From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp | 1612 ++++++++++---------- 1 file changed, 805 insertions(+), 807 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp index 31fbf88603..ebe583b5d4 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp @@ -78,7 +78,6 @@ void a64_hybrid_u8qa_dot_4x16 ( flags |= 0x20; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x4\n" "bge 91f\n" @@ -102,11 +101,11 @@ void a64_hybrid_u8qa_dot_4x16 ( "4:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 5f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" "cbnz x26, 6f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20\n" @@ -128,32 +127,32 @@ void a64_hybrid_u8qa_dot_4x16 ( "blt 9f\n" "7:" // Height 1: Multiply loop: Main loop head ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - "ldr q4, [x28, #0x70]\n" + "ldr q21, [x28, #0x70]\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" - "ldr q5, [x28, #0x80]\n" + "ldr q20, [x28, #0x80]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x28, #0x90]\n" + "ldr q26, [x28, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x28, #0xa0]\n" + "ldr q25, [x28, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" - "ldr q8, [x28, #0xb0]\n" + "ldr q24, [x28, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" - "ldr q9, [x28, #0xc0]\n" + "ldr q23, [x28, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" - "ldr q10, [x28, #0xd0]\n" - ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - "ldr q4, [x28, #0xe0]\n" - ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - "ldr q5, [x28, #0xf0]\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "ldr q22, [x28, #0xd0]\n" + ".inst 0x6fa0e2b3 // udot v19.4s, v21.16b, v0.4b[1]\n" + "ldr q21, [x28, #0xe0]\n" + ".inst 0x6f80ea90 // udot v16.4s, v20.16b, v0.4b[2]\n" + "ldr q20, [x28, #0xf0]\n" + ".inst 0x6f80eb51 // udot v17.4s, v26.16b, v0.4b[2]\n" "add x24, x24, #0x10\n" - ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80eb32 // udot v18.4s, v25.16b, v0.4b[2]\n" + ".inst 0x6f80eb13 // udot v19.4s, v24.16b, v0.4b[2]\n" "add x28, x28, #0x100\n" - ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" + ".inst 0x6fa0eaf0 // udot v16.4s, v23.16b, v0.4b[3]\n" + ".inst 0x6fa0ead1 // udot v17.4s, v22.16b, v0.4b[3]\n" + ".inst 0x6fa0eab2 // udot v18.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa0ea93 // udot v19.4s, v20.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 8f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" "8:" // Height 1: Multiply loop: unique 1: skip row sum @@ -171,33 +170,33 @@ void a64_hybrid_u8qa_dot_4x16 ( "bge 7b\n" "9:" // Height 1: Multiply loop: Single iteration only ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - "ldr q4, [x28, #0x70]\n" + "ldr q21, [x28, #0x70]\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" - "ldr q5, [x28, #0x80]\n" + "ldr q20, [x28, #0x80]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x28, #0x90]\n" + "ldr q26, [x28, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x28, #0xa0]\n" + "ldr q25, [x28, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" - "ldr q8, [x28, #0xb0]\n" + "ldr q24, [x28, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" - "ldr q9, [x28, #0xc0]\n" + "ldr q23, [x28, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" - "ldr q10, [x28, #0xd0]\n" - ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - "ldr q4, [x28, #0xe0]\n" - ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - "ldr q5, [x28, #0xf0]\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "ldr q22, [x28, #0xd0]\n" + ".inst 0x6fa0e2b3 // udot v19.4s, v21.16b, v0.4b[1]\n" + "ldr q21, [x28, #0xe0]\n" + ".inst 0x6f80ea90 // udot v16.4s, v20.16b, v0.4b[2]\n" + "ldr q20, [x28, #0xf0]\n" + ".inst 0x6f80eb51 // udot v17.4s, v26.16b, v0.4b[2]\n" "sub x25, x25, #0x10\n" - ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" + ".inst 0x6f80eb32 // udot v18.4s, v25.16b, v0.4b[2]\n" + ".inst 0x6f80eb13 // udot v19.4s, v24.16b, v0.4b[2]\n" "add x24, x24, #0x10\n" "add x28, x28, #0x100\n" - ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" + ".inst 0x6fa0eaf0 // udot v16.4s, v23.16b, v0.4b[3]\n" + ".inst 0x6fa0ead1 // udot v17.4s, v22.16b, v0.4b[3]\n" + ".inst 0x6fa0eab2 // udot v18.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa0ea93 // udot v19.4s, v20.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 10f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" "10:" // Height 1: Multiply loop: unique 2: skip row sum @@ -211,16 +210,16 @@ void a64_hybrid_u8qa_dot_4x16 ( "tbnz %x[flags], #31, 13f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" "13:" // Height 1: Multiply loop: unique 3: skip row sum - "ldr q6, [x28, #0x0]\n" - "ldr q7, [x28, #0x10]\n" + "ldr q23, [x28, #0x0]\n" + "ldr q22, [x28, #0x10]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x4\n" - "ldr q8, [x28, #0x20]\n" - "ldr q9, [x28, #0x30]\n" - ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" - ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" + "ldr q21, [x28, #0x20]\n" + "ldr q20, [x28, #0x30]\n" + ".inst 0x6f80e2f0 // udot v16.4s, v23.16b, v0.4b[0]\n" + ".inst 0x6f80e2d1 // udot v17.4s, v22.16b, v0.4b[0]\n" + ".inst 0x6f80e2b2 // udot v18.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f80e293 // udot v19.4s, v20.16b, v0.4b[0]\n" "add x28, x28, #0x40\n" "bge 12b\n" "14:" // Height 1: Multiply loop: Skip odd blocks @@ -236,14 +235,14 @@ void a64_hybrid_u8qa_dot_4x16 ( "tbnz %x[flags], #31, 17f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" "17:" // Height 1: Multiply loop: unique 4: skip row sum - "ldr q10, [x28, #0x0]\n" - "ldr q4, [x28, #0x10]\n" - ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" - ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" - "ldr q5, [x28, #0x20]\n" - "ldr q6, [x28, #0x30]\n" - ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" - ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" + "ldr q21, [x28, #0x0]\n" + "ldr q20, [x28, #0x10]\n" + ".inst 0x6f80e2b0 // udot v16.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f80e291 // udot v17.4s, v20.16b, v0.4b[0]\n" + "ldr q21, [x28, #0x20]\n" + "ldr q20, [x28, #0x30]\n" + ".inst 0x6f80e2b2 // udot v18.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f80e293 // udot v19.4s, v20.16b, v0.4b[0]\n" "add x28, x28, #0x40\n" "18:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -252,72 +251,72 @@ void a64_hybrid_u8qa_dot_4x16 ( "bne 4b\n" "prfm pstl1keep, [x27, #0x0]\n" "tbnz %x[flags], #31, 19f\n" - "add x23, %x[qp], %[b_offset]\n" - "ld1r { v1.4s }, [x23]\n" + "add x20, %x[qp], %[b_offset]\n" + "ld1r { v20.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" - "neg v1.4s, v1.4s\n" + "neg v20.4s, v20.4s\n" "addp v11.4s, v11.4s, v11.4s\n" - "mul v11.4s, v11.4s, v1.4s\n" + "mul v11.4s, v11.4s, v20.4s\n" "19:" // Height 1: skip row sum fixup - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x10, #0x10]\n" + "ldr q24, [x10, #0x0]\n" + "ldr q23, [x10, #0x10]\n" "add v16.4s, v16.4s, v11.4s\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q2, [x10, #0x20]\n" - "ldr q3, [x10, #0x30]\n" + "ldr q22, [x10, #0x20]\n" + "ldr q21, [x10, #0x30]\n" "add v18.4s, v18.4s, v11.4s\n" "add v19.4s, v19.4s, v11.4s\n" - "add x23, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v20.4s }, [x20]\n" "orr %x[flags], %x[flags], #0x80000000\n" - "add v16.4s, v16.4s, v0.4s\n" - "add v17.4s, v17.4s, v1.4s\n" - "add v18.4s, v18.4s, v2.4s\n" - "add x23, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x23]\n" - "add v19.4s, v19.4s, v3.4s\n" - "sqrdmulh v16.4s, v16.4s, v4.4s\n" + "add v16.4s, v16.4s, v24.4s\n" + "add v17.4s, v17.4s, v23.4s\n" + "add v18.4s, v18.4s, v22.4s\n" + "add x20, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x20]\n" + "add v19.4s, v19.4s, v21.4s\n" + "sqrdmulh v16.4s, v16.4s, v20.4s\n" "add x10, x10, #0x40\n" - "sqrdmulh v17.4s, v17.4s, v4.4s\n" - "sqrdmulh v18.4s, v18.4s, v4.4s\n" - "sqrdmulh v19.4s, v19.4s, v4.4s\n" + "sqrdmulh v17.4s, v17.4s, v20.4s\n" + "sqrdmulh v18.4s, v18.4s, v20.4s\n" + "sqrdmulh v19.4s, v19.4s, v20.4s\n" "tbz %x[flags], #5, 20f\n" - "and v4.16b, v16.16b, v0.16b\n" - "and v5.16b, v17.16b, v0.16b\n" - "and v6.16b, v18.16b, v0.16b\n" - "and v7.16b, v19.16b, v0.16b\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sshr v5.4s, v5.4s, #0x1f\n" - "sshr v6.4s, v6.4s, #0x1f\n" - "sshr v7.4s, v7.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v4.4s\n" - "sqadd v17.4s, v17.4s, v5.4s\n" - "sqadd v18.4s, v18.4s, v6.4s\n" - "sqadd v19.4s, v19.4s, v7.4s\n" + "and v23.16b, v16.16b, v0.16b\n" + "and v22.16b, v17.16b, v0.16b\n" + "and v21.16b, v18.16b, v0.16b\n" + "and v20.16b, v19.16b, v0.16b\n" + "sshr v23.4s, v23.4s, #0x1f\n" + "sshr v22.4s, v22.4s, #0x1f\n" + "sshr v21.4s, v21.4s, #0x1f\n" + "sshr v20.4s, v20.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v23.4s\n" + "sqadd v17.4s, v17.4s, v22.4s\n" + "sqadd v18.4s, v18.4s, v21.4s\n" + "sqadd v19.4s, v19.4s, v20.4s\n" "20:" // Height 1: no shift correction - "add x23, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[c_offset]\n" + "ld1r { v22.4s }, [x20]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" "srshl v18.4s, v18.4s, v0.4s\n" "srshl v19.4s, v19.4s, v0.4s\n" - "add x23, %x[qp], %[maxval]\n" - "ld1r { v6.4s }, [x23]\n" - "add v16.4s, v16.4s, v4.4s\n" - "add v17.4s, v17.4s, v4.4s\n" - "add x23, %x[qp], %[minval]\n" - "ld1r { v5.4s }, [x23]\n" - "add v18.4s, v18.4s, v4.4s\n" - "add v19.4s, v19.4s, v4.4s\n" + "add x20, %x[qp], %[maxval]\n" + "ld1r { v21.4s }, [x20]\n" + "add v16.4s, v16.4s, v22.4s\n" + "add v17.4s, v17.4s, v22.4s\n" + "add x20, %x[qp], %[minval]\n" + "ld1r { v20.4s }, [x20]\n" + "add v18.4s, v18.4s, v22.4s\n" + "add v19.4s, v19.4s, v22.4s\n" "cmp x9, #0x10\n" - "smin v16.4s, v16.4s, v6.4s\n" - "smin v17.4s, v17.4s, v6.4s\n" - "smin v18.4s, v18.4s, v6.4s\n" - "smin v19.4s, v19.4s, v6.4s\n" - "smax v16.4s, v16.4s, v5.4s\n" - "smax v17.4s, v17.4s, v5.4s\n" - "smax v18.4s, v18.4s, v5.4s\n" - "smax v19.4s, v19.4s, v5.4s\n" + "smin v16.4s, v16.4s, v21.4s\n" + "smin v17.4s, v17.4s, v21.4s\n" + "smin v18.4s, v18.4s, v21.4s\n" + "smin v19.4s, v19.4s, v21.4s\n" + "smax v16.4s, v16.4s, v20.4s\n" + "smax v17.4s, v17.4s, v20.4s\n" + "smax v18.4s, v18.4s, v20.4s\n" + "smax v19.4s, v19.4s, v20.4s\n" "uzp1 v16.8h, v16.8h, v17.8h\n" "uzp1 v17.8h, v18.8h, v19.8h\n" "uzp1 v16.16b, v16.16b, v17.16b\n" @@ -397,12 +396,12 @@ void a64_hybrid_u8qa_dot_4x16 ( "34:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 35f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" "cbnz x26, 36f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20\n" @@ -410,7 +409,7 @@ void a64_hybrid_u8qa_dot_4x16 ( "b 36f\n" "35:" // Height 2: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20\n" + "add x23, x24, x21\n" "36:" // Height 2: input setup done "cmp x25, #0x10\n" "blt 41f\n" @@ -428,48 +427,48 @@ void a64_hybrid_u8qa_dot_4x16 ( "37:" // Height 2: Multiply loop: Main loop head ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" - "ldr q4, [x28, #0x70]\n" + "ldr q25, [x28, #0x70]\n" "add x24, x24, #0x10\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" - "ldr q5, [x28, #0x80]\n" + "ldr q24, [x28, #0x80]\n" "add x23, x23, #0x10\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x28, #0x90]\n" + "ldr q30, [x28, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x28, #0xa0]\n" + "ldr q29, [x28, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" - "ldr q8, [x28, #0xb0]\n" + "ldr q28, [x28, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" - "ldr q9, [x28, #0xc0]\n" + "ldr q27, [x28, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" - "ldr q10, [x28, #0xd0]\n" - ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" - "ldr q4, [x28, #0xe0]\n" - ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" - "ldr q5, [x28, #0xf0]\n" + "ldr q26, [x28, #0xd0]\n" + ".inst 0x6fa0e333 // udot v19.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa1e337 // udot v23.4s, v25.16b, v1.4b[1]\n" + "ldr q25, [x28, #0xe0]\n" + ".inst 0x6f80eb10 // udot v16.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f81eb14 // udot v20.4s, v24.16b, v1.4b[2]\n" + "ldr q24, [x28, #0xf0]\n" "add x28, x28, #0x100\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" - ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" - ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" - ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" - ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" + ".inst 0x6f80ebd1 // udot v17.4s, v30.16b, v0.4b[2]\n" + ".inst 0x6f81ebd5 // udot v21.4s, v30.16b, v1.4b[2]\n" + ".inst 0x6f80ebb2 // udot v18.4s, v29.16b, v0.4b[2]\n" + ".inst 0x6f81ebb6 // udot v22.4s, v29.16b, v1.4b[2]\n" + ".inst 0x6f80eb93 // udot v19.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb97 // udot v23.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6fa0eb70 // udot v16.4s, v27.16b, v0.4b[3]\n" + ".inst 0x6fa1eb74 // udot v20.4s, v27.16b, v1.4b[3]\n" + ".inst 0x6fa0eb51 // udot v17.4s, v26.16b, v0.4b[3]\n" + ".inst 0x6fa1eb55 // udot v21.4s, v26.16b, v1.4b[3]\n" + ".inst 0x6fa0eb32 // udot v18.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa1eb36 // udot v22.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa0eb13 // udot v19.4s, v24.16b, v0.4b[3]\n" + ".inst 0x6fa1eb17 // udot v23.4s, v24.16b, v1.4b[3]\n" "tbnz %x[flags], #31, 38f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" @@ -491,49 +490,49 @@ void a64_hybrid_u8qa_dot_4x16 ( "39:" // Height 2: Multiply loop: Single iteration only ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" - "ldr q4, [x28, #0x70]\n" + "ldr q25, [x28, #0x70]\n" "sub x25, x25, #0x10\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" - "ldr q5, [x28, #0x80]\n" + "ldr q24, [x28, #0x80]\n" "add x24, x24, #0x10\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x28, #0x90]\n" + "ldr q30, [x28, #0x90]\n" "add x23, x23, #0x10\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x28, #0xa0]\n" + "ldr q29, [x28, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" - "ldr q8, [x28, #0xb0]\n" + "ldr q28, [x28, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" - "ldr q9, [x28, #0xc0]\n" + "ldr q27, [x28, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" - "ldr q10, [x28, #0xd0]\n" - ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" - "ldr q4, [x28, #0xe0]\n" - ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" - "ldr q5, [x28, #0xf0]\n" + "ldr q26, [x28, #0xd0]\n" + ".inst 0x6fa0e333 // udot v19.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa1e337 // udot v23.4s, v25.16b, v1.4b[1]\n" + "ldr q25, [x28, #0xe0]\n" + ".inst 0x6f80eb10 // udot v16.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f81eb14 // udot v20.4s, v24.16b, v1.4b[2]\n" + "ldr q24, [x28, #0xf0]\n" "add x28, x28, #0x100\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" - ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" - ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" - ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" - ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" + ".inst 0x6f80ebd1 // udot v17.4s, v30.16b, v0.4b[2]\n" + ".inst 0x6f81ebd5 // udot v21.4s, v30.16b, v1.4b[2]\n" + ".inst 0x6f80ebb2 // udot v18.4s, v29.16b, v0.4b[2]\n" + ".inst 0x6f81ebb6 // udot v22.4s, v29.16b, v1.4b[2]\n" + ".inst 0x6f80eb93 // udot v19.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb97 // udot v23.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6fa0eb70 // udot v16.4s, v27.16b, v0.4b[3]\n" + ".inst 0x6fa1eb74 // udot v20.4s, v27.16b, v1.4b[3]\n" + ".inst 0x6fa0eb51 // udot v17.4s, v26.16b, v0.4b[3]\n" + ".inst 0x6fa1eb55 // udot v21.4s, v26.16b, v1.4b[3]\n" + ".inst 0x6fa0eb32 // udot v18.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa1eb36 // udot v22.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa0eb13 // udot v19.4s, v24.16b, v0.4b[3]\n" + ".inst 0x6fa1eb17 // udot v23.4s, v24.16b, v1.4b[3]\n" "tbnz %x[flags], #31, 40f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" @@ -551,21 +550,21 @@ void a64_hybrid_u8qa_dot_4x16 ( ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" "43:" // Height 2: Multiply loop: unique 7: skip row sum - "ldr q6, [x28, #0x0]\n" - "ldr q7, [x28, #0x10]\n" + "ldr q27, [x28, #0x0]\n" + "ldr q26, [x28, #0x10]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x4\n" - "ldr q8, [x28, #0x20]\n" - "ldr q9, [x28, #0x30]\n" - ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" + "ldr q25, [x28, #0x20]\n" + "ldr q24, [x28, #0x30]\n" + ".inst 0x6f80e370 // udot v16.4s, v27.16b, v0.4b[0]\n" + ".inst 0x6f81e374 // udot v20.4s, v27.16b, v1.4b[0]\n" + ".inst 0x6f80e351 // udot v17.4s, v26.16b, v0.4b[0]\n" + ".inst 0x6f81e355 // udot v21.4s, v26.16b, v1.4b[0]\n" "add x28, x28, #0x40\n" - ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" - ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" - ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" - ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" + ".inst 0x6f80e332 // udot v18.4s, v25.16b, v0.4b[0]\n" + ".inst 0x6f81e336 // udot v22.4s, v25.16b, v1.4b[0]\n" + ".inst 0x6f80e313 // udot v19.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e317 // udot v23.4s, v24.16b, v1.4b[0]\n" "bge 42b\n" "44:" // Height 2: Multiply loop: Skip odd blocks "cbz x25, 48f\n" @@ -584,209 +583,209 @@ void a64_hybrid_u8qa_dot_4x16 ( ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" "47:" // Height 2: Multiply loop: unique 8: skip row sum - "ldr q10, [x28, #0x0]\n" - "ldr q4, [x28, #0x10]\n" - ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" - ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" - "ldr q5, [x28, #0x20]\n" - "ldr q6, [x28, #0x30]\n" - ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" - ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" - ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" + "ldr q24, [x28, #0x0]\n" + "ldr q26, [x28, #0x10]\n" + ".inst 0x6f80e310 // udot v16.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e314 // udot v20.4s, v24.16b, v1.4b[0]\n" + "ldr q25, [x28, #0x20]\n" + "ldr q24, [x28, #0x30]\n" + ".inst 0x6f80e351 // udot v17.4s, v26.16b, v0.4b[0]\n" + ".inst 0x6f81e355 // udot v21.4s, v26.16b, v1.4b[0]\n" + ".inst 0x6f80e332 // udot v18.4s, v25.16b, v0.4b[0]\n" + ".inst 0x6f81e336 // udot v22.4s, v25.16b, v1.4b[0]\n" "add x28, x28, #0x40\n" - ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f80e313 // udot v19.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e317 // udot v23.4s, v24.16b, v1.4b[0]\n" "48:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" "cmp x26, x20\n" "bne 34b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x22, x27, x20\n" + "add x23, x27, x20\n" "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" "tbnz %x[flags], #31, 49f\n" - "add x23, %x[qp], %[b_offset]\n" - "ld1r { v2.4s }, [x23]\n" + "add x20, %x[qp], %[b_offset]\n" + "ld1r { v24.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" - "neg v2.4s, v2.4s\n" + "neg v24.4s, v24.4s\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" - "mul v11.4s, v11.4s, v2.4s\n" - "mul v12.4s, v12.4s, v2.4s\n" + "mul v11.4s, v11.4s, v24.4s\n" + "mul v12.4s, v12.4s, v24.4s\n" "49:" // Height 2: skip row sum fixup - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x10, #0x10]\n" + "ldr q28, [x10, #0x0]\n" + "ldr q27, [x10, #0x10]\n" "add v16.4s, v16.4s, v11.4s\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q2, [x10, #0x20]\n" - "ldr q3, [x10, #0x30]\n" + "ldr q26, [x10, #0x20]\n" + "ldr q25, [x10, #0x30]\n" "add v18.4s, v18.4s, v11.4s\n" "add v19.4s, v19.4s, v11.4s\n" "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" - "add x23, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v24.4s }, [x20]\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" "orr %x[flags], %x[flags], #0x80000000\n" - "add x23, %x[qp], %[per_layer_right_shift]\n" - "add v16.4s, v16.4s, v0.4s\n" - "add v17.4s, v17.4s, v1.4s\n" + "add x20, %x[qp], %[per_layer_right_shift]\n" + "add v16.4s, v16.4s, v28.4s\n" + "add v17.4s, v17.4s, v27.4s\n" "add x10, x10, #0x40\n" - "add v18.4s, v18.4s, v2.4s\n" - "add v19.4s, v19.4s, v3.4s\n" - "add v20.4s, v20.4s, v0.4s\n" - "ld1r { v0.4s }, [x23]\n" - "add v21.4s, v21.4s, v1.4s\n" - "add v22.4s, v22.4s, v2.4s\n" - "add v23.4s, v23.4s, v3.4s\n" - "sqrdmulh v16.4s, v16.4s, v4.4s\n" - "sqrdmulh v17.4s, v17.4s, v4.4s\n" - "sqrdmulh v18.4s, v18.4s, v4.4s\n" - "sqrdmulh v19.4s, v19.4s, v4.4s\n" - "sqrdmulh v20.4s, v20.4s, v4.4s\n" - "sqrdmulh v21.4s, v21.4s, v4.4s\n" - "sqrdmulh v22.4s, v22.4s, v4.4s\n" - "sqrdmulh v23.4s, v23.4s, v4.4s\n" + "add v18.4s, v18.4s, v26.4s\n" + "add v19.4s, v19.4s, v25.4s\n" + "add v20.4s, v20.4s, v28.4s\n" + "ld1r { v0.4s }, [x20]\n" + "add v21.4s, v21.4s, v27.4s\n" + "add v22.4s, v22.4s, v26.4s\n" + "add v23.4s, v23.4s, v25.4s\n" + "sqrdmulh v16.4s, v16.4s, v24.4s\n" + "sqrdmulh v17.4s, v17.4s, v24.4s\n" + "sqrdmulh v18.4s, v18.4s, v24.4s\n" + "sqrdmulh v19.4s, v19.4s, v24.4s\n" + "sqrdmulh v20.4s, v20.4s, v24.4s\n" + "sqrdmulh v21.4s, v21.4s, v24.4s\n" + "sqrdmulh v22.4s, v22.4s, v24.4s\n" + "sqrdmulh v23.4s, v23.4s, v24.4s\n" "tbz %x[flags], #5, 50f\n" - "and v4.16b, v16.16b, v0.16b\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v4.4s\n" - "and v5.16b, v17.16b, v0.16b\n" - "and v6.16b, v18.16b, v0.16b\n" - "and v7.16b, v19.16b, v0.16b\n" - "and v8.16b, v20.16b, v0.16b\n" - "and v9.16b, v21.16b, v0.16b\n" - "and v10.16b, v22.16b, v0.16b\n" - "and v4.16b, v23.16b, v0.16b\n" - "sshr v5.4s, v5.4s, #0x1f\n" - "sshr v6.4s, v6.4s, #0x1f\n" - "sshr v7.4s, v7.4s, #0x1f\n" - "sshr v8.4s, v8.4s, #0x1f\n" - "sshr v9.4s, v9.4s, #0x1f\n" - "sshr v10.4s, v10.4s, #0x1f\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sqadd v17.4s, v17.4s, v5.4s\n" - "sqadd v18.4s, v18.4s, v6.4s\n" - "sqadd v19.4s, v19.4s, v7.4s\n" - "sqadd v20.4s, v20.4s, v8.4s\n" - "sqadd v21.4s, v21.4s, v9.4s\n" - "sqadd v22.4s, v22.4s, v10.4s\n" - "sqadd v23.4s, v23.4s, v4.4s\n" + "and v24.16b, v16.16b, v0.16b\n" + "sshr v24.4s, v24.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v24.4s\n" + "and v30.16b, v17.16b, v0.16b\n" + "and v29.16b, v18.16b, v0.16b\n" + "and v28.16b, v19.16b, v0.16b\n" + "and v27.16b, v20.16b, v0.16b\n" + "and v26.16b, v21.16b, v0.16b\n" + "and v25.16b, v22.16b, v0.16b\n" + "and v24.16b, v23.16b, v0.16b\n" + "sshr v30.4s, v30.4s, #0x1f\n" + "sshr v29.4s, v29.4s, #0x1f\n" + "sshr v28.4s, v28.4s, #0x1f\n" + "sshr v27.4s, v27.4s, #0x1f\n" + "sshr v26.4s, v26.4s, #0x1f\n" + "sshr v25.4s, v25.4s, #0x1f\n" + "sshr v24.4s, v24.4s, #0x1f\n" + "sqadd v17.4s, v17.4s, v30.4s\n" + "sqadd v18.4s, v18.4s, v29.4s\n" + "sqadd v19.4s, v19.4s, v28.4s\n" + "sqadd v20.4s, v20.4s, v27.4s\n" + "sqadd v21.4s, v21.4s, v26.4s\n" + "sqadd v22.4s, v22.4s, v25.4s\n" + "sqadd v23.4s, v23.4s, v24.4s\n" "50:" // Height 2: no shift correction - "add x23, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[c_offset]\n" + "ld1r { v26.4s }, [x20]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" "srshl v18.4s, v18.4s, v0.4s\n" "srshl v19.4s, v19.4s, v0.4s\n" - "add x23, %x[qp], %[maxval]\n" - "ld1r { v6.4s }, [x23]\n" + "add x20, %x[qp], %[maxval]\n" + "ld1r { v25.4s }, [x20]\n" "srshl v20.4s, v20.4s, v0.4s\n" "srshl v21.4s, v21.4s, v0.4s\n" - "add x23, %x[qp], %[minval]\n" - "ld1r { v5.4s }, [x23]\n" + "add x20, %x[qp], %[minval]\n" + "ld1r { v24.4s }, [x20]\n" "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" "cmp x9, #0x10\n" - "add v16.4s, v16.4s, v4.4s\n" - "add v17.4s, v17.4s, v4.4s\n" - "add v18.4s, v18.4s, v4.4s\n" - "add v19.4s, v19.4s, v4.4s\n" - "add v20.4s, v20.4s, v4.4s\n" - "add v21.4s, v21.4s, v4.4s\n" - "add v22.4s, v22.4s, v4.4s\n" - "add v23.4s, v23.4s, v4.4s\n" - "smin v16.4s, v16.4s, v6.4s\n" - "smin v17.4s, v17.4s, v6.4s\n" - "smin v18.4s, v18.4s, v6.4s\n" - "smin v19.4s, v19.4s, v6.4s\n" - "smin v20.4s, v20.4s, v6.4s\n" - "smin v21.4s, v21.4s, v6.4s\n" - "smin v22.4s, v22.4s, v6.4s\n" - "smin v23.4s, v23.4s, v6.4s\n" - "smax v16.4s, v16.4s, v5.4s\n" - "smax v17.4s, v17.4s, v5.4s\n" - "smax v18.4s, v18.4s, v5.4s\n" - "smax v19.4s, v19.4s, v5.4s\n" - "smax v20.4s, v20.4s, v5.4s\n" - "smax v21.4s, v21.4s, v5.4s\n" - "smax v22.4s, v22.4s, v5.4s\n" - "smax v23.4s, v23.4s, v5.4s\n" + "add v16.4s, v16.4s, v26.4s\n" + "add v17.4s, v17.4s, v26.4s\n" + "add v18.4s, v18.4s, v26.4s\n" + "add v19.4s, v19.4s, v26.4s\n" + "add v20.4s, v20.4s, v26.4s\n" + "add v21.4s, v21.4s, v26.4s\n" + "add v22.4s, v22.4s, v26.4s\n" + "add v23.4s, v23.4s, v26.4s\n" + "smin v16.4s, v16.4s, v25.4s\n" + "smin v17.4s, v17.4s, v25.4s\n" + "smin v18.4s, v18.4s, v25.4s\n" + "smin v19.4s, v19.4s, v25.4s\n" + "smin v20.4s, v20.4s, v25.4s\n" + "smin v21.4s, v21.4s, v25.4s\n" + "smin v22.4s, v22.4s, v25.4s\n" + "smin v23.4s, v23.4s, v25.4s\n" + "smax v16.4s, v16.4s, v24.4s\n" + "smax v17.4s, v17.4s, v24.4s\n" + "smax v18.4s, v18.4s, v24.4s\n" + "smax v19.4s, v19.4s, v24.4s\n" + "smax v20.4s, v20.4s, v24.4s\n" + "smax v21.4s, v21.4s, v24.4s\n" + "smax v22.4s, v22.4s, v24.4s\n" + "smax v23.4s, v23.4s, v24.4s\n" "uzp1 v16.8h, v16.8h, v17.8h\n" - "uzp1 v17.8h, v18.8h, v19.8h\n" + "uzp1 v18.8h, v18.8h, v19.8h\n" "uzp1 v20.8h, v20.8h, v21.8h\n" - "uzp1 v21.8h, v22.8h, v23.8h\n" - "uzp1 v16.16b, v16.16b, v17.16b\n" - "uzp1 v20.16b, v20.16b, v21.16b\n" + "uzp1 v17.8h, v22.8h, v23.8h\n" + "uzp1 v16.16b, v16.16b, v18.16b\n" + "uzp1 v20.16b, v20.16b, v17.16b\n" "bge 59f\n" "tbz x9, #3, 54f\n" "str d16, [x27], #0x8\n" - "str d20, [x22], #0x8\n" + "str d20, [x23], #0x8\n" "tbz x9, #2, 52f\n" "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x22], #0x4\n" + "st1 { v20.s }[2], [x23], #0x4\n" "tbz x9, #1, 51f\n" "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x22], #0x2\n" + "st1 { v20.h }[6], [x23], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x22]\n" + "st1 { v20.b }[14], [x23]\n" "b 58f\n" "51:" // Height 2: Partial direct writeback: partial_1_12 "tbz x9, #0, 58f\n" "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x22]\n" + "st1 { v20.b }[12], [x23]\n" "b 58f\n" "52:" // Height 2: Partial direct writeback: partial_2_8 "tbz x9, #1, 53f\n" "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x22], #0x2\n" + "st1 { v20.h }[4], [x23], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x22]\n" + "st1 { v20.b }[10], [x23]\n" "b 58f\n" "53:" // Height 2: Partial direct writeback: partial_1_8 "tbz x9, #0, 58f\n" "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x22]\n" + "st1 { v20.b }[8], [x23]\n" "b 58f\n" "54:" // Height 2: Partial direct writeback: partial_4_0 "tbz x9, #2, 56f\n" "str s16, [x27], #0x4\n" - "str s20, [x22], #0x4\n" + "str s20, [x23], #0x4\n" "tbz x9, #1, 55f\n" "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x22], #0x2\n" + "st1 { v20.h }[2], [x23], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x22]\n" + "st1 { v20.b }[6], [x23]\n" "b 58f\n" "55:" // Height 2: Partial direct writeback: partial_1_4 "tbz x9, #0, 58f\n" "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x22]\n" + "st1 { v20.b }[4], [x23]\n" "b 58f\n" "56:" // Height 2: Partial direct writeback: partial_2_0 "tbz x9, #1, 57f\n" "str h16, [x27], #0x2\n" - "str h20, [x22], #0x2\n" + "str h20, [x23], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x22]\n" + "st1 { v20.b }[2], [x23]\n" "b 58f\n" "57:" // Height 2: Partial direct writeback: partial_1_0 "str b16, [x27, #0x0]\n" - "str b20, [x22, #0x0]\n" + "str b20, [x23, #0x0]\n" "58:" // Height 2: Partial direct writeback: Done "b 60f\n" "59:" // Height 2: Full writeback "str q16, [x27, #0x0]\n" "add x27, x27, #0x10\n" - "str q20, [x22, #0x0]\n" + "str q20, [x23, #0x0]\n" "60:" // Height 2: Writeback done "subs x9, x9, #0x10\n" "bgt 32b\n" @@ -819,13 +818,13 @@ void a64_hybrid_u8qa_dot_4x16 ( "64:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 65f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" "cbnz x26, 66f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20\n" @@ -834,8 +833,8 @@ void a64_hybrid_u8qa_dot_4x16 ( "b 66f\n" "65:" // Height 3: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" "66:" // Height 3: input setup done "cmp x25, #0x10\n" "blt 71f\n" @@ -857,62 +856,62 @@ void a64_hybrid_u8qa_dot_4x16 ( "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" - "ldr q4, [x28, #0x70]\n" + "ldr q29, [x28, #0x70]\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" "add x22, x22, #0x10\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" - "ldr q5, [x28, #0x80]\n" + "ldr q28, [x28, #0x80]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x28, #0x90]\n" + "ldr q5, [x28, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x28, #0xa0]\n" + "ldr q4, [x28, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" - "ldr q8, [x28, #0xb0]\n" + "ldr q3, [x28, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" - "ldr q9, [x28, #0xc0]\n" + "ldr q31, [x28, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" - "ldr q10, [x28, #0xd0]\n" - ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" - ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" - "ldr q4, [x28, #0xe0]\n" - ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" - ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" - "ldr q5, [x28, #0xf0]\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "ldr q30, [x28, #0xd0]\n" + ".inst 0x6fa0e3b3 // udot v19.4s, v29.16b, v0.4b[1]\n" + ".inst 0x6fa1e3b7 // udot v23.4s, v29.16b, v1.4b[1]\n" + ".inst 0x6fa2e3bb // udot v27.4s, v29.16b, v2.4b[1]\n" + "ldr q29, [x28, #0xe0]\n" + ".inst 0x6f80eb90 // udot v16.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb94 // udot v20.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6f82eb98 // udot v24.4s, v28.16b, v2.4b[2]\n" + "ldr q28, [x28, #0xf0]\n" + ".inst 0x6f80e8b1 // udot v17.4s, v5.16b, v0.4b[2]\n" "add x28, x28, #0x100\n" - ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" - ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" - ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" - ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" - ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" - ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" - ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" - ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" - ".inst 0x6fa2e8bb // udot v27.4s, v5.16b, v2.4b[3]\n" + ".inst 0x6f81e8b5 // udot v21.4s, v5.16b, v1.4b[2]\n" + ".inst 0x6f82e8b9 // udot v25.4s, v5.16b, v2.4b[2]\n" + ".inst 0x6f80e892 // udot v18.4s, v4.16b, v0.4b[2]\n" + ".inst 0x6f81e896 // udot v22.4s, v4.16b, v1.4b[2]\n" + ".inst 0x6f82e89a // udot v26.4s, v4.16b, v2.4b[2]\n" + ".inst 0x6f80e873 // udot v19.4s, v3.16b, v0.4b[2]\n" + ".inst 0x6f81e877 // udot v23.4s, v3.16b, v1.4b[2]\n" + ".inst 0x6f82e87b // udot v27.4s, v3.16b, v2.4b[2]\n" + ".inst 0x6fa0ebf0 // udot v16.4s, v31.16b, v0.4b[3]\n" + ".inst 0x6fa1ebf4 // udot v20.4s, v31.16b, v1.4b[3]\n" + ".inst 0x6fa2ebf8 // udot v24.4s, v31.16b, v2.4b[3]\n" + ".inst 0x6fa0ebd1 // udot v17.4s, v30.16b, v0.4b[3]\n" + ".inst 0x6fa1ebd5 // udot v21.4s, v30.16b, v1.4b[3]\n" + ".inst 0x6fa2ebd9 // udot v25.4s, v30.16b, v2.4b[3]\n" + ".inst 0x6fa0ebb2 // udot v18.4s, v29.16b, v0.4b[3]\n" + ".inst 0x6fa1ebb6 // udot v22.4s, v29.16b, v1.4b[3]\n" + ".inst 0x6fa2ebba // udot v26.4s, v29.16b, v2.4b[3]\n" + ".inst 0x6fa0eb93 // udot v19.4s, v28.16b, v0.4b[3]\n" + ".inst 0x6fa1eb97 // udot v23.4s, v28.16b, v1.4b[3]\n" + ".inst 0x6fa2eb9b // udot v27.4s, v28.16b, v2.4b[3]\n" "tbnz %x[flags], #31, 68f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" @@ -940,63 +939,63 @@ void a64_hybrid_u8qa_dot_4x16 ( "sub x25, x25, #0x10\n" "add x24, x24, #0x10\n" ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" - "ldr q4, [x28, #0x70]\n" + "ldr q29, [x28, #0x70]\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" "add x23, x23, #0x10\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" - "ldr q5, [x28, #0x80]\n" + "ldr q28, [x28, #0x80]\n" "add x22, x22, #0x10\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x28, #0x90]\n" + "ldr q5, [x28, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x28, #0xa0]\n" + "ldr q4, [x28, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" - "ldr q8, [x28, #0xb0]\n" + "ldr q3, [x28, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" - "ldr q9, [x28, #0xc0]\n" + "ldr q31, [x28, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" - "ldr q10, [x28, #0xd0]\n" - ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" - ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" - "ldr q4, [x28, #0xe0]\n" - ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" - ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" - "ldr q5, [x28, #0xf0]\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "ldr q30, [x28, #0xd0]\n" + ".inst 0x6fa0e3b3 // udot v19.4s, v29.16b, v0.4b[1]\n" + ".inst 0x6fa1e3b7 // udot v23.4s, v29.16b, v1.4b[1]\n" + ".inst 0x6fa2e3bb // udot v27.4s, v29.16b, v2.4b[1]\n" + "ldr q29, [x28, #0xe0]\n" + ".inst 0x6f80eb90 // udot v16.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb94 // udot v20.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6f82eb98 // udot v24.4s, v28.16b, v2.4b[2]\n" + "ldr q28, [x28, #0xf0]\n" + ".inst 0x6f80e8b1 // udot v17.4s, v5.16b, v0.4b[2]\n" "add x28, x28, #0x100\n" - ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" - ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" - ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" - ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" - ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" - ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" - ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" - ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" - ".inst 0x6fa2e8bb // udot v27.4s, v5.16b, v2.4b[3]\n" + ".inst 0x6f81e8b5 // udot v21.4s, v5.16b, v1.4b[2]\n" + ".inst 0x6f82e8b9 // udot v25.4s, v5.16b, v2.4b[2]\n" + ".inst 0x6f80e892 // udot v18.4s, v4.16b, v0.4b[2]\n" + ".inst 0x6f81e896 // udot v22.4s, v4.16b, v1.4b[2]\n" + ".inst 0x6f82e89a // udot v26.4s, v4.16b, v2.4b[2]\n" + ".inst 0x6f80e873 // udot v19.4s, v3.16b, v0.4b[2]\n" + ".inst 0x6f81e877 // udot v23.4s, v3.16b, v1.4b[2]\n" + ".inst 0x6f82e87b // udot v27.4s, v3.16b, v2.4b[2]\n" + ".inst 0x6fa0ebf0 // udot v16.4s, v31.16b, v0.4b[3]\n" + ".inst 0x6fa1ebf4 // udot v20.4s, v31.16b, v1.4b[3]\n" + ".inst 0x6fa2ebf8 // udot v24.4s, v31.16b, v2.4b[3]\n" + ".inst 0x6fa0ebd1 // udot v17.4s, v30.16b, v0.4b[3]\n" + ".inst 0x6fa1ebd5 // udot v21.4s, v30.16b, v1.4b[3]\n" + ".inst 0x6fa2ebd9 // udot v25.4s, v30.16b, v2.4b[3]\n" + ".inst 0x6fa0ebb2 // udot v18.4s, v29.16b, v0.4b[3]\n" + ".inst 0x6fa1ebb6 // udot v22.4s, v29.16b, v1.4b[3]\n" + ".inst 0x6fa2ebba // udot v26.4s, v29.16b, v2.4b[3]\n" + ".inst 0x6fa0eb93 // udot v19.4s, v28.16b, v0.4b[3]\n" + ".inst 0x6fa1eb97 // udot v23.4s, v28.16b, v1.4b[3]\n" + ".inst 0x6fa2eb9b // udot v27.4s, v28.16b, v2.4b[3]\n" "tbnz %x[flags], #31, 70f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" @@ -1018,25 +1017,25 @@ void a64_hybrid_u8qa_dot_4x16 ( ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" "73:" // Height 3: Multiply loop: unique 11: skip row sum - "ldr q6, [x28, #0x0]\n" - "ldr q7, [x28, #0x10]\n" + "ldr q31, [x28, #0x0]\n" + "ldr q30, [x28, #0x10]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x4\n" - "ldr q8, [x28, #0x20]\n" - "ldr q9, [x28, #0x30]\n" - ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d8 // udot v24.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" + "ldr q29, [x28, #0x20]\n" + "ldr q28, [x28, #0x30]\n" + ".inst 0x6f80e3f0 // udot v16.4s, v31.16b, v0.4b[0]\n" + ".inst 0x6f81e3f4 // udot v20.4s, v31.16b, v1.4b[0]\n" + ".inst 0x6f82e3f8 // udot v24.4s, v31.16b, v2.4b[0]\n" + ".inst 0x6f80e3d1 // udot v17.4s, v30.16b, v0.4b[0]\n" "add x28, x28, #0x40\n" - ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f9 // udot v25.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" - ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" - ".inst 0x6f82e11a // udot v26.4s, v8.16b, v2.4b[0]\n" - ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" - ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" - ".inst 0x6f82e13b // udot v27.4s, v9.16b, v2.4b[0]\n" + ".inst 0x6f81e3d5 // udot v21.4s, v30.16b, v1.4b[0]\n" + ".inst 0x6f82e3d9 // udot v25.4s, v30.16b, v2.4b[0]\n" + ".inst 0x6f80e3b2 // udot v18.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f81e3b6 // udot v22.4s, v29.16b, v1.4b[0]\n" + ".inst 0x6f82e3ba // udot v26.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f80e393 // udot v19.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f81e397 // udot v23.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f82e39b // udot v27.4s, v28.16b, v2.4b[0]\n" "bge 72b\n" "74:" // Height 3: Multiply loop: Skip odd blocks "cbz x25, 78f\n" @@ -1059,144 +1058,144 @@ void a64_hybrid_u8qa_dot_4x16 ( ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" "77:" // Height 3: Multiply loop: unique 12: skip row sum - "ldr q10, [x28, #0x0]\n" - "ldr q4, [x28, #0x10]\n" - ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" - ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" - "ldr q5, [x28, #0x20]\n" - "ldr q6, [x28, #0x30]\n" - ".inst 0x6f82e158 // udot v24.4s, v10.16b, v2.4b[0]\n" - ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" - ".inst 0x6f82e099 // udot v25.4s, v4.16b, v2.4b[0]\n" + "ldr q31, [x28, #0x0]\n" + "ldr q30, [x28, #0x10]\n" + ".inst 0x6f80e3f0 // udot v16.4s, v31.16b, v0.4b[0]\n" + ".inst 0x6f81e3f4 // udot v20.4s, v31.16b, v1.4b[0]\n" + "ldr q29, [x28, #0x20]\n" + "ldr q28, [x28, #0x30]\n" + ".inst 0x6f82e3f8 // udot v24.4s, v31.16b, v2.4b[0]\n" + ".inst 0x6f80e3d1 // udot v17.4s, v30.16b, v0.4b[0]\n" + ".inst 0x6f81e3d5 // udot v21.4s, v30.16b, v1.4b[0]\n" + ".inst 0x6f82e3d9 // udot v25.4s, v30.16b, v2.4b[0]\n" "add x28, x28, #0x40\n" - ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" - ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" - ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" - ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0db // udot v27.4s, v6.16b, v2.4b[0]\n" + ".inst 0x6f80e3b2 // udot v18.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f81e3b6 // udot v22.4s, v29.16b, v1.4b[0]\n" + ".inst 0x6f82e3ba // udot v26.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f80e393 // udot v19.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f81e397 // udot v23.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f82e39b // udot v27.4s, v28.16b, v2.4b[0]\n" "78:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" "cmp x26, x20\n" "bne 64b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x22, x27, x20\n" - "add x21, x22, x20\n" + "add x23, x27, x20\n" + "add x22, x23, x20\n" "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbnz %x[flags], #31, 79f\n" - "add x23, %x[qp], %[b_offset]\n" - "ld1r { v3.4s }, [x23]\n" + "add x20, %x[qp], %[b_offset]\n" + "ld1r { v28.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" "addp v13.4s, v13.4s, v13.4s\n" - "neg v3.4s, v3.4s\n" + "neg v28.4s, v28.4s\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" "addp v13.4s, v13.4s, v13.4s\n" - "mul v11.4s, v11.4s, v3.4s\n" - "mul v12.4s, v12.4s, v3.4s\n" - "mul v13.4s, v13.4s, v3.4s\n" + "mul v11.4s, v11.4s, v28.4s\n" + "mul v12.4s, v12.4s, v28.4s\n" + "mul v13.4s, v13.4s, v28.4s\n" "79:" // Height 3: skip row sum fixup "ldr q0, [x10, #0x0]\n" - "ldr q1, [x10, #0x10]\n" + "ldr q31, [x10, #0x10]\n" "add v16.4s, v16.4s, v11.4s\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q2, [x10, #0x20]\n" - "ldr q3, [x10, #0x30]\n" + "ldr q30, [x10, #0x20]\n" + "ldr q29, [x10, #0x30]\n" "add v18.4s, v18.4s, v11.4s\n" "add v19.4s, v19.4s, v11.4s\n" "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" - "add x23, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v28.4s }, [x20]\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" "orr %x[flags], %x[flags], #0x80000000\n" - "add x23, %x[qp], %[per_layer_right_shift]\n" + "add x20, %x[qp], %[per_layer_right_shift]\n" "add v24.4s, v24.4s, v13.4s\n" "add v25.4s, v25.4s, v13.4s\n" "add x10, x10, #0x40\n" "add v26.4s, v26.4s, v13.4s\n" "add v27.4s, v27.4s, v13.4s\n" "add v16.4s, v16.4s, v0.4s\n" - "add v17.4s, v17.4s, v1.4s\n" - "add v18.4s, v18.4s, v2.4s\n" - "add v19.4s, v19.4s, v3.4s\n" + "add v17.4s, v17.4s, v31.4s\n" + "add v18.4s, v18.4s, v30.4s\n" + "add v19.4s, v19.4s, v29.4s\n" "add v20.4s, v20.4s, v0.4s\n" - "add v21.4s, v21.4s, v1.4s\n" - "add v22.4s, v22.4s, v2.4s\n" - "add v23.4s, v23.4s, v3.4s\n" + "add v21.4s, v21.4s, v31.4s\n" + "add v22.4s, v22.4s, v30.4s\n" + "add v23.4s, v23.4s, v29.4s\n" "add v24.4s, v24.4s, v0.4s\n" - "ld1r { v0.4s }, [x23]\n" - "add v25.4s, v25.4s, v1.4s\n" - "add v26.4s, v26.4s, v2.4s\n" - "add v27.4s, v27.4s, v3.4s\n" - "sqrdmulh v16.4s, v16.4s, v4.4s\n" - "sqrdmulh v17.4s, v17.4s, v4.4s\n" - "sqrdmulh v18.4s, v18.4s, v4.4s\n" - "sqrdmulh v19.4s, v19.4s, v4.4s\n" - "sqrdmulh v20.4s, v20.4s, v4.4s\n" - "sqrdmulh v21.4s, v21.4s, v4.4s\n" - "sqrdmulh v22.4s, v22.4s, v4.4s\n" - "sqrdmulh v23.4s, v23.4s, v4.4s\n" - "sqrdmulh v24.4s, v24.4s, v4.4s\n" - "sqrdmulh v25.4s, v25.4s, v4.4s\n" - "sqrdmulh v26.4s, v26.4s, v4.4s\n" - "sqrdmulh v27.4s, v27.4s, v4.4s\n" + "ld1r { v0.4s }, [x20]\n" + "add v25.4s, v25.4s, v31.4s\n" + "add v26.4s, v26.4s, v30.4s\n" + "add v27.4s, v27.4s, v29.4s\n" + "sqrdmulh v16.4s, v16.4s, v28.4s\n" + "sqrdmulh v17.4s, v17.4s, v28.4s\n" + "sqrdmulh v18.4s, v18.4s, v28.4s\n" + "sqrdmulh v19.4s, v19.4s, v28.4s\n" + "sqrdmulh v20.4s, v20.4s, v28.4s\n" + "sqrdmulh v21.4s, v21.4s, v28.4s\n" + "sqrdmulh v22.4s, v22.4s, v28.4s\n" + "sqrdmulh v23.4s, v23.4s, v28.4s\n" + "sqrdmulh v24.4s, v24.4s, v28.4s\n" + "sqrdmulh v25.4s, v25.4s, v28.4s\n" + "sqrdmulh v26.4s, v26.4s, v28.4s\n" + "sqrdmulh v27.4s, v27.4s, v28.4s\n" "tbz %x[flags], #5, 80f\n" - "and v4.16b, v16.16b, v0.16b\n" - "and v5.16b, v17.16b, v0.16b\n" - "and v6.16b, v18.16b, v0.16b\n" - "and v7.16b, v19.16b, v0.16b\n" - "and v8.16b, v20.16b, v0.16b\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sshr v5.4s, v5.4s, #0x1f\n" - "sshr v6.4s, v6.4s, #0x1f\n" - "sshr v7.4s, v7.4s, #0x1f\n" - "sshr v8.4s, v8.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v4.4s\n" - "sqadd v17.4s, v17.4s, v5.4s\n" - "sqadd v18.4s, v18.4s, v6.4s\n" - "sqadd v19.4s, v19.4s, v7.4s\n" - "sqadd v20.4s, v20.4s, v8.4s\n" - "and v9.16b, v21.16b, v0.16b\n" - "and v10.16b, v22.16b, v0.16b\n" - "and v4.16b, v23.16b, v0.16b\n" - "and v5.16b, v24.16b, v0.16b\n" - "and v6.16b, v25.16b, v0.16b\n" - "and v7.16b, v26.16b, v0.16b\n" - "and v8.16b, v27.16b, v0.16b\n" - "sshr v9.4s, v9.4s, #0x1f\n" - "sshr v10.4s, v10.4s, #0x1f\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sshr v5.4s, v5.4s, #0x1f\n" - "sshr v6.4s, v6.4s, #0x1f\n" - "sshr v7.4s, v7.4s, #0x1f\n" - "sshr v8.4s, v8.4s, #0x1f\n" - "sqadd v21.4s, v21.4s, v9.4s\n" - "sqadd v22.4s, v22.4s, v10.4s\n" - "sqadd v23.4s, v23.4s, v4.4s\n" - "sqadd v24.4s, v24.4s, v5.4s\n" - "sqadd v25.4s, v25.4s, v6.4s\n" - "sqadd v26.4s, v26.4s, v7.4s\n" - "sqadd v27.4s, v27.4s, v8.4s\n" + "and v1.16b, v16.16b, v0.16b\n" + "and v31.16b, v17.16b, v0.16b\n" + "and v30.16b, v18.16b, v0.16b\n" + "and v29.16b, v19.16b, v0.16b\n" + "and v28.16b, v20.16b, v0.16b\n" + "sshr v1.4s, v1.4s, #0x1f\n" + "sshr v31.4s, v31.4s, #0x1f\n" + "sshr v30.4s, v30.4s, #0x1f\n" + "sshr v29.4s, v29.4s, #0x1f\n" + "sshr v28.4s, v28.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v1.4s\n" + "sqadd v17.4s, v17.4s, v31.4s\n" + "sqadd v18.4s, v18.4s, v30.4s\n" + "sqadd v19.4s, v19.4s, v29.4s\n" + "sqadd v20.4s, v20.4s, v28.4s\n" + "and v3.16b, v21.16b, v0.16b\n" + "and v2.16b, v22.16b, v0.16b\n" + "and v1.16b, v23.16b, v0.16b\n" + "and v31.16b, v24.16b, v0.16b\n" + "and v30.16b, v25.16b, v0.16b\n" + "and v29.16b, v26.16b, v0.16b\n" + "and v28.16b, v27.16b, v0.16b\n" + "sshr v3.4s, v3.4s, #0x1f\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" + "sshr v31.4s, v31.4s, #0x1f\n" + "sshr v30.4s, v30.4s, #0x1f\n" + "sshr v29.4s, v29.4s, #0x1f\n" + "sshr v28.4s, v28.4s, #0x1f\n" + "sqadd v21.4s, v21.4s, v3.4s\n" + "sqadd v22.4s, v22.4s, v2.4s\n" + "sqadd v23.4s, v23.4s, v1.4s\n" + "sqadd v24.4s, v24.4s, v31.4s\n" + "sqadd v25.4s, v25.4s, v30.4s\n" + "sqadd v26.4s, v26.4s, v29.4s\n" + "sqadd v27.4s, v27.4s, v28.4s\n" "80:" // Height 3: no shift correction - "add x23, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[c_offset]\n" + "ld1r { v30.4s }, [x20]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" "srshl v18.4s, v18.4s, v0.4s\n" "srshl v19.4s, v19.4s, v0.4s\n" - "add x23, %x[qp], %[maxval]\n" - "ld1r { v6.4s }, [x23]\n" + "add x20, %x[qp], %[maxval]\n" + "ld1r { v29.4s }, [x20]\n" "srshl v20.4s, v20.4s, v0.4s\n" "srshl v21.4s, v21.4s, v0.4s\n" - "add x23, %x[qp], %[minval]\n" - "ld1r { v5.4s }, [x23]\n" + "add x20, %x[qp], %[minval]\n" + "ld1r { v28.4s }, [x20]\n" "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" "cmp x9, #0x10\n" @@ -1204,132 +1203,132 @@ void a64_hybrid_u8qa_dot_4x16 ( "srshl v25.4s, v25.4s, v0.4s\n" "srshl v26.4s, v26.4s, v0.4s\n" "srshl v27.4s, v27.4s, v0.4s\n" - "add v16.4s, v16.4s, v4.4s\n" - "add v17.4s, v17.4s, v4.4s\n" - "add v18.4s, v18.4s, v4.4s\n" - "add v19.4s, v19.4s, v4.4s\n" - "add v20.4s, v20.4s, v4.4s\n" - "add v21.4s, v21.4s, v4.4s\n" - "add v22.4s, v22.4s, v4.4s\n" - "add v23.4s, v23.4s, v4.4s\n" - "add v24.4s, v24.4s, v4.4s\n" - "add v25.4s, v25.4s, v4.4s\n" - "add v26.4s, v26.4s, v4.4s\n" - "add v27.4s, v27.4s, v4.4s\n" - "smin v16.4s, v16.4s, v6.4s\n" - "smin v17.4s, v17.4s, v6.4s\n" - "smin v18.4s, v18.4s, v6.4s\n" - "smin v19.4s, v19.4s, v6.4s\n" - "smin v20.4s, v20.4s, v6.4s\n" - "smin v21.4s, v21.4s, v6.4s\n" - "smin v22.4s, v22.4s, v6.4s\n" - "smin v23.4s, v23.4s, v6.4s\n" - "smin v24.4s, v24.4s, v6.4s\n" - "smin v25.4s, v25.4s, v6.4s\n" - "smin v26.4s, v26.4s, v6.4s\n" - "smin v27.4s, v27.4s, v6.4s\n" - "smax v16.4s, v16.4s, v5.4s\n" - "smax v17.4s, v17.4s, v5.4s\n" - "smax v18.4s, v18.4s, v5.4s\n" - "smax v19.4s, v19.4s, v5.4s\n" - "smax v20.4s, v20.4s, v5.4s\n" - "smax v21.4s, v21.4s, v5.4s\n" - "smax v22.4s, v22.4s, v5.4s\n" - "smax v23.4s, v23.4s, v5.4s\n" - "smax v24.4s, v24.4s, v5.4s\n" - "smax v25.4s, v25.4s, v5.4s\n" - "smax v26.4s, v26.4s, v5.4s\n" - "smax v27.4s, v27.4s, v5.4s\n" + "add v16.4s, v16.4s, v30.4s\n" + "add v17.4s, v17.4s, v30.4s\n" + "add v18.4s, v18.4s, v30.4s\n" + "add v19.4s, v19.4s, v30.4s\n" + "add v20.4s, v20.4s, v30.4s\n" + "add v21.4s, v21.4s, v30.4s\n" + "add v22.4s, v22.4s, v30.4s\n" + "add v23.4s, v23.4s, v30.4s\n" + "add v24.4s, v24.4s, v30.4s\n" + "add v25.4s, v25.4s, v30.4s\n" + "add v26.4s, v26.4s, v30.4s\n" + "add v27.4s, v27.4s, v30.4s\n" + "smin v16.4s, v16.4s, v29.4s\n" + "smin v17.4s, v17.4s, v29.4s\n" + "smin v18.4s, v18.4s, v29.4s\n" + "smin v19.4s, v19.4s, v29.4s\n" + "smin v20.4s, v20.4s, v29.4s\n" + "smin v21.4s, v21.4s, v29.4s\n" + "smin v22.4s, v22.4s, v29.4s\n" + "smin v23.4s, v23.4s, v29.4s\n" + "smin v24.4s, v24.4s, v29.4s\n" + "smin v25.4s, v25.4s, v29.4s\n" + "smin v26.4s, v26.4s, v29.4s\n" + "smin v27.4s, v27.4s, v29.4s\n" + "smax v16.4s, v16.4s, v28.4s\n" + "smax v17.4s, v17.4s, v28.4s\n" + "smax v18.4s, v18.4s, v28.4s\n" + "smax v19.4s, v19.4s, v28.4s\n" + "smax v20.4s, v20.4s, v28.4s\n" + "smax v21.4s, v21.4s, v28.4s\n" + "smax v22.4s, v22.4s, v28.4s\n" + "smax v23.4s, v23.4s, v28.4s\n" + "smax v24.4s, v24.4s, v28.4s\n" + "smax v25.4s, v25.4s, v28.4s\n" + "smax v26.4s, v26.4s, v28.4s\n" + "smax v27.4s, v27.4s, v28.4s\n" "uzp1 v16.8h, v16.8h, v17.8h\n" - "uzp1 v17.8h, v18.8h, v19.8h\n" + "uzp1 v19.8h, v18.8h, v19.8h\n" "uzp1 v20.8h, v20.8h, v21.8h\n" - "uzp1 v21.8h, v22.8h, v23.8h\n" + "uzp1 v18.8h, v22.8h, v23.8h\n" "uzp1 v24.8h, v24.8h, v25.8h\n" - "uzp1 v25.8h, v26.8h, v27.8h\n" - "uzp1 v16.16b, v16.16b, v17.16b\n" - "uzp1 v20.16b, v20.16b, v21.16b\n" - "uzp1 v24.16b, v24.16b, v25.16b\n" + "uzp1 v17.8h, v26.8h, v27.8h\n" + "uzp1 v16.16b, v16.16b, v19.16b\n" + "uzp1 v20.16b, v20.16b, v18.16b\n" + "uzp1 v24.16b, v24.16b, v17.16b\n" "bge 89f\n" "tbz x9, #3, 84f\n" "str d16, [x27], #0x8\n" - "str d20, [x22], #0x8\n" - "str d24, [x21], #0x8\n" + "str d20, [x23], #0x8\n" + "str d24, [x22], #0x8\n" "tbz x9, #2, 82f\n" "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x22], #0x4\n" - "st1 { v24.s }[2], [x21], #0x4\n" + "st1 { v20.s }[2], [x23], #0x4\n" + "st1 { v24.s }[2], [x22], #0x4\n" "tbz x9, #1, 81f\n" "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x22], #0x2\n" - "st1 { v24.h }[6], [x21], #0x2\n" + "st1 { v20.h }[6], [x23], #0x2\n" + "st1 { v24.h }[6], [x22], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x22]\n" - "st1 { v24.b }[14], [x21]\n" + "st1 { v20.b }[14], [x23]\n" + "st1 { v24.b }[14], [x22]\n" "b 88f\n" "81:" // Height 3: Partial direct writeback: partial_1_12 "tbz x9, #0, 88f\n" "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x22]\n" - "st1 { v24.b }[12], [x21]\n" + "st1 { v20.b }[12], [x23]\n" + "st1 { v24.b }[12], [x22]\n" "b 88f\n" "82:" // Height 3: Partial direct writeback: partial_2_8 "tbz x9, #1, 83f\n" "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x22], #0x2\n" - "st1 { v24.h }[4], [x21], #0x2\n" + "st1 { v20.h }[4], [x23], #0x2\n" + "st1 { v24.h }[4], [x22], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x22]\n" - "st1 { v24.b }[10], [x21]\n" + "st1 { v20.b }[10], [x23]\n" + "st1 { v24.b }[10], [x22]\n" "b 88f\n" "83:" // Height 3: Partial direct writeback: partial_1_8 "tbz x9, #0, 88f\n" "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x22]\n" - "st1 { v24.b }[8], [x21]\n" + "st1 { v20.b }[8], [x23]\n" + "st1 { v24.b }[8], [x22]\n" "b 88f\n" "84:" // Height 3: Partial direct writeback: partial_4_0 "tbz x9, #2, 86f\n" "str s16, [x27], #0x4\n" - "str s20, [x22], #0x4\n" - "str s24, [x21], #0x4\n" + "str s20, [x23], #0x4\n" + "str s24, [x22], #0x4\n" "tbz x9, #1, 85f\n" "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x22], #0x2\n" - "st1 { v24.h }[2], [x21], #0x2\n" + "st1 { v20.h }[2], [x23], #0x2\n" + "st1 { v24.h }[2], [x22], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x22]\n" - "st1 { v24.b }[6], [x21]\n" + "st1 { v20.b }[6], [x23]\n" + "st1 { v24.b }[6], [x22]\n" "b 88f\n" "85:" // Height 3: Partial direct writeback: partial_1_4 "tbz x9, #0, 88f\n" "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x22]\n" - "st1 { v24.b }[4], [x21]\n" + "st1 { v20.b }[4], [x23]\n" + "st1 { v24.b }[4], [x22]\n" "b 88f\n" "86:" // Height 3: Partial direct writeback: partial_2_0 "tbz x9, #1, 87f\n" "str h16, [x27], #0x2\n" - "str h20, [x22], #0x2\n" - "str h24, [x21], #0x2\n" + "str h20, [x23], #0x2\n" + "str h24, [x22], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x22]\n" - "st1 { v24.b }[2], [x21]\n" + "st1 { v20.b }[2], [x23]\n" + "st1 { v24.b }[2], [x22]\n" "b 88f\n" "87:" // Height 3: Partial direct writeback: partial_1_0 "str b16, [x27, #0x0]\n" - "str b20, [x22, #0x0]\n" - "str b24, [x21, #0x0]\n" + "str b20, [x23, #0x0]\n" + "str b24, [x22, #0x0]\n" "88:" // Height 3: Partial direct writeback: Done "b 90f\n" "89:" // Height 3: Full writeback "str q16, [x27, #0x0]\n" "add x27, x27, #0x10\n" - "str q20, [x22, #0x0]\n" - "str q24, [x21, #0x0]\n" + "str q20, [x23, #0x0]\n" + "str q24, [x22, #0x0]\n" "90:" // Height 3: Writeback done "subs x9, x9, #0x10\n" "bgt 62b\n" @@ -1370,14 +1369,14 @@ void a64_hybrid_u8qa_dot_4x16 ( "94:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 95f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" - "ldr x21, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" + "ldr x21, [x20, #0x18]\n" "cbnz x26, 96f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20\n" @@ -1387,9 +1386,9 @@ void a64_hybrid_u8qa_dot_4x16 ( "b 96f\n" "95:" // Height 4: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" - "add x21, x22, x20\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" + "add x21, x22, x21\n" "96:" // Height 4: input setup done "cmp x25, #0x10\n" "blt 101f\n" @@ -1614,29 +1613,29 @@ void a64_hybrid_u8qa_dot_4x16 ( ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" ".inst 0x6e8f946e // udot v14.4s, v3.16b, v15.16b\n" "103:" // Height 4: Multiply loop: unique 15: skip row sum - "ldr q6, [x28, #0x0]\n" - "ldr q7, [x28, #0x10]\n" + "ldr q7, [x28, #0x0]\n" + "ldr q6, [x28, #0x10]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x4\n" - "ldr q8, [x28, #0x20]\n" - "ldr q9, [x28, #0x30]\n" - ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d8 // udot v24.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0dc // udot v28.4s, v6.16b, v3.4b[0]\n" + "ldr q5, [x28, #0x20]\n" + "ldr q4, [x28, #0x30]\n" + ".inst 0x6f80e0f0 // udot v16.4s, v7.16b, v0.4b[0]\n" + ".inst 0x6f81e0f4 // udot v20.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f82e0f8 // udot v24.4s, v7.16b, v2.4b[0]\n" + ".inst 0x6f83e0fc // udot v28.4s, v7.16b, v3.4b[0]\n" "add x28, x28, #0x40\n" - ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f9 // udot v25.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0fd // udot v29.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" - ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" - ".inst 0x6f82e11a // udot v26.4s, v8.16b, v2.4b[0]\n" - ".inst 0x6f83e11e // udot v30.4s, v8.16b, v3.4b[0]\n" - ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" - ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" - ".inst 0x6f82e13b // udot v27.4s, v9.16b, v2.4b[0]\n" - ".inst 0x6f83e13f // udot v31.4s, v9.16b, v3.4b[0]\n" + ".inst 0x6f80e0d1 // udot v17.4s, v6.16b, v0.4b[0]\n" + ".inst 0x6f81e0d5 // udot v21.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f82e0d9 // udot v25.4s, v6.16b, v2.4b[0]\n" + ".inst 0x6f83e0dd // udot v29.4s, v6.16b, v3.4b[0]\n" + ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" + ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" + ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" + ".inst 0x6f83e0be // udot v30.4s, v5.16b, v3.4b[0]\n" + ".inst 0x6f80e093 // udot v19.4s, v4.16b, v0.4b[0]\n" + ".inst 0x6f81e097 // udot v23.4s, v4.16b, v1.4b[0]\n" + ".inst 0x6f82e09b // udot v27.4s, v4.16b, v2.4b[0]\n" + ".inst 0x6f83e09f // udot v31.4s, v4.16b, v3.4b[0]\n" "bge 102b\n" "104:" // Height 4: Multiply loop: Skip odd blocks "cbz x25, 108f\n" @@ -1663,73 +1662,73 @@ void a64_hybrid_u8qa_dot_4x16 ( ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" ".inst 0x6e8f946e // udot v14.4s, v3.16b, v15.16b\n" "107:" // Height 4: Multiply loop: unique 16: skip row sum - "ldr q10, [x28, #0x0]\n" - "ldr q4, [x28, #0x10]\n" - ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" - ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" + "ldr q7, [x28, #0x0]\n" + "ldr q6, [x28, #0x10]\n" + ".inst 0x6f80e0f0 // udot v16.4s, v7.16b, v0.4b[0]\n" + ".inst 0x6f81e0f4 // udot v20.4s, v7.16b, v1.4b[0]\n" "ldr q5, [x28, #0x20]\n" - "ldr q6, [x28, #0x30]\n" - ".inst 0x6f82e158 // udot v24.4s, v10.16b, v2.4b[0]\n" - ".inst 0x6f83e15c // udot v28.4s, v10.16b, v3.4b[0]\n" - ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" + "ldr q4, [x28, #0x30]\n" + ".inst 0x6f82e0f8 // udot v24.4s, v7.16b, v2.4b[0]\n" + ".inst 0x6f83e0fc // udot v28.4s, v7.16b, v3.4b[0]\n" + ".inst 0x6f80e0d1 // udot v17.4s, v6.16b, v0.4b[0]\n" + ".inst 0x6f81e0d5 // udot v21.4s, v6.16b, v1.4b[0]\n" "add x28, x28, #0x40\n" - ".inst 0x6f82e099 // udot v25.4s, v4.16b, v2.4b[0]\n" - ".inst 0x6f83e09d // udot v29.4s, v4.16b, v3.4b[0]\n" + ".inst 0x6f82e0d9 // udot v25.4s, v6.16b, v2.4b[0]\n" + ".inst 0x6f83e0dd // udot v29.4s, v6.16b, v3.4b[0]\n" ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" ".inst 0x6f83e0be // udot v30.4s, v5.16b, v3.4b[0]\n" - ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0db // udot v27.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0df // udot v31.4s, v6.16b, v3.4b[0]\n" + ".inst 0x6f80e093 // udot v19.4s, v4.16b, v0.4b[0]\n" + ".inst 0x6f81e097 // udot v23.4s, v4.16b, v1.4b[0]\n" + ".inst 0x6f82e09b // udot v27.4s, v4.16b, v2.4b[0]\n" + ".inst 0x6f83e09f // udot v31.4s, v4.16b, v3.4b[0]\n" "108:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" "cmp x26, x20\n" "bne 94b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x22, x27, x20\n" - "add x21, x22, x20\n" + "add x23, x27, x20\n" + "add x22, x23, x20\n" "prfm pstl1keep, [x27, #0x0]\n" - "add x20, x21, x20\n" + "add x21, x22, x20\n" + "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" "prfm pstl1keep, [x21, #0x0]\n" - "prfm pstl1keep, [x20, #0x0]\n" "tbnz %x[flags], #31, 109f\n" - "add x23, %x[qp], %[b_offset]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[b_offset]\n" + "ld1r { v0.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" "addp v13.4s, v13.4s, v13.4s\n" "addp v14.4s, v14.4s, v14.4s\n" - "neg v4.4s, v4.4s\n" + "neg v0.4s, v0.4s\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" "addp v13.4s, v13.4s, v13.4s\n" "addp v14.4s, v14.4s, v14.4s\n" - "mul v11.4s, v11.4s, v4.4s\n" - "mul v12.4s, v12.4s, v4.4s\n" - "mul v13.4s, v13.4s, v4.4s\n" - "mul v14.4s, v14.4s, v4.4s\n" + "mul v11.4s, v11.4s, v0.4s\n" + "mul v12.4s, v12.4s, v0.4s\n" + "mul v13.4s, v13.4s, v0.4s\n" + "mul v14.4s, v14.4s, v0.4s\n" "109:" // Height 4: skip row sum fixup "ldr q0, [x10, #0x0]\n" - "ldr q1, [x10, #0x10]\n" + "ldr q4, [x10, #0x10]\n" "add v16.4s, v16.4s, v11.4s\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q2, [x10, #0x20]\n" - "ldr q3, [x10, #0x30]\n" + "ldr q3, [x10, #0x20]\n" + "ldr q2, [x10, #0x30]\n" "add v18.4s, v18.4s, v11.4s\n" "add v19.4s, v19.4s, v11.4s\n" "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" - "add x23, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v1.4s }, [x20]\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" "orr %x[flags], %x[flags], #0x80000000\n" - "add x23, %x[qp], %[per_layer_right_shift]\n" + "add x20, %x[qp], %[per_layer_right_shift]\n" "add v24.4s, v24.4s, v13.4s\n" "add v25.4s, v25.4s, v13.4s\n" "add x10, x10, #0x40\n" @@ -1740,100 +1739,100 @@ void a64_hybrid_u8qa_dot_4x16 ( "add v30.4s, v30.4s, v14.4s\n" "add v31.4s, v31.4s, v14.4s\n" "add v16.4s, v16.4s, v0.4s\n" - "add v17.4s, v17.4s, v1.4s\n" - "add v18.4s, v18.4s, v2.4s\n" - "add v19.4s, v19.4s, v3.4s\n" + "add v17.4s, v17.4s, v4.4s\n" + "add v18.4s, v18.4s, v3.4s\n" + "add v19.4s, v19.4s, v2.4s\n" "add v20.4s, v20.4s, v0.4s\n" - "add v21.4s, v21.4s, v1.4s\n" - "add v22.4s, v22.4s, v2.4s\n" - "add v23.4s, v23.4s, v3.4s\n" + "add v21.4s, v21.4s, v4.4s\n" + "add v22.4s, v22.4s, v3.4s\n" + "add v23.4s, v23.4s, v2.4s\n" "add v24.4s, v24.4s, v0.4s\n" - "add v25.4s, v25.4s, v1.4s\n" - "add v26.4s, v26.4s, v2.4s\n" - "add v27.4s, v27.4s, v3.4s\n" + "add v25.4s, v25.4s, v4.4s\n" + "add v26.4s, v26.4s, v3.4s\n" + "add v27.4s, v27.4s, v2.4s\n" "add v28.4s, v28.4s, v0.4s\n" - "ld1r { v0.4s }, [x23]\n" - "add v29.4s, v29.4s, v1.4s\n" - "add v30.4s, v30.4s, v2.4s\n" - "add v31.4s, v31.4s, v3.4s\n" - "sqrdmulh v16.4s, v16.4s, v4.4s\n" - "sqrdmulh v17.4s, v17.4s, v4.4s\n" - "sqrdmulh v18.4s, v18.4s, v4.4s\n" - "sqrdmulh v19.4s, v19.4s, v4.4s\n" - "sqrdmulh v20.4s, v20.4s, v4.4s\n" - "sqrdmulh v21.4s, v21.4s, v4.4s\n" - "sqrdmulh v22.4s, v22.4s, v4.4s\n" - "sqrdmulh v23.4s, v23.4s, v4.4s\n" - "sqrdmulh v24.4s, v24.4s, v4.4s\n" - "sqrdmulh v25.4s, v25.4s, v4.4s\n" - "sqrdmulh v26.4s, v26.4s, v4.4s\n" - "sqrdmulh v27.4s, v27.4s, v4.4s\n" - "sqrdmulh v28.4s, v28.4s, v4.4s\n" - "sqrdmulh v29.4s, v29.4s, v4.4s\n" - "sqrdmulh v30.4s, v30.4s, v4.4s\n" - "sqrdmulh v31.4s, v31.4s, v4.4s\n" + "ld1r { v0.4s }, [x20]\n" + "add v29.4s, v29.4s, v4.4s\n" + "add v30.4s, v30.4s, v3.4s\n" + "add v31.4s, v31.4s, v2.4s\n" + "sqrdmulh v16.4s, v16.4s, v1.4s\n" + "sqrdmulh v17.4s, v17.4s, v1.4s\n" + "sqrdmulh v18.4s, v18.4s, v1.4s\n" + "sqrdmulh v19.4s, v19.4s, v1.4s\n" + "sqrdmulh v20.4s, v20.4s, v1.4s\n" + "sqrdmulh v21.4s, v21.4s, v1.4s\n" + "sqrdmulh v22.4s, v22.4s, v1.4s\n" + "sqrdmulh v23.4s, v23.4s, v1.4s\n" + "sqrdmulh v24.4s, v24.4s, v1.4s\n" + "sqrdmulh v25.4s, v25.4s, v1.4s\n" + "sqrdmulh v26.4s, v26.4s, v1.4s\n" + "sqrdmulh v27.4s, v27.4s, v1.4s\n" + "sqrdmulh v28.4s, v28.4s, v1.4s\n" + "sqrdmulh v29.4s, v29.4s, v1.4s\n" + "sqrdmulh v30.4s, v30.4s, v1.4s\n" + "sqrdmulh v31.4s, v31.4s, v1.4s\n" "tbz %x[flags], #5, 110f\n" - "and v4.16b, v16.16b, v0.16b\n" - "and v5.16b, v17.16b, v0.16b\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sshr v5.4s, v5.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v4.4s\n" - "sqadd v17.4s, v17.4s, v5.4s\n" - "and v6.16b, v18.16b, v0.16b\n" - "and v7.16b, v19.16b, v0.16b\n" - "and v8.16b, v20.16b, v0.16b\n" - "and v9.16b, v21.16b, v0.16b\n" - "and v10.16b, v22.16b, v0.16b\n" - "and v4.16b, v23.16b, v0.16b\n" - "and v5.16b, v24.16b, v0.16b\n" - "sshr v6.4s, v6.4s, #0x1f\n" + "and v2.16b, v16.16b, v0.16b\n" + "and v1.16b, v17.16b, v0.16b\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v2.4s\n" + "sqadd v17.4s, v17.4s, v1.4s\n" + "and v7.16b, v18.16b, v0.16b\n" + "and v6.16b, v19.16b, v0.16b\n" + "and v5.16b, v20.16b, v0.16b\n" + "and v4.16b, v21.16b, v0.16b\n" + "and v3.16b, v22.16b, v0.16b\n" + "and v2.16b, v23.16b, v0.16b\n" + "and v1.16b, v24.16b, v0.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" - "sshr v8.4s, v8.4s, #0x1f\n" - "sshr v9.4s, v9.4s, #0x1f\n" - "sshr v10.4s, v10.4s, #0x1f\n" - "sshr v4.4s, v4.4s, #0x1f\n" - "sshr v5.4s, v5.4s, #0x1f\n" - "sqadd v18.4s, v18.4s, v6.4s\n" - "sqadd v19.4s, v19.4s, v7.4s\n" - "sqadd v20.4s, v20.4s, v8.4s\n" - "sqadd v21.4s, v21.4s, v9.4s\n" - "sqadd v22.4s, v22.4s, v10.4s\n" - "sqadd v23.4s, v23.4s, v4.4s\n" - "sqadd v24.4s, v24.4s, v5.4s\n" - "and v6.16b, v25.16b, v0.16b\n" - "and v7.16b, v26.16b, v0.16b\n" - "and v8.16b, v27.16b, v0.16b\n" - "and v9.16b, v28.16b, v0.16b\n" - "and v10.16b, v29.16b, v0.16b\n" - "and v4.16b, v30.16b, v0.16b\n" - "and v5.16b, v31.16b, v0.16b\n" "sshr v6.4s, v6.4s, #0x1f\n" - "sshr v7.4s, v7.4s, #0x1f\n" - "sshr v8.4s, v8.4s, #0x1f\n" - "sshr v9.4s, v9.4s, #0x1f\n" - "sshr v10.4s, v10.4s, #0x1f\n" + "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" + "sshr v3.4s, v3.4s, #0x1f\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" + "sqadd v18.4s, v18.4s, v7.4s\n" + "sqadd v19.4s, v19.4s, v6.4s\n" + "sqadd v20.4s, v20.4s, v5.4s\n" + "sqadd v21.4s, v21.4s, v4.4s\n" + "sqadd v22.4s, v22.4s, v3.4s\n" + "sqadd v23.4s, v23.4s, v2.4s\n" + "sqadd v24.4s, v24.4s, v1.4s\n" + "and v7.16b, v25.16b, v0.16b\n" + "and v6.16b, v26.16b, v0.16b\n" + "and v5.16b, v27.16b, v0.16b\n" + "and v4.16b, v28.16b, v0.16b\n" + "and v3.16b, v29.16b, v0.16b\n" + "and v2.16b, v30.16b, v0.16b\n" + "and v1.16b, v31.16b, v0.16b\n" + "sshr v7.4s, v7.4s, #0x1f\n" + "sshr v6.4s, v6.4s, #0x1f\n" "sshr v5.4s, v5.4s, #0x1f\n" - "sqadd v25.4s, v25.4s, v6.4s\n" - "sqadd v26.4s, v26.4s, v7.4s\n" - "sqadd v27.4s, v27.4s, v8.4s\n" - "sqadd v28.4s, v28.4s, v9.4s\n" - "sqadd v29.4s, v29.4s, v10.4s\n" - "sqadd v30.4s, v30.4s, v4.4s\n" - "sqadd v31.4s, v31.4s, v5.4s\n" + "sshr v4.4s, v4.4s, #0x1f\n" + "sshr v3.4s, v3.4s, #0x1f\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" + "sqadd v25.4s, v25.4s, v7.4s\n" + "sqadd v26.4s, v26.4s, v6.4s\n" + "sqadd v27.4s, v27.4s, v5.4s\n" + "sqadd v28.4s, v28.4s, v4.4s\n" + "sqadd v29.4s, v29.4s, v3.4s\n" + "sqadd v30.4s, v30.4s, v2.4s\n" + "sqadd v31.4s, v31.4s, v1.4s\n" "110:" // Height 4: no shift correction - "add x23, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x23]\n" + "add x20, %x[qp], %[c_offset]\n" + "ld1r { v3.4s }, [x20]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" "srshl v18.4s, v18.4s, v0.4s\n" "srshl v19.4s, v19.4s, v0.4s\n" - "add x23, %x[qp], %[maxval]\n" - "ld1r { v6.4s }, [x23]\n" + "add x20, %x[qp], %[maxval]\n" + "ld1r { v2.4s }, [x20]\n" "srshl v20.4s, v20.4s, v0.4s\n" "srshl v21.4s, v21.4s, v0.4s\n" - "add x23, %x[qp], %[minval]\n" - "ld1r { v5.4s }, [x23]\n" + "add x20, %x[qp], %[minval]\n" + "ld1r { v1.4s }, [x20]\n" "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" "cmp x9, #0x10\n" @@ -1845,163 +1844,163 @@ void a64_hybrid_u8qa_dot_4x16 ( "srshl v29.4s, v29.4s, v0.4s\n" "srshl v30.4s, v30.4s, v0.4s\n" "srshl v31.4s, v31.4s, v0.4s\n" - "add v16.4s, v16.4s, v4.4s\n" - "add v17.4s, v17.4s, v4.4s\n" - "add v18.4s, v18.4s, v4.4s\n" - "add v19.4s, v19.4s, v4.4s\n" - "add v20.4s, v20.4s, v4.4s\n" - "add v21.4s, v21.4s, v4.4s\n" - "add v22.4s, v22.4s, v4.4s\n" - "add v23.4s, v23.4s, v4.4s\n" - "add v24.4s, v24.4s, v4.4s\n" - "add v25.4s, v25.4s, v4.4s\n" - "add v26.4s, v26.4s, v4.4s\n" - "add v27.4s, v27.4s, v4.4s\n" - "add v28.4s, v28.4s, v4.4s\n" - "add v29.4s, v29.4s, v4.4s\n" - "add v30.4s, v30.4s, v4.4s\n" - "add v31.4s, v31.4s, v4.4s\n" - "smin v16.4s, v16.4s, v6.4s\n" - "smin v17.4s, v17.4s, v6.4s\n" - "smin v18.4s, v18.4s, v6.4s\n" - "smin v19.4s, v19.4s, v6.4s\n" - "smin v20.4s, v20.4s, v6.4s\n" - "smin v21.4s, v21.4s, v6.4s\n" - "smin v22.4s, v22.4s, v6.4s\n" - "smin v23.4s, v23.4s, v6.4s\n" - "smin v24.4s, v24.4s, v6.4s\n" - "smin v25.4s, v25.4s, v6.4s\n" - "smin v26.4s, v26.4s, v6.4s\n" - "smin v27.4s, v27.4s, v6.4s\n" - "smin v28.4s, v28.4s, v6.4s\n" - "smin v29.4s, v29.4s, v6.4s\n" - "smin v30.4s, v30.4s, v6.4s\n" - "smin v31.4s, v31.4s, v6.4s\n" - "smax v16.4s, v16.4s, v5.4s\n" - "smax v17.4s, v17.4s, v5.4s\n" - "smax v18.4s, v18.4s, v5.4s\n" - "smax v19.4s, v19.4s, v5.4s\n" - "smax v20.4s, v20.4s, v5.4s\n" - "smax v21.4s, v21.4s, v5.4s\n" - "smax v22.4s, v22.4s, v5.4s\n" - "smax v23.4s, v23.4s, v5.4s\n" - "smax v24.4s, v24.4s, v5.4s\n" - "smax v25.4s, v25.4s, v5.4s\n" - "smax v26.4s, v26.4s, v5.4s\n" - "smax v27.4s, v27.4s, v5.4s\n" - "smax v28.4s, v28.4s, v5.4s\n" - "smax v29.4s, v29.4s, v5.4s\n" - "smax v30.4s, v30.4s, v5.4s\n" - "smax v31.4s, v31.4s, v5.4s\n" + "add v16.4s, v16.4s, v3.4s\n" + "add v17.4s, v17.4s, v3.4s\n" + "add v18.4s, v18.4s, v3.4s\n" + "add v19.4s, v19.4s, v3.4s\n" + "add v20.4s, v20.4s, v3.4s\n" + "add v21.4s, v21.4s, v3.4s\n" + "add v22.4s, v22.4s, v3.4s\n" + "add v23.4s, v23.4s, v3.4s\n" + "add v24.4s, v24.4s, v3.4s\n" + "add v25.4s, v25.4s, v3.4s\n" + "add v26.4s, v26.4s, v3.4s\n" + "add v27.4s, v27.4s, v3.4s\n" + "add v28.4s, v28.4s, v3.4s\n" + "add v29.4s, v29.4s, v3.4s\n" + "add v30.4s, v30.4s, v3.4s\n" + "add v31.4s, v31.4s, v3.4s\n" + "smin v16.4s, v16.4s, v2.4s\n" + "smin v17.4s, v17.4s, v2.4s\n" + "smin v18.4s, v18.4s, v2.4s\n" + "smin v19.4s, v19.4s, v2.4s\n" + "smin v20.4s, v20.4s, v2.4s\n" + "smin v21.4s, v21.4s, v2.4s\n" + "smin v22.4s, v22.4s, v2.4s\n" + "smin v23.4s, v23.4s, v2.4s\n" + "smin v24.4s, v24.4s, v2.4s\n" + "smin v25.4s, v25.4s, v2.4s\n" + "smin v26.4s, v26.4s, v2.4s\n" + "smin v27.4s, v27.4s, v2.4s\n" + "smin v28.4s, v28.4s, v2.4s\n" + "smin v29.4s, v29.4s, v2.4s\n" + "smin v30.4s, v30.4s, v2.4s\n" + "smin v31.4s, v31.4s, v2.4s\n" + "smax v16.4s, v16.4s, v1.4s\n" + "smax v17.4s, v17.4s, v1.4s\n" + "smax v18.4s, v18.4s, v1.4s\n" + "smax v19.4s, v19.4s, v1.4s\n" + "smax v20.4s, v20.4s, v1.4s\n" + "smax v21.4s, v21.4s, v1.4s\n" + "smax v22.4s, v22.4s, v1.4s\n" + "smax v23.4s, v23.4s, v1.4s\n" + "smax v24.4s, v24.4s, v1.4s\n" + "smax v25.4s, v25.4s, v1.4s\n" + "smax v26.4s, v26.4s, v1.4s\n" + "smax v27.4s, v27.4s, v1.4s\n" + "smax v28.4s, v28.4s, v1.4s\n" + "smax v29.4s, v29.4s, v1.4s\n" + "smax v30.4s, v30.4s, v1.4s\n" + "smax v31.4s, v31.4s, v1.4s\n" "uzp1 v16.8h, v16.8h, v17.8h\n" - "uzp1 v17.8h, v18.8h, v19.8h\n" + "uzp1 v0.8h, v18.8h, v19.8h\n" "uzp1 v20.8h, v20.8h, v21.8h\n" - "uzp1 v21.8h, v22.8h, v23.8h\n" + "uzp1 v19.8h, v22.8h, v23.8h\n" "uzp1 v24.8h, v24.8h, v25.8h\n" - "uzp1 v25.8h, v26.8h, v27.8h\n" + "uzp1 v18.8h, v26.8h, v27.8h\n" "uzp1 v28.8h, v28.8h, v29.8h\n" - "uzp1 v29.8h, v30.8h, v31.8h\n" - "uzp1 v16.16b, v16.16b, v17.16b\n" - "uzp1 v20.16b, v20.16b, v21.16b\n" - "uzp1 v24.16b, v24.16b, v25.16b\n" - "uzp1 v28.16b, v28.16b, v29.16b\n" + "uzp1 v17.8h, v30.8h, v31.8h\n" + "uzp1 v16.16b, v16.16b, v0.16b\n" + "uzp1 v20.16b, v20.16b, v19.16b\n" + "uzp1 v24.16b, v24.16b, v18.16b\n" + "uzp1 v28.16b, v28.16b, v17.16b\n" "bge 119f\n" "tbz x9, #3, 114f\n" "str d16, [x27], #0x8\n" - "str d20, [x22], #0x8\n" - "str d24, [x21], #0x8\n" - "str d28, [x20], #0x8\n" + "str d20, [x23], #0x8\n" + "str d24, [x22], #0x8\n" + "str d28, [x21], #0x8\n" "tbz x9, #2, 112f\n" "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x22], #0x4\n" - "st1 { v24.s }[2], [x21], #0x4\n" - "st1 { v28.s }[2], [x20], #0x4\n" + "st1 { v20.s }[2], [x23], #0x4\n" + "st1 { v24.s }[2], [x22], #0x4\n" + "st1 { v28.s }[2], [x21], #0x4\n" "tbz x9, #1, 111f\n" "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x22], #0x2\n" - "st1 { v24.h }[6], [x21], #0x2\n" - "st1 { v28.h }[6], [x20], #0x2\n" + "st1 { v20.h }[6], [x23], #0x2\n" + "st1 { v24.h }[6], [x22], #0x2\n" + "st1 { v28.h }[6], [x21], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x22]\n" - "st1 { v24.b }[14], [x21]\n" - "st1 { v28.b }[14], [x20]\n" + "st1 { v20.b }[14], [x23]\n" + "st1 { v24.b }[14], [x22]\n" + "st1 { v28.b }[14], [x21]\n" "b 118f\n" "111:" // Height 4: Partial direct writeback: partial_1_12 "tbz x9, #0, 118f\n" "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x22]\n" - "st1 { v24.b }[12], [x21]\n" - "st1 { v28.b }[12], [x20]\n" + "st1 { v20.b }[12], [x23]\n" + "st1 { v24.b }[12], [x22]\n" + "st1 { v28.b }[12], [x21]\n" "b 118f\n" "112:" // Height 4: Partial direct writeback: partial_2_8 "tbz x9, #1, 113f\n" "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x22], #0x2\n" - "st1 { v24.h }[4], [x21], #0x2\n" - "st1 { v28.h }[4], [x20], #0x2\n" + "st1 { v20.h }[4], [x23], #0x2\n" + "st1 { v24.h }[4], [x22], #0x2\n" + "st1 { v28.h }[4], [x21], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x22]\n" - "st1 { v24.b }[10], [x21]\n" - "st1 { v28.b }[10], [x20]\n" + "st1 { v20.b }[10], [x23]\n" + "st1 { v24.b }[10], [x22]\n" + "st1 { v28.b }[10], [x21]\n" "b 118f\n" "113:" // Height 4: Partial direct writeback: partial_1_8 "tbz x9, #0, 118f\n" "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x22]\n" - "st1 { v24.b }[8], [x21]\n" - "st1 { v28.b }[8], [x20]\n" + "st1 { v20.b }[8], [x23]\n" + "st1 { v24.b }[8], [x22]\n" + "st1 { v28.b }[8], [x21]\n" "b 118f\n" "114:" // Height 4: Partial direct writeback: partial_4_0 "tbz x9, #2, 116f\n" "str s16, [x27], #0x4\n" - "str s20, [x22], #0x4\n" - "str s24, [x21], #0x4\n" - "str s28, [x20], #0x4\n" + "str s20, [x23], #0x4\n" + "str s24, [x22], #0x4\n" + "str s28, [x21], #0x4\n" "tbz x9, #1, 115f\n" "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x22], #0x2\n" - "st1 { v24.h }[2], [x21], #0x2\n" - "st1 { v28.h }[2], [x20], #0x2\n" + "st1 { v20.h }[2], [x23], #0x2\n" + "st1 { v24.h }[2], [x22], #0x2\n" + "st1 { v28.h }[2], [x21], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x22]\n" - "st1 { v24.b }[6], [x21]\n" - "st1 { v28.b }[6], [x20]\n" + "st1 { v20.b }[6], [x23]\n" + "st1 { v24.b }[6], [x22]\n" + "st1 { v28.b }[6], [x21]\n" "b 118f\n" "115:" // Height 4: Partial direct writeback: partial_1_4 "tbz x9, #0, 118f\n" "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x22]\n" - "st1 { v24.b }[4], [x21]\n" - "st1 { v28.b }[4], [x20]\n" + "st1 { v20.b }[4], [x23]\n" + "st1 { v24.b }[4], [x22]\n" + "st1 { v28.b }[4], [x21]\n" "b 118f\n" "116:" // Height 4: Partial direct writeback: partial_2_0 "tbz x9, #1, 117f\n" "str h16, [x27], #0x2\n" - "str h20, [x22], #0x2\n" - "str h24, [x21], #0x2\n" - "str h28, [x20], #0x2\n" + "str h20, [x23], #0x2\n" + "str h24, [x22], #0x2\n" + "str h28, [x21], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x22]\n" - "st1 { v24.b }[2], [x21]\n" - "st1 { v28.b }[2], [x20]\n" + "st1 { v20.b }[2], [x23]\n" + "st1 { v24.b }[2], [x22]\n" + "st1 { v28.b }[2], [x21]\n" "b 118f\n" "117:" // Height 4: Partial direct writeback: partial_1_0 "str b16, [x27, #0x0]\n" - "str b20, [x22, #0x0]\n" - "str b24, [x21, #0x0]\n" - "str b28, [x20, #0x0]\n" + "str b20, [x23, #0x0]\n" + "str b24, [x22, #0x0]\n" + "str b28, [x21, #0x0]\n" "118:" // Height 4: Partial direct writeback: Done "b 120f\n" "119:" // Height 4: Full writeback "str q16, [x27, #0x0]\n" "add x27, x27, #0x10\n" - "str q20, [x22, #0x0]\n" - "str q24, [x21, #0x0]\n" - "str q28, [x20, #0x0]\n" + "str q20, [x23, #0x0]\n" + "str q24, [x22, #0x0]\n" + "str q28, [x21, #0x0]\n" "120:" // Height 4: Writeback done "subs x9, x9, #0x10\n" "bgt 92b\n" @@ -2017,7 +2016,6 @@ void a64_hybrid_u8qa_dot_4x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "122:" // Exit - : [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", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" -- cgit v1.2.1