diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp | 177 |
1 files changed, 88 insertions, 89 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp index 8b7f5afb7e..d3367b959a 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp @@ -44,18 +44,18 @@ void a64_hybrid_u8u32_dot_6x16 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; - void *output_ptr = {}; } ka; unsigned long flags=0; + void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - ka.output_ptr=(void *)(output_arg.indirect.ptr); + output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - ka.output_ptr=(void *)(output_arg.direct.base); + output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -88,7 +88,7 @@ void a64_hybrid_u8u32_dot_6x16 ( "beq 35f\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x9, %x[output_ptr]\n" "2:" // Height 1: Column loop "tbz %x[flags], #0, 12f\n" "cmp x11, #0x10\n" @@ -163,8 +163,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "mov x28, #0x0\n" "14:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 15f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -188,10 +188,6 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr q17, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" "ldr q16, [x10, #0x30]\n" - "sub x27, x27, #0x10\n" - "add x26, x26, #0x10\n" - "cmp x27, #0x20\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" "ldr q17, [x10, #0x40]\n" ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" @@ -216,21 +212,22 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr q17, [x10, #0xe0]\n" ".inst 0x6fa0ea09 // udot v9.4s, v16.16b, v0.4b[3]\n" "ldr q16, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" + "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" ".inst 0x6fa0ea2a // udot v10.4s, v17.16b, v0.4b[3]\n" - "ldr q6, [x10, #0x0]\n" ".inst 0x6fa0ea0b // udot v11.4s, v16.16b, v0.4b[3]\n" "ldr q0, [x26, #0x0]\n" + "cmp x27, #0x20\n" + "add x10, x10, #0x100\n" + "ldr q6, [x10, #0x0]\n" "ldr q7, [x10, #0x10]\n" + "prfm pldl1keep, [x26, #0x80]\n" "bge 17b\n" "18:" // Height 1: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" "ldr q17, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" "ldr q16, [x10, #0x30]\n" - "add x26, x26, #0x10\n" - "sub x27, x27, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" "ldr q17, [x10, #0x40]\n" ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" @@ -255,26 +252,29 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr q17, [x10, #0xe0]\n" ".inst 0x6fa0ea09 // udot v9.4s, v16.16b, v0.4b[3]\n" "ldr q16, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" + "add x26, x26, #0x10\n" + "sub x27, x27, #0x10\n" ".inst 0x6fa0ea2a // udot v10.4s, v17.16b, v0.4b[3]\n" ".inst 0x6fa0ea0b // udot v11.4s, v16.16b, v0.4b[3]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "add x10, x10, #0x100\n" "19:" // Height 1: Multiply loop: Main loop skip "cbz x27, 24f\n" "cmp x27, #0x4\n" "blt 21f\n" "20:" // Height 1: Multiply loop: Odd block loop "ldr s18, [x26], #0x4\n" - "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x0]\n" + ".inst 0x6f92e208 // udot v8.4s, v16.16b, v18.4b[0]\n" "sub x27, x27, #0x4\n" "ldr q16, [x10, #0x10]\n" - "cmp x27, #0x4\n" - ".inst 0x6f92e228 // udot v8.4s, v17.16b, v18.4b[0]\n" "ldr q17, [x10, #0x20]\n" ".inst 0x6f92e209 // udot v9.4s, v16.16b, v18.4b[0]\n" + "cmp x27, #0x4\n" "ldr q16, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f92e22a // udot v10.4s, v17.16b, v18.4b[0]\n" ".inst 0x6f92e20b // udot v11.4s, v16.16b, v18.4b[0]\n" + "add x10, x10, #0x40\n" "bge 20b\n" "21:" // Height 1: Multiply loop: Skip odd blocks "cbz x27, 24f\n" @@ -289,12 +289,12 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr q17, [x10, #0x0]\n" "ldr q16, [x10, #0x10]\n" ".inst 0x6f80e228 // udot v8.4s, v17.16b, v0.4b[0]\n" - "ldr q17, [x10, #0x20]\n" ".inst 0x6f80e209 // udot v9.4s, v16.16b, v0.4b[0]\n" + "ldr q17, [x10, #0x20]\n" "ldr q16, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" + "add x10, x10, #0x40\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -362,7 +362,7 @@ void a64_hybrid_u8u32_dot_6x16 ( "35:" // Height 2 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x9, %x[output_ptr]\n" "36:" // Height 2: Column loop "tbz %x[flags], #0, 46f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" @@ -463,8 +463,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "mov x28, #0x0\n" "48:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 49f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -496,22 +496,22 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" - "add x25, x25, #0x10\n" - "cmp x27, #0x20\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" ".inst 0x6f81e22e // udot v14.4s, v17.16b, v1.4b[0]\n" "ldr q17, [x10, #0x40]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x25, x25, #0x10\n" ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" ".inst 0x6f81e20f // udot v15.4s, v16.16b, v1.4b[0]\n" "ldr q16, [x10, #0x50]\n" + "cmp x27, #0x20\n" ".inst 0x6fa0e228 // udot v8.4s, v17.16b, v0.4b[1]\n" ".inst 0x6fa1e22c // udot v12.4s, v17.16b, v1.4b[1]\n" "ldr q17, [x10, #0x60]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6fa0e209 // udot v9.4s, v16.16b, v0.4b[1]\n" ".inst 0x6fa1e20d // udot v13.4s, v16.16b, v1.4b[1]\n" "ldr q16, [x10, #0x70]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6fa0e22a // udot v10.4s, v17.16b, v0.4b[1]\n" ".inst 0x6fa1e22e // udot v14.4s, v17.16b, v1.4b[1]\n" "ldr q17, [x10, #0x80]\n" @@ -555,18 +555,18 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" - "sub x27, x27, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" ".inst 0x6f81e22e // udot v14.4s, v17.16b, v1.4b[0]\n" "ldr q17, [x10, #0x40]\n" + "sub x27, x27, #0x10\n" ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" ".inst 0x6f81e20f // udot v15.4s, v16.16b, v1.4b[0]\n" "ldr q16, [x10, #0x50]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6fa0e228 // udot v8.4s, v17.16b, v0.4b[1]\n" ".inst 0x6fa1e22c // udot v12.4s, v17.16b, v1.4b[1]\n" "ldr q17, [x10, #0x60]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6fa0e209 // udot v9.4s, v16.16b, v0.4b[1]\n" ".inst 0x6fa1e20d // udot v13.4s, v16.16b, v1.4b[1]\n" "ldr q16, [x10, #0x70]\n" @@ -607,18 +607,18 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr s19, [x26], #0x4\n" "ldr s18, [x25], #0x4\n" "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ldr q17, [x10, #0x0]\n" "ldr q16, [x10, #0x10]\n" - "cmp x27, #0x4\n" ".inst 0x6f93e228 // udot v8.4s, v17.16b, v19.4b[0]\n" ".inst 0x6f92e22c // udot v12.4s, v17.16b, v18.4b[0]\n" "ldr q17, [x10, #0x20]\n" ".inst 0x6f93e209 // udot v9.4s, v16.16b, v19.4b[0]\n" ".inst 0x6f92e20d // udot v13.4s, v16.16b, v18.4b[0]\n" "ldr q16, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f93e22a // udot v10.4s, v17.16b, v19.4b[0]\n" ".inst 0x6f92e22e // udot v14.4s, v17.16b, v18.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f93e20b // udot v11.4s, v16.16b, v19.4b[0]\n" ".inst 0x6f92e20f // udot v15.4s, v16.16b, v18.4b[0]\n" "bge 54b\n" @@ -643,9 +643,9 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f80e209 // udot v9.4s, v16.16b, v0.4b[0]\n" ".inst 0x6f81e20d // udot v13.4s, v16.16b, v1.4b[0]\n" "ldr q16, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" ".inst 0x6f81e22e // udot v14.4s, v17.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" ".inst 0x6f81e20f // udot v15.4s, v16.16b, v1.4b[0]\n" "58:" // Height 2: Multiply loop: No odd multiplies @@ -654,9 +654,9 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x28, x20\n" "bne 48b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x24, x9, x20, LSL #2\n" "cmp x11, #0x10\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x24, x9, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" "bge 67f\n" "tbz x11, #3, 62f\n" @@ -738,12 +738,12 @@ void a64_hybrid_u8u32_dot_6x16 ( "69:" // Height 3 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x9, %x[output_ptr]\n" "70:" // Height 3: Column loop "tbz %x[flags], #0, 80f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" "add x24, x9, x20, LSL #2\n" + "cmp x11, #0x10\n" "add x23, x24, x20, LSL #2\n" "bge 79f\n" "tbz x11, #3, 74f\n" @@ -864,8 +864,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "mov x28, #0x0\n" "82:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 83f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -905,18 +905,18 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - "cmp x27, #0x20\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f80e2aa // udot v10.4s, v21.16b, v0.4b[0]\n" ".inst 0x6f81e2ae // udot v14.4s, v21.16b, v1.4b[0]\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f82e2b2 // udot v18.4s, v21.16b, v2.4b[0]\n" "ldr q21, [x10, #0x40]\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f80e28b // udot v11.4s, v20.16b, v0.4b[0]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f81e28f // udot v15.4s, v20.16b, v1.4b[0]\n" ".inst 0x6f82e293 // udot v19.4s, v20.16b, v2.4b[0]\n" "ldr q20, [x10, #0x50]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6fa0e2a8 // udot v8.4s, v21.16b, v0.4b[1]\n" ".inst 0x6fa1e2ac // udot v12.4s, v21.16b, v1.4b[1]\n" ".inst 0x6fa2e2b0 // udot v16.4s, v21.16b, v2.4b[1]\n" @@ -983,14 +983,14 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f80e2aa // udot v10.4s, v21.16b, v0.4b[0]\n" ".inst 0x6f81e2ae // udot v14.4s, v21.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f82e2b2 // udot v18.4s, v21.16b, v2.4b[0]\n" "ldr q21, [x10, #0x40]\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f80e28b // udot v11.4s, v20.16b, v0.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f81e28f // udot v15.4s, v20.16b, v1.4b[0]\n" ".inst 0x6f82e293 // udot v19.4s, v20.16b, v2.4b[0]\n" "ldr q20, [x10, #0x50]\n" @@ -1049,12 +1049,12 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr s24, [x26], #0x4\n" "ldr s23, [x25], #0x4\n" "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ldr s22, [x24], #0x4\n" "ldr q21, [x10, #0x0]\n" - "cmp x27, #0x4\n" - "ldr q20, [x10, #0x10]\n" ".inst 0x6f98e2a8 // udot v8.4s, v21.16b, v24.4b[0]\n" ".inst 0x6f97e2ac // udot v12.4s, v21.16b, v23.4b[0]\n" + "ldr q20, [x10, #0x10]\n" ".inst 0x6f96e2b0 // udot v16.4s, v21.16b, v22.4b[0]\n" "ldr q21, [x10, #0x20]\n" ".inst 0x6f98e289 // udot v9.4s, v20.16b, v24.4b[0]\n" @@ -1108,11 +1108,11 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x28, x20\n" "bne 82b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" - "prfm pstl1keep, [x9, #0x0]\n" "add x24, x9, x20, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" "add x23, x24, x20, LSL #2\n" + "prfm pstl1keep, [x9, #0x0]\n" + "cmp x11, #0x10\n" + "prfm pstl1keep, [x24, #0x0]\n" "prfm pstl1keep, [x23, #0x0]\n" "bge 101f\n" "tbz x11, #3, 96f\n" @@ -1214,13 +1214,13 @@ void a64_hybrid_u8u32_dot_6x16 ( "103:" // Height 4 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x9, %x[output_ptr]\n" "104:" // Height 4: Column loop "tbz %x[flags], #0, 114f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "cmp x11, #0x10\n" "add x22, x23, x20, LSL #2\n" "bge 113f\n" "tbz x11, #3, 108f\n" @@ -1365,8 +1365,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "mov x28, #0x0\n" "116:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 117f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1506,14 +1506,14 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" "add x23, x23, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" + "sub x27, x27, #0x10\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" "ldr q24, [x10, #0x30]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f80e32a // udot v10.4s, v25.16b, v0.4b[0]\n" ".inst 0x6f81e32e // udot v14.4s, v25.16b, v1.4b[0]\n" - "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f82e332 // udot v18.4s, v25.16b, v2.4b[0]\n" ".inst 0x6f83e336 // udot v22.4s, v25.16b, v3.4b[0]\n" @@ -1591,9 +1591,9 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr s29, [x26], #0x4\n" "ldr s28, [x25], #0x4\n" "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ldr s27, [x24], #0x4\n" "ldr s26, [x23], #0x4\n" - "cmp x27, #0x4\n" "ldr q25, [x10, #0x0]\n" "ldr q24, [x10, #0x10]\n" ".inst 0x6f9de328 // udot v8.4s, v25.16b, v29.4b[0]\n" @@ -1662,13 +1662,13 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x28, x20\n" "bne 116b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" - "prfm pstl1keep, [x9, #0x0]\n" "add x24, x9, x20, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" "add x23, x24, x20, LSL #2\n" - "prfm pstl1keep, [x23, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" "add x22, x23, x20, LSL #2\n" + "cmp x11, #0x10\n" + "prfm pstl1keep, [x24, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" "bge 135f\n" "tbz x11, #3, 130f\n" @@ -1790,14 +1790,14 @@ void a64_hybrid_u8u32_dot_6x16 ( "137:" // Height 5 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x9, %x[output_ptr]\n" "138:" // Height 5: Column loop "tbz %x[flags], #0, 148f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "cmp x11, #0x10\n" "add x21, x22, x20, LSL #2\n" "bge 147f\n" "tbz x11, #3, 142f\n" @@ -1966,8 +1966,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "mov x28, #0x0\n" "150:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 151f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2133,12 +2133,12 @@ void a64_hybrid_u8u32_dot_6x16 ( "add x22, x22, #0x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "sub x27, x27, #0x10\n" "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" "ldr q28, [x10, #0x30]\n" - "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6f80e3aa // udot v10.4s, v29.16b, v0.4b[0]\n" ".inst 0x6f81e3ae // udot v14.4s, v29.16b, v1.4b[0]\n" "prfm pldl1keep, [x24, #0x80]\n" @@ -2233,14 +2233,14 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr s2, [x26], #0x4\n" "ldr s1, [x25], #0x4\n" "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ldr s0, [x24], #0x4\n" "ldr s31, [x23], #0x4\n" - "cmp x27, #0x4\n" "ldr s30, [x22], #0x4\n" "ldr q29, [x10, #0x0]\n" - "ldr q28, [x10, #0x10]\n" ".inst 0x6f82e3a8 // udot v8.4s, v29.16b, v2.4b[0]\n" ".inst 0x6f81e3ac // udot v12.4s, v29.16b, v1.4b[0]\n" + "ldr q28, [x10, #0x10]\n" ".inst 0x6f80e3b0 // udot v16.4s, v29.16b, v0.4b[0]\n" ".inst 0x6f9fe3b4 // udot v20.4s, v29.16b, v31.4b[0]\n" ".inst 0x6f9ee3b8 // udot v24.4s, v29.16b, v30.4b[0]\n" @@ -2316,15 +2316,15 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x28, x20\n" "bne 150b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" - "prfm pstl1keep, [x9, #0x0]\n" "add x24, x9, x20, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" "add x23, x24, x20, LSL #2\n" - "prfm pstl1keep, [x23, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" "add x22, x23, x20, LSL #2\n" - "prfm pstl1keep, [x22, #0x0]\n" "add x21, x22, x20, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "cmp x11, #0x10\n" + "prfm pstl1keep, [x22, #0x0]\n" "prfm pstl1keep, [x21, #0x0]\n" "bge 169f\n" "tbz x11, #3, 164f\n" @@ -2465,20 +2465,19 @@ void a64_hybrid_u8u32_dot_6x16 ( "b 206f\n" "171:" // Height 6 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x18\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[output_ptr]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "madd x20, x21, x20, x9\n" - "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" "172:" // Height 6: Column loop "tbz %x[flags], #0, 182f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" "add x21, x22, x20, LSL #2\n" + "cmp x11, #0x10\n" "add x20, x21, x20, LSL #2\n" "bge 181f\n" "tbz x11, #3, 176f\n" @@ -2671,8 +2670,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "mov x28, #0x0\n" "184:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 185f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2862,18 +2861,18 @@ void a64_hybrid_u8u32_dot_6x16 ( ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" "add x21, x21, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" + "sub x27, x27, #0x10\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" "ldr q7, [x10, #0x30]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -2979,9 +2978,9 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr s7, [x26], #0x4\n" "ldr s6, [x25], #0x4\n" "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ldr s5, [x24], #0x4\n" "ldr s4, [x23], #0x4\n" - "cmp x27, #0x4\n" "ldr s3, [x22], #0x4\n" "ldr s2, [x21], #0x4\n" "ldr q1, [x10, #0x0]\n" @@ -3074,16 +3073,16 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x28, x20\n" "bne 184b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x11, #0x10\n" - "prfm pstl1keep, [x9, #0x0]\n" "add x24, x9, x20, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" "add x23, x24, x20, LSL #2\n" - "prfm pstl1keep, [x23, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" "add x22, x23, x20, LSL #2\n" - "prfm pstl1keep, [x22, #0x0]\n" "add x21, x22, x20, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" "add x20, x21, x20, LSL #2\n" + "cmp x11, #0x10\n" + "prfm pstl1keep, [x22, #0x0]\n" "prfm pstl1keep, [x21, #0x0]\n" "prfm pstl1keep, [x20, #0x0]\n" "bge 203f\n" @@ -3254,8 +3253,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "206:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) - : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) + : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) + : [args_ptr] "r" (&ka), [flags] "r" (flags), [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", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |