aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp1186
1 files changed, 560 insertions, 626 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
index c820da3ef1..85944e9f6a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -96,13 +96,13 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"1:" // Row loop
"cmp %x[M], #0x6\n"
- "bge 186f\n"
+ "bge 181f\n"
"cmp %x[M], #0x4\n"
- "bgt 149f\n"
- "beq 112f\n"
+ "bgt 145f\n"
+ "beq 109f\n"
"cmp %x[M], #0x2\n"
- "bgt 75f\n"
- "beq 38f\n"
+ "bgt 73f\n"
+ "beq 37f\n"
"ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
"ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
"mov x14, %x[bias]\n"
@@ -291,7 +291,7 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n"
".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n"
"21:" // Height 1: Multiply loop: Main loop skip
- "cbz x11, 26f\n"
+ "cbz x11, 25f\n"
"cmp x11, #0x2\n"
"blt 23f\n"
"22:" // Height 1: Multiply loop: Odd block loop
@@ -308,16 +308,10 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n"
"cmp x11, #0x2\n"
"bge 22b\n"
- "cbz x11, 26f\n"
+ "cbz x11, 25f\n"
"23:" // Height 1: Multiply loop: Skip odd blocks
- "tbz x11, #1, 24f\n"
- "ldr s0, [x10], #0x4\n"
- "tbz x11, #0, 25f\n"
- "ld1 { v0.h }[2], [x10]\n"
- "b 25f\n"
- "24:" // Height 1: Multiply loop: Ragged operand read: partial_1_0
"ldr h0, [x10, #0x0]\n"
- "25:" // Height 1: Multiply loop: Ragged operand read: Done
+ "24:" // Height 1: Multiply loop: Ragged operand read: Done
"ldr q6, [x15, #0x0]\n"
".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n"
"ldr q7, [x15, #0x10]\n"
@@ -327,13 +321,13 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n"
"add x15, x15, #0x40\n"
".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n"
- "26:" // Height 1: Multiply loop: No odd multiplies
+ "25:" // Height 1: Multiply loop: No odd multiplies
"ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x12, x12, #0x1\n"
"cmp x12, x19\n"
"bne 16b\n"
"prfm pstl1keep, [x13, #0x0]\n"
- "tbz %x[flags], #1, 27f\n"
+ "tbz %x[flags], #1, 26f\n"
"add x19, %x[args_ptr], %[offset_min]\n"
"ld1r { v1.4s }, [x19]\n"
"add x19, %x[args_ptr], %[offset_max]\n"
@@ -346,81 +340,81 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"fmax v9.4s, v9.4s, v1.4s\n"
"fmax v10.4s, v10.4s, v1.4s\n"
"fmax v11.4s, v11.4s, v1.4s\n"
- "27:" // Height 1: No activation
+ "26:" // Height 1: No activation
"cmp x16, #0x10\n"
- "bge 36f\n"
- "tbz x16, #3, 31f\n"
+ "bge 35f\n"
+ "tbz x16, #3, 30f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v9.4s }, [x13], #0x10\n"
- "tbz x16, #2, 29f\n"
+ "tbz x16, #2, 28f\n"
"st1 { v10.4s }, [x13], #0x10\n"
- "tbz x16, #1, 28f\n"
+ "tbz x16, #1, 27f\n"
"str d11, [x13], #0x8\n"
- "tbz x16, #0, 35f\n"
+ "tbz x16, #0, 34f\n"
"st1 { v11.s }[2], [x13]\n"
- "b 35f\n"
- "28:" // Height 1: Partial direct writeback: partial_1_12
- "tbz x16, #0, 35f\n"
+ "b 34f\n"
+ "27:" // Height 1: Partial direct writeback: partial_1_12
+ "tbz x16, #0, 34f\n"
"str s11, [x13, #0x0]\n"
- "b 35f\n"
- "29:" // Height 1: Partial direct writeback: partial_2_8
- "tbz x16, #1, 30f\n"
+ "b 34f\n"
+ "28:" // Height 1: Partial direct writeback: partial_2_8
+ "tbz x16, #1, 29f\n"
"str d10, [x13], #0x8\n"
- "tbz x16, #0, 35f\n"
+ "tbz x16, #0, 34f\n"
"st1 { v10.s }[2], [x13]\n"
- "b 35f\n"
- "30:" // Height 1: Partial direct writeback: partial_1_8
- "tbz x16, #0, 35f\n"
+ "b 34f\n"
+ "29:" // Height 1: Partial direct writeback: partial_1_8
+ "tbz x16, #0, 34f\n"
"str s10, [x13, #0x0]\n"
- "b 35f\n"
- "31:" // Height 1: Partial direct writeback: partial_4_0
- "tbz x16, #2, 33f\n"
+ "b 34f\n"
+ "30:" // Height 1: Partial direct writeback: partial_4_0
+ "tbz x16, #2, 32f\n"
"st1 { v8.4s }, [x13], #0x10\n"
- "tbz x16, #1, 32f\n"
+ "tbz x16, #1, 31f\n"
"str d9, [x13], #0x8\n"
- "tbz x16, #0, 35f\n"
+ "tbz x16, #0, 34f\n"
"st1 { v9.s }[2], [x13]\n"
- "b 35f\n"
- "32:" // Height 1: Partial direct writeback: partial_1_4
- "tbz x16, #0, 35f\n"
+ "b 34f\n"
+ "31:" // Height 1: Partial direct writeback: partial_1_4
+ "tbz x16, #0, 34f\n"
"str s9, [x13, #0x0]\n"
- "b 35f\n"
- "33:" // Height 1: Partial direct writeback: partial_2_0
- "tbz x16, #1, 34f\n"
+ "b 34f\n"
+ "32:" // Height 1: Partial direct writeback: partial_2_0
+ "tbz x16, #1, 33f\n"
"str d8, [x13], #0x8\n"
- "tbz x16, #0, 35f\n"
+ "tbz x16, #0, 34f\n"
"st1 { v8.s }[2], [x13]\n"
- "b 35f\n"
- "34:" // Height 1: Partial direct writeback: partial_1_0
+ "b 34f\n"
+ "33:" // Height 1: Partial direct writeback: partial_1_0
"str s8, [x13, #0x0]\n"
- "35:" // Height 1: Partial direct writeback: Done
- "b 37f\n"
- "36:" // Height 1: Full writeback
+ "34:" // Height 1: Partial direct writeback: Done
+ "b 36f\n"
+ "35:" // Height 1: Full writeback
"str q8, [x13, #0x0]\n"
"str q9, [x13, #0x10]\n"
"str q10, [x13, #0x20]\n"
"str q11, [x13, #0x30]\n"
"add x13, x13, #0x40\n"
- "37:" // Height 1: Writeback done
+ "36:" // Height 1: Writeback done
"subs x16, x16, #0x10\n"
"bgt 3b\n"
- "b 224f\n"
- "38:" // Height 2
+ "b 218f\n"
+ "37:" // Height 2
"ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
"mov x14, %x[bias]\n"
"ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "tbz %x[flags], #2, 39f\n"
+ "tbz %x[flags], #2, 38f\n"
"ldr x13, [%x[output_ptr], #0x0]\n"
"add x13, x13, x19, LSL #2\n"
"ldr x9, [%x[output_ptr], #0x8]\n"
"add x9, x9, x19, LSL #2\n"
- "b 40f\n"
- "39:" // Height 2: setup direct output
+ "b 39f\n"
+ "38:" // Height 2: setup direct output
"mov x13, %x[output_ptr]\n"
"add x9, x13, x19, LSL #2\n"
- "40:" // Height 2: Column loop
- "cbz x14, 41f\n"
+ "39:" // Height 2: Column loop
+ "cbz x14, 40f\n"
"ldr q8, [x14, #0x0]\n"
"mov v12.16b, v8.16b\n"
"ldr q9, [x14, #0x10]\n"
@@ -430,84 +424,84 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"mov v14.16b, v10.16b\n"
"add x14, x14, #0x40\n"
"mov v15.16b, v11.16b\n"
- "b 52f\n"
- "41:" // Height 2: no bias
- "tbz %x[flags], #0, 51f\n"
+ "b 51f\n"
+ "40:" // Height 2: no bias
+ "tbz %x[flags], #0, 50f\n"
"cmp x16, #0x10\n"
- "bge 50f\n"
- "tbz x16, #3, 45f\n"
+ "bge 49f\n"
+ "tbz x16, #3, 44f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v9.4s }, [x13], #0x10\n"
"ld1 { v13.4s }, [x9], #0x10\n"
- "tbz x16, #2, 43f\n"
+ "tbz x16, #2, 42f\n"
"ld1 { v10.4s }, [x13], #0x10\n"
"ld1 { v14.4s }, [x9], #0x10\n"
- "tbz x16, #1, 42f\n"
+ "tbz x16, #1, 41f\n"
"mov x19, #0x38\n"
"ldr d11, [x13], #0x8\n"
"ldr d15, [x9], #0x8\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ld1 { v11.s }[2], [x13]\n"
"ld1 { v15.s }[2], [x9]\n"
- "b 49f\n"
- "42:" // Height 2: Partial accumulate: partial_1_12
+ "b 48f\n"
+ "41:" // Height 2: Partial accumulate: partial_1_12
"mov x19, #0x30\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ldr s11, [x13, #0x0]\n"
"ldr s15, [x9, #0x0]\n"
- "b 49f\n"
- "43:" // Height 2: Partial accumulate: partial_2_8
- "tbz x16, #1, 44f\n"
+ "b 48f\n"
+ "42:" // Height 2: Partial accumulate: partial_2_8
+ "tbz x16, #1, 43f\n"
"ldr d10, [x13], #0x8\n"
"ldr d14, [x9], #0x8\n"
"mov x19, #0x28\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ld1 { v10.s }[2], [x13]\n"
"ld1 { v14.s }[2], [x9]\n"
- "b 49f\n"
- "44:" // Height 2: Partial accumulate: partial_1_8
+ "b 48f\n"
+ "43:" // Height 2: Partial accumulate: partial_1_8
"mov x19, #0x20\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ldr s10, [x13, #0x0]\n"
"ldr s14, [x9, #0x0]\n"
- "b 49f\n"
- "45:" // Height 2: Partial accumulate: partial_4_0
- "tbz x16, #2, 47f\n"
+ "b 48f\n"
+ "44:" // Height 2: Partial accumulate: partial_4_0
+ "tbz x16, #2, 46f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
- "tbz x16, #1, 46f\n"
+ "tbz x16, #1, 45f\n"
"mov x19, #0x18\n"
"ldr d9, [x13], #0x8\n"
"ldr d13, [x9], #0x8\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ld1 { v9.s }[2], [x13]\n"
"ld1 { v13.s }[2], [x9]\n"
- "b 49f\n"
- "46:" // Height 2: Partial accumulate: partial_1_4
+ "b 48f\n"
+ "45:" // Height 2: Partial accumulate: partial_1_4
"mov x19, #0x10\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ldr s9, [x13, #0x0]\n"
"ldr s13, [x9, #0x0]\n"
- "b 49f\n"
- "47:" // Height 2: Partial accumulate: partial_2_0
- "tbz x16, #1, 48f\n"
+ "b 48f\n"
+ "46:" // Height 2: Partial accumulate: partial_2_0
+ "tbz x16, #1, 47f\n"
"ldr d8, [x13], #0x8\n"
"ldr d12, [x9], #0x8\n"
"mov x19, #0x8\n"
- "tbz x16, #0, 49f\n"
+ "tbz x16, #0, 48f\n"
"ld1 { v8.s }[2], [x13]\n"
"ld1 { v12.s }[2], [x9]\n"
- "b 49f\n"
- "48:" // Height 2: Partial accumulate: partial_1_0
+ "b 48f\n"
+ "47:" // Height 2: Partial accumulate: partial_1_0
"mov x19, #0x0\n"
"ldr s8, [x13, #0x0]\n"
"ldr s12, [x9, #0x0]\n"
- "49:" // Height 2: Partial accumulate: Done
+ "48:" // Height 2: Partial accumulate: Done
"sub x13, x13, x19\n"
"sub x9, x9, x19\n"
- "b 52f\n"
- "50:" // Height 2: full accumulate
+ "b 51f\n"
+ "49:" // Height 2: full accumulate
"ldr q8, [x13, #0x0]\n"
"ldr q9, [x13, #0x10]\n"
"ldr q10, [x13, #0x20]\n"
@@ -516,8 +510,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr q13, [x9, #0x10]\n"
"ldr q14, [x9, #0x20]\n"
"ldr q15, [x9, #0x30]\n"
- "b 52f\n"
- "51:" // Height 2: no accumulate
+ "b 51f\n"
+ "50:" // Height 2: no accumulate
"movi v8.16b, #0x0\n"
"movi v9.16b, #0x0\n"
"movi v10.16b, #0x0\n"
@@ -526,31 +520,31 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"movi v13.16b, #0x0\n"
"movi v14.16b, #0x0\n"
"movi v15.16b, #0x0\n"
- "52:" // Height 2: setup done
+ "51:" // Height 2: setup done
"mov x12, #0x0\n"
- "53:" // Height 2: String loop
+ "52:" // Height 2: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
"ldr w11, [x20, x12, LSL #0x2]\n"
- "tbz %x[flags], #3, 54f\n"
+ "tbz %x[flags], #3, 53f\n"
"ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
"add x20, x20, x19, LSL #3\n"
"ldr x10, [x20, #0x0]\n"
"ldr x28, [x20, #0x8]\n"
- "cbnz x12, 55f\n"
+ "cbnz x12, 54f\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x10, x10, x19, LSL #1\n"
"add x28, x28, x19, LSL #1\n"
- "b 55f\n"
- "54:" // Height 2: setup direct input
+ "b 54f\n"
+ "53:" // Height 2: setup direct input
"mov x10, %x[input_ptr]\n"
"add x28, x10, x19, LSL #1\n"
- "55:" // Height 2: input setup done
+ "54:" // Height 2: input setup done
"cmp x11, #0x8\n"
- "blt 58f\n"
- "cmp x11, #0x10\n"
"blt 57f\n"
- "56:" // Height 2: Multiply loop: Main loop head
+ "cmp x11, #0x10\n"
+ "blt 56f\n"
+ "55:" // Height 2: Multiply loop: Main loop head
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
"ldr q6, [x15, #0x0]\n"
@@ -608,8 +602,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n"
".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n"
".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n"
- "bge 56b\n"
- "57:" // Height 2: Multiply loop: Single iteration only
+ "bge 55b\n"
+ "56:" // Height 2: Multiply loop: Single iteration only
"sub x11, x11, #0x8\n"
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
@@ -666,11 +660,11 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n"
".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n"
".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n"
- "58:" // Height 2: Multiply loop: Main loop skip
- "cbz x11, 63f\n"
+ "57:" // Height 2: Multiply loop: Main loop skip
+ "cbz x11, 61f\n"
"cmp x11, #0x2\n"
- "blt 60f\n"
- "59:" // Height 2: Multiply loop: Odd block loop
+ "blt 59f\n"
+ "58:" // Height 2: Multiply loop: Odd block loop
"ldr s0, [x10], #0x4\n"
"ldr s1, [x28], #0x4\n"
"ldr q6, [x15, #0x0]\n"
@@ -688,20 +682,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n"
".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n"
".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n"
- "bge 59b\n"
- "cbz x11, 63f\n"
- "60:" // Height 2: Multiply loop: Skip odd blocks
- "tbz x11, #1, 61f\n"
- "ldr s0, [x10], #0x4\n"
- "ldr s1, [x28], #0x4\n"
- "tbz x11, #0, 62f\n"
- "ld1 { v0.h }[2], [x10]\n"
- "ld1 { v1.h }[2], [x28]\n"
- "b 62f\n"
- "61:" // Height 2: Multiply loop: Ragged operand read: partial_1_0
+ "bge 58b\n"
+ "cbz x11, 61f\n"
+ "59:" // Height 2: Multiply loop: Skip odd blocks
"ldr h0, [x10, #0x0]\n"
"ldr h1, [x28, #0x0]\n"
- "62:" // Height 2: Multiply loop: Ragged operand read: Done
+ "60:" // Height 2: Multiply loop: Ragged operand read: Done
"ldr q6, [x15, #0x0]\n"
".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n"
"ldr q7, [x15, #0x10]\n"
@@ -715,14 +701,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n"
".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n"
".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n"
- "63:" // Height 2: Multiply loop: No odd multiplies
+ "61:" // Height 2: Multiply loop: No odd multiplies
"ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x12, x12, #0x1\n"
"cmp x12, x19\n"
- "bne 53b\n"
+ "bne 52b\n"
"prfm pstl1keep, [x13, #0x0]\n"
"prfm pstl1keep, [x9, #0x0]\n"
- "tbz %x[flags], #1, 64f\n"
+ "tbz %x[flags], #1, 62f\n"
"add x19, %x[args_ptr], %[offset_min]\n"
"ld1r { v1.4s }, [x19]\n"
"add x19, %x[args_ptr], %[offset_max]\n"
@@ -743,72 +729,72 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"fmax v14.4s, v14.4s, v1.4s\n"
"fmin v15.4s, v15.4s, v0.4s\n"
"fmax v15.4s, v15.4s, v1.4s\n"
- "64:" // Height 2: No activation
+ "62:" // Height 2: No activation
"cmp x16, #0x10\n"
- "bge 73f\n"
- "tbz x16, #3, 68f\n"
+ "bge 71f\n"
+ "tbz x16, #3, 66f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v9.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
"st1 { v13.4s }, [x9], #0x10\n"
- "tbz x16, #2, 66f\n"
+ "tbz x16, #2, 64f\n"
"st1 { v10.4s }, [x13], #0x10\n"
"st1 { v14.4s }, [x9], #0x10\n"
- "tbz x16, #1, 65f\n"
+ "tbz x16, #1, 63f\n"
"str d11, [x13], #0x8\n"
"str d15, [x9], #0x8\n"
- "tbz x16, #0, 72f\n"
+ "tbz x16, #0, 70f\n"
"st1 { v11.s }[2], [x13]\n"
"st1 { v15.s }[2], [x9]\n"
- "b 72f\n"
- "65:" // Height 2: Partial direct writeback: partial_1_12
- "tbz x16, #0, 72f\n"
+ "b 70f\n"
+ "63:" // Height 2: Partial direct writeback: partial_1_12
+ "tbz x16, #0, 70f\n"
"str s11, [x13, #0x0]\n"
"str s15, [x9, #0x0]\n"
- "b 72f\n"
- "66:" // Height 2: Partial direct writeback: partial_2_8
- "tbz x16, #1, 67f\n"
+ "b 70f\n"
+ "64:" // Height 2: Partial direct writeback: partial_2_8
+ "tbz x16, #1, 65f\n"
"str d10, [x13], #0x8\n"
"str d14, [x9], #0x8\n"
- "tbz x16, #0, 72f\n"
+ "tbz x16, #0, 70f\n"
"st1 { v10.s }[2], [x13]\n"
"st1 { v14.s }[2], [x9]\n"
- "b 72f\n"
- "67:" // Height 2: Partial direct writeback: partial_1_8
- "tbz x16, #0, 72f\n"
+ "b 70f\n"
+ "65:" // Height 2: Partial direct writeback: partial_1_8
+ "tbz x16, #0, 70f\n"
"str s10, [x13, #0x0]\n"
"str s14, [x9, #0x0]\n"
- "b 72f\n"
- "68:" // Height 2: Partial direct writeback: partial_4_0
- "tbz x16, #2, 70f\n"
+ "b 70f\n"
+ "66:" // Height 2: Partial direct writeback: partial_4_0
+ "tbz x16, #2, 68f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
- "tbz x16, #1, 69f\n"
+ "tbz x16, #1, 67f\n"
"str d9, [x13], #0x8\n"
"str d13, [x9], #0x8\n"
- "tbz x16, #0, 72f\n"
+ "tbz x16, #0, 70f\n"
"st1 { v9.s }[2], [x13]\n"
"st1 { v13.s }[2], [x9]\n"
- "b 72f\n"
- "69:" // Height 2: Partial direct writeback: partial_1_4
- "tbz x16, #0, 72f\n"
+ "b 70f\n"
+ "67:" // Height 2: Partial direct writeback: partial_1_4
+ "tbz x16, #0, 70f\n"
"str s9, [x13, #0x0]\n"
"str s13, [x9, #0x0]\n"
- "b 72f\n"
- "70:" // Height 2: Partial direct writeback: partial_2_0
- "tbz x16, #1, 71f\n"
+ "b 70f\n"
+ "68:" // Height 2: Partial direct writeback: partial_2_0
+ "tbz x16, #1, 69f\n"
"str d8, [x13], #0x8\n"
"str d12, [x9], #0x8\n"
- "tbz x16, #0, 72f\n"
+ "tbz x16, #0, 70f\n"
"st1 { v8.s }[2], [x13]\n"
"st1 { v12.s }[2], [x9]\n"
- "b 72f\n"
- "71:" // Height 2: Partial direct writeback: partial_1_0
+ "b 70f\n"
+ "69:" // Height 2: Partial direct writeback: partial_1_0
"str s8, [x13, #0x0]\n"
"str s12, [x9, #0x0]\n"
- "72:" // Height 2: Partial direct writeback: Done
- "b 74f\n"
- "73:" // Height 2: Full writeback
+ "70:" // Height 2: Partial direct writeback: Done
+ "b 72f\n"
+ "71:" // Height 2: Full writeback
"str q8, [x13, #0x0]\n"
"str q9, [x13, #0x10]\n"
"str q10, [x13, #0x20]\n"
@@ -819,29 +805,29 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"str q15, [x9, #0x30]\n"
"add x13, x13, #0x40\n"
"add x9, x9, #0x40\n"
- "74:" // Height 2: Writeback done
+ "72:" // Height 2: Writeback done
"subs x16, x16, #0x10\n"
- "bgt 40b\n"
- "b 224f\n"
- "75:" // Height 3
+ "bgt 39b\n"
+ "b 218f\n"
+ "73:" // Height 3
"ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
"mov x14, %x[bias]\n"
"ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "tbz %x[flags], #2, 76f\n"
+ "tbz %x[flags], #2, 74f\n"
"ldr x13, [%x[output_ptr], #0x0]\n"
"add x13, x13, x19, LSL #2\n"
"ldr x9, [%x[output_ptr], #0x8]\n"
"ldr x27, [%x[output_ptr], #0x10]\n"
"add x9, x9, x19, LSL #2\n"
"add x27, x27, x19, LSL #2\n"
- "b 77f\n"
- "76:" // Height 3: setup direct output
+ "b 75f\n"
+ "74:" // Height 3: setup direct output
"mov x13, %x[output_ptr]\n"
"add x9, x13, x19, LSL #2\n"
"add x27, x9, x19, LSL #2\n"
- "77:" // Height 3: Column loop
- "cbz x14, 78f\n"
+ "75:" // Height 3: Column loop
+ "cbz x14, 76f\n"
"ldr q8, [x14, #0x0]\n"
"mov v12.16b, v8.16b\n"
"ldr q9, [x14, #0x10]\n"
@@ -855,101 +841,101 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"mov v15.16b, v11.16b\n"
"mov v18.16b, v10.16b\n"
"mov v19.16b, v11.16b\n"
- "b 89f\n"
- "78:" // Height 3: no bias
- "tbz %x[flags], #0, 88f\n"
+ "b 87f\n"
+ "76:" // Height 3: no bias
+ "tbz %x[flags], #0, 86f\n"
"cmp x16, #0x10\n"
- "bge 87f\n"
- "tbz x16, #3, 82f\n"
+ "bge 85f\n"
+ "tbz x16, #3, 80f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
"ld1 { v9.4s }, [x13], #0x10\n"
"ld1 { v13.4s }, [x9], #0x10\n"
"ld1 { v17.4s }, [x27], #0x10\n"
- "tbz x16, #2, 80f\n"
+ "tbz x16, #2, 78f\n"
"ld1 { v10.4s }, [x13], #0x10\n"
"ld1 { v14.4s }, [x9], #0x10\n"
"ld1 { v18.4s }, [x27], #0x10\n"
- "tbz x16, #1, 79f\n"
+ "tbz x16, #1, 77f\n"
"mov x19, #0x38\n"
"ldr d11, [x13], #0x8\n"
"ldr d15, [x9], #0x8\n"
"ldr d19, [x27], #0x8\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ld1 { v11.s }[2], [x13]\n"
"ld1 { v15.s }[2], [x9]\n"
"ld1 { v19.s }[2], [x27]\n"
- "b 86f\n"
- "79:" // Height 3: Partial accumulate: partial_1_12
+ "b 84f\n"
+ "77:" // Height 3: Partial accumulate: partial_1_12
"mov x19, #0x30\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ldr s11, [x13, #0x0]\n"
"ldr s15, [x9, #0x0]\n"
"ldr s19, [x27, #0x0]\n"
- "b 86f\n"
- "80:" // Height 3: Partial accumulate: partial_2_8
- "tbz x16, #1, 81f\n"
+ "b 84f\n"
+ "78:" // Height 3: Partial accumulate: partial_2_8
+ "tbz x16, #1, 79f\n"
"ldr d10, [x13], #0x8\n"
"ldr d14, [x9], #0x8\n"
"ldr d18, [x27], #0x8\n"
"mov x19, #0x28\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ld1 { v10.s }[2], [x13]\n"
"ld1 { v14.s }[2], [x9]\n"
"ld1 { v18.s }[2], [x27]\n"
- "b 86f\n"
- "81:" // Height 3: Partial accumulate: partial_1_8
+ "b 84f\n"
+ "79:" // Height 3: Partial accumulate: partial_1_8
"mov x19, #0x20\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ldr s10, [x13, #0x0]\n"
"ldr s14, [x9, #0x0]\n"
"ldr s18, [x27, #0x0]\n"
- "b 86f\n"
- "82:" // Height 3: Partial accumulate: partial_4_0
- "tbz x16, #2, 84f\n"
+ "b 84f\n"
+ "80:" // Height 3: Partial accumulate: partial_4_0
+ "tbz x16, #2, 82f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
- "tbz x16, #1, 83f\n"
+ "tbz x16, #1, 81f\n"
"mov x19, #0x18\n"
"ldr d9, [x13], #0x8\n"
"ldr d13, [x9], #0x8\n"
"ldr d17, [x27], #0x8\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ld1 { v9.s }[2], [x13]\n"
"ld1 { v13.s }[2], [x9]\n"
"ld1 { v17.s }[2], [x27]\n"
- "b 86f\n"
- "83:" // Height 3: Partial accumulate: partial_1_4
+ "b 84f\n"
+ "81:" // Height 3: Partial accumulate: partial_1_4
"mov x19, #0x10\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ldr s9, [x13, #0x0]\n"
"ldr s13, [x9, #0x0]\n"
"ldr s17, [x27, #0x0]\n"
- "b 86f\n"
- "84:" // Height 3: Partial accumulate: partial_2_0
- "tbz x16, #1, 85f\n"
+ "b 84f\n"
+ "82:" // Height 3: Partial accumulate: partial_2_0
+ "tbz x16, #1, 83f\n"
"ldr d8, [x13], #0x8\n"
"ldr d12, [x9], #0x8\n"
"ldr d16, [x27], #0x8\n"
"mov x19, #0x8\n"
- "tbz x16, #0, 86f\n"
+ "tbz x16, #0, 84f\n"
"ld1 { v8.s }[2], [x13]\n"
"ld1 { v12.s }[2], [x9]\n"
"ld1 { v16.s }[2], [x27]\n"
- "b 86f\n"
- "85:" // Height 3: Partial accumulate: partial_1_0
+ "b 84f\n"
+ "83:" // Height 3: Partial accumulate: partial_1_0
"mov x19, #0x0\n"
"ldr s8, [x13, #0x0]\n"
"ldr s12, [x9, #0x0]\n"
"ldr s16, [x27, #0x0]\n"
- "86:" // Height 3: Partial accumulate: Done
+ "84:" // Height 3: Partial accumulate: Done
"sub x13, x13, x19\n"
"sub x9, x9, x19\n"
"sub x27, x27, x19\n"
- "b 89f\n"
- "87:" // Height 3: full accumulate
+ "b 87f\n"
+ "85:" // Height 3: full accumulate
"ldr q8, [x13, #0x0]\n"
"ldr q9, [x13, #0x10]\n"
"ldr q10, [x13, #0x20]\n"
@@ -962,8 +948,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr q17, [x27, #0x10]\n"
"ldr q18, [x27, #0x20]\n"
"ldr q19, [x27, #0x30]\n"
- "b 89f\n"
- "88:" // Height 3: no accumulate
+ "b 87f\n"
+ "86:" // Height 3: no accumulate
"movi v8.16b, #0x0\n"
"movi v9.16b, #0x0\n"
"movi v10.16b, #0x0\n"
@@ -976,34 +962,34 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"movi v17.16b, #0x0\n"
"movi v18.16b, #0x0\n"
"movi v19.16b, #0x0\n"
- "89:" // Height 3: setup done
+ "87:" // Height 3: setup done
"mov x12, #0x0\n"
- "90:" // Height 3: String loop
+ "88:" // Height 3: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
"ldr w11, [x20, x12, LSL #0x2]\n"
- "tbz %x[flags], #3, 91f\n"
+ "tbz %x[flags], #3, 89f\n"
"ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
"add x20, x20, x19, LSL #3\n"
"ldr x10, [x20, #0x0]\n"
"ldr x28, [x20, #0x8]\n"
"ldr x26, [x20, #0x10]\n"
- "cbnz x12, 92f\n"
+ "cbnz x12, 90f\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x10, x10, x19, LSL #1\n"
"add x28, x28, x19, LSL #1\n"
"add x26, x26, x19, LSL #1\n"
- "b 92f\n"
- "91:" // Height 3: setup direct input
+ "b 90f\n"
+ "89:" // Height 3: setup direct input
"mov x10, %x[input_ptr]\n"
"add x28, x10, x19, LSL #1\n"
"add x26, x28, x19, LSL #1\n"
- "92:" // Height 3: input setup done
+ "90:" // Height 3: input setup done
"cmp x11, #0x8\n"
- "blt 95f\n"
+ "blt 93f\n"
"cmp x11, #0x10\n"
- "blt 94f\n"
- "93:" // Height 3: Multiply loop: Main loop head
+ "blt 92f\n"
+ "91:" // Height 3: Multiply loop: Main loop head
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
"ldr q2, [x26, #0x0]\n"
@@ -1080,8 +1066,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n"
".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n"
".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n"
- "bge 93b\n"
- "94:" // Height 3: Multiply loop: Single iteration only
+ "bge 91b\n"
+ "92:" // Height 3: Multiply loop: Single iteration only
"sub x11, x11, #0x8\n"
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
@@ -1157,11 +1143,11 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n"
".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n"
".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n"
- "95:" // Height 3: Multiply loop: Main loop skip
- "cbz x11, 100f\n"
+ "93:" // Height 3: Multiply loop: Main loop skip
+ "cbz x11, 97f\n"
"cmp x11, #0x2\n"
- "blt 97f\n"
- "96:" // Height 3: Multiply loop: Odd block loop
+ "blt 95f\n"
+ "94:" // Height 3: Multiply loop: Odd block loop
"ldr s0, [x10], #0x4\n"
"ldr s1, [x28], #0x4\n"
"ldr s2, [x26], #0x4\n"
@@ -1184,23 +1170,13 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n"
".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n"
".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n"
- "bge 96b\n"
- "cbz x11, 100f\n"
- "97:" // Height 3: Multiply loop: Skip odd blocks
- "tbz x11, #1, 98f\n"
- "ldr s0, [x10], #0x4\n"
- "ldr s1, [x28], #0x4\n"
- "ldr s2, [x26], #0x4\n"
- "tbz x11, #0, 99f\n"
- "ld1 { v0.h }[2], [x10]\n"
- "ld1 { v1.h }[2], [x28]\n"
- "ld1 { v2.h }[2], [x26]\n"
- "b 99f\n"
- "98:" // Height 3: Multiply loop: Ragged operand read: partial_1_0
+ "bge 94b\n"
+ "cbz x11, 97f\n"
+ "95:" // Height 3: Multiply loop: Skip odd blocks
"ldr h0, [x10, #0x0]\n"
"ldr h1, [x28, #0x0]\n"
"ldr h2, [x26, #0x0]\n"
- "99:" // Height 3: Multiply loop: Ragged operand read: Done
+ "96:" // Height 3: Multiply loop: Ragged operand read: Done
"ldr q6, [x15, #0x0]\n"
".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n"
"ldr q7, [x15, #0x10]\n"
@@ -1218,15 +1194,15 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n"
".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n"
".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n"
- "100:" // Height 3: Multiply loop: No odd multiplies
+ "97:" // Height 3: Multiply loop: No odd multiplies
"ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x12, x12, #0x1\n"
"cmp x12, x19\n"
- "bne 90b\n"
+ "bne 88b\n"
"prfm pstl1keep, [x13, #0x0]\n"
"prfm pstl1keep, [x9, #0x0]\n"
"prfm pstl1keep, [x27, #0x0]\n"
- "tbz %x[flags], #1, 101f\n"
+ "tbz %x[flags], #1, 98f\n"
"add x19, %x[args_ptr], %[offset_min]\n"
"ld1r { v1.4s }, [x19]\n"
"add x19, %x[args_ptr], %[offset_max]\n"
@@ -1255,88 +1231,88 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"fmin v19.4s, v19.4s, v0.4s\n"
"fmax v18.4s, v18.4s, v1.4s\n"
"fmax v19.4s, v19.4s, v1.4s\n"
- "101:" // Height 3: No activation
+ "98:" // Height 3: No activation
"cmp x16, #0x10\n"
- "bge 110f\n"
- "tbz x16, #3, 105f\n"
+ "bge 107f\n"
+ "tbz x16, #3, 102f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v9.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
"st1 { v13.4s }, [x9], #0x10\n"
"st1 { v16.4s }, [x27], #0x10\n"
"st1 { v17.4s }, [x27], #0x10\n"
- "tbz x16, #2, 103f\n"
+ "tbz x16, #2, 100f\n"
"st1 { v10.4s }, [x13], #0x10\n"
"st1 { v14.4s }, [x9], #0x10\n"
"st1 { v18.4s }, [x27], #0x10\n"
- "tbz x16, #1, 102f\n"
+ "tbz x16, #1, 99f\n"
"str d11, [x13], #0x8\n"
"str d15, [x9], #0x8\n"
"str d19, [x27], #0x8\n"
- "tbz x16, #0, 109f\n"
+ "tbz x16, #0, 106f\n"
"st1 { v11.s }[2], [x13]\n"
"st1 { v15.s }[2], [x9]\n"
"st1 { v19.s }[2], [x27]\n"
- "b 109f\n"
- "102:" // Height 3: Partial direct writeback: partial_1_12
- "tbz x16, #0, 109f\n"
+ "b 106f\n"
+ "99:" // Height 3: Partial direct writeback: partial_1_12
+ "tbz x16, #0, 106f\n"
"str s11, [x13, #0x0]\n"
"str s15, [x9, #0x0]\n"
"str s19, [x27, #0x0]\n"
- "b 109f\n"
- "103:" // Height 3: Partial direct writeback: partial_2_8
- "tbz x16, #1, 104f\n"
+ "b 106f\n"
+ "100:" // Height 3: Partial direct writeback: partial_2_8
+ "tbz x16, #1, 101f\n"
"str d10, [x13], #0x8\n"
"str d14, [x9], #0x8\n"
"str d18, [x27], #0x8\n"
- "tbz x16, #0, 109f\n"
+ "tbz x16, #0, 106f\n"
"st1 { v10.s }[2], [x13]\n"
"st1 { v14.s }[2], [x9]\n"
"st1 { v18.s }[2], [x27]\n"
- "b 109f\n"
- "104:" // Height 3: Partial direct writeback: partial_1_8
- "tbz x16, #0, 109f\n"
+ "b 106f\n"
+ "101:" // Height 3: Partial direct writeback: partial_1_8
+ "tbz x16, #0, 106f\n"
"str s10, [x13, #0x0]\n"
"str s14, [x9, #0x0]\n"
"str s18, [x27, #0x0]\n"
- "b 109f\n"
- "105:" // Height 3: Partial direct writeback: partial_4_0
- "tbz x16, #2, 107f\n"
+ "b 106f\n"
+ "102:" // Height 3: Partial direct writeback: partial_4_0
+ "tbz x16, #2, 104f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
"st1 { v16.4s }, [x27], #0x10\n"
- "tbz x16, #1, 106f\n"
+ "tbz x16, #1, 103f\n"
"str d9, [x13], #0x8\n"
"str d13, [x9], #0x8\n"
"str d17, [x27], #0x8\n"
- "tbz x16, #0, 109f\n"
+ "tbz x16, #0, 106f\n"
"st1 { v9.s }[2], [x13]\n"
"st1 { v13.s }[2], [x9]\n"
"st1 { v17.s }[2], [x27]\n"
- "b 109f\n"
- "106:" // Height 3: Partial direct writeback: partial_1_4
- "tbz x16, #0, 109f\n"
+ "b 106f\n"
+ "103:" // Height 3: Partial direct writeback: partial_1_4
+ "tbz x16, #0, 106f\n"
"str s9, [x13, #0x0]\n"
"str s13, [x9, #0x0]\n"
"str s17, [x27, #0x0]\n"
- "b 109f\n"
- "107:" // Height 3: Partial direct writeback: partial_2_0
- "tbz x16, #1, 108f\n"
+ "b 106f\n"
+ "104:" // Height 3: Partial direct writeback: partial_2_0
+ "tbz x16, #1, 105f\n"
"str d8, [x13], #0x8\n"
"str d12, [x9], #0x8\n"
"str d16, [x27], #0x8\n"
- "tbz x16, #0, 109f\n"
+ "tbz x16, #0, 106f\n"
"st1 { v8.s }[2], [x13]\n"
"st1 { v12.s }[2], [x9]\n"
"st1 { v16.s }[2], [x27]\n"
- "b 109f\n"
- "108:" // Height 3: Partial direct writeback: partial_1_0
+ "b 106f\n"
+ "105:" // Height 3: Partial direct writeback: partial_1_0
"str s8, [x13, #0x0]\n"
"str s12, [x9, #0x0]\n"
"str s16, [x27, #0x0]\n"
- "109:" // Height 3: Partial direct writeback: Done
- "b 111f\n"
- "110:" // Height 3: Full writeback
+ "106:" // Height 3: Partial direct writeback: Done
+ "b 108f\n"
+ "107:" // Height 3: Full writeback
"str q8, [x13, #0x0]\n"
"str q9, [x13, #0x10]\n"
"str q10, [x13, #0x20]\n"
@@ -1352,16 +1328,16 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x13, x13, #0x40\n"
"add x9, x9, #0x40\n"
"add x27, x27, #0x40\n"
- "111:" // Height 3: Writeback done
+ "108:" // Height 3: Writeback done
"subs x16, x16, #0x10\n"
- "bgt 77b\n"
- "b 224f\n"
- "112:" // Height 4
+ "bgt 75b\n"
+ "b 218f\n"
+ "109:" // Height 4
"ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
"mov x14, %x[bias]\n"
"ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "tbz %x[flags], #2, 113f\n"
+ "tbz %x[flags], #2, 110f\n"
"ldr x13, [%x[output_ptr], #0x0]\n"
"add x13, x13, x19, LSL #2\n"
"ldr x9, [%x[output_ptr], #0x8]\n"
@@ -1370,14 +1346,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr x25, [%x[output_ptr], #0x18]\n"
"add x27, x27, x19, LSL #2\n"
"add x25, x25, x19, LSL #2\n"
- "b 114f\n"
- "113:" // Height 4: setup direct output
+ "b 111f\n"
+ "110:" // Height 4: setup direct output
"mov x13, %x[output_ptr]\n"
"add x9, x13, x19, LSL #2\n"
"add x27, x9, x19, LSL #2\n"
"add x25, x27, x19, LSL #2\n"
- "114:" // Height 4: Column loop
- "cbz x14, 115f\n"
+ "111:" // Height 4: Column loop
+ "cbz x14, 112f\n"
"ldr q8, [x14, #0x0]\n"
"mov v12.16b, v8.16b\n"
"ldr q9, [x14, #0x10]\n"
@@ -1395,12 +1371,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"mov v21.16b, v9.16b\n"
"mov v22.16b, v10.16b\n"
"mov v23.16b, v11.16b\n"
- "b 126f\n"
- "115:" // Height 4: no bias
- "tbz %x[flags], #0, 125f\n"
+ "b 123f\n"
+ "112:" // Height 4: no bias
+ "tbz %x[flags], #0, 122f\n"
"cmp x16, #0x10\n"
- "bge 124f\n"
- "tbz x16, #3, 119f\n"
+ "bge 121f\n"
+ "tbz x16, #3, 116f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
@@ -1409,104 +1385,104 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ld1 { v13.4s }, [x9], #0x10\n"
"ld1 { v17.4s }, [x27], #0x10\n"
"ld1 { v21.4s }, [x25], #0x10\n"
- "tbz x16, #2, 117f\n"
+ "tbz x16, #2, 114f\n"
"ld1 { v10.4s }, [x13], #0x10\n"
"ld1 { v14.4s }, [x9], #0x10\n"
"ld1 { v18.4s }, [x27], #0x10\n"
"ld1 { v22.4s }, [x25], #0x10\n"
- "tbz x16, #1, 116f\n"
+ "tbz x16, #1, 113f\n"
"mov x19, #0x38\n"
"ldr d11, [x13], #0x8\n"
"ldr d15, [x9], #0x8\n"
"ldr d19, [x27], #0x8\n"
"ldr d23, [x25], #0x8\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ld1 { v11.s }[2], [x13]\n"
"ld1 { v15.s }[2], [x9]\n"
"ld1 { v19.s }[2], [x27]\n"
"ld1 { v23.s }[2], [x25]\n"
- "b 123f\n"
- "116:" // Height 4: Partial accumulate: partial_1_12
+ "b 120f\n"
+ "113:" // Height 4: Partial accumulate: partial_1_12
"mov x19, #0x30\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ldr s11, [x13, #0x0]\n"
"ldr s15, [x9, #0x0]\n"
"ldr s19, [x27, #0x0]\n"
"ldr s23, [x25, #0x0]\n"
- "b 123f\n"
- "117:" // Height 4: Partial accumulate: partial_2_8
- "tbz x16, #1, 118f\n"
+ "b 120f\n"
+ "114:" // Height 4: Partial accumulate: partial_2_8
+ "tbz x16, #1, 115f\n"
"ldr d10, [x13], #0x8\n"
"ldr d14, [x9], #0x8\n"
"ldr d18, [x27], #0x8\n"
"ldr d22, [x25], #0x8\n"
"mov x19, #0x28\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ld1 { v10.s }[2], [x13]\n"
"ld1 { v14.s }[2], [x9]\n"
"ld1 { v18.s }[2], [x27]\n"
"ld1 { v22.s }[2], [x25]\n"
- "b 123f\n"
- "118:" // Height 4: Partial accumulate: partial_1_8
+ "b 120f\n"
+ "115:" // Height 4: Partial accumulate: partial_1_8
"mov x19, #0x20\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ldr s10, [x13, #0x0]\n"
"ldr s14, [x9, #0x0]\n"
"ldr s18, [x27, #0x0]\n"
"ldr s22, [x25, #0x0]\n"
- "b 123f\n"
- "119:" // Height 4: Partial accumulate: partial_4_0
- "tbz x16, #2, 121f\n"
+ "b 120f\n"
+ "116:" // Height 4: Partial accumulate: partial_4_0
+ "tbz x16, #2, 118f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
"ld1 { v20.4s }, [x25], #0x10\n"
- "tbz x16, #1, 120f\n"
+ "tbz x16, #1, 117f\n"
"mov x19, #0x18\n"
"ldr d9, [x13], #0x8\n"
"ldr d13, [x9], #0x8\n"
"ldr d17, [x27], #0x8\n"
"ldr d21, [x25], #0x8\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ld1 { v9.s }[2], [x13]\n"
"ld1 { v13.s }[2], [x9]\n"
"ld1 { v17.s }[2], [x27]\n"
"ld1 { v21.s }[2], [x25]\n"
- "b 123f\n"
- "120:" // Height 4: Partial accumulate: partial_1_4
+ "b 120f\n"
+ "117:" // Height 4: Partial accumulate: partial_1_4
"mov x19, #0x10\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ldr s9, [x13, #0x0]\n"
"ldr s13, [x9, #0x0]\n"
"ldr s17, [x27, #0x0]\n"
"ldr s21, [x25, #0x0]\n"
- "b 123f\n"
- "121:" // Height 4: Partial accumulate: partial_2_0
- "tbz x16, #1, 122f\n"
+ "b 120f\n"
+ "118:" // Height 4: Partial accumulate: partial_2_0
+ "tbz x16, #1, 119f\n"
"ldr d8, [x13], #0x8\n"
"ldr d12, [x9], #0x8\n"
"ldr d16, [x27], #0x8\n"
"ldr d20, [x25], #0x8\n"
"mov x19, #0x8\n"
- "tbz x16, #0, 123f\n"
+ "tbz x16, #0, 120f\n"
"ld1 { v8.s }[2], [x13]\n"
"ld1 { v12.s }[2], [x9]\n"
"ld1 { v16.s }[2], [x27]\n"
"ld1 { v20.s }[2], [x25]\n"
- "b 123f\n"
- "122:" // Height 4: Partial accumulate: partial_1_0
+ "b 120f\n"
+ "119:" // Height 4: Partial accumulate: partial_1_0
"mov x19, #0x0\n"
"ldr s8, [x13, #0x0]\n"
"ldr s12, [x9, #0x0]\n"
"ldr s16, [x27, #0x0]\n"
"ldr s20, [x25, #0x0]\n"
- "123:" // Height 4: Partial accumulate: Done
+ "120:" // Height 4: Partial accumulate: Done
"sub x13, x13, x19\n"
"sub x9, x9, x19\n"
"sub x27, x27, x19\n"
"sub x25, x25, x19\n"
- "b 126f\n"
- "124:" // Height 4: full accumulate
+ "b 123f\n"
+ "121:" // Height 4: full accumulate
"ldr q8, [x13, #0x0]\n"
"ldr q9, [x13, #0x10]\n"
"ldr q10, [x13, #0x20]\n"
@@ -1523,8 +1499,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr q21, [x25, #0x10]\n"
"ldr q22, [x25, #0x20]\n"
"ldr q23, [x25, #0x30]\n"
- "b 126f\n"
- "125:" // Height 4: no accumulate
+ "b 123f\n"
+ "122:" // Height 4: no accumulate
"movi v8.16b, #0x0\n"
"movi v9.16b, #0x0\n"
"movi v10.16b, #0x0\n"
@@ -1541,37 +1517,37 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"movi v21.16b, #0x0\n"
"movi v22.16b, #0x0\n"
"movi v23.16b, #0x0\n"
- "126:" // Height 4: setup done
+ "123:" // Height 4: setup done
"mov x12, #0x0\n"
- "127:" // Height 4: String loop
+ "124:" // Height 4: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
"ldr w11, [x20, x12, LSL #0x2]\n"
- "tbz %x[flags], #3, 128f\n"
+ "tbz %x[flags], #3, 125f\n"
"ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
"add x20, x20, x19, LSL #3\n"
"ldr x10, [x20, #0x0]\n"
"ldr x28, [x20, #0x8]\n"
"ldr x26, [x20, #0x10]\n"
"ldr x24, [x20, #0x18]\n"
- "cbnz x12, 129f\n"
+ "cbnz x12, 126f\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x10, x10, x19, LSL #1\n"
"add x28, x28, x19, LSL #1\n"
"add x26, x26, x19, LSL #1\n"
"add x24, x24, x19, LSL #1\n"
- "b 129f\n"
- "128:" // Height 4: setup direct input
+ "b 126f\n"
+ "125:" // Height 4: setup direct input
"mov x10, %x[input_ptr]\n"
"add x28, x10, x19, LSL #1\n"
"add x26, x28, x19, LSL #1\n"
"add x24, x26, x19, LSL #1\n"
- "129:" // Height 4: input setup done
+ "126:" // Height 4: input setup done
"cmp x11, #0x8\n"
- "blt 132f\n"
+ "blt 129f\n"
"cmp x11, #0x10\n"
- "blt 131f\n"
- "130:" // Height 4: Multiply loop: Main loop head
+ "blt 128f\n"
+ "127:" // Height 4: Multiply loop: Main loop head
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
"ldr q2, [x26, #0x0]\n"
@@ -1667,8 +1643,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n"
".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n"
".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n"
- "bge 130b\n"
- "131:" // Height 4: Multiply loop: Single iteration only
+ "bge 127b\n"
+ "128:" // Height 4: Multiply loop: Single iteration only
"sub x11, x11, #0x8\n"
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
@@ -1763,11 +1739,11 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n"
".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n"
".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n"
- "132:" // Height 4: Multiply loop: Main loop skip
- "cbz x11, 137f\n"
+ "129:" // Height 4: Multiply loop: Main loop skip
+ "cbz x11, 133f\n"
"cmp x11, #0x2\n"
- "blt 134f\n"
- "133:" // Height 4: Multiply loop: Odd block loop
+ "blt 131f\n"
+ "130:" // Height 4: Multiply loop: Odd block loop
"ldr s0, [x10], #0x4\n"
"ldr s1, [x28], #0x4\n"
"ldr s2, [x26], #0x4\n"
@@ -1795,26 +1771,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n"
".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n"
".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n"
- "bge 133b\n"
- "cbz x11, 137f\n"
- "134:" // Height 4: Multiply loop: Skip odd blocks
- "tbz x11, #1, 135f\n"
- "ldr s0, [x10], #0x4\n"
- "ldr s1, [x28], #0x4\n"
- "ldr s2, [x26], #0x4\n"
- "ldr s3, [x24], #0x4\n"
- "tbz x11, #0, 136f\n"
- "ld1 { v0.h }[2], [x10]\n"
- "ld1 { v1.h }[2], [x28]\n"
- "ld1 { v2.h }[2], [x26]\n"
- "ld1 { v3.h }[2], [x24]\n"
- "b 136f\n"
- "135:" // Height 4: Multiply loop: Ragged operand read: partial_1_0
+ "bge 130b\n"
+ "cbz x11, 133f\n"
+ "131:" // Height 4: Multiply loop: Skip odd blocks
"ldr h0, [x10, #0x0]\n"
"ldr h1, [x28, #0x0]\n"
"ldr h2, [x26, #0x0]\n"
"ldr h3, [x24, #0x0]\n"
- "136:" // Height 4: Multiply loop: Ragged operand read: Done
+ "132:" // Height 4: Multiply loop: Ragged operand read: Done
"ldr q6, [x15, #0x0]\n"
".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n"
"ldr q7, [x15, #0x10]\n"
@@ -1836,16 +1800,16 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n"
".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n"
".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n"
- "137:" // Height 4: Multiply loop: No odd multiplies
+ "133:" // Height 4: Multiply loop: No odd multiplies
"ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x12, x12, #0x1\n"
"cmp x12, x19\n"
- "bne 127b\n"
+ "bne 124b\n"
"prfm pstl1keep, [x13, #0x0]\n"
"prfm pstl1keep, [x9, #0x0]\n"
"prfm pstl1keep, [x27, #0x0]\n"
"prfm pstl1keep, [x25, #0x0]\n"
- "tbz %x[flags], #1, 138f\n"
+ "tbz %x[flags], #1, 134f\n"
"add x19, %x[args_ptr], %[offset_min]\n"
"ld1r { v1.4s }, [x19]\n"
"add x19, %x[args_ptr], %[offset_max]\n"
@@ -1882,10 +1846,10 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"fmax v21.4s, v21.4s, v1.4s\n"
"fmax v22.4s, v22.4s, v1.4s\n"
"fmax v23.4s, v23.4s, v1.4s\n"
- "138:" // Height 4: No activation
+ "134:" // Height 4: No activation
"cmp x16, #0x10\n"
- "bge 147f\n"
- "tbz x16, #3, 142f\n"
+ "bge 143f\n"
+ "tbz x16, #3, 138f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v9.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
@@ -1894,92 +1858,92 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"st1 { v17.4s }, [x27], #0x10\n"
"st1 { v20.4s }, [x25], #0x10\n"
"st1 { v21.4s }, [x25], #0x10\n"
- "tbz x16, #2, 140f\n"
+ "tbz x16, #2, 136f\n"
"st1 { v10.4s }, [x13], #0x10\n"
"st1 { v14.4s }, [x9], #0x10\n"
"st1 { v18.4s }, [x27], #0x10\n"
"st1 { v22.4s }, [x25], #0x10\n"
- "tbz x16, #1, 139f\n"
+ "tbz x16, #1, 135f\n"
"str d11, [x13], #0x8\n"
"str d15, [x9], #0x8\n"
"str d19, [x27], #0x8\n"
"str d23, [x25], #0x8\n"
- "tbz x16, #0, 146f\n"
+ "tbz x16, #0, 142f\n"
"st1 { v11.s }[2], [x13]\n"
"st1 { v15.s }[2], [x9]\n"
"st1 { v19.s }[2], [x27]\n"
"st1 { v23.s }[2], [x25]\n"
- "b 146f\n"
- "139:" // Height 4: Partial direct writeback: partial_1_12
- "tbz x16, #0, 146f\n"
+ "b 142f\n"
+ "135:" // Height 4: Partial direct writeback: partial_1_12
+ "tbz x16, #0, 142f\n"
"str s11, [x13, #0x0]\n"
"str s15, [x9, #0x0]\n"
"str s19, [x27, #0x0]\n"
"str s23, [x25, #0x0]\n"
- "b 146f\n"
- "140:" // Height 4: Partial direct writeback: partial_2_8
- "tbz x16, #1, 141f\n"
+ "b 142f\n"
+ "136:" // Height 4: Partial direct writeback: partial_2_8
+ "tbz x16, #1, 137f\n"
"str d10, [x13], #0x8\n"
"str d14, [x9], #0x8\n"
"str d18, [x27], #0x8\n"
"str d22, [x25], #0x8\n"
- "tbz x16, #0, 146f\n"
+ "tbz x16, #0, 142f\n"
"st1 { v10.s }[2], [x13]\n"
"st1 { v14.s }[2], [x9]\n"
"st1 { v18.s }[2], [x27]\n"
"st1 { v22.s }[2], [x25]\n"
- "b 146f\n"
- "141:" // Height 4: Partial direct writeback: partial_1_8
- "tbz x16, #0, 146f\n"
+ "b 142f\n"
+ "137:" // Height 4: Partial direct writeback: partial_1_8
+ "tbz x16, #0, 142f\n"
"str s10, [x13, #0x0]\n"
"str s14, [x9, #0x0]\n"
"str s18, [x27, #0x0]\n"
"str s22, [x25, #0x0]\n"
- "b 146f\n"
- "142:" // Height 4: Partial direct writeback: partial_4_0
- "tbz x16, #2, 144f\n"
+ "b 142f\n"
+ "138:" // Height 4: Partial direct writeback: partial_4_0
+ "tbz x16, #2, 140f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
"st1 { v16.4s }, [x27], #0x10\n"
"st1 { v20.4s }, [x25], #0x10\n"
- "tbz x16, #1, 143f\n"
+ "tbz x16, #1, 139f\n"
"str d9, [x13], #0x8\n"
"str d13, [x9], #0x8\n"
"str d17, [x27], #0x8\n"
"str d21, [x25], #0x8\n"
- "tbz x16, #0, 146f\n"
+ "tbz x16, #0, 142f\n"
"st1 { v9.s }[2], [x13]\n"
"st1 { v13.s }[2], [x9]\n"
"st1 { v17.s }[2], [x27]\n"
"st1 { v21.s }[2], [x25]\n"
- "b 146f\n"
- "143:" // Height 4: Partial direct writeback: partial_1_4
- "tbz x16, #0, 146f\n"
+ "b 142f\n"
+ "139:" // Height 4: Partial direct writeback: partial_1_4
+ "tbz x16, #0, 142f\n"
"str s9, [x13, #0x0]\n"
"str s13, [x9, #0x0]\n"
"str s17, [x27, #0x0]\n"
"str s21, [x25, #0x0]\n"
- "b 146f\n"
- "144:" // Height 4: Partial direct writeback: partial_2_0
- "tbz x16, #1, 145f\n"
+ "b 142f\n"
+ "140:" // Height 4: Partial direct writeback: partial_2_0
+ "tbz x16, #1, 141f\n"
"str d8, [x13], #0x8\n"
"str d12, [x9], #0x8\n"
"str d16, [x27], #0x8\n"
"str d20, [x25], #0x8\n"
- "tbz x16, #0, 146f\n"
+ "tbz x16, #0, 142f\n"
"st1 { v8.s }[2], [x13]\n"
"st1 { v12.s }[2], [x9]\n"
"st1 { v16.s }[2], [x27]\n"
"st1 { v20.s }[2], [x25]\n"
- "b 146f\n"
- "145:" // Height 4: Partial direct writeback: partial_1_0
+ "b 142f\n"
+ "141:" // Height 4: Partial direct writeback: partial_1_0
"str s8, [x13, #0x0]\n"
"str s12, [x9, #0x0]\n"
"str s16, [x27, #0x0]\n"
"str s20, [x25, #0x0]\n"
- "146:" // Height 4: Partial direct writeback: Done
- "b 148f\n"
- "147:" // Height 4: Full writeback
+ "142:" // Height 4: Partial direct writeback: Done
+ "b 144f\n"
+ "143:" // Height 4: Full writeback
"str q8, [x13, #0x0]\n"
"str q9, [x13, #0x10]\n"
"str q10, [x13, #0x20]\n"
@@ -2000,16 +1964,16 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x9, x9, #0x40\n"
"add x27, x27, #0x40\n"
"add x25, x25, #0x40\n"
- "148:" // Height 4: Writeback done
+ "144:" // Height 4: Writeback done
"subs x16, x16, #0x10\n"
- "bgt 114b\n"
- "b 224f\n"
- "149:" // Height 5
+ "bgt 111b\n"
+ "b 218f\n"
+ "145:" // Height 5
"ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
"mov x14, %x[bias]\n"
"ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "tbz %x[flags], #2, 150f\n"
+ "tbz %x[flags], #2, 146f\n"
"ldr x13, [%x[output_ptr], #0x0]\n"
"add x13, x13, x19, LSL #2\n"
"ldr x9, [%x[output_ptr], #0x8]\n"
@@ -2020,15 +1984,15 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x27, x27, x19, LSL #2\n"
"add x25, x25, x19, LSL #2\n"
"add x23, x23, x19, LSL #2\n"
- "b 151f\n"
- "150:" // Height 5: setup direct output
+ "b 147f\n"
+ "146:" // Height 5: setup direct output
"mov x13, %x[output_ptr]\n"
"add x9, x13, x19, LSL #2\n"
"add x27, x9, x19, LSL #2\n"
"add x25, x27, x19, LSL #2\n"
"add x23, x25, x19, LSL #2\n"
- "151:" // Height 5: Column loop
- "cbz x14, 152f\n"
+ "147:" // Height 5: Column loop
+ "cbz x14, 148f\n"
"ldr q8, [x14, #0x0]\n"
"mov v12.16b, v8.16b\n"
"ldr q9, [x14, #0x10]\n"
@@ -2050,12 +2014,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"mov v25.16b, v9.16b\n"
"mov v26.16b, v10.16b\n"
"mov v27.16b, v11.16b\n"
- "b 163f\n"
- "152:" // Height 5: no bias
- "tbz %x[flags], #0, 162f\n"
+ "b 159f\n"
+ "148:" // Height 5: no bias
+ "tbz %x[flags], #0, 158f\n"
"cmp x16, #0x10\n"
- "bge 161f\n"
- "tbz x16, #3, 156f\n"
+ "bge 157f\n"
+ "tbz x16, #3, 152f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
@@ -2066,119 +2030,119 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ld1 { v17.4s }, [x27], #0x10\n"
"ld1 { v21.4s }, [x25], #0x10\n"
"ld1 { v25.4s }, [x23], #0x10\n"
- "tbz x16, #2, 154f\n"
+ "tbz x16, #2, 150f\n"
"ld1 { v10.4s }, [x13], #0x10\n"
"ld1 { v14.4s }, [x9], #0x10\n"
"ld1 { v18.4s }, [x27], #0x10\n"
"ld1 { v22.4s }, [x25], #0x10\n"
"ld1 { v26.4s }, [x23], #0x10\n"
- "tbz x16, #1, 153f\n"
+ "tbz x16, #1, 149f\n"
"mov x19, #0x38\n"
"ldr d11, [x13], #0x8\n"
"ldr d15, [x9], #0x8\n"
"ldr d19, [x27], #0x8\n"
"ldr d23, [x25], #0x8\n"
"ldr d27, [x23], #0x8\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ld1 { v11.s }[2], [x13]\n"
"ld1 { v15.s }[2], [x9]\n"
"ld1 { v19.s }[2], [x27]\n"
"ld1 { v23.s }[2], [x25]\n"
"ld1 { v27.s }[2], [x23]\n"
- "b 160f\n"
- "153:" // Height 5: Partial accumulate: partial_1_12
+ "b 156f\n"
+ "149:" // Height 5: Partial accumulate: partial_1_12
"mov x19, #0x30\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ldr s11, [x13, #0x0]\n"
"ldr s15, [x9, #0x0]\n"
"ldr s19, [x27, #0x0]\n"
"ldr s23, [x25, #0x0]\n"
"ldr s27, [x23, #0x0]\n"
- "b 160f\n"
- "154:" // Height 5: Partial accumulate: partial_2_8
- "tbz x16, #1, 155f\n"
+ "b 156f\n"
+ "150:" // Height 5: Partial accumulate: partial_2_8
+ "tbz x16, #1, 151f\n"
"ldr d10, [x13], #0x8\n"
"ldr d14, [x9], #0x8\n"
"ldr d18, [x27], #0x8\n"
"ldr d22, [x25], #0x8\n"
"ldr d26, [x23], #0x8\n"
"mov x19, #0x28\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ld1 { v10.s }[2], [x13]\n"
"ld1 { v14.s }[2], [x9]\n"
"ld1 { v18.s }[2], [x27]\n"
"ld1 { v22.s }[2], [x25]\n"
"ld1 { v26.s }[2], [x23]\n"
- "b 160f\n"
- "155:" // Height 5: Partial accumulate: partial_1_8
+ "b 156f\n"
+ "151:" // Height 5: Partial accumulate: partial_1_8
"mov x19, #0x20\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ldr s10, [x13, #0x0]\n"
"ldr s14, [x9, #0x0]\n"
"ldr s18, [x27, #0x0]\n"
"ldr s22, [x25, #0x0]\n"
"ldr s26, [x23, #0x0]\n"
- "b 160f\n"
- "156:" // Height 5: Partial accumulate: partial_4_0
- "tbz x16, #2, 158f\n"
+ "b 156f\n"
+ "152:" // Height 5: Partial accumulate: partial_4_0
+ "tbz x16, #2, 154f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
"ld1 { v20.4s }, [x25], #0x10\n"
"ld1 { v24.4s }, [x23], #0x10\n"
- "tbz x16, #1, 157f\n"
+ "tbz x16, #1, 153f\n"
"mov x19, #0x18\n"
"ldr d9, [x13], #0x8\n"
"ldr d13, [x9], #0x8\n"
"ldr d17, [x27], #0x8\n"
"ldr d21, [x25], #0x8\n"
"ldr d25, [x23], #0x8\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ld1 { v9.s }[2], [x13]\n"
"ld1 { v13.s }[2], [x9]\n"
"ld1 { v17.s }[2], [x27]\n"
"ld1 { v21.s }[2], [x25]\n"
"ld1 { v25.s }[2], [x23]\n"
- "b 160f\n"
- "157:" // Height 5: Partial accumulate: partial_1_4
+ "b 156f\n"
+ "153:" // Height 5: Partial accumulate: partial_1_4
"mov x19, #0x10\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ldr s9, [x13, #0x0]\n"
"ldr s13, [x9, #0x0]\n"
"ldr s17, [x27, #0x0]\n"
"ldr s21, [x25, #0x0]\n"
"ldr s25, [x23, #0x0]\n"
- "b 160f\n"
- "158:" // Height 5: Partial accumulate: partial_2_0
- "tbz x16, #1, 159f\n"
+ "b 156f\n"
+ "154:" // Height 5: Partial accumulate: partial_2_0
+ "tbz x16, #1, 155f\n"
"ldr d8, [x13], #0x8\n"
"ldr d12, [x9], #0x8\n"
"ldr d16, [x27], #0x8\n"
"ldr d20, [x25], #0x8\n"
"ldr d24, [x23], #0x8\n"
"mov x19, #0x8\n"
- "tbz x16, #0, 160f\n"
+ "tbz x16, #0, 156f\n"
"ld1 { v8.s }[2], [x13]\n"
"ld1 { v12.s }[2], [x9]\n"
"ld1 { v16.s }[2], [x27]\n"
"ld1 { v20.s }[2], [x25]\n"
"ld1 { v24.s }[2], [x23]\n"
- "b 160f\n"
- "159:" // Height 5: Partial accumulate: partial_1_0
+ "b 156f\n"
+ "155:" // Height 5: Partial accumulate: partial_1_0
"mov x19, #0x0\n"
"ldr s8, [x13, #0x0]\n"
"ldr s12, [x9, #0x0]\n"
"ldr s16, [x27, #0x0]\n"
"ldr s20, [x25, #0x0]\n"
"ldr s24, [x23, #0x0]\n"
- "160:" // Height 5: Partial accumulate: Done
+ "156:" // Height 5: Partial accumulate: Done
"sub x13, x13, x19\n"
"sub x9, x9, x19\n"
"sub x27, x27, x19\n"
"sub x25, x25, x19\n"
"sub x23, x23, x19\n"
- "b 163f\n"
- "161:" // Height 5: full accumulate
+ "b 159f\n"
+ "157:" // Height 5: full accumulate
"ldr q8, [x13, #0x0]\n"
"ldr q9, [x13, #0x10]\n"
"ldr q10, [x13, #0x20]\n"
@@ -2199,8 +2163,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr q25, [x23, #0x10]\n"
"ldr q26, [x23, #0x20]\n"
"ldr q27, [x23, #0x30]\n"
- "b 163f\n"
- "162:" // Height 5: no accumulate
+ "b 159f\n"
+ "158:" // Height 5: no accumulate
"movi v8.16b, #0x0\n"
"movi v9.16b, #0x0\n"
"movi v10.16b, #0x0\n"
@@ -2221,13 +2185,13 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"movi v25.16b, #0x0\n"
"movi v26.16b, #0x0\n"
"movi v27.16b, #0x0\n"
- "163:" // Height 5: setup done
+ "159:" // Height 5: setup done
"mov x12, #0x0\n"
- "164:" // Height 5: String loop
+ "160:" // Height 5: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
"ldr w11, [x20, x12, LSL #0x2]\n"
- "tbz %x[flags], #3, 165f\n"
+ "tbz %x[flags], #3, 161f\n"
"ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
"add x20, x20, x19, LSL #3\n"
"ldr x10, [x20, #0x0]\n"
@@ -2235,26 +2199,26 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr x26, [x20, #0x10]\n"
"ldr x24, [x20, #0x18]\n"
"ldr x22, [x20, #0x20]\n"
- "cbnz x12, 166f\n"
+ "cbnz x12, 162f\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x10, x10, x19, LSL #1\n"
"add x28, x28, x19, LSL #1\n"
"add x26, x26, x19, LSL #1\n"
"add x24, x24, x19, LSL #1\n"
"add x22, x22, x19, LSL #1\n"
- "b 166f\n"
- "165:" // Height 5: setup direct input
+ "b 162f\n"
+ "161:" // Height 5: setup direct input
"mov x10, %x[input_ptr]\n"
"add x28, x10, x19, LSL #1\n"
"add x26, x28, x19, LSL #1\n"
"add x24, x26, x19, LSL #1\n"
"add x22, x24, x19, LSL #1\n"
- "166:" // Height 5: input setup done
+ "162:" // Height 5: input setup done
"cmp x11, #0x8\n"
- "blt 169f\n"
+ "blt 165f\n"
"cmp x11, #0x10\n"
- "blt 168f\n"
- "167:" // Height 5: Multiply loop: Main loop head
+ "blt 164f\n"
+ "163:" // Height 5: Multiply loop: Main loop head
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
"ldr q2, [x26, #0x0]\n"
@@ -2369,8 +2333,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n"
".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n"
".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n"
- "bge 167b\n"
- "168:" // Height 5: Multiply loop: Single iteration only
+ "bge 163b\n"
+ "164:" // Height 5: Multiply loop: Single iteration only
"sub x11, x11, #0x8\n"
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
@@ -2484,11 +2448,11 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n"
".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n"
".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n"
- "169:" // Height 5: Multiply loop: Main loop skip
- "cbz x11, 174f\n"
+ "165:" // Height 5: Multiply loop: Main loop skip
+ "cbz x11, 169f\n"
"cmp x11, #0x2\n"
- "blt 171f\n"
- "170:" // Height 5: Multiply loop: Odd block loop
+ "blt 167f\n"
+ "166:" // Height 5: Multiply loop: Odd block loop
"ldr s0, [x10], #0x4\n"
"ldr s1, [x28], #0x4\n"
"ldr s2, [x26], #0x4\n"
@@ -2521,29 +2485,15 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n"
".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n"
".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n"
- "bge 170b\n"
- "cbz x11, 174f\n"
- "171:" // Height 5: Multiply loop: Skip odd blocks
- "tbz x11, #1, 172f\n"
- "ldr s0, [x10], #0x4\n"
- "ldr s1, [x28], #0x4\n"
- "ldr s2, [x26], #0x4\n"
- "ldr s3, [x24], #0x4\n"
- "ldr s4, [x22], #0x4\n"
- "tbz x11, #0, 173f\n"
- "ld1 { v0.h }[2], [x10]\n"
- "ld1 { v1.h }[2], [x28]\n"
- "ld1 { v2.h }[2], [x26]\n"
- "ld1 { v3.h }[2], [x24]\n"
- "ld1 { v4.h }[2], [x22]\n"
- "b 173f\n"
- "172:" // Height 5: Multiply loop: Ragged operand read: partial_1_0
+ "bge 166b\n"
+ "cbz x11, 169f\n"
+ "167:" // Height 5: Multiply loop: Skip odd blocks
"ldr h0, [x10, #0x0]\n"
"ldr h1, [x28, #0x0]\n"
"ldr h2, [x26, #0x0]\n"
"ldr h3, [x24, #0x0]\n"
"ldr h4, [x22, #0x0]\n"
- "173:" // Height 5: Multiply loop: Ragged operand read: Done
+ "168:" // Height 5: Multiply loop: Ragged operand read: Done
"ldr q6, [x15, #0x0]\n"
".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n"
"ldr q7, [x15, #0x10]\n"
@@ -2569,17 +2519,17 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n"
".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n"
".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n"
- "174:" // Height 5: Multiply loop: No odd multiplies
+ "169:" // Height 5: Multiply loop: No odd multiplies
"ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x12, x12, #0x1\n"
"cmp x12, x19\n"
- "bne 164b\n"
+ "bne 160b\n"
"prfm pstl1keep, [x13, #0x0]\n"
"prfm pstl1keep, [x9, #0x0]\n"
"prfm pstl1keep, [x27, #0x0]\n"
"prfm pstl1keep, [x25, #0x0]\n"
"prfm pstl1keep, [x23, #0x0]\n"
- "tbz %x[flags], #1, 175f\n"
+ "tbz %x[flags], #1, 170f\n"
"add x19, %x[args_ptr], %[offset_min]\n"
"ld1r { v1.4s }, [x19]\n"
"add x19, %x[args_ptr], %[offset_max]\n"
@@ -2624,10 +2574,10 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"fmax v26.4s, v26.4s, v1.4s\n"
"fmin v27.4s, v27.4s, v0.4s\n"
"fmax v27.4s, v27.4s, v1.4s\n"
- "175:" // Height 5: No activation
+ "170:" // Height 5: No activation
"cmp x16, #0x10\n"
- "bge 184f\n"
- "tbz x16, #3, 179f\n"
+ "bge 179f\n"
+ "tbz x16, #3, 174f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v9.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
@@ -2638,106 +2588,106 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"st1 { v21.4s }, [x25], #0x10\n"
"st1 { v24.4s }, [x23], #0x10\n"
"st1 { v25.4s }, [x23], #0x10\n"
- "tbz x16, #2, 177f\n"
+ "tbz x16, #2, 172f\n"
"st1 { v10.4s }, [x13], #0x10\n"
"st1 { v14.4s }, [x9], #0x10\n"
"st1 { v18.4s }, [x27], #0x10\n"
"st1 { v22.4s }, [x25], #0x10\n"
"st1 { v26.4s }, [x23], #0x10\n"
- "tbz x16, #1, 176f\n"
+ "tbz x16, #1, 171f\n"
"str d11, [x13], #0x8\n"
"str d15, [x9], #0x8\n"
"str d19, [x27], #0x8\n"
"str d23, [x25], #0x8\n"
"str d27, [x23], #0x8\n"
- "tbz x16, #0, 183f\n"
+ "tbz x16, #0, 178f\n"
"st1 { v11.s }[2], [x13]\n"
"st1 { v15.s }[2], [x9]\n"
"st1 { v19.s }[2], [x27]\n"
"st1 { v23.s }[2], [x25]\n"
"st1 { v27.s }[2], [x23]\n"
- "b 183f\n"
- "176:" // Height 5: Partial direct writeback: partial_1_12
- "tbz x16, #0, 183f\n"
+ "b 178f\n"
+ "171:" // Height 5: Partial direct writeback: partial_1_12
+ "tbz x16, #0, 178f\n"
"str s11, [x13, #0x0]\n"
"str s15, [x9, #0x0]\n"
"str s19, [x27, #0x0]\n"
"str s23, [x25, #0x0]\n"
"str s27, [x23, #0x0]\n"
- "b 183f\n"
- "177:" // Height 5: Partial direct writeback: partial_2_8
- "tbz x16, #1, 178f\n"
+ "b 178f\n"
+ "172:" // Height 5: Partial direct writeback: partial_2_8
+ "tbz x16, #1, 173f\n"
"str d10, [x13], #0x8\n"
"str d14, [x9], #0x8\n"
"str d18, [x27], #0x8\n"
"str d22, [x25], #0x8\n"
"str d26, [x23], #0x8\n"
- "tbz x16, #0, 183f\n"
+ "tbz x16, #0, 178f\n"
"st1 { v10.s }[2], [x13]\n"
"st1 { v14.s }[2], [x9]\n"
"st1 { v18.s }[2], [x27]\n"
"st1 { v22.s }[2], [x25]\n"
"st1 { v26.s }[2], [x23]\n"
- "b 183f\n"
- "178:" // Height 5: Partial direct writeback: partial_1_8
- "tbz x16, #0, 183f\n"
+ "b 178f\n"
+ "173:" // Height 5: Partial direct writeback: partial_1_8
+ "tbz x16, #0, 178f\n"
"str s10, [x13, #0x0]\n"
"str s14, [x9, #0x0]\n"
"str s18, [x27, #0x0]\n"
"str s22, [x25, #0x0]\n"
"str s26, [x23, #0x0]\n"
- "b 183f\n"
- "179:" // Height 5: Partial direct writeback: partial_4_0
- "tbz x16, #2, 181f\n"
+ "b 178f\n"
+ "174:" // Height 5: Partial direct writeback: partial_4_0
+ "tbz x16, #2, 176f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
"st1 { v16.4s }, [x27], #0x10\n"
"st1 { v20.4s }, [x25], #0x10\n"
"st1 { v24.4s }, [x23], #0x10\n"
- "tbz x16, #1, 180f\n"
+ "tbz x16, #1, 175f\n"
"str d9, [x13], #0x8\n"
"str d13, [x9], #0x8\n"
"str d17, [x27], #0x8\n"
"str d21, [x25], #0x8\n"
"str d25, [x23], #0x8\n"
- "tbz x16, #0, 183f\n"
+ "tbz x16, #0, 178f\n"
"st1 { v9.s }[2], [x13]\n"
"st1 { v13.s }[2], [x9]\n"
"st1 { v17.s }[2], [x27]\n"
"st1 { v21.s }[2], [x25]\n"
"st1 { v25.s }[2], [x23]\n"
- "b 183f\n"
- "180:" // Height 5: Partial direct writeback: partial_1_4
- "tbz x16, #0, 183f\n"
+ "b 178f\n"
+ "175:" // Height 5: Partial direct writeback: partial_1_4
+ "tbz x16, #0, 178f\n"
"str s9, [x13, #0x0]\n"
"str s13, [x9, #0x0]\n"
"str s17, [x27, #0x0]\n"
"str s21, [x25, #0x0]\n"
"str s25, [x23, #0x0]\n"
- "b 183f\n"
- "181:" // Height 5: Partial direct writeback: partial_2_0
- "tbz x16, #1, 182f\n"
+ "b 178f\n"
+ "176:" // Height 5: Partial direct writeback: partial_2_0
+ "tbz x16, #1, 177f\n"
"str d8, [x13], #0x8\n"
"str d12, [x9], #0x8\n"
"str d16, [x27], #0x8\n"
"str d20, [x25], #0x8\n"
"str d24, [x23], #0x8\n"
- "tbz x16, #0, 183f\n"
+ "tbz x16, #0, 178f\n"
"st1 { v8.s }[2], [x13]\n"
"st1 { v12.s }[2], [x9]\n"
"st1 { v16.s }[2], [x27]\n"
"st1 { v20.s }[2], [x25]\n"
"st1 { v24.s }[2], [x23]\n"
- "b 183f\n"
- "182:" // Height 5: Partial direct writeback: partial_1_0
+ "b 178f\n"
+ "177:" // Height 5: Partial direct writeback: partial_1_0
"str s8, [x13, #0x0]\n"
"str s12, [x9, #0x0]\n"
"str s16, [x27, #0x0]\n"
"str s20, [x25, #0x0]\n"
"str s24, [x23, #0x0]\n"
- "183:" // Height 5: Partial direct writeback: Done
- "b 185f\n"
- "184:" // Height 5: Full writeback
+ "178:" // Height 5: Partial direct writeback: Done
+ "b 180f\n"
+ "179:" // Height 5: Full writeback
"str q8, [x13, #0x0]\n"
"str q9, [x13, #0x10]\n"
"str q10, [x13, #0x20]\n"
@@ -2763,16 +2713,16 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x27, x27, #0x40\n"
"add x25, x25, #0x40\n"
"add x23, x23, #0x40\n"
- "185:" // Height 5: Writeback done
+ "180:" // Height 5: Writeback done
"subs x16, x16, #0x10\n"
- "bgt 151b\n"
- "b 224f\n"
- "186:" // Height 6
+ "bgt 147b\n"
+ "b 218f\n"
+ "181:" // Height 6
"ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
"mov x14, %x[bias]\n"
"ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "tbz %x[flags], #2, 187f\n"
+ "tbz %x[flags], #2, 182f\n"
"ldr x13, [%x[output_ptr], #0x0]\n"
"add x13, x13, x19, LSL #2\n"
"ldr x9, [%x[output_ptr], #0x8]\n"
@@ -2786,8 +2736,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x25, x25, x19, LSL #2\n"
"add x23, x23, x19, LSL #2\n"
"add x21, x21, x19, LSL #2\n"
- "b 188f\n"
- "187:" // Height 6: setup direct output
+ "b 183f\n"
+ "182:" // Height 6: setup direct output
"mov x13, %x[output_ptr]\n"
"add x9, x13, x19, LSL #2\n"
"add x27, x9, x19, LSL #2\n"
@@ -2795,8 +2745,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x23, x25, x19, LSL #2\n"
"add x21, x23, x19, LSL #2\n"
"add %x[output_ptr], x21, x19, LSL #2\n"
- "188:" // Height 6: Column loop
- "cbz x14, 189f\n"
+ "183:" // Height 6: Column loop
+ "cbz x14, 184f\n"
"ldr q8, [x14, #0x0]\n"
"mov v12.16b, v8.16b\n"
"ldr q9, [x14, #0x10]\n"
@@ -2822,12 +2772,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"mov v29.16b, v9.16b\n"
"mov v30.16b, v10.16b\n"
"mov v31.16b, v11.16b\n"
- "b 200f\n"
- "189:" // Height 6: no bias
- "tbz %x[flags], #0, 199f\n"
+ "b 195f\n"
+ "184:" // Height 6: no bias
+ "tbz %x[flags], #0, 194f\n"
"cmp x16, #0x10\n"
- "bge 198f\n"
- "tbz x16, #3, 193f\n"
+ "bge 193f\n"
+ "tbz x16, #3, 188f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
@@ -2840,14 +2790,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ld1 { v21.4s }, [x25], #0x10\n"
"ld1 { v25.4s }, [x23], #0x10\n"
"ld1 { v29.4s }, [x21], #0x10\n"
- "tbz x16, #2, 191f\n"
+ "tbz x16, #2, 186f\n"
"ld1 { v10.4s }, [x13], #0x10\n"
"ld1 { v14.4s }, [x9], #0x10\n"
"ld1 { v18.4s }, [x27], #0x10\n"
"ld1 { v22.4s }, [x25], #0x10\n"
"ld1 { v26.4s }, [x23], #0x10\n"
"ld1 { v30.4s }, [x21], #0x10\n"
- "tbz x16, #1, 190f\n"
+ "tbz x16, #1, 185f\n"
"mov x19, #0x38\n"
"ldr d11, [x13], #0x8\n"
"ldr d15, [x9], #0x8\n"
@@ -2855,26 +2805,26 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr d23, [x25], #0x8\n"
"ldr d27, [x23], #0x8\n"
"ldr d31, [x21], #0x8\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ld1 { v11.s }[2], [x13]\n"
"ld1 { v15.s }[2], [x9]\n"
"ld1 { v19.s }[2], [x27]\n"
"ld1 { v23.s }[2], [x25]\n"
"ld1 { v27.s }[2], [x23]\n"
"ld1 { v31.s }[2], [x21]\n"
- "b 197f\n"
- "190:" // Height 6: Partial accumulate: partial_1_12
+ "b 192f\n"
+ "185:" // Height 6: Partial accumulate: partial_1_12
"mov x19, #0x30\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ldr s11, [x13, #0x0]\n"
"ldr s15, [x9, #0x0]\n"
"ldr s19, [x27, #0x0]\n"
"ldr s23, [x25, #0x0]\n"
"ldr s27, [x23, #0x0]\n"
"ldr s31, [x21, #0x0]\n"
- "b 197f\n"
- "191:" // Height 6: Partial accumulate: partial_2_8
- "tbz x16, #1, 192f\n"
+ "b 192f\n"
+ "186:" // Height 6: Partial accumulate: partial_2_8
+ "tbz x16, #1, 187f\n"
"ldr d10, [x13], #0x8\n"
"ldr d14, [x9], #0x8\n"
"ldr d18, [x27], #0x8\n"
@@ -2882,33 +2832,33 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr d26, [x23], #0x8\n"
"ldr d30, [x21], #0x8\n"
"mov x19, #0x28\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ld1 { v10.s }[2], [x13]\n"
"ld1 { v14.s }[2], [x9]\n"
"ld1 { v18.s }[2], [x27]\n"
"ld1 { v22.s }[2], [x25]\n"
"ld1 { v26.s }[2], [x23]\n"
"ld1 { v30.s }[2], [x21]\n"
- "b 197f\n"
- "192:" // Height 6: Partial accumulate: partial_1_8
+ "b 192f\n"
+ "187:" // Height 6: Partial accumulate: partial_1_8
"mov x19, #0x20\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ldr s10, [x13, #0x0]\n"
"ldr s14, [x9, #0x0]\n"
"ldr s18, [x27, #0x0]\n"
"ldr s22, [x25, #0x0]\n"
"ldr s26, [x23, #0x0]\n"
"ldr s30, [x21, #0x0]\n"
- "b 197f\n"
- "193:" // Height 6: Partial accumulate: partial_4_0
- "tbz x16, #2, 195f\n"
+ "b 192f\n"
+ "188:" // Height 6: Partial accumulate: partial_4_0
+ "tbz x16, #2, 190f\n"
"ld1 { v8.4s }, [x13], #0x10\n"
"ld1 { v12.4s }, [x9], #0x10\n"
"ld1 { v16.4s }, [x27], #0x10\n"
"ld1 { v20.4s }, [x25], #0x10\n"
"ld1 { v24.4s }, [x23], #0x10\n"
"ld1 { v28.4s }, [x21], #0x10\n"
- "tbz x16, #1, 194f\n"
+ "tbz x16, #1, 189f\n"
"mov x19, #0x18\n"
"ldr d9, [x13], #0x8\n"
"ldr d13, [x9], #0x8\n"
@@ -2916,26 +2866,26 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr d21, [x25], #0x8\n"
"ldr d25, [x23], #0x8\n"
"ldr d29, [x21], #0x8\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ld1 { v9.s }[2], [x13]\n"
"ld1 { v13.s }[2], [x9]\n"
"ld1 { v17.s }[2], [x27]\n"
"ld1 { v21.s }[2], [x25]\n"
"ld1 { v25.s }[2], [x23]\n"
"ld1 { v29.s }[2], [x21]\n"
- "b 197f\n"
- "194:" // Height 6: Partial accumulate: partial_1_4
+ "b 192f\n"
+ "189:" // Height 6: Partial accumulate: partial_1_4
"mov x19, #0x10\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ldr s9, [x13, #0x0]\n"
"ldr s13, [x9, #0x0]\n"
"ldr s17, [x27, #0x0]\n"
"ldr s21, [x25, #0x0]\n"
"ldr s25, [x23, #0x0]\n"
"ldr s29, [x21, #0x0]\n"
- "b 197f\n"
- "195:" // Height 6: Partial accumulate: partial_2_0
- "tbz x16, #1, 196f\n"
+ "b 192f\n"
+ "190:" // Height 6: Partial accumulate: partial_2_0
+ "tbz x16, #1, 191f\n"
"ldr d8, [x13], #0x8\n"
"ldr d12, [x9], #0x8\n"
"ldr d16, [x27], #0x8\n"
@@ -2943,15 +2893,15 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr d24, [x23], #0x8\n"
"ldr d28, [x21], #0x8\n"
"mov x19, #0x8\n"
- "tbz x16, #0, 197f\n"
+ "tbz x16, #0, 192f\n"
"ld1 { v8.s }[2], [x13]\n"
"ld1 { v12.s }[2], [x9]\n"
"ld1 { v16.s }[2], [x27]\n"
"ld1 { v20.s }[2], [x25]\n"
"ld1 { v24.s }[2], [x23]\n"
"ld1 { v28.s }[2], [x21]\n"
- "b 197f\n"
- "196:" // Height 6: Partial accumulate: partial_1_0
+ "b 192f\n"
+ "191:" // Height 6: Partial accumulate: partial_1_0
"mov x19, #0x0\n"
"ldr s8, [x13, #0x0]\n"
"ldr s12, [x9, #0x0]\n"
@@ -2959,15 +2909,15 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr s20, [x25, #0x0]\n"
"ldr s24, [x23, #0x0]\n"
"ldr s28, [x21, #0x0]\n"
- "197:" // Height 6: Partial accumulate: Done
+ "192:" // Height 6: Partial accumulate: Done
"sub x13, x13, x19\n"
"sub x9, x9, x19\n"
"sub x27, x27, x19\n"
"sub x25, x25, x19\n"
"sub x23, x23, x19\n"
"sub x21, x21, x19\n"
- "b 200f\n"
- "198:" // Height 6: full accumulate
+ "b 195f\n"
+ "193:" // Height 6: full accumulate
"ldr q8, [x13, #0x0]\n"
"ldr q9, [x13, #0x10]\n"
"ldr q10, [x13, #0x20]\n"
@@ -2992,8 +2942,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr q29, [x21, #0x10]\n"
"ldr q30, [x21, #0x20]\n"
"ldr q31, [x21, #0x30]\n"
- "b 200f\n"
- "199:" // Height 6: no accumulate
+ "b 195f\n"
+ "194:" // Height 6: no accumulate
"movi v8.16b, #0x0\n"
"movi v9.16b, #0x0\n"
"movi v10.16b, #0x0\n"
@@ -3018,13 +2968,13 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"movi v29.16b, #0x0\n"
"movi v30.16b, #0x0\n"
"movi v31.16b, #0x0\n"
- "200:" // Height 6: setup done
+ "195:" // Height 6: setup done
"mov x12, #0x0\n"
- "201:" // Height 6: String loop
+ "196:" // Height 6: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
"ldr w11, [x20, x12, LSL #0x2]\n"
- "tbz %x[flags], #3, 202f\n"
+ "tbz %x[flags], #3, 197f\n"
"ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
"add x20, x20, x19, LSL #3\n"
"ldr x10, [x20, #0x0]\n"
@@ -3033,7 +2983,7 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"ldr x24, [x20, #0x18]\n"
"ldr x22, [x20, #0x20]\n"
"ldr x20, [x20, #0x28]\n"
- "cbnz x12, 203f\n"
+ "cbnz x12, 198f\n"
"ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x10, x10, x19, LSL #1\n"
"add x28, x28, x19, LSL #1\n"
@@ -3041,20 +2991,20 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x24, x24, x19, LSL #1\n"
"add x22, x22, x19, LSL #1\n"
"add x20, x20, x19, LSL #1\n"
- "b 203f\n"
- "202:" // Height 6: setup direct input
+ "b 198f\n"
+ "197:" // Height 6: setup direct input
"mov x10, %x[input_ptr]\n"
"add x28, x10, x19, LSL #1\n"
"add x26, x28, x19, LSL #1\n"
"add x24, x26, x19, LSL #1\n"
"add x22, x24, x19, LSL #1\n"
"add x20, x22, x19, LSL #1\n"
- "203:" // Height 6: input setup done
+ "198:" // Height 6: input setup done
"cmp x11, #0x8\n"
- "blt 206f\n"
+ "blt 201f\n"
"cmp x11, #0x10\n"
- "blt 205f\n"
- "204:" // Height 6: Multiply loop: Main loop head
+ "blt 200f\n"
+ "199:" // Height 6: Multiply loop: Main loop head
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
"ldr q2, [x26, #0x0]\n"
@@ -3188,8 +3138,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n"
".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n"
".inst 0x4f65f8ff // bfdot v31.4s, v7.8h, v5.h[3]\n"
- "bge 204b\n"
- "205:" // Height 6: Multiply loop: Single iteration only
+ "bge 199b\n"
+ "200:" // Height 6: Multiply loop: Single iteration only
"sub x11, x11, #0x8\n"
"ldr q0, [x10, #0x0]\n"
"ldr q1, [x28, #0x0]\n"
@@ -3322,11 +3272,11 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n"
".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n"
".inst 0x4f65f8ff // bfdot v31.4s, v7.8h, v5.h[3]\n"
- "206:" // Height 6: Multiply loop: Main loop skip
- "cbz x11, 211f\n"
+ "201:" // Height 6: Multiply loop: Main loop skip
+ "cbz x11, 205f\n"
"cmp x11, #0x2\n"
- "blt 208f\n"
- "207:" // Height 6: Multiply loop: Odd block loop
+ "blt 203f\n"
+ "202:" // Height 6: Multiply loop: Odd block loop
"ldr s0, [x10], #0x4\n"
"ldr s1, [x28], #0x4\n"
"ldr s2, [x26], #0x4\n"
@@ -3364,32 +3314,16 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n"
".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n"
".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n"
- "bge 207b\n"
- "cbz x11, 211f\n"
- "208:" // Height 6: Multiply loop: Skip odd blocks
- "tbz x11, #1, 209f\n"
- "ldr s0, [x10], #0x4\n"
- "ldr s1, [x28], #0x4\n"
- "ldr s2, [x26], #0x4\n"
- "ldr s3, [x24], #0x4\n"
- "ldr s4, [x22], #0x4\n"
- "ldr s5, [x20], #0x4\n"
- "tbz x11, #0, 210f\n"
- "ld1 { v0.h }[2], [x10]\n"
- "ld1 { v1.h }[2], [x28]\n"
- "ld1 { v2.h }[2], [x26]\n"
- "ld1 { v3.h }[2], [x24]\n"
- "ld1 { v4.h }[2], [x22]\n"
- "ld1 { v5.h }[2], [x20]\n"
- "b 210f\n"
- "209:" // Height 6: Multiply loop: Ragged operand read: partial_1_0
+ "bge 202b\n"
+ "cbz x11, 205f\n"
+ "203:" // Height 6: Multiply loop: Skip odd blocks
"ldr h0, [x10, #0x0]\n"
"ldr h1, [x28, #0x0]\n"
"ldr h2, [x26, #0x0]\n"
"ldr h3, [x24, #0x0]\n"
"ldr h4, [x22, #0x0]\n"
"ldr h5, [x20, #0x0]\n"
- "210:" // Height 6: Multiply loop: Ragged operand read: Done
+ "204:" // Height 6: Multiply loop: Ragged operand read: Done
"ldr q6, [x15, #0x0]\n"
".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n"
"ldr q7, [x15, #0x10]\n"
@@ -3419,18 +3353,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n"
".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n"
".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n"
- "211:" // Height 6: Multiply loop: No odd multiplies
+ "205:" // Height 6: Multiply loop: No odd multiplies
"ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x12, x12, #0x1\n"
"cmp x12, x19\n"
- "bne 201b\n"
+ "bne 196b\n"
"prfm pstl1keep, [x13, #0x0]\n"
"prfm pstl1keep, [x9, #0x0]\n"
"prfm pstl1keep, [x27, #0x0]\n"
"prfm pstl1keep, [x25, #0x0]\n"
"prfm pstl1keep, [x23, #0x0]\n"
"prfm pstl1keep, [x21, #0x0]\n"
- "tbz %x[flags], #1, 212f\n"
+ "tbz %x[flags], #1, 206f\n"
"add x19, %x[args_ptr], %[offset_min]\n"
"ld1r { v1.4s }, [x19]\n"
"add x19, %x[args_ptr], %[offset_max]\n"
@@ -3483,10 +3417,10 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"fmin v31.4s, v31.4s, v0.4s\n"
"fmax v30.4s, v30.4s, v1.4s\n"
"fmax v31.4s, v31.4s, v1.4s\n"
- "212:" // Height 6: No activation
+ "206:" // Height 6: No activation
"cmp x16, #0x10\n"
- "bge 221f\n"
- "tbz x16, #3, 216f\n"
+ "bge 215f\n"
+ "tbz x16, #3, 210f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v9.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
@@ -3499,120 +3433,120 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"st1 { v25.4s }, [x23], #0x10\n"
"st1 { v28.4s }, [x21], #0x10\n"
"st1 { v29.4s }, [x21], #0x10\n"
- "tbz x16, #2, 214f\n"
+ "tbz x16, #2, 208f\n"
"st1 { v10.4s }, [x13], #0x10\n"
"st1 { v14.4s }, [x9], #0x10\n"
"st1 { v18.4s }, [x27], #0x10\n"
"st1 { v22.4s }, [x25], #0x10\n"
"st1 { v26.4s }, [x23], #0x10\n"
"st1 { v30.4s }, [x21], #0x10\n"
- "tbz x16, #1, 213f\n"
+ "tbz x16, #1, 207f\n"
"str d11, [x13], #0x8\n"
"str d15, [x9], #0x8\n"
"str d19, [x27], #0x8\n"
"str d23, [x25], #0x8\n"
"str d27, [x23], #0x8\n"
"str d31, [x21], #0x8\n"
- "tbz x16, #0, 220f\n"
+ "tbz x16, #0, 214f\n"
"st1 { v11.s }[2], [x13]\n"
"st1 { v15.s }[2], [x9]\n"
"st1 { v19.s }[2], [x27]\n"
"st1 { v23.s }[2], [x25]\n"
"st1 { v27.s }[2], [x23]\n"
"st1 { v31.s }[2], [x21]\n"
- "b 220f\n"
- "213:" // Height 6: Partial direct writeback: partial_1_12
- "tbz x16, #0, 220f\n"
+ "b 214f\n"
+ "207:" // Height 6: Partial direct writeback: partial_1_12
+ "tbz x16, #0, 214f\n"
"str s11, [x13, #0x0]\n"
"str s15, [x9, #0x0]\n"
"str s19, [x27, #0x0]\n"
"str s23, [x25, #0x0]\n"
"str s27, [x23, #0x0]\n"
"str s31, [x21, #0x0]\n"
- "b 220f\n"
- "214:" // Height 6: Partial direct writeback: partial_2_8
- "tbz x16, #1, 215f\n"
+ "b 214f\n"
+ "208:" // Height 6: Partial direct writeback: partial_2_8
+ "tbz x16, #1, 209f\n"
"str d10, [x13], #0x8\n"
"str d14, [x9], #0x8\n"
"str d18, [x27], #0x8\n"
"str d22, [x25], #0x8\n"
"str d26, [x23], #0x8\n"
"str d30, [x21], #0x8\n"
- "tbz x16, #0, 220f\n"
+ "tbz x16, #0, 214f\n"
"st1 { v10.s }[2], [x13]\n"
"st1 { v14.s }[2], [x9]\n"
"st1 { v18.s }[2], [x27]\n"
"st1 { v22.s }[2], [x25]\n"
"st1 { v26.s }[2], [x23]\n"
"st1 { v30.s }[2], [x21]\n"
- "b 220f\n"
- "215:" // Height 6: Partial direct writeback: partial_1_8
- "tbz x16, #0, 220f\n"
+ "b 214f\n"
+ "209:" // Height 6: Partial direct writeback: partial_1_8
+ "tbz x16, #0, 214f\n"
"str s10, [x13, #0x0]\n"
"str s14, [x9, #0x0]\n"
"str s18, [x27, #0x0]\n"
"str s22, [x25, #0x0]\n"
"str s26, [x23, #0x0]\n"
"str s30, [x21, #0x0]\n"
- "b 220f\n"
- "216:" // Height 6: Partial direct writeback: partial_4_0
- "tbz x16, #2, 218f\n"
+ "b 214f\n"
+ "210:" // Height 6: Partial direct writeback: partial_4_0
+ "tbz x16, #2, 212f\n"
"st1 { v8.4s }, [x13], #0x10\n"
"st1 { v12.4s }, [x9], #0x10\n"
"st1 { v16.4s }, [x27], #0x10\n"
"st1 { v20.4s }, [x25], #0x10\n"
"st1 { v24.4s }, [x23], #0x10\n"
"st1 { v28.4s }, [x21], #0x10\n"
- "tbz x16, #1, 217f\n"
+ "tbz x16, #1, 211f\n"
"str d9, [x13], #0x8\n"
"str d13, [x9], #0x8\n"
"str d17, [x27], #0x8\n"
"str d21, [x25], #0x8\n"
"str d25, [x23], #0x8\n"
"str d29, [x21], #0x8\n"
- "tbz x16, #0, 220f\n"
+ "tbz x16, #0, 214f\n"
"st1 { v9.s }[2], [x13]\n"
"st1 { v13.s }[2], [x9]\n"
"st1 { v17.s }[2], [x27]\n"
"st1 { v21.s }[2], [x25]\n"
"st1 { v25.s }[2], [x23]\n"
"st1 { v29.s }[2], [x21]\n"
- "b 220f\n"
- "217:" // Height 6: Partial direct writeback: partial_1_4
- "tbz x16, #0, 220f\n"
+ "b 214f\n"
+ "211:" // Height 6: Partial direct writeback: partial_1_4
+ "tbz x16, #0, 214f\n"
"str s9, [x13, #0x0]\n"
"str s13, [x9, #0x0]\n"
"str s17, [x27, #0x0]\n"
"str s21, [x25, #0x0]\n"
"str s25, [x23, #0x0]\n"
"str s29, [x21, #0x0]\n"
- "b 220f\n"
- "218:" // Height 6: Partial direct writeback: partial_2_0
- "tbz x16, #1, 219f\n"
+ "b 214f\n"
+ "212:" // Height 6: Partial direct writeback: partial_2_0
+ "tbz x16, #1, 213f\n"
"str d8, [x13], #0x8\n"
"str d12, [x9], #0x8\n"
"str d16, [x27], #0x8\n"
"str d20, [x25], #0x8\n"
"str d24, [x23], #0x8\n"
"str d28, [x21], #0x8\n"
- "tbz x16, #0, 220f\n"
+ "tbz x16, #0, 214f\n"
"st1 { v8.s }[2], [x13]\n"
"st1 { v12.s }[2], [x9]\n"
"st1 { v16.s }[2], [x27]\n"
"st1 { v20.s }[2], [x25]\n"
"st1 { v24.s }[2], [x23]\n"
"st1 { v28.s }[2], [x21]\n"
- "b 220f\n"
- "219:" // Height 6: Partial direct writeback: partial_1_0
+ "b 214f\n"
+ "213:" // Height 6: Partial direct writeback: partial_1_0
"str s8, [x13, #0x0]\n"
"str s12, [x9, #0x0]\n"
"str s16, [x27, #0x0]\n"
"str s20, [x25, #0x0]\n"
"str s24, [x23, #0x0]\n"
"str s28, [x21, #0x0]\n"
- "220:" // Height 6: Partial direct writeback: Done
- "b 222f\n"
- "221:" // Height 6: Full writeback
+ "214:" // Height 6: Partial direct writeback: Done
+ "b 216f\n"
+ "215:" // Height 6: Full writeback
"str q8, [x13, #0x0]\n"
"str q9, [x13, #0x10]\n"
"str q10, [x13, #0x20]\n"
@@ -3643,23 +3577,23 @@ void a64_hybrid_bf16fp32_dot_6x16 (
"add x25, x25, #0x40\n"
"add x23, x23, #0x40\n"
"add x21, x21, #0x40\n"
- "222:" // Height 6: Writeback done
+ "216:" // Height 6: Writeback done
"subs x16, x16, #0x10\n"
- "bgt 188b\n"
+ "bgt 183b\n"
"subs %x[M], %x[M], #0x6\n"
- "beq 224f\n"
+ "beq 218f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
- "tbz %x[flags], #3, 223f\n"
+ "tbz %x[flags], #3, 217f\n"
"add x20, x20, #0x6\n"
"str x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
"b 1b\n"
- "223:" // Update direct input
+ "217:" // Update direct input
"mov x19, #0xc\n"
"madd %x[input_ptr], x19, x20, %x[input_ptr]\n"
"b 1b\n"
- "224:" // Exit
+ "218:" // Exit
- : [M] "+r" (M), [input_ptr] "+r" (input_ptr), [output_ptr] "+r" (output_ptr)
+ : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
: [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-18 03:03:36 +0000