diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp | 705 |
1 files changed, 352 insertions, 353 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp index b8e65e6999..0d5ea54cb7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp @@ -45,18 +45,18 @@ void sve_hybrid_s8qa_mmla_4x4VL ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; - void *output_ptr = {}; } ka; unsigned long flags=0; + void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - ka.output_ptr=(void *)(output_arg.indirect.ptr); + output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - ka.output_ptr=(void *)(output_arg.direct.base); + output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -91,24 +91,24 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bic %x[flags], %x[flags], #0x80000000\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x27, %x[output_ptr]\n" "2:" // Height 1: Column loop "mov x20, #0x0\n" + "whilelt p1.b, x20, x9\n" "mov z16.s, #0x0\n" "mov z17.s, #0x0\n" "mov z18.s, #0x0\n" "mov z19.s, #0x0\n" "mov z20.s, #0x0\n" "mov z21.s, #0x0\n" - "whilelt p1.b, x20, x9\n" "mov z22.s, #0x0\n" "mov z23.s, #0x0\n" "3:" // Height 1: setup done "mov x26, #0x0\n" "4:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 5f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -124,43 +124,43 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "ble 9f\n" "7:" // Height 1: Multiply loop: Main loop head "whilelt p0.b, XZR, x25\n" - "ld1b { z30.b }, p2/Z, [x28]\n" - "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z28.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z25.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n" - "add x24, x24, #0x10\n" - "trn1 z0.d, z1.d, z31.d\n" - ".inst 0x451e9810 // smmla z16.s, z0.b, z30.b\n" - "ld1b { z8.b }, p2/Z, [x28, #7, MUL VL]\n" + "trn1 z0.d, z1.d, z26.d\n" + "ld1b { z24.b }, p2/Z, [x28]\n" + ".inst 0x45189810 // smmla z16.s, z0.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n" + "trn2 z1.d, z1.d, z26.d\n" + "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45199814 // smmla z20.s, z0.b, z25.b\n" + ".inst 0x45189811 // smmla z17.s, z0.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x451a9815 // smmla z21.s, z0.b, z26.b\n" + ".inst 0x45199812 // smmla z18.s, z0.b, z25.b\n" + "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n" + "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - "trn2 z1.d, z1.d, z31.d\n" - ".inst 0x451d9814 // smmla z20.s, z0.b, z29.b\n" - ".inst 0x451c9811 // smmla z17.s, z0.b, z28.b\n" - ".inst 0x451b9815 // smmla z21.s, z0.b, z27.b\n" - ".inst 0x451a9812 // smmla z18.s, z0.b, z26.b\n" - "ld1b { z31.b }, p2/Z, [x28, #-8, MUL VL]\n" - ".inst 0x45199816 // smmla z22.s, z0.b, z25.b\n" - ".inst 0x45189813 // smmla z19.s, z0.b, z24.b\n" - "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n" - "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n" - ".inst 0x45089817 // smmla z23.s, z0.b, z8.b\n" - "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n" + ".inst 0x45189816 // smmla z22.s, z0.b, z24.b\n" + "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n" + ".inst 0x451a9813 // smmla z19.s, z0.b, z26.b\n" + ".inst 0x45199817 // smmla z23.s, z0.b, z25.b\n" + "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n" + ".inst 0x45189830 // smmla z16.s, z1.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n" + ".inst 0x451a9834 // smmla z20.s, z1.b, z26.b\n" "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n" - ".inst 0x451f9830 // smmla z16.s, z1.b, z31.b\n" "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n" + ".inst 0x45199831 // smmla z17.s, z1.b, z25.b\n" + ".inst 0x45189835 // smmla z21.s, z1.b, z24.b\n" "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n" "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n" - ".inst 0x451e9834 // smmla z20.s, z1.b, z30.b\n" - ".inst 0x451d9831 // smmla z17.s, z1.b, z29.b\n" - ".inst 0x451c9835 // smmla z21.s, z1.b, z28.b\n" ".inst 0x451b9832 // smmla z18.s, z1.b, z27.b\n" ".inst 0x451a9836 // smmla z22.s, z1.b, z26.b\n" ".inst 0x45199833 // smmla z19.s, z1.b, z25.b\n" ".inst 0x45189837 // smmla z23.s, z1.b, z24.b\n" + "add x24, x24, #0x10\n" "tbnz %x[flags], #31, 8f\n" "sdot z11.s, z0.b, z15.b\n" "sdot z11.s, z1.b, z15.b\n" @@ -170,45 +170,45 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bgt 7b\n" "9:" // Height 1: Multiply loop: Single iteration only "whilelt p0.b, XZR, x25\n" + "ld1rqb { z1.b }, p0/Z, [x24]\n" + "trn1 z0.d, z1.d, z27.d\n" "ld1b { z24.b }, p2/Z, [x28]\n" - "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45189810 // smmla z16.s, z0.b, z24.b\n" + "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n" "subs x25, x25, #0x8\n" + "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n" + "trn2 z1.d, z1.d, z27.d\n" + ".inst 0x451a9814 // smmla z20.s, z0.b, z26.b\n" "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n" "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n" - "ld1rqb { z1.b }, p0/Z, [x24]\n" + ".inst 0x45199811 // smmla z17.s, z0.b, z25.b\n" + ".inst 0x45189815 // smmla z21.s, z0.b, z24.b\n" "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n" - "trn1 z0.d, z1.d, z31.d\n" - ".inst 0x45189810 // smmla z16.s, z0.b, z24.b\n" "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n" - "addvl x28, x28, #8\n" - "trn2 z1.d, z1.d, z31.d\n" - ".inst 0x451e9814 // smmla z20.s, z0.b, z30.b\n" - ".inst 0x451d9811 // smmla z17.s, z0.b, z29.b\n" - ".inst 0x451c9815 // smmla z21.s, z0.b, z28.b\n" ".inst 0x451b9812 // smmla z18.s, z0.b, z27.b\n" ".inst 0x451a9816 // smmla z22.s, z0.b, z26.b\n" ".inst 0x45199813 // smmla z19.s, z0.b, z25.b\n" ".inst 0x45189817 // smmla z23.s, z0.b, z24.b\n" + "addvl x28, x28, #8\n" "ble 10f\n" "ld1b { z24.b }, p2/Z, [x28]\n" - "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n" ".inst 0x45189830 // smmla z16.s, z1.b, z24.b\n" + "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n" + ".inst 0x45189834 // smmla z20.s, z1.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45199831 // smmla z17.s, z1.b, z25.b\n" + ".inst 0x45189835 // smmla z21.s, z1.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x45199832 // smmla z18.s, z1.b, z25.b\n" + ".inst 0x45189836 // smmla z22.s, z1.b, z24.b\n" "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n" "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n" - ".inst 0x451e9834 // smmla z20.s, z1.b, z30.b\n" - "addvl x28, x28, #8\n" - ".inst 0x451d9831 // smmla z17.s, z1.b, z29.b\n" - ".inst 0x451c9835 // smmla z21.s, z1.b, z28.b\n" - ".inst 0x451b9832 // smmla z18.s, z1.b, z27.b\n" - ".inst 0x451a9836 // smmla z22.s, z1.b, z26.b\n" ".inst 0x45199833 // smmla z19.s, z1.b, z25.b\n" ".inst 0x45189837 // smmla z23.s, z1.b, z24.b\n" + "addvl x28, x28, #8\n" "10:" // Height 1: Multiply loop: multiply skip "tbnz %x[flags], #31, 11f\n" "sdot z11.s, z0.b, z15.b\n" @@ -225,32 +225,32 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov z23.d, z16.d\n" "tbnz %x[flags], #31, 12f\n" "add x20, %x[qp], %[b_offset]\n" + "ld1rw { z16.s }, p2/Z, [x20]\n" ".inst 0x4491a96b // addp z11.s, p2/m, z11.s, z11.s\n" - "ld1rw { z9.s }, p2/Z, [x20]\n" - "neg z9.s, p2/M, z9.s\n" + "neg z16.s, p2/M, z16.s\n" "mov z11.s, z11.s[0]\n" - "mul z11.s, p2/M, z11.s, z9.s\n" + "mul z11.s, p2/M, z11.s, z16.s\n" "12:" // Height 1: skip row sum fixup "add z23.s, z23.s, z11.s\n" "add z17.s, z17.s, z11.s\n" "ld1w { z22.s }, p2/Z, [x10]\n" - "ld1w { z24.s }, p2/Z, [x10, #1, MUL VL]\n" + "ld1w { z21.s }, p2/Z, [x10, #1, MUL VL]\n" "add z18.s, z18.s, z11.s\n" "add z19.s, z19.s, z11.s\n" - "ld1w { z21.s }, p2/Z, [x10, #2, MUL VL]\n" - "ld1w { z20.s }, p2/Z, [x10, #3, MUL VL]\n" + "ld1w { z20.s }, p2/Z, [x10, #2, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x10, #3, MUL VL]\n" "add x20, %x[qp], %[per_layer_mul]\n" "orr %x[flags], %x[flags], #0x80000000\n" "add z23.s, z23.s, z22.s\n" - "add z17.s, z17.s, z24.s\n" + "add z17.s, z17.s, z21.s\n" + "add z18.s, z18.s, z20.s\n" + "add z19.s, z19.s, z16.s\n" "ld1rw { z16.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - "addvl x10, x10, #4\n" - "add z18.s, z18.s, z21.s\n" - "add z19.s, z19.s, z20.s\n" "ld1rw { z0.s }, p2/Z, [x20]\n" ".inst 0x04b076f7 // sqrdmulh z23.s, z23.s, z16.s\n" ".inst 0x04b07631 // sqrdmulh z17.s, z17.s, z16.s\n" + "addvl x10, x10, #4\n" ".inst 0x04b07652 // sqrdmulh z18.s, z18.s, z16.s\n" ".inst 0x04b07673 // sqrdmulh z19.s, z19.s, z16.s\n" "tbz %x[flags], #5, 13f\n" @@ -268,19 +268,19 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "sqadd z19.s, z19.s, z16.s\n" "13:" // Height 1: no shift correction "add x20, %x[qp], %[c_offset]\n" + "ld1rw { z16.s }, p2/Z, [x20]\n" ".inst 0x44828817 // srshl z23.s, p2/M, z23.s, z0.s\n" - "ld1rw { z21.s }, p2/Z, [x20]\n" + "add z23.s, z23.s, z16.s\n" ".inst 0x44828811 // srshl z17.s, p2/M, z17.s, z0.s\n" ".inst 0x44828812 // srshl z18.s, p2/M, z18.s, z0.s\n" + "add z17.s, z17.s, z16.s\n" + "add z18.s, z18.s, z16.s\n" ".inst 0x44828813 // srshl z19.s, p2/M, z19.s, z0.s\n" "add x20, %x[qp], %[maxval]\n" "ld1rw { z20.s }, p2/Z, [x20]\n" - "add z23.s, z23.s, z21.s\n" + "add z19.s, z19.s, z16.s\n" "add x20, %x[qp], %[minval]\n" - "add z17.s, z17.s, z21.s\n" - "add z18.s, z18.s, z21.s\n" "ld1rw { z16.s }, p2/Z, [x20]\n" - "add z19.s, z19.s, z21.s\n" "smin z23.s, p2/M, z23.s, z20.s\n" "smin z17.s, p2/M, z17.s, z20.s\n" "smin z18.s, p2/M, z18.s, z20.s\n" @@ -288,8 +288,8 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "smax z23.s, p2/M, z23.s, z16.s\n" "smax z17.s, p2/M, z17.s, z16.s\n" "smax z18.s, p2/M, z18.s, z16.s\n" - "smax z19.s, p2/M, z19.s, z16.s\n" "uzp1 z23.h, z23.h, z17.h\n" + "smax z19.s, p2/M, z19.s, z16.s\n" "uzp1 z16.h, z18.h, z19.h\n" "uzp1 z23.b, z23.b, z16.b\n" "st1b { z23.b }, p1, [x27]\n" @@ -307,24 +307,24 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov z15.b, #0x1\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x27, %x[output_ptr]\n" "16:" // Height 2: Column loop "mov x20, #0x0\n" + "whilelt p1.b, x20, x9\n" "mov z16.s, #0x0\n" "mov z17.s, #0x0\n" "mov z18.s, #0x0\n" "mov z19.s, #0x0\n" "mov z20.s, #0x0\n" "mov z21.s, #0x0\n" - "whilelt p1.b, x20, x9\n" "mov z22.s, #0x0\n" "mov z23.s, #0x0\n" "17:" // Height 2: setup done "mov x26, #0x0\n" "18:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 19f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -343,45 +343,45 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "ble 23f\n" "21:" // Height 2: Multiply loop: Main loop head "whilelt p0.b, XZR, x25\n" - "ld1b { z31.b }, p2/Z, [x28]\n" - "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "ld1rqb { z25.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" - "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n" - "trn1 z0.d, z1.d, z25.d\n" - "trn2 z1.d, z1.d, z25.d\n" - ".inst 0x451f9810 // smmla z16.s, z0.b, z31.b\n" + "ld1rqb { z26.b }, p0/Z, [x23]\n" + "trn1 z0.d, z1.d, z26.d\n" + "ld1b { z24.b }, p2/Z, [x28]\n" + ".inst 0x45189810 // smmla z16.s, z0.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n" + "trn2 z1.d, z1.d, z26.d\n" + "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45199814 // smmla z20.s, z0.b, z25.b\n" + ".inst 0x45189811 // smmla z17.s, z0.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x451a9815 // smmla z21.s, z0.b, z26.b\n" + ".inst 0x45199812 // smmla z18.s, z0.b, z25.b\n" + "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n" "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x451e9814 // smmla z20.s, z0.b, z30.b\n" - ".inst 0x451d9811 // smmla z17.s, z0.b, z29.b\n" - ".inst 0x451c9815 // smmla z21.s, z0.b, z28.b\n" - ".inst 0x451b9812 // smmla z18.s, z0.b, z27.b\n" - ".inst 0x451a9816 // smmla z22.s, z0.b, z26.b\n" - ".inst 0x45189813 // smmla z19.s, z0.b, z24.b\n" + ".inst 0x45189816 // smmla z22.s, z0.b, z24.b\n" "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n" + ".inst 0x451a9813 // smmla z19.s, z0.b, z26.b\n" ".inst 0x45199817 // smmla z23.s, z0.b, z25.b\n" - "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n" - "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n" - "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n" - "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n" + "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n" ".inst 0x45189830 // smmla z16.s, z1.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n" + ".inst 0x451a9834 // smmla z20.s, z1.b, z26.b\n" + "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n" "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n" + ".inst 0x45199831 // smmla z17.s, z1.b, z25.b\n" + ".inst 0x45189835 // smmla z21.s, z1.b, z24.b\n" "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n" "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n" - ".inst 0x451e9834 // smmla z20.s, z1.b, z30.b\n" - ".inst 0x451d9831 // smmla z17.s, z1.b, z29.b\n" - ".inst 0x451c9835 // smmla z21.s, z1.b, z28.b\n" ".inst 0x451b9832 // smmla z18.s, z1.b, z27.b\n" ".inst 0x451a9836 // smmla z22.s, z1.b, z26.b\n" ".inst 0x45199833 // smmla z19.s, z1.b, z25.b\n" ".inst 0x45189837 // smmla z23.s, z1.b, z24.b\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" "tbnz %x[flags], #31, 22f\n" "sdot z11.s, z0.b, z15.b\n" "sdot z11.s, z1.b, z15.b\n" @@ -391,46 +391,46 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bgt 21b\n" "23:" // Height 2: Multiply loop: Single iteration only "whilelt p0.b, XZR, x25\n" - "ld1b { z29.b }, p2/Z, [x28]\n" - "ld1b { z28.b }, p2/Z, [x28, #1, MUL VL]\n" - "subs x25, x25, #0x8\n" - "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" - "ld1rqb { z24.b }, p0/Z, [x23]\n" + "ld1rqb { z27.b }, p0/Z, [x23]\n" + "trn1 z0.d, z1.d, z27.d\n" + "ld1b { z24.b }, p2/Z, [x28]\n" + ".inst 0x45189810 // smmla z16.s, z0.b, z24.b\n" + "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n" + "subs x25, x25, #0x8\n" + "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n" + "trn2 z1.d, z1.d, z27.d\n" + ".inst 0x451a9814 // smmla z20.s, z0.b, z26.b\n" + "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x45199811 // smmla z17.s, z0.b, z25.b\n" + ".inst 0x45189815 // smmla z21.s, z0.b, z24.b\n" "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n" - "trn1 z0.d, z1.d, z24.d\n" - "trn2 z1.d, z1.d, z24.d\n" - ".inst 0x451d9810 // smmla z16.s, z0.b, z29.b\n" "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n" - "addvl x28, x28, #8\n" - ".inst 0x451c9814 // smmla z20.s, z0.b, z28.b\n" - ".inst 0x45049811 // smmla z17.s, z0.b, z4.b\n" - ".inst 0x451b9815 // smmla z21.s, z0.b, z27.b\n" - ".inst 0x451a9812 // smmla z18.s, z0.b, z26.b\n" - ".inst 0x45069816 // smmla z22.s, z0.b, z6.b\n" + ".inst 0x451b9812 // smmla z18.s, z0.b, z27.b\n" + ".inst 0x451a9816 // smmla z22.s, z0.b, z26.b\n" ".inst 0x45199813 // smmla z19.s, z0.b, z25.b\n" ".inst 0x45189817 // smmla z23.s, z0.b, z24.b\n" + "addvl x28, x28, #8\n" "ble 24f\n" "ld1b { z24.b }, p2/Z, [x28]\n" - "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n" ".inst 0x45189830 // smmla z16.s, z1.b, z24.b\n" + "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n" + ".inst 0x45189834 // smmla z20.s, z1.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45199831 // smmla z17.s, z1.b, z25.b\n" + ".inst 0x45189835 // smmla z21.s, z1.b, z24.b\n" + "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x45199832 // smmla z18.s, z1.b, z25.b\n" + ".inst 0x45189836 // smmla z22.s, z1.b, z24.b\n" "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n" "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n" - ".inst 0x451e9834 // smmla z20.s, z1.b, z30.b\n" - "addvl x28, x28, #8\n" - ".inst 0x451d9831 // smmla z17.s, z1.b, z29.b\n" - ".inst 0x451c9835 // smmla z21.s, z1.b, z28.b\n" - ".inst 0x451b9832 // smmla z18.s, z1.b, z27.b\n" - ".inst 0x451a9836 // smmla z22.s, z1.b, z26.b\n" ".inst 0x45199833 // smmla z19.s, z1.b, z25.b\n" ".inst 0x45189837 // smmla z23.s, z1.b, z24.b\n" + "addvl x28, x28, #8\n" "24:" // Height 2: Multiply loop: multiply skip "tbnz %x[flags], #31, 25f\n" "sdot z11.s, z0.b, z15.b\n" @@ -443,18 +443,18 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "uzp1 z24.d, z16.d, z20.d\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp2 z16.d, z16.d, z20.d\n" + "add x23, x27, x20\n" "uzp1 z20.d, z17.d, z21.d\n" "uzp2 z17.d, z17.d, z21.d\n" "uzp1 z21.d, z18.d, z22.d\n" "uzp2 z18.d, z18.d, z22.d\n" "uzp1 z22.d, z19.d, z23.d\n" "uzp2 z19.d, z19.d, z23.d\n" - "add x23, x27, x20\n" "mov z23.d, z24.d\n" "tbnz %x[flags], #31, 26f\n" "add x20, %x[qp], %[b_offset]\n" - ".inst 0x4491a96b // addp z11.s, p2/m, z11.s, z11.s\n" "ld1rw { z24.s }, p2/Z, [x20]\n" + ".inst 0x4491a96b // addp z11.s, p2/m, z11.s, z11.s\n" "neg z24.s, p2/M, z24.s\n" "mov z12.s, z11.s[3]\n" "mov z11.s, z11.s[0]\n" @@ -497,24 +497,24 @@ void sve_hybrid_s8qa_mmla_4x4VL ( ".inst 0x04b87673 // sqrdmulh z19.s, z19.s, z24.s\n" "tbz %x[flags], #5, 27f\n" "and z24.d, z23.d, z0.d\n" + "asr z24.s, z24.s, #0x1f\n" + "sqadd z23.s, z23.s, z24.s\n" "and z30.d, z20.d, z0.d\n" "and z29.d, z21.d, z0.d\n" "and z28.d, z22.d, z0.d\n" "and z27.d, z16.d, z0.d\n" "and z26.d, z17.d, z0.d\n" - "asr z24.s, z24.s, #0x1f\n" "and z25.d, z18.d, z0.d\n" + "and z24.d, z19.d, z0.d\n" "asr z30.s, z30.s, #0x1f\n" "asr z29.s, z29.s, #0x1f\n" "asr z28.s, z28.s, #0x1f\n" "asr z27.s, z27.s, #0x1f\n" - "sqadd z23.s, z23.s, z24.s\n" - "and z24.d, z19.d, z0.d\n" "asr z26.s, z26.s, #0x1f\n" "asr z25.s, z25.s, #0x1f\n" + "asr z24.s, z24.s, #0x1f\n" "sqadd z20.s, z20.s, z30.s\n" "sqadd z21.s, z21.s, z29.s\n" - "asr z24.s, z24.s, #0x1f\n" "sqadd z22.s, z22.s, z28.s\n" "sqadd z16.s, z16.s, z27.s\n" "sqadd z17.s, z17.s, z26.s\n" @@ -522,27 +522,27 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "sqadd z19.s, z19.s, z24.s\n" "27:" // Height 2: no shift correction "add x20, %x[qp], %[c_offset]\n" + "ld1rw { z24.s }, p2/Z, [x20]\n" ".inst 0x44828817 // srshl z23.s, p2/M, z23.s, z0.s\n" - "ld1rw { z26.s }, p2/Z, [x20]\n" + "add z23.s, z23.s, z24.s\n" ".inst 0x44828814 // srshl z20.s, p2/M, z20.s, z0.s\n" ".inst 0x44828815 // srshl z21.s, p2/M, z21.s, z0.s\n" + "add z20.s, z20.s, z24.s\n" + "add z21.s, z21.s, z24.s\n" ".inst 0x44828816 // srshl z22.s, p2/M, z22.s, z0.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" + "add z22.s, z22.s, z24.s\n" + "add z16.s, z16.s, z24.s\n" ".inst 0x44828811 // srshl z17.s, p2/M, z17.s, z0.s\n" ".inst 0x44828812 // srshl z18.s, p2/M, z18.s, z0.s\n" - "add z23.s, z23.s, z26.s\n" + "add z17.s, z17.s, z24.s\n" + "add z18.s, z18.s, z24.s\n" ".inst 0x44828813 // srshl z19.s, p2/M, z19.s, z0.s\n" "add x20, %x[qp], %[maxval]\n" - "add z20.s, z20.s, z26.s\n" - "add z21.s, z21.s, z26.s\n" "ld1rw { z25.s }, p2/Z, [x20]\n" - "add z22.s, z22.s, z26.s\n" - "add z16.s, z16.s, z26.s\n" + "add z19.s, z19.s, z24.s\n" "add x20, %x[qp], %[minval]\n" - "add z17.s, z17.s, z26.s\n" - "add z18.s, z18.s, z26.s\n" "ld1rw { z24.s }, p2/Z, [x20]\n" - "add z19.s, z19.s, z26.s\n" "smin z23.s, p2/M, z23.s, z25.s\n" "smin z20.s, p2/M, z20.s, z25.s\n" "smin z21.s, p2/M, z21.s, z25.s\n" @@ -554,20 +554,20 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "smax z23.s, p2/M, z23.s, z24.s\n" "smax z20.s, p2/M, z20.s, z24.s\n" "smax z21.s, p2/M, z21.s, z24.s\n" + "uzp1 z23.h, z23.h, z20.h\n" "smax z22.s, p2/M, z22.s, z24.s\n" "smax z16.s, p2/M, z16.s, z24.s\n" + "uzp1 z20.h, z21.h, z22.h\n" + "uzp1 z23.b, z23.b, z20.b\n" "smax z17.s, p2/M, z17.s, z24.s\n" "smax z18.s, p2/M, z18.s, z24.s\n" - "smax z19.s, p2/M, z19.s, z24.s\n" - "uzp1 z23.h, z23.h, z20.h\n" - "uzp1 z20.h, z21.h, z22.h\n" "uzp1 z16.h, z16.h, z17.h\n" + "st1b { z23.b }, p1, [x27]\n" + "smax z19.s, p2/M, z19.s, z24.s\n" "uzp1 z17.h, z18.h, z19.h\n" - "uzp1 z23.b, z23.b, z20.b\n" "uzp1 z16.b, z16.b, z17.b\n" - "st1b { z23.b }, p1, [x27]\n" - "addvl x27, x27, #1\n" "st1b { z16.b }, p1, [x23]\n" + "addvl x27, x27, #1\n" "28:" // Height 2: Writeback done "decw x9, ALL, MUL #4\n" "cmp x9, XZR\n" @@ -582,16 +582,16 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov z15.b, #0x1\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x27, %x[output_ptr]\n" "30:" // Height 3: Column loop "mov x20, #0x0\n" + "whilelt p1.b, x20, x9\n" "mov z16.s, #0x0\n" "mov z17.s, #0x0\n" "mov z18.s, #0x0\n" "mov z19.s, #0x0\n" "mov z20.s, #0x0\n" "mov z21.s, #0x0\n" - "whilelt p1.b, x20, x9\n" "mov z22.s, #0x0\n" "mov z23.s, #0x0\n" "mov z24.s, #0x0\n" @@ -606,8 +606,8 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov x26, #0x0\n" "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 33f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -629,49 +629,49 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "ble 37f\n" "35:" // Height 3: Multiply loop: Main loop head "whilelt p0.b, XZR, x25\n" - "ld1b { z5.b }, p2/Z, [x28]\n" - "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z4.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" - "add x22, x22, #0x10\n" "trn1 z0.d, z1.d, z2.d\n" "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z6.d\n" - "trn2 z3.d, z3.d, z6.d\n" + "ld1b { z4.b }, p2/Z, [x28]\n" + "trn1 z2.d, z3.d, z5.d\n" + "trn2 z3.d, z3.d, z5.d\n" + ".inst 0x45049810 // smmla z16.s, z0.b, z4.b\n" + "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n" + ".inst 0x45049858 // smmla z24.s, z2.b, z4.b\n" + "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45059814 // smmla z20.s, z0.b, z5.b\n" + "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x4505985c // smmla z28.s, z2.b, z5.b\n" + ".inst 0x45049811 // smmla z17.s, z0.b, z4.b\n" "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" - ".inst 0x45059810 // smmla z16.s, z0.b, z5.b\n" - ".inst 0x450a9814 // smmla z20.s, z0.b, z10.b\n" - ".inst 0x45099811 // smmla z17.s, z0.b, z9.b\n" - ".inst 0x45089815 // smmla z21.s, z0.b, z8.b\n" - ".inst 0x45049812 // smmla z18.s, z0.b, z4.b\n" - ".inst 0x45059858 // smmla z24.s, z2.b, z5.b\n" "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x450a985c // smmla z28.s, z2.b, z10.b\n" - ".inst 0x45099859 // smmla z25.s, z2.b, z9.b\n" - ".inst 0x4508985d // smmla z29.s, z2.b, z8.b\n" - ".inst 0x4504985a // smmla z26.s, z2.b, z4.b\n" - ".inst 0x45079816 // smmla z22.s, z0.b, z7.b\n" + ".inst 0x45049859 // smmla z25.s, z2.b, z4.b\n" "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n" - ".inst 0x4507985e // smmla z30.s, z2.b, z7.b\n" - ".inst 0x45069813 // smmla z19.s, z0.b, z6.b\n" + ".inst 0x45099815 // smmla z21.s, z0.b, z9.b\n" + ".inst 0x4509985d // smmla z29.s, z2.b, z9.b\n" "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n" + ".inst 0x45089812 // smmla z18.s, z0.b, z8.b\n" + ".inst 0x4508985a // smmla z26.s, z2.b, z8.b\n" "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n" - ".inst 0x4506985b // smmla z27.s, z2.b, z6.b\n" - ".inst 0x45059817 // smmla z23.s, z0.b, z5.b\n" "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n" + ".inst 0x45079816 // smmla z22.s, z0.b, z7.b\n" + ".inst 0x4507985e // smmla z30.s, z2.b, z7.b\n" "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n" - ".inst 0x4505985f // smmla z31.s, z2.b, z5.b\n" - ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" + "add x24, x24, #0x10\n" + ".inst 0x45069813 // smmla z19.s, z0.b, z6.b\n" + ".inst 0x4506985b // smmla z27.s, z2.b, z6.b\n" "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n" + "add x23, x23, #0x10\n" + ".inst 0x45059817 // smmla z23.s, z0.b, z5.b\n" + ".inst 0x4505985f // smmla z31.s, z2.b, z5.b\n" "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n" + "add x22, x22, #0x10\n" + ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" ".inst 0x45049878 // smmla z24.s, z3.b, z4.b\n" "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n" ".inst 0x450a9834 // smmla z20.s, z1.b, z10.b\n" @@ -699,32 +699,32 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bgt 35b\n" "37:" // Height 3: Multiply loop: Single iteration only "whilelt p0.b, XZR, x25\n" - "ld1b { z4.b }, p2/Z, [x28]\n" - "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n" - "subs x25, x25, #0x8\n" - "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" "trn1 z0.d, z1.d, z2.d\n" "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z5.d\n" - "trn2 z3.d, z3.d, z5.d\n" + "ld1b { z5.b }, p2/Z, [x28]\n" + "trn1 z2.d, z3.d, z4.d\n" + "trn2 z3.d, z3.d, z4.d\n" + ".inst 0x45059810 // smmla z16.s, z0.b, z5.b\n" + "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n" + ".inst 0x45059858 // smmla z24.s, z2.b, z5.b\n" + "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n" + "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n" + "subs x25, x25, #0x8\n" + "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n" + ".inst 0x45049814 // smmla z20.s, z0.b, z4.b\n" + ".inst 0x4504985c // smmla z28.s, z2.b, z4.b\n" "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n" - ".inst 0x45049810 // smmla z16.s, z0.b, z4.b\n" - ".inst 0x450a9814 // smmla z20.s, z0.b, z10.b\n" - ".inst 0x45099811 // smmla z17.s, z0.b, z9.b\n" - ".inst 0x45089815 // smmla z21.s, z0.b, z8.b\n" - ".inst 0x45079812 // smmla z18.s, z0.b, z7.b\n" - ".inst 0x45049858 // smmla z24.s, z2.b, z4.b\n" "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n" - ".inst 0x450a985c // smmla z28.s, z2.b, z10.b\n" - "addvl x28, x28, #8\n" + ".inst 0x45099811 // smmla z17.s, z0.b, z9.b\n" ".inst 0x45099859 // smmla z25.s, z2.b, z9.b\n" + ".inst 0x45089815 // smmla z21.s, z0.b, z8.b\n" ".inst 0x4508985d // smmla z29.s, z2.b, z8.b\n" + "addvl x28, x28, #8\n" + ".inst 0x45079812 // smmla z18.s, z0.b, z7.b\n" ".inst 0x4507985a // smmla z26.s, z2.b, z7.b\n" ".inst 0x45069816 // smmla z22.s, z0.b, z6.b\n" ".inst 0x4506985e // smmla z30.s, z2.b, z6.b\n" @@ -734,24 +734,24 @@ void sve_hybrid_s8qa_mmla_4x4VL ( ".inst 0x4504985f // smmla z31.s, z2.b, z4.b\n" "ble 38f\n" "ld1b { z4.b }, p2/Z, [x28]\n" - "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n" + ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" + ".inst 0x45049878 // smmla z24.s, z3.b, z4.b\n" + "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n" + "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n" "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45059834 // smmla z20.s, z1.b, z5.b\n" + ".inst 0x4505987c // smmla z28.s, z3.b, z5.b\n" "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n" "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n" - ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" - ".inst 0x45049878 // smmla z24.s, z3.b, z4.b\n" + ".inst 0x45049831 // smmla z17.s, z1.b, z4.b\n" + ".inst 0x45049879 // smmla z25.s, z3.b, z4.b\n" "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n" "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n" - ".inst 0x450a9834 // smmla z20.s, z1.b, z10.b\n" - ".inst 0x450a987c // smmla z28.s, z3.b, z10.b\n" - ".inst 0x45099831 // smmla z17.s, z1.b, z9.b\n" - ".inst 0x45099879 // smmla z25.s, z3.b, z9.b\n" - "addvl x28, x28, #8\n" ".inst 0x45089835 // smmla z21.s, z1.b, z8.b\n" ".inst 0x4508987d // smmla z29.s, z3.b, z8.b\n" ".inst 0x45079832 // smmla z18.s, z1.b, z7.b\n" ".inst 0x4507987a // smmla z26.s, z3.b, z7.b\n" + "addvl x28, x28, #8\n" ".inst 0x45069836 // smmla z22.s, z1.b, z6.b\n" ".inst 0x4506987e // smmla z30.s, z3.b, z6.b\n" ".inst 0x45059833 // smmla z19.s, z1.b, z5.b\n" @@ -771,15 +771,15 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bne 32b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 z0.d, z16.d, z20.d\n" + "add x23, x27, x20\n" "uzp2 z16.d, z16.d, z20.d\n" "uzp1 z20.d, z17.d, z21.d\n" "uzp2 z17.d, z17.d, z21.d\n" + "add x22, x23, x20\n" "uzp1 z21.d, z18.d, z22.d\n" "uzp2 z18.d, z18.d, z22.d\n" - "add x23, x27, x20\n" "uzp1 z22.d, z19.d, z23.d\n" "uzp2 z19.d, z19.d, z23.d\n" - "add x22, x23, x20\n" "uzp1 z24.d, z24.d, z28.d\n" "uzp1 z25.d, z25.d, z29.d\n" "uzp1 z26.d, z26.d, z30.d\n" @@ -787,14 +787,14 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov z31.d, z0.d\n" "tbnz %x[flags], #31, 40f\n" "add x20, %x[qp], %[b_offset]\n" + "ld1rw { z23.s }, p2/Z, [x20]\n" ".inst 0x4491a96b // addp z11.s, p2/m, z11.s, z11.s\n" ".inst 0x4491a9ad // addp z13.s, p2/m, z13.s, z13.s\n" - "ld1rw { z23.s }, p2/Z, [x20]\n" "neg z23.s, p2/M, z23.s\n" "mov z12.s, z11.s[3]\n" "mov z11.s, z11.s[0]\n" - "mov z13.s, z13.s[0]\n" "mul z11.s, p2/M, z11.s, z23.s\n" + "mov z13.s, z13.s[0]\n" "mul z12.s, p2/M, z12.s, z23.s\n" "mul z13.s, p2/M, z13.s, z23.s\n" "40:" // Height 3: skip row sum fixup @@ -850,18 +850,18 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "and z29.d, z21.d, z0.d\n" "and z28.d, z22.d, z0.d\n" "and z23.d, z16.d, z0.d\n" - "and z3.d, z17.d, z0.d\n" "asr z1.s, z1.s, #0x1f\n" "asr z30.s, z30.s, #0x1f\n" "asr z29.s, z29.s, #0x1f\n" "asr z28.s, z28.s, #0x1f\n" "asr z23.s, z23.s, #0x1f\n" - "and z2.d, z18.d, z0.d\n" "sqadd z31.s, z31.s, z1.s\n" "sqadd z20.s, z20.s, z30.s\n" "sqadd z21.s, z21.s, z29.s\n" "sqadd z22.s, z22.s, z28.s\n" "sqadd z16.s, z16.s, z23.s\n" + "and z3.d, z17.d, z0.d\n" + "and z2.d, z18.d, z0.d\n" "and z1.d, z19.d, z0.d\n" "and z30.d, z24.d, z0.d\n" "and z29.d, z25.d, z0.d\n" @@ -883,35 +883,35 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "sqadd z27.s, z27.s, z23.s\n" "41:" // Height 3: no shift correction "add x20, %x[qp], %[c_offset]\n" + "ld1rw { z23.s }, p2/Z, [x20]\n" ".inst 0x4482881f // srshl z31.s, p2/M, z31.s, z0.s\n" - "ld1rw { z29.s }, p2/Z, [x20]\n" + "add z31.s, z31.s, z23.s\n" ".inst 0x44828814 // srshl z20.s, p2/M, z20.s, z0.s\n" ".inst 0x44828815 // srshl z21.s, p2/M, z21.s, z0.s\n" + "add z20.s, z20.s, z23.s\n" + "add z21.s, z21.s, z23.s\n" ".inst 0x44828816 // srshl z22.s, p2/M, z22.s, z0.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" + "add z22.s, z22.s, z23.s\n" + "add z16.s, z16.s, z23.s\n" ".inst 0x44828811 // srshl z17.s, p2/M, z17.s, z0.s\n" ".inst 0x44828812 // srshl z18.s, p2/M, z18.s, z0.s\n" - "add z31.s, z31.s, z29.s\n" + "add z17.s, z17.s, z23.s\n" + "add z18.s, z18.s, z23.s\n" ".inst 0x44828813 // srshl z19.s, p2/M, z19.s, z0.s\n" ".inst 0x44828818 // srshl z24.s, p2/M, z24.s, z0.s\n" - "add z20.s, z20.s, z29.s\n" - "add z21.s, z21.s, z29.s\n" + "add z19.s, z19.s, z23.s\n" + "add z24.s, z24.s, z23.s\n" ".inst 0x44828819 // srshl z25.s, p2/M, z25.s, z0.s\n" ".inst 0x4482881a // srshl z26.s, p2/M, z26.s, z0.s\n" - "add z22.s, z22.s, z29.s\n" - "add z16.s, z16.s, z29.s\n" + "add z25.s, z25.s, z23.s\n" + "add z26.s, z26.s, z23.s\n" ".inst 0x4482881b // srshl z27.s, p2/M, z27.s, z0.s\n" "add x20, %x[qp], %[maxval]\n" - "add z17.s, z17.s, z29.s\n" - "add z18.s, z18.s, z29.s\n" "ld1rw { z28.s }, p2/Z, [x20]\n" - "add z19.s, z19.s, z29.s\n" - "add z24.s, z24.s, z29.s\n" + "add z27.s, z27.s, z23.s\n" "add x20, %x[qp], %[minval]\n" - "add z25.s, z25.s, z29.s\n" - "add z26.s, z26.s, z29.s\n" "ld1rw { z23.s }, p2/Z, [x20]\n" - "add z27.s, z27.s, z29.s\n" "smin z31.s, p2/M, z31.s, z28.s\n" "smin z20.s, p2/M, z20.s, z28.s\n" "smin z21.s, p2/M, z21.s, z28.s\n" @@ -927,28 +927,28 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "smax z31.s, p2/M, z31.s, z23.s\n" "smax z20.s, p2/M, z20.s, z23.s\n" "smax z21.s, p2/M, z21.s, z23.s\n" + "uzp1 z31.h, z31.h, z20.h\n" "smax z22.s, p2/M, z22.s, z23.s\n" "smax z16.s, p2/M, z16.s, z23.s\n" + "uzp1 z20.h, z21.h, z22.h\n" + "uzp1 z31.b, z31.b, z20.b\n" "smax z17.s, p2/M, z17.s, z23.s\n" "smax z18.s, p2/M, z18.s, z23.s\n" + "uzp1 z16.h, z16.h, z17.h\n" + "st1b { z31.b }, p1, [x27]\n" "smax z19.s, p2/M, z19.s, z23.s\n" - "uzp1 z31.h, z31.h, z20.h\n" "smax z24.s, p2/M, z24.s, z23.s\n" + "uzp1 z17.h, z18.h, z19.h\n" + "uzp1 z16.b, z16.b, z17.b\n" "smax z25.s, p2/M, z25.s, z23.s\n" - "uzp1 z20.h, z21.h, z22.h\n" "smax z26.s, p2/M, z26.s, z23.s\n" - "smax z27.s, p2/M, z27.s, z23.s\n" - "uzp1 z16.h, z16.h, z17.h\n" - "uzp1 z18.h, z18.h, z19.h\n" "uzp1 z24.h, z24.h, z25.h\n" - "uzp1 z31.b, z31.b, z20.b\n" - "uzp1 z17.h, z26.h, z27.h\n" - "uzp1 z16.b, z16.b, z18.b\n" - "st1b { z31.b }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "uzp1 z24.b, z24.b, z17.b\n" "st1b { z16.b }, p1, [x23]\n" + "smax z27.s, p2/M, z27.s, z23.s\n" + "uzp1 z16.h, z26.h, z27.h\n" + "uzp1 z24.b, z24.b, z16.b\n" "st1b { z24.b }, p1, [x22]\n" + "addvl x27, x27, #1\n" "42:" // Height 3: Writeback done "decw x9, ALL, MUL #4\n" "cmp x9, XZR\n" @@ -956,28 +956,27 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "b 58f\n" "43:" // Height 4 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x4\n" "mov x10, %x[col_bias]\n" "mov z11.s, #0x0\n" "mov z12.s, #0x0\n" + "mov z13.s, #0x0\n" "bic %x[flags], %x[flags], #0x80000000\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" - "mov z13.s, #0x0\n" "mov z14.s, #0x0\n" - "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "madd x20, x21, x20, x27\n" "mov z15.b, #0x1\n" - "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x27, %x[output_ptr]\n" + "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" "44:" // Height 4: Column loop "mov x20, #0x0\n" + "whilelt p1.b, x20, x9\n" "mov z16.s, #0x0\n" "mov z17.s, #0x0\n" "mov z18.s, #0x0\n" "mov z19.s, #0x0\n" "mov z20.s, #0x0\n" "mov z21.s, #0x0\n" - "whilelt p1.b, x20, x9\n" "mov z22.s, #0x0\n" "mov z23.s, #0x0\n" "mov z24.s, #0x0\n" @@ -992,8 +991,8 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov x26, #0x0\n" "46:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 47f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1018,56 +1017,56 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "ble 51f\n" "49:" // Height 4: Multiply loop: Main loop head "whilelt p0.b, XZR, x25\n" - "ld1b { z5.b }, p2/Z, [x28]\n" - "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" - "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" - "ld1rqb { z3.b }, p0/Z, [x22]\n" - "ld1rqb { z6.b }, p0/Z, [x21]\n" - "add x22, x22, #0x10\n" - "add x21, x21, #0x10\n" "trn1 z0.d, z1.d, z2.d\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "ld1rqb { z5.b }, p0/Z, [x21]\n" "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z6.d\n" - "trn2 z3.d, z3.d, z6.d\n" - "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" - ".inst 0x45059810 // smmla z16.s, z0.b, z5.b\n" + "trn1 z2.d, z3.d, z5.d\n" + "ld1b { z4.b }, p2/Z, [x28]\n" + "trn2 z3.d, z3.d, z5.d\n" + ".inst 0x45049810 // smmla z16.s, z0.b, z4.b\n" + ".inst 0x45049858 // smmla z24.s, z2.b, z4.b\n" + "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n" + "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n" ".inst 0x45049814 // smmla z20.s, z0.b, z4.b\n" - ".inst 0x45079811 // smmla z17.s, z0.b, z7.b\n" - ".inst 0x45099815 // smmla z21.s, z0.b, z9.b\n" - ".inst 0x45089812 // smmla z18.s, z0.b, z8.b\n" - ".inst 0x45059858 // smmla z24.s, z2.b, z5.b\n" + ".inst 0x4504985c // smmla z28.s, z2.b, z4.b\n" + "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n" + "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n" + ".inst 0x45059811 // smmla z17.s, z0.b, z5.b\n" + ".inst 0x45059859 // smmla z25.s, z2.b, z5.b\n" + "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n" + "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n" + ".inst 0x45049815 // smmla z21.s, z0.b, z4.b\n" + ".inst 0x4504985d // smmla z29.s, z2.b, z4.b\n" "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n" "addvl x28, x28, #16\n" - ".inst 0x4504985c // smmla z28.s, z2.b, z4.b\n" - ".inst 0x45079859 // smmla z25.s, z2.b, z7.b\n" - ".inst 0x4509985d // smmla z29.s, z2.b, z9.b\n" - ".inst 0x4508985a // smmla z26.s, z2.b, z8.b\n" - ".inst 0x450a9816 // smmla z22.s, z0.b, z10.b\n" "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n" - ".inst 0x450a985e // smmla z30.s, z2.b, z10.b\n" - ".inst 0x45069813 // smmla z19.s, z0.b, z6.b\n" + ".inst 0x45089812 // smmla z18.s, z0.b, z8.b\n" + ".inst 0x4508985a // smmla z26.s, z2.b, z8.b\n" + ".inst 0x45079816 // smmla z22.s, z0.b, z7.b\n" "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n" "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n" - ".inst 0x4506985b // smmla z27.s, z2.b, z6.b\n" - ".inst 0x45059817 // smmla z23.s, z0.b, z5.b\n" + ".inst 0x4507985e // smmla z30.s, z2.b, z7.b\n" + ".inst 0x45069813 // smmla z19.s, z0.b, z6.b\n" "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n" "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n" + ".inst 0x4506985b // smmla z27.s, z2.b, z6.b\n" + ".inst 0x45059817 // smmla z23.s, z0.b, z5.b\n" + "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n" + "add x24, x24, #0x10\n" ".inst 0x4505985f // smmla z31.s, z2.b, z5.b\n" ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" - "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n" "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n" + "add x23, x23, #0x10\n" ".inst 0x45049878 // smmla z24.s, z3.b, z4.b\n" "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n" ".inst 0x450a9834 // smmla z20.s, z1.b, z10.b\n" + "add x22, x22, #0x10\n" ".inst 0x450a987c // smmla z28.s, z3.b, z10.b\n" ".inst 0x45099831 // smmla z17.s, z1.b, z9.b\n" + "add x21, x21, #0x10\n" ".inst 0x45099879 // smmla z25.s, z3.b, z9.b\n" ".inst 0x45089835 // smmla z21.s, z1.b, z8.b\n" ".inst 0x4508987d // smmla z29.s, z3.b, z8.b\n" @@ -1090,60 +1089,60 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bgt 49b\n" "51:" // Height 4: Multiply loop: Single iteration only "whilelt p0.b, XZR, x25\n" - "ld1b { z6.b }, p2/Z, [x28]\n" - "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n" - "subs x25, x25, #0x8\n" - "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n" - "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n" - "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n" - "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1rqb { z1.b }, p0/Z, [x24]\n" "ld1rqb { z2.b }, p0/Z, [x23]\n" + "trn1 z0.d, z1.d, z2.d\n" "ld1rqb { z3.b }, p0/Z, [x22]\n" "ld1rqb { z5.b }, p0/Z, [x21]\n" - "trn1 z0.d, z1.d, z2.d\n" "trn2 z1.d, z1.d, z2.d\n" "trn1 z2.d, z3.d, z5.d\n" + "ld1b { z4.b }, p2/Z, [x28]\n" "trn2 z3.d, z3.d, z5.d\n" + ".inst 0x45049810 // smmla z16.s, z0.b, z4.b\n" + ".inst 0x45049858 // smmla z24.s, z2.b, z4.b\n" + "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n" + "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n" + "subs x25, x25, #0x8\n" + ".inst 0x45059814 // smmla z20.s, z0.b, z5.b\n" + "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n" + "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n" + ".inst 0x4505985c // smmla z28.s, z2.b, z5.b\n" + ".inst 0x45049811 // smmla z17.s, z0.b, z4.b\n" + "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n" "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n" - ".inst 0x45069810 // smmla z16.s, z0.b, z6.b\n" - ".inst 0x45049814 // smmla z20.s, z0.b, z4.b\n" - ".inst 0x45079811 // smmla z17.s, z0.b, z7.b\n" - ".inst 0x45099815 // smmla z21.s, z0.b, z9.b\n" - ".inst 0x45089812 // smmla z18.s, z0.b, z8.b\n" - ".inst 0x45069858 // smmla z24.s, z2.b, z6.b\n" - "ld1b { z6.b }, p2/Z, [x28, #7, MUL VL]\n" - ".inst 0x4504985c // smmla z28.s, z2.b, z4.b\n" + ".inst 0x45049859 // smmla z25.s, z2.b, z4.b\n" + ".inst 0x45089815 // smmla z21.s, z0.b, z8.b\n" + "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n" + ".inst 0x4508985d // smmla z29.s, z2.b, z8.b\n" + ".inst 0x45079812 // smmla z18.s, z0.b, z7.b\n" "addvl x28, x28, #8\n" - ".inst 0x45079859 // smmla z25.s, z2.b, z7.b\n" - ".inst 0x4509985d // smmla z29.s, z2.b, z9.b\n" - ".inst 0x4508985a // smmla z26.s, z2.b, z8.b\n" - ".inst 0x450a9816 // smmla z22.s, z0.b, z10.b\n" - ".inst 0x450a985e // smmla z30.s, z2.b, z10.b\n" + ".inst 0x4507985a // smmla z26.s, z2.b, z7.b\n" + ".inst 0x45069816 // smmla z22.s, z0.b, z6.b\n" + ".inst 0x4506985e // smmla z30.s, z2.b, z6.b\n" ".inst 0x45059813 // smmla z19.s, z0.b, z5.b\n" ".inst 0x4505985b // smmla z27.s, z2.b, z5.b\n" - ".inst 0x45069817 // smmla z23.s, z0.b, z6.b\n" - ".inst 0x4506985f // smmla z31.s, z2.b, z6.b\n" + ".inst 0x45049817 // smmla z23.s, z0.b, z4.b\n" + ".inst 0x4504985f // smmla z31.s, z2.b, z4.b\n" "ble 52f\n" "ld1b { z4.b }, p2/Z, [x28]\n" - "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n" - "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n" + ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" + ".inst 0x45049878 // smmla z24.s, z3.b, z4.b\n" + "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n" + "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n" "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n" + ".inst 0x45059834 // smmla z20.s, z1.b, z5.b\n" + ".inst 0x4505987c // smmla z28.s, z3.b, z5.b\n" "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n" "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n" - ".inst 0x45049830 // smmla z16.s, z1.b, z4.b\n" - ".inst 0x45049878 // smmla z24.s, z3.b, z4.b\n" + ".inst 0x45049831 // smmla z17.s, z1.b, z4.b\n" + ".inst 0x45049879 // smmla z25.s, z3.b, z4.b\n" "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n" "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n" - ".inst 0x450a9834 // smmla z20.s, z1.b, z10.b\n" - ".inst 0x450a987c // smmla z28.s, z3.b, z10.b\n" - ".inst 0x45099831 // smmla z17.s, z1.b, z9.b\n" - ".inst 0x45099879 // smmla z25.s, z3.b, z9.b\n" - "addvl x28, x28, #8\n" ".inst 0x45089835 // smmla z21.s, z1.b, z8.b\n" ".inst 0x4508987d // smmla z29.s, z3.b, z8.b\n" ".inst 0x45079832 // smmla z18.s, z1.b, z7.b\n" ".inst 0x4507987a // smmla z26.s, z3.b, z7.b\n" + "addvl x28, x28, #8\n" ".inst 0x45069836 // smmla z22.s, z1.b, z6.b\n" ".inst 0x4506987e // smmla z30.s, z3.b, z6.b\n" ".inst 0x45059833 // smmla z19.s, z1.b, z5.b\n" @@ -1163,16 +1162,16 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "bne 46b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 z0.d, z16.d, z20.d\n" + "add x23, x27, x20\n" + "add x22, x23, x20\n" "uzp2 z16.d, z16.d, z20.d\n" "uzp1 z20.d, z17.d, z21.d\n" + "add x21, x22, x20\n" "uzp2 z17.d, z17.d, z21.d\n" "uzp1 z21.d, z18.d, z22.d\n" "uzp2 z18.d, z18.d, z22.d\n" - "add x23, x27, x20\n" - "add x22, x23, x20\n" "uzp1 z22.d, z19.d, z23.d\n" "uzp2 z19.d, z19.d, z23.d\n" - "add x21, x22, x20\n" "uzp1 z23.d, z24.d, z28.d\n" "uzp2 z24.d, z24.d, z28.d\n" "uzp1 z28.d, z25.d, z29.d\n" @@ -1184,15 +1183,15 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "mov z31.d, z0.d\n" "tbnz %x[flags], #31, 54f\n" "add x20, %x[qp], %[b_offset]\n" + "ld1rw { z0.s }, p2/Z, [x20]\n" ".inst 0x4491a96b // addp z11.s, p2/m, z11.s, z11.s\n" ".inst 0x4491a9ad // addp z13.s, p2/m, z13.s, z13.s\n" - "ld1rw { z0.s }, p2/Z, [x20]\n" "neg z0.s, p2/M, z0.s\n" "mov z12.s, z11.s[3]\n" "mov z11.s, z11.s[0]\n" + "mul z11.s, p2/M, z11.s, z0.s\n" "mov z14.s, z13.s[3]\n" "mov z13.s, z13.s[0]\n" - "mul z11.s, p2/M, z11.s, z0.s\n" "mul z12.s, p2/M, z12.s, z0.s\n" "mul z13.s, p2/M, z13.s, z0.s\n" "mul z14.s, p2/M, z14.s, z0.s\n" @@ -1258,32 +1257,32 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "tbz %x[flags], #5, 55f\n" "and z2.d, z31.d, z0.d\n" "and z1.d, z20.d, z0.d\n" + "asr z2.s, z2.s, #0x1f\n" + "asr z1.s, z1.s, #0x1f\n" + "sqadd z31.s, z31.s, z2.s\n" + "sqadd z20.s, z20.s, z1.s\n" "and z7.d, z21.d, z0.d\n" "and z6.d, z22.d, z0.d\n" "and z5.d, z16.d, z0.d\n" "and z4.d, z17.d, z0.d\n" - "asr z2.s, z2.s, #0x1f\n" - "asr z1.s, z1.s, #0x1f\n" "and z3.d, z18.d, z0.d\n" + "and z2.d, z19.d, z0.d\n" + "and z1.d, z23.d, z0.d\n" "asr z7.s, z7.s, #0x1f\n" "asr z6.s, z6.s, #0x1f\n" "asr z5.s, z5.s, #0x1f\n" - "sqadd z31.s, z31.s, z2.s\n" - "sqadd z20.s, z20.s, z1.s\n" - "and z2.d, z19.d, z0.d\n" - "and z1.d, z23.d, z0.d\n" "asr z4.s, z4.s, #0x1f\n" "asr z3.s, z3.s, #0x1f\n" - "sqadd z21.s, z21.s, z7.s\n" - "sqadd z22.s, z22.s, z6.s\n" "asr z2.s, z2.s, #0x1f\n" "asr z1.s, z1.s, #0x1f\n" + "sqadd z21.s, z21.s, z7.s\n" + "sqadd z22.s, z22.s, z6.s\n" "sqadd z16.s, z16.s, z5.s\n" "sqadd z17.s, z17.s, z4.s\n" "sqadd z18.s, z18.s, z3.s\n" - "and z7.d, z28.d, z0.d\n" "sqadd z19.s, z19.s, z2.s\n" "sqadd z23.s, z23.s, z1.s\n" + "and z7.d, z28.d, z0.d\n" "and z6.d, z29.d, z0.d\n" "and z5.d, z30.d, z0.d\n" "and z4.d, z24.d, z0.d\n" @@ -1306,43 +1305,43 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "sqadd z27.s, z27.s, z1.s\n" "55:" // Height 4: no shift correction "add x20, %x[qp], %[c_offset]\n" - ".inst 0x4482881f // srshl z31.s, p2/M, z31.s, z0.s\n" "ld1rw { z2.s }, p2/Z, [x20]\n" + ".inst 0x4482881f // srshl z31.s, p2/M, z31.s, z0.s\n" + "add z31.s, z31.s, z2.s\n" ".inst 0x44828814 // srshl z20.s, p2/M, z20.s, z0.s\n" ".inst 0x44828815 // srshl z21.s, p2/M, z21.s, z0.s\n" + "add z20.s, z20.s, z2.s\n" + "add z21.s, z21.s, z2.s\n" ".inst 0x44828816 // srshl z22.s, p2/M, z22.s, z0.s\n" ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n" + "add z22.s, z22.s, z2.s\n" + "add z16.s, z16.s, z2.s\n" ".inst 0x44828811 // srshl z17.s, p2/M, z17.s, z0.s\n" ".inst 0x44828812 // srshl z18.s, p2/M, z18.s, z0.s\n" - "add z31.s, z31.s, z2.s\n" + "add z17.s, z17.s, z2.s\n" + "add z18.s, z18.s, z2.s\n" ".inst 0x44828813 // srshl z19.s, p2/M, z19.s, z0.s\n" ".inst 0x44828817 // srshl z23.s, p2/M, z23.s, z0.s\n" - "add z20.s, z20.s, z2.s\n" - "add z21.s, z21.s, z2.s\n" + "add z19.s, z19.s, z2.s\n" + "add z23.s, z23.s, z2.s\n" ".inst 0x4482881c // srshl z28.s, p2/M, z28.s, z0.s\n" ".inst 0x4482881d // srshl z29.s, p2/M, z29.s, z0.s\n" - "add z22.s, z22.s, z2.s\n" - "add z16.s, z16.s, z2.s\n" + "add z28.s, z28.s, z2.s\n" + "add z29.s, z29.s, z2.s\n" ".inst 0x4482881e // srshl z30.s, p2/M, z30.s, z0.s\n" ".inst 0x44828818 // srshl z24.s, p2/M, z24.s, z0.s\n" - "add z17.s, z17.s, z2.s\n" - "add z18.s, z18.s, z2.s\n" + "add z30.s, z30.s, z2.s\n" + "add z24.s, z24.s, z2.s\n" ".inst 0x44828819 // srshl z25.s, p2/M, z25.s, z0.s\n" ".inst 0x4482881a // srshl z26.s, p2/M, z26.s, z0.s\n" - "add z19.s, z19.s, z2.s\n" - "add z23.s, z23.s, z2.s\n" + "add z25.s, z25.s, z2.s\n" + "add z26.s, z26.s, z2.s\n" ".inst 0x4482881b // srshl z27.s, p2/M, z27.s, z0.s\n" "add x20, %x[qp], %[maxval]\n" - "add z28.s, z28.s, z2.s\n" - "add z29.s, z29.s, z2.s\n" "ld1rw { z1.s }, p2/Z, [x20]\n" - "add z30.s, z30.s, z2.s\n" - "add z24.s, z24.s, z2.s\n" + "add z27.s, z27.s, z2.s\n" "add x20, %x[qp], %[minval]\n" - "add z25.s, z25.s, z2.s\n" - "add z26.s, z26.s, z2.s\n" "ld1rw { z0.s }, p2/Z, [x20]\n" - "add z27.s, z27.s, z2.s\n" "smin z31.s, p2/M, z31.s, z1.s\n" "smin z20.s, p2/M, z20.s, z1.s\n" "smin z21.s, p2/M, z21.s, z1.s\n" @@ -1362,36 +1361,36 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "smax z31.s, p2/M, z31.s, z0.s\n" "smax z20.s, p2/M, z20.s, z0.s\n" "smax z21.s, p2/M, z21.s, z0.s\n" + "uzp1 z31.h, z31.h, z20.h\n" "smax z22.s, p2/M, z22.s, z0.s\n" "smax z16.s, p2/M, z16.s, z0.s\n" + "uzp1 z20.h, z21.h, z22.h\n" + "uzp1 z31.b, z31.b, z20.b\n" "smax z17.s, p2/M, z17.s, z0.s\n" "smax z18.s, p2/M, z18.s, z0.s\n" + "uzp1 z16.h, z16.h, z17.h\n" + "st1b { z31.b }, p1, [x27]\n" "smax z19.s, p2/M, z19.s, z0.s\n" - "uzp1 z31.h, z31.h, z20.h\n" "smax z23.s, p2/M, z23.s, z0.s\n" + "uzp1 z17.h, z18.h, z19.h\n" + "uzp1 z16.b, z16.b, z17.b\n" "smax z28.s, p2/M, z28.s, z0.s\n" - "uzp1 z20.h, z21.h, z22.h\n" "smax z29.s, p2/M, z29.s, z0.s\n" + "uzp1 z23.h, z23.h, z28.h\n" + "st1b { z16.b }, p1, [x23]\n" "smax z30.s, p2/M, z30.s, z0.s\n" - "uzp1 z16.h, z16.h, z17.h\n" "smax z24.s, p2/M, z24.s, z0.s\n" + "uzp1 z16.h, z29.h, z30.h\n" + "uzp1 z23.b, z23.b, z16.b\n" "smax z25.s, p2/M, z25.s, z0.s\n" - "uzp1 z17.h, z18.h, z19.h\n" "smax z26.s, p2/M, z26.s, z0.s\n" - "smax z27.s, p2/M, z27.s, z0.s\n" - "uzp1 z23.h, z23.h, z28.h\n" - "uzp1 z31.b, z31.b, z20.b\n" - "uzp1 z18.h, z29.h, z30.h\n" "uzp1 z24.h, z24.h, z25.h\n" - "uzp1 z16.b, z16.b, z17.b\n" - "uzp1 z17.h, z26.h, z27.h\n" - "st1b { z31.b }, p1, [x27]\n" - "addvl x27, x27, #1\n" - "uzp1 z23.b, z23.b, z18.b\n" - "uzp1 z24.b, z24.b, z17.b\n" - "st1b { z16.b }, p1, [x23]\n" "st1b { z23.b }, p1, [x22]\n" + "smax z27.s, p2/M, z27.s, z0.s\n" + "uzp1 z16.h, z26.h, z27.h\n" + "uzp1 z24.b, z24.b, z16.b\n" "st1b { z24.b }, p1, [x21]\n" + "addvl x27, x27, #1\n" "56:" // Height 4: Writeback done "decw x9, ALL, MUL #4\n" "cmp x9, XZR\n" @@ -1408,8 +1407,8 @@ void sve_hybrid_s8qa_mmla_4x4VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "58:" // Exit - : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr) - : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) + : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) + : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |