diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp | 610 |
1 files changed, 305 insertions, 305 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp index 7ed04da9df..9747587495 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp @@ -64,22 +64,22 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( __asm__ __volatile__( "ptrue p8.b\n" ".inst 0xd503477f // SMSTART ZA\n" - "mov x10, #0x0\n" - "cntw x9, ALL, MUL #4\n" - "mov x28, #0x4\n" - "add x27, %x[N], x9\n" + "cntw x10, ALL, MUL #4\n" + "add x28, %x[N], x10\n" + "sub x28, x28, #0x1\n" + "udiv x28, x28, x10\n" + "add x22, x28, #0x3\n" + "and x22, x22, #0xfffffffffffffffc\n" + "mul x22, x22, x10\n" + "mul x22, x22, %x[K]\n" + "mov x9, #0x0\n" + "mov x27, #0x4\n" "mov x26, %x[B_ptr]\n" - "sub x27, x27, #0x1\n" "mov x25, %x[output_ptr]\n" - "udiv x27, x27, x9\n" "ptrue p2.b\n" - "add x22, x27, #0x3\n" ".inst 0x25207811 // ptrue pn9.b\n" - "and x22, x22, #0xfffffffffffffffc\n" - "mov x21, #0x1\n" - "mul x22, x22, x9\n" - "mul x22, x22, %x[K]\n" "lsl x22, x22, #0x1\n" + "mov x21, #0x1\n" "1:" // RHS size check loop "cmp x22, #0x200000\n" "blt 2f\n" @@ -97,9 +97,9 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "3:" // RHS prefetch exit "mov x24, %x[bias]\n" "4:" // Column loop - "cmp x27, #0x4\n" + "cmp x28, #0x4\n" "bge 28f\n" - "cmp x27, #0x2\n" + "cmp x28, #0x2\n" "bgt 20f\n" "beq 12f\n" "mov x23, %x[A_ptr]\n" @@ -110,7 +110,7 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 5f\n" ".inst 0xa040c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24]\n" - ".inst 0xc0044e00 // mova za.d[x10, #0], { z16.d-z19.d }\n" + ".inst 0xc0042e00 // mova za.d[x9, #0], { z16.d-z19.d }\n" "b 6f\n" "5:" // Width 1: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -119,72 +119,72 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "ble 8f\n" "7:" // Width 1: Multiply loop: Main loop head "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z8.s }, p1/Z, [x23]\n" - "addvl x26, x26, #16\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z10.s }, p1/Z, [x23]\n" + ".inst 0x658aa94a // bfcvt z10.h, p2/M, z10.s\n" + "ld1rqw { z16.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" + "uzp1 z10.h, z10.h, z10.h\n" "sub x22, x22, #0x8\n" - "ld1rqw { z11.s }, p0/Z, [x23, #16]\n" - "cmp x22, #0x8\n" - "add x23, x23, #0x20\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #16\n" - ".inst 0x658aa908 // bfcvt z8.h, p2/M, z8.s\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" + "uzp1 z16.h, z16.h, z16.h\n" + "trn1 z10.d, z10.d, z16.d\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" "addvl x26, x26, #16\n" - ".inst 0x658aa96b // bfcvt z11.h, p2/M, z11.s\n" + ".inst 0xc15ab198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[0]\n" ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" "addvl x26, x26, #16\n" - "uzp1 z8.h, z8.h, z8.h\n" - "uzp1 z11.h, z11.h, z11.h\n" - "trn1 z8.d, z8.d, z11.d\n" - ".inst 0xc158d098 // bfdot za.s[x10, 0], { z4.h-z7.h }, z8.h[0]\n" - ".inst 0xc158d698 // bfdot za.s[x10, 0], { z20.h-z23.h }, z8.h[1]\n" - ".inst 0xc158da18 // bfdot za.s[x10, 0], { z16.h-z19.h }, z8.h[2]\n" - ".inst 0xc158dd98 // bfdot za.s[x10, 0], { z12.h-z15.h }, z8.h[3]\n" + "cmp x22, #0x8\n" + ".inst 0xc15ab598 // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[1]\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + "addvl x26, x26, #16\n" + "add x23, x23, #0x20\n" + ".inst 0xc15ab818 // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[2]\n" + ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + "addvl x26, x26, #16\n" + ".inst 0xc15abf98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z10.h[3]\n" "bgt 7b\n" "8:" // Width 1: Multiply loop: Single iteration only "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z3.s }, p1/Z, [x23]\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z15.s }, p1/Z, [x23]\n" + ".inst 0x658aa9ef // bfcvt z15.h, p2/M, z15.s\n" + "ld1rqw { z17.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aaa31 // bfcvt z17.h, p2/M, z17.s\n" + "uzp1 z15.h, z15.h, z15.h\n" "subs x22, x22, #0x2\n" - "addvl x26, x26, #16\n" - "ld1rqw { z24.s }, p0/Z, [x23, #16]\n" + "uzp1 z17.h, z17.h, z17.h\n" + "trn1 z15.d, z15.d, z17.d\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" "add x23, x23, #0x20\n" - ".inst 0x658aa863 // bfcvt z3.h, p2/M, z3.s\n" - ".inst 0x658aab18 // bfcvt z24.h, p2/M, z24.s\n" - "uzp1 z3.h, z3.h, z3.h\n" - "uzp1 z24.h, z24.h, z24.h\n" - "trn1 z3.d, z3.d, z24.d\n" - ".inst 0xc153d198 // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[0]\n" + ".inst 0xc15fb218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[0]\n" + "addvl x26, x26, #16\n" "ble 9f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" + ".inst 0xc15fb418 // bfdot za.s[x9, 0], { z0.h-z3.h }, z15.h[1]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d598 // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[1]\n" "ble 9f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" + ".inst 0xc15fb898 // bfdot za.s[x9, 0], { z4.h-z7.h }, z15.h[2]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d998 // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[2]\n" "ble 9f\n" - ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n" + ".inst 0xc15fbd18 // bfdot za.s[x9, 0], { z8.h-z11.h }, z15.h[3]\n" "addvl x26, x26, #16\n" - ".inst 0xc153df98 // bfdot za.s[x10, 0], { z28.h-z31.h }, z3.h[3]\n" "9:" // Width 1: Multiply loop: multiply skip "tbz %x[flags], #1, 10f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0064c1c // mova { z28.d-z31.d }, za.d[x10, #0]\n" - "ld1rw { z4.s }, p2/Z, [x21]\n" - "ld1rw { z18.s }, p2/Z, [x20]\n" - ".inst 0xc1b2c89c // fclamp { z28.s-z31.s }, z4.s, z18.s\n" - ".inst 0xa060c33c // st1w { z28.s-z31.s }, p8, [x25]\n" + ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" + "ld1rw { z8.s }, p2/Z, [x21]\n" + "ld1rw { z26.s }, p2/Z, [x20]\n" + ".inst 0xc1bac900 // fclamp { z0.s-z3.s }, z8.s, z26.s\n" + ".inst 0xa060c320 // st1w { z0.s-z3.s }, p8, [x25]\n" "addvl x25, x25, #4\n" "b 11f\n" "10:" // Width 1: No activation - ".inst 0xc0064c04 // mova { z4.d-z7.d }, za.d[x10, #0]\n" + ".inst 0xc0062c04 // mova { z4.d-z7.d }, za.d[x9, #0]\n" ".inst 0xa060c324 // st1w { z4.s-z7.s }, p8, [x25]\n" "addvl x25, x25, #4\n" "11:" // Width 1: Output done @@ -192,15 +192,15 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "12:" // Width 2 "mov x23, %x[A_ptr]\n" "lsl x21, %x[K], #0x2\n" - "sub x20, %x[N], x9\n" + "sub x20, %x[N], x10\n" "mov x22, %x[K]\n" ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 13f\n" - ".inst 0xa040c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24]\n" - ".inst 0xa041c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0044e00 // mova za.d[x10, #0], { z16.d-z19.d }\n" - ".inst 0xc0044c01 // mova za.d[x10, #1], { z0.d-z3.d }\n" + ".inst 0xa040c718 // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n" + ".inst 0xc0042f00 // mova za.d[x9, #0], { z24.d-z27.d }\n" + ".inst 0xa041c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042e01 // mova za.d[x9, #1], { z16.d-z19.d }\n" "b 14f\n" "13:" // Width 2: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -209,94 +209,94 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "ble 16f\n" "15:" // Width 2: Multiply loop: Main loop head "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z8.s }, p1/Z, [x23]\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z13.s }, p1/Z, [x23]\n" + ".inst 0x658aa9ad // bfcvt z13.h, p2/M, z13.s\n" + "ld1rqw { z27.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aab7b // bfcvt z27.h, p2/M, z27.s\n" + "uzp1 z13.h, z13.h, z13.h\n" "sub x22, x22, #0x8\n" - "ld1rqw { z9.s }, p0/Z, [x23, #16]\n" + "uzp1 z27.h, z27.h, z27.h\n" + "trn1 z13.d, z13.d, z27.d\n" + ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" "cmp x22, #0x8\n" + ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15db298 // bfdot za.s[x9, 0], { z20.h-z23.h }, z13.h[0]\n" + "addvl x26, x26, #16\n" "add x23, x23, #0x20\n" + ".inst 0xc15db019 // bfdot za.s[x9, 1], { z0.h-z3.h }, z13.h[0]\n" + ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15db698 // bfdot za.s[x9, 0], { z20.h-z23.h }, z13.h[1]\n" "addvl x26, x26, #16\n" - ".inst 0x658aa908 // bfcvt z8.h, p2/M, z8.s\n" - ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" - ".inst 0x658aa929 // bfcvt z9.h, p2/M, z9.s\n" + ".inst 0xc15db719 // bfdot za.s[x9, 1], { z24.h-z27.h }, z13.h[1]\n" + ".inst 0xa040a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n" ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15db918 // bfdot za.s[x9, 0], { z8.h-z11.h }, z13.h[2]\n" "addvl x26, x26, #16\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" - "uzp1 z8.h, z8.h, z8.h\n" - ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" - "uzp1 z9.h, z9.h, z9.h\n" - "trn1 z8.d, z8.d, z9.d\n" - ".inst 0xc158d298 // bfdot za.s[x10, 0], { z20.h-z23.h }, z8.h[0]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15dba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z13.h[2]\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15dbc18 // bfdot za.s[x9, 0], { z0.h-z3.h }, z13.h[3]\n" "addvl x26, x26, #16\n" - ".inst 0xc158d319 // bfdot za.s[x10, 1], { z24.h-z27.h }, z8.h[0]\n" - ".inst 0xc158d418 // bfdot za.s[x10, 0], { z0.h-z3.h }, z8.h[1]\n" - ".inst 0xc158d619 // bfdot za.s[x10, 1], { z16.h-z19.h }, z8.h[1]\n" - ".inst 0xc158d898 // bfdot za.s[x10, 0], { z4.h-z7.h }, z8.h[2]\n" - ".inst 0xc158d999 // bfdot za.s[x10, 1], { z12.h-z15.h }, z8.h[2]\n" - ".inst 0xc158df98 // bfdot za.s[x10, 0], { z28.h-z31.h }, z8.h[3]\n" - ".inst 0xc158de99 // bfdot za.s[x10, 1], { z20.h-z23.h }, z8.h[3]\n" + ".inst 0xc15dbc99 // bfdot za.s[x9, 1], { z4.h-z7.h }, z13.h[3]\n" "bgt 15b\n" "16:" // Width 2: Multiply loop: Single iteration only "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z3.s }, p1/Z, [x23]\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z15.s }, p1/Z, [x23]\n" + ".inst 0x658aa9ef // bfcvt z15.h, p2/M, z15.s\n" + "ld1rqw { z5.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aa8a5 // bfcvt z5.h, p2/M, z5.s\n" + "uzp1 z15.h, z15.h, z15.h\n" "subs x22, x22, #0x2\n" - "ld1rqw { z16.s }, p0/Z, [x23, #16]\n" + "uzp1 z5.h, z5.h, z5.h\n" + "trn1 z15.d, z15.d, z5.d\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" "add x23, x23, #0x20\n" - ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[0]\n" "addvl x26, x26, #16\n" - ".inst 0x658aa863 // bfcvt z3.h, p2/M, z3.s\n" - ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" - "uzp1 z3.h, z3.h, z3.h\n" - "uzp1 z16.h, z16.h, z16.h\n" - "trn1 z3.d, z3.d, z16.d\n" - ".inst 0xc153d318 // bfdot za.s[x10, 0], { z24.h-z27.h }, z3.h[0]\n" - ".inst 0xc153d099 // bfdot za.s[x10, 1], { z4.h-z7.h }, z3.h[0]\n" + ".inst 0xc15fb319 // bfdot za.s[x9, 1], { z24.h-z27.h }, z15.h[0]\n" "ble 17f\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" - ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb798 // bfdot za.s[x9, 0], { z28.h-z31.h }, z15.h[1]\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb499 // bfdot za.s[x9, 1], { z4.h-z7.h }, z15.h[1]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d498 // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[1]\n" - ".inst 0xc153d719 // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[1]\n" "ble 17f\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fba18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[2]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[2]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d898 // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[2]\n" - ".inst 0xc153d999 // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[2]\n" "ble 17f\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + ".inst 0xc15fbf98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z15.h[3]\n" + ".inst 0xa041a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fbd19 // bfdot za.s[x9, 1], { z8.h-z11.h }, z15.h[3]\n" "addvl x26, x26, #16\n" - ".inst 0xc153dc98 // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[3]\n" - ".inst 0xc153de99 // bfdot za.s[x10, 1], { z20.h-z23.h }, z3.h[3]\n" "17:" // Width 2: Multiply loop: multiply skip "tbz %x[flags], #1, 18f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0064c08 // mova { z8.d-z11.d }, za.d[x10, #0]\n" - ".inst 0xc0064c2c // mova { z12.d-z15.d }, za.d[x10, #1]\n" - "ld1rw { z3.s }, p2/Z, [x21]\n" + ".inst 0xc0062c14 // mova { z20.d-z23.d }, za.d[x9, #0]\n" + "ld1rw { z11.s }, p2/Z, [x21]\n" + ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" "ld1rw { z28.s }, p2/Z, [x20]\n" - ".inst 0xc1bcc868 // fclamp { z8.s-z11.s }, z3.s, z28.s\n" - ".inst 0xc1bcc86c // fclamp { z12.s-z15.s }, z3.s, z28.s\n" - ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" + ".inst 0xc1bcc974 // fclamp { z20.s-z23.s }, z11.s, z28.s\n" + ".inst 0xa060c734 // st1w { z20.s-z23.s }, pn9.b, [x25]\n" + ".inst 0xc1bcc96c // fclamp { z12.s-z15.s }, z11.s, z28.s\n" ".inst 0xa061c32c // st1w { z12.s-z15.s }, p8, [x25, #0x4, MUL VL]\n" "addvl x25, x25, #8\n" "b 19f\n" "18:" // Width 2: No activation - ".inst 0xc0064c04 // mova { z4.d-z7.d }, za.d[x10, #0]\n" - ".inst 0xc0064c38 // mova { z24.d-z27.d }, za.d[x10, #1]\n" - ".inst 0xa060c724 // st1w { z4.s-z7.s }, pn9.b, [x25]\n" - ".inst 0xa061c338 // st1w { z24.s-z27.s }, p8, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" + ".inst 0xa060c720 // st1w { z0.s-z3.s }, pn9.b, [x25]\n" + ".inst 0xc0062c20 // mova { z0.d-z3.d }, za.d[x9, #1]\n" + ".inst 0xa061c320 // st1w { z0.s-z3.s }, p8, [x25, #0x4, MUL VL]\n" "addvl x25, x25, #8\n" "19:" // Width 2: Output done "b 36f\n" @@ -304,17 +304,17 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "mov x20, #0x2\n" "mov x23, %x[A_ptr]\n" "lsl x21, %x[K], #0x2\n" - "msub x20, x9, x20, %x[N]\n" + "msub x20, x10, x20, %x[N]\n" "mov x22, %x[K]\n" ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 21f\n" - ".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n" - ".inst 0xa041c718 // ld1w { z24.s-z27.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xa042c71c // ld1w { z28.s-z31.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0044c00 // mova za.d[x10, #0], { z0.d-z3.d }\n" - ".inst 0xc0044f01 // mova za.d[x10, #1], { z24.d-z27.d }\n" - ".inst 0xc0044f82 // mova za.d[x10, #2], { z28.d-z31.d }\n" + ".inst 0xa040c71c // ld1w { z28.s-z31.s }, pn9.b/Z, [x24]\n" + ".inst 0xc0042f80 // mova za.d[x9, #0], { z28.d-z31.d }\n" + ".inst 0xa041c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042c81 // mova za.d[x9, #1], { z4.d-z7.d }\n" + ".inst 0xa042c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" + ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n" "b 22f\n" "21:" // Width 3: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -323,114 +323,114 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "ble 24f\n" "23:" // Width 3: Multiply loop: Main loop head "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z3.s }, p1/Z, [x23]\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z14.s }, p1/Z, [x23]\n" + ".inst 0x658aa9ce // bfcvt z14.h, p2/M, z14.s\n" + "ld1rqw { z16.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" + "uzp1 z14.h, z14.h, z14.h\n" "sub x22, x22, #0x8\n" - "ld1rqw { z0.s }, p0/Z, [x23, #16]\n" + "uzp1 z16.h, z16.h, z16.h\n" + "trn1 z14.d, z14.d, z16.d\n" + ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" "cmp x22, #0x8\n" - "add x23, x23, #0x20\n" ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0x658aa863 // bfcvt z3.h, p2/M, z3.s\n" - ".inst 0xa042a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15eb098 // bfdot za.s[x9, 0], { z4.h-z7.h }, z14.h[0]\n" + "add x23, x23, #0x20\n" + ".inst 0xa042a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15eb319 // bfdot za.s[x9, 1], { z24.h-z27.h }, z14.h[0]\n" "addvl x26, x26, #16\n" - ".inst 0x658aa800 // bfcvt z0.h, p2/M, z0.s\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" + ".inst 0xc15eb01a // bfdot za.s[x9, 2], { z0.h-z3.h }, z14.h[0]\n" + ".inst 0xa040a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15eb518 // bfdot za.s[x9, 0], { z8.h-z11.h }, z14.h[1]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15eb499 // bfdot za.s[x9, 1], { z4.h-z7.h }, z14.h[1]\n" + "addvl x26, x26, #16\n" + ".inst 0xc15eb61a // bfdot za.s[x9, 2], { z16.h-z19.h }, z14.h[1]\n" + ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15eb818 // bfdot za.s[x9, 0], { z0.h-z3.h }, z14.h[2]\n" + ".inst 0xa042a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15ebb99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z14.h[2]\n" "addvl x26, x26, #16\n" - "uzp1 z3.h, z3.h, z3.h\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" - "uzp1 z0.h, z0.h, z0.h\n" - "trn1 z3.d, z3.d, z0.d\n" - ".inst 0xc153d198 // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[0]\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153d319 // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[0]\n" - ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15eb81a // bfdot za.s[x9, 2], { z0.h-z3.h }, z14.h[2]\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15ebf18 // bfdot za.s[x9, 0], { z24.h-z27.h }, z14.h[3]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15ebf99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z14.h[3]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d09a // bfdot za.s[x10, 2], { z4.h-z7.h }, z3.h[0]\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - ".inst 0xc153d618 // bfdot za.s[x10, 0], { z16.h-z19.h }, z3.h[1]\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc153d799 // bfdot za.s[x10, 1], { z28.h-z31.h }, z3.h[1]\n" - ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" - ".inst 0xc153d51a // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[1]\n" - ".inst 0xc153da98 // bfdot za.s[x10, 0], { z20.h-z23.h }, z3.h[2]\n" - ".inst 0xc153d999 // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[2]\n" - ".inst 0xc153db1a // bfdot za.s[x10, 2], { z24.h-z27.h }, z3.h[2]\n" - ".inst 0xc153dc98 // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[3]\n" - ".inst 0xc153de19 // bfdot za.s[x10, 1], { z16.h-z19.h }, z3.h[3]\n" - ".inst 0xc153df9a // bfdot za.s[x10, 2], { z28.h-z31.h }, z3.h[3]\n" + ".inst 0xc15ebe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z14.h[3]\n" "bgt 23b\n" "24:" // Width 3: Multiply loop: Single iteration only "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z3.s }, p1/Z, [x23]\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z15.s }, p1/Z, [x23]\n" + ".inst 0x658aa9ef // bfcvt z15.h, p2/M, z15.s\n" + "ld1rqw { z31.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aabff // bfcvt z31.h, p2/M, z31.s\n" + "uzp1 z15.h, z15.h, z15.h\n" "subs x22, x22, #0x2\n" - "ld1rqw { z20.s }, p0/Z, [x23, #16]\n" + "uzp1 z31.h, z31.h, z31.h\n" + "trn1 z15.d, z15.d, z31.d\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" "add x23, x23, #0x20\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0x658aa863 // bfcvt z3.h, p2/M, z3.s\n" - ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[0]\n" + ".inst 0xa042a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fb019 // bfdot za.s[x9, 1], { z0.h-z3.h }, z15.h[0]\n" "addvl x26, x26, #16\n" - ".inst 0x658aaa94 // bfcvt z20.h, p2/M, z20.s\n" - "uzp1 z3.h, z3.h, z3.h\n" - "uzp1 z20.h, z20.h, z20.h\n" - "trn1 z3.d, z3.d, z20.d\n" - ".inst 0xc153d098 // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[0]\n" - ".inst 0xc153d199 // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[0]\n" - ".inst 0xc153d31a // bfdot za.s[x10, 2], { z24.h-z27.h }, z3.h[0]\n" + ".inst 0xc15fb09a // bfdot za.s[x9, 2], { z4.h-z7.h }, z15.h[0]\n" "ble 25f\n" ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fb698 // bfdot za.s[x9, 0], { z20.h-z23.h }, z15.h[1]\n" + ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb699 // bfdot za.s[x9, 1], { z20.h-z23.h }, z15.h[1]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fb61a // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[1]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d698 // bfdot za.s[x10, 0], { z20.h-z23.h }, z3.h[1]\n" - ".inst 0xc153d619 // bfdot za.s[x10, 1], { z16.h-z19.h }, z3.h[1]\n" - ".inst 0xc153d79a // bfdot za.s[x10, 2], { z28.h-z31.h }, z3.h[1]\n" "ble 25f\n" - ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" - ".inst 0xa041a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fb898 // bfdot za.s[x9, 0], { z4.h-z7.h }, z15.h[2]\n" + ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb819 // bfdot za.s[x9, 1], { z0.h-z3.h }, z15.h[2]\n" + ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fbb1a // bfdot za.s[x9, 2], { z24.h-z27.h }, z15.h[2]\n" "addvl x26, x26, #16\n" - ".inst 0xc153db98 // bfdot za.s[x10, 0], { z28.h-z31.h }, z3.h[2]\n" - ".inst 0xc153d919 // bfdot za.s[x10, 1], { z8.h-z11.h }, z3.h[2]\n" - ".inst 0xc153d99a // bfdot za.s[x10, 2], { z12.h-z15.h }, z3.h[2]\n" "ble 25f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + ".inst 0xc15fbf98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z15.h[3]\n" + ".inst 0xa041a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fbd19 // bfdot za.s[x9, 1], { z8.h-z11.h }, z15.h[3]\n" + ".inst 0xa042a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fbc9a // bfdot za.s[x9, 2], { z4.h-z7.h }, z15.h[3]\n" "addvl x26, x26, #16\n" - ".inst 0xc153dd98 // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[3]\n" - ".inst 0xc153df19 // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[3]\n" - ".inst 0xc153dd1a // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[3]\n" "25:" // Width 3: Multiply loop: multiply skip "tbz %x[flags], #1, 26f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0064c08 // mova { z8.d-z11.d }, za.d[x10, #0]\n" - ".inst 0xc0064c20 // mova { z0.d-z3.d }, za.d[x10, #1]\n" + ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" "ld1rw { z17.s }, p2/Z, [x21]\n" - ".inst 0xc0064c44 // mova { z4.d-z7.d }, za.d[x10, #2]\n" + ".inst 0xc0062c24 // mova { z4.d-z7.d }, za.d[x9, #1]\n" "ld1rw { z16.s }, p2/Z, [x20]\n" - ".inst 0xc1b0ca28 // fclamp { z8.s-z11.s }, z17.s, z16.s\n" - ".inst 0xc1b0ca20 // fclamp { z0.s-z3.s }, z17.s, z16.s\n" + ".inst 0xc1b0ca3c // fclamp { z28.s-z31.s }, z17.s, z16.s\n" + ".inst 0xc0062c4c // mova { z12.d-z15.d }, za.d[x9, #2]\n" + ".inst 0xa060c73c // st1w { z28.s-z31.s }, pn9.b, [x25]\n" ".inst 0xc1b0ca24 // fclamp { z4.s-z7.s }, z17.s, z16.s\n" - ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" - ".inst 0xa061c720 // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n" - ".inst 0xa062c324 // st1w { z4.s-z7.s }, p8, [x25, #0x8, MUL VL]\n" + ".inst 0xa061c724 // st1w { z4.s-z7.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc1b0ca2c // fclamp { z12.s-z15.s }, z17.s, z16.s\n" + ".inst 0xa062c32c // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n" "addvl x25, x25, #12\n" "b 27f\n" "26:" // Width 3: No activation - ".inst 0xc0064c08 // mova { z8.d-z11.d }, za.d[x10, #0]\n" - ".inst 0xc0064c20 // mova { z0.d-z3.d }, za.d[x10, #1]\n" - ".inst 0xc0064c50 // mova { z16.d-z19.d }, za.d[x10, #2]\n" - ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" - ".inst 0xa061c720 // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" + ".inst 0xa060c720 // st1w { z0.s-z3.s }, pn9.b, [x25]\n" + ".inst 0xc0062c30 // mova { z16.d-z19.d }, za.d[x9, #1]\n" + ".inst 0xa061c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" ".inst 0xa062c330 // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n" "addvl x25, x25, #12\n" "27:" // Width 3: Output done @@ -439,20 +439,20 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "mov x20, #0x3\n" "mov x23, %x[A_ptr]\n" "lsl x21, %x[K], #0x2\n" - "msub x20, x9, x20, %x[N]\n" + "msub x20, x10, x20, %x[N]\n" "mov x22, %x[K]\n" ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 29f\n" ".inst 0xa040c70c // ld1w { z12.s-z15.s }, pn9.b/Z, [x24]\n" - ".inst 0xa041c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xa042c714 // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xa043c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n" - ".inst 0xc0044d80 // mova za.d[x10, #0], { z12.d-z15.d }\n" + ".inst 0xc0042d80 // mova za.d[x9, #0], { z12.d-z15.d }\n" + ".inst 0xa041c70c // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042d81 // mova za.d[x9, #1], { z12.d-z15.d }\n" + ".inst 0xa042c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" + ".inst 0xc0042e02 // mova za.d[x9, #2], { z16.d-z19.d }\n" + ".inst 0xa043c714 // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n" + ".inst 0xc0042e83 // mova za.d[x9, #3], { z20.d-z23.d }\n" "addvl x24, x24, #16\n" - ".inst 0xc0044c81 // mova za.d[x10, #1], { z4.d-z7.d }\n" - ".inst 0xc0044e82 // mova za.d[x10, #2], { z20.d-z23.d }\n" - ".inst 0xc0044e03 // mova za.d[x10, #3], { z16.d-z19.d }\n" "b 30f\n" "29:" // Width 4: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -461,140 +461,140 @@ void sme2_gemv_fp32bf16fp32_dot_16VL ( "ble 32f\n" "31:" // Width 4: Multiply loop: Main loop head "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z7.s }, p1/Z, [x23]\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z6.s }, p1/Z, [x23]\n" + ".inst 0x658aa8c6 // bfcvt z6.h, p2/M, z6.s\n" + "ld1rqw { z16.s }, p0/Z, [x23, #16]\n" + ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" + "uzp1 z6.h, z6.h, z6.h\n" "sub x22, x22, #0x8\n" - "ld1rqw { z4.s }, p0/Z, [x23, #16]\n" + "uzp1 z16.h, z16.h, z16.h\n" + "trn1 z6.d, z6.d, z16.d\n" + ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" "cmp x22, #0x8\n" - "add x23, x23, #0x20\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0x658aa8e7 // bfcvt z7.h, p2/M, z7.s\n" - ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0x658aa884 // bfcvt z4.h, p2/M, z4.s\n" - ".inst 0xa043a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" ".inst 0xa041a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "uzp1 z7.h, z7.h, z7.h\n" - ".inst 0xa042a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "uzp1 z4.h, z4.h, z4.h\n" - "trn1 z7.d, z7.d, z4.d\n" - ".inst 0xc157d218 // bfdot za.s[x10, 0], { z16.h-z19.h }, z7.h[0]\n" + ".inst 0xc156b198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z6.h[0]\n" + "add x23, x23, #0x20\n" + ".inst 0xa042a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc156b119 // bfdot za.s[x9, 1], { z8.h-z11.h }, z6.h[0]\n" ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc156b19a // bfdot za.s[x9, 2], { z12.h-z15.h }, z6.h[0]\n" "addvl x26, x26, #16\n" - ".inst 0xc157d199 // bfdot za.s[x10, 1], { z12.h-z15.h }, z7.h[0]\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - ".inst 0xc157d39a // bfdot za.s[x10, 2], { z28.h-z31.h }, z7.h[0]\n" - ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc157d31b // bfdot za.s[x10, 3], { z24.h-z27.h }, z7.h[0]\n" - ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc157d698 // bfdot za.s[x10, 0], { z20.h-z23.h }, z7.h[1]\n" + ".inst 0xc156b21b // bfdot za.s[x9, 3], { z16.h-z19.h }, z6.h[0]\n" + ".inst 0xa040a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n" + ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc156b518 // bfdot za.s[x9, 0], { z8.h-z11.h }, z6.h[1]\n" + ".inst 0xa042a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc156b599 // bfdot za.s[x9, 1], { z12.h-z15.h }, z6.h[1]\n" ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc156b41a // bfdot za.s[x9, 2], { z0.h-z3.h }, z6.h[1]\n" "addvl x26, x26, #16\n" - ".inst 0xc157d519 // bfdot za.s[x10, 1], { z8.h-z11.h }, z7.h[1]\n" + ".inst 0xc156b69b // bfdot za.s[x9, 3], { z20.h-z23.h }, z6.h[1]\n" ".inst 0xa040a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n" - ".inst 0xc157d41a // bfdot za.s[x10, 2], { z0.h-z3.h }, z7.h[1]\n" - ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc157d61b // bfdot za.s[x10, 3], { z16.h-z19.h }, z7.h[1]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc157d998 // bfdot za.s[x10, 0], { z12.h-z15.h }, z7.h[2]\n" - ".inst 0xa043a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - "addvl x26, x26, #16\n" - ".inst 0xc157db99 // bfdot za.s[x10, 1], { z28.h-z31.h }, z7.h[2]\n" - ".inst 0xc157db1a // bfdot za.s[x10, 2], { z24.h-z27.h }, z7.h[2]\n" - ".inst 0xc157da9b // bfdot za.s[x10, 3], { z20.h-z23.h }, z7.h[2]\n" - ".inst 0xc157dd18 // bfdot za.s[x10, 0], { z8.h-z11.h }, z7.h[3]\n" - ".inst 0xc157dc19 // bfdot za.s[x10, 1], { z0.h-z3.h }, z7.h[3]\n" - ".inst 0xc157de1a // bfdot za.s[x10, 2], { z16.h-z19.h }, z7.h[3]\n" - ".inst 0xc157dd9b // bfdot za.s[x10, 3], { z12.h-z15.h }, z7.h[3]\n" + ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc156b918 // bfdot za.s[x9, 0], { z8.h-z11.h }, z6.h[2]\n" + ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc156b999 // bfdot za.s[x9, 1], { z12.h-z15.h }, z6.h[2]\n" + ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc156b91a // bfdot za.s[x9, 2], { z8.h-z11.h }, z6.h[2]\n" + "addvl x26, x26, #16\n" + ".inst 0xc156ba9b // bfdot za.s[x9, 3], { z20.h-z23.h }, z6.h[2]\n" + ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc156bf98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z6.h[3]\n" + ".inst 0xa042a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc156bd99 // bfdot za.s[x9, 1], { z12.h-z15.h }, z6.h[3]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc156bf1a // bfdot za.s[x9, 2], { z24.h-z27.h }, z6.h[3]\n" + "addvl x26, x26, #16\n" + ".inst 0xc156be1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z6.h[3]\n" "bgt 31b\n" "32:" // Width 4: Multiply loop: Single iteration only "whilelt p1.s, XZR, x22\n" - "whilelt p0.s, x28, x22\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" - "ld1rqw { z3.s }, p1/Z, [x23]\n" - "subs x22, x22, #0x2\n" + "whilelt p0.s, x27, x22\n" + "ld1rqw { z15.s }, p1/Z, [x23]\n" + ".inst 0x658aa9ef // bfcvt z15.h, p2/M, z15.s\n" "ld1rqw { z16.s }, p0/Z, [x23, #16]\n" - "add x23, x23, #0x20\n" - ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0x658aa863 // bfcvt z3.h, p2/M, z3.s\n" - ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" - ".inst 0xa043a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - "addvl x26, x26, #16\n" - "uzp1 z3.h, z3.h, z3.h\n" + "uzp1 z15.h, z15.h, z15.h\n" + "subs x22, x22, #0x2\n" "uzp1 z16.h, z16.h, z16.h\n" - "trn1 z3.d, z3.d, z16.d\n" - ".inst 0xc153d318 // bfdot za.s[x10, 0], { z24.h-z27.h }, z3.h[0]\n" - ".inst 0xc153d399 // bfdot za.s[x10, 1], { z28.h-z31.h }, z3.h[0]\n" - ".inst 0xc153d11a // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[0]\n" - ".inst 0xc153d19b // bfdot za.s[x10, 3], { z12.h-z15.h }, z3.h[0]\n" + "trn1 z15.d, z15.d, z16.d\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + "add x23, x23, #0x20\n" + ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb318 // bfdot za.s[x9, 0], { z24.h-z27.h }, z15.h[0]\n" + ".inst 0xa042a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fb099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z15.h[0]\n" + ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc15fb01a // bfdot za.s[x9, 2], { z0.h-z3.h }, z15.h[0]\n" + "addvl x26, x26, #16\n" + ".inst 0xc15fb21b // bfdot za.s[x9, 3], { z16.h-z19.h }, z15.h[0]\n" "ble 33f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" - ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153d598 // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[1]\n" + ".inst 0xc15fb718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z15.h[1]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fb619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[1]\n" + ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fb69a // bfdot za.s[x9, 2], { z20.h-z23.h }, z15.h[1]\n" + ".inst 0xa043a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc15fb41b // bfdot za.s[x9, 3], { z0.h-z3.h }, z15.h[1]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d799 // bfdot za.s[x10, 1], { z28.h-z31.h }, z3.h[1]\n" - ".inst 0xc153d51a // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[1]\n" - ".inst 0xc153d71b // bfdot za.s[x10, 3], { z24.h-z27.h }, z3.h[1]\n" "ble 33f\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" "subs x22, x22, #0x2\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153db18 // bfdot za.s[x10, 0], { z24.h-z27.h }, z3.h[2]\n" + ".inst 0xc15fba18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[2]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[2]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fba1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[2]\n" + ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xc15fba9b // bfdot za.s[x9, 3], { z20.h-z23.h }, z15.h[2]\n" "addvl x26, x26, #16\n" - ".inst 0xc153d999 // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[2]\n" - ".inst 0xc153db9a // bfdot za.s[x10, 2], { z28.h-z31.h }, z3.h[2]\n" - ".inst 0xc153da1b // bfdot za.s[x10, 3], { z16.h-z19.h }, z3.h[2]\n" "ble 33f\n" - ".inst 0xa040a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n" - ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" + ".inst 0xc15fbe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[3]\n" + ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xc15fbe19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[3]\n" + ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xc15fbe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[3]\n" ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc153dd18 // bfdot za.s[x10, 0], { z8.h-z11.h }, z3.h[3]\n" + ".inst 0xc15fbe1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z15.h[3]\n" "addvl x26, x26, #16\n" - ".inst 0xc153df19 // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[3]\n" - ".inst 0xc153de9a // bfdot za.s[x10, 2], { z20.h-z23.h }, z3.h[3]\n" - ".inst 0xc153de1b // bfdot za.s[x10, 3], { z16.h-z19.h }, z3.h[3]\n" "33:" // Width 4: Multiply loop: multiply skip "tbz %x[flags], #1, 34f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0064c04 // mova { z4.d-z7.d }, za.d[x10, #0]\n" - ".inst 0xc0064c28 // mova { z8.d-z11.d }, za.d[x10, #1]\n" + ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\n" "ld1rw { z21.s }, p2/Z, [x21]\n" - ".inst 0xc0064c4c // mova { z12.d-z15.d }, za.d[x10, #2]\n" + ".inst 0xc0062c38 // mova { z24.d-z27.d }, za.d[x9, #1]\n" "ld1rw { z20.s }, p2/Z, [x20]\n" - ".inst 0xc0064c70 // mova { z16.d-z19.d }, za.d[x10, #3]\n" - ".inst 0xc1b4caa4 // fclamp { z4.s-z7.s }, z21.s, z20.s\n" - ".inst 0xc1b4caa8 // fclamp { z8.s-z11.s }, z21.s, z20.s\n" ".inst 0xc1b4caac // fclamp { z12.s-z15.s }, z21.s, z20.s\n" + ".inst 0xc0062c40 // mova { z0.d-z3.d }, za.d[x9, #2]\n" + ".inst 0xa060c72c // st1w { z12.s-z15.s }, pn9.b, [x25]\n" + ".inst 0xc1b4cab8 // fclamp { z24.s-z27.s }, z21.s, z20.s\n" + ".inst 0xc0062c70 // mova { z16.d-z19.d }, za.d[x9, #3]\n" + ".inst 0xa061c738 // st1w { z24.s-z27.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc1b4caa0 // fclamp { z0.s-z3.s }, z21.s, z20.s\n" + ".inst 0xa062c720 // st1w { z0.s-z3.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" - ".inst 0xa060c724 // st1w { z4.s-z7.s }, pn9.b, [x25]\n" - ".inst 0xa061c728 // st1w { z8.s-z11.s }, pn9.b, [x25, #0x4, MUL VL]\n" - ".inst 0xa062c72c // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xa063c330 // st1w { z16.s-z19.s }, p8, [x25, #0xc, MUL VL]\n" "addvl x25, x25, #16\n" "b 35f\n" "34:" // Width 4: No activation - ".inst 0xc0064c00 // mova { z0.d-z3.d }, za.d[x10, #0]\n" - ".inst 0xc0064c24 // mova { z4.d-z7.d }, za.d[x10, #1]\n" - ".inst 0xc0064c4c // mova { z12.d-z15.d }, za.d[x10, #2]\n" - ".inst 0xc0064c78 // mova { z24.d-z27.d }, za.d[x10, #3]\n" - ".inst 0xa060c720 // st1w { z0.s-z3.s }, pn9.b, [x25]\n" - ".inst 0xa061c724 // st1w { z4.s-z7.s }, pn9.b, [x25, #0x4, MUL VL]\n" - ".inst 0xa062c72c // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n" + ".inst 0xc0062c10 // mova { z16.d-z19.d }, za.d[x9, #0]\n" + ".inst 0xa060c730 // st1w { z16.s-z19.s }, pn9.b, [x25]\n" + ".inst 0xc0062c30 // mova { z16.d-z19.d }, za.d[x9, #1]\n" + ".inst 0xa061c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c54 // mova { z20.d-z23.d }, za.d[x9, #2]\n" + ".inst 0xa062c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x8, MUL VL]\n" + ".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n" ".inst 0xa063c338 // st1w { z24.s-z27.s }, p8, [x25, #0xc, MUL VL]\n" "addvl x25, x25, #16\n" "35:" // Width 4: Output done - "subs x27, x27, #0x4\n" - "sub %x[N], %x[N], x9, LSL #2\n" + "subs x28, x28, #0x4\n" + "sub %x[N], %x[N], x10, LSL #2\n" "bgt 4b\n" "36:" // Exit ".inst 0xd503467f // SMSTOP\n" |