diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp | 300 |
1 files changed, 149 insertions, 151 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp index a51b3db4b0..70c94d32a3 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_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" @@ -113,14 +112,14 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "mov x12, #0x0\n" "cntw x20\n" "1:" // Initial accumulator load from buffer: 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 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 0xa042c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840402 // mova za2h.s[x12], { z0.s-z3.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 0xa040c1fc // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n" + ".inst 0xc0840780 // mova za0h.s[x12], { z28.s-z31.s }\n" + ".inst 0xa041c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840401 // mova za1h.s[x12], { z0.s-z3.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 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" @@ -138,12 +137,12 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "ldr x20, [%x[args], %[offsetof_bias]]\n" ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" "cbz x20, 5f\n" - "fmov z8.s, #1.0\n" - "ldnt1w { z27.s }, p0/Z, [x20, x10, LSL #2]\n" - ".inst 0x809b2500 // fmopa za0.s, p1/M, p1/M, z8.s, z27.s\n" - ".inst 0x809b2501 // fmopa za1.s, p1/M, p1/M, z8.s, z27.s\n" - ".inst 0x809b2502 // fmopa za2.s, p1/M, p1/M, z8.s, z27.s\n" - ".inst 0x809b2503 // fmopa za3.s, p1/M, p1/M, z8.s, z27.s\n" + "fmov z11.s, #1.0\n" + "ldnt1w { z13.s }, p0/Z, [x20, x10, LSL #2]\n" + ".inst 0x808d2560 // fmopa za0.s, p1/M, p1/M, z11.s, z13.s\n" + ".inst 0x808d2561 // fmopa za1.s, p1/M, p1/M, z11.s, z13.s\n" + ".inst 0x808d2562 // fmopa za2.s, p1/M, p1/M, z11.s, z13.s\n" + ".inst 0x808d2563 // fmopa za3.s, p1/M, p1/M, z11.s, z13.s\n" "4:" // Prepare accumulators: Test for last block "mov x20, x10\n" "mov x21, x11\n" @@ -166,75 +165,75 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "madd x23, x10, x20, x23\n" // bptr = B + n * kstride_bytes "cbz x22, 8f\n" "subs x22, x22, #0x1\n" - ".inst 0xa040a364 // ld1h { z4.h-z7.h }, pn8.b/Z, [x27]\n" - "ldnt1h { z29.h }, p1/Z, [x23]\n" - ".inst 0xa041a36c // ld1h { z12.h-z15.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n" - "ldnt1h { z23.h }, p1/Z, [x23, #1, MUL VL]\n" - ".inst 0xa042a360 // ld1h { z0.h-z3.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n" - "ldnt1h { z21.h }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0xa143a372 // ld1h { z18.h, z22.h, z26.h, z30.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xa140a360 // ld1h { z0.h, z4.h, z8.h, z12.h }, pn8.b/Z, [x27]\n" + "ldnt1h { z19.h }, p1/Z, [x23]\n" + ".inst 0xa141a371 // ld1h { z17.h, z21.h, z25.h, z29.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "ldnt1h { z22.h }, p1/Z, [x23, #1, MUL VL]\n" + ".inst 0xa142a370 // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "ldnt1h { z23.h }, p1/Z, [x23, #2, MUL VL]\n" + ".inst 0xa143a363 // ld1h { z3.h, z7.h, z11.h, z15.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n" "addvl x27, x27, #16\n" - "ldnt1h { z27.h }, p1/Z, [x23, #3, MUL VL]\n" + "ldnt1h { z2.h }, p1/Z, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "ble 7f\n" "6:" // K loop - ".inst 0x819d2480 // bfmopa za0.s, p1/M, p1/M, z4.h, z29.h\n" + ".inst 0x81932400 // bfmopa za0.s, p1/M, p1/M, z0.h, z19.h\n" "subs x22, x22, #0x1\n" - ".inst 0x819d24a1 // bfmopa za1.s, p1/M, p1/M, z5.h, z29.h\n" - ".inst 0x819d24c2 // bfmopa za2.s, p1/M, p1/M, z6.h, z29.h\n" - ".inst 0x819d24e3 // bfmopa za3.s, p1/M, p1/M, z7.h, z29.h\n" - ".inst 0xa040a364 // ld1h { z4.h-z7.h }, pn8.b/Z, [x27]\n" - ".inst 0x81972580 // bfmopa za0.s, p1/M, p1/M, z12.h, z23.h\n" - "ldnt1h { z29.h }, p1/Z, [x23]\n" - ".inst 0x819725a1 // bfmopa za1.s, p1/M, p1/M, z13.h, z23.h\n" - ".inst 0x819725c2 // bfmopa za2.s, p1/M, p1/M, z14.h, z23.h\n" - ".inst 0x819725e3 // bfmopa za3.s, p1/M, p1/M, z15.h, z23.h\n" - ".inst 0xa041a36c // ld1h { z12.h-z15.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n" - ".inst 0x81952400 // bfmopa za0.s, p1/M, p1/M, z0.h, z21.h\n" - "ldnt1h { z23.h }, p1/Z, [x23, #1, MUL VL]\n" - ".inst 0x81952421 // bfmopa za1.s, p1/M, p1/M, z1.h, z21.h\n" - ".inst 0x81952442 // bfmopa za2.s, p1/M, p1/M, z2.h, z21.h\n" - ".inst 0x81952463 // bfmopa za3.s, p1/M, p1/M, z3.h, z21.h\n" - ".inst 0xa042a360 // ld1h { z0.h-z3.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n" - "ldnt1h { z21.h }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0x819b2640 // bfmopa za0.s, p1/M, p1/M, z18.h, z27.h\n" - ".inst 0x819b26c1 // bfmopa za1.s, p1/M, p1/M, z22.h, z27.h\n" - ".inst 0x819b2742 // bfmopa za2.s, p1/M, p1/M, z26.h, z27.h\n" - ".inst 0x819b27c3 // bfmopa za3.s, p1/M, p1/M, z30.h, z27.h\n" - ".inst 0xa143a372 // ld1h { z18.h, z22.h, z26.h, z30.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0x81932481 // bfmopa za1.s, p1/M, p1/M, z4.h, z19.h\n" + ".inst 0x81932502 // bfmopa za2.s, p1/M, p1/M, z8.h, z19.h\n" + ".inst 0x81932583 // bfmopa za3.s, p1/M, p1/M, z12.h, z19.h\n" + ".inst 0xa140a360 // ld1h { z0.h, z4.h, z8.h, z12.h }, pn8.b/Z, [x27]\n" + ".inst 0x81962620 // bfmopa za0.s, p1/M, p1/M, z17.h, z22.h\n" + "ldnt1h { z19.h }, p1/Z, [x23]\n" + ".inst 0x819626a1 // bfmopa za1.s, p1/M, p1/M, z21.h, z22.h\n" + ".inst 0x81962722 // bfmopa za2.s, p1/M, p1/M, z25.h, z22.h\n" + ".inst 0x819627a3 // bfmopa za3.s, p1/M, p1/M, z29.h, z22.h\n" + ".inst 0xa141a371 // ld1h { z17.h, z21.h, z25.h, z29.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0x81972600 // bfmopa za0.s, p1/M, p1/M, z16.h, z23.h\n" + "ldnt1h { z22.h }, p1/Z, [x23, #1, MUL VL]\n" + ".inst 0x81972681 // bfmopa za1.s, p1/M, p1/M, z20.h, z23.h\n" + ".inst 0x81972702 // bfmopa za2.s, p1/M, p1/M, z24.h, z23.h\n" + ".inst 0x81972783 // bfmopa za3.s, p1/M, p1/M, z28.h, z23.h\n" + ".inst 0xa142a370 // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "ldnt1h { z23.h }, p1/Z, [x23, #2, MUL VL]\n" + ".inst 0x81822460 // bfmopa za0.s, p1/M, p1/M, z3.h, z2.h\n" + ".inst 0x818224e1 // bfmopa za1.s, p1/M, p1/M, z7.h, z2.h\n" + ".inst 0x81822562 // bfmopa za2.s, p1/M, p1/M, z11.h, z2.h\n" + ".inst 0x818225e3 // bfmopa za3.s, p1/M, p1/M, z15.h, z2.h\n" + ".inst 0xa143a363 // ld1h { z3.h, z7.h, z11.h, z15.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n" "addvl x27, x27, #16\n" - "ldnt1h { z27.h }, p1/Z, [x23, #3, MUL VL]\n" + "ldnt1h { z2.h }, p1/Z, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "bgt 6b\n" "7:" // K loop tail - ".inst 0x819d2480 // bfmopa za0.s, p1/M, p1/M, z4.h, z29.h\n" - ".inst 0x819d24a1 // bfmopa za1.s, p1/M, p1/M, z5.h, z29.h\n" - ".inst 0x819d24c2 // bfmopa za2.s, p1/M, p1/M, z6.h, z29.h\n" - ".inst 0x819d24e3 // bfmopa za3.s, p1/M, p1/M, z7.h, z29.h\n" - ".inst 0x81972580 // bfmopa za0.s, p1/M, p1/M, z12.h, z23.h\n" - ".inst 0x819725a1 // bfmopa za1.s, p1/M, p1/M, z13.h, z23.h\n" - ".inst 0x819725c2 // bfmopa za2.s, p1/M, p1/M, z14.h, z23.h\n" - ".inst 0x819725e3 // bfmopa za3.s, p1/M, p1/M, z15.h, z23.h\n" - ".inst 0x81952400 // bfmopa za0.s, p1/M, p1/M, z0.h, z21.h\n" - ".inst 0x81952421 // bfmopa za1.s, p1/M, p1/M, z1.h, z21.h\n" - ".inst 0x81952442 // bfmopa za2.s, p1/M, p1/M, z2.h, z21.h\n" - ".inst 0x81952463 // bfmopa za3.s, p1/M, p1/M, z3.h, z21.h\n" - ".inst 0x819b2640 // bfmopa za0.s, p1/M, p1/M, z18.h, z27.h\n" - ".inst 0x819b26c1 // bfmopa za1.s, p1/M, p1/M, z22.h, z27.h\n" - ".inst 0x819b2742 // bfmopa za2.s, p1/M, p1/M, z26.h, z27.h\n" - ".inst 0x819b27c3 // bfmopa za3.s, p1/M, p1/M, z30.h, z27.h\n" + ".inst 0x81932400 // bfmopa za0.s, p1/M, p1/M, z0.h, z19.h\n" + ".inst 0x81932481 // bfmopa za1.s, p1/M, p1/M, z4.h, z19.h\n" + ".inst 0x81932502 // bfmopa za2.s, p1/M, p1/M, z8.h, z19.h\n" + ".inst 0x81932583 // bfmopa za3.s, p1/M, p1/M, z12.h, z19.h\n" + ".inst 0x81962620 // bfmopa za0.s, p1/M, p1/M, z17.h, z22.h\n" + ".inst 0x819626a1 // bfmopa za1.s, p1/M, p1/M, z21.h, z22.h\n" + ".inst 0x81962722 // bfmopa za2.s, p1/M, p1/M, z25.h, z22.h\n" + ".inst 0x819627a3 // bfmopa za3.s, p1/M, p1/M, z29.h, z22.h\n" + ".inst 0x81972600 // bfmopa za0.s, p1/M, p1/M, z16.h, z23.h\n" + ".inst 0x81972681 // bfmopa za1.s, p1/M, p1/M, z20.h, z23.h\n" + ".inst 0x81972702 // bfmopa za2.s, p1/M, p1/M, z24.h, z23.h\n" + ".inst 0x81972783 // bfmopa za3.s, p1/M, p1/M, z28.h, z23.h\n" + ".inst 0x81822460 // bfmopa za0.s, p1/M, p1/M, z3.h, z2.h\n" + ".inst 0x818224e1 // bfmopa za1.s, p1/M, p1/M, z7.h, z2.h\n" + ".inst 0x81822562 // bfmopa za2.s, p1/M, p1/M, z11.h, z2.h\n" + ".inst 0x818225e3 // bfmopa za3.s, p1/M, p1/M, z15.h, z2.h\n" "8:" // K oddments "cbz x21, 10f\n" "9:" // K oddments: Loop - ".inst 0xa040a364 // ld1h { z4.h-z7.h }, pn8.b/Z, [x27]\n" + ".inst 0xa140a373 // ld1h { z19.h, z23.h, z27.h, z31.h }, pn8.b/Z, [x27]\n" "subs x21, x21, #0x1\n" "addvl x27, x27, #4\n" - "ld1h { z29.h }, p1/Z, [x23]\n" + "ld1h { z11.h }, p1/Z, [x23]\n" "addvl x23, x23, #1\n" - ".inst 0x819d2480 // bfmopa za0.s, p1/M, p1/M, z4.h, z29.h\n" - ".inst 0x819d24a1 // bfmopa za1.s, p1/M, p1/M, z5.h, z29.h\n" - ".inst 0x819d24c2 // bfmopa za2.s, p1/M, p1/M, z6.h, z29.h\n" - ".inst 0x819d24e3 // bfmopa za3.s, p1/M, p1/M, z7.h, z29.h\n" + ".inst 0x818b2660 // bfmopa za0.s, p1/M, p1/M, z19.h, z11.h\n" + ".inst 0x818b26e1 // bfmopa za1.s, p1/M, p1/M, z23.h, z11.h\n" + ".inst 0x818b2762 // bfmopa za2.s, p1/M, p1/M, z27.h, z11.h\n" + ".inst 0x818b27e3 // bfmopa za3.s, p1/M, p1/M, z31.h, z11.h\n" "bgt 9b\n" "10:" // K oddments: End "tbz x16, #1, 14f\n" @@ -242,25 +241,25 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "mov x12, #0x0\n" "cntw x20\n" "11:" // Store to partial result buffer: Store and refill: Loop - ".inst 0xa040c1e8 // ld1w { z8.s-z11.s }, pn8.b/Z, [x15]\n" + ".inst 0xa040c1f4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n" ".inst 0xc0860418 // mova { z24.s-z27.s }, za0h.s[x12]\n" - ".inst 0xc0840500 // mova za0h.s[x12], { z8.s-z11.s }\n" + ".inst 0xc0840680 // mova za0h.s[x12], { z20.s-z23.s }\n" ".inst 0xc0860424 // mova { z4.s-z7.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 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" - ".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n" - ".inst 0xa042c1e8 // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" - ".inst 0xc0840502 // mova za2h.s[x12], { z8.s-z11.s }\n" - ".inst 0xa043c1fc // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n" - ".inst 0xc0840783 // mova za3h.s[x12], { z28.s-z31.s }\n" + ".inst 0xa041c1e8 // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n" + ".inst 0xc0840501 // mova za1h.s[x12], { z8.s-z11.s }\n" + ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" + ".inst 0xc086047c // mova { z28.s-z31.s }, za3h.s[x12]\n" + ".inst 0xa042c1ec // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n" + ".inst 0xc0840582 // mova za2h.s[x12], { z12.s-z15.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 0xa060c1d8 // st1w { z24.s-z27.s }, pn8.b, [x14]\n" "addvl x15, x15, #16\n" ".inst 0xa061c1c4 // st1w { z4.s-z7.s }, pn8.b, [x14, #0x4, MUL VL]\n" - ".inst 0xa062c1cc // st1w { z12.s-z15.s }, pn8.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa062c1d0 // st1w { z16.s-z19.s }, pn8.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c1dc // st1w { z28.s-z31.s }, pn8.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 11b\n" "b 42f\n" @@ -269,15 +268,15 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "cntw x20\n" "13:" // Store to partial result buffer: Store only: Loop ".inst 0xc086040c // mova { z12.s-z15.s }, za0h.s[x12]\n" - ".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n" + ".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n" ".inst 0xa060c1cc // st1w { z12.s-z15.s }, pn8.b, [x14]\n" ".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n" - ".inst 0xc0860468 // mova { z8.s-z11.s }, za3h.s[x12]\n" - ".inst 0xa061c1d8 // st1w { z24.s-z27.s }, pn8.b, [x14, #0x4, MUL VL]\n" + ".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n" + ".inst 0xa061c1dc // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n" "add x12, x12, #0x4\n" "cmp x12, x20\n" ".inst 0xa062c1c0 // st1w { z0.s-z3.s }, pn8.b, [x14, #0x8, MUL VL]\n" - ".inst 0xa063c1c8 // st1w { z8.s-z11.s }, pn8.b, [x14, #0xc, MUL VL]\n" + ".inst 0xa063c1cc // st1w { z12.s-z15.s }, pn8.b, [x14, #0xc, MUL VL]\n" "addvl x14, x14, #16\n" "blt 13b\n" "b 42f\n" @@ -296,16 +295,16 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "and x20, x22, #0x3\n" "cbz x21, 16f\n" "15:" // Store to output array: Skip activation: Accumulator row 0 loop - ".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n" + "st1w { z4.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z5.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z6.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z19.s }, p0, [x26]\n" + "st1w { z7.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 15b\n" "16:" // Store to output array: Skip activation: Accumulator row 0 oddments @@ -331,30 +330,30 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "and x20, x22, #0x3\n" "cbz x21, 19f\n" "18:" // Store to output array: Skip activation: Accumulator row 1 loop - ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc0860428 // mova { z8.s-z11.s }, za1h.s[x12]\n" + "st1w { z8.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z9.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z10.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z19.s }, p0, [x26]\n" + "st1w { z11.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 18b\n" "19:" // Store to output array: Skip activation: Accumulator row 1 oddments "cbz x20, 20f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" - "st1w { z4.s }, p0, [x26]\n" + ".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n" + "st1w { z24.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 20f\n" "subs x20, x20, #0x1\n" - "st1w { z5.s }, p0, [x26]\n" + "st1w { z25.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 20f\n" - "st1w { z6.s }, p0, [x26]\n" + "st1w { z26.s }, p0, [x26]\n" "add x26, x26, x24\n" "20:" // Store to output array: Skip activation: Accumulator row 1 oddments: End "subs x25, x25, x22\n" @@ -366,30 +365,30 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "and x20, x22, #0x3\n" "cbz x21, 22f\n" "21:" // Store to output array: Skip activation: Accumulator row 2 loop - ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc0860444 // mova { z4.s-z7.s }, za2h.s[x12]\n" + "st1w { z4.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z5.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z6.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z19.s }, p0, [x26]\n" + "st1w { z7.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 21b\n" "22:" // Store to output array: Skip activation: Accumulator row 2 oddments "cbz x20, 23f\n" "subs x20, x20, #0x1\n" - ".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n" - "st1w { z20.s }, p0, [x26]\n" + ".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n" + "st1w { z12.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 23f\n" "subs x20, x20, #0x1\n" - "st1w { z21.s }, p0, [x26]\n" + "st1w { z13.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 23f\n" - "st1w { z22.s }, p0, [x26]\n" + "st1w { z14.s }, p0, [x26]\n" "add x26, x26, x24\n" "23:" // Store to output array: Skip activation: Accumulator row 2 oddments: End "subs x25, x25, x22\n" @@ -401,30 +400,30 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "and x20, x22, #0x3\n" "cbz x21, 25f\n" "24:" // Store to output array: Skip activation: Accumulator row 3 loop - ".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n" - "st1w { z4.s }, p0, [x26]\n" + ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" + "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z5.s }, p0, [x26]\n" + "st1w { z17.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z6.s }, p0, [x26]\n" + "st1w { z18.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z7.s }, p0, [x26]\n" + "st1w { z19.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 24b\n" "25:" // Store to output array: Skip activation: Accumulator row 3 oddments "cbz x20, 26f\n" "subs x20, x20, #0x1\n" - ".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, x24\n" "beq 26f\n" "subs x20, x20, #0x1\n" - "st1w { z13.s }, p0, [x26]\n" + "st1w { z17.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 26f\n" - "st1w { z14.s }, p0, [x26]\n" + "st1w { z18.s }, p0, [x26]\n" "add x26, x26, x24\n" "26:" // Store to output array: Skip activation: Accumulator row 3 oddments: End "subs x25, x25, x22\n" @@ -433,40 +432,40 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "27:" // Store to output array: Skip activation: End "cntw x23\n" "cmp x25, x23\n" - "ld1rw { z25.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n" + "ld1rw { z21.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n" "csel x22, x25, x23, LT\n" "lsr x21, x22, #0x2\n" - "ld1rw { z24.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n" + "ld1rw { z20.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n" "mov x12, #0x0\n" "and x20, x22, #0x3\n" "cbz x21, 29f\n" "28:" // Store to output array: Accumulator row 0 loop - ".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n" - ".inst 0xc1b8cb34 // fclamp { z20.s-z23.s }, z25.s, z24.s\n" - "st1w { z20.s }, p0, [x26]\n" + ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" + ".inst 0xc1b4cabc // fclamp { z28.s-z31.s }, z21.s, z20.s\n" + "st1w { z28.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z21.s }, p0, [x26]\n" + "st1w { z29.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z22.s }, p0, [x26]\n" + "st1w { z30.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z23.s }, p0, [x26]\n" + "st1w { z31.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 28b\n" "29:" // Store to output array: Accumulator row 0 oddments "cbz x20, 30f\n" - ".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n" + ".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n" "subs x20, x20, #0x1\n" - ".inst 0xc1b8cb28 // fclamp { z8.s-z11.s }, z25.s, z24.s\n" - "st1w { z8.s }, p0, [x26]\n" + ".inst 0xc1b4cabc // fclamp { z28.s-z31.s }, z21.s, z20.s\n" + "st1w { z28.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 30f\n" "subs x20, x20, #0x1\n" - "st1w { z9.s }, p0, [x26]\n" + "st1w { z29.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 30f\n" - "st1w { z10.s }, p0, [x26]\n" + "st1w { z30.s }, p0, [x26]\n" "add x26, x26, x24\n" "30:" // Store to output array: Accumulator row 0 oddments: End "subs x25, x25, x22\n" @@ -478,24 +477,24 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "and x20, x22, #0x3\n" "cbz x21, 32f\n" "31:" // Store to output array: Accumulator row 1 loop - ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" - ".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n" - "st1w { z16.s }, p0, [x26]\n" + ".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n" + ".inst 0xc1b4caa4 // fclamp { z4.s-z7.s }, z21.s, z20.s\n" + "st1w { z4.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z17.s }, p0, [x26]\n" + "st1w { z5.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z18.s }, p0, [x26]\n" + "st1w { z6.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z19.s }, p0, [x26]\n" + "st1w { z7.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 31b\n" "32:" // Store to output array: Accumulator row 1 oddments "cbz x20, 33f\n" ".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n" "subs x20, x20, #0x1\n" - ".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n" + ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 33f\n" @@ -516,7 +515,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "cbz x21, 35f\n" "34:" // Store to output array: Accumulator row 2 loop ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" - ".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n" + ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" "st1w { z17.s }, p0, [x26]\n" @@ -532,7 +531,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "cbz x20, 36f\n" ".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n" "subs x20, x20, #0x1\n" - ".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n" + ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 36f\n" @@ -552,24 +551,24 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con "and x20, x20, #0x3\n" "cbz x21, 38f\n" "37:" // Store to output array: Accumulator row 3 loop - ".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\n" - ".inst 0xc1b8cb34 // fclamp { z20.s-z23.s }, z25.s, z24.s\n" - "st1w { z20.s }, p0, [x26]\n" + ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" + ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" + "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" - "st1w { z21.s }, p0, [x26]\n" + "st1w { z17.s }, p0, [x26]\n" "add x26, x26, x24\n" "add x12, x12, #0x4\n" - "st1w { z22.s }, p0, [x26]\n" + "st1w { z18.s }, p0, [x26]\n" "add x26, x26, x24\n" "cmp x12, x21, LSL #2\n" - "st1w { z23.s }, p0, [x26]\n" + "st1w { z19.s }, p0, [x26]\n" "add x26, x26, x24\n" "blt 37b\n" "38:" // Store to output array: Accumulator row 3 oddments "cbz x20, 39f\n" ".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n" "subs x20, x20, #0x1\n" - ".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n" + ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" "st1w { z16.s }, p0, [x26]\n" "add x26, x26, x24\n" "beq 39f\n" @@ -588,10 +587,10 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con ".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.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 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 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 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 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" "addvl x15, x15, #16\n" @@ -615,4 +614,3 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con } // namespace arm_gemm #endif // ARM_COMPUTE_ENABLE_SME2 -#endif // __ARM_FEATURE_SVE |