s, p1/M, p1/M, z27.b, z23.b\n" - ".inst 0xa0438378 // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xa08c26c1 // smopa za1.s, p1/M, p1/M, z22.b, z12.b\n" + ".inst 0xa08c2742 // smopa za2.s, p1/M, p1/M, z26.b, z12.b\n" + ".inst 0xa08c27c3 // smopa za3.s, p1/M, p1/M, z30.b, z12.b\n" + ".inst 0xa1408372 // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n" + ".inst 0xa0852600 // smopa za0.s, p1/M, p1/M, z16.b, z5.b\n" + "ldnt1b { z12.b }, p1/Z, [x23]\n" + ".inst 0xa0852681 // smopa za1.s, p1/M, p1/M, z20.b, z5.b\n" + ".inst 0xa0852702 // smopa za2.s, p1/M, p1/M, z24.b, z5.b\n" + ".inst 0xa0852783 // smopa za3.s, p1/M, p1/M, z28.b, z5.b\n" + ".inst 0xa1418370 // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa0842460 // smopa za0.s, p1/M, p1/M, z3.b, z4.b\n" + "ldnt1b { z5.b }, p1/Z, [x23, #1, MUL VL]\n" + ".inst 0xa08424e1 // smopa za1.s, p1/M, p1/M, z7.b, z4.b\n" + ".inst 0xa0842562 // smopa za2.s, p1/M, p1/M, z11.b, z4.b\n" + ".inst 0xa08425e3 // smopa za3.s, p1/M, p1/M, z15.b, z4.b\n" + ".inst 0xa1428363 // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "ldnt1b { z4.b }, p1/Z, [x23, #2, MUL VL]\n" + ".inst 0xa0932440 // smopa za0.s, p1/M, p1/M, z2.b, z19.b\n" + ".inst 0xa09324c1 // smopa za1.s, p1/M, p1/M, z6.b, z19.b\n" + ".inst 0xa0932542 // smopa za2.s, p1/M, p1/M, z10.b, z19.b\n" + ".inst 0xa09325c3 // smopa za3.s, p1/M, p1/M, z14.b, z19.b\n" + ".inst 0xa1438362 // ld1b { z2.b, z6.b, z10.b, z14.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" "addvl x27, x27, #16\n" - "ldnt1b { z23.b }, p1/Z, [x23, #3, MUL VL]\n" + "ldnt1b { z19.b }, p1/Z, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0xa0872600 // smopa za0.s, p1/M, p1/M, z16.b, z7.b\n" - ".inst 0xa0872621 // smopa za1.s, p1/M, p1/M, z17.b, z7.b\n" - ".inst 0xa0872642 // smopa za2.s, p1/M, p1/M, z18.b, z7.b\n" - ".inst 0xa0872663 // smopa za3.s, p1/M, p1/M, z19.b, z7.b\n" - ".inst 0xa08d2780 // smopa za0.s, p1/M, p1/M, z28.b, z13.b\n" - ".inst 0xa08d27a1 // smopa za1.s, p1/M, p1/M, z29.b, z13.b\n" - ".inst 0xa08d27c2 // smopa za2.s, p1/M, p1/M, z30.b, z13.b\n" - ".inst 0xa08d27e3 // smopa za3.s, p1/M, p1/M, z31.b, z13.b\n" - ".inst 0xa08c2400 // smopa za0.s, p1/M, p1/M, z0.b, z12.b\n" - ".inst 0xa08c2421 // smopa za1.s, p1/M, p1/M, z1.b, z12.b\n" - ".inst 0xa08c2442 // smopa za2.s, p1/M, p1/M, z2.b, z12.b\n" - ".inst 0xa08c2463 // smopa za3.s, p1/M, p1/M, z3.b, z12.b\n" - ".inst 0xa0972700 // smopa za0.s, p1/M, p1/M, z24.b, z23.b\n" - ".inst 0xa0972721 // smopa za1.s, p1/M, p1/M, z25.b, z23.b\n" - ".inst 0xa0972742 // smopa za2.s, p1/M, p1/M, z26.b, z23.b\n" - ".inst 0xa0972763 // smopa za3.s, p1/M, p1/M, z27.b, z23.b\n" + ".inst 0xa08c2640 // smopa za0.s, p1/M, p1/M, z18.b, z12.b\n" + ".inst 0xa08c26c1 // smopa za1.s, p1/M, p1/M, z22.b, z12.b\n" + ".inst 0xa08c2742 // smopa za2.s, p1/M, p1/M, z26.b, z12.b\n" + ".inst 0xa08c27c3 // smopa za3.s, p1/M, p1/M, z30.b, z12.b\n" + ".inst 0xa0852600 // smopa za0.s, p1/M, p1/M, z16.b, z5.b\n" + ".inst 0xa0852681 // smopa za1.s, p1/M, p1/M, z20.b, z5.b\n" + ".inst 0xa0852702 // smopa za2.s, p1/M, p1/M, z24.b, z5.b\n" + ".inst 0xa0852783 // smopa za3.s, p1/M, p1/M, z28.b, z5.b\n" + ".inst 0xa0842460 // smopa za0.s, p1/M, p1/M, z3.b, z4.b\n" + ".inst 0xa08424e1 // smopa za1.s, p1/M, p1/M, z7.b, z4.b\n" + ".inst 0xa0842562 // smopa za2.s, p1/M, p1/M, z11.b, z4.b\n" + ".inst 0xa08425e3 // smopa za3.s, p1/M, p1/M, z15.b, z4.b\n" + ".inst 0xa0932440 // smopa za0.s, p1/M, p1/M, z2.b, z19.b\n" + ".inst 0xa09324c1 // smopa za1.s, p1/M, p1/M, z6.b, z19.b\n" + ".inst 0xa0932542 // smopa za2.s, p1/M, p1/M, z10.b, z19.b\n" + ".inst 0xa09325c3 // smopa za3.s, p1/M, p1/M, z14.b, z19.b\n" "8:" // K oddments "cbz x21, 10f\n" "9:" // K oddments: Loop - ".inst 0xa0408370 // ld1b { z16.b-z19.b }, pn8.b/Z, [x27]\n" + ".inst 0xa0408368 // ld1b { z8.b-z11.b }, pn8.b/Z, [x27]\n" "subs x21, x21, #0x1\n" "addvl x27, x27, #4\n" - "ld1b { z7.b }, p1/Z, [x23]\n" + "ld1b { z15.b }, p1/Z, [x23]\n" "addvl x23, x23, #1\n" - ".inst 0xa0872600 // smopa za0.s, p1/M, p1/M, z16.b, z7.b\n" - ".inst 0xa0872621 // smopa za1.s, p1/M, p1/M, z17.b, z7.b\n" - ".inst 0xa0872642 // smopa za2.s, p1/M, p1/M, z18.b, z7.b\n" - ".inst 0xa0872663 // smopa za3.s, p1/M, p1/M, z19.b, z7.b\n" + ".inst 0xa08f2500 // smopa za0.s, p1/M, p1/M, z8.b, z15.b\n" + ".inst 0xa08f2521 // smopa za1.s, p1/M, p1/M, z9.b, z15.b\n" + ".inst 0xa08f2542 // smopa za2.s, p1/M, p1/M, z10.b, z15.b\n" + ".inst 0xa08f2563 // smopa za3.s, p1/M, p1/M, z11.b, z15.b\n" "bgt 9b\n" "10:" // K oddments: End "tbz x16, #1, 14f\n" @@ -222,25 +219,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(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 0xa040c1f4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n" - ".inst 0xc0860400 // mova { z0.s-z3.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 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 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n" - ".inst 0xc0860470 // mova { z16.s-z19.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 0xa040c1ec // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n" + ".inst 0xc0860404 // mova { z4.s-z7.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 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 0xc0860448 // mova { z8.s-z11.s }, za2h.s[x12]\n" + ".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\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 0xa043c1ec // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n" + ".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" - ".inst 0xa060c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14]\n" + ".inst 0xa060c1c4 // st1w { z4.s-z7.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 0xa062c1dc // st1w { z28.s-z31.s }, pn8.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c1d0 // st1w { z16.s-z19.s }, pn8.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa061c1dc // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c1c8 // st1w { z8.s-z11.s }, pn8.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c1d4 // st1w { z20.s-z23.s }, pn8.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 11b\n" "b 29f\n" @@ -248,12 +245,12 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "mov x12, #0x0\n" "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop - ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" - ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" - ".inst 0xa060c1c8 // st1w { z8.s-z11.s }, pn8.b, [x14]\n" + ".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n" + ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" + ".inst 0xa060c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14]\n" ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" ".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" - ".inst 0xa061c1c4 // st1w { z4.s-z7.s }, pn8.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa061c1c8 // st1w { z8.s-z11.s }, pn8.b, [x14, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" ".inst 0xa062c1cc // st1w { z12.s-z15.s }, pn8.b, [x14, #0x8, MUL VL]\n" @@ -275,30 +272,30 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "and x20, x22, #0x3\n" "cbz x21, 16f\n" "15:" // Store to output array: Accumulator row 0 loop - ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" - "st1w { z28.s }, p0, [x26]\n" + ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" + "st1w { z8.s }, p0, [x26]\n" "add x26, x26, x23\n" - "st1w { z29.s }, p0, [x26]\n" + "st1w { z9.s }, p0, [x26]\n" "add x26, x26, x23\n" "add x12, x12, #0x4\n" - "st1w { z30.s }, p0, [x26]\n" + "st1w { z10.s }, p0, [x26]\n" "add x26, x26, x23\n" "cmp x12, x21, LSL #2\n" - "st1w { z31.s }, p0, [x26]\n" + "st1w { z11.s }, p0, [x26]\n" "add x26, x26, x23\n" "blt 15b\n" "16:" // Store to output array: Accumulator row 0 oddments "cbz x20, 17f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" - "st1w { z8.s }, p0, [x26]\n" + ".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n" + "st1w { z4.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 17f\n" "subs x20, x20, #0x1\n" - "st1w { z9.s }, p0, [x26]\n" + "st1w { z5.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 17f\n" - "st1w { z10.s }, p0, [x26]\n" + "st1w { z6.s }, p0, [x26]\n" "add x26, x26, x23\n" "17:" // Store to output array: Accumulator row 0 oddments: End "subs x25, x25, x22\n" @@ -310,30 +307,30 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "and x20, x22, #0x3\n" "cbz x21, 19f\n" "18:" // Store to output array: Accumulator row 1 loop - ".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n" - "st1w { z0.s }, p0, [x26]\n" + ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" + "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x23\n" - "st1w { z1.s }, p0, [x26]\n" + "st1w { z17.s }, p0, [x26]\n" "add x26, x26, x23\n" "add x12, x12, #0x4\n" - "st1w { z2.s }, p0, [x26]\n" + "st1w { z18.s }, p0, [x26]\n" "add x26, x26, x23\n" "cmp x12, x21, LSL #2\n" - "st1w { z3.s }, p0, [x26]\n" + "st1w { z19.s }, p0, [x26]\n" "add x26, x26, x23\n" "blt 18b\n" "19:" // Store to output array: Accumulator row 1 oddments "cbz x20, 20f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n" + "st1w { z20.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 20f\n" "subs x20, x20, #0x1\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z21.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 20f\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z22.s }, p0, [x26]\n" "add x26, x26, x23\n" "20:" // Store to output array: Accumulator row 1 oddments: End "subs x25, x25, x22\n" @@ -345,30 +342,30 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "and x20, x22, #0x3\n" "cbz x21, 22f\n" "21:" // Store to output array: Accumulator row 2 loop - ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n" + "st1w { z24.s }, p0, [x26]\n" "add x26, x26, x23\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z25.s }, p0, [x26]\n" "add x26, x26, x23\n" "add x12, x12, #0x4\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z26.s }, p0, [x26]\n" "add x26, x26, x23\n" "cmp x12, x21, LSL #2\n" - "st1w { z19.s }, p0, [x26]\n" + "st1w { z27.s }, p0, [x26]\n" "add x26, x26, x23\n" "blt 21b\n" "22:" // Store to output array: Accumulator row 2 oddments "cbz x20, 23f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n" - "st1w { z0.s }, p0, [x26]\n" + ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" + "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 23f\n" "subs x20, x20, #0x1\n" - "st1w { z1.s }, p0, [x26]\n" + "st1w { z17.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 23f\n" - "st1w { z2.s }, p0, [x26]\n" + "st1w { z18.s }, p0, [x26]\n" "add x26, x26, x23\n" "23:" // Store to output array: Accumulator row 2 oddments: End "subs x25, x25, x22\n" @@ -380,44 +377,44 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in "and x20, x20, #0x3\n" "cbz x21, 25f\n" "24:" // Store to output array: Accumulator row 3 loop - ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" - "st1w { z12.s }, p0, [x26]\n" + ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" + "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x23\n" - "st1w { z13.s }, p0, [x26]\n" + "st1w { z17.s }, p0, [x26]\n" "add x26, x26, x23\n" "add x12, x12, #0x4\n" - "st1w { z14.s }, p0, [x26]\n" + "st1w { z18.s }, p0, [x26]\n" "add x26, x26, x23\n" "cmp x12, x21, LSL #2\n" - "st1w { z15.s }, p0, [x26]\n" + "st1w { z19.s }, p0, [x26]\n" "add x26, x26, x23\n" "blt 24b\n" "25:" // Store to output array: Accumulator row 3 oddments "cbz x20, 26f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" + "st1w { z12.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 26f\n" "subs x20, x20, #0x1\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z13.s }, p0, [x26]\n" "add x26, x26, x23\n" "beq 26f\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z14.s }, p0, [x26]\n" "26:" // Store to output array: Accumulator row 3 oddments: End "27:" // Store to output array: End "tbz x16, #0, 29f\n" "mov x12, #0x0\n" "cntw x20\n" "28:" // Store to output array: Refill accumulators: Loop - ".inst 0xa040c1f0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n" - ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n" + ".inst 0xa040c1ec // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n" + ".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.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 0xa042c1f8 // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840702 // mova za2h.s[x12], { z24.s-z27.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" + ".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" "add x12, x12, #0x4\n" "cmp x12, x20\n" "addvl x15, x15, #16\n" @@ -441,4 +438,3 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |