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 | 338 |
1 files changed, 168 insertions, 170 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 bd22336c8d..004e5d7f23 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 @@ -92,7 +92,6 @@ void a64_hybrid_fp32_mla_8x4 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x8\n" "bge 148f\n" @@ -140,11 +139,11 @@ void a64_hybrid_fp32_mla_8x4 ( "9:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 10f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" "cbnz x10, 11f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -189,10 +188,10 @@ void a64_hybrid_fp32_mla_8x4 ( "14:" // Height 1: Multiply loop: Main loop skip "cbz x9, 16f\n" "15:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr q12, [x12, #0x0]\n" + "ldr s17, [x28], #0x4\n" + "ldr q16, [x12, #0x0]\n" "sub x9, x9, #0x1\n" - "fmla v24.4s, v12.4s, v0.s[0]\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 @@ -271,12 +270,12 @@ void a64_hybrid_fp32_mla_8x4 ( "30:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 31f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" "cbnz x10, 32f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -284,7 +283,7 @@ void a64_hybrid_fp32_mla_8x4 ( "b 32f\n" "31:" // Height 2: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" "32:" // Height 2: input setup done "cmp x9, #0x4\n" "blt 35f\n" @@ -337,12 +336,12 @@ void a64_hybrid_fp32_mla_8x4 ( "35:" // Height 2: Multiply loop: Main loop skip "cbz x9, 37f\n" "36:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s18, [x28], #0x4\n" + "ldr s17, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr q16, [x12, #0x0]\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 @@ -437,13 +436,13 @@ void a64_hybrid_fp32_mla_8x4 ( "51:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 52f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" "cbnz x10, 53f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -452,8 +451,8 @@ void a64_hybrid_fp32_mla_8x4 ( "b 53f\n" "52:" // Height 3: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" "53:" // Height 3: input setup done "cmp x9, #0x4\n" "blt 56f\n" @@ -520,14 +519,14 @@ void a64_hybrid_fp32_mla_8x4 ( "56:" // Height 3: Multiply loop: Main loop skip "cbz x9, 58f\n" "57:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s19, [x28], #0x4\n" + "ldr s18, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" + "ldr s17, [x26], #0x4\n" + "ldr q16, [x12, #0x0]\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 @@ -637,14 +636,14 @@ void a64_hybrid_fp32_mla_8x4 ( "72:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 73f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" "cbnz x10, 74f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -654,9 +653,9 @@ void a64_hybrid_fp32_mla_8x4 ( "b 74f\n" "73:" // Height 4: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" - "add x25, x26, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "74:" // Height 4: input setup done "cmp x9, #0x4\n" "blt 77f\n" @@ -737,17 +736,17 @@ void a64_hybrid_fp32_mla_8x4 ( "77:" // Height 4: Multiply loop: Main loop skip "cbz x9, 79f\n" "78:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s20, [x28], #0x4\n" + "ldr s19, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr s18, [x26], #0x4\n" + "ldr s17, [x25], #0x4\n" + "ldr q16, [x12, #0x0]\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, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" + "fmla v26.4s, v16.4s, v18.s[0]\n" + "fmla v27.4s, v16.4s, v17.s[0]\n" "cbnz x9, 78b\n" "79:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -871,15 +870,15 @@ void a64_hybrid_fp32_mla_8x4 ( "93:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 94f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" "cbnz x10, 95f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -890,10 +889,10 @@ void a64_hybrid_fp32_mla_8x4 ( "b 95f\n" "94:" // Height 5: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, 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 x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "95:" // Height 5: input setup done "cmp x9, #0x4\n" "blt 98f\n" @@ -988,19 +987,19 @@ void a64_hybrid_fp32_mla_8x4 ( "98:" // Height 5: Multiply loop: Main loop skip "cbz x9, 100f\n" "99:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s21, [x28], #0x4\n" + "ldr s20, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" + "ldr s19, [x26], #0x4\n" + "ldr s18, [x25], #0x4\n" + "ldr s17, [x24], #0x4\n" + "ldr q16, [x12, #0x0]\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, v12.4s, v4.s[0]\n" + "fmla v28.4s, v16.4s, v17.s[0]\n" "cbnz x9, 99b\n" "100:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1139,16 +1138,16 @@ void a64_hybrid_fp32_mla_8x4 ( "114:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 115f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" - "ldr x23, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" + "ldr x23, [x20, #0x28]\n" "cbnz x10, 116f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -1160,11 +1159,11 @@ void a64_hybrid_fp32_mla_8x4 ( "b 116f\n" "115:" // Height 6: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, 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 x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "116:" // Height 6: input setup done "cmp x9, #0x4\n" "blt 119f\n" @@ -1273,21 +1272,21 @@ void a64_hybrid_fp32_mla_8x4 ( "119:" // Height 6: Multiply loop: Main loop skip "cbz x9, 121f\n" "120:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s22, [x28], #0x4\n" + "ldr s21, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr s20, [x26], #0x4\n" + "ldr s19, [x25], #0x4\n" + "ldr s18, [x24], #0x4\n" + "ldr s17, [x23], #0x4\n" + "ldr q16, [x12, #0x0]\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, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\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" + "fmla v29.4s, v16.4s, v17.s[0]\n" "cbnz x9, 120b\n" "121:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1441,17 +1440,17 @@ void a64_hybrid_fp32_mla_8x4 ( "135:" // Height 7: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 136f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" - "ldr x23, [x21, #0x28]\n" - "ldr x22, [x21, #0x30]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" + "ldr x23, [x20, #0x28]\n" + "ldr x22, [x20, #0x30]\n" "cbnz x10, 137f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -1464,12 +1463,12 @@ void a64_hybrid_fp32_mla_8x4 ( "b 137f\n" "136:" // Height 7: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, 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" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "137:" // Height 7: input setup done "cmp x9, #0x4\n" "blt 140f\n" @@ -1592,23 +1591,23 @@ void a64_hybrid_fp32_mla_8x4 ( "140:" // Height 7: Multiply loop: Main loop skip "cbz x9, 142f\n" "141:" // Height 7: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s23, [x28], #0x4\n" + "ldr s22, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr s6, [x22], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" + "ldr s21, [x26], #0x4\n" + "ldr s20, [x25], #0x4\n" + "ldr s19, [x24], #0x4\n" + "ldr s18, [x23], #0x4\n" + "ldr s17, [x22], #0x4\n" + "ldr q16, [x12, #0x0]\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, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "fmla v30.4s, v12.4s, v6.s[0]\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" "cbnz x9, 141b\n" "142:" // Height 7: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1780,18 +1779,18 @@ void a64_hybrid_fp32_mla_8x4 ( "156:" // Height 8: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, 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, 157f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" - "ldr x23, [x21, #0x28]\n" - "ldr x22, [x21, #0x30]\n" - "ldr x21, [x21, #0x38]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" + "ldr x23, [x20, #0x28]\n" + "ldr x22, [x20, #0x30]\n" + "ldr x21, [x20, #0x38]\n" "cbnz x10, 158f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -1805,13 +1804,13 @@ void a64_hybrid_fp32_mla_8x4 ( "b 158f\n" "157:" // Height 8: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, 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" - "add x21, x22, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "158:" // Height 8: input setup done "cmp x9, #0x4\n" "blt 161f\n" @@ -1949,24 +1948,24 @@ void a64_hybrid_fp32_mla_8x4 ( "cbz x9, 163f\n" "162:" // Height 8: Multiply loop: Odd block loop "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s23, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr s6, [x22], #0x4\n" - "ldr s7, [x21], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr s22, [x26], #0x4\n" + "ldr s21, [x25], #0x4\n" + "ldr s20, [x24], #0x4\n" + "ldr s19, [x23], #0x4\n" + "ldr s18, [x22], #0x4\n" + "ldr s17, [x21], #0x4\n" + "ldr q16, [x12, #0x0]\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, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "fmla v30.4s, v12.4s, v6.s[0]\n" - "fmla v31.4s, v12.4s, v7.s[0]\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" + "fmla v29.4s, v16.4s, v19.s[0]\n" + "fmla v30.4s, v16.4s, v18.s[0]\n" + "fmla v31.4s, v16.4s, v17.s[0]\n" "cbnz x9, 162b\n" "163:" // Height 8: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -2068,10 +2067,9 @@ 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), [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", "v16", "v17", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "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" ); } |