diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp | 873 |
1 files changed, 435 insertions, 438 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp index 6401b01607..36356dfb35 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp @@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_8x4 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; - void *output_ptr = {}; - const float *bias = {}; } 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; } @@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_8x4 ( ka.string_lengths = string_lengths; ka.N = N; ka.B_ptr = B_ptr; - ka.bias = bias; switch(act.type) { default: case Activation::Type::None: @@ -106,10 +104,10 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp %x[M], #0x2\n" "bgt 43f\n" "beq 22f\n" - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "2:" // Height 1: Column loop "cbz x14, 3f\n" "ldr q24, [x14, #0x0]\n" @@ -140,8 +138,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "9:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 10f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -164,37 +162,37 @@ void a64_hybrid_fp32_mla_8x4 ( "blt 13f\n" "12:" // Height 1: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "sub x9, x9, #0x4\n" "add x28, x28, #0x10\n" + "fmla v24.4s, v10.4s, v0.s[2]\n" "cmp x9, #0x8\n" "add x12, x12, #0x40\n" - "prfm pldl1keep, [x28, #0x80]\n" "ldr q8, [x12, #0x0]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" "ldr q9, [x12, #0x10]\n" - "fmla v24.4s, v10.4s, v0.s[2]\n" "ldr q10, [x12, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "ldr q0, [x28, #0x0]\n" "ldr q11, [x12, #0x30]\n" + "prfm pldl1keep, [x28, #0x80]\n" "bge 12b\n" "13:" // Height 1: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "add x28, x28, #0x10\n" "sub x9, x9, #0x4\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x28, #0x80]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "add x12, x12, #0x40\n" "14:" // Height 1: Multiply loop: Main loop skip "cbz x9, 16f\n" "15:" // Height 1: Multiply loop: Odd block loop "ldr s17, [x28], #0x4\n" "ldr q16, [x12, #0x0]\n" "sub x9, x9, #0x1\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v17.s[0]\n" + "add x12, x12, #0x10\n" "cbnz x9, 15b\n" "16:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -203,9 +201,9 @@ void a64_hybrid_fp32_mla_8x4 ( "bne 9b\n" "prfm pstl1keep, [x11, #0x0]\n" "tbz %x[flags], #1, 17f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmax v24.4s, v24.4s, v16.4s\n" @@ -229,40 +227,40 @@ void a64_hybrid_fp32_mla_8x4 ( "bgt 2b\n" "b 170f\n" "22:" // Height 2 - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "23:" // Height 2: Column loop "cbz x14, 24f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" + "add x14, x14, #0x10\n" "b 29f\n" "24:" // Height 2: no bias "tbz %x[flags], #0, 28f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" + "add x27, x11, x20, LSL #2\n" "bge 27f\n" "tbz x13, #1, 25f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" "tbz x13, #0, 26f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" + "ld1 { v25.s }[2], [x27]\n" "b 26f\n" "25:" // Height 2: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" "26:" // Height 2: Partial accumulate: Done "sub x11, x11, x20\n" "b 29f\n" "27:" // Height 2: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" + "ldr q25, [x27, #0x0]\n" "b 29f\n" "28:" // Height 2: no accumulate "movi v24.16b, #0x0\n" @@ -271,8 +269,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "30:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 31f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -302,37 +300,37 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v25.4s, v8.4s, v1.s[0]\n" "sub x9, x9, #0x4\n" "add x28, x28, #0x10\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" + "fmla v25.4s, v9.4s, v1.s[1]\n" "add x27, x27, #0x10\n" "cmp x9, #0x8\n" - "prfm pldl1keep, [x28, #0x80]\n" + "fmla v24.4s, v10.4s, v0.s[2]\n" + "fmla v25.4s, v10.4s, v1.s[2]\n" "add x12, x12, #0x40\n" - "prfm pldl1keep, [x27, #0x80]\n" "ldr q8, [x12, #0x0]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" - "fmla v25.4s, v9.4s, v1.s[1]\n" "ldr q9, [x12, #0x10]\n" - "fmla v24.4s, v10.4s, v0.s[2]\n" - "fmla v25.4s, v10.4s, v1.s[2]\n" "ldr q10, [x12, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr q0, [x28, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" + "ldr q0, [x28, #0x0]\n" "ldr q1, [x27, #0x0]\n" "ldr q11, [x12, #0x30]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "bge 33b\n" "34:" // Height 2: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" "fmla v25.4s, v8.4s, v1.s[0]\n" "add x28, x28, #0x10\n" "add x27, x27, #0x10\n" - "sub x9, x9, #0x4\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x27, #0x80]\n" "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" + "sub x9, x9, #0x4\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x27, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "35:" // Height 2: Multiply loop: Main loop skip @@ -342,9 +340,9 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr s17, [x27], #0x4\n" "sub x9, x9, #0x1\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v18.s[0]\n" "fmla v25.4s, v16.4s, v17.s[0]\n" + "add x12, x12, #0x10\n" "cbnz x9, 36b\n" "37:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -352,13 +350,13 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 30b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x27, x11, x20, LSL #2\n" "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" + "prfm pstl1keep, [x27, #0x0]\n" "tbz %x[flags], #1, 38f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -369,65 +367,65 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 41f\n" "tbz x13, #1, 39f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" + "str d25, [x27], #0x8\n" "tbz x13, #0, 40f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" + "st1 { v25.s }[2], [x27]\n" "b 40f\n" "39:" // Height 2: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" + "str s25, [x27, #0x0]\n" "40:" // Height 2: Partial direct writeback: Done "b 42f\n" "41:" // Height 2: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" + "str q25, [x27, #0x0]\n" "42:" // Height 2: Writeback done "subs x13, x13, #0x4\n" "bgt 23b\n" "b 170f\n" "43:" // Height 3 - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "44:" // Height 3: Column loop "cbz x14, 45f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x14, x14, #0x10\n" "b 50f\n" "45:" // Height 3: no bias "tbz %x[flags], #0, 49f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x27, x11, x20, LSL #2\n" "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" - "add x27, x28, x20, LSL #2\n" + "add x26, x27, x20, LSL #2\n" "bge 48f\n" "tbz x13, #1, 46f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" - "ldr d26, [x27], #0x8\n" + "ldr d26, [x26], #0x8\n" "tbz x13, #0, 47f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" - "ld1 { v26.s }[2], [x27]\n" + "ld1 { v25.s }[2], [x27]\n" + "ld1 { v26.s }[2], [x26]\n" "b 47f\n" "46:" // Height 3: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" - "ldr s26, [x27, #0x0]\n" + "ldr s26, [x26, #0x0]\n" "47:" // Height 3: Partial accumulate: Done "sub x11, x11, x20\n" "b 50f\n" "48:" // Height 3: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" - "ldr q26, [x27, #0x0]\n" + "ldr q25, [x27, #0x0]\n" + "ldr q26, [x26, #0x0]\n" "b 50f\n" "49:" // Height 3: no accumulate "movi v24.16b, #0x0\n" @@ -437,8 +435,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "51:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 52f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -473,17 +471,14 @@ void a64_hybrid_fp32_mla_8x4 ( "sub x9, x9, #0x4\n" "add x28, x28, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "add x27, x27, #0x10\n" "add x26, x26, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" + "fmla v25.4s, v9.4s, v1.s[1]\n" + "fmla v26.4s, v9.4s, v2.s[1]\n" "cmp x9, #0x8\n" "add x12, x12, #0x40\n" - "prfm pldl1keep, [x27, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" "ldr q8, [x12, #0x0]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" - "fmla v25.4s, v9.4s, v1.s[1]\n" - "fmla v26.4s, v9.4s, v2.s[1]\n" "ldr q9, [x12, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" @@ -491,11 +486,14 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr q10, [x12, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "ldr q0, [x28, #0x0]\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "ldr q1, [x27, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" "ldr q2, [x26, #0x0]\n" "ldr q11, [x12, #0x30]\n" + "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "bge 54b\n" "55:" // Height 3: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" @@ -503,17 +501,17 @@ void a64_hybrid_fp32_mla_8x4 ( "add x28, x28, #0x10\n" "add x27, x27, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "add x26, x26, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" "sub x9, x9, #0x4\n" - "prfm pldl1keep, [x27, #0x80]\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x26, #0x80]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" @@ -526,10 +524,10 @@ void a64_hybrid_fp32_mla_8x4 ( "sub x9, x9, #0x1\n" "ldr s17, [x26], #0x4\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v19.s[0]\n" "fmla v25.4s, v16.4s, v18.s[0]\n" "fmla v26.4s, v16.4s, v17.s[0]\n" + "add x12, x12, #0x10\n" "cbnz x9, 57b\n" "58:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -537,15 +535,15 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 51b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x27, x11, x20, LSL #2\n" + "add x26, x27, x20, LSL #2\n" "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" - "add x27, x28, x20, LSL #2\n" "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x26, #0x0]\n" "tbz %x[flags], #1, 59f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -558,75 +556,75 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 62f\n" "tbz x13, #1, 60f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" - "str d26, [x27], #0x8\n" + "str d25, [x27], #0x8\n" + "str d26, [x26], #0x8\n" "tbz x13, #0, 61f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" - "st1 { v26.s }[2], [x27]\n" + "st1 { v25.s }[2], [x27]\n" + "st1 { v26.s }[2], [x26]\n" "b 61f\n" "60:" // Height 3: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" - "str s26, [x27, #0x0]\n" + "str s25, [x27, #0x0]\n" + "str s26, [x26, #0x0]\n" "61:" // Height 3: Partial direct writeback: Done "b 63f\n" "62:" // Height 3: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" - "str q26, [x27, #0x0]\n" + "str q25, [x27, #0x0]\n" + "str q26, [x26, #0x0]\n" "63:" // Height 3: Writeback done "subs x13, x13, #0x4\n" "bgt 44b\n" "b 170f\n" "64:" // Height 4 - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "65:" // Height 4: Column loop "cbz x14, 66f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x14, x14, #0x10\n" "mov v27.16b, v24.16b\n" "b 71f\n" "66:" // Height 4: no bias "tbz %x[flags], #0, 70f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" + "cmp x13, #0x4\n" + "add x25, x26, x20, LSL #2\n" "bge 69f\n" "tbz x13, #1, 67f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" - "ldr d26, [x27], #0x8\n" - "ldr d27, [x26], #0x8\n" + "ldr d26, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" "tbz x13, #0, 68f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" - "ld1 { v26.s }[2], [x27]\n" - "ld1 { v27.s }[2], [x26]\n" + "ld1 { v25.s }[2], [x27]\n" + "ld1 { v26.s }[2], [x26]\n" + "ld1 { v27.s }[2], [x25]\n" "b 68f\n" "67:" // Height 4: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" - "ldr s26, [x27, #0x0]\n" - "ldr s27, [x26, #0x0]\n" + "ldr s26, [x26, #0x0]\n" + "ldr s27, [x25, #0x0]\n" "68:" // Height 4: Partial accumulate: Done "sub x11, x11, x20\n" "b 71f\n" "69:" // Height 4: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" - "ldr q26, [x27, #0x0]\n" - "ldr q27, [x26, #0x0]\n" + "ldr q25, [x27, #0x0]\n" + "ldr q26, [x26, #0x0]\n" + "ldr q27, [x25, #0x0]\n" "b 71f\n" "70:" // Height 4: no accumulate "movi v24.16b, #0x0\n" @@ -637,8 +635,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "72:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 73f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -680,24 +678,23 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v27.4s, v8.4s, v3.s[0]\n" "add x27, x27, #0x10\n" "add x26, x26, #0x10\n" - "add x25, x25, #0x10\n" - "cmp x9, #0x8\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "ldr q8, [x12, #0x0]\n" "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" + "add x25, x25, #0x10\n" + "cmp x9, #0x8\n" "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" + "add x12, x12, #0x40\n" + "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" "ldr q10, [x12, #0x20]\n" + "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "ldr q0, [x28, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" @@ -707,6 +704,7 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v27.4s, v11.4s, v3.s[3]\n" "ldr q3, [x25, #0x0]\n" "ldr q11, [x12, #0x30]\n" + "prfm pldl1keep, [x25, #0x80]\n" "bge 75b\n" "76:" // Height 4: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" @@ -717,18 +715,18 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v27.4s, v8.4s, v3.s[0]\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" - "sub x9, x9, #0x4\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" + "sub x9, x9, #0x4\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" + "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" @@ -744,9 +742,9 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr s18, [x26], #0x4\n" "ldr s17, [x25], #0x4\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v20.s[0]\n" "fmla v25.4s, v16.4s, v19.s[0]\n" + "add x12, x12, #0x10\n" "fmla v26.4s, v16.4s, v18.s[0]\n" "fmla v27.4s, v16.4s, v17.s[0]\n" "cbnz x9, 78b\n" @@ -756,17 +754,17 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 72b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x27, x11, x20, LSL #2\n" + "add x26, x27, x20, LSL #2\n" "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" - "add x27, x28, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x20, LSL #2\n" "prfm pstl1keep, [x26, #0x0]\n" + "prfm pstl1keep, [x25, #0x0]\n" "tbz %x[flags], #1, 80f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -781,85 +779,85 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 83f\n" "tbz x13, #1, 81f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" - "str d26, [x27], #0x8\n" - "str d27, [x26], #0x8\n" + "str d25, [x27], #0x8\n" + "str d26, [x26], #0x8\n" + "str d27, [x25], #0x8\n" "tbz x13, #0, 82f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" - "st1 { v26.s }[2], [x27]\n" - "st1 { v27.s }[2], [x26]\n" + "st1 { v25.s }[2], [x27]\n" + "st1 { v26.s }[2], [x26]\n" + "st1 { v27.s }[2], [x25]\n" "b 82f\n" "81:" // Height 4: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" - "str s26, [x27, #0x0]\n" - "str s27, [x26, #0x0]\n" + "str s25, [x27, #0x0]\n" + "str s26, [x26, #0x0]\n" + "str s27, [x25, #0x0]\n" "82:" // Height 4: Partial direct writeback: Done "b 84f\n" "83:" // Height 4: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" - "str q26, [x27, #0x0]\n" - "str q27, [x26, #0x0]\n" + "str q25, [x27, #0x0]\n" + "str q26, [x26, #0x0]\n" + "str q27, [x25, #0x0]\n" "84:" // Height 4: Writeback done "subs x13, x13, #0x4\n" "bgt 65b\n" "b 170f\n" "85:" // Height 5 - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "86:" // Height 5: Column loop "cbz x14, 87f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x14, x14, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "b 92f\n" "87:" // Height 5: no bias "tbz %x[flags], #0, 91f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" "add x25, x26, x20, LSL #2\n" + "cmp x13, #0x4\n" + "add x24, x25, x20, LSL #2\n" "bge 90f\n" "tbz x13, #1, 88f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" - "ldr d26, [x27], #0x8\n" - "ldr d27, [x26], #0x8\n" - "ldr d28, [x25], #0x8\n" + "ldr d26, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" + "ldr d28, [x24], #0x8\n" "tbz x13, #0, 89f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" - "ld1 { v26.s }[2], [x27]\n" - "ld1 { v27.s }[2], [x26]\n" - "ld1 { v28.s }[2], [x25]\n" + "ld1 { v25.s }[2], [x27]\n" + "ld1 { v26.s }[2], [x26]\n" + "ld1 { v27.s }[2], [x25]\n" + "ld1 { v28.s }[2], [x24]\n" "b 89f\n" "88:" // Height 5: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" - "ldr s26, [x27, #0x0]\n" - "ldr s27, [x26, #0x0]\n" - "ldr s28, [x25, #0x0]\n" + "ldr s26, [x26, #0x0]\n" + "ldr s27, [x25, #0x0]\n" + "ldr s28, [x24, #0x0]\n" "89:" // Height 5: Partial accumulate: Done "sub x11, x11, x20\n" "b 92f\n" "90:" // Height 5: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" - "ldr q26, [x27, #0x0]\n" - "ldr q27, [x26, #0x0]\n" - "ldr q28, [x25, #0x0]\n" + "ldr q25, [x27, #0x0]\n" + "ldr q26, [x26, #0x0]\n" + "ldr q27, [x25, #0x0]\n" + "ldr q28, [x24, #0x0]\n" "b 92f\n" "91:" // Height 5: no accumulate "movi v24.16b, #0x0\n" @@ -871,8 +869,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "93:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 94f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -919,30 +917,30 @@ void a64_hybrid_fp32_mla_8x4 ( "add x27, x27, #0x10\n" "add x26, x26, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" + "fmla v25.4s, v9.4s, v1.s[1]\n" + "fmla v26.4s, v9.4s, v2.s[1]\n" "cmp x9, #0x8\n" "add x12, x12, #0x40\n" - "prfm pldl1keep, [x27, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" "ldr q8, [x12, #0x0]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" - "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" "ldr q9, [x12, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" "ldr q10, [x12, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "ldr q0, [x28, #0x0]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "ldr q1, [x27, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" @@ -963,21 +961,21 @@ void a64_hybrid_fp32_mla_8x4 ( "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "add x24, x24, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" "sub x9, x9, #0x4\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x25, #0x80]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" @@ -996,11 +994,11 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr s18, [x25], #0x4\n" "ldr s17, [x24], #0x4\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v21.s[0]\n" "fmla v25.4s, v16.4s, v20.s[0]\n" "fmla v26.4s, v16.4s, v19.s[0]\n" "fmla v27.4s, v16.4s, v18.s[0]\n" + "add x12, x12, #0x10\n" "fmla v28.4s, v16.4s, v17.s[0]\n" "cbnz x9, 99b\n" "100:" // Height 5: Multiply loop: No odd multiplies @@ -1009,19 +1007,19 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 93b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x27, x11, x20, LSL #2\n" + "add x26, x27, x20, LSL #2\n" "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" - "add x27, x28, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x20, LSL #2\n" "prfm pstl1keep, [x26, #0x0]\n" - "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "prfm pstl1keep, [x24, #0x0]\n" "tbz %x[flags], #1, 101f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -1038,47 +1036,47 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 104f\n" "tbz x13, #1, 102f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" - "str d26, [x27], #0x8\n" - "str d27, [x26], #0x8\n" - "str d28, [x25], #0x8\n" + "str d25, [x27], #0x8\n" + "str d26, [x26], #0x8\n" + "str d27, [x25], #0x8\n" + "str d28, [x24], #0x8\n" "tbz x13, #0, 103f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" - "st1 { v26.s }[2], [x27]\n" - "st1 { v27.s }[2], [x26]\n" - "st1 { v28.s }[2], [x25]\n" + "st1 { v25.s }[2], [x27]\n" + "st1 { v26.s }[2], [x26]\n" + "st1 { v27.s }[2], [x25]\n" + "st1 { v28.s }[2], [x24]\n" "b 103f\n" "102:" // Height 5: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" - "str s26, [x27, #0x0]\n" - "str s27, [x26, #0x0]\n" - "str s28, [x25, #0x0]\n" + "str s25, [x27, #0x0]\n" + "str s26, [x26, #0x0]\n" + "str s27, [x25, #0x0]\n" + "str s28, [x24, #0x0]\n" "103:" // Height 5: Partial direct writeback: Done "b 105f\n" "104:" // Height 5: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" - "str q26, [x27, #0x0]\n" - "str q27, [x26, #0x0]\n" - "str q28, [x25, #0x0]\n" + "str q25, [x27, #0x0]\n" + "str q26, [x26, #0x0]\n" + "str q27, [x25, #0x0]\n" + "str q28, [x24, #0x0]\n" "105:" // Height 5: Writeback done "subs x13, x13, #0x4\n" "bgt 86b\n" "b 170f\n" "106:" // Height 6 - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "107:" // Height 6: Column loop "cbz x14, 108f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x14, x14, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "mov v29.16b, v24.16b\n" @@ -1086,47 +1084,47 @@ void a64_hybrid_fp32_mla_8x4 ( "108:" // Height 6: no bias "tbz %x[flags], #0, 112f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" + "cmp x13, #0x4\n" + "add x23, x24, x20, LSL #2\n" "bge 111f\n" "tbz x13, #1, 109f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" - "ldr d26, [x27], #0x8\n" - "ldr d27, [x26], #0x8\n" - "ldr d28, [x25], #0x8\n" - "ldr d29, [x24], #0x8\n" + "ldr d26, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" + "ldr d28, [x24], #0x8\n" + "ldr d29, [x23], #0x8\n" "tbz x13, #0, 110f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" - "ld1 { v26.s }[2], [x27]\n" - "ld1 { v27.s }[2], [x26]\n" - "ld1 { v28.s }[2], [x25]\n" - "ld1 { v29.s }[2], [x24]\n" + "ld1 { v25.s }[2], [x27]\n" + "ld1 { v26.s }[2], [x26]\n" + "ld1 { v27.s }[2], [x25]\n" + "ld1 { v28.s }[2], [x24]\n" + "ld1 { v29.s }[2], [x23]\n" "b 110f\n" "109:" // Height 6: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" - "ldr s26, [x27, #0x0]\n" - "ldr s27, [x26, #0x0]\n" - "ldr s28, [x25, #0x0]\n" - "ldr s29, [x24, #0x0]\n" + "ldr s26, [x26, #0x0]\n" + "ldr s27, [x25, #0x0]\n" + "ldr s28, [x24, #0x0]\n" + "ldr s29, [x23, #0x0]\n" "110:" // Height 6: Partial accumulate: Done "sub x11, x11, x20\n" "b 113f\n" "111:" // Height 6: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" - "ldr q26, [x27, #0x0]\n" - "ldr q27, [x26, #0x0]\n" - "ldr q28, [x25, #0x0]\n" - "ldr q29, [x24, #0x0]\n" + "ldr q25, [x27, #0x0]\n" + "ldr q26, [x26, #0x0]\n" + "ldr q27, [x25, #0x0]\n" + "ldr q28, [x24, #0x0]\n" + "ldr q29, [x23, #0x0]\n" "b 113f\n" "112:" // Height 6: no accumulate "movi v24.16b, #0x0\n" @@ -1139,8 +1137,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "114:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 115f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1194,30 +1192,30 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v29.4s, v8.4s, v5.s[0]\n" "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" - "cmp x9, #0x8\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "add x12, x12, #0x40\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q8, [x12, #0x0]\n" + "add x23, x23, #0x10\n" + "cmp x9, #0x8\n" "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x12, x12, #0x40\n" + "ldr q8, [x12, #0x0]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" "ldr q9, [x12, #0x10]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" "fmla v29.4s, v10.4s, v5.s[2]\n" "ldr q10, [x12, #0x20]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "ldr q0, [x28, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" @@ -1245,22 +1243,22 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v29.4s, v8.4s, v5.s[0]\n" "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" - "sub x9, x9, #0x4\n" - "add x12, x12, #0x40\n" "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "sub x9, x9, #0x4\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" @@ -1282,9 +1280,9 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr s18, [x24], #0x4\n" "ldr s17, [x23], #0x4\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v22.s[0]\n" "fmla v25.4s, v16.4s, v21.s[0]\n" + "add x12, x12, #0x10\n" "fmla v26.4s, v16.4s, v20.s[0]\n" "fmla v27.4s, v16.4s, v19.s[0]\n" "fmla v28.4s, v16.4s, v18.s[0]\n" @@ -1296,21 +1294,21 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 114b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" - "add x27, x28, x20, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" + "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x23, x24, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" "prfm pstl1keep, [x24, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 122f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -1329,51 +1327,51 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 125f\n" "tbz x13, #1, 123f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" - "str d26, [x27], #0x8\n" - "str d27, [x26], #0x8\n" - "str d28, [x25], #0x8\n" - "str d29, [x24], #0x8\n" + "str d25, [x27], #0x8\n" + "str d26, [x26], #0x8\n" + "str d27, [x25], #0x8\n" + "str d28, [x24], #0x8\n" + "str d29, [x23], #0x8\n" "tbz x13, #0, 124f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" - "st1 { v26.s }[2], [x27]\n" - "st1 { v27.s }[2], [x26]\n" - "st1 { v28.s }[2], [x25]\n" - "st1 { v29.s }[2], [x24]\n" + "st1 { v25.s }[2], [x27]\n" + "st1 { v26.s }[2], [x26]\n" + "st1 { v27.s }[2], [x25]\n" + "st1 { v28.s }[2], [x24]\n" + "st1 { v29.s }[2], [x23]\n" "b 124f\n" "123:" // Height 6: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" - "str s26, [x27, #0x0]\n" - "str s27, [x26, #0x0]\n" - "str s28, [x25, #0x0]\n" - "str s29, [x24, #0x0]\n" + "str s25, [x27, #0x0]\n" + "str s26, [x26, #0x0]\n" + "str s27, [x25, #0x0]\n" + "str s28, [x24, #0x0]\n" + "str s29, [x23, #0x0]\n" "124:" // Height 6: Partial direct writeback: Done "b 126f\n" "125:" // Height 6: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" - "str q26, [x27, #0x0]\n" - "str q27, [x26, #0x0]\n" - "str q28, [x25, #0x0]\n" - "str q29, [x24, #0x0]\n" + "str q25, [x27, #0x0]\n" + "str q26, [x26, #0x0]\n" + "str q27, [x25, #0x0]\n" + "str q28, [x24, #0x0]\n" + "str q29, [x23, #0x0]\n" "126:" // Height 6: Writeback done "subs x13, x13, #0x4\n" "bgt 107b\n" "b 170f\n" "127:" // Height 7 - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" "128:" // Height 7: Column loop "cbz x14, 129f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x14, x14, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "mov v29.16b, v24.16b\n" @@ -1382,52 +1380,52 @@ void a64_hybrid_fp32_mla_8x4 ( "129:" // Height 7: no bias "tbz %x[flags], #0, 133f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "cmp x13, #0x4\n" + "add x22, x23, x20, LSL #2\n" "bge 132f\n" "tbz x13, #1, 130f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" - "ldr d26, [x27], #0x8\n" - "ldr d27, [x26], #0x8\n" - "ldr d28, [x25], #0x8\n" - "ldr d29, [x24], #0x8\n" - "ldr d30, [x23], #0x8\n" + "ldr d26, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" + "ldr d28, [x24], #0x8\n" + "ldr d29, [x23], #0x8\n" + "ldr d30, [x22], #0x8\n" "tbz x13, #0, 131f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" - "ld1 { v26.s }[2], [x27]\n" - "ld1 { v27.s }[2], [x26]\n" - "ld1 { v28.s }[2], [x25]\n" - "ld1 { v29.s }[2], [x24]\n" - "ld1 { v30.s }[2], [x23]\n" + "ld1 { v25.s }[2], [x27]\n" + "ld1 { v26.s }[2], [x26]\n" + "ld1 { v27.s }[2], [x25]\n" + "ld1 { v28.s }[2], [x24]\n" + "ld1 { v29.s }[2], [x23]\n" + "ld1 { v30.s }[2], [x22]\n" "b 131f\n" "130:" // Height 7: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" - "ldr s26, [x27, #0x0]\n" - "ldr s27, [x26, #0x0]\n" - "ldr s28, [x25, #0x0]\n" - "ldr s29, [x24, #0x0]\n" - "ldr s30, [x23, #0x0]\n" + "ldr s26, [x26, #0x0]\n" + "ldr s27, [x25, #0x0]\n" + "ldr s28, [x24, #0x0]\n" + "ldr s29, [x23, #0x0]\n" + "ldr s30, [x22, #0x0]\n" "131:" // Height 7: Partial accumulate: Done "sub x11, x11, x20\n" "b 134f\n" "132:" // Height 7: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" - "ldr q26, [x27, #0x0]\n" - "ldr q27, [x26, #0x0]\n" - "ldr q28, [x25, #0x0]\n" - "ldr q29, [x24, #0x0]\n" - "ldr q30, [x23, #0x0]\n" + "ldr q25, [x27, #0x0]\n" + "ldr q26, [x26, #0x0]\n" + "ldr q27, [x25, #0x0]\n" + "ldr q28, [x24, #0x0]\n" + "ldr q29, [x23, #0x0]\n" + "ldr q30, [x22, #0x0]\n" "b 134f\n" "133:" // Height 7: no accumulate "movi v24.16b, #0x0\n" @@ -1441,8 +1439,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "135:" // Height 7: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 136f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1501,32 +1499,32 @@ void a64_hybrid_fp32_mla_8x4 ( "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" "fmla v30.4s, v8.4s, v6.s[0]\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" + "fmla v26.4s, v9.4s, v2.s[1]\n" "cmp x9, #0x8\n" "add x12, x12, #0x40\n" "ldr q8, [x12, #0x0]\n" - "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" + "prfm pldl1keep, [x28, #0x80]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla v30.4s, v9.4s, v6.s[1]\n" "ldr q9, [x12, #0x10]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" "fmla v29.4s, v10.4s, v5.s[2]\n" + "prfm pldl1keep, [x22, #0x80]\n" "fmla v30.4s, v10.4s, v6.s[2]\n" "ldr q10, [x12, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" @@ -1559,25 +1557,25 @@ void a64_hybrid_fp32_mla_8x4 ( "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" "fmla v30.4s, v8.4s, v6.s[0]\n" - "add x22, x22, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "fmla v25.4s, v9.4s, v1.s[1]\n" + "add x22, x22, #0x10\n" "sub x9, x9, #0x4\n" - "add x12, x12, #0x40\n" + "fmla v25.4s, v9.4s, v1.s[1]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" + "fmla v28.4s, v9.4s, v4.s[1]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v28.4s, v9.4s, v4.s[1]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" + "fmla v30.4s, v9.4s, v6.s[1]\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v30.4s, v9.4s, v6.s[1]\n" - "prfm pldl1keep, [x22, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" @@ -1602,11 +1600,11 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr s18, [x23], #0x4\n" "ldr s17, [x22], #0x4\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v23.s[0]\n" "fmla v25.4s, v16.4s, v22.s[0]\n" "fmla v26.4s, v16.4s, v21.s[0]\n" "fmla v27.4s, v16.4s, v20.s[0]\n" + "add x12, x12, #0x10\n" "fmla v28.4s, v16.4s, v19.s[0]\n" "fmla v29.4s, v16.4s, v18.s[0]\n" "fmla v30.4s, v16.4s, v17.s[0]\n" @@ -1617,23 +1615,23 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 135b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" - "add x27, x28, x20, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" + "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x26, #0x0]\n" "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" "prfm pstl1keep, [x24, #0x0]\n" "prfm pstl1keep, [x23, #0x0]\n" + "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 143f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -1654,59 +1652,58 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 146f\n" "tbz x13, #1, 144f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" - "str d26, [x27], #0x8\n" - "str d27, [x26], #0x8\n" - "str d28, [x25], #0x8\n" - "str d29, [x24], #0x8\n" - "str d30, [x23], #0x8\n" + "str d25, [x27], #0x8\n" + "str d26, [x26], #0x8\n" + "str d27, [x25], #0x8\n" + "str d28, [x24], #0x8\n" + "str d29, [x23], #0x8\n" + "str d30, [x22], #0x8\n" "tbz x13, #0, 145f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" - "st1 { v26.s }[2], [x27]\n" - "st1 { v27.s }[2], [x26]\n" - "st1 { v28.s }[2], [x25]\n" - "st1 { v29.s }[2], [x24]\n" - "st1 { v30.s }[2], [x23]\n" + "st1 { v25.s }[2], [x27]\n" + "st1 { v26.s }[2], [x26]\n" + "st1 { v27.s }[2], [x25]\n" + "st1 { v28.s }[2], [x24]\n" + "st1 { v29.s }[2], [x23]\n" + "st1 { v30.s }[2], [x22]\n" "b 145f\n" "144:" // Height 7: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" - "str s26, [x27, #0x0]\n" - "str s27, [x26, #0x0]\n" - "str s28, [x25, #0x0]\n" - "str s29, [x24, #0x0]\n" - "str s30, [x23, #0x0]\n" + "str s25, [x27, #0x0]\n" + "str s26, [x26, #0x0]\n" + "str s27, [x25, #0x0]\n" + "str s28, [x24, #0x0]\n" + "str s29, [x23, #0x0]\n" + "str s30, [x22, #0x0]\n" "145:" // Height 7: Partial direct writeback: Done "b 147f\n" "146:" // Height 7: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" - "str q26, [x27, #0x0]\n" - "str q27, [x26, #0x0]\n" - "str q28, [x25, #0x0]\n" - "str q29, [x24, #0x0]\n" - "str q30, [x23, #0x0]\n" + "str q25, [x27, #0x0]\n" + "str q26, [x26, #0x0]\n" + "str q27, [x25, #0x0]\n" + "str q28, [x24, #0x0]\n" + "str q29, [x23, #0x0]\n" + "str q30, [x22, #0x0]\n" "147:" // Height 7: Writeback done "subs x13, x13, #0x4\n" "bgt 128b\n" "b 170f\n" "148:" // Height 8 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" - "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x20\n" - "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x14, %x[bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_N]]\n" "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "madd x20, x21, x20, x11\n" - "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x11, %x[output_ptr]\n" + "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" "149:" // Height 8: Column loop "cbz x14, 150f\n" "ldr q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x14, x14, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "mov v29.16b, v24.16b\n" @@ -1716,57 +1713,57 @@ void a64_hybrid_fp32_mla_8x4 ( "150:" // Height 8: no bias "tbz %x[flags], #0, 154f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x13, #0x4\n" - "add x28, x11, x20, LSL #2\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "cmp x13, #0x4\n" + "add x21, x22, x20, LSL #2\n" "bge 153f\n" "tbz x13, #1, 151f\n" "ldr d24, [x11], #0x8\n" - "ldr d25, [x28], #0x8\n" + "ldr d25, [x27], #0x8\n" "mov x20, #0x8\n" - "ldr d26, [x27], #0x8\n" - "ldr d27, [x26], #0x8\n" - "ldr d28, [x25], #0x8\n" - "ldr d29, [x24], #0x8\n" - "ldr d30, [x23], #0x8\n" - "ldr d31, [x22], #0x8\n" + "ldr d26, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" + "ldr d28, [x24], #0x8\n" + "ldr d29, [x23], #0x8\n" + "ldr d30, [x22], #0x8\n" + "ldr d31, [x21], #0x8\n" "tbz x13, #0, 152f\n" "ld1 { v24.s }[2], [x11]\n" - "ld1 { v25.s }[2], [x28]\n" - "ld1 { v26.s }[2], [x27]\n" - "ld1 { v27.s }[2], [x26]\n" - "ld1 { v28.s }[2], [x25]\n" - "ld1 { v29.s }[2], [x24]\n" - "ld1 { v30.s }[2], [x23]\n" - "ld1 { v31.s }[2], [x22]\n" + "ld1 { v25.s }[2], [x27]\n" + "ld1 { v26.s }[2], [x26]\n" + "ld1 { v27.s }[2], [x25]\n" + "ld1 { v28.s }[2], [x24]\n" + "ld1 { v29.s }[2], [x23]\n" + "ld1 { v30.s }[2], [x22]\n" + "ld1 { v31.s }[2], [x21]\n" "b 152f\n" "151:" // Height 8: Partial accumulate: partial_1_0 "ldr s24, [x11, #0x0]\n" - "ldr s25, [x28, #0x0]\n" + "ldr s25, [x27, #0x0]\n" "mov x20, #0x0\n" - "ldr s26, [x27, #0x0]\n" - "ldr s27, [x26, #0x0]\n" - "ldr s28, [x25, #0x0]\n" - "ldr s29, [x24, #0x0]\n" - "ldr s30, [x23, #0x0]\n" - "ldr s31, [x22, #0x0]\n" + "ldr s26, [x26, #0x0]\n" + "ldr s27, [x25, #0x0]\n" + "ldr s28, [x24, #0x0]\n" + "ldr s29, [x23, #0x0]\n" + "ldr s30, [x22, #0x0]\n" + "ldr s31, [x21, #0x0]\n" "152:" // Height 8: Partial accumulate: Done "sub x11, x11, x20\n" "b 155f\n" "153:" // Height 8: full accumulate "ldr q24, [x11, #0x0]\n" - "ldr q25, [x28, #0x0]\n" - "ldr q26, [x27, #0x0]\n" - "ldr q27, [x26, #0x0]\n" - "ldr q28, [x25, #0x0]\n" - "ldr q29, [x24, #0x0]\n" - "ldr q30, [x23, #0x0]\n" - "ldr q31, [x22, #0x0]\n" + "ldr q25, [x27, #0x0]\n" + "ldr q26, [x26, #0x0]\n" + "ldr q27, [x25, #0x0]\n" + "ldr q28, [x24, #0x0]\n" + "ldr q29, [x23, #0x0]\n" + "ldr q30, [x22, #0x0]\n" + "ldr q31, [x21, #0x0]\n" "b 155f\n" "154:" // Height 8: no accumulate "movi v24.16b, #0x0\n" @@ -1781,8 +1778,8 @@ void a64_hybrid_fp32_mla_8x4 ( "mov x10, #0x0\n" "156:" // Height 8: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 157f\n" "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1855,10 +1852,10 @@ void a64_hybrid_fp32_mla_8x4 ( "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" "add x12, x12, #0x40\n" - "prfm pldl1keep, [x28, #0x80]\n" "ldr q8, [x12, #0x0]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" + "prfm pldl1keep, [x28, #0x80]\n" "prfm pldl1keep, [x27, #0x80]\n" "fmla v30.4s, v9.4s, v6.s[1]\n" "fmla v31.4s, v9.4s, v7.s[1]\n" @@ -1915,24 +1912,24 @@ void a64_hybrid_fp32_mla_8x4 ( "add x21, x21, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" + "sub x9, x9, #0x4\n" "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "sub x9, x9, #0x4\n" - "add x12, x12, #0x40\n" + "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla v30.4s, v9.4s, v6.s[1]\n" "fmla v31.4s, v9.4s, v7.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x22, #0x80]\n" "prfm pldl1keep, [x21, #0x80]\n" + "add x12, x12, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" @@ -1960,9 +1957,9 @@ void a64_hybrid_fp32_mla_8x4 ( "ldr s18, [x22], #0x4\n" "ldr s17, [x21], #0x4\n" "ldr q16, [x12, #0x0]\n" - "add x12, x12, #0x10\n" "fmla v24.4s, v16.4s, v0.s[0]\n" "fmla v25.4s, v16.4s, v23.s[0]\n" + "add x12, x12, #0x10\n" "fmla v26.4s, v16.4s, v22.s[0]\n" "fmla v27.4s, v16.4s, v21.s[0]\n" "fmla v28.4s, v16.4s, v20.s[0]\n" @@ -1976,25 +1973,25 @@ void a64_hybrid_fp32_mla_8x4 ( "cmp x10, x20\n" "bne 156b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x11, #0x0]\n" - "add x28, x11, x20, LSL #2\n" - "prfm pstl1keep, [x28, #0x0]\n" - "add x27, x28, x20, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" + "add x27, x11, x20, LSL #2\n" "add x26, x27, x20, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" + "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x26, #0x0]\n" "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" "prfm pstl1keep, [x24, #0x0]\n" - "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" + "prfm pstl1keep, [x21, #0x0]\n" "tbz %x[flags], #1, 164f\n" - "add x21, %x[args_ptr], %[offset_max]\n" + "add x20, %x[args_ptr], %[offset_max]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v24.4s, v24.4s, v17.4s\n" "fmin v25.4s, v25.4s, v17.4s\n" @@ -2017,44 +2014,44 @@ void a64_hybrid_fp32_mla_8x4 ( "bge 167f\n" "tbz x13, #1, 165f\n" "str d24, [x11], #0x8\n" - "str d25, [x28], #0x8\n" - "str d26, [x27], #0x8\n" - "str d27, [x26], #0x8\n" - "str d28, [x25], #0x8\n" - "str d29, [x24], #0x8\n" - "str d30, [x23], #0x8\n" - "str d31, [x22], #0x8\n" + "str d25, [x27], #0x8\n" + "str d26, [x26], #0x8\n" + "str d27, [x25], #0x8\n" + "str d28, [x24], #0x8\n" + "str d29, [x23], #0x8\n" + "str d30, [x22], #0x8\n" + "str d31, [x21], #0x8\n" "tbz x13, #0, 166f\n" "st1 { v24.s }[2], [x11]\n" - "st1 { v25.s }[2], [x28]\n" - "st1 { v26.s }[2], [x27]\n" - "st1 { v27.s }[2], [x26]\n" - "st1 { v28.s }[2], [x25]\n" - "st1 { v29.s }[2], [x24]\n" - "st1 { v30.s }[2], [x23]\n" - "st1 { v31.s }[2], [x22]\n" + "st1 { v25.s }[2], [x27]\n" + "st1 { v26.s }[2], [x26]\n" + "st1 { v27.s }[2], [x25]\n" + "st1 { v28.s }[2], [x24]\n" + "st1 { v29.s }[2], [x23]\n" + "st1 { v30.s }[2], [x22]\n" + "st1 { v31.s }[2], [x21]\n" "b 166f\n" "165:" // Height 8: Partial direct writeback: partial_1_0 "str s24, [x11, #0x0]\n" - "str s25, [x28, #0x0]\n" - "str s26, [x27, #0x0]\n" - "str s27, [x26, #0x0]\n" - "str s28, [x25, #0x0]\n" - "str s29, [x24, #0x0]\n" - "str s30, [x23, #0x0]\n" - "str s31, [x22, #0x0]\n" + "str s25, [x27, #0x0]\n" + "str s26, [x26, #0x0]\n" + "str s27, [x25, #0x0]\n" + "str s28, [x24, #0x0]\n" + "str s29, [x23, #0x0]\n" + "str s30, [x22, #0x0]\n" + "str s31, [x21, #0x0]\n" "166:" // Height 8: Partial direct writeback: Done "b 168f\n" "167:" // Height 8: Full writeback "str q24, [x11, #0x0]\n" "add x11, x11, #0x10\n" - "str q25, [x28, #0x0]\n" - "str q26, [x27, #0x0]\n" - "str q27, [x26, #0x0]\n" - "str q28, [x25, #0x0]\n" - "str q29, [x24, #0x0]\n" - "str q30, [x23, #0x0]\n" - "str q31, [x22, #0x0]\n" + "str q25, [x27, #0x0]\n" + "str q26, [x26, #0x0]\n" + "str q27, [x25, #0x0]\n" + "str q28, [x24, #0x0]\n" + "str q29, [x23, #0x0]\n" + "str q30, [x22, #0x0]\n" + "str q31, [x21, #0x0]\n" "168:" // Height 8: Writeback done "subs x13, x13, #0x4\n" "bgt 149b\n" @@ -2070,8 +2067,8 @@ void a64_hybrid_fp32_mla_8x4 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "170:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) - : [args_ptr] "r" (&ka), [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_bias] "I" (offsetof(KernelArgs, bias)), [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), [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", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |