aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp3620
1 files changed, 1809 insertions, 1811 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
index 598d1524e8..f3942328a6 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
@@ -85,7 +85,6 @@ void a64_hybrid_s8qs_dot_6x16 (
flags |= 0x20;
}
__asm__ __volatile__(
-
"1:" // Row loop
"cmp %x[M], #0x6\n"
"bge 136f\n"
@@ -111,11 +110,11 @@ void a64_hybrid_s8qs_dot_6x16 (
"4:" // Height 1: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, 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, 5f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
"cbnz x28, 6f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -132,37 +131,37 @@ void a64_hybrid_s8qs_dot_6x16 (
"blt 8f\n"
"7:" // Height 1: Multiply loop: Main loop head
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q17, [x9, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ "ldr q16, [x9, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ "ldr q17, [x9, #0x40]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ "ldr q16, [x9, #0x50]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x9, #0x60]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x9, #0x70]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x9, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x9, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x9, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x9, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x9, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x9, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ "ldr q17, [x9, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ "ldr q16, [x9, #0xf0]\n"
"sub x27, x27, #0x10\n"
"add x26, x26, #0x10\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
"cmp x27, #0x20\n"
"add x9, x9, #0x100\n"
@@ -172,37 +171,37 @@ void a64_hybrid_s8qs_dot_6x16 (
"bge 7b\n"
"8:" // Height 1: Multiply loop: Single iteration only
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q17, [x9, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ "ldr q16, [x9, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ "ldr q17, [x9, #0x40]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ "ldr q16, [x9, #0x50]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x9, #0x60]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x9, #0x70]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x9, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x9, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x9, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x9, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x9, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x9, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ "ldr q17, [x9, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ "ldr q16, [x9, #0xf0]\n"
"add x26, x26, #0x10\n"
"sub x27, x27, #0x10\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
"prfm pldl1keep, [x26, #0x80]\n"
"add x9, x9, #0x100\n"
"9:" // Height 1: Multiply loop: Main loop skip
@@ -210,17 +209,17 @@ void a64_hybrid_s8qs_dot_6x16 (
"cmp x27, #0x4\n"
"blt 11f\n"
"10:" // Height 1: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr q6, [x9, #0x0]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
+ "ldr s18, [x26], #0x4\n"
+ "ldr q16, [x9, #0x0]\n"
+ ".inst 0x4f92e208 // sdot v8.4s, v16.16b, v18.4b[0]\n"
"sub x27, x27, #0x4\n"
- "ldr q7, [x9, #0x10]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
+ "ldr q16, [x9, #0x10]\n"
+ "ldr q17, [x9, #0x20]\n"
+ ".inst 0x4f92e209 // sdot v9.4s, v16.16b, v18.4b[0]\n"
"cmp x27, #0x4\n"
- "ldr q7, [x9, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ "ldr q16, [x9, #0x30]\n"
+ ".inst 0x4f92e22a // sdot v10.4s, v17.16b, v18.4b[0]\n"
+ ".inst 0x4f92e20b // sdot v11.4s, v16.16b, v18.4b[0]\n"
"add x9, x9, #0x40\n"
"bge 10b\n"
"11:" // Height 1: Multiply loop: Skip odd blocks
@@ -233,28 +232,28 @@ void a64_hybrid_s8qs_dot_6x16 (
"12:" // Height 1: Multiply loop: Ragged operand read: partial_1_0
"ldr b0, [x26, #0x0]\n"
"13:" // Height 1: Multiply loop: Ragged operand read: Done
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- "ldr q7, [x9, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ "ldr q17, [x9, #0x0]\n"
+ "ldr q16, [x9, #0x10]\n"
+ ".inst 0x4f80e228 // sdot v8.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f80e209 // sdot v9.4s, v16.16b, v0.4b[0]\n"
+ "ldr q17, [x9, #0x20]\n"
+ "ldr q16, [x9, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
"add x9, x9, #0x40\n"
"14:" // Height 1: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
"cmp x28, x20\n"
"bne 4b\n"
- "ldr q0, [x14, #0x0]\n"
- "ldr q1, [x14, #0x10]\n"
- "add v8.4s, v8.4s, v0.4s\n"
- "add v9.4s, v9.4s, v1.4s\n"
- "ldr q2, [x14, #0x20]\n"
- "ldr q3, [x14, #0x30]\n"
- "add v10.4s, v10.4s, v2.4s\n"
- "add v11.4s, v11.4s, v3.4s\n"
+ "ldr q17, [x14, #0x0]\n"
+ "ldr q16, [x14, #0x10]\n"
+ "add v8.4s, v8.4s, v17.4s\n"
+ "add v9.4s, v9.4s, v16.4s\n"
+ "ldr q17, [x14, #0x20]\n"
+ "ldr q16, [x14, #0x30]\n"
+ "add v10.4s, v10.4s, v17.4s\n"
+ "add v11.4s, v11.4s, v16.4s\n"
"prfm pstl1keep, [x11, #0x0]\n"
"add x14, x14, #0x40\n"
"tbz %x[flags], #4, 15f\n"
@@ -270,10 +269,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x13, x13, #0x40\n"
"b 16f\n"
"15:" // Height 1: per layer parameters
- "add x25, %x[qp], %[per_layer_right_shift]\n"
- "ld1r { v0.4s }, [x25]\n"
- "add x25, %x[qp], %[per_layer_mul]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v4.4s }, [x20]\n"
"mov v1.16b, v0.16b\n"
"mov v5.16b, v4.16b\n"
"mov v2.16b, v0.16b\n"
@@ -286,45 +285,45 @@ void a64_hybrid_s8qs_dot_6x16 (
"sqrdmulh v10.4s, v10.4s, v6.4s\n"
"sqrdmulh v11.4s, v11.4s, v7.4s\n"
"tbz %x[flags], #5, 17f\n"
- "and v4.16b, v8.16b, v0.16b\n"
- "and v5.16b, v9.16b, v1.16b\n"
- "and v6.16b, v10.16b, v2.16b\n"
- "and v7.16b, v11.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v8.4s, v8.4s, v4.4s\n"
- "sqadd v9.4s, v9.4s, v5.4s\n"
- "sqadd v10.4s, v10.4s, v6.4s\n"
- "sqadd v11.4s, v11.4s, v7.4s\n"
+ "and v19.16b, v8.16b, v0.16b\n"
+ "and v18.16b, v9.16b, v1.16b\n"
+ "and v17.16b, v10.16b, v2.16b\n"
+ "and v16.16b, v11.16b, v3.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sqadd v8.4s, v8.4s, v19.4s\n"
+ "sqadd v9.4s, v9.4s, v18.4s\n"
+ "sqadd v10.4s, v10.4s, v17.4s\n"
+ "sqadd v11.4s, v11.4s, v16.4s\n"
"17:" // Height 1: no shift correction
- "add x25, %x[qp], %[c_offset]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v18.4s }, [x20]\n"
"srshl v8.4s, v8.4s, v0.4s\n"
"srshl v9.4s, v9.4s, v1.4s\n"
"srshl v10.4s, v10.4s, v2.4s\n"
"srshl v11.4s, v11.4s, v3.4s\n"
- "add x25, %x[qp], %[maxval]\n"
- "ld1r { v6.4s }, [x25]\n"
- "add v8.4s, v8.4s, v4.4s\n"
- "add v9.4s, v9.4s, v4.4s\n"
- "add x25, %x[qp], %[minval]\n"
- "ld1r { v5.4s }, [x25]\n"
- "add v10.4s, v10.4s, v4.4s\n"
- "add v11.4s, v11.4s, v4.4s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v17.4s }, [x20]\n"
+ "add v8.4s, v8.4s, v18.4s\n"
+ "add v9.4s, v9.4s, v18.4s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v16.4s }, [x20]\n"
+ "add v10.4s, v10.4s, v18.4s\n"
+ "add v11.4s, v11.4s, v18.4s\n"
"cmp x10, #0x10\n"
- "smin v8.4s, v8.4s, v6.4s\n"
- "smin v9.4s, v9.4s, v6.4s\n"
- "smin v10.4s, v10.4s, v6.4s\n"
- "smin v11.4s, v11.4s, v6.4s\n"
- "smax v8.4s, v8.4s, v5.4s\n"
- "smax v9.4s, v9.4s, v5.4s\n"
- "smax v10.4s, v10.4s, v5.4s\n"
- "smax v11.4s, v11.4s, v5.4s\n"
+ "smin v8.4s, v8.4s, v17.4s\n"
+ "smin v9.4s, v9.4s, v17.4s\n"
+ "smin v10.4s, v10.4s, v17.4s\n"
+ "smin v11.4s, v11.4s, v17.4s\n"
+ "smax v8.4s, v8.4s, v16.4s\n"
+ "smax v9.4s, v9.4s, v16.4s\n"
+ "smax v10.4s, v10.4s, v16.4s\n"
+ "smax v11.4s, v11.4s, v16.4s\n"
"uzp1 v8.8h, v8.8h, v9.8h\n"
- "uzp1 v9.8h, v10.8h, v11.8h\n"
- "uzp1 v8.16b, v8.16b, v9.16b\n"
+ "uzp1 v16.8h, v10.8h, v11.8h\n"
+ "uzp1 v8.16b, v8.16b, v16.16b\n"
"bge 26f\n"
"tbz x10, #3, 21f\n"
"str d8, [x11], #0x8\n"
@@ -399,12 +398,12 @@ void a64_hybrid_s8qs_dot_6x16 (
"31:" // Height 2: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, 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, 32f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
"cbnz x28, 33f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -412,7 +411,7 @@ void a64_hybrid_s8qs_dot_6x16 (
"b 33f\n"
"32:" // Height 2: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
+ "add x25, x26, x21\n"
"33:" // Height 2: input setup done
"cmp x27, #0x10\n"
"blt 36f\n"
@@ -425,137 +424,137 @@ void a64_hybrid_s8qs_dot_6x16 (
"34:" // Height 2: Multiply loop: Main loop head
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q17, [x9, #0x20]\n"
"sub x27, x27, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q16, [x9, #0x30]\n"
"add x26, x26, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n"
+ "ldr q17, [x9, #0x40]\n"
"add x25, x25, #0x10\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n"
+ "ldr q16, [x9, #0x50]\n"
"cmp x27, #0x20\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22c // sdot v12.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x9, #0x60]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20d // sdot v13.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x9, #0x70]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22e // sdot v14.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x9, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20f // sdot v15.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x9, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2c // sdot v12.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x9, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0d // sdot v13.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x9, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2e // sdot v14.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x9, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0f // sdot v15.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x9, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2c // sdot v12.4s, v17.16b, v1.4b[3]\n"
+ "ldr q17, [x9, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea0d // sdot v13.4s, v16.16b, v1.4b[3]\n"
+ "ldr q16, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2e // sdot v14.4s, v17.16b, v1.4b[3]\n"
"ldr q6, [x9, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1ea0f // sdot v15.4s, v16.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
"ldr q7, [x9, #0x10]\n"
"bge 34b\n"
"35:" // Height 2: Multiply loop: Single iteration only
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q17, [x9, #0x20]\n"
"add x26, x26, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q16, [x9, #0x30]\n"
"add x25, x25, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n"
+ "ldr q17, [x9, #0x40]\n"
"sub x27, x27, #0x10\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n"
+ "ldr q16, [x9, #0x50]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22c // sdot v12.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x9, #0x60]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20d // sdot v13.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x9, #0x70]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22e // sdot v14.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x9, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20f // sdot v15.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x9, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2c // sdot v12.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x9, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0d // sdot v13.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x9, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2e // sdot v14.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x9, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0f // sdot v15.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x9, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2c // sdot v12.4s, v17.16b, v1.4b[3]\n"
+ "ldr q17, [x9, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea0d // sdot v13.4s, v16.16b, v1.4b[3]\n"
+ "ldr q16, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2e // sdot v14.4s, v17.16b, v1.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea0f // sdot v15.4s, v16.16b, v1.4b[3]\n"
"36:" // Height 2: Multiply loop: Main loop skip
"cbz x27, 41f\n"
"cmp x27, #0x4\n"
"blt 38f\n"
"37:" // Height 2: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s19, [x26], #0x4\n"
+ "ldr s18, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ "ldr q17, [x9, #0x0]\n"
+ "ldr q16, [x9, #0x10]\n"
+ ".inst 0x4f93e228 // sdot v8.4s, v17.16b, v19.4b[0]\n"
+ ".inst 0x4f92e22c // sdot v12.4s, v17.16b, v18.4b[0]\n"
+ "ldr q17, [x9, #0x20]\n"
+ ".inst 0x4f93e209 // sdot v9.4s, v16.16b, v19.4b[0]\n"
+ ".inst 0x4f92e20d // sdot v13.4s, v16.16b, v18.4b[0]\n"
+ "ldr q16, [x9, #0x30]\n"
+ ".inst 0x4f93e22a // sdot v10.4s, v17.16b, v19.4b[0]\n"
+ ".inst 0x4f92e22e // sdot v14.4s, v17.16b, v18.4b[0]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f93e20b // sdot v11.4s, v16.16b, v19.4b[0]\n"
+ ".inst 0x4f92e20f // sdot v15.4s, v16.16b, v18.4b[0]\n"
"bge 37b\n"
"38:" // Height 2: Multiply loop: Skip odd blocks
"cbz x27, 41f\n"
@@ -570,41 +569,41 @@ void a64_hybrid_s8qs_dot_6x16 (
"ldr b0, [x26, #0x0]\n"
"ldr b1, [x25, #0x0]\n"
"40:" // Height 2: Multiply loop: Ragged operand read: Done
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ "ldr q17, [x9, #0x0]\n"
+ "ldr q16, [x9, #0x10]\n"
+ ".inst 0x4f80e228 // sdot v8.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22c // sdot v12.4s, v17.16b, v1.4b[0]\n"
+ "ldr q17, [x9, #0x20]\n"
+ ".inst 0x4f80e209 // sdot v9.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20d // sdot v13.4s, v16.16b, v1.4b[0]\n"
+ "ldr q16, [x9, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n"
"41:" // Height 2: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
"cmp x28, x20\n"
"bne 31b\n"
- "ldr q0, [x14, #0x0]\n"
- "ldr q1, [x14, #0x10]\n"
- "add v8.4s, v8.4s, v0.4s\n"
- "add v9.4s, v9.4s, v1.4s\n"
- "ldr q2, [x14, #0x20]\n"
- "ldr q3, [x14, #0x30]\n"
- "add v10.4s, v10.4s, v2.4s\n"
- "add v11.4s, v11.4s, v3.4s\n"
+ "ldr q19, [x14, #0x0]\n"
+ "ldr q18, [x14, #0x10]\n"
+ "add v8.4s, v8.4s, v19.4s\n"
+ "add v9.4s, v9.4s, v18.4s\n"
+ "ldr q17, [x14, #0x20]\n"
+ "ldr q16, [x14, #0x30]\n"
+ "add v10.4s, v10.4s, v17.4s\n"
+ "add v11.4s, v11.4s, v16.4s\n"
"ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "add x24, x11, x20\n"
+ "add x25, x11, x20\n"
"prfm pstl1keep, [x11, #0x0]\n"
- "add v12.4s, v12.4s, v0.4s\n"
- "prfm pstl1keep, [x24, #0x0]\n"
- "add v13.4s, v13.4s, v1.4s\n"
- "add v14.4s, v14.4s, v2.4s\n"
+ "add v12.4s, v12.4s, v19.4s\n"
+ "prfm pstl1keep, [x25, #0x0]\n"
+ "add v13.4s, v13.4s, v18.4s\n"
+ "add v14.4s, v14.4s, v17.4s\n"
"add x14, x14, #0x40\n"
- "add v15.4s, v15.4s, v3.4s\n"
+ "add v15.4s, v15.4s, v16.4s\n"
"tbz %x[flags], #4, 42f\n"
"ldr q0, [x12, #0x0]\n"
"ldr q4, [x13, #0x0]\n"
@@ -618,10 +617,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x13, x13, #0x40\n"
"b 43f\n"
"42:" // Height 2: per layer parameters
- "add x25, %x[qp], %[per_layer_right_shift]\n"
- "ld1r { v0.4s }, [x25]\n"
- "add x25, %x[qp], %[per_layer_mul]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v4.4s }, [x20]\n"
"mov v1.16b, v0.16b\n"
"mov v5.16b, v4.16b\n"
"mov v2.16b, v0.16b\n"
@@ -638,141 +637,141 @@ void a64_hybrid_s8qs_dot_6x16 (
"sqrdmulh v14.4s, v14.4s, v6.4s\n"
"sqrdmulh v15.4s, v15.4s, v7.4s\n"
"tbz %x[flags], #5, 44f\n"
- "and v4.16b, v8.16b, v0.16b\n"
- "and v5.16b, v9.16b, v1.16b\n"
- "and v6.16b, v10.16b, v2.16b\n"
- "and v7.16b, v11.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v8.4s, v8.4s, v4.4s\n"
- "sqadd v9.4s, v9.4s, v5.4s\n"
- "sqadd v10.4s, v10.4s, v6.4s\n"
- "sqadd v11.4s, v11.4s, v7.4s\n"
- "and v4.16b, v12.16b, v0.16b\n"
- "and v5.16b, v13.16b, v1.16b\n"
- "and v6.16b, v14.16b, v2.16b\n"
- "and v7.16b, v15.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v12.4s, v12.4s, v4.4s\n"
- "sqadd v13.4s, v13.4s, v5.4s\n"
- "sqadd v14.4s, v14.4s, v6.4s\n"
- "sqadd v15.4s, v15.4s, v7.4s\n"
+ "and v19.16b, v8.16b, v0.16b\n"
+ "and v18.16b, v9.16b, v1.16b\n"
+ "and v17.16b, v10.16b, v2.16b\n"
+ "and v16.16b, v11.16b, v3.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sqadd v8.4s, v8.4s, v19.4s\n"
+ "sqadd v9.4s, v9.4s, v18.4s\n"
+ "sqadd v10.4s, v10.4s, v17.4s\n"
+ "sqadd v11.4s, v11.4s, v16.4s\n"
+ "and v19.16b, v12.16b, v0.16b\n"
+ "and v18.16b, v13.16b, v1.16b\n"
+ "and v17.16b, v14.16b, v2.16b\n"
+ "and v16.16b, v15.16b, v3.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sqadd v12.4s, v12.4s, v19.4s\n"
+ "sqadd v13.4s, v13.4s, v18.4s\n"
+ "sqadd v14.4s, v14.4s, v17.4s\n"
+ "sqadd v15.4s, v15.4s, v16.4s\n"
"44:" // Height 2: no shift correction
- "add x25, %x[qp], %[c_offset]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v18.4s }, [x20]\n"
"srshl v8.4s, v8.4s, v0.4s\n"
"srshl v9.4s, v9.4s, v1.4s\n"
"srshl v10.4s, v10.4s, v2.4s\n"
"srshl v11.4s, v11.4s, v3.4s\n"
- "add x25, %x[qp], %[maxval]\n"
- "ld1r { v6.4s }, [x25]\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v17.4s }, [x20]\n"
"srshl v12.4s, v12.4s, v0.4s\n"
"srshl v13.4s, v13.4s, v1.4s\n"
- "add x25, %x[qp], %[minval]\n"
- "ld1r { v5.4s }, [x25]\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v16.4s }, [x20]\n"
"srshl v14.4s, v14.4s, v2.4s\n"
"srshl v15.4s, v15.4s, v3.4s\n"
"cmp x10, #0x10\n"
- "add v8.4s, v8.4s, v4.4s\n"
- "add v9.4s, v9.4s, v4.4s\n"
- "add v10.4s, v10.4s, v4.4s\n"
- "add v11.4s, v11.4s, v4.4s\n"
- "add v12.4s, v12.4s, v4.4s\n"
- "add v13.4s, v13.4s, v4.4s\n"
- "add v14.4s, v14.4s, v4.4s\n"
- "add v15.4s, v15.4s, v4.4s\n"
- "smin v8.4s, v8.4s, v6.4s\n"
- "smin v9.4s, v9.4s, v6.4s\n"
- "smin v10.4s, v10.4s, v6.4s\n"
- "smin v11.4s, v11.4s, v6.4s\n"
- "smin v12.4s, v12.4s, v6.4s\n"
- "smin v13.4s, v13.4s, v6.4s\n"
- "smin v14.4s, v14.4s, v6.4s\n"
- "smin v15.4s, v15.4s, v6.4s\n"
- "smax v8.4s, v8.4s, v5.4s\n"
- "smax v9.4s, v9.4s, v5.4s\n"
- "smax v10.4s, v10.4s, v5.4s\n"
- "smax v11.4s, v11.4s, v5.4s\n"
- "smax v12.4s, v12.4s, v5.4s\n"
- "smax v13.4s, v13.4s, v5.4s\n"
- "smax v14.4s, v14.4s, v5.4s\n"
- "smax v15.4s, v15.4s, v5.4s\n"
+ "add v8.4s, v8.4s, v18.4s\n"
+ "add v9.4s, v9.4s, v18.4s\n"
+ "add v10.4s, v10.4s, v18.4s\n"
+ "add v11.4s, v11.4s, v18.4s\n"
+ "add v12.4s, v12.4s, v18.4s\n"
+ "add v13.4s, v13.4s, v18.4s\n"
+ "add v14.4s, v14.4s, v18.4s\n"
+ "add v15.4s, v15.4s, v18.4s\n"
+ "smin v8.4s, v8.4s, v17.4s\n"
+ "smin v9.4s, v9.4s, v17.4s\n"
+ "smin v10.4s, v10.4s, v17.4s\n"
+ "smin v11.4s, v11.4s, v17.4s\n"
+ "smin v12.4s, v12.4s, v17.4s\n"
+ "smin v13.4s, v13.4s, v17.4s\n"
+ "smin v14.4s, v14.4s, v17.4s\n"
+ "smin v15.4s, v15.4s, v17.4s\n"
+ "smax v8.4s, v8.4s, v16.4s\n"
+ "smax v9.4s, v9.4s, v16.4s\n"
+ "smax v10.4s, v10.4s, v16.4s\n"
+ "smax v11.4s, v11.4s, v16.4s\n"
+ "smax v12.4s, v12.4s, v16.4s\n"
+ "smax v13.4s, v13.4s, v16.4s\n"
+ "smax v14.4s, v14.4s, v16.4s\n"
+ "smax v15.4s, v15.4s, v16.4s\n"
"uzp1 v8.8h, v8.8h, v9.8h\n"
- "uzp1 v9.8h, v10.8h, v11.8h\n"
+ "uzp1 v17.8h, v10.8h, v11.8h\n"
"uzp1 v12.8h, v12.8h, v13.8h\n"
- "uzp1 v13.8h, v14.8h, v15.8h\n"
- "uzp1 v8.16b, v8.16b, v9.16b\n"
- "uzp1 v12.16b, v12.16b, v13.16b\n"
+ "uzp1 v16.8h, v14.8h, v15.8h\n"
+ "uzp1 v8.16b, v8.16b, v17.16b\n"
+ "uzp1 v12.16b, v12.16b, v16.16b\n"
"bge 53f\n"
"tbz x10, #3, 48f\n"
"str d8, [x11], #0x8\n"
- "str d12, [x24], #0x8\n"
+ "str d12, [x25], #0x8\n"
"tbz x10, #2, 46f\n"
"st1 { v8.s }[2], [x11], #0x4\n"
- "st1 { v12.s }[2], [x24], #0x4\n"
+ "st1 { v12.s }[2], [x25], #0x4\n"
"tbz x10, #1, 45f\n"
"st1 { v8.h }[6], [x11], #0x2\n"
- "st1 { v12.h }[6], [x24], #0x2\n"
+ "st1 { v12.h }[6], [x25], #0x2\n"
"tbz x10, #0, 52f\n"
"st1 { v8.b }[14], [x11]\n"
- "st1 { v12.b }[14], [x24]\n"
+ "st1 { v12.b }[14], [x25]\n"
"b 52f\n"
"45:" // Height 2: Partial direct writeback: partial_1_12
"tbz x10, #0, 52f\n"
"st1 { v8.b }[12], [x11]\n"
- "st1 { v12.b }[12], [x24]\n"
+ "st1 { v12.b }[12], [x25]\n"
"b 52f\n"
"46:" // Height 2: Partial direct writeback: partial_2_8
"tbz x10, #1, 47f\n"
"st1 { v8.h }[4], [x11], #0x2\n"
- "st1 { v12.h }[4], [x24], #0x2\n"
+ "st1 { v12.h }[4], [x25], #0x2\n"
"tbz x10, #0, 52f\n"
"st1 { v8.b }[10], [x11]\n"
- "st1 { v12.b }[10], [x24]\n"
+ "st1 { v12.b }[10], [x25]\n"
"b 52f\n"
"47:" // Height 2: Partial direct writeback: partial_1_8
"tbz x10, #0, 52f\n"
"st1 { v8.b }[8], [x11]\n"
- "st1 { v12.b }[8], [x24]\n"
+ "st1 { v12.b }[8], [x25]\n"
"b 52f\n"
"48:" // Height 2: Partial direct writeback: partial_4_0
"tbz x10, #2, 50f\n"
"str s8, [x11], #0x4\n"
- "str s12, [x24], #0x4\n"
+ "str s12, [x25], #0x4\n"
"tbz x10, #1, 49f\n"
"st1 { v8.h }[2], [x11], #0x2\n"
- "st1 { v12.h }[2], [x24], #0x2\n"
+ "st1 { v12.h }[2], [x25], #0x2\n"
"tbz x10, #0, 52f\n"
"st1 { v8.b }[6], [x11]\n"
- "st1 { v12.b }[6], [x24]\n"
+ "st1 { v12.b }[6], [x25]\n"
"b 52f\n"
"49:" // Height 2: Partial direct writeback: partial_1_4
"tbz x10, #0, 52f\n"
"st1 { v8.b }[4], [x11]\n"
- "st1 { v12.b }[4], [x24]\n"
+ "st1 { v12.b }[4], [x25]\n"
"b 52f\n"
"50:" // Height 2: Partial direct writeback: partial_2_0
"tbz x10, #1, 51f\n"
"str h8, [x11], #0x2\n"
- "str h12, [x24], #0x2\n"
+ "str h12, [x25], #0x2\n"
"tbz x10, #0, 52f\n"
"st1 { v8.b }[2], [x11]\n"
- "st1 { v12.b }[2], [x24]\n"
+ "st1 { v12.b }[2], [x25]\n"
"b 52f\n"
"51:" // Height 2: Partial direct writeback: partial_1_0
"str b8, [x11, #0x0]\n"
- "str b12, [x24, #0x0]\n"
+ "str b12, [x25, #0x0]\n"
"52:" // Height 2: Partial direct writeback: Done
"b 54f\n"
"53:" // Height 2: Full writeback
"str q8, [x11, #0x0]\n"
"add x11, x11, #0x10\n"
- "str q12, [x24, #0x0]\n"
+ "str q12, [x25, #0x0]\n"
"54:" // Height 2: Writeback done
"subs x10, x10, #0x10\n"
"bgt 29b\n"
@@ -802,13 +801,13 @@ void a64_hybrid_s8qs_dot_6x16 (
"58:" // Height 3: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, 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, 59f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
"cbnz x28, 60f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -817,8 +816,8 @@ void a64_hybrid_s8qs_dot_6x16 (
"b 60f\n"
"59:" // Height 3: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
"60:" // Height 3: input setup done
"cmp x27, #0x10\n"
"blt 63f\n"
@@ -835,75 +834,75 @@ void a64_hybrid_s8qs_dot_6x16 (
"sub x27, x27, #0x10\n"
"add x26, x26, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q21, [x9, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x25, x25, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q20, [x9, #0x30]\n"
"add x24, x24, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n"
"cmp x27, #0x20\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n"
+ "ldr q21, [x9, #0x40]\n"
+ ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
+ ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n"
+ "ldr q20, [x9, #0x50]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4fa0e2a8 // sdot v8.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ac // sdot v12.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b0 // sdot v16.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x9, #0x60]\n"
+ ".inst 0x4fa0e289 // sdot v9.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28d // sdot v13.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e291 // sdot v17.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x9, #0x70]\n"
+ ".inst 0x4fa0e2aa // sdot v10.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ae // sdot v14.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b2 // sdot v18.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x9, #0x80]\n"
+ ".inst 0x4fa0e28b // sdot v11.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28f // sdot v15.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e293 // sdot v19.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x9, #0x90]\n"
+ ".inst 0x4f80eaa8 // sdot v8.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaac // sdot v12.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab0 // sdot v16.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x9, #0xa0]\n"
+ ".inst 0x4f80ea89 // sdot v9.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8d // sdot v13.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea91 // sdot v17.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x9, #0xb0]\n"
+ ".inst 0x4f80eaaa // sdot v10.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaae // sdot v14.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab2 // sdot v18.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x9, #0xc0]\n"
+ ".inst 0x4f80ea8b // sdot v11.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8f // sdot v15.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea93 // sdot v19.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x9, #0xd0]\n"
+ ".inst 0x4fa0eaa8 // sdot v8.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaac // sdot v12.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab0 // sdot v16.4s, v21.16b, v2.4b[3]\n"
+ "ldr q21, [x9, #0xe0]\n"
+ ".inst 0x4fa0ea89 // sdot v9.4s, v20.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea8d // sdot v13.4s, v20.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea91 // sdot v17.4s, v20.16b, v2.4b[3]\n"
+ "ldr q20, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
+ ".inst 0x4fa0eaaa // sdot v10.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaae // sdot v14.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab2 // sdot v18.4s, v21.16b, v2.4b[3]\n"
"ldr q6, [x9, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea8b // sdot v11.4s, v20.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1ea8f // sdot v15.4s, v20.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa2ea93 // sdot v19.4s, v20.16b, v2.4b[3]\n"
"ldr q2, [x24, #0x0]\n"
"ldr q7, [x9, #0x10]\n"
"bge 61b\n"
@@ -913,98 +912,98 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x26, x26, #0x10\n"
"add x25, x25, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q21, [x9, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x24, x24, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q20, [x9, #0x30]\n"
"sub x27, x27, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n"
"prfm pldl1keep, [x26, #0x80]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n"
+ "ldr q21, [x9, #0x40]\n"
+ ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n"
+ "ldr q20, [x9, #0x50]\n"
+ ".inst 0x4fa0e2a8 // sdot v8.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ac // sdot v12.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b0 // sdot v16.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x9, #0x60]\n"
+ ".inst 0x4fa0e289 // sdot v9.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28d // sdot v13.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e291 // sdot v17.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x9, #0x70]\n"
+ ".inst 0x4fa0e2aa // sdot v10.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ae // sdot v14.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b2 // sdot v18.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x9, #0x80]\n"
+ ".inst 0x4fa0e28b // sdot v11.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28f // sdot v15.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e293 // sdot v19.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x9, #0x90]\n"
+ ".inst 0x4f80eaa8 // sdot v8.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaac // sdot v12.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab0 // sdot v16.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x9, #0xa0]\n"
+ ".inst 0x4f80ea89 // sdot v9.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8d // sdot v13.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea91 // sdot v17.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x9, #0xb0]\n"
+ ".inst 0x4f80eaaa // sdot v10.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaae // sdot v14.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab2 // sdot v18.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x9, #0xc0]\n"
+ ".inst 0x4f80ea8b // sdot v11.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8f // sdot v15.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea93 // sdot v19.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x9, #0xd0]\n"
+ ".inst 0x4fa0eaa8 // sdot v8.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaac // sdot v12.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab0 // sdot v16.4s, v21.16b, v2.4b[3]\n"
+ "ldr q21, [x9, #0xe0]\n"
+ ".inst 0x4fa0ea89 // sdot v9.4s, v20.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea8d // sdot v13.4s, v20.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea91 // sdot v17.4s, v20.16b, v2.4b[3]\n"
+ "ldr q20, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa0eaaa // sdot v10.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaae // sdot v14.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab2 // sdot v18.4s, v21.16b, v2.4b[3]\n"
+ ".inst 0x4fa0ea8b // sdot v11.4s, v20.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea8f // sdot v15.4s, v20.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea93 // sdot v19.4s, v20.16b, v2.4b[3]\n"
"63:" // Height 3: Multiply loop: Main loop skip
"cbz x27, 68f\n"
"cmp x27, #0x4\n"
"blt 65f\n"
"64:" // Height 3: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s24, [x26], #0x4\n"
+ "ldr s23, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr q6, [x9, #0x0]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr s22, [x24], #0x4\n"
+ "ldr q21, [x9, #0x0]\n"
+ ".inst 0x4f98e2a8 // sdot v8.4s, v21.16b, v24.4b[0]\n"
+ ".inst 0x4f97e2ac // sdot v12.4s, v21.16b, v23.4b[0]\n"
+ "ldr q20, [x9, #0x10]\n"
+ ".inst 0x4f96e2b0 // sdot v16.4s, v21.16b, v22.4b[0]\n"
+ "ldr q21, [x9, #0x20]\n"
+ ".inst 0x4f98e289 // sdot v9.4s, v20.16b, v24.4b[0]\n"
+ ".inst 0x4f97e28d // sdot v13.4s, v20.16b, v23.4b[0]\n"
+ ".inst 0x4f96e291 // sdot v17.4s, v20.16b, v22.4b[0]\n"
+ "ldr q20, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f98e2aa // sdot v10.4s, v21.16b, v24.4b[0]\n"
+ ".inst 0x4f97e2ae // sdot v14.4s, v21.16b, v23.4b[0]\n"
+ ".inst 0x4f96e2b2 // sdot v18.4s, v21.16b, v22.4b[0]\n"
+ ".inst 0x4f98e28b // sdot v11.4s, v20.16b, v24.4b[0]\n"
+ ".inst 0x4f97e28f // sdot v15.4s, v20.16b, v23.4b[0]\n"
+ ".inst 0x4f96e293 // sdot v19.4s, v20.16b, v22.4b[0]\n"
"bge 64b\n"
"65:" // Height 3: Multiply loop: Skip odd blocks
"cbz x27, 68f\n"
@@ -1022,51 +1021,51 @@ void a64_hybrid_s8qs_dot_6x16 (
"ldr b1, [x25, #0x0]\n"
"ldr b2, [x24, #0x0]\n"
"67:" // Height 3: Multiply loop: Ragged operand read: Done
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q21, [x9, #0x0]\n"
+ "ldr q20, [x9, #0x10]\n"
+ ".inst 0x4f80e2a8 // sdot v8.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ac // sdot v12.4s, v21.16b, v1.4b[0]\n"
+ ".inst 0x4f82e2b0 // sdot v16.4s, v21.16b, v2.4b[0]\n"
+ "ldr q21, [x9, #0x20]\n"
+ ".inst 0x4f80e289 // sdot v9.4s, v20.16b, v0.4b[0]\n"
+ ".inst 0x4f81e28d // sdot v13.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e291 // sdot v17.4s, v20.16b, v2.4b[0]\n"
+ "ldr q20, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n"
+ ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n"
+ ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n"
+ ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n"
"68:" // Height 3: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
"cmp x28, x20\n"
"bne 58b\n"
- "ldr q0, [x14, #0x0]\n"
- "ldr q1, [x14, #0x10]\n"
- "add v8.4s, v8.4s, v0.4s\n"
- "add v9.4s, v9.4s, v1.4s\n"
- "ldr q2, [x14, #0x20]\n"
- "ldr q3, [x14, #0x30]\n"
- "add v10.4s, v10.4s, v2.4s\n"
- "add v11.4s, v11.4s, v3.4s\n"
+ "ldr q23, [x14, #0x0]\n"
+ "ldr q22, [x14, #0x10]\n"
+ "add v8.4s, v8.4s, v23.4s\n"
+ "add v9.4s, v9.4s, v22.4s\n"
+ "ldr q21, [x14, #0x20]\n"
+ "ldr q20, [x14, #0x30]\n"
+ "add v10.4s, v10.4s, v21.4s\n"
+ "add v11.4s, v11.4s, v20.4s\n"
"ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "add x24, x11, x20\n"
- "add x23, x24, x20\n"
+ "add x25, x11, x20\n"
+ "add x24, x25, x20\n"
"prfm pstl1keep, [x11, #0x0]\n"
+ "prfm pstl1keep, [x25, #0x0]\n"
"prfm pstl1keep, [x24, #0x0]\n"
- "prfm pstl1keep, [x23, #0x0]\n"
- "add v12.4s, v12.4s, v0.4s\n"
- "add v13.4s, v13.4s, v1.4s\n"
- "add v14.4s, v14.4s, v2.4s\n"
- "add v15.4s, v15.4s, v3.4s\n"
+ "add v12.4s, v12.4s, v23.4s\n"
+ "add v13.4s, v13.4s, v22.4s\n"
+ "add v14.4s, v14.4s, v21.4s\n"
+ "add v15.4s, v15.4s, v20.4s\n"
"add x14, x14, #0x40\n"
- "add v16.4s, v16.4s, v0.4s\n"
- "add v17.4s, v17.4s, v1.4s\n"
- "add v18.4s, v18.4s, v2.4s\n"
- "add v19.4s, v19.4s, v3.4s\n"
+ "add v16.4s, v16.4s, v23.4s\n"
+ "add v17.4s, v17.4s, v22.4s\n"
+ "add v18.4s, v18.4s, v21.4s\n"
+ "add v19.4s, v19.4s, v20.4s\n"
"tbz %x[flags], #4, 69f\n"
"ldr q0, [x12, #0x0]\n"
"ldr q4, [x13, #0x0]\n"
@@ -1080,10 +1079,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x13, x13, #0x40\n"
"b 70f\n"
"69:" // Height 3: per layer parameters
- "add x25, %x[qp], %[per_layer_right_shift]\n"
- "ld1r { v0.4s }, [x25]\n"
- "add x25, %x[qp], %[per_layer_mul]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v4.4s }, [x20]\n"
"mov v1.16b, v0.16b\n"
"mov v5.16b, v4.16b\n"
"mov v2.16b, v0.16b\n"
@@ -1104,55 +1103,55 @@ void a64_hybrid_s8qs_dot_6x16 (
"sqrdmulh v18.4s, v18.4s, v6.4s\n"
"sqrdmulh v19.4s, v19.4s, v7.4s\n"
"tbz %x[flags], #5, 71f\n"
- "and v4.16b, v8.16b, v0.16b\n"
- "and v5.16b, v9.16b, v1.16b\n"
- "and v6.16b, v10.16b, v2.16b\n"
- "and v7.16b, v11.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v8.4s, v8.4s, v4.4s\n"
- "sqadd v9.4s, v9.4s, v5.4s\n"
- "sqadd v10.4s, v10.4s, v6.4s\n"
- "sqadd v11.4s, v11.4s, v7.4s\n"
- "and v4.16b, v12.16b, v0.16b\n"
- "and v5.16b, v13.16b, v1.16b\n"
- "and v6.16b, v14.16b, v2.16b\n"
- "and v7.16b, v15.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v12.4s, v12.4s, v4.4s\n"
- "sqadd v13.4s, v13.4s, v5.4s\n"
- "sqadd v14.4s, v14.4s, v6.4s\n"
- "sqadd v15.4s, v15.4s, v7.4s\n"
- "and v4.16b, v16.16b, v0.16b\n"
- "and v5.16b, v17.16b, v1.16b\n"
- "and v6.16b, v18.16b, v2.16b\n"
- "and v7.16b, v19.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v16.4s, v16.4s, v4.4s\n"
- "sqadd v17.4s, v17.4s, v5.4s\n"
- "sqadd v18.4s, v18.4s, v6.4s\n"
- "sqadd v19.4s, v19.4s, v7.4s\n"
+ "and v23.16b, v8.16b, v0.16b\n"
+ "and v22.16b, v9.16b, v1.16b\n"
+ "and v21.16b, v10.16b, v2.16b\n"
+ "and v20.16b, v11.16b, v3.16b\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "sqadd v8.4s, v8.4s, v23.4s\n"
+ "sqadd v9.4s, v9.4s, v22.4s\n"
+ "sqadd v10.4s, v10.4s, v21.4s\n"
+ "sqadd v11.4s, v11.4s, v20.4s\n"
+ "and v23.16b, v12.16b, v0.16b\n"
+ "and v22.16b, v13.16b, v1.16b\n"
+ "and v21.16b, v14.16b, v2.16b\n"
+ "and v20.16b, v15.16b, v3.16b\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "sqadd v12.4s, v12.4s, v23.4s\n"
+ "sqadd v13.4s, v13.4s, v22.4s\n"
+ "sqadd v14.4s, v14.4s, v21.4s\n"
+ "sqadd v15.4s, v15.4s, v20.4s\n"
+ "and v23.16b, v16.16b, v0.16b\n"
+ "and v22.16b, v17.16b, v1.16b\n"
+ "and v21.16b, v18.16b, v2.16b\n"
+ "and v20.16b, v19.16b, v3.16b\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "sqadd v16.4s, v16.4s, v23.4s\n"
+ "sqadd v17.4s, v17.4s, v22.4s\n"
+ "sqadd v18.4s, v18.4s, v21.4s\n"
+ "sqadd v19.4s, v19.4s, v20.4s\n"
"71:" // Height 3: no shift correction
- "add x25, %x[qp], %[c_offset]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v22.4s }, [x20]\n"
"srshl v8.4s, v8.4s, v0.4s\n"
"srshl v9.4s, v9.4s, v1.4s\n"
"srshl v10.4s, v10.4s, v2.4s\n"
"srshl v11.4s, v11.4s, v3.4s\n"
- "add x25, %x[qp], %[maxval]\n"
- "ld1r { v6.4s }, [x25]\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v21.4s }, [x20]\n"
"srshl v12.4s, v12.4s, v0.4s\n"
"srshl v13.4s, v13.4s, v1.4s\n"
- "add x25, %x[qp], %[minval]\n"
- "ld1r { v5.4s }, [x25]\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v20.4s }, [x20]\n"
"srshl v14.4s, v14.4s, v2.4s\n"
"srshl v15.4s, v15.4s, v3.4s\n"
"cmp x10, #0x10\n"
@@ -1160,132 +1159,132 @@ void a64_hybrid_s8qs_dot_6x16 (
"srshl v17.4s, v17.4s, v1.4s\n"
"srshl v18.4s, v18.4s, v2.4s\n"
"srshl v19.4s, v19.4s, v3.4s\n"
- "add v8.4s, v8.4s, v4.4s\n"
- "add v9.4s, v9.4s, v4.4s\n"
- "add v10.4s, v10.4s, v4.4s\n"
- "add v11.4s, v11.4s, v4.4s\n"
- "add v12.4s, v12.4s, v4.4s\n"
- "add v13.4s, v13.4s, v4.4s\n"
- "add v14.4s, v14.4s, v4.4s\n"
- "add v15.4s, v15.4s, v4.4s\n"
- "add v16.4s, v16.4s, v4.4s\n"
- "add v17.4s, v17.4s, v4.4s\n"
- "add v18.4s, v18.4s, v4.4s\n"
- "add v19.4s, v19.4s, v4.4s\n"
- "smin v8.4s, v8.4s, v6.4s\n"
- "smin v9.4s, v9.4s, v6.4s\n"
- "smin v10.4s, v10.4s, v6.4s\n"
- "smin v11.4s, v11.4s, v6.4s\n"
- "smin v12.4s, v12.4s, v6.4s\n"
- "smin v13.4s, v13.4s, v6.4s\n"
- "smin v14.4s, v14.4s, v6.4s\n"
- "smin v15.4s, v15.4s, v6.4s\n"
- "smin v16.4s, v16.4s, v6.4s\n"
- "smin v17.4s, v17.4s, v6.4s\n"
- "smin v18.4s, v18.4s, v6.4s\n"
- "smin v19.4s, v19.4s, v6.4s\n"
- "smax v8.4s, v8.4s, v5.4s\n"
- "smax v9.4s, v9.4s, v5.4s\n"
- "smax v10.4s, v10.4s, v5.4s\n"
- "smax v11.4s, v11.4s, v5.4s\n"
- "smax v12.4s, v12.4s, v5.4s\n"
- "smax v13.4s, v13.4s, v5.4s\n"
- "smax v14.4s, v14.4s, v5.4s\n"
- "smax v15.4s, v15.4s, v5.4s\n"
- "smax v16.4s, v16.4s, v5.4s\n"
- "smax v17.4s, v17.4s, v5.4s\n"
- "smax v18.4s, v18.4s, v5.4s\n"
- "smax v19.4s, v19.4s, v5.4s\n"
+ "add v8.4s, v8.4s, v22.4s\n"
+ "add v9.4s, v9.4s, v22.4s\n"
+ "add v10.4s, v10.4s, v22.4s\n"
+ "add v11.4s, v11.4s, v22.4s\n"
+ "add v12.4s, v12.4s, v22.4s\n"
+ "add v13.4s, v13.4s, v22.4s\n"
+ "add v14.4s, v14.4s, v22.4s\n"
+ "add v15.4s, v15.4s, v22.4s\n"
+ "add v16.4s, v16.4s, v22.4s\n"
+ "add v17.4s, v17.4s, v22.4s\n"
+ "add v18.4s, v18.4s, v22.4s\n"
+ "add v19.4s, v19.4s, v22.4s\n"
+ "smin v8.4s, v8.4s, v21.4s\n"
+ "smin v9.4s, v9.4s, v21.4s\n"
+ "smin v10.4s, v10.4s, v21.4s\n"
+ "smin v11.4s, v11.4s, v21.4s\n"
+ "smin v12.4s, v12.4s, v21.4s\n"
+ "smin v13.4s, v13.4s, v21.4s\n"
+ "smin v14.4s, v14.4s, v21.4s\n"
+ "smin v15.4s, v15.4s, v21.4s\n"
+ "smin v16.4s, v16.4s, v21.4s\n"
+ "smin v17.4s, v17.4s, v21.4s\n"
+ "smin v18.4s, v18.4s, v21.4s\n"
+ "smin v19.4s, v19.4s, v21.4s\n"
+ "smax v8.4s, v8.4s, v20.4s\n"
+ "smax v9.4s, v9.4s, v20.4s\n"
+ "smax v10.4s, v10.4s, v20.4s\n"
+ "smax v11.4s, v11.4s, v20.4s\n"
+ "smax v12.4s, v12.4s, v20.4s\n"
+ "smax v13.4s, v13.4s, v20.4s\n"
+ "smax v14.4s, v14.4s, v20.4s\n"
+ "smax v15.4s, v15.4s, v20.4s\n"
+ "smax v16.4s, v16.4s, v20.4s\n"
+ "smax v17.4s, v17.4s, v20.4s\n"
+ "smax v18.4s, v18.4s, v20.4s\n"
+ "smax v19.4s, v19.4s, v20.4s\n"
"uzp1 v8.8h, v8.8h, v9.8h\n"
- "uzp1 v9.8h, v10.8h, v11.8h\n"
+ "uzp1 v21.8h, v10.8h, v11.8h\n"
"uzp1 v12.8h, v12.8h, v13.8h\n"
- "uzp1 v13.8h, v14.8h, v15.8h\n"
+ "uzp1 v20.8h, v14.8h, v15.8h\n"
"uzp1 v16.8h, v16.8h, v17.8h\n"
"uzp1 v17.8h, v18.8h, v19.8h\n"
- "uzp1 v8.16b, v8.16b, v9.16b\n"
- "uzp1 v12.16b, v12.16b, v13.16b\n"
+ "uzp1 v8.16b, v8.16b, v21.16b\n"
+ "uzp1 v12.16b, v12.16b, v20.16b\n"
"uzp1 v16.16b, v16.16b, v17.16b\n"
"bge 80f\n"
"tbz x10, #3, 75f\n"
"str d8, [x11], #0x8\n"
- "str d12, [x24], #0x8\n"
- "str d16, [x23], #0x8\n"
+ "str d12, [x25], #0x8\n"
+ "str d16, [x24], #0x8\n"
"tbz x10, #2, 73f\n"
"st1 { v8.s }[2], [x11], #0x4\n"
- "st1 { v12.s }[2], [x24], #0x4\n"
- "st1 { v16.s }[2], [x23], #0x4\n"
+ "st1 { v12.s }[2], [x25], #0x4\n"
+ "st1 { v16.s }[2], [x24], #0x4\n"
"tbz x10, #1, 72f\n"
"st1 { v8.h }[6], [x11], #0x2\n"
- "st1 { v12.h }[6], [x24], #0x2\n"
- "st1 { v16.h }[6], [x23], #0x2\n"
+ "st1 { v12.h }[6], [x25], #0x2\n"
+ "st1 { v16.h }[6], [x24], #0x2\n"
"tbz x10, #0, 79f\n"
"st1 { v8.b }[14], [x11]\n"
- "st1 { v12.b }[14], [x24]\n"
- "st1 { v16.b }[14], [x23]\n"
+ "st1 { v12.b }[14], [x25]\n"
+ "st1 { v16.b }[14], [x24]\n"
"b 79f\n"
"72:" // Height 3: Partial direct writeback: partial_1_12
"tbz x10, #0, 79f\n"
"st1 { v8.b }[12], [x11]\n"
- "st1 { v12.b }[12], [x24]\n"
- "st1 { v16.b }[12], [x23]\n"
+ "st1 { v12.b }[12], [x25]\n"
+ "st1 { v16.b }[12], [x24]\n"
"b 79f\n"
"73:" // Height 3: Partial direct writeback: partial_2_8
"tbz x10, #1, 74f\n"
"st1 { v8.h }[4], [x11], #0x2\n"
- "st1 { v12.h }[4], [x24], #0x2\n"
- "st1 { v16.h }[4], [x23], #0x2\n"
+ "st1 { v12.h }[4], [x25], #0x2\n"
+ "st1 { v16.h }[4], [x24], #0x2\n"
"tbz x10, #0, 79f\n"
"st1 { v8.b }[10], [x11]\n"
- "st1 { v12.b }[10], [x24]\n"
- "st1 { v16.b }[10], [x23]\n"
+ "st1 { v12.b }[10], [x25]\n"
+ "st1 { v16.b }[10], [x24]\n"
"b 79f\n"
"74:" // Height 3: Partial direct writeback: partial_1_8
"tbz x10, #0, 79f\n"
"st1 { v8.b }[8], [x11]\n"
- "st1 { v12.b }[8], [x24]\n"
- "st1 { v16.b }[8], [x23]\n"
+ "st1 { v12.b }[8], [x25]\n"
+ "st1 { v16.b }[8], [x24]\n"
"b 79f\n"
"75:" // Height 3: Partial direct writeback: partial_4_0
"tbz x10, #2, 77f\n"
"str s8, [x11], #0x4\n"
- "str s12, [x24], #0x4\n"
- "str s16, [x23], #0x4\n"
+ "str s12, [x25], #0x4\n"
+ "str s16, [x24], #0x4\n"
"tbz x10, #1, 76f\n"
"st1 { v8.h }[2], [x11], #0x2\n"
- "st1 { v12.h }[2], [x24], #0x2\n"
- "st1 { v16.h }[2], [x23], #0x2\n"
+ "st1 { v12.h }[2], [x25], #0x2\n"
+ "st1 { v16.h }[2], [x24], #0x2\n"
"tbz x10, #0, 79f\n"
"st1 { v8.b }[6], [x11]\n"
- "st1 { v12.b }[6], [x24]\n"
- "st1 { v16.b }[6], [x23]\n"
+ "st1 { v12.b }[6], [x25]\n"
+ "st1 { v16.b }[6], [x24]\n"
"b 79f\n"
"76:" // Height 3: Partial direct writeback: partial_1_4
"tbz x10, #0, 79f\n"
"st1 { v8.b }[4], [x11]\n"
- "st1 { v12.b }[4], [x24]\n"
- "st1 { v16.b }[4], [x23]\n"
+ "st1 { v12.b }[4], [x25]\n"
+ "st1 { v16.b }[4], [x24]\n"
"b 79f\n"
"77:" // Height 3: Partial direct writeback: partial_2_0
"tbz x10, #1, 78f\n"
"str h8, [x11], #0x2\n"
- "str h12, [x24], #0x2\n"
- "str h16, [x23], #0x2\n"
+ "str h12, [x25], #0x2\n"
+ "str h16, [x24], #0x2\n"
"tbz x10, #0, 79f\n"
"st1 { v8.b }[2], [x11]\n"
- "st1 { v12.b }[2], [x24]\n"
- "st1 { v16.b }[2], [x23]\n"
+ "st1 { v12.b }[2], [x25]\n"
+ "st1 { v16.b }[2], [x24]\n"
"b 79f\n"
"78:" // Height 3: Partial direct writeback: partial_1_0
"str b8, [x11, #0x0]\n"
- "str b12, [x24, #0x0]\n"
- "str b16, [x23, #0x0]\n"
+ "str b12, [x25, #0x0]\n"
+ "str b16, [x24, #0x0]\n"
"79:" // Height 3: Partial direct writeback: Done
"b 81f\n"
"80:" // Height 3: Full writeback
"str q8, [x11, #0x0]\n"
"add x11, x11, #0x10\n"
- "str q12, [x24, #0x0]\n"
- "str q16, [x23, #0x0]\n"
+ "str q12, [x25, #0x0]\n"
+ "str q16, [x24, #0x0]\n"
"81:" // Height 3: Writeback done
"subs x10, x10, #0x10\n"
"bgt 56b\n"
@@ -1319,14 +1318,14 @@ void a64_hybrid_s8qs_dot_6x16 (
"85:" // Height 4: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, 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, 86f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
"cbnz x28, 87f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -1336,9 +1335,9 @@ void a64_hybrid_s8qs_dot_6x16 (
"b 87f\n"
"86:" // Height 4: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
- "add x23, x24, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
"87:" // Height 4: input setup done
"cmp x27, #0x10\n"
"blt 90f\n"
@@ -1357,7 +1356,7 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x26, x26, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q25, [x9, #0x20]\n"
"add x25, x25, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1365,85 +1364,85 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x23, x23, #0x10\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q24, [x9, #0x30]\n"
"cmp x27, #0x20\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n"
"prfm pldl1keep, [x26, #0x80]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
+ ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n"
+ "ldr q25, [x9, #0x40]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n"
"prfm pldl1keep, [x23, #0x80]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n"
+ "ldr q24, [x9, #0x50]\n"
+ ".inst 0x4fa0e328 // sdot v8.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32c // sdot v12.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e330 // sdot v16.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e334 // sdot v20.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x9, #0x60]\n"
+ ".inst 0x4fa0e309 // sdot v9.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30d // sdot v13.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e311 // sdot v17.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e315 // sdot v21.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x9, #0x70]\n"
+ ".inst 0x4fa0e32a // sdot v10.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32e // sdot v14.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e332 // sdot v18.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e336 // sdot v22.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x9, #0x80]\n"
+ ".inst 0x4fa0e30b // sdot v11.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30f // sdot v15.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e313 // sdot v19.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e317 // sdot v23.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x9, #0x90]\n"
+ ".inst 0x4f80eb28 // sdot v8.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2c // sdot v12.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb30 // sdot v16.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb34 // sdot v20.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x9, #0xa0]\n"
+ ".inst 0x4f80eb09 // sdot v9.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0d // sdot v13.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb11 // sdot v17.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb15 // sdot v21.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x9, #0xb0]\n"
+ ".inst 0x4f80eb2a // sdot v10.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2e // sdot v14.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb32 // sdot v18.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb36 // sdot v22.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x9, #0xc0]\n"
+ ".inst 0x4f80eb0b // sdot v11.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0f // sdot v15.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb13 // sdot v19.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb17 // sdot v23.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x9, #0xd0]\n"
+ ".inst 0x4fa0eb28 // sdot v8.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2c // sdot v12.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb30 // sdot v16.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb34 // sdot v20.4s, v25.16b, v3.4b[3]\n"
+ "ldr q25, [x9, #0xe0]\n"
+ ".inst 0x4fa0eb09 // sdot v9.4s, v24.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb0d // sdot v13.4s, v24.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb11 // sdot v17.4s, v24.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb15 // sdot v21.4s, v24.16b, v3.4b[3]\n"
+ "ldr q24, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
+ ".inst 0x4fa0eb2a // sdot v10.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2e // sdot v14.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb32 // sdot v18.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb36 // sdot v22.4s, v25.16b, v3.4b[3]\n"
"ldr q6, [x9, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0eb0b // sdot v11.4s, v24.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1eb0f // sdot v15.4s, v24.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb13 // sdot v19.4s, v24.16b, v2.4b[3]\n"
"ldr q2, [x24, #0x0]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
+ ".inst 0x4fa3eb17 // sdot v23.4s, v24.16b, v3.4b[3]\n"
"ldr q3, [x23, #0x0]\n"
"ldr q7, [x9, #0x10]\n"
"bge 88b\n"
@@ -1454,7 +1453,7 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x25, x25, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q25, [x9, #0x20]\n"
"add x24, x24, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1462,112 +1461,112 @@ void a64_hybrid_s8qs_dot_6x16 (
"sub x27, x27, #0x10\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q24, [x9, #0x30]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n"
"prfm pldl1keep, [x25, #0x80]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
+ ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n"
+ "ldr q25, [x9, #0x40]\n"
"prfm pldl1keep, [x23, #0x80]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n"
+ "ldr q24, [x9, #0x50]\n"
+ ".inst 0x4fa0e328 // sdot v8.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32c // sdot v12.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e330 // sdot v16.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e334 // sdot v20.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x9, #0x60]\n"
+ ".inst 0x4fa0e309 // sdot v9.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30d // sdot v13.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e311 // sdot v17.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e315 // sdot v21.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x9, #0x70]\n"
+ ".inst 0x4fa0e32a // sdot v10.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32e // sdot v14.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e332 // sdot v18.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e336 // sdot v22.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x9, #0x80]\n"
+ ".inst 0x4fa0e30b // sdot v11.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30f // sdot v15.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e313 // sdot v19.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e317 // sdot v23.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x9, #0x90]\n"
+ ".inst 0x4f80eb28 // sdot v8.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2c // sdot v12.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb30 // sdot v16.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb34 // sdot v20.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x9, #0xa0]\n"
+ ".inst 0x4f80eb09 // sdot v9.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0d // sdot v13.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb11 // sdot v17.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb15 // sdot v21.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x9, #0xb0]\n"
+ ".inst 0x4f80eb2a // sdot v10.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2e // sdot v14.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb32 // sdot v18.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb36 // sdot v22.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x9, #0xc0]\n"
+ ".inst 0x4f80eb0b // sdot v11.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0f // sdot v15.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb13 // sdot v19.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb17 // sdot v23.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x9, #0xd0]\n"
+ ".inst 0x4fa0eb28 // sdot v8.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2c // sdot v12.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb30 // sdot v16.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb34 // sdot v20.4s, v25.16b, v3.4b[3]\n"
+ "ldr q25, [x9, #0xe0]\n"
+ ".inst 0x4fa0eb09 // sdot v9.4s, v24.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb0d // sdot v13.4s, v24.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb11 // sdot v17.4s, v24.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb15 // sdot v21.4s, v24.16b, v3.4b[3]\n"
+ "ldr q24, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
+ ".inst 0x4fa0eb2a // sdot v10.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2e // sdot v14.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb32 // sdot v18.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb36 // sdot v22.4s, v25.16b, v3.4b[3]\n"
+ ".inst 0x4fa0eb0b // sdot v11.4s, v24.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb0f // sdot v15.4s, v24.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb13 // sdot v19.4s, v24.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb17 // sdot v23.4s, v24.16b, v3.4b[3]\n"
"90:" // Height 4: Multiply loop: Main loop skip
"cbz x27, 95f\n"
"cmp x27, #0x4\n"
"blt 92f\n"
"91:" // Height 4: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s29, [x26], #0x4\n"
+ "ldr s28, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr s3, [x23], #0x4\n"
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr s27, [x24], #0x4\n"
+ "ldr s26, [x23], #0x4\n"
+ "ldr q25, [x9, #0x0]\n"
+ "ldr q24, [x9, #0x10]\n"
+ ".inst 0x4f9de328 // sdot v8.4s, v25.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce32c // sdot v12.4s, v25.16b, v28.4b[0]\n"
+ ".inst 0x4f9be330 // sdot v16.4s, v25.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae334 // sdot v20.4s, v25.16b, v26.4b[0]\n"
+ "ldr q25, [x9, #0x20]\n"
+ ".inst 0x4f9de309 // sdot v9.4s, v24.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce30d // sdot v13.4s, v24.16b, v28.4b[0]\n"
+ ".inst 0x4f9be311 // sdot v17.4s, v24.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae315 // sdot v21.4s, v24.16b, v26.4b[0]\n"
+ "ldr q24, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f9de32a // sdot v10.4s, v25.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce32e // sdot v14.4s, v25.16b, v28.4b[0]\n"
+ ".inst 0x4f9be332 // sdot v18.4s, v25.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae336 // sdot v22.4s, v25.16b, v26.4b[0]\n"
+ ".inst 0x4f9de30b // sdot v11.4s, v24.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce30f // sdot v15.4s, v24.16b, v28.4b[0]\n"
+ ".inst 0x4f9be313 // sdot v19.4s, v24.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae317 // sdot v23.4s, v24.16b, v26.4b[0]\n"
"bge 91b\n"
"92:" // Height 4: Multiply loop: Skip odd blocks
"cbz x27, 95f\n"
@@ -1588,61 +1587,61 @@ void a64_hybrid_s8qs_dot_6x16 (
"ldr b2, [x24, #0x0]\n"
"ldr b3, [x23, #0x0]\n"
"94:" // Height 4: Multiply loop: Ragged operand read: Done
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q25, [x9, #0x0]\n"
+ "ldr q24, [x9, #0x10]\n"
+ ".inst 0x4f80e328 // sdot v8.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32c // sdot v12.4s, v25.16b, v1.4b[0]\n"
+ ".inst 0x4f82e330 // sdot v16.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e334 // sdot v20.4s, v25.16b, v3.4b[0]\n"
+ "ldr q25, [x9, #0x20]\n"
+ ".inst 0x4f80e309 // sdot v9.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30d // sdot v13.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f82e311 // sdot v17.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e315 // sdot v21.4s, v24.16b, v3.4b[0]\n"
+ "ldr q24, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n"
+ ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n"
+ ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n"
"95:" // Height 4: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
"cmp x28, x20\n"
"bne 85b\n"
- "ldr q0, [x14, #0x0]\n"
- "ldr q1, [x14, #0x10]\n"
- "add v8.4s, v8.4s, v0.4s\n"
- "add v9.4s, v9.4s, v1.4s\n"
- "ldr q2, [x14, #0x20]\n"
- "ldr q3, [x14, #0x30]\n"
- "add v10.4s, v10.4s, v2.4s\n"
- "add v11.4s, v11.4s, v3.4s\n"
+ "ldr q27, [x14, #0x0]\n"
+ "ldr q26, [x14, #0x10]\n"
+ "add v8.4s, v8.4s, v27.4s\n"
+ "add v9.4s, v9.4s, v26.4s\n"
+ "ldr q25, [x14, #0x20]\n"
+ "ldr q24, [x14, #0x30]\n"
+ "add v10.4s, v10.4s, v25.4s\n"
+ "add v11.4s, v11.4s, v24.4s\n"
"ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "add x24, x11, x20\n"
- "add x23, x24, x20\n"
+ "add x25, x11, x20\n"
+ "add x24, x25, x20\n"
"prfm pstl1keep, [x11, #0x0]\n"
- "add x22, x23, x20\n"
+ "add x23, x24, x20\n"
+ "prfm pstl1keep, [x25, #0x0]\n"
"prfm pstl1keep, [x24, #0x0]\n"
+ "add v12.4s, v12.4s, v27.4s\n"
"prfm pstl1keep, [x23, #0x0]\n"
- "add v12.4s, v12.4s, v0.4s\n"
- "prfm pstl1keep, [x22, #0x0]\n"
- "add v13.4s, v13.4s, v1.4s\n"
- "add v14.4s, v14.4s, v2.4s\n"
+ "add v13.4s, v13.4s, v26.4s\n"
+ "add v14.4s, v14.4s, v25.4s\n"
"add x14, x14, #0x40\n"
- "add v15.4s, v15.4s, v3.4s\n"
- "add v16.4s, v16.4s, v0.4s\n"
- "add v17.4s, v17.4s, v1.4s\n"
- "add v18.4s, v18.4s, v2.4s\n"
- "add v19.4s, v19.4s, v3.4s\n"
- "add v20.4s, v20.4s, v0.4s\n"
- "add v21.4s, v21.4s, v1.4s\n"
- "add v22.4s, v22.4s, v2.4s\n"
- "add v23.4s, v23.4s, v3.4s\n"
+ "add v15.4s, v15.4s, v24.4s\n"
+ "add v16.4s, v16.4s, v27.4s\n"
+ "add v17.4s, v17.4s, v26.4s\n"
+ "add v18.4s, v18.4s, v25.4s\n"
+ "add v19.4s, v19.4s, v24.4s\n"
+ "add v20.4s, v20.4s, v27.4s\n"
+ "add v21.4s, v21.4s, v26.4s\n"
+ "add v22.4s, v22.4s, v25.4s\n"
+ "add v23.4s, v23.4s, v24.4s\n"
"tbz %x[flags], #4, 96f\n"
"ldr q0, [x12, #0x0]\n"
"ldr q4, [x13, #0x0]\n"
@@ -1656,10 +1655,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x13, x13, #0x40\n"
"b 97f\n"
"96:" // Height 4: per layer parameters
- "add x25, %x[qp], %[per_layer_right_shift]\n"
- "ld1r { v0.4s }, [x25]\n"
- "add x25, %x[qp], %[per_layer_mul]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v4.4s }, [x20]\n"
"mov v1.16b, v0.16b\n"
"mov v5.16b, v4.16b\n"
"mov v2.16b, v0.16b\n"
@@ -1684,67 +1683,67 @@ void a64_hybrid_s8qs_dot_6x16 (
"sqrdmulh v22.4s, v22.4s, v6.4s\n"
"sqrdmulh v23.4s, v23.4s, v7.4s\n"
"tbz %x[flags], #5, 98f\n"
- "and v4.16b, v8.16b, v0.16b\n"
- "and v5.16b, v9.16b, v1.16b\n"
- "and v6.16b, v10.16b, v2.16b\n"
- "and v7.16b, v11.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v8.4s, v8.4s, v4.4s\n"
- "sqadd v9.4s, v9.4s, v5.4s\n"
- "sqadd v10.4s, v10.4s, v6.4s\n"
- "sqadd v11.4s, v11.4s, v7.4s\n"
- "and v4.16b, v12.16b, v0.16b\n"
- "and v5.16b, v13.16b, v1.16b\n"
- "and v6.16b, v14.16b, v2.16b\n"
- "and v7.16b, v15.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v12.4s, v12.4s, v4.4s\n"
- "sqadd v13.4s, v13.4s, v5.4s\n"
- "sqadd v14.4s, v14.4s, v6.4s\n"
- "sqadd v15.4s, v15.4s, v7.4s\n"
- "and v4.16b, v16.16b, v0.16b\n"
- "and v5.16b, v17.16b, v1.16b\n"
- "and v6.16b, v18.16b, v2.16b\n"
- "and v7.16b, v19.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v16.4s, v16.4s, v4.4s\n"
- "sqadd v17.4s, v17.4s, v5.4s\n"
- "sqadd v18.4s, v18.4s, v6.4s\n"
- "sqadd v19.4s, v19.4s, v7.4s\n"
- "and v4.16b, v20.16b, v0.16b\n"
- "and v5.16b, v21.16b, v1.16b\n"
- "and v6.16b, v22.16b, v2.16b\n"
- "and v7.16b, v23.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v20.4s, v20.4s, v4.4s\n"
- "sqadd v21.4s, v21.4s, v5.4s\n"
- "sqadd v22.4s, v22.4s, v6.4s\n"
- "sqadd v23.4s, v23.4s, v7.4s\n"
+ "and v27.16b, v8.16b, v0.16b\n"
+ "and v26.16b, v9.16b, v1.16b\n"
+ "and v25.16b, v10.16b, v2.16b\n"
+ "and v24.16b, v11.16b, v3.16b\n"
+ "sshr v27.4s, v27.4s, #0x1f\n"
+ "sshr v26.4s, v26.4s, #0x1f\n"
+ "sshr v25.4s, v25.4s, #0x1f\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sqadd v8.4s, v8.4s, v27.4s\n"
+ "sqadd v9.4s, v9.4s, v26.4s\n"
+ "sqadd v10.4s, v10.4s, v25.4s\n"
+ "sqadd v11.4s, v11.4s, v24.4s\n"
+ "and v27.16b, v12.16b, v0.16b\n"
+ "and v26.16b, v13.16b, v1.16b\n"
+ "and v25.16b, v14.16b, v2.16b\n"
+ "and v24.16b, v15.16b, v3.16b\n"
+ "sshr v27.4s, v27.4s, #0x1f\n"
+ "sshr v26.4s, v26.4s, #0x1f\n"
+ "sshr v25.4s, v25.4s, #0x1f\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sqadd v12.4s, v12.4s, v27.4s\n"
+ "sqadd v13.4s, v13.4s, v26.4s\n"
+ "sqadd v14.4s, v14.4s, v25.4s\n"
+ "sqadd v15.4s, v15.4s, v24.4s\n"
+ "and v27.16b, v16.16b, v0.16b\n"
+ "and v26.16b, v17.16b, v1.16b\n"
+ "and v25.16b, v18.16b, v2.16b\n"
+ "and v24.16b, v19.16b, v3.16b\n"
+ "sshr v27.4s, v27.4s, #0x1f\n"
+ "sshr v26.4s, v26.4s, #0x1f\n"
+ "sshr v25.4s, v25.4s, #0x1f\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sqadd v16.4s, v16.4s, v27.4s\n"
+ "sqadd v17.4s, v17.4s, v26.4s\n"
+ "sqadd v18.4s, v18.4s, v25.4s\n"
+ "sqadd v19.4s, v19.4s, v24.4s\n"
+ "and v27.16b, v20.16b, v0.16b\n"
+ "and v26.16b, v21.16b, v1.16b\n"
+ "and v25.16b, v22.16b, v2.16b\n"
+ "and v24.16b, v23.16b, v3.16b\n"
+ "sshr v27.4s, v27.4s, #0x1f\n"
+ "sshr v26.4s, v26.4s, #0x1f\n"
+ "sshr v25.4s, v25.4s, #0x1f\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v27.4s\n"
+ "sqadd v21.4s, v21.4s, v26.4s\n"
+ "sqadd v22.4s, v22.4s, v25.4s\n"
+ "sqadd v23.4s, v23.4s, v24.4s\n"
"98:" // Height 4: no shift correction
- "add x25, %x[qp], %[c_offset]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v26.4s }, [x20]\n"
"srshl v8.4s, v8.4s, v0.4s\n"
"srshl v9.4s, v9.4s, v1.4s\n"
"srshl v10.4s, v10.4s, v2.4s\n"
"srshl v11.4s, v11.4s, v3.4s\n"
- "add x25, %x[qp], %[maxval]\n"
- "ld1r { v6.4s }, [x25]\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v25.4s }, [x20]\n"
"srshl v12.4s, v12.4s, v0.4s\n"
"srshl v13.4s, v13.4s, v1.4s\n"
- "add x25, %x[qp], %[minval]\n"
- "ld1r { v5.4s }, [x25]\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v24.4s }, [x20]\n"
"srshl v14.4s, v14.4s, v2.4s\n"
"srshl v15.4s, v15.4s, v3.4s\n"
"cmp x10, #0x10\n"
@@ -1756,163 +1755,163 @@ void a64_hybrid_s8qs_dot_6x16 (
"srshl v21.4s, v21.4s, v1.4s\n"
"srshl v22.4s, v22.4s, v2.4s\n"
"srshl v23.4s, v23.4s, v3.4s\n"
- "add v8.4s, v8.4s, v4.4s\n"
- "add v9.4s, v9.4s, v4.4s\n"
- "add v10.4s, v10.4s, v4.4s\n"
- "add v11.4s, v11.4s, v4.4s\n"
- "add v12.4s, v12.4s, v4.4s\n"
- "add v13.4s, v13.4s, v4.4s\n"
- "add v14.4s, v14.4s, v4.4s\n"
- "add v15.4s, v15.4s, v4.4s\n"
- "add v16.4s, v16.4s, v4.4s\n"
- "add v17.4s, v17.4s, v4.4s\n"
- "add v18.4s, v18.4s, v4.4s\n"
- "add v19.4s, v19.4s, v4.4s\n"
- "add v20.4s, v20.4s, v4.4s\n"
- "add v21.4s, v21.4s, v4.4s\n"
- "add v22.4s, v22.4s, v4.4s\n"
- "add v23.4s, v23.4s, v4.4s\n"
- "smin v8.4s, v8.4s, v6.4s\n"
- "smin v9.4s, v9.4s, v6.4s\n"
- "smin v10.4s, v10.4s, v6.4s\n"
- "smin v11.4s, v11.4s, v6.4s\n"
- "smin v12.4s, v12.4s, v6.4s\n"
- "smin v13.4s, v13.4s, v6.4s\n"
- "smin v14.4s, v14.4s, v6.4s\n"
- "smin v15.4s, v15.4s, v6.4s\n"
- "smin v16.4s, v16.4s, v6.4s\n"
- "smin v17.4s, v17.4s, v6.4s\n"
- "smin v18.4s, v18.4s, v6.4s\n"
- "smin v19.4s, v19.4s, v6.4s\n"
- "smin v20.4s, v20.4s, v6.4s\n"
- "smin v21.4s, v21.4s, v6.4s\n"
- "smin v22.4s, v22.4s, v6.4s\n"
- "smin v23.4s, v23.4s, v6.4s\n"
- "smax v8.4s, v8.4s, v5.4s\n"
- "smax v9.4s, v9.4s, v5.4s\n"
- "smax v10.4s, v10.4s, v5.4s\n"
- "smax v11.4s, v11.4s, v5.4s\n"
- "smax v12.4s, v12.4s, v5.4s\n"
- "smax v13.4s, v13.4s, v5.4s\n"
- "smax v14.4s, v14.4s, v5.4s\n"
- "smax v15.4s, v15.4s, v5.4s\n"
- "smax v16.4s, v16.4s, v5.4s\n"
- "smax v17.4s, v17.4s, v5.4s\n"
- "smax v18.4s, v18.4s, v5.4s\n"
- "smax v19.4s, v19.4s, v5.4s\n"
- "smax v20.4s, v20.4s, v5.4s\n"
- "smax v21.4s, v21.4s, v5.4s\n"
- "smax v22.4s, v22.4s, v5.4s\n"
- "smax v23.4s, v23.4s, v5.4s\n"
+ "add v8.4s, v8.4s, v26.4s\n"
+ "add v9.4s, v9.4s, v26.4s\n"
+ "add v10.4s, v10.4s, v26.4s\n"
+ "add v11.4s, v11.4s, v26.4s\n"
+ "add v12.4s, v12.4s, v26.4s\n"
+ "add v13.4s, v13.4s, v26.4s\n"
+ "add v14.4s, v14.4s, v26.4s\n"
+ "add v15.4s, v15.4s, v26.4s\n"
+ "add v16.4s, v16.4s, v26.4s\n"
+ "add v17.4s, v17.4s, v26.4s\n"
+ "add v18.4s, v18.4s, v26.4s\n"
+ "add v19.4s, v19.4s, v26.4s\n"
+ "add v20.4s, v20.4s, v26.4s\n"
+ "add v21.4s, v21.4s, v26.4s\n"
+ "add v22.4s, v22.4s, v26.4s\n"
+ "add v23.4s, v23.4s, v26.4s\n"
+ "smin v8.4s, v8.4s, v25.4s\n"
+ "smin v9.4s, v9.4s, v25.4s\n"
+ "smin v10.4s, v10.4s, v25.4s\n"
+ "smin v11.4s, v11.4s, v25.4s\n"
+ "smin v12.4s, v12.4s, v25.4s\n"
+ "smin v13.4s, v13.4s, v25.4s\n"
+ "smin v14.4s, v14.4s, v25.4s\n"
+ "smin v15.4s, v15.4s, v25.4s\n"
+ "smin v16.4s, v16.4s, v25.4s\n"
+ "smin v17.4s, v17.4s, v25.4s\n"
+ "smin v18.4s, v18.4s, v25.4s\n"
+ "smin v19.4s, v19.4s, v25.4s\n"
+ "smin v20.4s, v20.4s, v25.4s\n"
+ "smin v21.4s, v21.4s, v25.4s\n"
+ "smin v22.4s, v22.4s, v25.4s\n"
+ "smin v23.4s, v23.4s, v25.4s\n"
+ "smax v8.4s, v8.4s, v24.4s\n"
+ "smax v9.4s, v9.4s, v24.4s\n"
+ "smax v10.4s, v10.4s, v24.4s\n"
+ "smax v11.4s, v11.4s, v24.4s\n"
+ "smax v12.4s, v12.4s, v24.4s\n"
+ "smax v13.4s, v13.4s, v24.4s\n"
+ "smax v14.4s, v14.4s, v24.4s\n"
+ "smax v15.4s, v15.4s, v24.4s\n"
+ "smax v16.4s, v16.4s, v24.4s\n"
+ "smax v17.4s, v17.4s, v24.4s\n"
+ "smax v18.4s, v18.4s, v24.4s\n"
+ "smax v19.4s, v19.4s, v24.4s\n"
+ "smax v20.4s, v20.4s, v24.4s\n"
+ "smax v21.4s, v21.4s, v24.4s\n"
+ "smax v22.4s, v22.4s, v24.4s\n"
+ "smax v23.4s, v23.4s, v24.4s\n"
"uzp1 v8.8h, v8.8h, v9.8h\n"
- "uzp1 v9.8h, v10.8h, v11.8h\n"
+ "uzp1 v25.8h, v10.8h, v11.8h\n"
"uzp1 v12.8h, v12.8h, v13.8h\n"
- "uzp1 v13.8h, v14.8h, v15.8h\n"
+ "uzp1 v24.8h, v14.8h, v15.8h\n"
"uzp1 v16.8h, v16.8h, v17.8h\n"
- "uzp1 v17.8h, v18.8h, v19.8h\n"
+ "uzp1 v18.8h, v18.8h, v19.8h\n"
"uzp1 v20.8h, v20.8h, v21.8h\n"
- "uzp1 v21.8h, v22.8h, v23.8h\n"
- "uzp1 v8.16b, v8.16b, v9.16b\n"
- "uzp1 v12.16b, v12.16b, v13.16b\n"
- "uzp1 v16.16b, v16.16b, v17.16b\n"
- "uzp1 v20.16b, v20.16b, v21.16b\n"
+ "uzp1 v17.8h, v22.8h, v23.8h\n"
+ "uzp1 v8.16b, v8.16b, v25.16b\n"
+ "uzp1 v12.16b, v12.16b, v24.16b\n"
+ "uzp1 v16.16b, v16.16b, v18.16b\n"
+ "uzp1 v20.16b, v20.16b, v17.16b\n"
"bge 107f\n"
"tbz x10, #3, 102f\n"
"str d8, [x11], #0x8\n"
- "str d12, [x24], #0x8\n"
- "str d16, [x23], #0x8\n"
- "str d20, [x22], #0x8\n"
+ "str d12, [x25], #0x8\n"
+ "str d16, [x24], #0x8\n"
+ "str d20, [x23], #0x8\n"
"tbz x10, #2, 100f\n"
"st1 { v8.s }[2], [x11], #0x4\n"
- "st1 { v12.s }[2], [x24], #0x4\n"
- "st1 { v16.s }[2], [x23], #0x4\n"
- "st1 { v20.s }[2], [x22], #0x4\n"
+ "st1 { v12.s }[2], [x25], #0x4\n"
+ "st1 { v16.s }[2], [x24], #0x4\n"
+ "st1 { v20.s }[2], [x23], #0x4\n"
"tbz x10, #1, 99f\n"
"st1 { v8.h }[6], [x11], #0x2\n"
- "st1 { v12.h }[6], [x24], #0x2\n"
- "st1 { v16.h }[6], [x23], #0x2\n"
- "st1 { v20.h }[6], [x22], #0x2\n"
+ "st1 { v12.h }[6], [x25], #0x2\n"
+ "st1 { v16.h }[6], [x24], #0x2\n"
+ "st1 { v20.h }[6], [x23], #0x2\n"
"tbz x10, #0, 106f\n"
"st1 { v8.b }[14], [x11]\n"
- "st1 { v12.b }[14], [x24]\n"
- "st1 { v16.b }[14], [x23]\n"
- "st1 { v20.b }[14], [x22]\n"
+ "st1 { v12.b }[14], [x25]\n"
+ "st1 { v16.b }[14], [x24]\n"
+ "st1 { v20.b }[14], [x23]\n"
"b 106f\n"
"99:" // Height 4: Partial direct writeback: partial_1_12
"tbz x10, #0, 106f\n"
"st1 { v8.b }[12], [x11]\n"
- "st1 { v12.b }[12], [x24]\n"
- "st1 { v16.b }[12], [x23]\n"
- "st1 { v20.b }[12], [x22]\n"
+ "st1 { v12.b }[12], [x25]\n"
+ "st1 { v16.b }[12], [x24]\n"
+ "st1 { v20.b }[12], [x23]\n"
"b 106f\n"
"100:" // Height 4: Partial direct writeback: partial_2_8
"tbz x10, #1, 101f\n"
"st1 { v8.h }[4], [x11], #0x2\n"
- "st1 { v12.h }[4], [x24], #0x2\n"
- "st1 { v16.h }[4], [x23], #0x2\n"
- "st1 { v20.h }[4], [x22], #0x2\n"
+ "st1 { v12.h }[4], [x25], #0x2\n"
+ "st1 { v16.h }[4], [x24], #0x2\n"
+ "st1 { v20.h }[4], [x23], #0x2\n"
"tbz x10, #0, 106f\n"
"st1 { v8.b }[10], [x11]\n"
- "st1 { v12.b }[10], [x24]\n"
- "st1 { v16.b }[10], [x23]\n"
- "st1 { v20.b }[10], [x22]\n"
+ "st1 { v12.b }[10], [x25]\n"
+ "st1 { v16.b }[10], [x24]\n"
+ "st1 { v20.b }[10], [x23]\n"
"b 106f\n"
"101:" // Height 4: Partial direct writeback: partial_1_8
"tbz x10, #0, 106f\n"
"st1 { v8.b }[8], [x11]\n"
- "st1 { v12.b }[8], [x24]\n"
- "st1 { v16.b }[8], [x23]\n"
- "st1 { v20.b }[8], [x22]\n"
+ "st1 { v12.b }[8], [x25]\n"
+ "st1 { v16.b }[8], [x24]\n"
+ "st1 { v20.b }[8], [x23]\n"
"b 106f\n"
"102:" // Height 4: Partial direct writeback: partial_4_0
"tbz x10, #2, 104f\n"
"str s8, [x11], #0x4\n"
- "str s12, [x24], #0x4\n"
- "str s16, [x23], #0x4\n"
- "str s20, [x22], #0x4\n"
+ "str s12, [x25], #0x4\n"
+ "str s16, [x24], #0x4\n"
+ "str s20, [x23], #0x4\n"
"tbz x10, #1, 103f\n"
"st1 { v8.h }[2], [x11], #0x2\n"
- "st1 { v12.h }[2], [x24], #0x2\n"
- "st1 { v16.h }[2], [x23], #0x2\n"
- "st1 { v20.h }[2], [x22], #0x2\n"
+ "st1 { v12.h }[2], [x25], #0x2\n"
+ "st1 { v16.h }[2], [x24], #0x2\n"
+ "st1 { v20.h }[2], [x23], #0x2\n"
"tbz x10, #0, 106f\n"
"st1 { v8.b }[6], [x11]\n"
- "st1 { v12.b }[6], [x24]\n"
- "st1 { v16.b }[6], [x23]\n"
- "st1 { v20.b }[6], [x22]\n"
+ "st1 { v12.b }[6], [x25]\n"
+ "st1 { v16.b }[6], [x24]\n"
+ "st1 { v20.b }[6], [x23]\n"
"b 106f\n"
"103:" // Height 4: Partial direct writeback: partial_1_4
"tbz x10, #0, 106f\n"
"st1 { v8.b }[4], [x11]\n"
- "st1 { v12.b }[4], [x24]\n"
- "st1 { v16.b }[4], [x23]\n"
- "st1 { v20.b }[4], [x22]\n"
+ "st1 { v12.b }[4], [x25]\n"
+ "st1 { v16.b }[4], [x24]\n"
+ "st1 { v20.b }[4], [x23]\n"
"b 106f\n"
"104:" // Height 4: Partial direct writeback: partial_2_0
"tbz x10, #1, 105f\n"
"str h8, [x11], #0x2\n"
- "str h12, [x24], #0x2\n"
- "str h16, [x23], #0x2\n"
- "str h20, [x22], #0x2\n"
+ "str h12, [x25], #0x2\n"
+ "str h16, [x24], #0x2\n"
+ "str h20, [x23], #0x2\n"
"tbz x10, #0, 106f\n"
"st1 { v8.b }[2], [x11]\n"
- "st1 { v12.b }[2], [x24]\n"
- "st1 { v16.b }[2], [x23]\n"
- "st1 { v20.b }[2], [x22]\n"
+ "st1 { v12.b }[2], [x25]\n"
+ "st1 { v16.b }[2], [x24]\n"
+ "st1 { v20.b }[2], [x23]\n"
"b 106f\n"
"105:" // Height 4: Partial direct writeback: partial_1_0
"str b8, [x11, #0x0]\n"
- "str b12, [x24, #0x0]\n"
- "str b16, [x23, #0x0]\n"
- "str b20, [x22, #0x0]\n"
+ "str b12, [x25, #0x0]\n"
+ "str b16, [x24, #0x0]\n"
+ "str b20, [x23, #0x0]\n"
"106:" // Height 4: Partial direct writeback: Done
"b 108f\n"
"107:" // Height 4: Full writeback
"str q8, [x11, #0x0]\n"
"add x11, x11, #0x10\n"
- "str q12, [x24, #0x0]\n"
- "str q16, [x23, #0x0]\n"
- "str q20, [x22, #0x0]\n"
+ "str q12, [x25, #0x0]\n"
+ "str q16, [x24, #0x0]\n"
+ "str q20, [x23, #0x0]\n"
"108:" // Height 4: Writeback done
"subs x10, x10, #0x10\n"
"bgt 83b\n"
@@ -1950,15 +1949,15 @@ void a64_hybrid_s8qs_dot_6x16 (
"112:" // Height 5: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, 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, 113f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
- "ldr x22, [x21, #0x20]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
"cbnz x28, 114f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -1969,10 +1968,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"b 114f\n"
"113:" // Height 5: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
- "add x23, x24, x20\n"
- "add x22, x23, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
"114:" // Height 5: input setup done
"cmp x27, #0x10\n"
"blt 117f\n"
@@ -1995,7 +1994,7 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x25, x25, #0x10\n"
"add x24, x24, #0x10\n"
".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q29, [x9, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x23, x23, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2004,100 +2003,100 @@ void a64_hybrid_s8qs_dot_6x16 (
"cmp x27, #0x20\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q28, [x9, #0x30]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
"prfm pldl1keep, [x25, #0x80]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n"
"prfm pldl1keep, [x23, #0x80]\n"
"prfm pldl1keep, [x22, #0x80]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n"
+ "ldr q29, [x9, #0x40]\n"
+ ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n"
+ "ldr q28, [x9, #0x50]\n"
+ ".inst 0x4fa0e3a8 // sdot v8.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ac // sdot v12.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b0 // sdot v16.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b4 // sdot v20.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x9, #0x60]\n"
+ ".inst 0x4fa0e389 // sdot v9.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38d // sdot v13.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e391 // sdot v17.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e395 // sdot v21.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x9, #0x70]\n"
+ ".inst 0x4fa0e3aa // sdot v10.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ae // sdot v14.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b2 // sdot v18.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b6 // sdot v22.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3ba // sdot v26.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x9, #0x80]\n"
+ ".inst 0x4fa0e38b // sdot v11.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38f // sdot v15.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e393 // sdot v19.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e397 // sdot v23.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e39b // sdot v27.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x9, #0x90]\n"
+ ".inst 0x4f80eba8 // sdot v8.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebac // sdot v12.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb0 // sdot v16.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb4 // sdot v20.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebb8 // sdot v24.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x9, #0xa0]\n"
+ ".inst 0x4f80eb89 // sdot v9.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8d // sdot v13.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb91 // sdot v17.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb95 // sdot v21.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb99 // sdot v25.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x9, #0xb0]\n"
+ ".inst 0x4f80ebaa // sdot v10.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebae // sdot v14.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb2 // sdot v18.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb6 // sdot v22.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebba // sdot v26.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x9, #0xc0]\n"
+ ".inst 0x4f80eb8b // sdot v11.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8f // sdot v15.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb93 // sdot v19.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb97 // sdot v23.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb9b // sdot v27.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x9, #0xd0]\n"
+ ".inst 0x4fa0eba8 // sdot v8.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebac // sdot v12.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb0 // sdot v16.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb4 // sdot v20.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebb8 // sdot v24.4s, v29.16b, v4.4b[3]\n"
+ "ldr q29, [x9, #0xe0]\n"
+ ".inst 0x4fa0eb89 // sdot v9.4s, v28.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb8d // sdot v13.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb91 // sdot v17.4s, v28.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb95 // sdot v21.4s, v28.16b, v3.4b[3]\n"
+ ".inst 0x4fa4eb99 // sdot v25.4s, v28.16b, v4.4b[3]\n"
+ "ldr q28, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ebaa // sdot v10.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebae // sdot v14.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb2 // sdot v18.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb6 // sdot v22.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebba // sdot v26.4s, v29.16b, v4.4b[3]\n"
"ldr q6, [x9, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0eb8b // sdot v11.4s, v28.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1eb8f // sdot v15.4s, v28.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb93 // sdot v19.4s, v28.16b, v2.4b[3]\n"
"ldr q2, [x24, #0x0]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
+ ".inst 0x4fa3eb97 // sdot v23.4s, v28.16b, v3.4b[3]\n"
"ldr q3, [x23, #0x0]\n"
- ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n"
+ ".inst 0x4fa4eb9b // sdot v27.4s, v28.16b, v4.4b[3]\n"
"ldr q4, [x22, #0x0]\n"
"ldr q7, [x9, #0x10]\n"
"bge 115b\n"
@@ -2111,7 +2110,7 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x24, x24, #0x10\n"
"add x23, x23, #0x10\n"
".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
+ "ldr q29, [x9, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x22, x22, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2120,131 +2119,131 @@ void a64_hybrid_s8qs_dot_6x16 (
"prfm pldl1keep, [x26, #0x80]\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q28, [x9, #0x30]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
"prfm pldl1keep, [x24, #0x80]\n"
"prfm pldl1keep, [x23, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n"
"prfm pldl1keep, [x22, #0x80]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x9, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x9, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x9, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x9, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x9, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x9, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x9, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x9, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x9, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x9, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n"
- "ldr q6, [x9, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n"
- "ldr q7, [x9, #0xf0]\n"
+ ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n"
+ "ldr q29, [x9, #0x40]\n"
+ ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n"
+ "ldr q28, [x9, #0x50]\n"
+ ".inst 0x4fa0e3a8 // sdot v8.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ac // sdot v12.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b0 // sdot v16.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b4 // sdot v20.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x9, #0x60]\n"
+ ".inst 0x4fa0e389 // sdot v9.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38d // sdot v13.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e391 // sdot v17.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e395 // sdot v21.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x9, #0x70]\n"
+ ".inst 0x4fa0e3aa // sdot v10.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ae // sdot v14.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b2 // sdot v18.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b6 // sdot v22.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3ba // sdot v26.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x9, #0x80]\n"
+ ".inst 0x4fa0e38b // sdot v11.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38f // sdot v15.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e393 // sdot v19.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e397 // sdot v23.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e39b // sdot v27.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x9, #0x90]\n"
+ ".inst 0x4f80eba8 // sdot v8.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebac // sdot v12.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb0 // sdot v16.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb4 // sdot v20.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebb8 // sdot v24.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x9, #0xa0]\n"
+ ".inst 0x4f80eb89 // sdot v9.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8d // sdot v13.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb91 // sdot v17.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb95 // sdot v21.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb99 // sdot v25.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x9, #0xb0]\n"
+ ".inst 0x4f80ebaa // sdot v10.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebae // sdot v14.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb2 // sdot v18.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb6 // sdot v22.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebba // sdot v26.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x9, #0xc0]\n"
+ ".inst 0x4f80eb8b // sdot v11.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8f // sdot v15.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb93 // sdot v19.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb97 // sdot v23.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb9b // sdot v27.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x9, #0xd0]\n"
+ ".inst 0x4fa0eba8 // sdot v8.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebac // sdot v12.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb0 // sdot v16.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb4 // sdot v20.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebb8 // sdot v24.4s, v29.16b, v4.4b[3]\n"
+ "ldr q29, [x9, #0xe0]\n"
+ ".inst 0x4fa0eb89 // sdot v9.4s, v28.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb8d // sdot v13.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb91 // sdot v17.4s, v28.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb95 // sdot v21.4s, v28.16b, v3.4b[3]\n"
+ ".inst 0x4fa4eb99 // sdot v25.4s, v28.16b, v4.4b[3]\n"
+ "ldr q28, [x9, #0xf0]\n"
"add x9, x9, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ebaa // sdot v10.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebae // sdot v14.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb2 // sdot v18.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb6 // sdot v22.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebba // sdot v26.4s, v29.16b, v4.4b[3]\n"
+ ".inst 0x4fa0eb8b // sdot v11.4s, v28.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb8f // sdot v15.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb93 // sdot v19.4s, v28.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb97 // sdot v23.4s, v28.16b, v3.4b[3]\n"
+ ".inst 0x4fa4eb9b // sdot v27.4s, v28.16b, v4.4b[3]\n"
"117:" // Height 5: Multiply loop: Main loop skip
"cbz x27, 122f\n"
"cmp x27, #0x4\n"
"blt 119f\n"
"118:" // Height 5: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
+ "ldr s2, [x26], #0x4\n"
"ldr s1, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr s3, [x23], #0x4\n"
- "ldr s4, [x22], #0x4\n"
- "ldr q6, [x9, #0x0]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr s0, [x24], #0x4\n"
+ "ldr s31, [x23], #0x4\n"
+ "ldr s30, [x22], #0x4\n"
+ "ldr q29, [x9, #0x0]\n"
+ ".inst 0x4f82e3a8 // sdot v8.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f81e3ac // sdot v12.4s, v29.16b, v1.4b[0]\n"
+ "ldr q28, [x9, #0x10]\n"
+ ".inst 0x4f80e3b0 // sdot v16.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe3b4 // sdot v20.4s, v29.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee3b8 // sdot v24.4s, v29.16b, v30.4b[0]\n"
+ "ldr q29, [x9, #0x20]\n"
+ ".inst 0x4f82e389 // sdot v9.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f81e38d // sdot v13.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f80e391 // sdot v17.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe395 // sdot v21.4s, v28.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee399 // sdot v25.4s, v28.16b, v30.4b[0]\n"
+ "ldr q28, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f82e3aa // sdot v10.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
+ ".inst 0x4f80e3b2 // sdot v18.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe3b6 // sdot v22.4s, v29.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee3ba // sdot v26.4s, v29.16b, v30.4b[0]\n"
+ ".inst 0x4f82e38b // sdot v11.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f80e393 // sdot v19.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe397 // sdot v23.4s, v28.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee39b // sdot v27.4s, v28.16b, v30.4b[0]\n"
"bge 118b\n"
"119:" // Height 5: Multiply loop: Skip odd blocks
"cbz x27, 122f\n"
@@ -2268,71 +2267,71 @@ void a64_hybrid_s8qs_dot_6x16 (
"ldr b3, [x23, #0x0]\n"
"ldr b4, [x22, #0x0]\n"
"121:" // Height 5: Multiply loop: Ragged operand read: Done
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q29, [x9, #0x0]\n"
+ "ldr q28, [x9, #0x10]\n"
+ ".inst 0x4f80e3a8 // sdot v8.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ac // sdot v12.4s, v29.16b, v1.4b[0]\n"
+ ".inst 0x4f82e3b0 // sdot v16.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b4 // sdot v20.4s, v29.16b, v3.4b[0]\n"
+ ".inst 0x4f84e3b8 // sdot v24.4s, v29.16b, v4.4b[0]\n"
+ "ldr q29, [x9, #0x20]\n"
+ ".inst 0x4f80e389 // sdot v9.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38d // sdot v13.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e391 // sdot v17.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e395 // sdot v21.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e399 // sdot v25.4s, v28.16b, v4.4b[0]\n"
+ "ldr q28, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
+ ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n"
+ ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n"
+ ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n"
"122:" // Height 5: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
"cmp x28, x20\n"
"bne 112b\n"
- "ldr q0, [x14, #0x0]\n"
- "ldr q1, [x14, #0x10]\n"
- "add v8.4s, v8.4s, v0.4s\n"
- "add v9.4s, v9.4s, v1.4s\n"
- "ldr q2, [x14, #0x20]\n"
- "ldr q3, [x14, #0x30]\n"
- "add v10.4s, v10.4s, v2.4s\n"
- "add v11.4s, v11.4s, v3.4s\n"
+ "ldr q31, [x14, #0x0]\n"
+ "ldr q30, [x14, #0x10]\n"
+ "add v8.4s, v8.4s, v31.4s\n"
+ "add v9.4s, v9.4s, v30.4s\n"
+ "ldr q29, [x14, #0x20]\n"
+ "ldr q28, [x14, #0x30]\n"
+ "add v10.4s, v10.4s, v29.4s\n"
+ "add v11.4s, v11.4s, v28.4s\n"
"ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "add x24, x11, x20\n"
- "add x23, x24, x20\n"
+ "add x25, x11, x20\n"
+ "add x24, x25, x20\n"
"prfm pstl1keep, [x11, #0x0]\n"
+ "add x23, x24, x20\n"
"add x22, x23, x20\n"
- "add x21, x22, x20\n"
+ "prfm pstl1keep, [x25, #0x0]\n"
"prfm pstl1keep, [x24, #0x0]\n"
"prfm pstl1keep, [x23, #0x0]\n"
"prfm pstl1keep, [x22, #0x0]\n"
- "prfm pstl1keep, [x21, #0x0]\n"
- "add v12.4s, v12.4s, v0.4s\n"
- "add v13.4s, v13.4s, v1.4s\n"
- "add v14.4s, v14.4s, v2.4s\n"
- "add v15.4s, v15.4s, v3.4s\n"
+ "add v12.4s, v12.4s, v31.4s\n"
+ "add v13.4s, v13.4s, v30.4s\n"
+ "add v14.4s, v14.4s, v29.4s\n"
+ "add v15.4s, v15.4s, v28.4s\n"
"add x14, x14, #0x40\n"
- "add v16.4s, v16.4s, v0.4s\n"
- "add v17.4s, v17.4s, v1.4s\n"
- "add v18.4s, v18.4s, v2.4s\n"
- "add v19.4s, v19.4s, v3.4s\n"
- "add v20.4s, v20.4s, v0.4s\n"
- "add v21.4s, v21.4s, v1.4s\n"
- "add v22.4s, v22.4s, v2.4s\n"
- "add v23.4s, v23.4s, v3.4s\n"
- "add v24.4s, v24.4s, v0.4s\n"
- "add v25.4s, v25.4s, v1.4s\n"
- "add v26.4s, v26.4s, v2.4s\n"
- "add v27.4s, v27.4s, v3.4s\n"
+ "add v16.4s, v16.4s, v31.4s\n"
+ "add v17.4s, v17.4s, v30.4s\n"
+ "add v18.4s, v18.4s, v29.4s\n"
+ "add v19.4s, v19.4s, v28.4s\n"
+ "add v20.4s, v20.4s, v31.4s\n"
+ "add v21.4s, v21.4s, v30.4s\n"
+ "add v22.4s, v22.4s, v29.4s\n"
+ "add v23.4s, v23.4s, v28.4s\n"
+ "add v24.4s, v24.4s, v31.4s\n"
+ "add v25.4s, v25.4s, v30.4s\n"
+ "add v26.4s, v26.4s, v29.4s\n"
+ "add v27.4s, v27.4s, v28.4s\n"
"tbz %x[flags], #4, 123f\n"
"ldr q0, [x12, #0x0]\n"
"ldr q4, [x13, #0x0]\n"
@@ -2346,10 +2345,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x13, x13, #0x40\n"
"b 124f\n"
"123:" // Height 5: per layer parameters
- "add x25, %x[qp], %[per_layer_right_shift]\n"
- "ld1r { v0.4s }, [x25]\n"
- "add x25, %x[qp], %[per_layer_mul]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v4.4s }, [x20]\n"
"mov v1.16b, v0.16b\n"
"mov v5.16b, v4.16b\n"
"mov v2.16b, v0.16b\n"
@@ -2378,79 +2377,79 @@ void a64_hybrid_s8qs_dot_6x16 (
"sqrdmulh v26.4s, v26.4s, v6.4s\n"
"sqrdmulh v27.4s, v27.4s, v7.4s\n"
"tbz %x[flags], #5, 125f\n"
- "and v4.16b, v8.16b, v0.16b\n"
- "and v5.16b, v9.16b, v1.16b\n"
- "and v6.16b, v10.16b, v2.16b\n"
- "and v7.16b, v11.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v8.4s, v8.4s, v4.4s\n"
- "sqadd v9.4s, v9.4s, v5.4s\n"
- "sqadd v10.4s, v10.4s, v6.4s\n"
- "sqadd v11.4s, v11.4s, v7.4s\n"
- "and v4.16b, v12.16b, v0.16b\n"
- "and v5.16b, v13.16b, v1.16b\n"
- "and v6.16b, v14.16b, v2.16b\n"
- "and v7.16b, v15.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v12.4s, v12.4s, v4.4s\n"
- "sqadd v13.4s, v13.4s, v5.4s\n"
- "sqadd v14.4s, v14.4s, v6.4s\n"
- "sqadd v15.4s, v15.4s, v7.4s\n"
- "and v4.16b, v16.16b, v0.16b\n"
- "and v5.16b, v17.16b, v1.16b\n"
- "and v6.16b, v18.16b, v2.16b\n"
- "and v7.16b, v19.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v16.4s, v16.4s, v4.4s\n"
- "sqadd v17.4s, v17.4s, v5.4s\n"
- "sqadd v18.4s, v18.4s, v6.4s\n"
- "sqadd v19.4s, v19.4s, v7.4s\n"
- "and v4.16b, v20.16b, v0.16b\n"
- "and v5.16b, v21.16b, v1.16b\n"
- "and v6.16b, v22.16b, v2.16b\n"
- "and v7.16b, v23.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v20.4s, v20.4s, v4.4s\n"
- "sqadd v21.4s, v21.4s, v5.4s\n"
- "sqadd v22.4s, v22.4s, v6.4s\n"
- "sqadd v23.4s, v23.4s, v7.4s\n"
- "and v4.16b, v24.16b, v0.16b\n"
- "and v5.16b, v25.16b, v1.16b\n"
- "and v6.16b, v26.16b, v2.16b\n"
- "and v7.16b, v27.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v24.4s, v24.4s, v4.4s\n"
- "sqadd v25.4s, v25.4s, v5.4s\n"
- "sqadd v26.4s, v26.4s, v6.4s\n"
- "sqadd v27.4s, v27.4s, v7.4s\n"
+ "and v31.16b, v8.16b, v0.16b\n"
+ "and v30.16b, v9.16b, v1.16b\n"
+ "and v29.16b, v10.16b, v2.16b\n"
+ "and v28.16b, v11.16b, v3.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sqadd v8.4s, v8.4s, v31.4s\n"
+ "sqadd v9.4s, v9.4s, v30.4s\n"
+ "sqadd v10.4s, v10.4s, v29.4s\n"
+ "sqadd v11.4s, v11.4s, v28.4s\n"
+ "and v31.16b, v12.16b, v0.16b\n"
+ "and v30.16b, v13.16b, v1.16b\n"
+ "and v29.16b, v14.16b, v2.16b\n"
+ "and v28.16b, v15.16b, v3.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sqadd v12.4s, v12.4s, v31.4s\n"
+ "sqadd v13.4s, v13.4s, v30.4s\n"
+ "sqadd v14.4s, v14.4s, v29.4s\n"
+ "sqadd v15.4s, v15.4s, v28.4s\n"
+ "and v31.16b, v16.16b, v0.16b\n"
+ "and v30.16b, v17.16b, v1.16b\n"
+ "and v29.16b, v18.16b, v2.16b\n"
+ "and v28.16b, v19.16b, v3.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sqadd v16.4s, v16.4s, v31.4s\n"
+ "sqadd v17.4s, v17.4s, v30.4s\n"
+ "sqadd v18.4s, v18.4s, v29.4s\n"
+ "sqadd v19.4s, v19.4s, v28.4s\n"
+ "and v31.16b, v20.16b, v0.16b\n"
+ "and v30.16b, v21.16b, v1.16b\n"
+ "and v29.16b, v22.16b, v2.16b\n"
+ "and v28.16b, v23.16b, v3.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v31.4s\n"
+ "sqadd v21.4s, v21.4s, v30.4s\n"
+ "sqadd v22.4s, v22.4s, v29.4s\n"
+ "sqadd v23.4s, v23.4s, v28.4s\n"
+ "and v31.16b, v24.16b, v0.16b\n"
+ "and v30.16b, v25.16b, v1.16b\n"
+ "and v29.16b, v26.16b, v2.16b\n"
+ "and v28.16b, v27.16b, v3.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sqadd v24.4s, v24.4s, v31.4s\n"
+ "sqadd v25.4s, v25.4s, v30.4s\n"
+ "sqadd v26.4s, v26.4s, v29.4s\n"
+ "sqadd v27.4s, v27.4s, v28.4s\n"
"125:" // Height 5: no shift correction
- "add x25, %x[qp], %[c_offset]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v30.4s }, [x20]\n"
"srshl v8.4s, v8.4s, v0.4s\n"
"srshl v9.4s, v9.4s, v1.4s\n"
"srshl v10.4s, v10.4s, v2.4s\n"
"srshl v11.4s, v11.4s, v3.4s\n"
- "add x25, %x[qp], %[maxval]\n"
- "ld1r { v6.4s }, [x25]\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v29.4s }, [x20]\n"
"srshl v12.4s, v12.4s, v0.4s\n"
"srshl v13.4s, v13.4s, v1.4s\n"
- "add x25, %x[qp], %[minval]\n"
- "ld1r { v5.4s }, [x25]\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v28.4s }, [x20]\n"
"srshl v14.4s, v14.4s, v2.4s\n"
"srshl v15.4s, v15.4s, v3.4s\n"
"cmp x10, #0x10\n"
@@ -2466,194 +2465,194 @@ void a64_hybrid_s8qs_dot_6x16 (
"srshl v25.4s, v25.4s, v1.4s\n"
"srshl v26.4s, v26.4s, v2.4s\n"
"srshl v27.4s, v27.4s, v3.4s\n"
- "add v8.4s, v8.4s, v4.4s\n"
- "add v9.4s, v9.4s, v4.4s\n"
- "add v10.4s, v10.4s, v4.4s\n"
- "add v11.4s, v11.4s, v4.4s\n"
- "add v12.4s, v12.4s, v4.4s\n"
- "add v13.4s, v13.4s, v4.4s\n"
- "add v14.4s, v14.4s, v4.4s\n"
- "add v15.4s, v15.4s, v4.4s\n"
- "add v16.4s, v16.4s, v4.4s\n"
- "add v17.4s, v17.4s, v4.4s\n"
- "add v18.4s, v18.4s, v4.4s\n"
- "add v19.4s, v19.4s, v4.4s\n"
- "add v20.4s, v20.4s, v4.4s\n"
- "add v21.4s, v21.4s, v4.4s\n"
- "add v22.4s, v22.4s, v4.4s\n"
- "add v23.4s, v23.4s, v4.4s\n"
- "add v24.4s, v24.4s, v4.4s\n"
- "add v25.4s, v25.4s, v4.4s\n"
- "add v26.4s, v26.4s, v4.4s\n"
- "add v27.4s, v27.4s, v4.4s\n"
- "smin v8.4s, v8.4s, v6.4s\n"
- "smin v9.4s, v9.4s, v6.4s\n"
- "smin v10.4s, v10.4s, v6.4s\n"
- "smin v11.4s, v11.4s, v6.4s\n"
- "smin v12.4s, v12.4s, v6.4s\n"
- "smin v13.4s, v13.4s, v6.4s\n"
- "smin v14.4s, v14.4s, v6.4s\n"
- "smin v15.4s, v15.4s, v6.4s\n"
- "smin v16.4s, v16.4s, v6.4s\n"
- "smin v17.4s, v17.4s, v6.4s\n"
- "smin v18.4s, v18.4s, v6.4s\n"
- "smin v19.4s, v19.4s, v6.4s\n"
- "smin v20.4s, v20.4s, v6.4s\n"
- "smin v21.4s, v21.4s, v6.4s\n"
- "smin v22.4s, v22.4s, v6.4s\n"
- "smin v23.4s, v23.4s, v6.4s\n"
- "smin v24.4s, v24.4s, v6.4s\n"
- "smin v25.4s, v25.4s, v6.4s\n"
- "smin v26.4s, v26.4s, v6.4s\n"
- "smin v27.4s, v27.4s, v6.4s\n"
- "smax v8.4s, v8.4s, v5.4s\n"
- "smax v9.4s, v9.4s, v5.4s\n"
- "smax v10.4s, v10.4s, v5.4s\n"
- "smax v11.4s, v11.4s, v5.4s\n"
- "smax v12.4s, v12.4s, v5.4s\n"
- "smax v13.4s, v13.4s, v5.4s\n"
- "smax v14.4s, v14.4s, v5.4s\n"
- "smax v15.4s, v15.4s, v5.4s\n"
- "smax v16.4s, v16.4s, v5.4s\n"
- "smax v17.4s, v17.4s, v5.4s\n"
- "smax v18.4s, v18.4s, v5.4s\n"
- "smax v19.4s, v19.4s, v5.4s\n"
- "smax v20.4s, v20.4s, v5.4s\n"
- "smax v21.4s, v21.4s, v5.4s\n"
- "smax v22.4s, v22.4s, v5.4s\n"
- "smax v23.4s, v23.4s, v5.4s\n"
- "smax v24.4s, v24.4s, v5.4s\n"
- "smax v25.4s, v25.4s, v5.4s\n"
- "smax v26.4s, v26.4s, v5.4s\n"
- "smax v27.4s, v27.4s, v5.4s\n"
+ "add v8.4s, v8.4s, v30.4s\n"
+ "add v9.4s, v9.4s, v30.4s\n"
+ "add v10.4s, v10.4s, v30.4s\n"
+ "add v11.4s, v11.4s, v30.4s\n"
+ "add v12.4s, v12.4s, v30.4s\n"
+ "add v13.4s, v13.4s, v30.4s\n"
+ "add v14.4s, v14.4s, v30.4s\n"
+ "add v15.4s, v15.4s, v30.4s\n"
+ "add v16.4s, v16.4s, v30.4s\n"
+ "add v17.4s, v17.4s, v30.4s\n"
+ "add v18.4s, v18.4s, v30.4s\n"
+ "add v19.4s, v19.4s, v30.4s\n"
+ "add v20.4s, v20.4s, v30.4s\n"
+ "add v21.4s, v21.4s, v30.4s\n"
+ "add v22.4s, v22.4s, v30.4s\n"
+ "add v23.4s, v23.4s, v30.4s\n"
+ "add v24.4s, v24.4s, v30.4s\n"
+ "add v25.4s, v25.4s, v30.4s\n"
+ "add v26.4s, v26.4s, v30.4s\n"
+ "add v27.4s, v27.4s, v30.4s\n"
+ "smin v8.4s, v8.4s, v29.4s\n"
+ "smin v9.4s, v9.4s, v29.4s\n"
+ "smin v10.4s, v10.4s, v29.4s\n"
+ "smin v11.4s, v11.4s, v29.4s\n"
+ "smin v12.4s, v12.4s, v29.4s\n"
+ "smin v13.4s, v13.4s, v29.4s\n"
+ "smin v14.4s, v14.4s, v29.4s\n"
+ "smin v15.4s, v15.4s, v29.4s\n"
+ "smin v16.4s, v16.4s, v29.4s\n"
+ "smin v17.4s, v17.4s, v29.4s\n"
+ "smin v18.4s, v18.4s, v29.4s\n"
+ "smin v19.4s, v19.4s, v29.4s\n"
+ "smin v20.4s, v20.4s, v29.4s\n"
+ "smin v21.4s, v21.4s, v29.4s\n"
+ "smin v22.4s, v22.4s, v29.4s\n"
+ "smin v23.4s, v23.4s, v29.4s\n"
+ "smin v24.4s, v24.4s, v29.4s\n"
+ "smin v25.4s, v25.4s, v29.4s\n"
+ "smin v26.4s, v26.4s, v29.4s\n"
+ "smin v27.4s, v27.4s, v29.4s\n"
+ "smax v8.4s, v8.4s, v28.4s\n"
+ "smax v9.4s, v9.4s, v28.4s\n"
+ "smax v10.4s, v10.4s, v28.4s\n"
+ "smax v11.4s, v11.4s, v28.4s\n"
+ "smax v12.4s, v12.4s, v28.4s\n"
+ "smax v13.4s, v13.4s, v28.4s\n"
+ "smax v14.4s, v14.4s, v28.4s\n"
+ "smax v15.4s, v15.4s, v28.4s\n"
+ "smax v16.4s, v16.4s, v28.4s\n"
+ "smax v17.4s, v17.4s, v28.4s\n"
+ "smax v18.4s, v18.4s, v28.4s\n"
+ "smax v19.4s, v19.4s, v28.4s\n"
+ "smax v20.4s, v20.4s, v28.4s\n"
+ "smax v21.4s, v21.4s, v28.4s\n"
+ "smax v22.4s, v22.4s, v28.4s\n"
+ "smax v23.4s, v23.4s, v28.4s\n"
+ "smax v24.4s, v24.4s, v28.4s\n"
+ "smax v25.4s, v25.4s, v28.4s\n"
+ "smax v26.4s, v26.4s, v28.4s\n"
+ "smax v27.4s, v27.4s, v28.4s\n"
"uzp1 v8.8h, v8.8h, v9.8h\n"
- "uzp1 v9.8h, v10.8h, v11.8h\n"
+ "uzp1 v29.8h, v10.8h, v11.8h\n"
"uzp1 v12.8h, v12.8h, v13.8h\n"
- "uzp1 v13.8h, v14.8h, v15.8h\n"
+ "uzp1 v28.8h, v14.8h, v15.8h\n"
"uzp1 v16.8h, v16.8h, v17.8h\n"
- "uzp1 v17.8h, v18.8h, v19.8h\n"
+ "uzp1 v19.8h, v18.8h, v19.8h\n"
"uzp1 v20.8h, v20.8h, v21.8h\n"
- "uzp1 v21.8h, v22.8h, v23.8h\n"
+ "uzp1 v18.8h, v22.8h, v23.8h\n"
"uzp1 v24.8h, v24.8h, v25.8h\n"
- "uzp1 v25.8h, v26.8h, v27.8h\n"
- "uzp1 v8.16b, v8.16b, v9.16b\n"
- "uzp1 v12.16b, v12.16b, v13.16b\n"
- "uzp1 v16.16b, v16.16b, v17.16b\n"
- "uzp1 v20.16b, v20.16b, v21.16b\n"
- "uzp1 v24.16b, v24.16b, v25.16b\n"
+ "uzp1 v17.8h, v26.8h, v27.8h\n"
+ "uzp1 v8.16b, v8.16b, v29.16b\n"
+ "uzp1 v12.16b, v12.16b, v28.16b\n"
+ "uzp1 v16.16b, v16.16b, v19.16b\n"
+ "uzp1 v20.16b, v20.16b, v18.16b\n"
+ "uzp1 v24.16b, v24.16b, v17.16b\n"
"bge 134f\n"
"tbz x10, #3, 129f\n"
"str d8, [x11], #0x8\n"
- "str d12, [x24], #0x8\n"
- "str d16, [x23], #0x8\n"
- "str d20, [x22], #0x8\n"
- "str d24, [x21], #0x8\n"
+ "str d12, [x25], #0x8\n"
+ "str d16, [x24], #0x8\n"
+ "str d20, [x23], #0x8\n"
+ "str d24, [x22], #0x8\n"
"tbz x10, #2, 127f\n"
"st1 { v8.s }[2], [x11], #0x4\n"
- "st1 { v12.s }[2], [x24], #0x4\n"
- "st1 { v16.s }[2], [x23], #0x4\n"
- "st1 { v20.s }[2], [x22], #0x4\n"
- "st1 { v24.s }[2], [x21], #0x4\n"
+ "st1 { v12.s }[2], [x25], #0x4\n"
+ "st1 { v16.s }[2], [x24], #0x4\n"
+ "st1 { v20.s }[2], [x23], #0x4\n"
+ "st1 { v24.s }[2], [x22], #0x4\n"
"tbz x10, #1, 126f\n"
"st1 { v8.h }[6], [x11], #0x2\n"
- "st1 { v12.h }[6], [x24], #0x2\n"
- "st1 { v16.h }[6], [x23], #0x2\n"
- "st1 { v20.h }[6], [x22], #0x2\n"
- "st1 { v24.h }[6], [x21], #0x2\n"
+ "st1 { v12.h }[6], [x25], #0x2\n"
+ "st1 { v16.h }[6], [x24], #0x2\n"
+ "st1 { v20.h }[6], [x23], #0x2\n"
+ "st1 { v24.h }[6], [x22], #0x2\n"
"tbz x10, #0, 133f\n"
"st1 { v8.b }[14], [x11]\n"
- "st1 { v12.b }[14], [x24]\n"
- "st1 { v16.b }[14], [x23]\n"
- "st1 { v20.b }[14], [x22]\n"
- "st1 { v24.b }[14], [x21]\n"
+ "st1 { v12.b }[14], [x25]\n"
+ "st1 { v16.b }[14], [x24]\n"
+ "st1 { v20.b }[14], [x23]\n"
+ "st1 { v24.b }[14], [x22]\n"
"b 133f\n"
"126:" // Height 5: Partial direct writeback: partial_1_12
"tbz x10, #0, 133f\n"
"st1 { v8.b }[12], [x11]\n"
- "st1 { v12.b }[12], [x24]\n"
- "st1 { v16.b }[12], [x23]\n"
- "st1 { v20.b }[12], [x22]\n"
- "st1 { v24.b }[12], [x21]\n"
+ "st1 { v12.b }[12], [x25]\n"
+ "st1 { v16.b }[12], [x24]\n"
+ "st1 { v20.b }[12], [x23]\n"
+ "st1 { v24.b }[12], [x22]\n"
"b 133f\n"
"127:" // Height 5: Partial direct writeback: partial_2_8
"tbz x10, #1, 128f\n"
"st1 { v8.h }[4], [x11], #0x2\n"
- "st1 { v12.h }[4], [x24], #0x2\n"
- "st1 { v16.h }[4], [x23], #0x2\n"
- "st1 { v20.h }[4], [x22], #0x2\n"
- "st1 { v24.h }[4], [x21], #0x2\n"
+ "st1 { v12.h }[4], [x25], #0x2\n"
+ "st1 { v16.h }[4], [x24], #0x2\n"
+ "st1 { v20.h }[4], [x23], #0x2\n"
+ "st1 { v24.h }[4], [x22], #0x2\n"
"tbz x10, #0, 133f\n"
"st1 { v8.b }[10], [x11]\n"
- "st1 { v12.b }[10], [x24]\n"
- "st1 { v16.b }[10], [x23]\n"
- "st1 { v20.b }[10], [x22]\n"
- "st1 { v24.b }[10], [x21]\n"
+ "st1 { v12.b }[10], [x25]\n"
+ "st1 { v16.b }[10], [x24]\n"
+ "st1 { v20.b }[10], [x23]\n"
+ "st1 { v24.b }[10], [x22]\n"
"b 133f\n"
"128:" // Height 5: Partial direct writeback: partial_1_8
"tbz x10, #0, 133f\n"
"st1 { v8.b }[8], [x11]\n"
- "st1 { v12.b }[8], [x24]\n"
- "st1 { v16.b }[8], [x23]\n"
- "st1 { v20.b }[8], [x22]\n"
- "st1 { v24.b }[8], [x21]\n"
+ "st1 { v12.b }[8], [x25]\n"
+ "st1 { v16.b }[8], [x24]\n"
+ "st1 { v20.b }[8], [x23]\n"
+ "st1 { v24.b }[8], [x22]\n"
"b 133f\n"
"129:" // Height 5: Partial direct writeback: partial_4_0
"tbz x10, #2, 131f\n"
"str s8, [x11], #0x4\n"
- "str s12, [x24], #0x4\n"
- "str s16, [x23], #0x4\n"
- "str s20, [x22], #0x4\n"
- "str s24, [x21], #0x4\n"
+ "str s12, [x25], #0x4\n"
+ "str s16, [x24], #0x4\n"
+ "str s20, [x23], #0x4\n"
+ "str s24, [x22], #0x4\n"
"tbz x10, #1, 130f\n"
"st1 { v8.h }[2], [x11], #0x2\n"
- "st1 { v12.h }[2], [x24], #0x2\n"
- "st1 { v16.h }[2], [x23], #0x2\n"
- "st1 { v20.h }[2], [x22], #0x2\n"
- "st1 { v24.h }[2], [x21], #0x2\n"
+ "st1 { v12.h }[2], [x25], #0x2\n"
+ "st1 { v16.h }[2], [x24], #0x2\n"
+ "st1 { v20.h }[2], [x23], #0x2\n"
+ "st1 { v24.h }[2], [x22], #0x2\n"
"tbz x10, #0, 133f\n"
"st1 { v8.b }[6], [x11]\n"
- "st1 { v12.b }[6], [x24]\n"
- "st1 { v16.b }[6], [x23]\n"
- "st1 { v20.b }[6], [x22]\n"
- "st1 { v24.b }[6], [x21]\n"
+ "st1 { v12.b }[6], [x25]\n"
+ "st1 { v16.b }[6], [x24]\n"
+ "st1 { v20.b }[6], [x23]\n"
+ "st1 { v24.b }[6], [x22]\n"
"b 133f\n"
"130:" // Height 5: Partial direct writeback: partial_1_4
"tbz x10, #0, 133f\n"
"st1 { v8.b }[4], [x11]\n"
- "st1 { v12.b }[4], [x24]\n"
- "st1 { v16.b }[4], [x23]\n"
- "st1 { v20.b }[4], [x22]\n"
- "st1 { v24.b }[4], [x21]\n"
+ "st1 { v12.b }[4], [x25]\n"
+ "st1 { v16.b }[4], [x24]\n"
+ "st1 { v20.b }[4], [x23]\n"
+ "st1 { v24.b }[4], [x22]\n"
"b 133f\n"
"131:" // Height 5: Partial direct writeback: partial_2_0
"tbz x10, #1, 132f\n"
"str h8, [x11], #0x2\n"
- "str h12, [x24], #0x2\n"
- "str h16, [x23], #0x2\n"
- "str h20, [x22], #0x2\n"
- "str h24, [x21], #0x2\n"
+ "str h12, [x25], #0x2\n"
+ "str h16, [x24], #0x2\n"
+ "str h20, [x23], #0x2\n"
+ "str h24, [x22], #0x2\n"
"tbz x10, #0, 133f\n"
"st1 { v8.b }[2], [x11]\n"
- "st1 { v12.b }[2], [x24]\n"
- "st1 { v16.b }[2], [x23]\n"
- "st1 { v20.b }[2], [x22]\n"
- "st1 { v24.b }[2], [x21]\n"
+ "st1 { v12.b }[2], [x25]\n"
+ "st1 { v16.b }[2], [x24]\n"
+ "st1 { v20.b }[2], [x23]\n"
+ "st1 { v24.b }[2], [x22]\n"
"b 133f\n"
"132:" // Height 5: Partial direct writeback: partial_1_0
"str b8, [x11, #0x0]\n"
- "str b12, [x24, #0x0]\n"
- "str b16, [x23, #0x0]\n"
- "str b20, [x22, #0x0]\n"
- "str b24, [x21, #0x0]\n"
+ "str b12, [x25, #0x0]\n"
+ "str b16, [x24, #0x0]\n"
+ "str b20, [x23, #0x0]\n"
+ "str b24, [x22, #0x0]\n"
"133:" // Height 5: Partial direct writeback: Done
"b 135f\n"
"134:" // Height 5: Full writeback
"str q8, [x11, #0x0]\n"
"add x11, x11, #0x10\n"
- "str q12, [x24, #0x0]\n"
- "str q16, [x23, #0x0]\n"
- "str q20, [x22, #0x0]\n"
- "str q24, [x21, #0x0]\n"
+ "str q12, [x25, #0x0]\n"
+ "str q16, [x24, #0x0]\n"
+ "str q20, [x23, #0x0]\n"
+ "str q24, [x22, #0x0]\n"
"135:" // Height 5: Writeback done
"subs x10, x10, #0x10\n"
"bgt 110b\n"
@@ -2698,16 +2697,16 @@ void a64_hybrid_s8qs_dot_6x16 (
"139:" // Height 6: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, 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, 140f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
- "ldr x22, [x21, #0x20]\n"
- "ldr x21, [x21, #0x28]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
+ "ldr x21, [x20, #0x28]\n"
"cbnz x28, 141f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -2719,11 +2718,11 @@ void a64_hybrid_s8qs_dot_6x16 (
"b 141f\n"
"140:" // Height 6: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
- "add x23, x24, x20\n"
- "add x22, x23, x20\n"
- "add x21, x22, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
+ "add x21, x22, x21\n"
"141:" // Height 6: input setup done
"cmp x27, #0x10\n"
"blt 144f\n"
@@ -3002,43 +3001,43 @@ void a64_hybrid_s8qs_dot_6x16 (
"cmp x27, #0x4\n"
"blt 146f\n"
"145:" // Height 6: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s7, [x26], #0x4\n"
+ "ldr s6, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr s3, [x23], #0x4\n"
- "ldr s4, [x22], #0x4\n"
- "ldr s5, [x21], #0x4\n"
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr s5, [x24], #0x4\n"
+ "ldr s4, [x23], #0x4\n"
+ "ldr s3, [x22], #0x4\n"
+ "ldr s2, [x21], #0x4\n"
+ "ldr q1, [x9, #0x0]\n"
+ "ldr q0, [x9, #0x10]\n"
+ ".inst 0x4f87e028 // sdot v8.4s, v1.16b, v7.4b[0]\n"
+ ".inst 0x4f86e02c // sdot v12.4s, v1.16b, v6.4b[0]\n"
+ ".inst 0x4f85e030 // sdot v16.4s, v1.16b, v5.4b[0]\n"
+ ".inst 0x4f84e034 // sdot v20.4s, v1.16b, v4.4b[0]\n"
+ ".inst 0x4f83e038 // sdot v24.4s, v1.16b, v3.4b[0]\n"
+ ".inst 0x4f82e03c // sdot v28.4s, v1.16b, v2.4b[0]\n"
+ "ldr q1, [x9, #0x20]\n"
+ ".inst 0x4f87e009 // sdot v9.4s, v0.16b, v7.4b[0]\n"
+ ".inst 0x4f86e00d // sdot v13.4s, v0.16b, v6.4b[0]\n"
+ ".inst 0x4f85e011 // sdot v17.4s, v0.16b, v5.4b[0]\n"
+ ".inst 0x4f84e015 // sdot v21.4s, v0.16b, v4.4b[0]\n"
+ ".inst 0x4f83e019 // sdot v25.4s, v0.16b, v3.4b[0]\n"
+ ".inst 0x4f82e01d // sdot v29.4s, v0.16b, v2.4b[0]\n"
+ "ldr q0, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n"
+ ".inst 0x4f87e02a // sdot v10.4s, v1.16b, v7.4b[0]\n"
+ ".inst 0x4f86e02e // sdot v14.4s, v1.16b, v6.4b[0]\n"
+ ".inst 0x4f85e032 // sdot v18.4s, v1.16b, v5.4b[0]\n"
+ ".inst 0x4f84e036 // sdot v22.4s, v1.16b, v4.4b[0]\n"
+ ".inst 0x4f83e03a // sdot v26.4s, v1.16b, v3.4b[0]\n"
+ ".inst 0x4f82e03e // sdot v30.4s, v1.16b, v2.4b[0]\n"
+ ".inst 0x4f87e00b // sdot v11.4s, v0.16b, v7.4b[0]\n"
+ ".inst 0x4f86e00f // sdot v15.4s, v0.16b, v6.4b[0]\n"
+ ".inst 0x4f85e013 // sdot v19.4s, v0.16b, v5.4b[0]\n"
+ ".inst 0x4f84e017 // sdot v23.4s, v0.16b, v4.4b[0]\n"
+ ".inst 0x4f83e01b // sdot v27.4s, v0.16b, v3.4b[0]\n"
+ ".inst 0x4f82e01f // sdot v31.4s, v0.16b, v2.4b[0]\n"
"bge 145b\n"
"146:" // Height 6: Multiply loop: Skip odd blocks
"cbz x27, 149f\n"
@@ -3065,81 +3064,81 @@ void a64_hybrid_s8qs_dot_6x16 (
"ldr b4, [x22, #0x0]\n"
"ldr b5, [x21, #0x0]\n"
"148:" // Height 6: Multiply loop: Ragged operand read: Done
- "ldr q6, [x9, #0x0]\n"
- "ldr q7, [x9, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n"
- "ldr q6, [x9, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n"
- "ldr q7, [x9, #0x30]\n"
+ "ldr q7, [x9, #0x0]\n"
+ "ldr q6, [x9, #0x10]\n"
+ ".inst 0x4f80e0e8 // sdot v8.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0ec // sdot v12.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0f0 // sdot v16.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0f4 // sdot v20.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0f8 // sdot v24.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0fc // sdot v28.4s, v7.16b, v5.4b[0]\n"
+ "ldr q7, [x9, #0x20]\n"
+ ".inst 0x4f80e0c9 // sdot v9.4s, v6.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0cd // sdot v13.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0d1 // sdot v17.4s, v6.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0d5 // sdot v21.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0d9 // sdot v25.4s, v6.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0dd // sdot v29.4s, v6.16b, v5.4b[0]\n"
+ "ldr q6, [x9, #0x30]\n"
"add x9, x9, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n"
+ ".inst 0x4f80e0ea // sdot v10.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0ee // sdot v14.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0f2 // sdot v18.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0f6 // sdot v22.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0fa // sdot v26.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0fe // sdot v30.4s, v7.16b, v5.4b[0]\n"
+ ".inst 0x4f80e0cb // sdot v11.4s, v6.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0cf // sdot v15.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0d3 // sdot v19.4s, v6.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0d7 // sdot v23.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0db // sdot v27.4s, v6.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0df // sdot v31.4s, v6.16b, v5.4b[0]\n"
"149:" // Height 6: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
"cmp x28, x20\n"
"bne 139b\n"
- "ldr q0, [x14, #0x0]\n"
- "ldr q1, [x14, #0x10]\n"
- "add v8.4s, v8.4s, v0.4s\n"
- "add v9.4s, v9.4s, v1.4s\n"
- "ldr q2, [x14, #0x20]\n"
- "ldr q3, [x14, #0x30]\n"
- "add v10.4s, v10.4s, v2.4s\n"
- "add v11.4s, v11.4s, v3.4s\n"
+ "ldr q3, [x14, #0x0]\n"
+ "ldr q2, [x14, #0x10]\n"
+ "add v8.4s, v8.4s, v3.4s\n"
+ "add v9.4s, v9.4s, v2.4s\n"
+ "ldr q1, [x14, #0x20]\n"
+ "ldr q0, [x14, #0x30]\n"
+ "add v10.4s, v10.4s, v1.4s\n"
+ "add v11.4s, v11.4s, v0.4s\n"
"ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
- "add x24, x11, x20\n"
- "add x23, x24, x20\n"
+ "add x25, x11, x20\n"
+ "add x24, x25, x20\n"
"prfm pstl1keep, [x11, #0x0]\n"
+ "add x23, x24, x20\n"
"add x22, x23, x20\n"
- "add x21, x22, x20\n"
+ "prfm pstl1keep, [x25, #0x0]\n"
"prfm pstl1keep, [x24, #0x0]\n"
+ "add x21, x22, x20\n"
"prfm pstl1keep, [x23, #0x0]\n"
- "add x20, x21, x20\n"
"prfm pstl1keep, [x22, #0x0]\n"
+ "add v12.4s, v12.4s, v3.4s\n"
"prfm pstl1keep, [x21, #0x0]\n"
- "add v12.4s, v12.4s, v0.4s\n"
- "prfm pstl1keep, [x20, #0x0]\n"
- "add v13.4s, v13.4s, v1.4s\n"
- "add v14.4s, v14.4s, v2.4s\n"
+ "add v13.4s, v13.4s, v2.4s\n"
+ "add v14.4s, v14.4s, v1.4s\n"
"add x14, x14, #0x40\n"
- "add v15.4s, v15.4s, v3.4s\n"
- "add v16.4s, v16.4s, v0.4s\n"
- "add v17.4s, v17.4s, v1.4s\n"
- "add v18.4s, v18.4s, v2.4s\n"
- "add v19.4s, v19.4s, v3.4s\n"
- "add v20.4s, v20.4s, v0.4s\n"
- "add v21.4s, v21.4s, v1.4s\n"
- "add v22.4s, v22.4s, v2.4s\n"
- "add v23.4s, v23.4s, v3.4s\n"
- "add v24.4s, v24.4s, v0.4s\n"
- "add v25.4s, v25.4s, v1.4s\n"
- "add v26.4s, v26.4s, v2.4s\n"
- "add v27.4s, v27.4s, v3.4s\n"
- "add v28.4s, v28.4s, v0.4s\n"
- "add v29.4s, v29.4s, v1.4s\n"
- "add v30.4s, v30.4s, v2.4s\n"
- "add v31.4s, v31.4s, v3.4s\n"
+ "add v15.4s, v15.4s, v0.4s\n"
+ "add v16.4s, v16.4s, v3.4s\n"
+ "add v17.4s, v17.4s, v2.4s\n"
+ "add v18.4s, v18.4s, v1.4s\n"
+ "add v19.4s, v19.4s, v0.4s\n"
+ "add v20.4s, v20.4s, v3.4s\n"
+ "add v21.4s, v21.4s, v2.4s\n"
+ "add v22.4s, v22.4s, v1.4s\n"
+ "add v23.4s, v23.4s, v0.4s\n"
+ "add v24.4s, v24.4s, v3.4s\n"
+ "add v25.4s, v25.4s, v2.4s\n"
+ "add v26.4s, v26.4s, v1.4s\n"
+ "add v27.4s, v27.4s, v0.4s\n"
+ "add v28.4s, v28.4s, v3.4s\n"
+ "add v29.4s, v29.4s, v2.4s\n"
+ "add v30.4s, v30.4s, v1.4s\n"
+ "add v31.4s, v31.4s, v0.4s\n"
"tbz %x[flags], #4, 150f\n"
"ldr q0, [x12, #0x0]\n"
"ldr q4, [x13, #0x0]\n"
@@ -3153,10 +3152,10 @@ void a64_hybrid_s8qs_dot_6x16 (
"add x13, x13, #0x40\n"
"b 151f\n"
"150:" // Height 6: per layer parameters
- "add x25, %x[qp], %[per_layer_right_shift]\n"
- "ld1r { v0.4s }, [x25]\n"
- "add x25, %x[qp], %[per_layer_mul]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v4.4s }, [x20]\n"
"mov v1.16b, v0.16b\n"
"mov v5.16b, v4.16b\n"
"mov v2.16b, v0.16b\n"
@@ -3189,91 +3188,91 @@ void a64_hybrid_s8qs_dot_6x16 (
"sqrdmulh v30.4s, v30.4s, v6.4s\n"
"sqrdmulh v31.4s, v31.4s, v7.4s\n"
"tbz %x[flags], #5, 152f\n"
- "and v4.16b, v8.16b, v0.16b\n"
- "and v5.16b, v9.16b, v1.16b\n"
- "and v6.16b, v10.16b, v2.16b\n"
- "and v7.16b, v11.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
+ "and v7.16b, v8.16b, v0.16b\n"
+ "and v6.16b, v9.16b, v1.16b\n"
+ "and v5.16b, v10.16b, v2.16b\n"
+ "and v4.16b, v11.16b, v3.16b\n"
"sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v8.4s, v8.4s, v4.4s\n"
- "sqadd v9.4s, v9.4s, v5.4s\n"
- "sqadd v10.4s, v10.4s, v6.4s\n"
- "sqadd v11.4s, v11.4s, v7.4s\n"
- "and v4.16b, v12.16b, v0.16b\n"
- "and v5.16b, v13.16b, v1.16b\n"
- "and v6.16b, v14.16b, v2.16b\n"
- "and v7.16b, v15.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
"sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v12.4s, v12.4s, v4.4s\n"
- "sqadd v13.4s, v13.4s, v5.4s\n"
- "sqadd v14.4s, v14.4s, v6.4s\n"
- "sqadd v15.4s, v15.4s, v7.4s\n"
- "and v4.16b, v16.16b, v0.16b\n"
- "and v5.16b, v17.16b, v1.16b\n"
- "and v6.16b, v18.16b, v2.16b\n"
- "and v7.16b, v19.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
"sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v16.4s, v16.4s, v4.4s\n"
- "sqadd v17.4s, v17.4s, v5.4s\n"
- "sqadd v18.4s, v18.4s, v6.4s\n"
- "sqadd v19.4s, v19.4s, v7.4s\n"
- "and v4.16b, v20.16b, v0.16b\n"
- "and v5.16b, v21.16b, v1.16b\n"
- "and v6.16b, v22.16b, v2.16b\n"
- "and v7.16b, v23.16b, v3.16b\n"
"sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v6.4s, v6.4s, #0x1f\n"
+ "sqadd v8.4s, v8.4s, v7.4s\n"
+ "sqadd v9.4s, v9.4s, v6.4s\n"
+ "sqadd v10.4s, v10.4s, v5.4s\n"
+ "sqadd v11.4s, v11.4s, v4.4s\n"
+ "and v7.16b, v12.16b, v0.16b\n"
+ "and v6.16b, v13.16b, v1.16b\n"
+ "and v5.16b, v14.16b, v2.16b\n"
+ "and v4.16b, v15.16b, v3.16b\n"
"sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v20.4s, v20.4s, v4.4s\n"
- "sqadd v21.4s, v21.4s, v5.4s\n"
- "sqadd v22.4s, v22.4s, v6.4s\n"
- "sqadd v23.4s, v23.4s, v7.4s\n"
- "and v4.16b, v24.16b, v0.16b\n"
- "and v5.16b, v25.16b, v1.16b\n"
- "and v6.16b, v26.16b, v2.16b\n"
- "and v7.16b, v27.16b, v3.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
"sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "sqadd v12.4s, v12.4s, v7.4s\n"
+ "sqadd v13.4s, v13.4s, v6.4s\n"
+ "sqadd v14.4s, v14.4s, v5.4s\n"
+ "sqadd v15.4s, v15.4s, v4.4s\n"
+ "and v7.16b, v16.16b, v0.16b\n"
+ "and v6.16b, v17.16b, v1.16b\n"
+ "and v5.16b, v18.16b, v2.16b\n"
+ "and v4.16b, v19.16b, v3.16b\n"
"sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v24.4s, v24.4s, v4.4s\n"
- "sqadd v25.4s, v25.4s, v5.4s\n"
- "sqadd v26.4s, v26.4s, v6.4s\n"
- "sqadd v27.4s, v27.4s, v7.4s\n"
- "and v4.16b, v28.16b, v0.16b\n"
- "and v5.16b, v29.16b, v1.16b\n"
- "and v6.16b, v30.16b, v2.16b\n"
- "and v7.16b, v31.16b, v3.16b\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
"sshr v4.4s, v4.4s, #0x1f\n"
+ "sqadd v16.4s, v16.4s, v7.4s\n"
+ "sqadd v17.4s, v17.4s, v6.4s\n"
+ "sqadd v18.4s, v18.4s, v5.4s\n"
+ "sqadd v19.4s, v19.4s, v4.4s\n"
+ "and v7.16b, v20.16b, v0.16b\n"
+ "and v6.16b, v21.16b, v1.16b\n"
+ "and v5.16b, v22.16b, v2.16b\n"
+ "and v4.16b, v23.16b, v3.16b\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
"sshr v5.4s, v5.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v7.4s\n"
+ "sqadd v21.4s, v21.4s, v6.4s\n"
+ "sqadd v22.4s, v22.4s, v5.4s\n"
+ "sqadd v23.4s, v23.4s, v4.4s\n"
+ "and v7.16b, v24.16b, v0.16b\n"
+ "and v6.16b, v25.16b, v1.16b\n"
+ "and v5.16b, v26.16b, v2.16b\n"
+ "and v4.16b, v27.16b, v3.16b\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
"sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "sqadd v24.4s, v24.4s, v7.4s\n"
+ "sqadd v25.4s, v25.4s, v6.4s\n"
+ "sqadd v26.4s, v26.4s, v5.4s\n"
+ "sqadd v27.4s, v27.4s, v4.4s\n"
+ "and v7.16b, v28.16b, v0.16b\n"
+ "and v6.16b, v29.16b, v1.16b\n"
+ "and v5.16b, v30.16b, v2.16b\n"
+ "and v4.16b, v31.16b, v3.16b\n"
"sshr v7.4s, v7.4s, #0x1f\n"
- "sqadd v28.4s, v28.4s, v4.4s\n"
- "sqadd v29.4s, v29.4s, v5.4s\n"
- "sqadd v30.4s, v30.4s, v6.4s\n"
- "sqadd v31.4s, v31.4s, v7.4s\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "sqadd v28.4s, v28.4s, v7.4s\n"
+ "sqadd v29.4s, v29.4s, v6.4s\n"
+ "sqadd v30.4s, v30.4s, v5.4s\n"
+ "sqadd v31.4s, v31.4s, v4.4s\n"
"152:" // Height 6: no shift correction
- "add x25, %x[qp], %[c_offset]\n"
- "ld1r { v4.4s }, [x25]\n"
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v6.4s }, [x20]\n"
"srshl v8.4s, v8.4s, v0.4s\n"
"srshl v9.4s, v9.4s, v1.4s\n"
"srshl v10.4s, v10.4s, v2.4s\n"
"srshl v11.4s, v11.4s, v3.4s\n"
- "add x25, %x[qp], %[maxval]\n"
- "ld1r { v6.4s }, [x25]\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v5.4s }, [x20]\n"
"srshl v12.4s, v12.4s, v0.4s\n"
"srshl v13.4s, v13.4s, v1.4s\n"
- "add x25, %x[qp], %[minval]\n"
- "ld1r { v5.4s }, [x25]\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v4.4s }, [x20]\n"
"srshl v14.4s, v14.4s, v2.4s\n"
"srshl v15.4s, v15.4s, v3.4s\n"
"cmp x10, #0x10\n"
@@ -3293,225 +3292,225 @@ void a64_hybrid_s8qs_dot_6x16 (
"srshl v29.4s, v29.4s, v1.4s\n"
"srshl v30.4s, v30.4s, v2.4s\n"
"srshl v31.4s, v31.4s, v3.4s\n"
- "add v8.4s, v8.4s, v4.4s\n"
- "add v9.4s, v9.4s, v4.4s\n"
- "add v10.4s, v10.4s, v4.4s\n"
- "add v11.4s, v11.4s, v4.4s\n"
- "add v12.4s, v12.4s, v4.4s\n"
- "add v13.4s, v13.4s, v4.4s\n"
- "add v14.4s, v14.4s, v4.4s\n"
- "add v15.4s, v15.4s, v4.4s\n"
- "add v16.4s, v16.4s, v4.4s\n"
- "add v17.4s, v17.4s, v4.4s\n"
- "add v18.4s, v18.4s, v4.4s\n"
- "add v19.4s, v19.4s, v4.4s\n"
- "add v20.4s, v20.4s, v4.4s\n"
- "add v21.4s, v21.4s, v4.4s\n"
- "add v22.4s, v22.4s, v4.4s\n"
- "add v23.4s, v23.4s, v4.4s\n"
- "add v24.4s, v24.4s, v4.4s\n"
- "add v25.4s, v25.4s, v4.4s\n"
- "add v26.4s, v26.4s, v4.4s\n"
- "add v27.4s, v27.4s, v4.4s\n"
- "add v28.4s, v28.4s, v4.4s\n"
- "add v29.4s, v29.4s, v4.4s\n"
- "add v30.4s, v30.4s, v4.4s\n"
- "add v31.4s, v31.4s, v4.4s\n"
- "smin v8.4s, v8.4s, v6.4s\n"
- "smin v9.4s, v9.4s, v6.4s\n"
- "smin v10.4s, v10.4s, v6.4s\n"
- "smin v11.4s, v11.4s, v6.4s\n"
- "smin v12.4s, v12.4s, v6.4s\n"
- "smin v13.4s, v13.4s, v6.4s\n"
- "smin v14.4s, v14.4s, v6.4s\n"
- "smin v15.4s, v15.4s, v6.4s\n"
- "smin v16.4s, v16.4s, v6.4s\n"
- "smin v17.4s, v17.4s, v6.4s\n"
- "smin v18.4s, v18.4s, v6.4s\n"
- "smin v19.4s, v19.4s, v6.4s\n"
- "smin v20.4s, v20.4s, v6.4s\n"
- "smin v21.4s, v21.4s, v6.4s\n"
- "smin v22.4s, v22.4s, v6.4s\n"
- "smin v23.4s, v23.4s, v6.4s\n"
- "smin v24.4s, v24.4s, v6.4s\n"
- "smin v25.4s, v25.4s, v6.4s\n"
- "smin v26.4s, v26.4s, v6.4s\n"
- "smin v27.4s, v27.4s, v6.4s\n"
- "smin v28.4s, v28.4s, v6.4s\n"
- "smin v29.4s, v29.4s, v6.4s\n"
- "smin v30.4s, v30.4s, v6.4s\n"
- "smin v31.4s, v31.4s, v6.4s\n"
- "smax v8.4s, v8.4s, v5.4s\n"
- "smax v9.4s, v9.4s, v5.4s\n"
- "smax v10.4s, v10.4s, v5.4s\n"
- "smax v11.4s, v11.4s, v5.4s\n"
- "smax v12.4s, v12.4s, v5.4s\n"
- "smax v13.4s, v13.4s, v5.4s\n"
- "smax v14.4s, v14.4s, v5.4s\n"
- "smax v15.4s, v15.4s, v5.4s\n"
- "smax v16.4s, v16.4s, v5.4s\n"
- "smax v17.4s, v17.4s, v5.4s\n"
- "smax v18.4s, v18.4s, v5.4s\n"
- "smax v19.4s, v19.4s, v5.4s\n"
- "smax v20.4s, v20.4s, v5.4s\n"
- "smax v21.4s, v21.4s, v5.4s\n"
- "smax v22.4s, v22.4s, v5.4s\n"
- "smax v23.4s, v23.4s, v5.4s\n"
- "smax v24.4s, v24.4s, v5.4s\n"
- "smax v25.4s, v25.4s, v5.4s\n"
- "smax v26.4s, v26.4s, v5.4s\n"
- "smax v27.4s, v27.4s, v5.4s\n"
- "smax v28.4s, v28.4s, v5.4s\n"
- "smax v29.4s, v29.4s, v5.4s\n"
- "smax v30.4s, v30.4s, v5.4s\n"
- "smax v31.4s, v31.4s, v5.4s\n"
+ "add v8.4s, v8.4s, v6.4s\n"
+ "add v9.4s, v9.4s, v6.4s\n"
+ "add v10.4s, v10.4s, v6.4s\n"
+ "add v11.4s, v11.4s, v6.4s\n"
+ "add v12.4s, v12.4s, v6.4s\n"
+ "add v13.4s, v13.4s, v6.4s\n"
+ "add v14.4s, v14.4s, v6.4s\n"
+ "add v15.4s, v15.4s, v6.4s\n"
+ "add v16.4s, v16.4s, v6.4s\n"
+ "add v17.4s, v17.4s, v6.4s\n"
+ "add v18.4s, v18.4s, v6.4s\n"
+ "add v19.4s, v19.4s, v6.4s\n"
+ "add v20.4s, v20.4s, v6.4s\n"
+ "add v21.4s, v21.4s, v6.4s\n"
+ "add v22.4s, v22.4s, v6.4s\n"
+ "add v23.4s, v23.4s, v6.4s\n"
+ "add v24.4s, v24.4s, v6.4s\n"
+ "add v25.4s, v25.4s, v6.4s\n"
+ "add v26.4s, v26.4s, v6.4s\n"
+ "add v27.4s, v27.4s, v6.4s\n"
+ "add v28.4s, v28.4s, v6.4s\n"
+ "add v29.4s, v29.4s, v6.4s\n"
+ "add v30.4s, v30.4s, v6.4s\n"
+ "add v31.4s, v31.4s, v6.4s\n"
+ "smin v8.4s, v8.4s, v5.4s\n"
+ "smin v9.4s, v9.4s, v5.4s\n"
+ "smin v10.4s, v10.4s, v5.4s\n"
+ "smin v11.4s, v11.4s, v5.4s\n"
+ "smin v12.4s, v12.4s, v5.4s\n"
+ "smin v13.4s, v13.4s, v5.4s\n"
+ "smin v14.4s, v14.4s, v5.4s\n"
+ "smin v15.4s, v15.4s, v5.4s\n"
+ "smin v16.4s, v16.4s, v5.4s\n"
+ "smin v17.4s, v17.4s, v5.4s\n"
+ "smin v18.4s, v18.4s, v5.4s\n"
+ "smin v19.4s, v19.4s, v5.4s\n"
+ "smin v20.4s, v20.4s, v5.4s\n"
+ "smin v21.4s, v21.4s, v5.4s\n"
+ "smin v22.4s, v22.4s, v5.4s\n"
+ "smin v23.4s, v23.4s, v5.4s\n"
+ "smin v24.4s, v24.4s, v5.4s\n"
+ "smin v25.4s, v25.4s, v5.4s\n"
+ "smin v26.4s, v26.4s, v5.4s\n"
+ "smin v27.4s, v27.4s, v5.4s\n"
+ "smin v28.4s, v28.4s, v5.4s\n"
+ "smin v29.4s, v29.4s, v5.4s\n"
+ "smin v30.4s, v30.4s, v5.4s\n"
+ "smin v31.4s, v31.4s, v5.4s\n"
+ "smax v8.4s, v8.4s, v4.4s\n"
+ "smax v9.4s, v9.4s, v4.4s\n"
+ "smax v10.4s, v10.4s, v4.4s\n"
+ "smax v11.4s, v11.4s, v4.4s\n"
+ "smax v12.4s, v12.4s, v4.4s\n"
+ "smax v13.4s, v13.4s, v4.4s\n"
+ "smax v14.4s, v14.4s, v4.4s\n"
+ "smax v15.4s, v15.4s, v4.4s\n"
+ "smax v16.4s, v16.4s, v4.4s\n"
+ "smax v17.4s, v17.4s, v4.4s\n"
+ "smax v18.4s, v18.4s, v4.4s\n"
+ "smax v19.4s, v19.4s, v4.4s\n"
+ "smax v20.4s, v20.4s, v4.4s\n"
+ "smax v21.4s, v21.4s, v4.4s\n"
+ "smax v22.4s, v22.4s, v4.4s\n"
+ "smax v23.4s, v23.4s, v4.4s\n"
+ "smax v24.4s, v24.4s, v4.4s\n"
+ "smax v25.4s, v25.4s, v4.4s\n"
+ "smax v26.4s, v26.4s, v4.4s\n"
+ "smax v27.4s, v27.4s, v4.4s\n"
+ "smax v28.4s, v28.4s, v4.4s\n"
+ "smax v29.4s, v29.4s, v4.4s\n"
+ "smax v30.4s, v30.4s, v4.4s\n"
+ "smax v31.4s, v31.4s, v4.4s\n"
"uzp1 v8.8h, v8.8h, v9.8h\n"
- "uzp1 v9.8h, v10.8h, v11.8h\n"
+ "uzp1 v2.8h, v10.8h, v11.8h\n"
"uzp1 v12.8h, v12.8h, v13.8h\n"
- "uzp1 v13.8h, v14.8h, v15.8h\n"
+ "uzp1 v1.8h, v14.8h, v15.8h\n"
"uzp1 v16.8h, v16.8h, v17.8h\n"
- "uzp1 v17.8h, v18.8h, v19.8h\n"
+ "uzp1 v0.8h, v18.8h, v19.8h\n"
"uzp1 v20.8h, v20.8h, v21.8h\n"
- "uzp1 v21.8h, v22.8h, v23.8h\n"
+ "uzp1 v19.8h, v22.8h, v23.8h\n"
"uzp1 v24.8h, v24.8h, v25.8h\n"
- "uzp1 v25.8h, v26.8h, v27.8h\n"
+ "uzp1 v18.8h, v26.8h, v27.8h\n"
"uzp1 v28.8h, v28.8h, v29.8h\n"
- "uzp1 v29.8h, v30.8h, v31.8h\n"
- "uzp1 v8.16b, v8.16b, v9.16b\n"
- "uzp1 v12.16b, v12.16b, v13.16b\n"
- "uzp1 v16.16b, v16.16b, v17.16b\n"
- "uzp1 v20.16b, v20.16b, v21.16b\n"
- "uzp1 v24.16b, v24.16b, v25.16b\n"
- "uzp1 v28.16b, v28.16b, v29.16b\n"
+ "uzp1 v17.8h, v30.8h, v31.8h\n"
+ "uzp1 v8.16b, v8.16b, v2.16b\n"
+ "uzp1 v12.16b, v12.16b, v1.16b\n"
+ "uzp1 v16.16b, v16.16b, v0.16b\n"
+ "uzp1 v20.16b, v20.16b, v19.16b\n"
+ "uzp1 v24.16b, v24.16b, v18.16b\n"
+ "uzp1 v28.16b, v28.16b, v17.16b\n"
"bge 161f\n"
"tbz x10, #3, 156f\n"
"str d8, [x11], #0x8\n"
- "str d12, [x24], #0x8\n"
- "str d16, [x23], #0x8\n"
- "str d20, [x22], #0x8\n"
- "str d24, [x21], #0x8\n"
- "str d28, [x20], #0x8\n"
+ "str d12, [x25], #0x8\n"
+ "str d16, [x24], #0x8\n"
+ "str d20, [x23], #0x8\n"
+ "str d24, [x22], #0x8\n"
+ "str d28, [x21], #0x8\n"
"tbz x10, #2, 154f\n"
"st1 { v8.s }[2], [x11], #0x4\n"
- "st1 { v12.s }[2], [x24], #0x4\n"
- "st1 { v16.s }[2], [x23], #0x4\n"
- "st1 { v20.s }[2], [x22], #0x4\n"
- "st1 { v24.s }[2], [x21], #0x4\n"
- "st1 { v28.s }[2], [x20], #0x4\n"
+ "st1 { v12.s }[2], [x25], #0x4\n"
+ "st1 { v16.s }[2], [x24], #0x4\n"
+ "st1 { v20.s }[2], [x23], #0x4\n"
+ "st1 { v24.s }[2], [x22], #0x4\n"
+ "st1 { v28.s }[2], [x21], #0x4\n"
"tbz x10, #1, 153f\n"
"st1 { v8.h }[6], [x11], #0x2\n"
- "st1 { v12.h }[6], [x24], #0x2\n"
- "st1 { v16.h }[6], [x23], #0x2\n"
- "st1 { v20.h }[6], [x22], #0x2\n"
- "st1 { v24.h }[6], [x21], #0x2\n"
- "st1 { v28.h }[6], [x20], #0x2\n"
+ "st1 { v12.h }[6], [x25], #0x2\n"
+ "st1 { v16.h }[6], [x24], #0x2\n"
+ "st1 { v20.h }[6], [x23], #0x2\n"
+ "st1 { v24.h }[6], [x22], #0x2\n"
+ "st1 { v28.h }[6], [x21], #0x2\n"
"tbz x10, #0, 160f\n"
"st1 { v8.b }[14], [x11]\n"
- "st1 { v12.b }[14], [x24]\n"
- "st1 { v16.b }[14], [x23]\n"
- "st1 { v20.b }[14], [x22]\n"
- "st1 { v24.b }[14], [x21]\n"
- "st1 { v28.b }[14], [x20]\n"
+ "st1 { v12.b }[14], [x25]\n"
+ "st1 { v16.b }[14], [x24]\n"
+ "st1 { v20.b }[14], [x23]\n"
+ "st1 { v24.b }[14], [x22]\n"
+ "st1 { v28.b }[14], [x21]\n"
"b 160f\n"
"153:" // Height 6: Partial direct writeback: partial_1_12
"tbz x10, #0, 160f\n"
"st1 { v8.b }[12], [x11]\n"
- "st1 { v12.b }[12], [x24]\n"
- "st1 { v16.b }[12], [x23]\n"
- "st1 { v20.b }[12], [x22]\n"
- "st1 { v24.b }[12], [x21]\n"
- "st1 { v28.b }[12], [x20]\n"
+ "st1 { v12.b }[12], [x25]\n"
+ "st1 { v16.b }[12], [x24]\n"
+ "st1 { v20.b }[12], [x23]\n"
+ "st1 { v24.b }[12], [x22]\n"
+ "st1 { v28.b }[12], [x21]\n"
"b 160f\n"
"154:" // Height 6: Partial direct writeback: partial_2_8
"tbz x10, #1, 155f\n"
"st1 { v8.h }[4], [x11], #0x2\n"
- "st1 { v12.h }[4], [x24], #0x2\n"
- "st1 { v16.h }[4], [x23], #0x2\n"
- "st1 { v20.h }[4], [x22], #0x2\n"
- "st1 { v24.h }[4], [x21], #0x2\n"
- "st1 { v28.h }[4], [x20], #0x2\n"
+ "st1 { v12.h }[4], [x25], #0x2\n"
+ "st1 { v16.h }[4], [x24], #0x2\n"
+ "st1 { v20.h }[4], [x23], #0x2\n"
+ "st1 { v24.h }[4], [x22], #0x2\n"
+ "st1 { v28.h }[4], [x21], #0x2\n"
"tbz x10, #0, 160f\n"
"st1 { v8.b }[10], [x11]\n"
- "st1 { v12.b }[10], [x24]\n"
- "st1 { v16.b }[10], [x23]\n"
- "st1 { v20.b }[10], [x22]\n"
- "st1 { v24.b }[10], [x21]\n"
- "st1 { v28.b }[10], [x20]\n"
+ "st1 { v12.b }[10], [x25]\n"
+ "st1 { v16.b }[10], [x24]\n"
+ "st1 { v20.b }[10], [x23]\n"
+ "st1 { v24.b }[10], [x22]\n"
+ "st1 { v28.b }[10], [x21]\n"
"b 160f\n"
"155:" // Height 6: Partial direct writeback: partial_1_8
"tbz x10, #0, 160f\n"
"st1 { v8.b }[8], [x11]\n"
- "st1 { v12.b }[8], [x24]\n"
- "st1 { v16.b }[8], [x23]\n"
- "st1 { v20.b }[8], [x22]\n"
- "st1 { v24.b }[8], [x21]\n"
- "st1 { v28.b }[8], [x20]\n"
+ "st1 { v12.b }[8], [x25]\n"
+ "st1 { v16.b }[8], [x24]\n"
+ "st1 { v20.b }[8], [x23]\n"
+ "st1 { v24.b }[8], [x22]\n"
+ "st1 { v28.b }[8], [x21]\n"
"b 160f\n"
"156:" // Height 6: Partial direct writeback: partial_4_0
"tbz x10, #2, 158f\n"
"str s8, [x11], #0x4\n"
- "str s12, [x24], #0x4\n"
- "str s16, [x23], #0x4\n"
- "str s20, [x22], #0x4\n"
- "str s24, [x21], #0x4\n"
- "str s28, [x20], #0x4\n"
+ "str s12, [x25], #0x4\n"
+ "str s16, [x24], #0x4\n"
+ "str s20, [x23], #0x4\n"
+ "str s24, [x22], #0x4\n"
+ "str s28, [x21], #0x4\n"
"tbz x10, #1, 157f\n"
"st1 { v8.h }[2], [x11], #0x2\n"
- "st1 { v12.h }[2], [x24], #0x2\n"
- "st1 { v16.h }[2], [x23], #0x2\n"
- "st1 { v20.h }[2], [x22], #0x2\n"
- "st1 { v24.h }[2], [x21], #0x2\n"
- "st1 { v28.h }[2], [x20], #0x2\n"
+ "st1 { v12.h }[2], [x25], #0x2\n"
+ "st1 { v16.h }[2], [x24], #0x2\n"
+ "st1 { v20.h }[2], [x23], #0x2\n"
+ "st1 { v24.h }[2], [x22], #0x2\n"
+ "st1 { v28.h }[2], [x21], #0x2\n"
"tbz x10, #0, 160f\n"
"st1 { v8.b }[6], [x11]\n"
- "st1 { v12.b }[6], [x24]\n"
- "st1 { v16.b }[6], [x23]\n"
- "st1 { v20.b }[6], [x22]\n"
- "st1 { v24.b }[6], [x21]\n"
- "st1 { v28.b }[6], [x20]\n"
+ "st1 { v12.b }[6], [x25]\n"
+ "st1 { v16.b }[6], [x24]\n"
+ "st1 { v20.b }[6], [x23]\n"
+ "st1 { v24.b }[6], [x22]\n"
+ "st1 { v28.b }[6], [x21]\n"
"b 160f\n"
"157:" // Height 6: Partial direct writeback: partial_1_4
"tbz x10, #0, 160f\n"
"st1 { v8.b }[4], [x11]\n"
- "st1 { v12.b }[4], [x24]\n"
- "st1 { v16.b }[4], [x23]\n"
- "st1 { v20.b }[4], [x22]\n"
- "st1 { v24.b }[4], [x21]\n"
- "st1 { v28.b }[4], [x20]\n"
+ "st1 { v12.b }[4], [x25]\n"
+ "st1 { v16.b }[4], [x24]\n"
+ "st1 { v20.b }[4], [x23]\n"
+ "st1 { v24.b }[4], [x22]\n"
+ "st1 { v28.b }[4], [x21]\n"
"b 160f\n"
"158:" // Height 6: Partial direct writeback: partial_2_0
"tbz x10, #1, 159f\n"
"str h8, [x11], #0x2\n"
- "str h12, [x24], #0x2\n"
- "str h16, [x23], #0x2\n"
- "str h20, [x22], #0x2\n"
- "str h24, [x21], #0x2\n"
- "str h28, [x20], #0x2\n"
+ "str h12, [x25], #0x2\n"
+ "str h16, [x24], #0x2\n"
+ "str h20, [x23], #0x2\n"
+ "str h24, [x22], #0x2\n"
+ "str h28, [x21], #0x2\n"
"tbz x10, #0, 160f\n"
"st1 { v8.b }[2], [x11]\n"
- "st1 { v12.b }[2], [x24]\n"
- "st1 { v16.b }[2], [x23]\n"
- "st1 { v20.b }[2], [x22]\n"
- "st1 { v24.b }[2], [x21]\n"
- "st1 { v28.b }[2], [x20]\n"
+ "st1 { v12.b }[2], [x25]\n"
+ "st1 { v16.b }[2], [x24]\n"
+ "st1 { v20.b }[2], [x23]\n"
+ "st1 { v24.b }[2], [x22]\n"
+ "st1 { v28.b }[2], [x21]\n"
"b 160f\n"
"159:" // Height 6: Partial direct writeback: partial_1_0
"str b8, [x11, #0x0]\n"
- "str b12, [x24, #0x0]\n"
- "str b16, [x23, #0x0]\n"
- "str b20, [x22, #0x0]\n"
- "str b24, [x21, #0x0]\n"
- "str b28, [x20, #0x0]\n"
+ "str b12, [x25, #0x0]\n"
+ "str b16, [x24, #0x0]\n"
+ "str b20, [x23, #0x0]\n"
+ "str b24, [x22, #0x0]\n"
+ "str b28, [x21, #0x0]\n"
"160:" // Height 6: Partial direct writeback: Done
"b 162f\n"
"161:" // Height 6: Full writeback
"str q8, [x11, #0x0]\n"
"add x11, x11, #0x10\n"
- "str q12, [x24, #0x0]\n"
- "str q16, [x23, #0x0]\n"
- "str q20, [x22, #0x0]\n"
- "str q24, [x21, #0x0]\n"
- "str q28, [x20, #0x0]\n"
+ "str q12, [x25, #0x0]\n"
+ "str q16, [x24, #0x0]\n"
+ "str q20, [x23, #0x0]\n"
+ "str q24, [x22, #0x0]\n"
+ "str q28, [x21, #0x0]\n"
"162:" // Height 6: Writeback done
"subs x10, x10, #0x10\n"
"bgt 137b\n"
@@ -3527,7 +3526,6 @@ void a64_hybrid_s8qs_dot_6x16 (
"madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
"b 1b\n"
"164:" // Exit
-
: [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
: [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, 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_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
: "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", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"