diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp | 1120 |
1 files changed, 559 insertions, 561 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp index 5fb71c95b7..dbd45460e8 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp @@ -92,7 +92,6 @@ void a64_hybrid_fp32_mla_4x24 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x4\n" "bge 124f\n" @@ -223,11 +222,11 @@ void a64_hybrid_fp32_mla_4x24 ( "19:" // 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, 20f\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, 21f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -246,126 +245,126 @@ void a64_hybrid_fp32_mla_4x24 ( "blt 23f\n" "22:" // Height 1: Multiply loop: Main loop head "fmla v8.4s, v4.4s, v0.s[0]\n" - "ldr q4, [x28, #0x40]\n" + "ldr q19, [x28, #0x40]\n" "fmla v9.4s, v5.4s, v0.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr q18, [x28, #0x50]\n" "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x28, #0x60]\n" + "ldr q17, [x28, #0x60]\n" "fmla v11.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x28, #0x70]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "ldr q4, [x28, #0x80]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "ldr q5, [x28, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x28, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x28, #0xb0]\n" - "fmla v10.4s, v4.4s, v0.s[1]\n" - "ldr q4, [x28, #0xc0]\n" - "fmla v11.4s, v5.4s, v0.s[1]\n" - "ldr q5, [x28, #0xd0]\n" - "fmla v12.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x28, #0xe0]\n" - "fmla v13.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x28, #0xf0]\n" - "fmla v8.4s, v4.4s, v0.s[2]\n" - "ldr q4, [x28, #0x100]\n" - "fmla v9.4s, v5.4s, v0.s[2]\n" - "ldr q5, [x28, #0x110]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x28, #0x120]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x28, #0x130]\n" - "fmla v12.4s, v4.4s, v0.s[2]\n" - "ldr q4, [x28, #0x140]\n" - "fmla v13.4s, v5.4s, v0.s[2]\n" - "ldr q5, [x28, #0x150]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "ldr q6, [x28, #0x160]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "ldr q7, [x28, #0x170]\n" + "ldr q16, [x28, #0x70]\n" + "fmla v12.4s, v19.4s, v0.s[0]\n" + "ldr q19, [x28, #0x80]\n" + "fmla v13.4s, v18.4s, v0.s[0]\n" + "ldr q18, [x28, #0x90]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x28, #0xa0]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x28, #0xb0]\n" + "fmla v10.4s, v19.4s, v0.s[1]\n" + "ldr q19, [x28, #0xc0]\n" + "fmla v11.4s, v18.4s, v0.s[1]\n" + "ldr q18, [x28, #0xd0]\n" + "fmla v12.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x28, #0xe0]\n" + "fmla v13.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x28, #0xf0]\n" + "fmla v8.4s, v19.4s, v0.s[2]\n" + "ldr q19, [x28, #0x100]\n" + "fmla v9.4s, v18.4s, v0.s[2]\n" + "ldr q18, [x28, #0x110]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x28, #0x120]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x28, #0x130]\n" + "fmla v12.4s, v19.4s, v0.s[2]\n" + "ldr q19, [x28, #0x140]\n" + "fmla v13.4s, v18.4s, v0.s[2]\n" + "ldr q18, [x28, #0x150]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "ldr q17, [x28, #0x160]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "ldr q16, [x28, #0x170]\n" "sub x25, x25, #0x4\n" "add x24, x24, #0x10\n" - "fmla v10.4s, v4.4s, v0.s[3]\n" - "fmla v11.4s, v5.4s, v0.s[3]\n" + "fmla v10.4s, v19.4s, v0.s[3]\n" + "fmla v11.4s, v18.4s, v0.s[3]\n" "cmp x25, #0x8\n" "add x28, x28, #0x180\n" "ldr q4, [x28, #0x0]\n" "ldr q5, [x28, #0x10]\n" - "fmla v12.4s, v6.4s, v0.s[3]\n" + "fmla v12.4s, v17.4s, v0.s[3]\n" "ldr q6, [x28, #0x20]\n" - "fmla v13.4s, v7.4s, v0.s[3]\n" + "fmla v13.4s, v16.4s, v0.s[3]\n" "ldr q0, [x24, #0x0]\n" "ldr q7, [x28, #0x30]\n" "prfm pldl1keep, [x24, #0x80]\n" "bge 22b\n" "23:" // Height 1: Multiply loop: Single iteration only "fmla v8.4s, v4.4s, v0.s[0]\n" - "ldr q4, [x28, #0x40]\n" + "ldr q19, [x28, #0x40]\n" "fmla v9.4s, v5.4s, v0.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr q18, [x28, #0x50]\n" "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x28, #0x60]\n" + "ldr q17, [x28, #0x60]\n" "fmla v11.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x28, #0x70]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "ldr q4, [x28, #0x80]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "ldr q5, [x28, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x28, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x28, #0xb0]\n" - "fmla v10.4s, v4.4s, v0.s[1]\n" - "ldr q4, [x28, #0xc0]\n" - "fmla v11.4s, v5.4s, v0.s[1]\n" - "ldr q5, [x28, #0xd0]\n" - "fmla v12.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x28, #0xe0]\n" - "fmla v13.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x28, #0xf0]\n" - "fmla v8.4s, v4.4s, v0.s[2]\n" - "ldr q4, [x28, #0x100]\n" - "fmla v9.4s, v5.4s, v0.s[2]\n" - "ldr q5, [x28, #0x110]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x28, #0x120]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x28, #0x130]\n" - "fmla v12.4s, v4.4s, v0.s[2]\n" - "ldr q4, [x28, #0x140]\n" - "fmla v13.4s, v5.4s, v0.s[2]\n" - "ldr q5, [x28, #0x150]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "ldr q6, [x28, #0x160]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "ldr q7, [x28, #0x170]\n" + "ldr q16, [x28, #0x70]\n" + "fmla v12.4s, v19.4s, v0.s[0]\n" + "ldr q19, [x28, #0x80]\n" + "fmla v13.4s, v18.4s, v0.s[0]\n" + "ldr q18, [x28, #0x90]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x28, #0xa0]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x28, #0xb0]\n" + "fmla v10.4s, v19.4s, v0.s[1]\n" + "ldr q19, [x28, #0xc0]\n" + "fmla v11.4s, v18.4s, v0.s[1]\n" + "ldr q18, [x28, #0xd0]\n" + "fmla v12.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x28, #0xe0]\n" + "fmla v13.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x28, #0xf0]\n" + "fmla v8.4s, v19.4s, v0.s[2]\n" + "ldr q19, [x28, #0x100]\n" + "fmla v9.4s, v18.4s, v0.s[2]\n" + "ldr q18, [x28, #0x110]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x28, #0x120]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x28, #0x130]\n" + "fmla v12.4s, v19.4s, v0.s[2]\n" + "ldr q19, [x28, #0x140]\n" + "fmla v13.4s, v18.4s, v0.s[2]\n" + "ldr q18, [x28, #0x150]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "ldr q17, [x28, #0x160]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "ldr q16, [x28, #0x170]\n" "add x24, x24, #0x10\n" "sub x25, x25, #0x4\n" - "fmla v10.4s, v4.4s, v0.s[3]\n" - "fmla v11.4s, v5.4s, v0.s[3]\n" + "fmla v10.4s, v19.4s, v0.s[3]\n" + "fmla v11.4s, v18.4s, v0.s[3]\n" "prfm pldl1keep, [x24, #0x80]\n" "add x28, x28, #0x180\n" - "fmla v12.4s, v6.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v0.s[3]\n" + "fmla v12.4s, v17.4s, v0.s[3]\n" + "fmla v13.4s, v16.4s, v0.s[3]\n" "24:" // Height 1: Multiply loop: Main loop skip "cbz x25, 26f\n" "25:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x24], #0x4\n" - "ldr q4, [x28, #0x0]\n" - "fmla v8.4s, v4.4s, v0.s[0]\n" + "ldr s18, [x24], #0x4\n" + "ldr q16, [x28, #0x0]\n" + "fmla v8.4s, v16.4s, v18.s[0]\n" "sub x25, x25, #0x1\n" - "ldr q5, [x28, #0x10]\n" - "ldr q6, [x28, #0x20]\n" - "fmla v9.4s, v5.4s, v0.s[0]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q7, [x28, #0x30]\n" - "ldr q4, [x28, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "ldr q5, [x28, #0x50]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" + "ldr q17, [x28, #0x10]\n" + "ldr q16, [x28, #0x20]\n" + "fmla v9.4s, v17.4s, v18.s[0]\n" + "fmla v10.4s, v16.4s, v18.s[0]\n" + "ldr q17, [x28, #0x30]\n" + "ldr q16, [x28, #0x40]\n" + "fmla v11.4s, v17.4s, v18.s[0]\n" + "fmla v12.4s, v16.4s, v18.s[0]\n" + "ldr q16, [x28, #0x50]\n" + "fmla v13.4s, v16.4s, v18.s[0]\n" "add x28, x28, #0x60\n" "cbnz x25, 25b\n" "26:" // Height 1: Multiply loop: No odd multiplies @@ -376,21 +375,21 @@ void a64_hybrid_fp32_mla_4x24 ( "prfm pstl1keep, [x27, #0x0]\n" "tbz %x[flags], #1, 27f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" + "ld1r { v16.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v17.4s\n" + "fmin v9.4s, v9.4s, v17.4s\n" + "fmin v10.4s, v10.4s, v17.4s\n" + "fmin v11.4s, v11.4s, v17.4s\n" + "fmin v12.4s, v12.4s, v17.4s\n" + "fmin v13.4s, v13.4s, v17.4s\n" + "fmax v8.4s, v8.4s, v16.4s\n" + "fmax v9.4s, v9.4s, v16.4s\n" + "fmax v10.4s, v10.4s, v16.4s\n" + "fmax v11.4s, v11.4s, v16.4s\n" + "fmax v12.4s, v12.4s, v16.4s\n" + "fmax v13.4s, v13.4s, v16.4s\n" "27:" // Height 1: No activation "cmp x9, #0x18\n" "bge 40f\n" @@ -651,12 +650,12 @@ void a64_hybrid_fp32_mla_4x24 ( "60:" // 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, 61f\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, 62f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -664,7 +663,7 @@ void a64_hybrid_fp32_mla_4x24 ( "b 62f\n" "61:" // Height 2: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "62:" // Height 2: input setup done "cmp x25, #0x4\n" "blt 65f\n" @@ -679,186 +678,186 @@ void a64_hybrid_fp32_mla_4x24 ( "63:" // Height 2: Multiply loop: Main loop head "fmla v8.4s, v4.4s, v0.s[0]\n" "fmla v14.4s, v4.4s, v1.s[0]\n" - "ldr q4, [x28, #0x40]\n" + "ldr q23, [x28, #0x40]\n" "sub x25, x25, #0x4\n" "fmla v9.4s, v5.4s, v0.s[0]\n" "fmla v15.4s, v5.4s, v1.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr q22, [x28, #0x50]\n" "add x24, x24, #0x10\n" "fmla v10.4s, v6.4s, v0.s[0]\n" "fmla v16.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x28, #0x60]\n" + "ldr q21, [x28, #0x60]\n" "add x23, x23, #0x10\n" "fmla v11.4s, v7.4s, v0.s[0]\n" "fmla v17.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x28, #0x70]\n" + "ldr q20, [x28, #0x70]\n" "cmp x25, #0x8\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "ldr q4, [x28, #0x80]\n" + "fmla v12.4s, v23.4s, v0.s[0]\n" + "fmla v18.4s, v23.4s, v1.s[0]\n" + "ldr q23, [x28, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" - "ldr q5, [x28, #0x90]\n" + "fmla v13.4s, v22.4s, v0.s[0]\n" + "fmla v19.4s, v22.4s, v1.s[0]\n" + "ldr q22, [x28, #0x90]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x28, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x28, #0xb0]\n" - "fmla v10.4s, v4.4s, v0.s[1]\n" - "fmla v16.4s, v4.4s, v1.s[1]\n" - "ldr q4, [x28, #0xc0]\n" - "fmla v11.4s, v5.4s, v0.s[1]\n" - "fmla v17.4s, v5.4s, v1.s[1]\n" - "ldr q5, [x28, #0xd0]\n" - "fmla v12.4s, v6.4s, v0.s[1]\n" - "fmla v18.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x28, #0xe0]\n" - "fmla v13.4s, v7.4s, v0.s[1]\n" - "fmla v19.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x28, #0xf0]\n" - "fmla v8.4s, v4.4s, v0.s[2]\n" - "fmla v14.4s, v4.4s, v1.s[2]\n" - "ldr q4, [x28, #0x100]\n" - "fmla v9.4s, v5.4s, v0.s[2]\n" - "fmla v15.4s, v5.4s, v1.s[2]\n" - "ldr q5, [x28, #0x110]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v16.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x28, #0x120]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v17.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x28, #0x130]\n" - "fmla v12.4s, v4.4s, v0.s[2]\n" - "fmla v18.4s, v4.4s, v1.s[2]\n" - "ldr q4, [x28, #0x140]\n" - "fmla v13.4s, v5.4s, v0.s[2]\n" - "fmla v19.4s, v5.4s, v1.s[2]\n" - "ldr q5, [x28, #0x150]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "ldr q6, [x28, #0x160]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x28, #0x170]\n" + "fmla v8.4s, v21.4s, v0.s[1]\n" + "fmla v14.4s, v21.4s, v1.s[1]\n" + "ldr q21, [x28, #0xa0]\n" + "fmla v9.4s, v20.4s, v0.s[1]\n" + "fmla v15.4s, v20.4s, v1.s[1]\n" + "ldr q20, [x28, #0xb0]\n" + "fmla v10.4s, v23.4s, v0.s[1]\n" + "fmla v16.4s, v23.4s, v1.s[1]\n" + "ldr q23, [x28, #0xc0]\n" + "fmla v11.4s, v22.4s, v0.s[1]\n" + "fmla v17.4s, v22.4s, v1.s[1]\n" + "ldr q22, [x28, #0xd0]\n" + "fmla v12.4s, v21.4s, v0.s[1]\n" + "fmla v18.4s, v21.4s, v1.s[1]\n" + "ldr q21, [x28, #0xe0]\n" + "fmla v13.4s, v20.4s, v0.s[1]\n" + "fmla v19.4s, v20.4s, v1.s[1]\n" + "ldr q20, [x28, #0xf0]\n" + "fmla v8.4s, v23.4s, v0.s[2]\n" + "fmla v14.4s, v23.4s, v1.s[2]\n" + "ldr q23, [x28, #0x100]\n" + "fmla v9.4s, v22.4s, v0.s[2]\n" + "fmla v15.4s, v22.4s, v1.s[2]\n" + "ldr q22, [x28, #0x110]\n" + "fmla v10.4s, v21.4s, v0.s[2]\n" + "fmla v16.4s, v21.4s, v1.s[2]\n" + "ldr q21, [x28, #0x120]\n" + "fmla v11.4s, v20.4s, v0.s[2]\n" + "fmla v17.4s, v20.4s, v1.s[2]\n" + "ldr q20, [x28, #0x130]\n" + "fmla v12.4s, v23.4s, v0.s[2]\n" + "fmla v18.4s, v23.4s, v1.s[2]\n" + "ldr q23, [x28, #0x140]\n" + "fmla v13.4s, v22.4s, v0.s[2]\n" + "fmla v19.4s, v22.4s, v1.s[2]\n" + "ldr q22, [x28, #0x150]\n" + "fmla v8.4s, v21.4s, v0.s[3]\n" + "fmla v14.4s, v21.4s, v1.s[3]\n" + "ldr q21, [x28, #0x160]\n" + "fmla v9.4s, v20.4s, v0.s[3]\n" + "fmla v15.4s, v20.4s, v1.s[3]\n" + "ldr q20, [x28, #0x170]\n" "add x28, x28, #0x180\n" - "fmla v10.4s, v4.4s, v0.s[3]\n" - "fmla v16.4s, v4.4s, v1.s[3]\n" + "fmla v10.4s, v23.4s, v0.s[3]\n" + "fmla v16.4s, v23.4s, v1.s[3]\n" "ldr q4, [x28, #0x0]\n" - "fmla v11.4s, v5.4s, v0.s[3]\n" - "fmla v17.4s, v5.4s, v1.s[3]\n" + "fmla v11.4s, v22.4s, v0.s[3]\n" + "fmla v17.4s, v22.4s, v1.s[3]\n" "ldr q5, [x28, #0x10]\n" - "fmla v12.4s, v6.4s, v0.s[3]\n" - "fmla v18.4s, v6.4s, v1.s[3]\n" + "fmla v12.4s, v21.4s, v0.s[3]\n" + "fmla v18.4s, v21.4s, v1.s[3]\n" "ldr q6, [x28, #0x20]\n" - "fmla v13.4s, v7.4s, v0.s[3]\n" + "fmla v13.4s, v20.4s, v0.s[3]\n" "ldr q0, [x24, #0x0]\n" - "fmla v19.4s, v7.4s, v1.s[3]\n" + "fmla v19.4s, v20.4s, v1.s[3]\n" "ldr q1, [x23, #0x0]\n" "ldr q7, [x28, #0x30]\n" "bge 63b\n" "64:" // Height 2: Multiply loop: Single iteration only "fmla v8.4s, v4.4s, v0.s[0]\n" "fmla v14.4s, v4.4s, v1.s[0]\n" - "ldr q4, [x28, #0x40]\n" + "ldr q23, [x28, #0x40]\n" "add x24, x24, #0x10\n" "fmla v9.4s, v5.4s, v0.s[0]\n" "fmla v15.4s, v5.4s, v1.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr q22, [x28, #0x50]\n" "add x23, x23, #0x10\n" "fmla v10.4s, v6.4s, v0.s[0]\n" "fmla v16.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x28, #0x60]\n" + "ldr q21, [x28, #0x60]\n" "sub x25, x25, #0x4\n" "fmla v11.4s, v7.4s, v0.s[0]\n" "fmla v17.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x28, #0x70]\n" + "ldr q20, [x28, #0x70]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "ldr q4, [x28, #0x80]\n" + "fmla v12.4s, v23.4s, v0.s[0]\n" + "fmla v18.4s, v23.4s, v1.s[0]\n" + "ldr q23, [x28, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" - "ldr q5, [x28, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x28, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x28, #0xb0]\n" - "fmla v10.4s, v4.4s, v0.s[1]\n" - "fmla v16.4s, v4.4s, v1.s[1]\n" - "ldr q4, [x28, #0xc0]\n" - "fmla v11.4s, v5.4s, v0.s[1]\n" - "fmla v17.4s, v5.4s, v1.s[1]\n" - "ldr q5, [x28, #0xd0]\n" - "fmla v12.4s, v6.4s, v0.s[1]\n" - "fmla v18.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x28, #0xe0]\n" - "fmla v13.4s, v7.4s, v0.s[1]\n" - "fmla v19.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x28, #0xf0]\n" - "fmla v8.4s, v4.4s, v0.s[2]\n" - "fmla v14.4s, v4.4s, v1.s[2]\n" - "ldr q4, [x28, #0x100]\n" - "fmla v9.4s, v5.4s, v0.s[2]\n" - "fmla v15.4s, v5.4s, v1.s[2]\n" - "ldr q5, [x28, #0x110]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v16.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x28, #0x120]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v17.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x28, #0x130]\n" - "fmla v12.4s, v4.4s, v0.s[2]\n" - "fmla v18.4s, v4.4s, v1.s[2]\n" - "ldr q4, [x28, #0x140]\n" - "fmla v13.4s, v5.4s, v0.s[2]\n" - "fmla v19.4s, v5.4s, v1.s[2]\n" - "ldr q5, [x28, #0x150]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "ldr q6, [x28, #0x160]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x28, #0x170]\n" + "fmla v13.4s, v22.4s, v0.s[0]\n" + "fmla v19.4s, v22.4s, v1.s[0]\n" + "ldr q22, [x28, #0x90]\n" + "fmla v8.4s, v21.4s, v0.s[1]\n" + "fmla v14.4s, v21.4s, v1.s[1]\n" + "ldr q21, [x28, #0xa0]\n" + "fmla v9.4s, v20.4s, v0.s[1]\n" + "fmla v15.4s, v20.4s, v1.s[1]\n" + "ldr q20, [x28, #0xb0]\n" + "fmla v10.4s, v23.4s, v0.s[1]\n" + "fmla v16.4s, v23.4s, v1.s[1]\n" + "ldr q23, [x28, #0xc0]\n" + "fmla v11.4s, v22.4s, v0.s[1]\n" + "fmla v17.4s, v22.4s, v1.s[1]\n" + "ldr q22, [x28, #0xd0]\n" + "fmla v12.4s, v21.4s, v0.s[1]\n" + "fmla v18.4s, v21.4s, v1.s[1]\n" + "ldr q21, [x28, #0xe0]\n" + "fmla v13.4s, v20.4s, v0.s[1]\n" + "fmla v19.4s, v20.4s, v1.s[1]\n" + "ldr q20, [x28, #0xf0]\n" + "fmla v8.4s, v23.4s, v0.s[2]\n" + "fmla v14.4s, v23.4s, v1.s[2]\n" + "ldr q23, [x28, #0x100]\n" + "fmla v9.4s, v22.4s, v0.s[2]\n" + "fmla v15.4s, v22.4s, v1.s[2]\n" + "ldr q22, [x28, #0x110]\n" + "fmla v10.4s, v21.4s, v0.s[2]\n" + "fmla v16.4s, v21.4s, v1.s[2]\n" + "ldr q21, [x28, #0x120]\n" + "fmla v11.4s, v20.4s, v0.s[2]\n" + "fmla v17.4s, v20.4s, v1.s[2]\n" + "ldr q20, [x28, #0x130]\n" + "fmla v12.4s, v23.4s, v0.s[2]\n" + "fmla v18.4s, v23.4s, v1.s[2]\n" + "ldr q23, [x28, #0x140]\n" + "fmla v13.4s, v22.4s, v0.s[2]\n" + "fmla v19.4s, v22.4s, v1.s[2]\n" + "ldr q22, [x28, #0x150]\n" + "fmla v8.4s, v21.4s, v0.s[3]\n" + "fmla v14.4s, v21.4s, v1.s[3]\n" + "ldr q21, [x28, #0x160]\n" + "fmla v9.4s, v20.4s, v0.s[3]\n" + "fmla v15.4s, v20.4s, v1.s[3]\n" + "ldr q20, [x28, #0x170]\n" "add x28, x28, #0x180\n" - "fmla v10.4s, v4.4s, v0.s[3]\n" - "fmla v16.4s, v4.4s, v1.s[3]\n" - "fmla v11.4s, v5.4s, v0.s[3]\n" - "fmla v17.4s, v5.4s, v1.s[3]\n" - "fmla v12.4s, v6.4s, v0.s[3]\n" - "fmla v18.4s, v6.4s, v1.s[3]\n" - "fmla v13.4s, v7.4s, v0.s[3]\n" - "fmla v19.4s, v7.4s, v1.s[3]\n" + "fmla v10.4s, v23.4s, v0.s[3]\n" + "fmla v16.4s, v23.4s, v1.s[3]\n" + "fmla v11.4s, v22.4s, v0.s[3]\n" + "fmla v17.4s, v22.4s, v1.s[3]\n" + "fmla v12.4s, v21.4s, v0.s[3]\n" + "fmla v18.4s, v21.4s, v1.s[3]\n" + "fmla v13.4s, v20.4s, v0.s[3]\n" + "fmla v19.4s, v20.4s, v1.s[3]\n" "65:" // Height 2: Multiply loop: Main loop skip "cbz x25, 67f\n" "66:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x24], #0x4\n" - "ldr s1, [x23], #0x4\n" + "ldr s25, [x24], #0x4\n" + "ldr s24, [x23], #0x4\n" "sub x25, x25, #0x1\n" - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" - "fmla v8.4s, v4.4s, v0.s[0]\n" - "fmla v14.4s, v4.4s, v1.s[0]\n" - "ldr q6, [x28, #0x20]\n" - "ldr q7, [x28, #0x30]\n" - "fmla v9.4s, v5.4s, v0.s[0]\n" - "fmla v15.4s, v5.4s, v1.s[0]\n" - "ldr q4, [x28, #0x40]\n" - "ldr q5, [x28, #0x50]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v16.4s, v6.4s, v1.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v17.4s, v7.4s, v1.s[0]\n" + "ldr q21, [x28, #0x0]\n" + "ldr q20, [x28, #0x10]\n" + "fmla v8.4s, v21.4s, v25.s[0]\n" + "fmla v14.4s, v21.4s, v24.s[0]\n" + "ldr q23, [x28, #0x20]\n" + "ldr q22, [x28, #0x30]\n" + "fmla v9.4s, v20.4s, v25.s[0]\n" + "fmla v15.4s, v20.4s, v24.s[0]\n" + "ldr q21, [x28, #0x40]\n" + "ldr q20, [x28, #0x50]\n" + "fmla v10.4s, v23.4s, v25.s[0]\n" + "fmla v16.4s, v23.4s, v24.s[0]\n" + "fmla v11.4s, v22.4s, v25.s[0]\n" + "fmla v17.4s, v22.4s, v24.s[0]\n" "add x28, x28, #0x60\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" + "fmla v12.4s, v21.4s, v25.s[0]\n" + "fmla v18.4s, v21.4s, v24.s[0]\n" + "fmla v13.4s, v20.4s, v25.s[0]\n" + "fmla v19.4s, v20.4s, v24.s[0]\n" "cbnz x25, 66b\n" "67:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -871,33 +870,33 @@ void a64_hybrid_fp32_mla_4x24 ( "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 68f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v21.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" + "ld1r { v20.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v21.4s\n" + "fmin v9.4s, v9.4s, v21.4s\n" + "fmin v10.4s, v10.4s, v21.4s\n" + "fmin v11.4s, v11.4s, v21.4s\n" + "fmin v12.4s, v12.4s, v21.4s\n" + "fmin v13.4s, v13.4s, v21.4s\n" + "fmin v14.4s, v14.4s, v21.4s\n" + "fmin v15.4s, v15.4s, v21.4s\n" + "fmin v16.4s, v16.4s, v21.4s\n" + "fmin v17.4s, v17.4s, v21.4s\n" + "fmin v18.4s, v18.4s, v21.4s\n" + "fmin v19.4s, v19.4s, v21.4s\n" + "fmax v8.4s, v8.4s, v20.4s\n" + "fmax v9.4s, v9.4s, v20.4s\n" + "fmax v10.4s, v10.4s, v20.4s\n" + "fmax v11.4s, v11.4s, v20.4s\n" + "fmax v12.4s, v12.4s, v20.4s\n" + "fmax v13.4s, v13.4s, v20.4s\n" + "fmax v14.4s, v14.4s, v20.4s\n" + "fmax v15.4s, v15.4s, v20.4s\n" + "fmax v16.4s, v16.4s, v20.4s\n" + "fmax v17.4s, v17.4s, v20.4s\n" + "fmax v18.4s, v18.4s, v20.4s\n" + "fmax v19.4s, v19.4s, v20.4s\n" "68:" // Height 2: No activation "cmp x9, #0x18\n" "bge 81f\n" @@ -1237,13 +1236,13 @@ void a64_hybrid_fp32_mla_4x24 ( "101:" // 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, 102f\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, 103f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -1252,8 +1251,8 @@ void a64_hybrid_fp32_mla_4x24 ( "b 103f\n" "102:" // Height 3: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "103:" // Height 3: input setup done "cmp x25, #0x4\n" "blt 106f\n" @@ -1272,107 +1271,107 @@ void a64_hybrid_fp32_mla_4x24 ( "sub x25, x25, #0x4\n" "add x24, x24, #0x10\n" "fmla v20.4s, v4.4s, v2.s[0]\n" - "ldr q4, [x28, #0x40]\n" + "ldr q29, [x28, #0x40]\n" "fmla v9.4s, v5.4s, v0.s[0]\n" "add x23, x23, #0x10\n" "fmla v15.4s, v5.4s, v1.s[0]\n" "fmla v21.4s, v5.4s, v2.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr q28, [x28, #0x50]\n" "add x22, x22, #0x10\n" "fmla v10.4s, v6.4s, v0.s[0]\n" "fmla v16.4s, v6.4s, v1.s[0]\n" "cmp x25, #0x8\n" "prfm pldl1keep, [x24, #0x80]\n" "fmla v22.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x28, #0x60]\n" + "ldr q27, [x28, #0x60]\n" "fmla v11.4s, v7.4s, v0.s[0]\n" "prfm pldl1keep, [x23, #0x80]\n" "fmla v17.4s, v7.4s, v1.s[0]\n" "fmla v23.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x28, #0x70]\n" + "ldr q26, [x28, #0x70]\n" "prfm pldl1keep, [x22, #0x80]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "fmla v24.4s, v4.4s, v2.s[0]\n" - "ldr q4, [x28, #0x80]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" - "fmla v25.4s, v5.4s, v2.s[0]\n" - "ldr q5, [x28, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v20.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x28, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v21.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x28, #0xb0]\n" - "fmla v10.4s, v4.4s, v0.s[1]\n" - "fmla v16.4s, v4.4s, v1.s[1]\n" - "fmla v22.4s, v4.4s, v2.s[1]\n" - "ldr q4, [x28, #0xc0]\n" - "fmla v11.4s, v5.4s, v0.s[1]\n" - "fmla v17.4s, v5.4s, v1.s[1]\n" - "fmla v23.4s, v5.4s, v2.s[1]\n" - "ldr q5, [x28, #0xd0]\n" - "fmla v12.4s, v6.4s, v0.s[1]\n" - "fmla v18.4s, v6.4s, v1.s[1]\n" - "fmla v24.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x28, #0xe0]\n" - "fmla v13.4s, v7.4s, v0.s[1]\n" - "fmla v19.4s, v7.4s, v1.s[1]\n" - "fmla v25.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x28, #0xf0]\n" - "fmla v8.4s, v4.4s, v0.s[2]\n" - "fmla v14.4s, v4.4s, v1.s[2]\n" - "fmla v20.4s, v4.4s, v2.s[2]\n" - "ldr q4, [x28, #0x100]\n" - "fmla v9.4s, v5.4s, v0.s[2]\n" - "fmla v15.4s, v5.4s, v1.s[2]\n" - "fmla v21.4s, v5.4s, v2.s[2]\n" - "ldr q5, [x28, #0x110]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v16.4s, v6.4s, v1.s[2]\n" - "fmla v22.4s, v6.4s, v2.s[2]\n" - "ldr q6, [x28, #0x120]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v17.4s, v7.4s, v1.s[2]\n" - "fmla v23.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x28, #0x130]\n" - "fmla v12.4s, v4.4s, v0.s[2]\n" - "fmla v18.4s, v4.4s, v1.s[2]\n" - "fmla v24.4s, v4.4s, v2.s[2]\n" - "ldr q4, [x28, #0x140]\n" - "fmla v13.4s, v5.4s, v0.s[2]\n" - "fmla v19.4s, v5.4s, v1.s[2]\n" - "fmla v25.4s, v5.4s, v2.s[2]\n" - "ldr q5, [x28, #0x150]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v20.4s, v6.4s, v2.s[3]\n" - "ldr q6, [x28, #0x160]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "fmla v21.4s, v7.4s, v2.s[3]\n" - "ldr q7, [x28, #0x170]\n" + "fmla v12.4s, v29.4s, v0.s[0]\n" + "fmla v18.4s, v29.4s, v1.s[0]\n" + "fmla v24.4s, v29.4s, v2.s[0]\n" + "ldr q29, [x28, #0x80]\n" + "fmla v13.4s, v28.4s, v0.s[0]\n" + "fmla v19.4s, v28.4s, v1.s[0]\n" + "fmla v25.4s, v28.4s, v2.s[0]\n" + "ldr q28, [x28, #0x90]\n" + "fmla v8.4s, v27.4s, v0.s[1]\n" + "fmla v14.4s, v27.4s, v1.s[1]\n" + "fmla v20.4s, v27.4s, v2.s[1]\n" + "ldr q27, [x28, #0xa0]\n" + "fmla v9.4s, v26.4s, v0.s[1]\n" + "fmla v15.4s, v26.4s, v1.s[1]\n" + "fmla v21.4s, v26.4s, v2.s[1]\n" + "ldr q26, [x28, #0xb0]\n" + "fmla v10.4s, v29.4s, v0.s[1]\n" + "fmla v16.4s, v29.4s, v1.s[1]\n" + "fmla v22.4s, v29.4s, v2.s[1]\n" + "ldr q29, [x28, #0xc0]\n" + "fmla v11.4s, v28.4s, v0.s[1]\n" + "fmla v17.4s, v28.4s, v1.s[1]\n" + "fmla v23.4s, v28.4s, v2.s[1]\n" + "ldr q28, [x28, #0xd0]\n" + "fmla v12.4s, v27.4s, v0.s[1]\n" + "fmla v18.4s, v27.4s, v1.s[1]\n" + "fmla v24.4s, v27.4s, v2.s[1]\n" + "ldr q27, [x28, #0xe0]\n" + "fmla v13.4s, v26.4s, v0.s[1]\n" + "fmla v19.4s, v26.4s, v1.s[1]\n" + "fmla v25.4s, v26.4s, v2.s[1]\n" + "ldr q26, [x28, #0xf0]\n" + "fmla v8.4s, v29.4s, v0.s[2]\n" + "fmla v14.4s, v29.4s, v1.s[2]\n" + "fmla v20.4s, v29.4s, v2.s[2]\n" + "ldr q29, [x28, #0x100]\n" + "fmla v9.4s, v28.4s, v0.s[2]\n" + "fmla v15.4s, v28.4s, v1.s[2]\n" + "fmla v21.4s, v28.4s, v2.s[2]\n" + "ldr q28, [x28, #0x110]\n" + "fmla v10.4s, v27.4s, v0.s[2]\n" + "fmla v16.4s, v27.4s, v1.s[2]\n" + "fmla v22.4s, v27.4s, v2.s[2]\n" + "ldr q27, [x28, #0x120]\n" + "fmla v11.4s, v26.4s, v0.s[2]\n" + "fmla v17.4s, v26.4s, v1.s[2]\n" + "fmla v23.4s, v26.4s, v2.s[2]\n" + "ldr q26, [x28, #0x130]\n" + "fmla v12.4s, v29.4s, v0.s[2]\n" + "fmla v18.4s, v29.4s, v1.s[2]\n" + "fmla v24.4s, v29.4s, v2.s[2]\n" + "ldr q29, [x28, #0x140]\n" + "fmla v13.4s, v28.4s, v0.s[2]\n" + "fmla v19.4s, v28.4s, v1.s[2]\n" + "fmla v25.4s, v28.4s, v2.s[2]\n" + "ldr q28, [x28, #0x150]\n" + "fmla v8.4s, v27.4s, v0.s[3]\n" + "fmla v14.4s, v27.4s, v1.s[3]\n" + "fmla v20.4s, v27.4s, v2.s[3]\n" + "ldr q27, [x28, #0x160]\n" + "fmla v9.4s, v26.4s, v0.s[3]\n" + "fmla v15.4s, v26.4s, v1.s[3]\n" + "fmla v21.4s, v26.4s, v2.s[3]\n" + "ldr q26, [x28, #0x170]\n" "add x28, x28, #0x180\n" - "fmla v10.4s, v4.4s, v0.s[3]\n" - "fmla v16.4s, v4.4s, v1.s[3]\n" - "fmla v22.4s, v4.4s, v2.s[3]\n" + "fmla v10.4s, v29.4s, v0.s[3]\n" + "fmla v16.4s, v29.4s, v1.s[3]\n" + "fmla v22.4s, v29.4s, v2.s[3]\n" "ldr q4, [x28, #0x0]\n" - "fmla v11.4s, v5.4s, v0.s[3]\n" - "fmla v17.4s, v5.4s, v1.s[3]\n" - "fmla v23.4s, v5.4s, v2.s[3]\n" + "fmla v11.4s, v28.4s, v0.s[3]\n" + "fmla v17.4s, v28.4s, v1.s[3]\n" + "fmla v23.4s, v28.4s, v2.s[3]\n" "ldr q5, [x28, #0x10]\n" - "fmla v12.4s, v6.4s, v0.s[3]\n" - "fmla v18.4s, v6.4s, v1.s[3]\n" - "fmla v24.4s, v6.4s, v2.s[3]\n" + "fmla v12.4s, v27.4s, v0.s[3]\n" + "fmla v18.4s, v27.4s, v1.s[3]\n" + "fmla v24.4s, v27.4s, v2.s[3]\n" "ldr q6, [x28, #0x20]\n" - "fmla v13.4s, v7.4s, v0.s[3]\n" + "fmla v13.4s, v26.4s, v0.s[3]\n" "ldr q0, [x24, #0x0]\n" - "fmla v19.4s, v7.4s, v1.s[3]\n" + "fmla v19.4s, v26.4s, v1.s[3]\n" "ldr q1, [x23, #0x0]\n" - "fmla v25.4s, v7.4s, v2.s[3]\n" + "fmla v25.4s, v26.4s, v2.s[3]\n" "ldr q2, [x22, #0x0]\n" "ldr q7, [x28, #0x30]\n" "bge 104b\n" @@ -1382,133 +1381,133 @@ void a64_hybrid_fp32_mla_4x24 ( "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" "fmla v20.4s, v4.4s, v2.s[0]\n" - "ldr q4, [x28, #0x40]\n" + "ldr q29, [x28, #0x40]\n" "fmla v9.4s, v5.4s, v0.s[0]\n" "add x22, x22, #0x10\n" "fmla v15.4s, v5.4s, v1.s[0]\n" "fmla v21.4s, v5.4s, v2.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr q28, [x28, #0x50]\n" "sub x25, x25, #0x4\n" "fmla v10.4s, v6.4s, v0.s[0]\n" "fmla v16.4s, v6.4s, v1.s[0]\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" "fmla v22.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x28, #0x60]\n" + "ldr q27, [x28, #0x60]\n" "fmla v11.4s, v7.4s, v0.s[0]\n" "prfm pldl1keep, [x22, #0x80]\n" "fmla v17.4s, v7.4s, v1.s[0]\n" "fmla v23.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x28, #0x70]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "fmla v24.4s, v4.4s, v2.s[0]\n" - "ldr q4, [x28, #0x80]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" - "fmla v25.4s, v5.4s, v2.s[0]\n" - "ldr q5, [x28, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v20.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x28, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v21.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x28, #0xb0]\n" - "fmla v10.4s, v4.4s, v0.s[1]\n" - "fmla v16.4s, v4.4s, v1.s[1]\n" - "fmla v22.4s, v4.4s, v2.s[1]\n" - "ldr q4, [x28, #0xc0]\n" - "fmla v11.4s, v5.4s, v0.s[1]\n" - "fmla v17.4s, v5.4s, v1.s[1]\n" - "fmla v23.4s, v5.4s, v2.s[1]\n" - "ldr q5, [x28, #0xd0]\n" - "fmla v12.4s, v6.4s, v0.s[1]\n" - "fmla v18.4s, v6.4s, v1.s[1]\n" - "fmla v24.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x28, #0xe0]\n" - "fmla v13.4s, v7.4s, v0.s[1]\n" - "fmla v19.4s, v7.4s, v1.s[1]\n" - "fmla v25.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x28, #0xf0]\n" - "fmla v8.4s, v4.4s, v0.s[2]\n" - "fmla v14.4s, v4.4s, v1.s[2]\n" - "fmla v20.4s, v4.4s, v2.s[2]\n" - "ldr q4, [x28, #0x100]\n" - "fmla v9.4s, v5.4s, v0.s[2]\n" - "fmla v15.4s, v5.4s, v1.s[2]\n" - "fmla v21.4s, v5.4s, v2.s[2]\n" - "ldr q5, [x28, #0x110]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v16.4s, v6.4s, v1.s[2]\n" - "fmla v22.4s, v6.4s, v2.s[2]\n" - "ldr q6, [x28, #0x120]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v17.4s, v7.4s, v1.s[2]\n" - "fmla v23.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x28, #0x130]\n" - "fmla v12.4s, v4.4s, v0.s[2]\n" - "fmla v18.4s, v4.4s, v1.s[2]\n" - "fmla v24.4s, v4.4s, v2.s[2]\n" - "ldr q4, [x28, #0x140]\n" - "fmla v13.4s, v5.4s, v0.s[2]\n" - "fmla v19.4s, v5.4s, v1.s[2]\n" - "fmla v25.4s, v5.4s, v2.s[2]\n" - "ldr q5, [x28, #0x150]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v20.4s, v6.4s, v2.s[3]\n" - "ldr q6, [x28, #0x160]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "fmla v21.4s, v7.4s, v2.s[3]\n" - "ldr q7, [x28, #0x170]\n" + "ldr q26, [x28, #0x70]\n" + "fmla v12.4s, v29.4s, v0.s[0]\n" + "fmla v18.4s, v29.4s, v1.s[0]\n" + "fmla v24.4s, v29.4s, v2.s[0]\n" + "ldr q29, [x28, #0x80]\n" + "fmla v13.4s, v28.4s, v0.s[0]\n" + "fmla v19.4s, v28.4s, v1.s[0]\n" + "fmla v25.4s, v28.4s, v2.s[0]\n" + "ldr q28, [x28, #0x90]\n" + "fmla v8.4s, v27.4s, v0.s[1]\n" + "fmla v14.4s, v27.4s, v1.s[1]\n" + "fmla v20.4s, v27.4s, v2.s[1]\n" + "ldr q27, [x28, #0xa0]\n" + "fmla v9.4s, v26.4s, v0.s[1]\n" + "fmla v15.4s, v26.4s, v1.s[1]\n" + "fmla v21.4s, v26.4s, v2.s[1]\n" + "ldr q26, [x28, #0xb0]\n" + "fmla v10.4s, v29.4s, v0.s[1]\n" + "fmla v16.4s, v29.4s, v1.s[1]\n" + "fmla v22.4s, v29.4s, v2.s[1]\n" + "ldr q29, [x28, #0xc0]\n" + "fmla v11.4s, v28.4s, v0.s[1]\n" + "fmla v17.4s, v28.4s, v1.s[1]\n" + "fmla v23.4s, v28.4s, v2.s[1]\n" + "ldr q28, [x28, #0xd0]\n" + "fmla v12.4s, v27.4s, v0.s[1]\n" + "fmla v18.4s, v27.4s, v1.s[1]\n" + "fmla v24.4s, v27.4s, v2.s[1]\n" + "ldr q27, [x28, #0xe0]\n" + "fmla v13.4s, v26.4s, v0.s[1]\n" + "fmla v19.4s, v26.4s, v1.s[1]\n" + "fmla v25.4s, v26.4s, v2.s[1]\n" + "ldr q26, [x28, #0xf0]\n" + "fmla v8.4s, v29.4s, v0.s[2]\n" + "fmla v14.4s, v29.4s, v1.s[2]\n" + "fmla v20.4s, v29.4s, v2.s[2]\n" + "ldr q29, [x28, #0x100]\n" + "fmla v9.4s, v28.4s, v0.s[2]\n" + "fmla v15.4s, v28.4s, v1.s[2]\n" + "fmla v21.4s, v28.4s, v2.s[2]\n" + "ldr q28, [x28, #0x110]\n" + "fmla v10.4s, v27.4s, v0.s[2]\n" + "fmla v16.4s, v27.4s, v1.s[2]\n" + "fmla v22.4s, v27.4s, v2.s[2]\n" + "ldr q27, [x28, #0x120]\n" + "fmla v11.4s, v26.4s, v0.s[2]\n" + "fmla v17.4s, v26.4s, v1.s[2]\n" + "fmla v23.4s, v26.4s, v2.s[2]\n" + "ldr q26, [x28, #0x130]\n" + "fmla v12.4s, v29.4s, v0.s[2]\n" + "fmla v18.4s, v29.4s, v1.s[2]\n" + "fmla v24.4s, v29.4s, v2.s[2]\n" + "ldr q29, [x28, #0x140]\n" + "fmla v13.4s, v28.4s, v0.s[2]\n" + "fmla v19.4s, v28.4s, v1.s[2]\n" + "fmla v25.4s, v28.4s, v2.s[2]\n" + "ldr q28, [x28, #0x150]\n" + "fmla v8.4s, v27.4s, v0.s[3]\n" + "fmla v14.4s, v27.4s, v1.s[3]\n" + "fmla v20.4s, v27.4s, v2.s[3]\n" + "ldr q27, [x28, #0x160]\n" + "fmla v9.4s, v26.4s, v0.s[3]\n" + "fmla v15.4s, v26.4s, v1.s[3]\n" + "fmla v21.4s, v26.4s, v2.s[3]\n" + "ldr q26, [x28, #0x170]\n" "add x28, x28, #0x180\n" - "fmla v10.4s, v4.4s, v0.s[3]\n" - "fmla v16.4s, v4.4s, v1.s[3]\n" - "fmla v22.4s, v4.4s, v2.s[3]\n" - "fmla v11.4s, v5.4s, v0.s[3]\n" - "fmla v17.4s, v5.4s, v1.s[3]\n" - "fmla v23.4s, v5.4s, v2.s[3]\n" - "fmla v12.4s, v6.4s, v0.s[3]\n" - "fmla v18.4s, v6.4s, v1.s[3]\n" - "fmla v24.4s, v6.4s, v2.s[3]\n" - "fmla v13.4s, v7.4s, v0.s[3]\n" - "fmla v19.4s, v7.4s, v1.s[3]\n" - "fmla v25.4s, v7.4s, v2.s[3]\n" + "fmla v10.4s, v29.4s, v0.s[3]\n" + "fmla v16.4s, v29.4s, v1.s[3]\n" + "fmla v22.4s, v29.4s, v2.s[3]\n" + "fmla v11.4s, v28.4s, v0.s[3]\n" + "fmla v17.4s, v28.4s, v1.s[3]\n" + "fmla v23.4s, v28.4s, v2.s[3]\n" + "fmla v12.4s, v27.4s, v0.s[3]\n" + "fmla v18.4s, v27.4s, v1.s[3]\n" + "fmla v24.4s, v27.4s, v2.s[3]\n" + "fmla v13.4s, v26.4s, v0.s[3]\n" + "fmla v19.4s, v26.4s, v1.s[3]\n" + "fmla v25.4s, v26.4s, v2.s[3]\n" "106:" // Height 3: Multiply loop: Main loop skip "cbz x25, 108f\n" "107:" // Height 3: Multiply loop: Odd block loop "ldr s0, [x24], #0x4\n" - "ldr s1, [x23], #0x4\n" + "ldr s31, [x23], #0x4\n" "sub x25, x25, #0x1\n" - "ldr s2, [x22], #0x4\n" - "ldr q4, [x28, #0x0]\n" - "fmla v8.4s, v4.4s, v0.s[0]\n" - "fmla v14.4s, v4.4s, v1.s[0]\n" - "ldr q5, [x28, #0x10]\n" - "ldr q6, [x28, #0x20]\n" - "fmla v20.4s, v4.4s, v2.s[0]\n" - "fmla v9.4s, v5.4s, v0.s[0]\n" - "ldr q7, [x28, #0x30]\n" - "ldr q4, [x28, #0x40]\n" - "fmla v15.4s, v5.4s, v1.s[0]\n" - "fmla v21.4s, v5.4s, v2.s[0]\n" - "ldr q5, [x28, #0x50]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v16.4s, v6.4s, v1.s[0]\n" + "ldr s30, [x22], #0x4\n" + "ldr q27, [x28, #0x0]\n" + "fmla v8.4s, v27.4s, v0.s[0]\n" + "fmla v14.4s, v27.4s, v31.s[0]\n" + "ldr q26, [x28, #0x10]\n" + "ldr q29, [x28, #0x20]\n" + "fmla v20.4s, v27.4s, v30.s[0]\n" + "fmla v9.4s, v26.4s, v0.s[0]\n" + "ldr q28, [x28, #0x30]\n" + "ldr q27, [x28, #0x40]\n" + "fmla v15.4s, v26.4s, v31.s[0]\n" + "fmla v21.4s, v26.4s, v30.s[0]\n" + "ldr q26, [x28, #0x50]\n" + "fmla v10.4s, v29.4s, v0.s[0]\n" + "fmla v16.4s, v29.4s, v31.s[0]\n" "add x28, x28, #0x60\n" - "fmla v22.4s, v6.4s, v2.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v17.4s, v7.4s, v1.s[0]\n" - "fmla v23.4s, v7.4s, v2.s[0]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "fmla v24.4s, v4.4s, v2.s[0]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" - "fmla v25.4s, v5.4s, v2.s[0]\n" + "fmla v22.4s, v29.4s, v30.s[0]\n" + "fmla v11.4s, v28.4s, v0.s[0]\n" + "fmla v17.4s, v28.4s, v31.s[0]\n" + "fmla v23.4s, v28.4s, v30.s[0]\n" + "fmla v12.4s, v27.4s, v0.s[0]\n" + "fmla v18.4s, v27.4s, v31.s[0]\n" + "fmla v24.4s, v27.4s, v30.s[0]\n" + "fmla v13.4s, v26.4s, v0.s[0]\n" + "fmla v19.4s, v26.4s, v31.s[0]\n" + "fmla v25.4s, v26.4s, v30.s[0]\n" "cbnz x25, 107b\n" "108:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1523,45 +1522,45 @@ void a64_hybrid_fp32_mla_4x24 ( "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 109f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v27.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmin v20.4s, v20.4s, v1.4s\n" - "fmin v21.4s, v21.4s, v1.4s\n" - "fmin v22.4s, v22.4s, v1.4s\n" - "fmin v23.4s, v23.4s, v1.4s\n" - "fmin v24.4s, v24.4s, v1.4s\n" - "fmin v25.4s, v25.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" - "fmax v20.4s, v20.4s, v0.4s\n" - "fmax v21.4s, v21.4s, v0.4s\n" - "fmax v22.4s, v22.4s, v0.4s\n" - "fmax v23.4s, v23.4s, v0.4s\n" - "fmax v24.4s, v24.4s, v0.4s\n" - "fmax v25.4s, v25.4s, v0.4s\n" + "ld1r { v26.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v27.4s\n" + "fmin v9.4s, v9.4s, v27.4s\n" + "fmin v10.4s, v10.4s, v27.4s\n" + "fmin v11.4s, v11.4s, v27.4s\n" + "fmin v12.4s, v12.4s, v27.4s\n" + "fmin v13.4s, v13.4s, v27.4s\n" + "fmin v14.4s, v14.4s, v27.4s\n" + "fmin v15.4s, v15.4s, v27.4s\n" + "fmin v16.4s, v16.4s, v27.4s\n" + "fmin v17.4s, v17.4s, v27.4s\n" + "fmin v18.4s, v18.4s, v27.4s\n" + "fmin v19.4s, v19.4s, v27.4s\n" + "fmin v20.4s, v20.4s, v27.4s\n" + "fmin v21.4s, v21.4s, v27.4s\n" + "fmin v22.4s, v22.4s, v27.4s\n" + "fmin v23.4s, v23.4s, v27.4s\n" + "fmin v24.4s, v24.4s, v27.4s\n" + "fmin v25.4s, v25.4s, v27.4s\n" + "fmax v8.4s, v8.4s, v26.4s\n" + "fmax v9.4s, v9.4s, v26.4s\n" + "fmax v10.4s, v10.4s, v26.4s\n" + "fmax v11.4s, v11.4s, v26.4s\n" + "fmax v12.4s, v12.4s, v26.4s\n" + "fmax v13.4s, v13.4s, v26.4s\n" + "fmax v14.4s, v14.4s, v26.4s\n" + "fmax v15.4s, v15.4s, v26.4s\n" + "fmax v16.4s, v16.4s, v26.4s\n" + "fmax v17.4s, v17.4s, v26.4s\n" + "fmax v18.4s, v18.4s, v26.4s\n" + "fmax v19.4s, v19.4s, v26.4s\n" + "fmax v20.4s, v20.4s, v26.4s\n" + "fmax v21.4s, v21.4s, v26.4s\n" + "fmax v22.4s, v22.4s, v26.4s\n" + "fmax v23.4s, v23.4s, v26.4s\n" + "fmax v24.4s, v24.4s, v26.4s\n" + "fmax v25.4s, v25.4s, v26.4s\n" "109:" // Height 3: No activation "cmp x9, #0x18\n" "bge 122f\n" @@ -1983,14 +1982,14 @@ void a64_hybrid_fp32_mla_4x24 ( "142:" // 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, 143f\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, 144f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -2000,9 +1999,9 @@ void a64_hybrid_fp32_mla_4x24 ( "b 144f\n" "143:" // Height 4: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "144:" // Height 4: input setup done "cmp x25, #0x4\n" "blt 147f\n" @@ -2283,42 +2282,42 @@ void a64_hybrid_fp32_mla_4x24 ( "147:" // Height 4: Multiply loop: Main loop skip "cbz x25, 149f\n" "148:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x24], #0x4\n" - "ldr s1, [x23], #0x4\n" + "ldr s7, [x24], #0x4\n" + "ldr s6, [x23], #0x4\n" "sub x25, x25, #0x1\n" - "ldr s2, [x22], #0x4\n" - "ldr s3, [x21], #0x4\n" - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" - "fmla v8.4s, v4.4s, v0.s[0]\n" - "fmla v14.4s, v4.4s, v1.s[0]\n" - "ldr q6, [x28, #0x20]\n" - "ldr q7, [x28, #0x30]\n" - "fmla v20.4s, v4.4s, v2.s[0]\n" - "fmla v26.4s, v4.4s, v3.s[0]\n" - "ldr q4, [x28, #0x40]\n" - "fmla v9.4s, v5.4s, v0.s[0]\n" - "fmla v15.4s, v5.4s, v1.s[0]\n" - "fmla v21.4s, v5.4s, v2.s[0]\n" - "fmla v27.4s, v5.4s, v3.s[0]\n" - "ldr q5, [x28, #0x50]\n" + "ldr s5, [x22], #0x4\n" + "ldr s4, [x21], #0x4\n" + "ldr q1, [x28, #0x0]\n" + "ldr q0, [x28, #0x10]\n" + "fmla v8.4s, v1.4s, v7.s[0]\n" + "fmla v14.4s, v1.4s, v6.s[0]\n" + "ldr q3, [x28, #0x20]\n" + "ldr q2, [x28, #0x30]\n" + "fmla v20.4s, v1.4s, v5.s[0]\n" + "fmla v26.4s, v1.4s, v4.s[0]\n" + "ldr q1, [x28, #0x40]\n" + "fmla v9.4s, v0.4s, v7.s[0]\n" + "fmla v15.4s, v0.4s, v6.s[0]\n" + "fmla v21.4s, v0.4s, v5.s[0]\n" + "fmla v27.4s, v0.4s, v4.s[0]\n" + "ldr q0, [x28, #0x50]\n" "add x28, x28, #0x60\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v16.4s, v6.4s, v1.s[0]\n" - "fmla v22.4s, v6.4s, v2.s[0]\n" - "fmla v28.4s, v6.4s, v3.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v17.4s, v7.4s, v1.s[0]\n" - "fmla v23.4s, v7.4s, v2.s[0]\n" - "fmla v29.4s, v7.4s, v3.s[0]\n" - "fmla v12.4s, v4.4s, v0.s[0]\n" - "fmla v18.4s, v4.4s, v1.s[0]\n" - "fmla v24.4s, v4.4s, v2.s[0]\n" - "fmla v30.4s, v4.4s, v3.s[0]\n" - "fmla v13.4s, v5.4s, v0.s[0]\n" - "fmla v19.4s, v5.4s, v1.s[0]\n" - "fmla v25.4s, v5.4s, v2.s[0]\n" - "fmla v31.4s, v5.4s, v3.s[0]\n" + "fmla v10.4s, v3.4s, v7.s[0]\n" + "fmla v16.4s, v3.4s, v6.s[0]\n" + "fmla v22.4s, v3.4s, v5.s[0]\n" + "fmla v28.4s, v3.4s, v4.s[0]\n" + "fmla v11.4s, v2.4s, v7.s[0]\n" + "fmla v17.4s, v2.4s, v6.s[0]\n" + "fmla v23.4s, v2.4s, v5.s[0]\n" + "fmla v29.4s, v2.4s, v4.s[0]\n" + "fmla v12.4s, v1.4s, v7.s[0]\n" + "fmla v18.4s, v1.4s, v6.s[0]\n" + "fmla v24.4s, v1.4s, v5.s[0]\n" + "fmla v30.4s, v1.4s, v4.s[0]\n" + "fmla v13.4s, v0.4s, v7.s[0]\n" + "fmla v19.4s, v0.4s, v6.s[0]\n" + "fmla v25.4s, v0.4s, v5.s[0]\n" + "fmla v31.4s, v0.4s, v4.s[0]\n" "cbnz x25, 148b\n" "149:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -2584,7 +2583,6 @@ void a64_hybrid_fp32_mla_4x24 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "166:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, 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)) : "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" |