aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp4077
1 files changed, 2038 insertions, 2039 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
index 19636548a0..b049ed45f9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
@@ -244,11 +244,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"23:" // Height 1: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w14, [x20, x15, 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, 24f\n"
- "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x13, [x21, #0x0]\n"
+ "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x13, [x20, #0x0]\n"
"cbnz x15, 25f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x13, x13, x20, LSL #1\n"
@@ -265,222 +265,222 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"blt 27f\n"
"26:" // Height 1: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr d6, [x17, #0x20]\n"
- "ldr x12, [x17, #0x28]\n"
+ "ldr d17, [x17, #0x20]\n"
+ "ldr x20, [x17, #0x28]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "ldr d7, [x17, #0x30]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x38]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "ldr d6, [x17, #0x40]\n"
- "ldr x12, [x17, #0x48]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "ldr d7, [x17, #0x50]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x58]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "ldr d6, [x17, #0x60]\n"
- "ldr x12, [x17, #0x68]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "ldr d7, [x17, #0x70]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x78]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "ldr d6, [x17, #0x80]\n"
- "ldr x12, [x17, #0x88]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "ldr d7, [x17, #0x90]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x98]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "ldr d6, [x17, #0xa0]\n"
- "ldr x12, [x17, #0xa8]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "ldr d7, [x17, #0xb0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0xb8]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "ldr d6, [x17, #0xc0]\n"
- "ldr x12, [x17, #0xc8]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "ldr d7, [x17, #0xd0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0xd8]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "ldr d6, [x17, #0xe0]\n"
- "ldr x12, [x17, #0xe8]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "ldr d7, [x17, #0xf0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0xf8]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "ldr d6, [x17, #0x100]\n"
- "ldr x12, [x17, #0x108]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "ldr d7, [x17, #0x110]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x118]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "ldr d6, [x17, #0x120]\n"
- "ldr x12, [x17, #0x128]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "ldr d7, [x17, #0x130]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x138]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "ldr d6, [x17, #0x140]\n"
- "ldr x12, [x17, #0x148]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "ldr d7, [x17, #0x150]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x158]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "ldr d6, [x17, #0x160]\n"
- "ldr x12, [x17, #0x168]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "ldr d7, [x17, #0x170]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x178]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "ldr d6, [x17, #0x180]\n"
- "ldr x12, [x17, #0x188]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "ldr d7, [x17, #0x190]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x198]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "ldr d6, [x17, #0x1a0]\n"
- "ldr x12, [x17, #0x1a8]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "ldr d7, [x17, #0x1b0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x1b8]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "ldr d6, [x17, #0x1c0]\n"
- "ldr x12, [x17, #0x1c8]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "ldr d7, [x17, #0x1d0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x1d8]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "ldr d6, [x17, #0x1e0]\n"
- "ldr x12, [x17, #0x1e8]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "ldr d7, [x17, #0x1f0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x1f8]\n"
- "mov v7.d[1], x11\n"
+ "ldr d16, [x17, #0x30]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x38]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "ldr d17, [x17, #0x40]\n"
+ "ldr x20, [x17, #0x48]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "ldr d16, [x17, #0x50]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x58]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "ldr d17, [x17, #0x60]\n"
+ "ldr x20, [x17, #0x68]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "ldr d16, [x17, #0x70]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x78]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "ldr d17, [x17, #0x80]\n"
+ "ldr x20, [x17, #0x88]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "ldr d16, [x17, #0x90]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x98]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "ldr d17, [x17, #0xa0]\n"
+ "ldr x20, [x17, #0xa8]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "ldr d16, [x17, #0xb0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0xb8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "ldr d17, [x17, #0xc0]\n"
+ "ldr x20, [x17, #0xc8]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "ldr d16, [x17, #0xd0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0xd8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "ldr d17, [x17, #0xe0]\n"
+ "ldr x20, [x17, #0xe8]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "ldr d16, [x17, #0xf0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0xf8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "ldr d17, [x17, #0x100]\n"
+ "ldr x20, [x17, #0x108]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "ldr d16, [x17, #0x110]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x118]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "ldr d17, [x17, #0x120]\n"
+ "ldr x20, [x17, #0x128]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "ldr d16, [x17, #0x130]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x138]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "ldr d17, [x17, #0x140]\n"
+ "ldr x20, [x17, #0x148]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "ldr d16, [x17, #0x150]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x158]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "ldr d17, [x17, #0x160]\n"
+ "ldr x20, [x17, #0x168]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "ldr d16, [x17, #0x170]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x178]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "ldr d17, [x17, #0x180]\n"
+ "ldr x20, [x17, #0x188]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "ldr d16, [x17, #0x190]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x198]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "ldr d17, [x17, #0x1a0]\n"
+ "ldr x20, [x17, #0x1a8]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "ldr d16, [x17, #0x1b0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x1b8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "ldr d17, [x17, #0x1c0]\n"
+ "ldr x20, [x17, #0x1c8]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "ldr d16, [x17, #0x1d0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x1d8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "ldr d17, [x17, #0x1e0]\n"
+ "ldr x20, [x17, #0x1e8]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "ldr d16, [x17, #0x1f0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x1f8]\n"
+ "mov v16.d[1], x20\n"
"add x13, x13, #0x10\n"
"add x17, x17, #0x200\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
+ "fmla v10.8h, v17.8h, v0.h[7]\n"
"ldr d6, [x17, #0x0]\n"
- "ldr x12, [x17, #0x8]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "ldr x20, [x17, #0x8]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
"ldr d0, [x13, #0x0]\n"
"sub x14, x14, #0x8\n"
"ldr d7, [x17, #0x10]\n"
"cmp x14, #0x10\n"
- "ldr x10, [x13, #0x8]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x18]\n"
- "mov v0.d[1], x10\n"
- "mov v7.d[1], x11\n"
+ "ldr x21, [x13, #0x8]\n"
+ "mov v6.d[1], x20\n"
+ "ldr x20, [x17, #0x18]\n"
+ "mov v0.d[1], x21\n"
+ "mov v7.d[1], x20\n"
"prfm pldl1keep, [x13, #0x80]\n"
"bge 26b\n"
"27:" // Height 1: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
+ "ldr q17, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "ldr q6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "ldr q6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "ldr q7, [x17, #0x70]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "ldr q6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "ldr q7, [x17, #0x90]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "ldr q6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "ldr q7, [x17, #0xb0]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "ldr q6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "ldr q7, [x17, #0xd0]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "ldr q6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "ldr q7, [x17, #0xf0]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "ldr q6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "ldr q7, [x17, #0x110]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "ldr q6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "ldr q7, [x17, #0x130]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "ldr q6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "ldr q7, [x17, #0x150]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "ldr q6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "ldr q7, [x17, #0x170]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "ldr q6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "ldr q7, [x17, #0x190]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "ldr q6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "ldr q7, [x17, #0x1b0]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "ldr q6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "ldr q7, [x17, #0x1d0]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "ldr q6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "ldr q7, [x17, #0x1f0]\n"
+ "ldr q16, [x17, #0x30]\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "ldr q17, [x17, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x50]\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "ldr q17, [x17, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "ldr q16, [x17, #0x70]\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "ldr q17, [x17, #0x80]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "ldr q16, [x17, #0x90]\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "ldr q17, [x17, #0xa0]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "ldr q16, [x17, #0xb0]\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "ldr q17, [x17, #0xc0]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "ldr q16, [x17, #0xd0]\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "ldr q17, [x17, #0xe0]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "ldr q16, [x17, #0xf0]\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "ldr q17, [x17, #0x100]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "ldr q16, [x17, #0x110]\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "ldr q17, [x17, #0x120]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "ldr q16, [x17, #0x130]\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "ldr q17, [x17, #0x140]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "ldr q16, [x17, #0x150]\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "ldr q17, [x17, #0x160]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "ldr q16, [x17, #0x170]\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "ldr q17, [x17, #0x180]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "ldr q16, [x17, #0x190]\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "ldr q17, [x17, #0x1a0]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "ldr q16, [x17, #0x1b0]\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "ldr q17, [x17, #0x1c0]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "ldr q16, [x17, #0x1d0]\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "ldr q17, [x17, #0x1e0]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "ldr q16, [x17, #0x1f0]\n"
"add x13, x13, #0x10\n"
"sub x14, x14, #0x8\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
+ "fmla v10.8h, v17.8h, v0.h[7]\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
"add x17, x17, #0x200\n"
"28:" // Height 1: Multiply loop: Main loop skip
"cbz x14, 30f\n"
"29:" // Height 1: Multiply loop: Odd block loop
"ldr h0, [x13], #0x2\n"
"sub x14, x14, #0x1\n"
- "ldr q6, [x17, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x10]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x0]\n"
+ "fmla v8.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x10]\n"
+ "fmla v9.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x20]\n"
+ "fmla v10.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x30]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
"add x17, x17, #0x40\n"
"cbnz x14, 29b\n"
"30:" // Height 1: Multiply loop: No odd multiplies
@@ -491,17 +491,17 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"prfm pstl1keep, [x16, #0x0]\n"
"tbz %x[flags], #1, 31f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v0.8h\n"
- "fmin v9.8h, v9.8h, v0.8h\n"
- "fmin v10.8h, v10.8h, v0.8h\n"
- "fmin v11.8h, v11.8h, v0.8h\n"
+ "ld1r { v16.8h }, [x20]\n"
+ "fmin v8.8h, v8.8h, v16.8h\n"
+ "fmin v9.8h, v9.8h, v16.8h\n"
+ "fmin v10.8h, v10.8h, v16.8h\n"
+ "fmin v11.8h, v11.8h, v16.8h\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmax v8.8h, v8.8h, v0.8h\n"
- "fmax v9.8h, v9.8h, v0.8h\n"
- "fmax v10.8h, v10.8h, v0.8h\n"
- "fmax v11.8h, v11.8h, v0.8h\n"
+ "ld1r { v16.8h }, [x20]\n"
+ "fmax v8.8h, v8.8h, v16.8h\n"
+ "fmax v9.8h, v9.8h, v16.8h\n"
+ "fmax v10.8h, v10.8h, v16.8h\n"
+ "fmax v11.8h, v11.8h, v16.8h\n"
"31:" // Height 1: No activation
"cmp x8, #0x20\n"
"bge 48f\n"
@@ -799,324 +799,324 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"72:" // Height 2: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w14, [x20, x15, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 73f\n"
- "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x13, [x21, #0x0]\n"
- "ldr x9, [x21, #0x8]\n"
+ "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x13, [x20, #0x0]\n"
+ "ldr x12, [x20, #0x8]\n"
"cbnz x15, 74f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x13, x13, x20, LSL #1\n"
- "add x9, x9, x20, LSL #1\n"
+ "add x12, x12, x20, LSL #1\n"
"b 74f\n"
"73:" // Height 2: setup direct input
"mov x13, %x[input_ptr]\n"
- "add x9, x13, x20, LSL #1\n"
+ "add x12, x13, x21, LSL #1\n"
"74:" // Height 2: input setup done
"cmp x14, #0x8\n"
"blt 77f\n"
"ldr q0, [x13, #0x0]\n"
"cmp x14, #0x10\n"
- "ldr q1, [x9, #0x0]\n"
+ "ldr q1, [x12, #0x0]\n"
"ldr q6, [x17, #0x0]\n"
"ldr q7, [x17, #0x10]\n"
"blt 76f\n"
"75:" // Height 2: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr x12, [x17, #0x28]\n"
+ "ldr x21, [x17, #0x28]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr d6, [x17, #0x20]\n"
+ "ldr d17, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "ldr x11, [x17, #0x38]\n"
+ "ldr x20, [x17, #0x38]\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr d7, [x17, #0x30]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr d6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "ldr x12, [x17, #0x48]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr d7, [x17, #0x50]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x58]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "ldr x12, [x17, #0x68]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr d6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "ldr x11, [x17, #0x78]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr d7, [x17, #0x70]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr d6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "ldr x12, [x17, #0x88]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr d7, [x17, #0x90]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x98]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "ldr x12, [x17, #0xa8]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr d6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "ldr x11, [x17, #0xb8]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr d7, [x17, #0xb0]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr d6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "ldr x12, [x17, #0xc8]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr d7, [x17, #0xd0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0xd8]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "ldr x12, [x17, #0xe8]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr d6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "ldr x11, [x17, #0xf8]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr d7, [x17, #0xf0]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr d6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "ldr x12, [x17, #0x108]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr d7, [x17, #0x110]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x118]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "ldr x12, [x17, #0x128]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr d6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "ldr x11, [x17, #0x138]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr d7, [x17, #0x130]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr d6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "ldr x12, [x17, #0x148]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr d7, [x17, #0x150]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x158]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "ldr x12, [x17, #0x168]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr d6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "ldr x11, [x17, #0x178]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr d7, [x17, #0x170]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr d6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "ldr x12, [x17, #0x188]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr d7, [x17, #0x190]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x198]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "ldr x12, [x17, #0x1a8]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr d6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "ldr x11, [x17, #0x1b8]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr d7, [x17, #0x1b0]\n"
- "mov v6.d[1], x12\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "mov v7.d[1], x11\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr d6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "ldr x12, [x17, #0x1c8]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr d7, [x17, #0x1d0]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x1d8]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "ldr x12, [x17, #0x1e8]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr d6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "ldr x11, [x17, #0x1f8]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "ldr d7, [x17, #0x1f0]\n"
- "mov v6.d[1], x12\n"
+ "ldr d16, [x17, #0x30]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[0]\n"
+ "ldr d17, [x17, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "ldr x20, [x17, #0x48]\n"
+ "fmla v15.8h, v16.8h, v1.h[0]\n"
+ "ldr d16, [x17, #0x50]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x58]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "ldr x21, [x17, #0x68]\n"
+ "fmla v12.8h, v17.8h, v1.h[1]\n"
+ "ldr d17, [x17, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "ldr x20, [x17, #0x78]\n"
+ "fmla v13.8h, v16.8h, v1.h[1]\n"
+ "ldr d16, [x17, #0x70]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[1]\n"
+ "ldr d17, [x17, #0x80]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "ldr x20, [x17, #0x88]\n"
+ "fmla v15.8h, v16.8h, v1.h[1]\n"
+ "ldr d16, [x17, #0x90]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x98]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "ldr x21, [x17, #0xa8]\n"
+ "fmla v12.8h, v17.8h, v1.h[2]\n"
+ "ldr d17, [x17, #0xa0]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "ldr x20, [x17, #0xb8]\n"
+ "fmla v13.8h, v16.8h, v1.h[2]\n"
+ "ldr d16, [x17, #0xb0]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[2]\n"
+ "ldr d17, [x17, #0xc0]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "ldr x20, [x17, #0xc8]\n"
+ "fmla v15.8h, v16.8h, v1.h[2]\n"
+ "ldr d16, [x17, #0xd0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0xd8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "ldr x21, [x17, #0xe8]\n"
+ "fmla v12.8h, v17.8h, v1.h[3]\n"
+ "ldr d17, [x17, #0xe0]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "ldr x20, [x17, #0xf8]\n"
+ "fmla v13.8h, v16.8h, v1.h[3]\n"
+ "ldr d16, [x17, #0xf0]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[3]\n"
+ "ldr d17, [x17, #0x100]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "ldr x20, [x17, #0x108]\n"
+ "fmla v15.8h, v16.8h, v1.h[3]\n"
+ "ldr d16, [x17, #0x110]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x118]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "ldr x21, [x17, #0x128]\n"
+ "fmla v12.8h, v17.8h, v1.h[4]\n"
+ "ldr d17, [x17, #0x120]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "ldr x20, [x17, #0x138]\n"
+ "fmla v13.8h, v16.8h, v1.h[4]\n"
+ "ldr d16, [x17, #0x130]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[4]\n"
+ "ldr d17, [x17, #0x140]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "ldr x20, [x17, #0x148]\n"
+ "fmla v15.8h, v16.8h, v1.h[4]\n"
+ "ldr d16, [x17, #0x150]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x158]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "ldr x21, [x17, #0x168]\n"
+ "fmla v12.8h, v17.8h, v1.h[5]\n"
+ "ldr d17, [x17, #0x160]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "ldr x20, [x17, #0x178]\n"
+ "fmla v13.8h, v16.8h, v1.h[5]\n"
+ "ldr d16, [x17, #0x170]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[5]\n"
+ "ldr d17, [x17, #0x180]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "ldr x20, [x17, #0x188]\n"
+ "fmla v15.8h, v16.8h, v1.h[5]\n"
+ "ldr d16, [x17, #0x190]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x198]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "ldr x21, [x17, #0x1a8]\n"
+ "fmla v12.8h, v17.8h, v1.h[6]\n"
+ "ldr d17, [x17, #0x1a0]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "ldr x20, [x17, #0x1b8]\n"
+ "fmla v13.8h, v16.8h, v1.h[6]\n"
+ "ldr d16, [x17, #0x1b0]\n"
+ "mov v17.d[1], x21\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v14.8h, v17.8h, v1.h[6]\n"
+ "ldr d17, [x17, #0x1c0]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "ldr x20, [x17, #0x1c8]\n"
+ "fmla v15.8h, v16.8h, v1.h[6]\n"
+ "ldr d16, [x17, #0x1d0]\n"
+ "mov v17.d[1], x20\n"
+ "ldr x20, [x17, #0x1d8]\n"
+ "mov v16.d[1], x20\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "ldr x21, [x17, #0x1e8]\n"
+ "fmla v12.8h, v17.8h, v1.h[7]\n"
+ "ldr d17, [x17, #0x1e0]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "ldr x20, [x17, #0x1f8]\n"
+ "fmla v13.8h, v16.8h, v1.h[7]\n"
+ "ldr d16, [x17, #0x1f0]\n"
+ "mov v17.d[1], x21\n"
"add x13, x13, #0x10\n"
- "mov v7.d[1], x11\n"
- "add x9, x9, #0x10\n"
+ "mov v16.d[1], x20\n"
+ "add x12, x12, #0x10\n"
"add x17, x17, #0x200\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
+ "fmla v10.8h, v17.8h, v0.h[7]\n"
+ "fmla v14.8h, v17.8h, v1.h[7]\n"
"ldr d6, [x17, #0x0]\n"
- "ldr x12, [x17, #0x8]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "ldr x21, [x17, #0x8]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
"ldr d0, [x13, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "ldr d1, [x9, #0x0]\n"
+ "fmla v15.8h, v16.8h, v1.h[7]\n"
+ "ldr d1, [x12, #0x0]\n"
"sub x14, x14, #0x8\n"
"ldr d7, [x17, #0x10]\n"
"cmp x14, #0x10\n"
- "ldr x10, [x13, #0x8]\n"
- "mov v6.d[1], x12\n"
- "ldr x28, [x9, #0x8]\n"
- "mov v0.d[1], x10\n"
- "ldr x11, [x17, #0x18]\n"
- "mov v1.d[1], x28\n"
+ "ldr x20, [x13, #0x8]\n"
+ "mov v6.d[1], x21\n"
+ "ldr x21, [x12, #0x8]\n"
+ "mov v0.d[1], x20\n"
+ "ldr x20, [x17, #0x18]\n"
+ "mov v1.d[1], x21\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "mov v7.d[1], x11\n"
- "prfm pldl1keep, [x9, #0x80]\n"
+ "mov v7.d[1], x20\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
"bge 75b\n"
"76:" // Height 2: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
+ "ldr q17, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x30]\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
"sub x14, x14, #0x8\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
+ "fmla v14.8h, v17.8h, v1.h[0]\n"
+ "ldr q17, [x17, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x17, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr q6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr q7, [x17, #0x70]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr q6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr q7, [x17, #0x90]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr q6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr q7, [x17, #0xb0]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr q6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr q7, [x17, #0xd0]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr q6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr q7, [x17, #0xf0]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr q6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr q7, [x17, #0x110]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr q6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr q7, [x17, #0x130]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr q6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr q7, [x17, #0x150]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr q6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr q7, [x17, #0x170]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr q6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr q7, [x17, #0x190]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr q6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr q7, [x17, #0x1b0]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr q6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr q7, [x17, #0x1d0]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr q6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "ldr q7, [x17, #0x1f0]\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
+ "fmla v15.8h, v16.8h, v1.h[0]\n"
+ "ldr q16, [x17, #0x50]\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
+ "fmla v12.8h, v17.8h, v1.h[1]\n"
+ "ldr q17, [x17, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "fmla v13.8h, v16.8h, v1.h[1]\n"
+ "ldr q16, [x17, #0x70]\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "fmla v14.8h, v17.8h, v1.h[1]\n"
+ "ldr q17, [x17, #0x80]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "fmla v15.8h, v16.8h, v1.h[1]\n"
+ "ldr q16, [x17, #0x90]\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "fmla v12.8h, v17.8h, v1.h[2]\n"
+ "ldr q17, [x17, #0xa0]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "fmla v13.8h, v16.8h, v1.h[2]\n"
+ "ldr q16, [x17, #0xb0]\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "fmla v14.8h, v17.8h, v1.h[2]\n"
+ "ldr q17, [x17, #0xc0]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "fmla v15.8h, v16.8h, v1.h[2]\n"
+ "ldr q16, [x17, #0xd0]\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "fmla v12.8h, v17.8h, v1.h[3]\n"
+ "ldr q17, [x17, #0xe0]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "fmla v13.8h, v16.8h, v1.h[3]\n"
+ "ldr q16, [x17, #0xf0]\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "fmla v14.8h, v17.8h, v1.h[3]\n"
+ "ldr q17, [x17, #0x100]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "fmla v15.8h, v16.8h, v1.h[3]\n"
+ "ldr q16, [x17, #0x110]\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "fmla v12.8h, v17.8h, v1.h[4]\n"
+ "ldr q17, [x17, #0x120]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "fmla v13.8h, v16.8h, v1.h[4]\n"
+ "ldr q16, [x17, #0x130]\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "fmla v14.8h, v17.8h, v1.h[4]\n"
+ "ldr q17, [x17, #0x140]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "fmla v15.8h, v16.8h, v1.h[4]\n"
+ "ldr q16, [x17, #0x150]\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "fmla v12.8h, v17.8h, v1.h[5]\n"
+ "ldr q17, [x17, #0x160]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "fmla v13.8h, v16.8h, v1.h[5]\n"
+ "ldr q16, [x17, #0x170]\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "fmla v14.8h, v17.8h, v1.h[5]\n"
+ "ldr q17, [x17, #0x180]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "fmla v15.8h, v16.8h, v1.h[5]\n"
+ "ldr q16, [x17, #0x190]\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "fmla v12.8h, v17.8h, v1.h[6]\n"
+ "ldr q17, [x17, #0x1a0]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "fmla v13.8h, v16.8h, v1.h[6]\n"
+ "ldr q16, [x17, #0x1b0]\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "fmla v14.8h, v17.8h, v1.h[6]\n"
+ "ldr q17, [x17, #0x1c0]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "fmla v15.8h, v16.8h, v1.h[6]\n"
+ "ldr q16, [x17, #0x1d0]\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "fmla v12.8h, v17.8h, v1.h[7]\n"
+ "ldr q17, [x17, #0x1e0]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "fmla v13.8h, v16.8h, v1.h[7]\n"
+ "ldr q16, [x17, #0x1f0]\n"
+ "fmla v10.8h, v17.8h, v0.h[7]\n"
"add x17, x17, #0x200\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
+ "fmla v14.8h, v17.8h, v1.h[7]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
+ "fmla v15.8h, v16.8h, v1.h[7]\n"
"77:" // Height 2: Multiply loop: Main loop skip
"cbz x14, 79f\n"
"78:" // Height 2: Multiply loop: Odd block loop
- "ldr h0, [x13], #0x2\n"
+ "ldr h1, [x13], #0x2\n"
"sub x14, x14, #0x1\n"
- "ldr h1, [x9], #0x2\n"
- "ldr q6, [x17, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x10]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr h0, [x12], #0x2\n"
+ "ldr q17, [x17, #0x0]\n"
+ "fmla v8.8h, v17.8h, v1.h[0]\n"
+ "ldr q16, [x17, #0x10]\n"
+ "fmla v12.8h, v17.8h, v0.h[0]\n"
+ "ldr q17, [x17, #0x20]\n"
+ "fmla v9.8h, v16.8h, v1.h[0]\n"
+ "fmla v13.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x17, #0x30]\n"
+ "fmla v10.8h, v17.8h, v1.h[0]\n"
"add x17, x17, #0x40\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
+ "fmla v14.8h, v17.8h, v0.h[0]\n"
+ "fmla v11.8h, v16.8h, v1.h[0]\n"
+ "fmla v15.8h, v16.8h, v0.h[0]\n"
"cbnz x14, 78b\n"
"79:" // Height 2: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -1129,25 +1129,25 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"prfm pstl1keep, [x25, #0x0]\n"
"tbz %x[flags], #1, 80f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v0.8h\n"
- "fmin v9.8h, v9.8h, v0.8h\n"
- "fmin v10.8h, v10.8h, v0.8h\n"
- "fmin v11.8h, v11.8h, v0.8h\n"
- "fmin v12.8h, v12.8h, v0.8h\n"
- "fmin v13.8h, v13.8h, v0.8h\n"
- "fmin v14.8h, v14.8h, v0.8h\n"
- "fmin v15.8h, v15.8h, v0.8h\n"
+ "ld1r { v16.8h }, [x20]\n"
+ "fmin v8.8h, v8.8h, v16.8h\n"
+ "fmin v9.8h, v9.8h, v16.8h\n"
+ "fmin v10.8h, v10.8h, v16.8h\n"
+ "fmin v11.8h, v11.8h, v16.8h\n"
+ "fmin v12.8h, v12.8h, v16.8h\n"
+ "fmin v13.8h, v13.8h, v16.8h\n"
+ "fmin v14.8h, v14.8h, v16.8h\n"
+ "fmin v15.8h, v15.8h, v16.8h\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmax v8.8h, v8.8h, v0.8h\n"
- "fmax v9.8h, v9.8h, v0.8h\n"
- "fmax v10.8h, v10.8h, v0.8h\n"
- "fmax v11.8h, v11.8h, v0.8h\n"
- "fmax v12.8h, v12.8h, v0.8h\n"
- "fmax v13.8h, v13.8h, v0.8h\n"
- "fmax v14.8h, v14.8h, v0.8h\n"
- "fmax v15.8h, v15.8h, v0.8h\n"
+ "ld1r { v16.8h }, [x20]\n"
+ "fmax v8.8h, v8.8h, v16.8h\n"
+ "fmax v9.8h, v9.8h, v16.8h\n"
+ "fmax v10.8h, v10.8h, v16.8h\n"
+ "fmax v11.8h, v11.8h, v16.8h\n"
+ "fmax v12.8h, v12.8h, v16.8h\n"
+ "fmax v13.8h, v13.8h, v16.8h\n"
+ "fmax v14.8h, v14.8h, v16.8h\n"
+ "fmax v15.8h, v15.8h, v16.8h\n"
"80:" // Height 2: No activation
"cmp x8, #0x20\n"
"bge 97f\n"
@@ -1526,404 +1526,404 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"121:" // Height 3: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w14, [x20, x15, 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, 122f\n"
- "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x13, [x21, #0x0]\n"
- "ldr x9, [x21, #0x8]\n"
- "ldr x27, [x21, #0x10]\n"
+ "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x13, [x20, #0x0]\n"
+ "ldr x12, [x20, #0x8]\n"
+ "ldr x11, [x20, #0x10]\n"
"cbnz x15, 123f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x13, x13, x20, LSL #1\n"
- "add x9, x9, x20, LSL #1\n"
- "add x27, x27, x20, LSL #1\n"
+ "add x12, x12, x20, LSL #1\n"
+ "add x11, x11, x20, LSL #1\n"
"b 123f\n"
"122:" // Height 3: setup direct input
"mov x13, %x[input_ptr]\n"
- "add x9, x13, x20, LSL #1\n"
- "add x27, x9, x20, LSL #1\n"
+ "add x12, x13, x21, LSL #1\n"
+ "add x11, x12, x21, LSL #1\n"
"123:" // Height 3: input setup done
"cmp x14, #0x8\n"
"blt 126f\n"
"ldr q0, [x13, #0x0]\n"
"cmp x14, #0x10\n"
- "ldr q1, [x9, #0x0]\n"
- "ldr q2, [x27, #0x0]\n"
+ "ldr q1, [x12, #0x0]\n"
+ "ldr q2, [x11, #0x0]\n"
"ldr q6, [x17, #0x0]\n"
"ldr q7, [x17, #0x10]\n"
"blt 125f\n"
"124:" // Height 3: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr x12, [x17, #0x28]\n"
+ "ldr x21, [x17, #0x28]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x38]\n"
+ "ldr x20, [x17, #0x38]\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "ldr d6, [x17, #0x20]\n"
+ "ldr d21, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
+ "mov v21.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x48]\n"
+ "ldr x21, [x17, #0x48]\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "ldr d7, [x17, #0x30]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x58]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "ldr d6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x68]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "ldr d7, [x17, #0x50]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x78]\n"
- "fmla v16.8h, v6.8h, v2.h[1]\n"
- "ldr d6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0x88]\n"
- "fmla v17.8h, v7.8h, v2.h[1]\n"
- "ldr d7, [x17, #0x70]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x98]\n"
- "fmla v18.8h, v6.8h, v2.h[1]\n"
- "ldr d6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0xa8]\n"
- "fmla v19.8h, v7.8h, v2.h[1]\n"
- "ldr d7, [x17, #0x90]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xb8]\n"
- "fmla v16.8h, v6.8h, v2.h[2]\n"
- "ldr d6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xc8]\n"
- "fmla v17.8h, v7.8h, v2.h[2]\n"
- "ldr d7, [x17, #0xb0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xd8]\n"
- "fmla v18.8h, v6.8h, v2.h[2]\n"
- "ldr d6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xe8]\n"
- "fmla v19.8h, v7.8h, v2.h[2]\n"
- "ldr d7, [x17, #0xd0]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0xf8]\n"
- "fmla v16.8h, v6.8h, v2.h[3]\n"
- "ldr d6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x108]\n"
- "fmla v17.8h, v7.8h, v2.h[3]\n"
- "ldr d7, [x17, #0xf0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0x118]\n"
- "fmla v18.8h, v6.8h, v2.h[3]\n"
- "ldr d6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x128]\n"
- "fmla v19.8h, v7.8h, v2.h[3]\n"
- "ldr d7, [x17, #0x110]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x138]\n"
- "fmla v16.8h, v6.8h, v2.h[4]\n"
- "ldr d6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x148]\n"
- "fmla v17.8h, v7.8h, v2.h[4]\n"
- "ldr d7, [x17, #0x130]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x158]\n"
- "fmla v18.8h, v6.8h, v2.h[4]\n"
- "ldr d6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x168]\n"
- "fmla v19.8h, v7.8h, v2.h[4]\n"
- "ldr d7, [x17, #0x150]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x178]\n"
- "fmla v16.8h, v6.8h, v2.h[5]\n"
- "ldr d6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x188]\n"
- "fmla v17.8h, v7.8h, v2.h[5]\n"
- "ldr d7, [x17, #0x170]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x198]\n"
- "fmla v18.8h, v6.8h, v2.h[5]\n"
- "ldr d6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x1a8]\n"
- "fmla v19.8h, v7.8h, v2.h[5]\n"
- "ldr d7, [x17, #0x190]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1b8]\n"
- "fmla v16.8h, v6.8h, v2.h[6]\n"
- "ldr d6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1c8]\n"
- "fmla v17.8h, v7.8h, v2.h[6]\n"
- "ldr d7, [x17, #0x1b0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1d8]\n"
- "fmla v18.8h, v6.8h, v2.h[6]\n"
- "ldr d6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1e8]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "ldr d7, [x17, #0x1d0]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x1f8]\n"
- "fmla v16.8h, v6.8h, v2.h[7]\n"
- "ldr d6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
+ "ldr d20, [x17, #0x30]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[0]\n"
+ "fmla v14.8h, v21.8h, v1.h[0]\n"
+ "ldr x20, [x17, #0x58]\n"
+ "fmla v18.8h, v21.8h, v2.h[0]\n"
+ "ldr d21, [x17, #0x40]\n"
+ "fmla v11.8h, v20.8h, v0.h[0]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[0]\n"
+ "ldr x21, [x17, #0x68]\n"
+ "fmla v19.8h, v20.8h, v2.h[0]\n"
+ "ldr d20, [x17, #0x50]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[1]\n"
+ "fmla v12.8h, v21.8h, v1.h[1]\n"
+ "ldr x20, [x17, #0x78]\n"
+ "fmla v16.8h, v21.8h, v2.h[1]\n"
+ "ldr d21, [x17, #0x60]\n"
+ "fmla v9.8h, v20.8h, v0.h[1]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[1]\n"
+ "ldr x21, [x17, #0x88]\n"
+ "fmla v17.8h, v20.8h, v2.h[1]\n"
+ "ldr d20, [x17, #0x70]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[1]\n"
+ "fmla v14.8h, v21.8h, v1.h[1]\n"
+ "ldr x20, [x17, #0x98]\n"
+ "fmla v18.8h, v21.8h, v2.h[1]\n"
+ "ldr d21, [x17, #0x80]\n"
+ "fmla v11.8h, v20.8h, v0.h[1]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[1]\n"
+ "ldr x21, [x17, #0xa8]\n"
+ "fmla v19.8h, v20.8h, v2.h[1]\n"
+ "ldr d20, [x17, #0x90]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[2]\n"
+ "fmla v12.8h, v21.8h, v1.h[2]\n"
+ "ldr x20, [x17, #0xb8]\n"
+ "fmla v16.8h, v21.8h, v2.h[2]\n"
+ "ldr d21, [x17, #0xa0]\n"
+ "fmla v9.8h, v20.8h, v0.h[2]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[2]\n"
+ "ldr x21, [x17, #0xc8]\n"
+ "fmla v17.8h, v20.8h, v2.h[2]\n"
+ "ldr d20, [x17, #0xb0]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[2]\n"
+ "fmla v14.8h, v21.8h, v1.h[2]\n"
+ "ldr x20, [x17, #0xd8]\n"
+ "fmla v18.8h, v21.8h, v2.h[2]\n"
+ "ldr d21, [x17, #0xc0]\n"
+ "fmla v11.8h, v20.8h, v0.h[2]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[2]\n"
+ "ldr x21, [x17, #0xe8]\n"
+ "fmla v19.8h, v20.8h, v2.h[2]\n"
+ "ldr d20, [x17, #0xd0]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[3]\n"
+ "fmla v12.8h, v21.8h, v1.h[3]\n"
+ "ldr x20, [x17, #0xf8]\n"
+ "fmla v16.8h, v21.8h, v2.h[3]\n"
+ "ldr d21, [x17, #0xe0]\n"
+ "fmla v9.8h, v20.8h, v0.h[3]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[3]\n"
+ "ldr x21, [x17, #0x108]\n"
+ "fmla v17.8h, v20.8h, v2.h[3]\n"
+ "ldr d20, [x17, #0xf0]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[3]\n"
+ "fmla v14.8h, v21.8h, v1.h[3]\n"
+ "ldr x20, [x17, #0x118]\n"
+ "fmla v18.8h, v21.8h, v2.h[3]\n"
+ "ldr d21, [x17, #0x100]\n"
+ "fmla v11.8h, v20.8h, v0.h[3]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[3]\n"
+ "ldr x21, [x17, #0x128]\n"
+ "fmla v19.8h, v20.8h, v2.h[3]\n"
+ "ldr d20, [x17, #0x110]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[4]\n"
+ "fmla v12.8h, v21.8h, v1.h[4]\n"
+ "ldr x20, [x17, #0x138]\n"
+ "fmla v16.8h, v21.8h, v2.h[4]\n"
+ "ldr d21, [x17, #0x120]\n"
+ "fmla v9.8h, v20.8h, v0.h[4]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[4]\n"
+ "ldr x21, [x17, #0x148]\n"
+ "fmla v17.8h, v20.8h, v2.h[4]\n"
+ "ldr d20, [x17, #0x130]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[4]\n"
+ "fmla v14.8h, v21.8h, v1.h[4]\n"
+ "ldr x20, [x17, #0x158]\n"
+ "fmla v18.8h, v21.8h, v2.h[4]\n"
+ "ldr d21, [x17, #0x140]\n"
+ "fmla v11.8h, v20.8h, v0.h[4]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[4]\n"
+ "ldr x21, [x17, #0x168]\n"
+ "fmla v19.8h, v20.8h, v2.h[4]\n"
+ "ldr d20, [x17, #0x150]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[5]\n"
+ "fmla v12.8h, v21.8h, v1.h[5]\n"
+ "ldr x20, [x17, #0x178]\n"
+ "fmla v16.8h, v21.8h, v2.h[5]\n"
+ "ldr d21, [x17, #0x160]\n"
+ "fmla v9.8h, v20.8h, v0.h[5]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[5]\n"
+ "ldr x21, [x17, #0x188]\n"
+ "fmla v17.8h, v20.8h, v2.h[5]\n"
+ "ldr d20, [x17, #0x170]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[5]\n"
+ "fmla v14.8h, v21.8h, v1.h[5]\n"
+ "ldr x20, [x17, #0x198]\n"
+ "fmla v18.8h, v21.8h, v2.h[5]\n"
+ "ldr d21, [x17, #0x180]\n"
+ "fmla v11.8h, v20.8h, v0.h[5]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[5]\n"
+ "ldr x21, [x17, #0x1a8]\n"
+ "fmla v19.8h, v20.8h, v2.h[5]\n"
+ "ldr d20, [x17, #0x190]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[6]\n"
+ "fmla v12.8h, v21.8h, v1.h[6]\n"
+ "ldr x20, [x17, #0x1b8]\n"
+ "fmla v16.8h, v21.8h, v2.h[6]\n"
+ "ldr d21, [x17, #0x1a0]\n"
+ "fmla v9.8h, v20.8h, v0.h[6]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[6]\n"
+ "ldr x21, [x17, #0x1c8]\n"
+ "fmla v17.8h, v20.8h, v2.h[6]\n"
+ "ldr d20, [x17, #0x1b0]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v10.8h, v21.8h, v0.h[6]\n"
+ "fmla v14.8h, v21.8h, v1.h[6]\n"
+ "ldr x20, [x17, #0x1d8]\n"
+ "fmla v18.8h, v21.8h, v2.h[6]\n"
+ "ldr d21, [x17, #0x1c0]\n"
+ "fmla v11.8h, v20.8h, v0.h[6]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v15.8h, v20.8h, v1.h[6]\n"
+ "ldr x21, [x17, #0x1e8]\n"
+ "fmla v19.8h, v20.8h, v2.h[6]\n"
+ "ldr d20, [x17, #0x1d0]\n"
+ "mov v20.d[1], x20\n"
+ "fmla v8.8h, v21.8h, v0.h[7]\n"
+ "fmla v12.8h, v21.8h, v1.h[7]\n"
+ "ldr x20, [x17, #0x1f8]\n"
+ "fmla v16.8h, v21.8h, v2.h[7]\n"
+ "ldr d21, [x17, #0x1e0]\n"
+ "fmla v9.8h, v20.8h, v0.h[7]\n"
+ "mov v21.d[1], x21\n"
+ "fmla v13.8h, v20.8h, v1.h[7]\n"
"add x13, x13, #0x10\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "ldr d7, [x17, #0x1f0]\n"
- "mov v7.d[1], x11\n"
- "add x9, x9, #0x10\n"
- "add x27, x27, #0x10\n"
+ "fmla v17.8h, v20.8h, v2.h[7]\n"
+ "ldr d20, [x17, #0x1f0]\n"
+ "mov v20.d[1], x20\n"
+ "add x12, x12, #0x10\n"
+ "add x11, x11, #0x10\n"
"add x17, x17, #0x200\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "ldr x12, [x17, #0x8]\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "ldr x10, [x13, #0x8]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
+ "fmla v10.8h, v21.8h, v0.h[7]\n"
+ "ldr x20, [x17, #0x8]\n"
+ "fmla v14.8h, v21.8h, v1.h[7]\n"
+ "ldr x23, [x13, #0x8]\n"
+ "fmla v18.8h, v21.8h, v2.h[7]\n"
"ldr d6, [x17, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v20.8h, v0.h[7]\n"
"ldr d0, [x13, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "ldr d1, [x9, #0x0]\n"
- "ldr x28, [x9, #0x8]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
- "ldr d2, [x27, #0x0]\n"
+ "fmla v15.8h, v20.8h, v1.h[7]\n"
+ "ldr d1, [x12, #0x0]\n"
+ "ldr x22, [x12, #0x8]\n"
+ "fmla v19.8h, v20.8h, v2.h[7]\n"
+ "ldr d2, [x11, #0x0]\n"
"sub x14, x14, #0x8\n"
"ldr d7, [x17, #0x10]\n"
"cmp x14, #0x10\n"
- "ldr x26, [x27, #0x8]\n"
- "mov v6.d[1], x12\n"
- "ldr x11, [x17, #0x18]\n"
- "mov v0.d[1], x10\n"
+ "ldr x21, [x11, #0x8]\n"
+ "mov v6.d[1], x20\n"
+ "ldr x20, [x17, #0x18]\n"
+ "mov v0.d[1], x23\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "mov v1.d[1], x28\n"
- "prfm pldl1keep, [x9, #0x80]\n"
- "mov v2.d[1], x26\n"
- "prfm pldl1keep, [x27, #0x80]\n"
- "mov v7.d[1], x11\n"
+ "mov v1.d[1], x22\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
+ "mov v2.d[1], x21\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
+ "mov v7.d[1], x20\n"
"bge 124b\n"
"125:" // Height 3: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
+ "ldr q21, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "add x27, x27, #0x10\n"
+ "add x11, x11, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
"sub x14, x14, #0x8\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr q20, [x17, #0x30]\n"
+ "fmla v10.8h, v21.8h, v0.h[0]\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "ldr q7, [x17, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "fmla v16.8h, v6.8h, v2.h[1]\n"
- "ldr q6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "fmla v17.8h, v7.8h, v2.h[1]\n"
- "ldr q7, [x17, #0x70]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "fmla v18.8h, v6.8h, v2.h[1]\n"
- "ldr q6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "fmla v19.8h, v7.8h, v2.h[1]\n"
- "ldr q7, [x17, #0x90]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "fmla v16.8h, v6.8h, v2.h[2]\n"
- "ldr q6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "fmla v17.8h, v7.8h, v2.h[2]\n"
- "ldr q7, [x17, #0xb0]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "fmla v18.8h, v6.8h, v2.h[2]\n"
- "ldr q6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "fmla v19.8h, v7.8h, v2.h[2]\n"
- "ldr q7, [x17, #0xd0]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "fmla v16.8h, v6.8h, v2.h[3]\n"
- "ldr q6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "fmla v17.8h, v7.8h, v2.h[3]\n"
- "ldr q7, [x17, #0xf0]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "fmla v18.8h, v6.8h, v2.h[3]\n"
- "ldr q6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "fmla v19.8h, v7.8h, v2.h[3]\n"
- "ldr q7, [x17, #0x110]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "fmla v16.8h, v6.8h, v2.h[4]\n"
- "ldr q6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "fmla v17.8h, v7.8h, v2.h[4]\n"
- "ldr q7, [x17, #0x130]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "fmla v18.8h, v6.8h, v2.h[4]\n"
- "ldr q6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "fmla v19.8h, v7.8h, v2.h[4]\n"
- "ldr q7, [x17, #0x150]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "fmla v16.8h, v6.8h, v2.h[5]\n"
- "ldr q6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "fmla v17.8h, v7.8h, v2.h[5]\n"
- "ldr q7, [x17, #0x170]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "fmla v18.8h, v6.8h, v2.h[5]\n"
- "ldr q6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "fmla v19.8h, v7.8h, v2.h[5]\n"
- "ldr q7, [x17, #0x190]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "fmla v16.8h, v6.8h, v2.h[6]\n"
- "ldr q6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "fmla v17.8h, v7.8h, v2.h[6]\n"
- "ldr q7, [x17, #0x1b0]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "fmla v18.8h, v6.8h, v2.h[6]\n"
- "ldr q6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "ldr q7, [x17, #0x1d0]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "fmla v16.8h, v6.8h, v2.h[7]\n"
- "ldr q6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "ldr q7, [x17, #0x1f0]\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
+ "fmla v14.8h, v21.8h, v1.h[0]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
+ "fmla v18.8h, v21.8h, v2.h[0]\n"
+ "ldr q21, [x17, #0x40]\n"
+ "fmla v11.8h, v20.8h, v0.h[0]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
+ "fmla v15.8h, v20.8h, v1.h[0]\n"
+ "fmla v19.8h, v20.8h, v2.h[0]\n"
+ "ldr q20, [x17, #0x50]\n"
+ "fmla v8.8h, v21.8h, v0.h[1]\n"
+ "fmla v12.8h, v21.8h, v1.h[1]\n"
+ "fmla v16.8h, v21.8h, v2.h[1]\n"
+ "ldr q21, [x17, #0x60]\n"
+ "fmla v9.8h, v20.8h, v0.h[1]\n"
+ "fmla v13.8h, v20.8h, v1.h[1]\n"
+ "fmla v17.8h, v20.8h, v2.h[1]\n"
+ "ldr q20, [x17, #0x70]\n"
+ "fmla v10.8h, v21.8h, v0.h[1]\n"
+ "fmla v14.8h, v21.8h, v1.h[1]\n"
+ "fmla v18.8h, v21.8h, v2.h[1]\n"
+ "ldr q21, [x17, #0x80]\n"
+ "fmla v11.8h, v20.8h, v0.h[1]\n"
+ "fmla v15.8h, v20.8h, v1.h[1]\n"
+ "fmla v19.8h, v20.8h, v2.h[1]\n"
+ "ldr q20, [x17, #0x90]\n"
+ "fmla v8.8h, v21.8h, v0.h[2]\n"
+ "fmla v12.8h, v21.8h, v1.h[2]\n"
+ "fmla v16.8h, v21.8h, v2.h[2]\n"
+ "ldr q21, [x17, #0xa0]\n"
+ "fmla v9.8h, v20.8h, v0.h[2]\n"
+ "fmla v13.8h, v20.8h, v1.h[2]\n"
+ "fmla v17.8h, v20.8h, v2.h[2]\n"
+ "ldr q20, [x17, #0xb0]\n"
+ "fmla v10.8h, v21.8h, v0.h[2]\n"
+ "fmla v14.8h, v21.8h, v1.h[2]\n"
+ "fmla v18.8h, v21.8h, v2.h[2]\n"
+ "ldr q21, [x17, #0xc0]\n"
+ "fmla v11.8h, v20.8h, v0.h[2]\n"
+ "fmla v15.8h, v20.8h, v1.h[2]\n"
+ "fmla v19.8h, v20.8h, v2.h[2]\n"
+ "ldr q20, [x17, #0xd0]\n"
+ "fmla v8.8h, v21.8h, v0.h[3]\n"
+ "fmla v12.8h, v21.8h, v1.h[3]\n"
+ "fmla v16.8h, v21.8h, v2.h[3]\n"
+ "ldr q21, [x17, #0xe0]\n"
+ "fmla v9.8h, v20.8h, v0.h[3]\n"
+ "fmla v13.8h, v20.8h, v1.h[3]\n"
+ "fmla v17.8h, v20.8h, v2.h[3]\n"
+ "ldr q20, [x17, #0xf0]\n"
+ "fmla v10.8h, v21.8h, v0.h[3]\n"
+ "fmla v14.8h, v21.8h, v1.h[3]\n"
+ "fmla v18.8h, v21.8h, v2.h[3]\n"
+ "ldr q21, [x17, #0x100]\n"
+ "fmla v11.8h, v20.8h, v0.h[3]\n"
+ "fmla v15.8h, v20.8h, v1.h[3]\n"
+ "fmla v19.8h, v20.8h, v2.h[3]\n"
+ "ldr q20, [x17, #0x110]\n"
+ "fmla v8.8h, v21.8h, v0.h[4]\n"
+ "fmla v12.8h, v21.8h, v1.h[4]\n"
+ "fmla v16.8h, v21.8h, v2.h[4]\n"
+ "ldr q21, [x17, #0x120]\n"
+ "fmla v9.8h, v20.8h, v0.h[4]\n"
+ "fmla v13.8h, v20.8h, v1.h[4]\n"
+ "fmla v17.8h, v20.8h, v2.h[4]\n"
+ "ldr q20, [x17, #0x130]\n"
+ "fmla v10.8h, v21.8h, v0.h[4]\n"
+ "fmla v14.8h, v21.8h, v1.h[4]\n"
+ "fmla v18.8h, v21.8h, v2.h[4]\n"
+ "ldr q21, [x17, #0x140]\n"
+ "fmla v11.8h, v20.8h, v0.h[4]\n"
+ "fmla v15.8h, v20.8h, v1.h[4]\n"
+ "fmla v19.8h, v20.8h, v2.h[4]\n"
+ "ldr q20, [x17, #0x150]\n"
+ "fmla v8.8h, v21.8h, v0.h[5]\n"
+ "fmla v12.8h, v21.8h, v1.h[5]\n"
+ "fmla v16.8h, v21.8h, v2.h[5]\n"
+ "ldr q21, [x17, #0x160]\n"
+ "fmla v9.8h, v20.8h, v0.h[5]\n"
+ "fmla v13.8h, v20.8h, v1.h[5]\n"
+ "fmla v17.8h, v20.8h, v2.h[5]\n"
+ "ldr q20, [x17, #0x170]\n"
+ "fmla v10.8h, v21.8h, v0.h[5]\n"
+ "fmla v14.8h, v21.8h, v1.h[5]\n"
+ "fmla v18.8h, v21.8h, v2.h[5]\n"
+ "ldr q21, [x17, #0x180]\n"
+ "fmla v11.8h, v20.8h, v0.h[5]\n"
+ "fmla v15.8h, v20.8h, v1.h[5]\n"
+ "fmla v19.8h, v20.8h, v2.h[5]\n"
+ "ldr q20, [x17, #0x190]\n"
+ "fmla v8.8h, v21.8h, v0.h[6]\n"
+ "fmla v12.8h, v21.8h, v1.h[6]\n"
+ "fmla v16.8h, v21.8h, v2.h[6]\n"
+ "ldr q21, [x17, #0x1a0]\n"
+ "fmla v9.8h, v20.8h, v0.h[6]\n"
+ "fmla v13.8h, v20.8h, v1.h[6]\n"
+ "fmla v17.8h, v20.8h, v2.h[6]\n"
+ "ldr q20, [x17, #0x1b0]\n"
+ "fmla v10.8h, v21.8h, v0.h[6]\n"
+ "fmla v14.8h, v21.8h, v1.h[6]\n"
+ "fmla v18.8h, v21.8h, v2.h[6]\n"
+ "ldr q21, [x17, #0x1c0]\n"
+ "fmla v11.8h, v20.8h, v0.h[6]\n"
+ "fmla v15.8h, v20.8h, v1.h[6]\n"
+ "fmla v19.8h, v20.8h, v2.h[6]\n"
+ "ldr q20, [x17, #0x1d0]\n"
+ "fmla v8.8h, v21.8h, v0.h[7]\n"
+ "fmla v12.8h, v21.8h, v1.h[7]\n"
+ "fmla v16.8h, v21.8h, v2.h[7]\n"
+ "ldr q21, [x17, #0x1e0]\n"
+ "fmla v9.8h, v20.8h, v0.h[7]\n"
+ "fmla v13.8h, v20.8h, v1.h[7]\n"
+ "fmla v17.8h, v20.8h, v2.h[7]\n"
+ "ldr q20, [x17, #0x1f0]\n"
+ "fmla v10.8h, v21.8h, v0.h[7]\n"
"add x17, x17, #0x200\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
+ "fmla v14.8h, v21.8h, v1.h[7]\n"
+ "fmla v18.8h, v21.8h, v2.h[7]\n"
+ "fmla v11.8h, v20.8h, v0.h[7]\n"
+ "fmla v15.8h, v20.8h, v1.h[7]\n"
+ "fmla v19.8h, v20.8h, v2.h[7]\n"
"126:" // Height 3: Multiply loop: Main loop skip
"cbz x14, 128f\n"
"127:" // Height 3: Multiply loop: Odd block loop
- "ldr h0, [x13], #0x2\n"
+ "ldr h2, [x13], #0x2\n"
"sub x14, x14, #0x1\n"
- "ldr h1, [x9], #0x2\n"
- "ldr h2, [x27], #0x2\n"
- "ldr q6, [x17, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x10]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "fmla v16.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "fmla v17.8h, v7.8h, v2.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr h1, [x12], #0x2\n"
+ "ldr h0, [x11], #0x2\n"
+ "ldr q21, [x17, #0x0]\n"
+ "fmla v8.8h, v21.8h, v2.h[0]\n"
+ "ldr q20, [x17, #0x10]\n"
+ "fmla v12.8h, v21.8h, v1.h[0]\n"
+ "fmla v16.8h, v21.8h, v0.h[0]\n"
+ "ldr q21, [x17, #0x20]\n"
+ "fmla v9.8h, v20.8h, v2.h[0]\n"
+ "fmla v13.8h, v20.8h, v1.h[0]\n"
+ "fmla v17.8h, v20.8h, v0.h[0]\n"
+ "ldr q20, [x17, #0x30]\n"
+ "fmla v10.8h, v21.8h, v2.h[0]\n"
"add x17, x17, #0x40\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
+ "fmla v14.8h, v21.8h, v1.h[0]\n"
+ "fmla v18.8h, v21.8h, v0.h[0]\n"
+ "fmla v11.8h, v20.8h, v2.h[0]\n"
+ "fmla v15.8h, v20.8h, v1.h[0]\n"
+ "fmla v19.8h, v20.8h, v0.h[0]\n"
"cbnz x14, 127b\n"
"128:" // Height 3: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -1938,33 +1938,33 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"prfm pstl1keep, [x24, #0x0]\n"
"tbz %x[flags], #1, 129f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v0.8h\n"
- "fmin v9.8h, v9.8h, v0.8h\n"
- "fmin v10.8h, v10.8h, v0.8h\n"
- "fmin v11.8h, v11.8h, v0.8h\n"
- "fmin v12.8h, v12.8h, v0.8h\n"
- "fmin v13.8h, v13.8h, v0.8h\n"
- "fmin v14.8h, v14.8h, v0.8h\n"
- "fmin v15.8h, v15.8h, v0.8h\n"
- "fmin v16.8h, v16.8h, v0.8h\n"
- "fmin v17.8h, v17.8h, v0.8h\n"
- "fmin v18.8h, v18.8h, v0.8h\n"
- "fmin v19.8h, v19.8h, v0.8h\n"
+ "ld1r { v20.8h }, [x20]\n"
+ "fmin v8.8h, v8.8h, v20.8h\n"
+ "fmin v9.8h, v9.8h, v20.8h\n"
+ "fmin v10.8h, v10.8h, v20.8h\n"
+ "fmin v11.8h, v11.8h, v20.8h\n"
+ "fmin v12.8h, v12.8h, v20.8h\n"
+ "fmin v13.8h, v13.8h, v20.8h\n"
+ "fmin v14.8h, v14.8h, v20.8h\n"
+ "fmin v15.8h, v15.8h, v20.8h\n"
+ "fmin v16.8h, v16.8h, v20.8h\n"
+ "fmin v17.8h, v17.8h, v20.8h\n"
+ "fmin v18.8h, v18.8h, v20.8h\n"
+ "fmin v19.8h, v19.8h, v20.8h\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmax v8.8h, v8.8h, v0.8h\n"
- "fmax v9.8h, v9.8h, v0.8h\n"
- "fmax v10.8h, v10.8h, v0.8h\n"
- "fmax v11.8h, v11.8h, v0.8h\n"
- "fmax v12.8h, v12.8h, v0.8h\n"
- "fmax v13.8h, v13.8h, v0.8h\n"
- "fmax v14.8h, v14.8h, v0.8h\n"
- "fmax v15.8h, v15.8h, v0.8h\n"
- "fmax v16.8h, v16.8h, v0.8h\n"
- "fmax v17.8h, v17.8h, v0.8h\n"
- "fmax v18.8h, v18.8h, v0.8h\n"
- "fmax v19.8h, v19.8h, v0.8h\n"
+ "ld1r { v20.8h }, [x20]\n"
+ "fmax v8.8h, v8.8h, v20.8h\n"
+ "fmax v9.8h, v9.8h, v20.8h\n"
+ "fmax v10.8h, v10.8h, v20.8h\n"
+ "fmax v11.8h, v11.8h, v20.8h\n"
+ "fmax v12.8h, v12.8h, v20.8h\n"
+ "fmax v13.8h, v13.8h, v20.8h\n"
+ "fmax v14.8h, v14.8h, v20.8h\n"
+ "fmax v15.8h, v15.8h, v20.8h\n"
+ "fmax v16.8h, v16.8h, v20.8h\n"
+ "fmax v17.8h, v17.8h, v20.8h\n"
+ "fmax v18.8h, v18.8h, v20.8h\n"
+ "fmax v19.8h, v19.8h, v20.8h\n"
"129:" // Height 3: No activation
"cmp x8, #0x20\n"
"bge 146f\n"
@@ -2424,484 +2424,484 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"170:" // Height 4: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w14, [x20, x15, 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, 171f\n"
- "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x13, [x21, #0x0]\n"
- "ldr x9, [x21, #0x8]\n"
- "ldr x27, [x21, #0x10]\n"
- "ldr x25, [x21, #0x18]\n"
+ "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x13, [x20, #0x0]\n"
+ "ldr x12, [x20, #0x8]\n"
+ "ldr x11, [x20, #0x10]\n"
+ "ldr x10, [x20, #0x18]\n"
"cbnz x15, 172f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x13, x13, x20, LSL #1\n"
- "add x9, x9, x20, LSL #1\n"
- "add x27, x27, x20, LSL #1\n"
- "add x25, x25, x20, LSL #1\n"
+ "add x12, x12, x20, LSL #1\n"
+ "add x11, x11, x20, LSL #1\n"
+ "add x10, x10, x20, LSL #1\n"
"b 172f\n"
"171:" // Height 4: setup direct input
"mov x13, %x[input_ptr]\n"
- "add x9, x13, x20, LSL #1\n"
- "add x27, x9, x20, LSL #1\n"
- "add x25, x27, x20, LSL #1\n"
+ "add x12, x13, x21, LSL #1\n"
+ "add x11, x12, x21, LSL #1\n"
+ "add x10, x11, x21, LSL #1\n"
"172:" // Height 4: input setup done
"cmp x14, #0x8\n"
"blt 175f\n"
"ldr q0, [x13, #0x0]\n"
"cmp x14, #0x10\n"
- "ldr q1, [x9, #0x0]\n"
- "ldr q2, [x27, #0x0]\n"
- "ldr q3, [x25, #0x0]\n"
+ "ldr q1, [x12, #0x0]\n"
+ "ldr q2, [x11, #0x0]\n"
+ "ldr q3, [x10, #0x0]\n"
"ldr q6, [x17, #0x0]\n"
"ldr q7, [x17, #0x10]\n"
"blt 174f\n"
"173:" // Height 4: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr x12, [x17, #0x28]\n"
+ "ldr x21, [x17, #0x28]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x38]\n"
+ "ldr x20, [x17, #0x38]\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "ldr d6, [x17, #0x20]\n"
+ "ldr d25, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
+ "mov v25.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x48]\n"
+ "ldr x21, [x17, #0x48]\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "ldr d7, [x17, #0x30]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x58]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "add x27, x27, #0x10\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "ldr d6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x68]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "add x25, x25, #0x10\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
- "ldr d7, [x17, #0x50]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x78]\n"
- "fmla v16.8h, v6.8h, v2.h[1]\n"
- "ldr x10, [x13, #0x8]\n"
- "fmla v20.8h, v6.8h, v3.h[1]\n"
- "ldr d6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0x88]\n"
- "fmla v17.8h, v7.8h, v2.h[1]\n"
- "ldr x28, [x9, #0x8]\n"
- "fmla v21.8h, v7.8h, v3.h[1]\n"
- "ldr d7, [x17, #0x70]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x98]\n"
- "fmla v18.8h, v6.8h, v2.h[1]\n"
- "ldr x26, [x27, #0x8]\n"
- "fmla v22.8h, v6.8h, v3.h[1]\n"
- "ldr d6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0xa8]\n"
- "fmla v19.8h, v7.8h, v2.h[1]\n"
- "ldr x24, [x25, #0x8]\n"
- "fmla v23.8h, v7.8h, v3.h[1]\n"
- "ldr d7, [x17, #0x90]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xb8]\n"
- "fmla v16.8h, v6.8h, v2.h[2]\n"
+ "ldr d24, [x17, #0x30]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[0]\n"
+ "fmla v14.8h, v25.8h, v1.h[0]\n"
+ "ldr x20, [x17, #0x58]\n"
+ "fmla v18.8h, v25.8h, v2.h[0]\n"
+ "add x11, x11, #0x10\n"
+ "fmla v22.8h, v25.8h, v3.h[0]\n"
+ "ldr d25, [x17, #0x40]\n"
+ "fmla v11.8h, v24.8h, v0.h[0]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[0]\n"
+ "ldr x21, [x17, #0x68]\n"
+ "fmla v19.8h, v24.8h, v2.h[0]\n"
+ "add x10, x10, #0x10\n"
+ "fmla v23.8h, v24.8h, v3.h[0]\n"
+ "ldr d24, [x17, #0x50]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[1]\n"
+ "fmla v12.8h, v25.8h, v1.h[1]\n"
+ "ldr x20, [x17, #0x78]\n"
+ "fmla v16.8h, v25.8h, v2.h[1]\n"
+ "ldr x25, [x13, #0x8]\n"
+ "fmla v20.8h, v25.8h, v3.h[1]\n"
+ "ldr d25, [x17, #0x60]\n"
+ "fmla v9.8h, v24.8h, v0.h[1]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[1]\n"
+ "ldr x21, [x17, #0x88]\n"
+ "fmla v17.8h, v24.8h, v2.h[1]\n"
+ "ldr x24, [x12, #0x8]\n"
+ "fmla v21.8h, v24.8h, v3.h[1]\n"
+ "ldr d24, [x17, #0x70]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[1]\n"
+ "fmla v14.8h, v25.8h, v1.h[1]\n"
+ "ldr x20, [x17, #0x98]\n"
+ "fmla v18.8h, v25.8h, v2.h[1]\n"
+ "ldr x23, [x11, #0x8]\n"
+ "fmla v22.8h, v25.8h, v3.h[1]\n"
+ "ldr d25, [x17, #0x80]\n"
+ "fmla v11.8h, v24.8h, v0.h[1]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[1]\n"
+ "ldr x21, [x17, #0xa8]\n"
+ "fmla v19.8h, v24.8h, v2.h[1]\n"
+ "ldr x22, [x10, #0x8]\n"
+ "fmla v23.8h, v24.8h, v3.h[1]\n"
+ "ldr d24, [x17, #0x90]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[2]\n"
+ "fmla v12.8h, v25.8h, v1.h[2]\n"
+ "ldr x20, [x17, #0xb8]\n"
+ "fmla v16.8h, v25.8h, v2.h[2]\n"
"sub x14, x14, #0x8\n"
- "fmla v20.8h, v6.8h, v3.h[2]\n"
- "ldr d6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xc8]\n"
- "fmla v17.8h, v7.8h, v2.h[2]\n"
+ "fmla v20.8h, v25.8h, v3.h[2]\n"
+ "ldr d25, [x17, #0xa0]\n"
+ "fmla v9.8h, v24.8h, v0.h[2]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[2]\n"
+ "ldr x21, [x17, #0xc8]\n"
+ "fmla v17.8h, v24.8h, v2.h[2]\n"
"cmp x14, #0x10\n"
- "fmla v21.8h, v7.8h, v3.h[2]\n"
- "ldr d7, [x17, #0xb0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xd8]\n"
- "fmla v18.8h, v6.8h, v2.h[2]\n"
+ "fmla v21.8h, v24.8h, v3.h[2]\n"
+ "ldr d24, [x17, #0xb0]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[2]\n"
+ "fmla v14.8h, v25.8h, v1.h[2]\n"
+ "ldr x20, [x17, #0xd8]\n"
+ "fmla v18.8h, v25.8h, v2.h[2]\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "fmla v22.8h, v6.8h, v3.h[2]\n"
- "ldr d6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xe8]\n"
- "fmla v19.8h, v7.8h, v2.h[2]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
- "fmla v23.8h, v7.8h, v3.h[2]\n"
- "ldr d7, [x17, #0xd0]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0xf8]\n"
- "fmla v16.8h, v6.8h, v2.h[3]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
- "fmla v20.8h, v6.8h, v3.h[3]\n"
- "ldr d6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x108]\n"
- "fmla v17.8h, v7.8h, v2.h[3]\n"
- "prfm pldl1keep, [x25, #0x80]\n"
- "fmla v21.8h, v7.8h, v3.h[3]\n"
- "ldr d7, [x17, #0xf0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0x118]\n"
- "fmla v18.8h, v6.8h, v2.h[3]\n"
- "fmla v22.8h, v6.8h, v3.h[3]\n"
- "ldr d6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x128]\n"
- "fmla v19.8h, v7.8h, v2.h[3]\n"
- "fmla v23.8h, v7.8h, v3.h[3]\n"
- "ldr d7, [x17, #0x110]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x138]\n"
- "fmla v16.8h, v6.8h, v2.h[4]\n"
- "fmla v20.8h, v6.8h, v3.h[4]\n"
- "ldr d6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x148]\n"
- "fmla v17.8h, v7.8h, v2.h[4]\n"
- "fmla v21.8h, v7.8h, v3.h[4]\n"
- "ldr d7, [x17, #0x130]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x158]\n"
- "fmla v18.8h, v6.8h, v2.h[4]\n"
- "fmla v22.8h, v6.8h, v3.h[4]\n"
- "ldr d6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x168]\n"
- "fmla v19.8h, v7.8h, v2.h[4]\n"
- "fmla v23.8h, v7.8h, v3.h[4]\n"
- "ldr d7, [x17, #0x150]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x178]\n"
- "fmla v16.8h, v6.8h, v2.h[5]\n"
- "fmla v20.8h, v6.8h, v3.h[5]\n"
- "ldr d6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x188]\n"
- "fmla v17.8h, v7.8h, v2.h[5]\n"
- "fmla v21.8h, v7.8h, v3.h[5]\n"
- "ldr d7, [x17, #0x170]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x198]\n"
- "fmla v18.8h, v6.8h, v2.h[5]\n"
- "fmla v22.8h, v6.8h, v3.h[5]\n"
- "ldr d6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x1a8]\n"
- "fmla v19.8h, v7.8h, v2.h[5]\n"
- "fmla v23.8h, v7.8h, v3.h[5]\n"
- "ldr d7, [x17, #0x190]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1b8]\n"
- "fmla v16.8h, v6.8h, v2.h[6]\n"
- "fmla v20.8h, v6.8h, v3.h[6]\n"
- "ldr d6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1c8]\n"
- "fmla v17.8h, v7.8h, v2.h[6]\n"
- "fmla v21.8h, v7.8h, v3.h[6]\n"
- "ldr d7, [x17, #0x1b0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1d8]\n"
- "fmla v18.8h, v6.8h, v2.h[6]\n"
- "fmla v22.8h, v6.8h, v3.h[6]\n"
- "ldr d6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1e8]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "fmla v23.8h, v7.8h, v3.h[6]\n"
- "ldr d7, [x17, #0x1d0]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x1f8]\n"
- "fmla v16.8h, v6.8h, v2.h[7]\n"
- "fmla v20.8h, v6.8h, v3.h[7]\n"
- "ldr d6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "fmla v21.8h, v7.8h, v3.h[7]\n"
- "ldr d7, [x17, #0x1f0]\n"
- "mov v7.d[1], x11\n"
+ "fmla v22.8h, v25.8h, v3.h[2]\n"
+ "ldr d25, [x17, #0xc0]\n"
+ "fmla v11.8h, v24.8h, v0.h[2]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[2]\n"
+ "ldr x21, [x17, #0xe8]\n"
+ "fmla v19.8h, v24.8h, v2.h[2]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
+ "fmla v23.8h, v24.8h, v3.h[2]\n"
+ "ldr d24, [x17, #0xd0]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[3]\n"
+ "fmla v12.8h, v25.8h, v1.h[3]\n"
+ "ldr x20, [x17, #0xf8]\n"
+ "fmla v16.8h, v25.8h, v2.h[3]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
+ "fmla v20.8h, v25.8h, v3.h[3]\n"
+ "ldr d25, [x17, #0xe0]\n"
+ "fmla v9.8h, v24.8h, v0.h[3]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[3]\n"
+ "ldr x21, [x17, #0x108]\n"
+ "fmla v17.8h, v24.8h, v2.h[3]\n"
+ "prfm pldl1keep, [x10, #0x80]\n"
+ "fmla v21.8h, v24.8h, v3.h[3]\n"
+ "ldr d24, [x17, #0xf0]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[3]\n"
+ "fmla v14.8h, v25.8h, v1.h[3]\n"
+ "ldr x20, [x17, #0x118]\n"
+ "fmla v18.8h, v25.8h, v2.h[3]\n"
+ "fmla v22.8h, v25.8h, v3.h[3]\n"
+ "ldr d25, [x17, #0x100]\n"
+ "fmla v11.8h, v24.8h, v0.h[3]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[3]\n"
+ "ldr x21, [x17, #0x128]\n"
+ "fmla v19.8h, v24.8h, v2.h[3]\n"
+ "fmla v23.8h, v24.8h, v3.h[3]\n"
+ "ldr d24, [x17, #0x110]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[4]\n"
+ "fmla v12.8h, v25.8h, v1.h[4]\n"
+ "ldr x20, [x17, #0x138]\n"
+ "fmla v16.8h, v25.8h, v2.h[4]\n"
+ "fmla v20.8h, v25.8h, v3.h[4]\n"
+ "ldr d25, [x17, #0x120]\n"
+ "fmla v9.8h, v24.8h, v0.h[4]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[4]\n"
+ "ldr x21, [x17, #0x148]\n"
+ "fmla v17.8h, v24.8h, v2.h[4]\n"
+ "fmla v21.8h, v24.8h, v3.h[4]\n"
+ "ldr d24, [x17, #0x130]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[4]\n"
+ "fmla v14.8h, v25.8h, v1.h[4]\n"
+ "ldr x20, [x17, #0x158]\n"
+ "fmla v18.8h, v25.8h, v2.h[4]\n"
+ "fmla v22.8h, v25.8h, v3.h[4]\n"
+ "ldr d25, [x17, #0x140]\n"
+ "fmla v11.8h, v24.8h, v0.h[4]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[4]\n"
+ "ldr x21, [x17, #0x168]\n"
+ "fmla v19.8h, v24.8h, v2.h[4]\n"
+ "fmla v23.8h, v24.8h, v3.h[4]\n"
+ "ldr d24, [x17, #0x150]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[5]\n"
+ "fmla v12.8h, v25.8h, v1.h[5]\n"
+ "ldr x20, [x17, #0x178]\n"
+ "fmla v16.8h, v25.8h, v2.h[5]\n"
+ "fmla v20.8h, v25.8h, v3.h[5]\n"
+ "ldr d25, [x17, #0x160]\n"
+ "fmla v9.8h, v24.8h, v0.h[5]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[5]\n"
+ "ldr x21, [x17, #0x188]\n"
+ "fmla v17.8h, v24.8h, v2.h[5]\n"
+ "fmla v21.8h, v24.8h, v3.h[5]\n"
+ "ldr d24, [x17, #0x170]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[5]\n"
+ "fmla v14.8h, v25.8h, v1.h[5]\n"
+ "ldr x20, [x17, #0x198]\n"
+ "fmla v18.8h, v25.8h, v2.h[5]\n"
+ "fmla v22.8h, v25.8h, v3.h[5]\n"
+ "ldr d25, [x17, #0x180]\n"
+ "fmla v11.8h, v24.8h, v0.h[5]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[5]\n"
+ "ldr x21, [x17, #0x1a8]\n"
+ "fmla v19.8h, v24.8h, v2.h[5]\n"
+ "fmla v23.8h, v24.8h, v3.h[5]\n"
+ "ldr d24, [x17, #0x190]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[6]\n"
+ "fmla v12.8h, v25.8h, v1.h[6]\n"
+ "ldr x20, [x17, #0x1b8]\n"
+ "fmla v16.8h, v25.8h, v2.h[6]\n"
+ "fmla v20.8h, v25.8h, v3.h[6]\n"
+ "ldr d25, [x17, #0x1a0]\n"
+ "fmla v9.8h, v24.8h, v0.h[6]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[6]\n"
+ "ldr x21, [x17, #0x1c8]\n"
+ "fmla v17.8h, v24.8h, v2.h[6]\n"
+ "fmla v21.8h, v24.8h, v3.h[6]\n"
+ "ldr d24, [x17, #0x1b0]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v10.8h, v25.8h, v0.h[6]\n"
+ "fmla v14.8h, v25.8h, v1.h[6]\n"
+ "ldr x20, [x17, #0x1d8]\n"
+ "fmla v18.8h, v25.8h, v2.h[6]\n"
+ "fmla v22.8h, v25.8h, v3.h[6]\n"
+ "ldr d25, [x17, #0x1c0]\n"
+ "fmla v11.8h, v24.8h, v0.h[6]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v15.8h, v24.8h, v1.h[6]\n"
+ "ldr x21, [x17, #0x1e8]\n"
+ "fmla v19.8h, v24.8h, v2.h[6]\n"
+ "fmla v23.8h, v24.8h, v3.h[6]\n"
+ "ldr d24, [x17, #0x1d0]\n"
+ "mov v24.d[1], x20\n"
+ "fmla v8.8h, v25.8h, v0.h[7]\n"
+ "fmla v12.8h, v25.8h, v1.h[7]\n"
+ "ldr x20, [x17, #0x1f8]\n"
+ "fmla v16.8h, v25.8h, v2.h[7]\n"
+ "fmla v20.8h, v25.8h, v3.h[7]\n"
+ "ldr d25, [x17, #0x1e0]\n"
+ "fmla v9.8h, v24.8h, v0.h[7]\n"
+ "mov v25.d[1], x21\n"
+ "fmla v13.8h, v24.8h, v1.h[7]\n"
+ "fmla v17.8h, v24.8h, v2.h[7]\n"
+ "fmla v21.8h, v24.8h, v3.h[7]\n"
+ "ldr d24, [x17, #0x1f0]\n"
+ "mov v24.d[1], x20\n"
"add x17, x17, #0x200\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "ldr x12, [x17, #0x8]\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x18]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
- "fmla v22.8h, v6.8h, v3.h[7]\n"
+ "fmla v10.8h, v25.8h, v0.h[7]\n"
+ "ldr x21, [x17, #0x8]\n"
+ "fmla v14.8h, v25.8h, v1.h[7]\n"
+ "ldr x20, [x17, #0x18]\n"
+ "fmla v18.8h, v25.8h, v2.h[7]\n"
+ "fmla v22.8h, v25.8h, v3.h[7]\n"
"ldr d6, [x17, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v24.8h, v0.h[7]\n"
"ldr d0, [x13, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "ldr d1, [x9, #0x0]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
- "ldr d2, [x27, #0x0]\n"
- "fmla v23.8h, v7.8h, v3.h[7]\n"
- "ldr d3, [x25, #0x0]\n"
+ "fmla v15.8h, v24.8h, v1.h[7]\n"
+ "ldr d1, [x12, #0x0]\n"
+ "fmla v19.8h, v24.8h, v2.h[7]\n"
+ "ldr d2, [x11, #0x0]\n"
+ "fmla v23.8h, v24.8h, v3.h[7]\n"
+ "ldr d3, [x10, #0x0]\n"
"ldr d7, [x17, #0x10]\n"
- "mov v6.d[1], x12\n"
- "mov v0.d[1], x10\n"
- "mov v1.d[1], x28\n"
- "mov v2.d[1], x26\n"
- "mov v3.d[1], x24\n"
- "mov v7.d[1], x11\n"
+ "mov v6.d[1], x21\n"
+ "mov v0.d[1], x25\n"
+ "mov v1.d[1], x24\n"
+ "mov v2.d[1], x23\n"
+ "mov v3.d[1], x22\n"
+ "mov v7.d[1], x20\n"
"bge 173b\n"
"174:" // Height 4: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "add x27, x27, #0x10\n"
+ "add x11, x11, #0x10\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
+ "ldr q25, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "add x25, x25, #0x10\n"
+ "add x10, x10, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
"sub x14, x14, #0x8\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
"prfm pldl1keep, [x13, #0x80]\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "prfm pldl1keep, [x25, #0x80]\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "ldr q6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
- "ldr q7, [x17, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "fmla v16.8h, v6.8h, v2.h[1]\n"
- "fmla v20.8h, v6.8h, v3.h[1]\n"
- "ldr q6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "fmla v17.8h, v7.8h, v2.h[1]\n"
- "fmla v21.8h, v7.8h, v3.h[1]\n"
- "ldr q7, [x17, #0x70]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "fmla v18.8h, v6.8h, v2.h[1]\n"
- "fmla v22.8h, v6.8h, v3.h[1]\n"
- "ldr q6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "fmla v19.8h, v7.8h, v2.h[1]\n"
- "fmla v23.8h, v7.8h, v3.h[1]\n"
- "ldr q7, [x17, #0x90]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "fmla v16.8h, v6.8h, v2.h[2]\n"
- "fmla v20.8h, v6.8h, v3.h[2]\n"
- "ldr q6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "fmla v17.8h, v7.8h, v2.h[2]\n"
- "fmla v21.8h, v7.8h, v3.h[2]\n"
- "ldr q7, [x17, #0xb0]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "fmla v18.8h, v6.8h, v2.h[2]\n"
- "fmla v22.8h, v6.8h, v3.h[2]\n"
- "ldr q6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "fmla v19.8h, v7.8h, v2.h[2]\n"
- "fmla v23.8h, v7.8h, v3.h[2]\n"
- "ldr q7, [x17, #0xd0]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "fmla v16.8h, v6.8h, v2.h[3]\n"
- "fmla v20.8h, v6.8h, v3.h[3]\n"
- "ldr q6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "fmla v17.8h, v7.8h, v2.h[3]\n"
- "fmla v21.8h, v7.8h, v3.h[3]\n"
- "ldr q7, [x17, #0xf0]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "fmla v18.8h, v6.8h, v2.h[3]\n"
- "fmla v22.8h, v6.8h, v3.h[3]\n"
- "ldr q6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "fmla v19.8h, v7.8h, v2.h[3]\n"
- "fmla v23.8h, v7.8h, v3.h[3]\n"
- "ldr q7, [x17, #0x110]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "fmla v16.8h, v6.8h, v2.h[4]\n"
- "fmla v20.8h, v6.8h, v3.h[4]\n"
- "ldr q6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "fmla v17.8h, v7.8h, v2.h[4]\n"
- "fmla v21.8h, v7.8h, v3.h[4]\n"
- "ldr q7, [x17, #0x130]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "fmla v18.8h, v6.8h, v2.h[4]\n"
- "fmla v22.8h, v6.8h, v3.h[4]\n"
- "ldr q6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "fmla v19.8h, v7.8h, v2.h[4]\n"
- "fmla v23.8h, v7.8h, v3.h[4]\n"
- "ldr q7, [x17, #0x150]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "fmla v16.8h, v6.8h, v2.h[5]\n"
- "fmla v20.8h, v6.8h, v3.h[5]\n"
- "ldr q6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "fmla v17.8h, v7.8h, v2.h[5]\n"
- "fmla v21.8h, v7.8h, v3.h[5]\n"
- "ldr q7, [x17, #0x170]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "fmla v18.8h, v6.8h, v2.h[5]\n"
- "fmla v22.8h, v6.8h, v3.h[5]\n"
- "ldr q6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "fmla v19.8h, v7.8h, v2.h[5]\n"
- "fmla v23.8h, v7.8h, v3.h[5]\n"
- "ldr q7, [x17, #0x190]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "fmla v16.8h, v6.8h, v2.h[6]\n"
- "fmla v20.8h, v6.8h, v3.h[6]\n"
- "ldr q6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "fmla v17.8h, v7.8h, v2.h[6]\n"
- "fmla v21.8h, v7.8h, v3.h[6]\n"
- "ldr q7, [x17, #0x1b0]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "fmla v18.8h, v6.8h, v2.h[6]\n"
- "fmla v22.8h, v6.8h, v3.h[6]\n"
- "ldr q6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "fmla v23.8h, v7.8h, v3.h[6]\n"
- "ldr q7, [x17, #0x1d0]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "fmla v16.8h, v6.8h, v2.h[7]\n"
- "fmla v20.8h, v6.8h, v3.h[7]\n"
- "ldr q6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "fmla v21.8h, v7.8h, v3.h[7]\n"
- "ldr q7, [x17, #0x1f0]\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
+ "ldr q24, [x17, #0x30]\n"
+ "fmla v10.8h, v25.8h, v0.h[0]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
+ "fmla v14.8h, v25.8h, v1.h[0]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
+ "fmla v18.8h, v25.8h, v2.h[0]\n"
+ "prfm pldl1keep, [x10, #0x80]\n"
+ "fmla v22.8h, v25.8h, v3.h[0]\n"
+ "ldr q25, [x17, #0x40]\n"
+ "fmla v11.8h, v24.8h, v0.h[0]\n"
+ "fmla v15.8h, v24.8h, v1.h[0]\n"
+ "fmla v19.8h, v24.8h, v2.h[0]\n"
+ "fmla v23.8h, v24.8h, v3.h[0]\n"
+ "ldr q24, [x17, #0x50]\n"
+ "fmla v8.8h, v25.8h, v0.h[1]\n"
+ "fmla v12.8h, v25.8h, v1.h[1]\n"
+ "fmla v16.8h, v25.8h, v2.h[1]\n"
+ "fmla v20.8h, v25.8h, v3.h[1]\n"
+ "ldr q25, [x17, #0x60]\n"
+ "fmla v9.8h, v24.8h, v0.h[1]\n"
+ "fmla v13.8h, v24.8h, v1.h[1]\n"
+ "fmla v17.8h, v24.8h, v2.h[1]\n"
+ "fmla v21.8h, v24.8h, v3.h[1]\n"
+ "ldr q24, [x17, #0x70]\n"
+ "fmla v10.8h, v25.8h, v0.h[1]\n"
+ "fmla v14.8h, v25.8h, v1.h[1]\n"
+ "fmla v18.8h, v25.8h, v2.h[1]\n"
+ "fmla v22.8h, v25.8h, v3.h[1]\n"
+ "ldr q25, [x17, #0x80]\n"
+ "fmla v11.8h, v24.8h, v0.h[1]\n"
+ "fmla v15.8h, v24.8h, v1.h[1]\n"
+ "fmla v19.8h, v24.8h, v2.h[1]\n"
+ "fmla v23.8h, v24.8h, v3.h[1]\n"
+ "ldr q24, [x17, #0x90]\n"
+ "fmla v8.8h, v25.8h, v0.h[2]\n"
+ "fmla v12.8h, v25.8h, v1.h[2]\n"
+ "fmla v16.8h, v25.8h, v2.h[2]\n"
+ "fmla v20.8h, v25.8h, v3.h[2]\n"
+ "ldr q25, [x17, #0xa0]\n"
+ "fmla v9.8h, v24.8h, v0.h[2]\n"
+ "fmla v13.8h, v24.8h, v1.h[2]\n"
+ "fmla v17.8h, v24.8h, v2.h[2]\n"
+ "fmla v21.8h, v24.8h, v3.h[2]\n"
+ "ldr q24, [x17, #0xb0]\n"
+ "fmla v10.8h, v25.8h, v0.h[2]\n"
+ "fmla v14.8h, v25.8h, v1.h[2]\n"
+ "fmla v18.8h, v25.8h, v2.h[2]\n"
+ "fmla v22.8h, v25.8h, v3.h[2]\n"
+ "ldr q25, [x17, #0xc0]\n"
+ "fmla v11.8h, v24.8h, v0.h[2]\n"
+ "fmla v15.8h, v24.8h, v1.h[2]\n"
+ "fmla v19.8h, v24.8h, v2.h[2]\n"
+ "fmla v23.8h, v24.8h, v3.h[2]\n"
+ "ldr q24, [x17, #0xd0]\n"
+ "fmla v8.8h, v25.8h, v0.h[3]\n"
+ "fmla v12.8h, v25.8h, v1.h[3]\n"
+ "fmla v16.8h, v25.8h, v2.h[3]\n"
+ "fmla v20.8h, v25.8h, v3.h[3]\n"
+ "ldr q25, [x17, #0xe0]\n"
+ "fmla v9.8h, v24.8h, v0.h[3]\n"
+ "fmla v13.8h, v24.8h, v1.h[3]\n"
+ "fmla v17.8h, v24.8h, v2.h[3]\n"
+ "fmla v21.8h, v24.8h, v3.h[3]\n"
+ "ldr q24, [x17, #0xf0]\n"
+ "fmla v10.8h, v25.8h, v0.h[3]\n"
+ "fmla v14.8h, v25.8h, v1.h[3]\n"
+ "fmla v18.8h, v25.8h, v2.h[3]\n"
+ "fmla v22.8h, v25.8h, v3.h[3]\n"
+ "ldr q25, [x17, #0x100]\n"
+ "fmla v11.8h, v24.8h, v0.h[3]\n"
+ "fmla v15.8h, v24.8h, v1.h[3]\n"
+ "fmla v19.8h, v24.8h, v2.h[3]\n"
+ "fmla v23.8h, v24.8h, v3.h[3]\n"
+ "ldr q24, [x17, #0x110]\n"
+ "fmla v8.8h, v25.8h, v0.h[4]\n"
+ "fmla v12.8h, v25.8h, v1.h[4]\n"
+ "fmla v16.8h, v25.8h, v2.h[4]\n"
+ "fmla v20.8h, v25.8h, v3.h[4]\n"
+ "ldr q25, [x17, #0x120]\n"
+ "fmla v9.8h, v24.8h, v0.h[4]\n"
+ "fmla v13.8h, v24.8h, v1.h[4]\n"
+ "fmla v17.8h, v24.8h, v2.h[4]\n"
+ "fmla v21.8h, v24.8h, v3.h[4]\n"
+ "ldr q24, [x17, #0x130]\n"
+ "fmla v10.8h, v25.8h, v0.h[4]\n"
+ "fmla v14.8h, v25.8h, v1.h[4]\n"
+ "fmla v18.8h, v25.8h, v2.h[4]\n"
+ "fmla v22.8h, v25.8h, v3.h[4]\n"
+ "ldr q25, [x17, #0x140]\n"
+ "fmla v11.8h, v24.8h, v0.h[4]\n"
+ "fmla v15.8h, v24.8h, v1.h[4]\n"
+ "fmla v19.8h, v24.8h, v2.h[4]\n"
+ "fmla v23.8h, v24.8h, v3.h[4]\n"
+ "ldr q24, [x17, #0x150]\n"
+ "fmla v8.8h, v25.8h, v0.h[5]\n"
+ "fmla v12.8h, v25.8h, v1.h[5]\n"
+ "fmla v16.8h, v25.8h, v2.h[5]\n"
+ "fmla v20.8h, v25.8h, v3.h[5]\n"
+ "ldr q25, [x17, #0x160]\n"
+ "fmla v9.8h, v24.8h, v0.h[5]\n"
+ "fmla v13.8h, v24.8h, v1.h[5]\n"
+ "fmla v17.8h, v24.8h, v2.h[5]\n"
+ "fmla v21.8h, v24.8h, v3.h[5]\n"
+ "ldr q24, [x17, #0x170]\n"
+ "fmla v10.8h, v25.8h, v0.h[5]\n"
+ "fmla v14.8h, v25.8h, v1.h[5]\n"
+ "fmla v18.8h, v25.8h, v2.h[5]\n"
+ "fmla v22.8h, v25.8h, v3.h[5]\n"
+ "ldr q25, [x17, #0x180]\n"
+ "fmla v11.8h, v24.8h, v0.h[5]\n"
+ "fmla v15.8h, v24.8h, v1.h[5]\n"
+ "fmla v19.8h, v24.8h, v2.h[5]\n"
+ "fmla v23.8h, v24.8h, v3.h[5]\n"
+ "ldr q24, [x17, #0x190]\n"
+ "fmla v8.8h, v25.8h, v0.h[6]\n"
+ "fmla v12.8h, v25.8h, v1.h[6]\n"
+ "fmla v16.8h, v25.8h, v2.h[6]\n"
+ "fmla v20.8h, v25.8h, v3.h[6]\n"
+ "ldr q25, [x17, #0x1a0]\n"
+ "fmla v9.8h, v24.8h, v0.h[6]\n"
+ "fmla v13.8h, v24.8h, v1.h[6]\n"
+ "fmla v17.8h, v24.8h, v2.h[6]\n"
+ "fmla v21.8h, v24.8h, v3.h[6]\n"
+ "ldr q24, [x17, #0x1b0]\n"
+ "fmla v10.8h, v25.8h, v0.h[6]\n"
+ "fmla v14.8h, v25.8h, v1.h[6]\n"
+ "fmla v18.8h, v25.8h, v2.h[6]\n"
+ "fmla v22.8h, v25.8h, v3.h[6]\n"
+ "ldr q25, [x17, #0x1c0]\n"
+ "fmla v11.8h, v24.8h, v0.h[6]\n"
+ "fmla v15.8h, v24.8h, v1.h[6]\n"
+ "fmla v19.8h, v24.8h, v2.h[6]\n"
+ "fmla v23.8h, v24.8h, v3.h[6]\n"
+ "ldr q24, [x17, #0x1d0]\n"
+ "fmla v8.8h, v25.8h, v0.h[7]\n"
+ "fmla v12.8h, v25.8h, v1.h[7]\n"
+ "fmla v16.8h, v25.8h, v2.h[7]\n"
+ "fmla v20.8h, v25.8h, v3.h[7]\n"
+ "ldr q25, [x17, #0x1e0]\n"
+ "fmla v9.8h, v24.8h, v0.h[7]\n"
+ "fmla v13.8h, v24.8h, v1.h[7]\n"
+ "fmla v17.8h, v24.8h, v2.h[7]\n"
+ "fmla v21.8h, v24.8h, v3.h[7]\n"
+ "ldr q24, [x17, #0x1f0]\n"
+ "fmla v10.8h, v25.8h, v0.h[7]\n"
"add x17, x17, #0x200\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
- "fmla v22.8h, v6.8h, v3.h[7]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
- "fmla v23.8h, v7.8h, v3.h[7]\n"
+ "fmla v14.8h, v25.8h, v1.h[7]\n"
+ "fmla v18.8h, v25.8h, v2.h[7]\n"
+ "fmla v22.8h, v25.8h, v3.h[7]\n"
+ "fmla v11.8h, v24.8h, v0.h[7]\n"
+ "fmla v15.8h, v24.8h, v1.h[7]\n"
+ "fmla v19.8h, v24.8h, v2.h[7]\n"
+ "fmla v23.8h, v24.8h, v3.h[7]\n"
"175:" // Height 4: Multiply loop: Main loop skip
"cbz x14, 177f\n"
"176:" // Height 4: Multiply loop: Odd block loop
- "ldr h0, [x13], #0x2\n"
+ "ldr h3, [x13], #0x2\n"
"sub x14, x14, #0x1\n"
- "ldr h1, [x9], #0x2\n"
- "ldr h2, [x27], #0x2\n"
- "ldr h3, [x25], #0x2\n"
- "ldr q6, [x17, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x10]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "fmla v16.8h, v6.8h, v2.h[0]\n"
- "fmla v20.8h, v6.8h, v3.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "fmla v17.8h, v7.8h, v2.h[0]\n"
- "fmla v21.8h, v7.8h, v3.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr h2, [x12], #0x2\n"
+ "ldr h1, [x11], #0x2\n"
+ "ldr h0, [x10], #0x2\n"
+ "ldr q25, [x17, #0x0]\n"
+ "fmla v8.8h, v25.8h, v3.h[0]\n"
+ "ldr q24, [x17, #0x10]\n"
+ "fmla v12.8h, v25.8h, v2.h[0]\n"
+ "fmla v16.8h, v25.8h, v1.h[0]\n"
+ "fmla v20.8h, v25.8h, v0.h[0]\n"
+ "ldr q25, [x17, #0x20]\n"
+ "fmla v9.8h, v24.8h, v3.h[0]\n"
+ "fmla v13.8h, v24.8h, v2.h[0]\n"
+ "fmla v17.8h, v24.8h, v1.h[0]\n"
+ "fmla v21.8h, v24.8h, v0.h[0]\n"
+ "ldr q24, [x17, #0x30]\n"
+ "fmla v10.8h, v25.8h, v3.h[0]\n"
"add x17, x17, #0x40\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
+ "fmla v14.8h, v25.8h, v2.h[0]\n"
+ "fmla v18.8h, v25.8h, v1.h[0]\n"
+ "fmla v22.8h, v25.8h, v0.h[0]\n"
+ "fmla v11.8h, v24.8h, v3.h[0]\n"
+ "fmla v15.8h, v24.8h, v2.h[0]\n"
+ "fmla v19.8h, v24.8h, v1.h[0]\n"
+ "fmla v23.8h, v24.8h, v0.h[0]\n"
"cbnz x14, 176b\n"
"177:" // Height 4: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -2918,41 +2918,41 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"prfm pstl1keep, [x23, #0x0]\n"
"tbz %x[flags], #1, 178f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v0.8h\n"
- "fmin v9.8h, v9.8h, v0.8h\n"
- "fmin v10.8h, v10.8h, v0.8h\n"
- "fmin v11.8h, v11.8h, v0.8h\n"
- "fmin v12.8h, v12.8h, v0.8h\n"
- "fmin v13.8h, v13.8h, v0.8h\n"
- "fmin v14.8h, v14.8h, v0.8h\n"
- "fmin v15.8h, v15.8h, v0.8h\n"
- "fmin v16.8h, v16.8h, v0.8h\n"
- "fmin v17.8h, v17.8h, v0.8h\n"
- "fmin v18.8h, v18.8h, v0.8h\n"
- "fmin v19.8h, v19.8h, v0.8h\n"
- "fmin v20.8h, v20.8h, v0.8h\n"
- "fmin v21.8h, v21.8h, v0.8h\n"
- "fmin v22.8h, v22.8h, v0.8h\n"
- "fmin v23.8h, v23.8h, v0.8h\n"
+ "ld1r { v24.8h }, [x20]\n"
+ "fmin v8.8h, v8.8h, v24.8h\n"
+ "fmin v9.8h, v9.8h, v24.8h\n"
+ "fmin v10.8h, v10.8h, v24.8h\n"
+ "fmin v11.8h, v11.8h, v24.8h\n"
+ "fmin v12.8h, v12.8h, v24.8h\n"
+ "fmin v13.8h, v13.8h, v24.8h\n"
+ "fmin v14.8h, v14.8h, v24.8h\n"
+ "fmin v15.8h, v15.8h, v24.8h\n"
+ "fmin v16.8h, v16.8h, v24.8h\n"
+ "fmin v17.8h, v17.8h, v24.8h\n"
+ "fmin v18.8h, v18.8h, v24.8h\n"
+ "fmin v19.8h, v19.8h, v24.8h\n"
+ "fmin v20.8h, v20.8h, v24.8h\n"
+ "fmin v21.8h, v21.8h, v24.8h\n"
+ "fmin v22.8h, v22.8h, v24.8h\n"
+ "fmin v23.8h, v23.8h, v24.8h\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmax v8.8h, v8.8h, v0.8h\n"
- "fmax v9.8h, v9.8h, v0.8h\n"
- "fmax v10.8h, v10.8h, v0.8h\n"
- "fmax v11.8h, v11.8h, v0.8h\n"
- "fmax v12.8h, v12.8h, v0.8h\n"
- "fmax v13.8h, v13.8h, v0.8h\n"
- "fmax v14.8h, v14.8h, v0.8h\n"
- "fmax v15.8h, v15.8h, v0.8h\n"
- "fmax v16.8h, v16.8h, v0.8h\n"
- "fmax v17.8h, v17.8h, v0.8h\n"
- "fmax v18.8h, v18.8h, v0.8h\n"
- "fmax v19.8h, v19.8h, v0.8h\n"
- "fmax v20.8h, v20.8h, v0.8h\n"
- "fmax v21.8h, v21.8h, v0.8h\n"
- "fmax v22.8h, v22.8h, v0.8h\n"
- "fmax v23.8h, v23.8h, v0.8h\n"
+ "ld1r { v24.8h }, [x20]\n"
+ "fmax v8.8h, v8.8h, v24.8h\n"
+ "fmax v9.8h, v9.8h, v24.8h\n"
+ "fmax v10.8h, v10.8h, v24.8h\n"
+ "fmax v11.8h, v11.8h, v24.8h\n"
+ "fmax v12.8h, v12.8h, v24.8h\n"
+ "fmax v13.8h, v13.8h, v24.8h\n"
+ "fmax v14.8h, v14.8h, v24.8h\n"
+ "fmax v15.8h, v15.8h, v24.8h\n"
+ "fmax v16.8h, v16.8h, v24.8h\n"
+ "fmax v17.8h, v17.8h, v24.8h\n"
+ "fmax v18.8h, v18.8h, v24.8h\n"
+ "fmax v19.8h, v19.8h, v24.8h\n"
+ "fmax v20.8h, v20.8h, v24.8h\n"
+ "fmax v21.8h, v21.8h, v24.8h\n"
+ "fmax v22.8h, v22.8h, v24.8h\n"
+ "fmax v23.8h, v23.8h, v24.8h\n"
"178:" // Height 4: No activation
"cmp x8, #0x20\n"
"bge 195f\n"
@@ -3493,564 +3493,564 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"219:" // Height 5: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w14, [x20, x15, 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, 220f\n"
- "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x13, [x21, #0x0]\n"
- "ldr x9, [x21, #0x8]\n"
- "ldr x27, [x21, #0x10]\n"
- "ldr x25, [x21, #0x18]\n"
- "ldr x23, [x21, #0x20]\n"
+ "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x13, [x20, #0x0]\n"
+ "ldr x12, [x20, #0x8]\n"
+ "ldr x11, [x20, #0x10]\n"
+ "ldr x10, [x20, #0x18]\n"
+ "ldr x9, [x20, #0x20]\n"
"cbnz x15, 221f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x13, x13, x20, LSL #1\n"
+ "add x12, x12, x20, LSL #1\n"
+ "add x11, x11, x20, LSL #1\n"
+ "add x10, x10, x20, LSL #1\n"
"add x9, x9, x20, LSL #1\n"
- "add x27, x27, x20, LSL #1\n"
- "add x25, x25, x20, LSL #1\n"
- "add x23, x23, x20, LSL #1\n"
"b 221f\n"
"220:" // Height 5: setup direct input
"mov x13, %x[input_ptr]\n"
- "add x9, x13, x20, LSL #1\n"
- "add x27, x9, x20, LSL #1\n"
- "add x25, x27, x20, LSL #1\n"
- "add x23, x25, x20, LSL #1\n"
+ "add x12, x13, x21, LSL #1\n"
+ "add x11, x12, x21, LSL #1\n"
+ "add x10, x11, x21, LSL #1\n"
+ "add x9, x10, x21, LSL #1\n"
"221:" // Height 5: input setup done
"cmp x14, #0x8\n"
"blt 224f\n"
"ldr q0, [x13, #0x0]\n"
"cmp x14, #0x10\n"
- "ldr q1, [x9, #0x0]\n"
- "ldr q2, [x27, #0x0]\n"
- "ldr q3, [x25, #0x0]\n"
- "ldr q4, [x23, #0x0]\n"
+ "ldr q1, [x12, #0x0]\n"
+ "ldr q2, [x11, #0x0]\n"
+ "ldr q3, [x10, #0x0]\n"
+ "ldr q4, [x9, #0x0]\n"
"ldr q6, [x17, #0x0]\n"
"ldr q7, [x17, #0x10]\n"
"blt 223f\n"
"222:" // Height 5: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr x12, [x17, #0x28]\n"
+ "ldr x21, [x17, #0x28]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x38]\n"
+ "ldr x20, [x17, #0x38]\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v24.8h, v6.8h, v4.h[0]\n"
- "ldr d6, [x17, #0x20]\n"
+ "ldr d29, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
+ "mov v29.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x48]\n"
+ "ldr x21, [x17, #0x48]\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "add x27, x27, #0x10\n"
+ "add x11, x11, #0x10\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "add x25, x25, #0x10\n"
+ "add x10, x10, #0x10\n"
"fmla v25.8h, v7.8h, v4.h[0]\n"
- "ldr d7, [x17, #0x30]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x58]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "add x23, x23, #0x10\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "ldr x10, [x13, #0x8]\n"
- "fmla v26.8h, v6.8h, v4.h[0]\n"
- "ldr d6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x68]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "ldr x28, [x9, #0x8]\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
- "ldr x26, [x27, #0x8]\n"
- "fmla v27.8h, v7.8h, v4.h[0]\n"
- "ldr d7, [x17, #0x50]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x78]\n"
- "fmla v16.8h, v6.8h, v2.h[1]\n"
- "ldr x24, [x25, #0x8]\n"
- "fmla v20.8h, v6.8h, v3.h[1]\n"
- "ldr x22, [x23, #0x8]\n"
- "fmla v24.8h, v6.8h, v4.h[1]\n"
- "ldr d6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0x88]\n"
- "fmla v17.8h, v7.8h, v2.h[1]\n"
+ "ldr d28, [x17, #0x30]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[0]\n"
+ "fmla v14.8h, v29.8h, v1.h[0]\n"
+ "ldr x20, [x17, #0x58]\n"
+ "fmla v18.8h, v29.8h, v2.h[0]\n"
+ "add x9, x9, #0x10\n"
+ "fmla v22.8h, v29.8h, v3.h[0]\n"
+ "ldr x26, [x13, #0x8]\n"
+ "fmla v26.8h, v29.8h, v4.h[0]\n"
+ "ldr d29, [x17, #0x40]\n"
+ "fmla v11.8h, v28.8h, v0.h[0]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[0]\n"
+ "ldr x21, [x17, #0x68]\n"
+ "fmla v19.8h, v28.8h, v2.h[0]\n"
+ "ldr x25, [x12, #0x8]\n"
+ "fmla v23.8h, v28.8h, v3.h[0]\n"
+ "ldr x24, [x11, #0x8]\n"
+ "fmla v27.8h, v28.8h, v4.h[0]\n"
+ "ldr d28, [x17, #0x50]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[1]\n"
+ "fmla v12.8h, v29.8h, v1.h[1]\n"
+ "ldr x20, [x17, #0x78]\n"
+ "fmla v16.8h, v29.8h, v2.h[1]\n"
+ "ldr x23, [x10, #0x8]\n"
+ "fmla v20.8h, v29.8h, v3.h[1]\n"
+ "ldr x22, [x9, #0x8]\n"
+ "fmla v24.8h, v29.8h, v4.h[1]\n"
+ "ldr d29, [x17, #0x60]\n"
+ "fmla v9.8h, v28.8h, v0.h[1]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[1]\n"
+ "ldr x21, [x17, #0x88]\n"
+ "fmla v17.8h, v28.8h, v2.h[1]\n"
"sub x14, x14, #0x8\n"
- "fmla v21.8h, v7.8h, v3.h[1]\n"
+ "fmla v21.8h, v28.8h, v3.h[1]\n"
"cmp x14, #0x10\n"
- "fmla v25.8h, v7.8h, v4.h[1]\n"
- "ldr d7, [x17, #0x70]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x98]\n"
- "fmla v18.8h, v6.8h, v2.h[1]\n"
+ "fmla v25.8h, v28.8h, v4.h[1]\n"
+ "ldr d28, [x17, #0x70]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[1]\n"
+ "fmla v14.8h, v29.8h, v1.h[1]\n"
+ "ldr x20, [x17, #0x98]\n"
+ "fmla v18.8h, v29.8h, v2.h[1]\n"
"prfm pldl1keep, [x13, #0x80]\n"
- "fmla v22.8h, v6.8h, v3.h[1]\n"
+ "fmla v22.8h, v29.8h, v3.h[1]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
+ "fmla v26.8h, v29.8h, v4.h[1]\n"
+ "ldr d29, [x17, #0x80]\n"
+ "fmla v11.8h, v28.8h, v0.h[1]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[1]\n"
+ "ldr x21, [x17, #0xa8]\n"
+ "fmla v19.8h, v28.8h, v2.h[1]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
+ "fmla v23.8h, v28.8h, v3.h[1]\n"
+ "prfm pldl1keep, [x10, #0x80]\n"
+ "fmla v27.8h, v28.8h, v4.h[1]\n"
+ "ldr d28, [x17, #0x90]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[2]\n"
+ "fmla v12.8h, v29.8h, v1.h[2]\n"
+ "ldr x20, [x17, #0xb8]\n"
+ "fmla v16.8h, v29.8h, v2.h[2]\n"
"prfm pldl1keep, [x9, #0x80]\n"
- "fmla v26.8h, v6.8h, v4.h[1]\n"
- "ldr d6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0xa8]\n"
- "fmla v19.8h, v7.8h, v2.h[1]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
- "fmla v23.8h, v7.8h, v3.h[1]\n"
- "prfm pldl1keep, [x25, #0x80]\n"
- "fmla v27.8h, v7.8h, v4.h[1]\n"
- "ldr d7, [x17, #0x90]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xb8]\n"
- "fmla v16.8h, v6.8h, v2.h[2]\n"
- "prfm pldl1keep, [x23, #0x80]\n"
- "fmla v20.8h, v6.8h, v3.h[2]\n"
- "fmla v24.8h, v6.8h, v4.h[2]\n"
- "ldr d6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xc8]\n"
- "fmla v17.8h, v7.8h, v2.h[2]\n"
- "fmla v21.8h, v7.8h, v3.h[2]\n"
- "fmla v25.8h, v7.8h, v4.h[2]\n"
- "ldr d7, [x17, #0xb0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xd8]\n"
- "fmla v18.8h, v6.8h, v2.h[2]\n"
- "fmla v22.8h, v6.8h, v3.h[2]\n"
- "fmla v26.8h, v6.8h, v4.h[2]\n"
- "ldr d6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xe8]\n"
- "fmla v19.8h, v7.8h, v2.h[2]\n"
- "fmla v23.8h, v7.8h, v3.h[2]\n"
- "fmla v27.8h, v7.8h, v4.h[2]\n"
- "ldr d7, [x17, #0xd0]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0xf8]\n"
- "fmla v16.8h, v6.8h, v2.h[3]\n"
- "fmla v20.8h, v6.8h, v3.h[3]\n"
- "fmla v24.8h, v6.8h, v4.h[3]\n"
- "ldr d6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x108]\n"
- "fmla v17.8h, v7.8h, v2.h[3]\n"
- "fmla v21.8h, v7.8h, v3.h[3]\n"
- "fmla v25.8h, v7.8h, v4.h[3]\n"
- "ldr d7, [x17, #0xf0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0x118]\n"
- "fmla v18.8h, v6.8h, v2.h[3]\n"
- "fmla v22.8h, v6.8h, v3.h[3]\n"
- "fmla v26.8h, v6.8h, v4.h[3]\n"
- "ldr d6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x128]\n"
- "fmla v19.8h, v7.8h, v2.h[3]\n"
- "fmla v23.8h, v7.8h, v3.h[3]\n"
- "fmla v27.8h, v7.8h, v4.h[3]\n"
- "ldr d7, [x17, #0x110]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x138]\n"
- "fmla v16.8h, v6.8h, v2.h[4]\n"
- "fmla v20.8h, v6.8h, v3.h[4]\n"
- "fmla v24.8h, v6.8h, v4.h[4]\n"
- "ldr d6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x148]\n"
- "fmla v17.8h, v7.8h, v2.h[4]\n"
- "fmla v21.8h, v7.8h, v3.h[4]\n"
- "fmla v25.8h, v7.8h, v4.h[4]\n"
- "ldr d7, [x17, #0x130]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x158]\n"
- "fmla v18.8h, v6.8h, v2.h[4]\n"
- "fmla v22.8h, v6.8h, v3.h[4]\n"
- "fmla v26.8h, v6.8h, v4.h[4]\n"
- "ldr d6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x168]\n"
- "fmla v19.8h, v7.8h, v2.h[4]\n"
- "fmla v23.8h, v7.8h, v3.h[4]\n"
- "fmla v27.8h, v7.8h, v4.h[4]\n"
- "ldr d7, [x17, #0x150]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x178]\n"
- "fmla v16.8h, v6.8h, v2.h[5]\n"
- "fmla v20.8h, v6.8h, v3.h[5]\n"
- "fmla v24.8h, v6.8h, v4.h[5]\n"
- "ldr d6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x188]\n"
- "fmla v17.8h, v7.8h, v2.h[5]\n"
- "fmla v21.8h, v7.8h, v3.h[5]\n"
- "fmla v25.8h, v7.8h, v4.h[5]\n"
- "ldr d7, [x17, #0x170]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x198]\n"
- "fmla v18.8h, v6.8h, v2.h[5]\n"
- "fmla v22.8h, v6.8h, v3.h[5]\n"
- "fmla v26.8h, v6.8h, v4.h[5]\n"
- "ldr d6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x1a8]\n"
- "fmla v19.8h, v7.8h, v2.h[5]\n"
- "fmla v23.8h, v7.8h, v3.h[5]\n"
- "fmla v27.8h, v7.8h, v4.h[5]\n"
- "ldr d7, [x17, #0x190]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1b8]\n"
- "fmla v16.8h, v6.8h, v2.h[6]\n"
- "fmla v20.8h, v6.8h, v3.h[6]\n"
- "fmla v24.8h, v6.8h, v4.h[6]\n"
- "ldr d6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1c8]\n"
- "fmla v17.8h, v7.8h, v2.h[6]\n"
- "fmla v21.8h, v7.8h, v3.h[6]\n"
- "fmla v25.8h, v7.8h, v4.h[6]\n"
- "ldr d7, [x17, #0x1b0]\n"
- "mov v7.d[1], x11\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1d8]\n"
- "fmla v18.8h, v6.8h, v2.h[6]\n"
- "fmla v22.8h, v6.8h, v3.h[6]\n"
- "fmla v26.8h, v6.8h, v4.h[6]\n"
- "ldr d6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1e8]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "fmla v23.8h, v7.8h, v3.h[6]\n"
- "fmla v27.8h, v7.8h, v4.h[6]\n"
- "ldr d7, [x17, #0x1d0]\n"
- "mov v7.d[1], x11\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x1f8]\n"
- "fmla v16.8h, v6.8h, v2.h[7]\n"
- "fmla v20.8h, v6.8h, v3.h[7]\n"
- "fmla v24.8h, v6.8h, v4.h[7]\n"
- "ldr d6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "mov v6.d[1], x12\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "fmla v21.8h, v7.8h, v3.h[7]\n"
- "fmla v25.8h, v7.8h, v4.h[7]\n"
- "ldr d7, [x17, #0x1f0]\n"
- "mov v7.d[1], x11\n"
+ "fmla v20.8h, v29.8h, v3.h[2]\n"
+ "fmla v24.8h, v29.8h, v4.h[2]\n"
+ "ldr d29, [x17, #0xa0]\n"
+ "fmla v9.8h, v28.8h, v0.h[2]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[2]\n"
+ "ldr x21, [x17, #0xc8]\n"
+ "fmla v17.8h, v28.8h, v2.h[2]\n"
+ "fmla v21.8h, v28.8h, v3.h[2]\n"
+ "fmla v25.8h, v28.8h, v4.h[2]\n"
+ "ldr d28, [x17, #0xb0]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[2]\n"
+ "fmla v14.8h, v29.8h, v1.h[2]\n"
+ "ldr x20, [x17, #0xd8]\n"
+ "fmla v18.8h, v29.8h, v2.h[2]\n"
+ "fmla v22.8h, v29.8h, v3.h[2]\n"
+ "fmla v26.8h, v29.8h, v4.h[2]\n"
+ "ldr d29, [x17, #0xc0]\n"
+ "fmla v11.8h, v28.8h, v0.h[2]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[2]\n"
+ "ldr x21, [x17, #0xe8]\n"
+ "fmla v19.8h, v28.8h, v2.h[2]\n"
+ "fmla v23.8h, v28.8h, v3.h[2]\n"
+ "fmla v27.8h, v28.8h, v4.h[2]\n"
+ "ldr d28, [x17, #0xd0]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[3]\n"
+ "fmla v12.8h, v29.8h, v1.h[3]\n"
+ "ldr x20, [x17, #0xf8]\n"
+ "fmla v16.8h, v29.8h, v2.h[3]\n"
+ "fmla v20.8h, v29.8h, v3.h[3]\n"
+ "fmla v24.8h, v29.8h, v4.h[3]\n"
+ "ldr d29, [x17, #0xe0]\n"
+ "fmla v9.8h, v28.8h, v0.h[3]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[3]\n"
+ "ldr x21, [x17, #0x108]\n"
+ "fmla v17.8h, v28.8h, v2.h[3]\n"
+ "fmla v21.8h, v28.8h, v3.h[3]\n"
+ "fmla v25.8h, v28.8h, v4.h[3]\n"
+ "ldr d28, [x17, #0xf0]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[3]\n"
+ "fmla v14.8h, v29.8h, v1.h[3]\n"
+ "ldr x20, [x17, #0x118]\n"
+ "fmla v18.8h, v29.8h, v2.h[3]\n"
+ "fmla v22.8h, v29.8h, v3.h[3]\n"
+ "fmla v26.8h, v29.8h, v4.h[3]\n"
+ "ldr d29, [x17, #0x100]\n"
+ "fmla v11.8h, v28.8h, v0.h[3]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[3]\n"
+ "ldr x21, [x17, #0x128]\n"
+ "fmla v19.8h, v28.8h, v2.h[3]\n"
+ "fmla v23.8h, v28.8h, v3.h[3]\n"
+ "fmla v27.8h, v28.8h, v4.h[3]\n"
+ "ldr d28, [x17, #0x110]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[4]\n"
+ "fmla v12.8h, v29.8h, v1.h[4]\n"
+ "ldr x20, [x17, #0x138]\n"
+ "fmla v16.8h, v29.8h, v2.h[4]\n"
+ "fmla v20.8h, v29.8h, v3.h[4]\n"
+ "fmla v24.8h, v29.8h, v4.h[4]\n"
+ "ldr d29, [x17, #0x120]\n"
+ "fmla v9.8h, v28.8h, v0.h[4]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[4]\n"
+ "ldr x21, [x17, #0x148]\n"
+ "fmla v17.8h, v28.8h, v2.h[4]\n"
+ "fmla v21.8h, v28.8h, v3.h[4]\n"
+ "fmla v25.8h, v28.8h, v4.h[4]\n"
+ "ldr d28, [x17, #0x130]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[4]\n"
+ "fmla v14.8h, v29.8h, v1.h[4]\n"
+ "ldr x20, [x17, #0x158]\n"
+ "fmla v18.8h, v29.8h, v2.h[4]\n"
+ "fmla v22.8h, v29.8h, v3.h[4]\n"
+ "fmla v26.8h, v29.8h, v4.h[4]\n"
+ "ldr d29, [x17, #0x140]\n"
+ "fmla v11.8h, v28.8h, v0.h[4]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[4]\n"
+ "ldr x21, [x17, #0x168]\n"
+ "fmla v19.8h, v28.8h, v2.h[4]\n"
+ "fmla v23.8h, v28.8h, v3.h[4]\n"
+ "fmla v27.8h, v28.8h, v4.h[4]\n"
+ "ldr d28, [x17, #0x150]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[5]\n"
+ "fmla v12.8h, v29.8h, v1.h[5]\n"
+ "ldr x20, [x17, #0x178]\n"
+ "fmla v16.8h, v29.8h, v2.h[5]\n"
+ "fmla v20.8h, v29.8h, v3.h[5]\n"
+ "fmla v24.8h, v29.8h, v4.h[5]\n"
+ "ldr d29, [x17, #0x160]\n"
+ "fmla v9.8h, v28.8h, v0.h[5]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[5]\n"
+ "ldr x21, [x17, #0x188]\n"
+ "fmla v17.8h, v28.8h, v2.h[5]\n"
+ "fmla v21.8h, v28.8h, v3.h[5]\n"
+ "fmla v25.8h, v28.8h, v4.h[5]\n"
+ "ldr d28, [x17, #0x170]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[5]\n"
+ "fmla v14.8h, v29.8h, v1.h[5]\n"
+ "ldr x20, [x17, #0x198]\n"
+ "fmla v18.8h, v29.8h, v2.h[5]\n"
+ "fmla v22.8h, v29.8h, v3.h[5]\n"
+ "fmla v26.8h, v29.8h, v4.h[5]\n"
+ "ldr d29, [x17, #0x180]\n"
+ "fmla v11.8h, v28.8h, v0.h[5]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[5]\n"
+ "ldr x21, [x17, #0x1a8]\n"
+ "fmla v19.8h, v28.8h, v2.h[5]\n"
+ "fmla v23.8h, v28.8h, v3.h[5]\n"
+ "fmla v27.8h, v28.8h, v4.h[5]\n"
+ "ldr d28, [x17, #0x190]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[6]\n"
+ "fmla v12.8h, v29.8h, v1.h[6]\n"
+ "ldr x20, [x17, #0x1b8]\n"
+ "fmla v16.8h, v29.8h, v2.h[6]\n"
+ "fmla v20.8h, v29.8h, v3.h[6]\n"
+ "fmla v24.8h, v29.8h, v4.h[6]\n"
+ "ldr d29, [x17, #0x1a0]\n"
+ "fmla v9.8h, v28.8h, v0.h[6]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[6]\n"
+ "ldr x21, [x17, #0x1c8]\n"
+ "fmla v17.8h, v28.8h, v2.h[6]\n"
+ "fmla v21.8h, v28.8h, v3.h[6]\n"
+ "fmla v25.8h, v28.8h, v4.h[6]\n"
+ "ldr d28, [x17, #0x1b0]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v10.8h, v29.8h, v0.h[6]\n"
+ "fmla v14.8h, v29.8h, v1.h[6]\n"
+ "ldr x20, [x17, #0x1d8]\n"
+ "fmla v18.8h, v29.8h, v2.h[6]\n"
+ "fmla v22.8h, v29.8h, v3.h[6]\n"
+ "fmla v26.8h, v29.8h, v4.h[6]\n"
+ "ldr d29, [x17, #0x1c0]\n"
+ "fmla v11.8h, v28.8h, v0.h[6]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v15.8h, v28.8h, v1.h[6]\n"
+ "ldr x21, [x17, #0x1e8]\n"
+ "fmla v19.8h, v28.8h, v2.h[6]\n"
+ "fmla v23.8h, v28.8h, v3.h[6]\n"
+ "fmla v27.8h, v28.8h, v4.h[6]\n"
+ "ldr d28, [x17, #0x1d0]\n"
+ "mov v28.d[1], x20\n"
+ "fmla v8.8h, v29.8h, v0.h[7]\n"
+ "fmla v12.8h, v29.8h, v1.h[7]\n"
+ "ldr x20, [x17, #0x1f8]\n"
+ "fmla v16.8h, v29.8h, v2.h[7]\n"
+ "fmla v20.8h, v29.8h, v3.h[7]\n"
+ "fmla v24.8h, v29.8h, v4.h[7]\n"
+ "ldr d29, [x17, #0x1e0]\n"
+ "fmla v9.8h, v28.8h, v0.h[7]\n"
+ "mov v29.d[1], x21\n"
+ "fmla v13.8h, v28.8h, v1.h[7]\n"
+ "fmla v17.8h, v28.8h, v2.h[7]\n"
+ "fmla v21.8h, v28.8h, v3.h[7]\n"
+ "fmla v25.8h, v28.8h, v4.h[7]\n"
+ "ldr d28, [x17, #0x1f0]\n"
+ "mov v28.d[1], x20\n"
"add x17, x17, #0x200\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "ldr x12, [x17, #0x8]\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x18]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
- "fmla v22.8h, v6.8h, v3.h[7]\n"
- "fmla v26.8h, v6.8h, v4.h[7]\n"
+ "fmla v10.8h, v29.8h, v0.h[7]\n"
+ "ldr x21, [x17, #0x8]\n"
+ "fmla v14.8h, v29.8h, v1.h[7]\n"
+ "ldr x20, [x17, #0x18]\n"
+ "fmla v18.8h, v29.8h, v2.h[7]\n"
+ "fmla v22.8h, v29.8h, v3.h[7]\n"
+ "fmla v26.8h, v29.8h, v4.h[7]\n"
"ldr d6, [x17, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v28.8h, v0.h[7]\n"
"ldr d0, [x13, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "ldr d1, [x9, #0x0]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
- "ldr d2, [x27, #0x0]\n"
- "fmla v23.8h, v7.8h, v3.h[7]\n"
- "ldr d3, [x25, #0x0]\n"
- "fmla v27.8h, v7.8h, v4.h[7]\n"
- "ldr d4, [x23, #0x0]\n"
+ "fmla v15.8h, v28.8h, v1.h[7]\n"
+ "ldr d1, [x12, #0x0]\n"
+ "fmla v19.8h, v28.8h, v2.h[7]\n"
+ "ldr d2, [x11, #0x0]\n"
+ "fmla v23.8h, v28.8h, v3.h[7]\n"
+ "ldr d3, [x10, #0x0]\n"
+ "fmla v27.8h, v28.8h, v4.h[7]\n"
+ "ldr d4, [x9, #0x0]\n"
"ldr d7, [x17, #0x10]\n"
- "mov v6.d[1], x12\n"
- "mov v0.d[1], x10\n"
- "mov v1.d[1], x28\n"
- "mov v2.d[1], x26\n"
- "mov v3.d[1], x24\n"
+ "mov v6.d[1], x21\n"
+ "mov v0.d[1], x26\n"
+ "mov v1.d[1], x25\n"
+ "mov v2.d[1], x24\n"
+ "mov v3.d[1], x23\n"
"mov v4.d[1], x22\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"bge 222b\n"
"223:" // Height 5: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "add x27, x27, #0x10\n"
+ "add x11, x11, #0x10\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "add x25, x25, #0x10\n"
+ "add x10, x10, #0x10\n"
"fmla v24.8h, v6.8h, v4.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
+ "ldr q29, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "add x23, x23, #0x10\n"
+ "add x9, x9, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
"sub x14, x14, #0x8\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
"prfm pldl1keep, [x13, #0x80]\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
"fmla v25.8h, v7.8h, v4.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "prfm pldl1keep, [x25, #0x80]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "prfm pldl1keep, [x23, #0x80]\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "fmla v26.8h, v6.8h, v4.h[0]\n"
- "ldr q6, [x17, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
- "fmla v27.8h, v7.8h, v4.h[0]\n"
- "ldr q7, [x17, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "fmla v16.8h, v6.8h, v2.h[1]\n"
- "fmla v20.8h, v6.8h, v3.h[1]\n"
- "fmla v24.8h, v6.8h, v4.h[1]\n"
- "ldr q6, [x17, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "fmla v17.8h, v7.8h, v2.h[1]\n"
- "fmla v21.8h, v7.8h, v3.h[1]\n"
- "fmla v25.8h, v7.8h, v4.h[1]\n"
- "ldr q7, [x17, #0x70]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "fmla v18.8h, v6.8h, v2.h[1]\n"
- "fmla v22.8h, v6.8h, v3.h[1]\n"
- "fmla v26.8h, v6.8h, v4.h[1]\n"
- "ldr q6, [x17, #0x80]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "fmla v19.8h, v7.8h, v2.h[1]\n"
- "fmla v23.8h, v7.8h, v3.h[1]\n"
- "fmla v27.8h, v7.8h, v4.h[1]\n"
- "ldr q7, [x17, #0x90]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "fmla v16.8h, v6.8h, v2.h[2]\n"
- "fmla v20.8h, v6.8h, v3.h[2]\n"
- "fmla v24.8h, v6.8h, v4.h[2]\n"
- "ldr q6, [x17, #0xa0]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "fmla v17.8h, v7.8h, v2.h[2]\n"
- "fmla v21.8h, v7.8h, v3.h[2]\n"
- "fmla v25.8h, v7.8h, v4.h[2]\n"
- "ldr q7, [x17, #0xb0]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "fmla v18.8h, v6.8h, v2.h[2]\n"
- "fmla v22.8h, v6.8h, v3.h[2]\n"
- "fmla v26.8h, v6.8h, v4.h[2]\n"
- "ldr q6, [x17, #0xc0]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "fmla v19.8h, v7.8h, v2.h[2]\n"
- "fmla v23.8h, v7.8h, v3.h[2]\n"
- "fmla v27.8h, v7.8h, v4.h[2]\n"
- "ldr q7, [x17, #0xd0]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "fmla v16.8h, v6.8h, v2.h[3]\n"
- "fmla v20.8h, v6.8h, v3.h[3]\n"
- "fmla v24.8h, v6.8h, v4.h[3]\n"
- "ldr q6, [x17, #0xe0]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "fmla v17.8h, v7.8h, v2.h[3]\n"
- "fmla v21.8h, v7.8h, v3.h[3]\n"
- "fmla v25.8h, v7.8h, v4.h[3]\n"
- "ldr q7, [x17, #0xf0]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "fmla v18.8h, v6.8h, v2.h[3]\n"
- "fmla v22.8h, v6.8h, v3.h[3]\n"
- "fmla v26.8h, v6.8h, v4.h[3]\n"
- "ldr q6, [x17, #0x100]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "fmla v19.8h, v7.8h, v2.h[3]\n"
- "fmla v23.8h, v7.8h, v3.h[3]\n"
- "fmla v27.8h, v7.8h, v4.h[3]\n"
- "ldr q7, [x17, #0x110]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "fmla v16.8h, v6.8h, v2.h[4]\n"
- "fmla v20.8h, v6.8h, v3.h[4]\n"
- "fmla v24.8h, v6.8h, v4.h[4]\n"
- "ldr q6, [x17, #0x120]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "fmla v17.8h, v7.8h, v2.h[4]\n"
- "fmla v21.8h, v7.8h, v3.h[4]\n"
- "fmla v25.8h, v7.8h, v4.h[4]\n"
- "ldr q7, [x17, #0x130]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "fmla v18.8h, v6.8h, v2.h[4]\n"
- "fmla v22.8h, v6.8h, v3.h[4]\n"
- "fmla v26.8h, v6.8h, v4.h[4]\n"
- "ldr q6, [x17, #0x140]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "fmla v19.8h, v7.8h, v2.h[4]\n"
- "fmla v23.8h, v7.8h, v3.h[4]\n"
- "fmla v27.8h, v7.8h, v4.h[4]\n"
- "ldr q7, [x17, #0x150]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "fmla v16.8h, v6.8h, v2.h[5]\n"
- "fmla v20.8h, v6.8h, v3.h[5]\n"
- "fmla v24.8h, v6.8h, v4.h[5]\n"
- "ldr q6, [x17, #0x160]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "fmla v17.8h, v7.8h, v2.h[5]\n"
- "fmla v21.8h, v7.8h, v3.h[5]\n"
- "fmla v25.8h, v7.8h, v4.h[5]\n"
- "ldr q7, [x17, #0x170]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "fmla v18.8h, v6.8h, v2.h[5]\n"
- "fmla v22.8h, v6.8h, v3.h[5]\n"
- "fmla v26.8h, v6.8h, v4.h[5]\n"
- "ldr q6, [x17, #0x180]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "fmla v19.8h, v7.8h, v2.h[5]\n"
- "fmla v23.8h, v7.8h, v3.h[5]\n"
- "fmla v27.8h, v7.8h, v4.h[5]\n"
- "ldr q7, [x17, #0x190]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "fmla v16.8h, v6.8h, v2.h[6]\n"
- "fmla v20.8h, v6.8h, v3.h[6]\n"
- "fmla v24.8h, v6.8h, v4.h[6]\n"
- "ldr q6, [x17, #0x1a0]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "fmla v17.8h, v7.8h, v2.h[6]\n"
- "fmla v21.8h, v7.8h, v3.h[6]\n"
- "fmla v25.8h, v7.8h, v4.h[6]\n"
- "ldr q7, [x17, #0x1b0]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "fmla v18.8h, v6.8h, v2.h[6]\n"
- "fmla v22.8h, v6.8h, v3.h[6]\n"
- "fmla v26.8h, v6.8h, v4.h[6]\n"
- "ldr q6, [x17, #0x1c0]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "fmla v23.8h, v7.8h, v3.h[6]\n"
- "fmla v27.8h, v7.8h, v4.h[6]\n"
- "ldr q7, [x17, #0x1d0]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "fmla v16.8h, v6.8h, v2.h[7]\n"
- "fmla v20.8h, v6.8h, v3.h[7]\n"
- "fmla v24.8h, v6.8h, v4.h[7]\n"
- "ldr q6, [x17, #0x1e0]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "fmla v21.8h, v7.8h, v3.h[7]\n"
- "fmla v25.8h, v7.8h, v4.h[7]\n"
- "ldr q7, [x17, #0x1f0]\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
+ "ldr q28, [x17, #0x30]\n"
+ "fmla v10.8h, v29.8h, v0.h[0]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
+ "fmla v14.8h, v29.8h, v1.h[0]\n"
+ "prfm pldl1keep, [x10, #0x80]\n"
+ "fmla v18.8h, v29.8h, v2.h[0]\n"
+ "prfm pldl1keep, [x9, #0x80]\n"
+ "fmla v22.8h, v29.8h, v3.h[0]\n"
+ "fmla v26.8h, v29.8h, v4.h[0]\n"
+ "ldr q29, [x17, #0x40]\n"
+ "fmla v11.8h, v28.8h, v0.h[0]\n"
+ "fmla v15.8h, v28.8h, v1.h[0]\n"
+ "fmla v19.8h, v28.8h, v2.h[0]\n"
+ "fmla v23.8h, v28.8h, v3.h[0]\n"
+ "fmla v27.8h, v28.8h, v4.h[0]\n"
+ "ldr q28, [x17, #0x50]\n"
+ "fmla v8.8h, v29.8h, v0.h[1]\n"
+ "fmla v12.8h, v29.8h, v1.h[1]\n"
+ "fmla v16.8h, v29.8h, v2.h[1]\n"
+ "fmla v20.8h, v29.8h, v3.h[1]\n"
+ "fmla v24.8h, v29.8h, v4.h[1]\n"
+ "ldr q29, [x17, #0x60]\n"
+ "fmla v9.8h, v28.8h, v0.h[1]\n"
+ "fmla v13.8h, v28.8h, v1.h[1]\n"
+ "fmla v17.8h, v28.8h, v2.h[1]\n"
+ "fmla v21.8h, v28.8h, v3.h[1]\n"
+ "fmla v25.8h, v28.8h, v4.h[1]\n"
+ "ldr q28, [x17, #0x70]\n"
+ "fmla v10.8h, v29.8h, v0.h[1]\n"
+ "fmla v14.8h, v29.8h, v1.h[1]\n"
+ "fmla v18.8h, v29.8h, v2.h[1]\n"
+ "fmla v22.8h, v29.8h, v3.h[1]\n"
+ "fmla v26.8h, v29.8h, v4.h[1]\n"
+ "ldr q29, [x17, #0x80]\n"
+ "fmla v11.8h, v28.8h, v0.h[1]\n"
+ "fmla v15.8h, v28.8h, v1.h[1]\n"
+ "fmla v19.8h, v28.8h, v2.h[1]\n"
+ "fmla v23.8h, v28.8h, v3.h[1]\n"
+ "fmla v27.8h, v28.8h, v4.h[1]\n"
+ "ldr q28, [x17, #0x90]\n"
+ "fmla v8.8h, v29.8h, v0.h[2]\n"
+ "fmla v12.8h, v29.8h, v1.h[2]\n"
+ "fmla v16.8h, v29.8h, v2.h[2]\n"
+ "fmla v20.8h, v29.8h, v3.h[2]\n"
+ "fmla v24.8h, v29.8h, v4.h[2]\n"
+ "ldr q29, [x17, #0xa0]\n"
+ "fmla v9.8h, v28.8h, v0.h[2]\n"
+ "fmla v13.8h, v28.8h, v1.h[2]\n"
+ "fmla v17.8h, v28.8h, v2.h[2]\n"
+ "fmla v21.8h, v28.8h, v3.h[2]\n"
+ "fmla v25.8h, v28.8h, v4.h[2]\n"
+ "ldr q28, [x17, #0xb0]\n"
+ "fmla v10.8h, v29.8h, v0.h[2]\n"
+ "fmla v14.8h, v29.8h, v1.h[2]\n"
+ "fmla v18.8h, v29.8h, v2.h[2]\n"
+ "fmla v22.8h, v29.8h, v3.h[2]\n"
+ "fmla v26.8h, v29.8h, v4.h[2]\n"
+ "ldr q29, [x17, #0xc0]\n"
+ "fmla v11.8h, v28.8h, v0.h[2]\n"
+ "fmla v15.8h, v28.8h, v1.h[2]\n"
+ "fmla v19.8h, v28.8h, v2.h[2]\n"
+ "fmla v23.8h, v28.8h, v3.h[2]\n"
+ "fmla v27.8h, v28.8h, v4.h[2]\n"
+ "ldr q28, [x17, #0xd0]\n"
+ "fmla v8.8h, v29.8h, v0.h[3]\n"
+ "fmla v12.8h, v29.8h, v1.h[3]\n"
+ "fmla v16.8h, v29.8h, v2.h[3]\n"
+ "fmla v20.8h, v29.8h, v3.h[3]\n"
+ "fmla v24.8h, v29.8h, v4.h[3]\n"
+ "ldr q29, [x17, #0xe0]\n"
+ "fmla v9.8h, v28.8h, v0.h[3]\n"
+ "fmla v13.8h, v28.8h, v1.h[3]\n"
+ "fmla v17.8h, v28.8h, v2.h[3]\n"
+ "fmla v21.8h, v28.8h, v3.h[3]\n"
+ "fmla v25.8h, v28.8h, v4.h[3]\n"
+ "ldr q28, [x17, #0xf0]\n"
+ "fmla v10.8h, v29.8h, v0.h[3]\n"
+ "fmla v14.8h, v29.8h, v1.h[3]\n"
+ "fmla v18.8h, v29.8h, v2.h[3]\n"
+ "fmla v22.8h, v29.8h, v3.h[3]\n"
+ "fmla v26.8h, v29.8h, v4.h[3]\n"
+ "ldr q29, [x17, #0x100]\n"
+ "fmla v11.8h, v28.8h, v0.h[3]\n"
+ "fmla v15.8h, v28.8h, v1.h[3]\n"
+ "fmla v19.8h, v28.8h, v2.h[3]\n"
+ "fmla v23.8h, v28.8h, v3.h[3]\n"
+ "fmla v27.8h, v28.8h, v4.h[3]\n"
+ "ldr q28, [x17, #0x110]\n"
+ "fmla v8.8h, v29.8h, v0.h[4]\n"
+ "fmla v12.8h, v29.8h, v1.h[4]\n"
+ "fmla v16.8h, v29.8h, v2.h[4]\n"
+ "fmla v20.8h, v29.8h, v3.h[4]\n"
+ "fmla v24.8h, v29.8h, v4.h[4]\n"
+ "ldr q29, [x17, #0x120]\n"
+ "fmla v9.8h, v28.8h, v0.h[4]\n"
+ "fmla v13.8h, v28.8h, v1.h[4]\n"
+ "fmla v17.8h, v28.8h, v2.h[4]\n"
+ "fmla v21.8h, v28.8h, v3.h[4]\n"
+ "fmla v25.8h, v28.8h, v4.h[4]\n"
+ "ldr q28, [x17, #0x130]\n"
+ "fmla v10.8h, v29.8h, v0.h[4]\n"
+ "fmla v14.8h, v29.8h, v1.h[4]\n"
+ "fmla v18.8h, v29.8h, v2.h[4]\n"
+ "fmla v22.8h, v29.8h, v3.h[4]\n"
+ "fmla v26.8h, v29.8h, v4.h[4]\n"
+ "ldr q29, [x17, #0x140]\n"
+ "fmla v11.8h, v28.8h, v0.h[4]\n"
+ "fmla v15.8h, v28.8h, v1.h[4]\n"
+ "fmla v19.8h, v28.8h, v2.h[4]\n"
+ "fmla v23.8h, v28.8h, v3.h[4]\n"
+ "fmla v27.8h, v28.8h, v4.h[4]\n"
+ "ldr q28, [x17, #0x150]\n"
+ "fmla v8.8h, v29.8h, v0.h[5]\n"
+ "fmla v12.8h, v29.8h, v1.h[5]\n"
+ "fmla v16.8h, v29.8h, v2.h[5]\n"
+ "fmla v20.8h, v29.8h, v3.h[5]\n"
+ "fmla v24.8h, v29.8h, v4.h[5]\n"
+ "ldr q29, [x17, #0x160]\n"
+ "fmla v9.8h, v28.8h, v0.h[5]\n"
+ "fmla v13.8h, v28.8h, v1.h[5]\n"
+ "fmla v17.8h, v28.8h, v2.h[5]\n"
+ "fmla v21.8h, v28.8h, v3.h[5]\n"
+ "fmla v25.8h, v28.8h, v4.h[5]\n"
+ "ldr q28, [x17, #0x170]\n"
+ "fmla v10.8h, v29.8h, v0.h[5]\n"
+ "fmla v14.8h, v29.8h, v1.h[5]\n"
+ "fmla v18.8h, v29.8h, v2.h[5]\n"
+ "fmla v22.8h, v29.8h, v3.h[5]\n"
+ "fmla v26.8h, v29.8h, v4.h[5]\n"
+ "ldr q29, [x17, #0x180]\n"
+ "fmla v11.8h, v28.8h, v0.h[5]\n"
+ "fmla v15.8h, v28.8h, v1.h[5]\n"
+ "fmla v19.8h, v28.8h, v2.h[5]\n"
+ "fmla v23.8h, v28.8h, v3.h[5]\n"
+ "fmla v27.8h, v28.8h, v4.h[5]\n"
+ "ldr q28, [x17, #0x190]\n"
+ "fmla v8.8h, v29.8h, v0.h[6]\n"
+ "fmla v12.8h, v29.8h, v1.h[6]\n"
+ "fmla v16.8h, v29.8h, v2.h[6]\n"
+ "fmla v20.8h, v29.8h, v3.h[6]\n"
+ "fmla v24.8h, v29.8h, v4.h[6]\n"
+ "ldr q29, [x17, #0x1a0]\n"
+ "fmla v9.8h, v28.8h, v0.h[6]\n"
+ "fmla v13.8h, v28.8h, v1.h[6]\n"
+ "fmla v17.8h, v28.8h, v2.h[6]\n"
+ "fmla v21.8h, v28.8h, v3.h[6]\n"
+ "fmla v25.8h, v28.8h, v4.h[6]\n"
+ "ldr q28, [x17, #0x1b0]\n"
+ "fmla v10.8h, v29.8h, v0.h[6]\n"
+ "fmla v14.8h, v29.8h, v1.h[6]\n"
+ "fmla v18.8h, v29.8h, v2.h[6]\n"
+ "fmla v22.8h, v29.8h, v3.h[6]\n"
+ "fmla v26.8h, v29.8h, v4.h[6]\n"
+ "ldr q29, [x17, #0x1c0]\n"
+ "fmla v11.8h, v28.8h, v0.h[6]\n"
+ "fmla v15.8h, v28.8h, v1.h[6]\n"
+ "fmla v19.8h, v28.8h, v2.h[6]\n"
+ "fmla v23.8h, v28.8h, v3.h[6]\n"
+ "fmla v27.8h, v28.8h, v4.h[6]\n"
+ "ldr q28, [x17, #0x1d0]\n"
+ "fmla v8.8h, v29.8h, v0.h[7]\n"
+ "fmla v12.8h, v29.8h, v1.h[7]\n"
+ "fmla v16.8h, v29.8h, v2.h[7]\n"
+ "fmla v20.8h, v29.8h, v3.h[7]\n"
+ "fmla v24.8h, v29.8h, v4.h[7]\n"
+ "ldr q29, [x17, #0x1e0]\n"
+ "fmla v9.8h, v28.8h, v0.h[7]\n"
+ "fmla v13.8h, v28.8h, v1.h[7]\n"
+ "fmla v17.8h, v28.8h, v2.h[7]\n"
+ "fmla v21.8h, v28.8h, v3.h[7]\n"
+ "fmla v25.8h, v28.8h, v4.h[7]\n"
+ "ldr q28, [x17, #0x1f0]\n"
+ "fmla v10.8h, v29.8h, v0.h[7]\n"
"add x17, x17, #0x200\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
- "fmla v22.8h, v6.8h, v3.h[7]\n"
- "fmla v26.8h, v6.8h, v4.h[7]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
- "fmla v23.8h, v7.8h, v3.h[7]\n"
- "fmla v27.8h, v7.8h, v4.h[7]\n"
+ "fmla v14.8h, v29.8h, v1.h[7]\n"
+ "fmla v18.8h, v29.8h, v2.h[7]\n"
+ "fmla v22.8h, v29.8h, v3.h[7]\n"
+ "fmla v26.8h, v29.8h, v4.h[7]\n"
+ "fmla v11.8h, v28.8h, v0.h[7]\n"
+ "fmla v15.8h, v28.8h, v1.h[7]\n"
+ "fmla v19.8h, v28.8h, v2.h[7]\n"
+ "fmla v23.8h, v28.8h, v3.h[7]\n"
+ "fmla v27.8h, v28.8h, v4.h[7]\n"
"224:" // Height 5: Multiply loop: Main loop skip
"cbz x14, 226f\n"
"225:" // Height 5: Multiply loop: Odd block loop
- "ldr h0, [x13], #0x2\n"
+ "ldr h4, [x13], #0x2\n"
"sub x14, x14, #0x1\n"
- "ldr h1, [x9], #0x2\n"
- "ldr h2, [x27], #0x2\n"
- "ldr h3, [x25], #0x2\n"
- "ldr h4, [x23], #0x2\n"
- "ldr q6, [x17, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x10]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "fmla v16.8h, v6.8h, v2.h[0]\n"
- "fmla v20.8h, v6.8h, v3.h[0]\n"
- "fmla v24.8h, v6.8h, v4.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "fmla v17.8h, v7.8h, v2.h[0]\n"
- "fmla v21.8h, v7.8h, v3.h[0]\n"
- "fmla v25.8h, v7.8h, v4.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr h3, [x12], #0x2\n"
+ "ldr h2, [x11], #0x2\n"
+ "ldr h1, [x10], #0x2\n"
+ "ldr h0, [x9], #0x2\n"
+ "ldr q29, [x17, #0x0]\n"
+ "fmla v8.8h, v29.8h, v4.h[0]\n"
+ "ldr q28, [x17, #0x10]\n"
+ "fmla v12.8h, v29.8h, v3.h[0]\n"
+ "fmla v16.8h, v29.8h, v2.h[0]\n"
+ "fmla v20.8h, v29.8h, v1.h[0]\n"
+ "fmla v24.8h, v29.8h, v0.h[0]\n"
+ "ldr q29, [x17, #0x20]\n"
+ "fmla v9.8h, v28.8h, v4.h[0]\n"
+ "fmla v13.8h, v28.8h, v3.h[0]\n"
+ "fmla v17.8h, v28.8h, v2.h[0]\n"
+ "fmla v21.8h, v28.8h, v1.h[0]\n"
+ "fmla v25.8h, v28.8h, v0.h[0]\n"
+ "ldr q28, [x17, #0x30]\n"
+ "fmla v10.8h, v29.8h, v4.h[0]\n"
"add x17, x17, #0x40\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "fmla v26.8h, v6.8h, v4.h[0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
- "fmla v27.8h, v7.8h, v4.h[0]\n"
+ "fmla v14.8h, v29.8h, v3.h[0]\n"
+ "fmla v18.8h, v29.8h, v2.h[0]\n"
+ "fmla v22.8h, v29.8h, v1.h[0]\n"
+ "fmla v26.8h, v29.8h, v0.h[0]\n"
+ "fmla v11.8h, v28.8h, v4.h[0]\n"
+ "fmla v15.8h, v28.8h, v3.h[0]\n"
+ "fmla v19.8h, v28.8h, v2.h[0]\n"
+ "fmla v23.8h, v28.8h, v1.h[0]\n"
+ "fmla v27.8h, v28.8h, v0.h[0]\n"
"cbnz x14, 225b\n"
"226:" // Height 5: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -4069,49 +4069,49 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"prfm pstl1keep, [x22, #0x0]\n"
"tbz %x[flags], #1, 227f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v0.8h\n"
- "fmin v9.8h, v9.8h, v0.8h\n"
- "fmin v10.8h, v10.8h, v0.8h\n"
- "fmin v11.8h, v11.8h, v0.8h\n"
- "fmin v12.8h, v12.8h, v0.8h\n"
- "fmin v13.8h, v13.8h, v0.8h\n"
- "fmin v14.8h, v14.8h, v0.8h\n"
- "fmin v15.8h, v15.8h, v0.8h\n"
- "fmin v16.8h, v16.8h, v0.8h\n"
- "fmin v17.8h, v17.8h, v0.8h\n"
- "fmin v18.8h, v18.8h, v0.8h\n"
- "fmin v19.8h, v19.8h, v0.8h\n"
- "fmin v20.8h, v20.8h, v0.8h\n"
- "fmin v21.8h, v21.8h, v0.8h\n"
- "fmin v22.8h, v22.8h, v0.8h\n"
- "fmin v23.8h, v23.8h, v0.8h\n"
- "fmin v24.8h, v24.8h, v0.8h\n"
- "fmin v25.8h, v25.8h, v0.8h\n"
- "fmin v26.8h, v26.8h, v0.8h\n"
- "fmin v27.8h, v27.8h, v0.8h\n"
+ "ld1r { v28.8h }, [x20]\n"
+ "fmin v8.8h, v8.8h, v28.8h\n"
+ "fmin v9.8h, v9.8h, v28.8h\n"
+ "fmin v10.8h, v10.8h, v28.8h\n"
+ "fmin v11.8h, v11.8h, v28.8h\n"
+ "fmin v12.8h, v12.8h, v28.8h\n"
+ "fmin v13.8h, v13.8h, v28.8h\n"
+ "fmin v14.8h, v14.8h, v28.8h\n"
+ "fmin v15.8h, v15.8h, v28.8h\n"
+ "fmin v16.8h, v16.8h, v28.8h\n"
+ "fmin v17.8h, v17.8h, v28.8h\n"
+ "fmin v18.8h, v18.8h, v28.8h\n"
+ "fmin v19.8h, v19.8h, v28.8h\n"
+ "fmin v20.8h, v20.8h, v28.8h\n"
+ "fmin v21.8h, v21.8h, v28.8h\n"
+ "fmin v22.8h, v22.8h, v28.8h\n"
+ "fmin v23.8h, v23.8h, v28.8h\n"
+ "fmin v24.8h, v24.8h, v28.8h\n"
+ "fmin v25.8h, v25.8h, v28.8h\n"
+ "fmin v26.8h, v26.8h, v28.8h\n"
+ "fmin v27.8h, v27.8h, v28.8h\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmax v8.8h, v8.8h, v0.8h\n"
- "fmax v9.8h, v9.8h, v0.8h\n"
- "fmax v10.8h, v10.8h, v0.8h\n"
- "fmax v11.8h, v11.8h, v0.8h\n"
- "fmax v12.8h, v12.8h, v0.8h\n"
- "fmax v13.8h, v13.8h, v0.8h\n"
- "fmax v14.8h, v14.8h, v0.8h\n"
- "fmax v15.8h, v15.8h, v0.8h\n"
- "fmax v16.8h, v16.8h, v0.8h\n"
- "fmax v17.8h, v17.8h, v0.8h\n"
- "fmax v18.8h, v18.8h, v0.8h\n"
- "fmax v19.8h, v19.8h, v0.8h\n"
- "fmax v20.8h, v20.8h, v0.8h\n"
- "fmax v21.8h, v21.8h, v0.8h\n"
- "fmax v22.8h, v22.8h, v0.8h\n"
- "fmax v23.8h, v23.8h, v0.8h\n"
- "fmax v24.8h, v24.8h, v0.8h\n"
- "fmax v25.8h, v25.8h, v0.8h\n"
- "fmax v26.8h, v26.8h, v0.8h\n"
- "fmax v27.8h, v27.8h, v0.8h\n"
+ "ld1r { v28.8h }, [x20]\n"
+ "fmax v8.8h, v8.8h, v28.8h\n"
+ "fmax v9.8h, v9.8h, v28.8h\n"
+ "fmax v10.8h, v10.8h, v28.8h\n"
+ "fmax v11.8h, v11.8h, v28.8h\n"
+ "fmax v12.8h, v12.8h, v28.8h\n"
+ "fmax v13.8h, v13.8h, v28.8h\n"
+ "fmax v14.8h, v14.8h, v28.8h\n"
+ "fmax v15.8h, v15.8h, v28.8h\n"
+ "fmax v16.8h, v16.8h, v28.8h\n"
+ "fmax v17.8h, v17.8h, v28.8h\n"
+ "fmax v18.8h, v18.8h, v28.8h\n"
+ "fmax v19.8h, v19.8h, v28.8h\n"
+ "fmax v20.8h, v20.8h, v28.8h\n"
+ "fmax v21.8h, v21.8h, v28.8h\n"
+ "fmax v22.8h, v22.8h, v28.8h\n"
+ "fmax v23.8h, v23.8h, v28.8h\n"
+ "fmax v24.8h, v24.8h, v28.8h\n"
+ "fmax v25.8h, v25.8h, v28.8h\n"
+ "fmax v26.8h, v26.8h, v28.8h\n"
+ "fmax v27.8h, v27.8h, v28.8h\n"
"227:" // Height 5: No activation
"cmp x8, #0x20\n"
"bge 244f\n"
@@ -4736,98 +4736,98 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"268:" // Height 6: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w14, [x20, x15, 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, 269f\n"
- "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x13, [x21, #0x0]\n"
- "ldr x9, [x21, #0x8]\n"
- "ldr x27, [x21, #0x10]\n"
- "ldr x25, [x21, #0x18]\n"
- "ldr x23, [x21, #0x20]\n"
- "ldr x21, [x21, #0x28]\n"
+ "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x13, [x20, #0x0]\n"
+ "ldr x12, [x20, #0x8]\n"
+ "ldr x11, [x20, #0x10]\n"
+ "ldr x10, [x20, #0x18]\n"
+ "ldr x9, [x20, #0x20]\n"
+ "ldr x28, [x20, #0x28]\n"
"cbnz x15, 270f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x13, x13, x20, LSL #1\n"
+ "add x12, x12, x20, LSL #1\n"
+ "add x11, x11, x20, LSL #1\n"
+ "add x10, x10, x20, LSL #1\n"
"add x9, x9, x20, LSL #1\n"
- "add x27, x27, x20, LSL #1\n"
- "add x25, x25, x20, LSL #1\n"
- "add x23, x23, x20, LSL #1\n"
- "add x21, x21, x20, LSL #1\n"
+ "add x28, x28, x20, LSL #1\n"
"b 270f\n"
"269:" // Height 6: setup direct input
"mov x13, %x[input_ptr]\n"
- "add x9, x13, x20, LSL #1\n"
- "add x27, x9, x20, LSL #1\n"
- "add x25, x27, x20, LSL #1\n"
- "add x23, x25, x20, LSL #1\n"
- "add x21, x23, x20, LSL #1\n"
+ "add x12, x13, x21, LSL #1\n"
+ "add x11, x12, x21, LSL #1\n"
+ "add x10, x11, x21, LSL #1\n"
+ "add x9, x10, x21, LSL #1\n"
+ "add x28, x9, x21, LSL #1\n"
"270:" // Height 6: input setup done
"cmp x14, #0x8\n"
"blt 273f\n"
"ldr q0, [x13, #0x0]\n"
"cmp x14, #0x10\n"
- "ldr q1, [x9, #0x0]\n"
- "ldr q2, [x27, #0x0]\n"
- "ldr q3, [x25, #0x0]\n"
- "ldr q4, [x23, #0x0]\n"
- "ldr q5, [x21, #0x0]\n"
+ "ldr q1, [x12, #0x0]\n"
+ "ldr q2, [x11, #0x0]\n"
+ "ldr q3, [x10, #0x0]\n"
+ "ldr q4, [x9, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
"ldr q6, [x17, #0x0]\n"
"ldr q7, [x17, #0x10]\n"
"blt 272f\n"
"271:" // Height 6: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr x12, [x17, #0x28]\n"
+ "ldr x21, [x17, #0x28]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x38]\n"
+ "ldr x20, [x17, #0x38]\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v24.8h, v6.8h, v4.h[0]\n"
- "add x27, x27, #0x10\n"
+ "add x11, x11, #0x10\n"
"fmla v28.8h, v6.8h, v5.h[0]\n"
"ldr d6, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x48]\n"
+ "ldr x21, [x17, #0x48]\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "add x25, x25, #0x10\n"
+ "add x10, x10, #0x10\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "add x23, x23, #0x10\n"
+ "add x9, x9, #0x10\n"
"fmla v25.8h, v7.8h, v4.h[0]\n"
- "add x21, x21, #0x10\n"
+ "add x28, x28, #0x10\n"
"fmla v29.8h, v7.8h, v5.h[0]\n"
"ldr d7, [x17, #0x30]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[0]\n"
"fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr x11, [x17, #0x58]\n"
+ "ldr x20, [x17, #0x58]\n"
"fmla v18.8h, v6.8h, v2.h[0]\n"
- "ldr x10, [x13, #0x8]\n"
+ "ldr x27, [x13, #0x8]\n"
"fmla v22.8h, v6.8h, v3.h[0]\n"
- "ldr x28, [x9, #0x8]\n"
+ "ldr x26, [x12, #0x8]\n"
"fmla v26.8h, v6.8h, v4.h[0]\n"
- "ldr x26, [x27, #0x8]\n"
+ "ldr x25, [x11, #0x8]\n"
"fmla v30.8h, v6.8h, v5.h[0]\n"
"ldr d6, [x17, #0x40]\n"
"fmla v11.8h, v7.8h, v0.h[0]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr x12, [x17, #0x68]\n"
+ "ldr x21, [x17, #0x68]\n"
"fmla v19.8h, v7.8h, v2.h[0]\n"
- "ldr x24, [x25, #0x8]\n"
+ "ldr x24, [x10, #0x8]\n"
"fmla v23.8h, v7.8h, v3.h[0]\n"
- "ldr x22, [x23, #0x8]\n"
+ "ldr x23, [x9, #0x8]\n"
"fmla v27.8h, v7.8h, v4.h[0]\n"
- "ldr x20, [x21, #0x8]\n"
+ "ldr x22, [x28, #0x8]\n"
"fmla v31.8h, v7.8h, v5.h[0]\n"
"ldr d7, [x17, #0x50]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[1]\n"
"fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x78]\n"
+ "ldr x20, [x17, #0x78]\n"
"fmla v16.8h, v6.8h, v2.h[1]\n"
"sub x14, x14, #0x8\n"
"fmla v20.8h, v6.8h, v3.h[1]\n"
@@ -4837,240 +4837,240 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"fmla v28.8h, v6.8h, v5.h[1]\n"
"ldr d6, [x17, #0x60]\n"
"fmla v9.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0x88]\n"
+ "ldr x21, [x17, #0x88]\n"
"fmla v17.8h, v7.8h, v2.h[1]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
"fmla v21.8h, v7.8h, v3.h[1]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
"fmla v25.8h, v7.8h, v4.h[1]\n"
- "prfm pldl1keep, [x25, #0x80]\n"
+ "prfm pldl1keep, [x10, #0x80]\n"
"fmla v29.8h, v7.8h, v5.h[1]\n"
"ldr d7, [x17, #0x70]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[1]\n"
"fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr x11, [x17, #0x98]\n"
+ "ldr x20, [x17, #0x98]\n"
"fmla v18.8h, v6.8h, v2.h[1]\n"
- "prfm pldl1keep, [x23, #0x80]\n"
+ "prfm pldl1keep, [x9, #0x80]\n"
"fmla v22.8h, v6.8h, v3.h[1]\n"
- "prfm pldl1keep, [x21, #0x80]\n"
+ "prfm pldl1keep, [x28, #0x80]\n"
"fmla v26.8h, v6.8h, v4.h[1]\n"
"fmla v30.8h, v6.8h, v5.h[1]\n"
"ldr d6, [x17, #0x80]\n"
"fmla v11.8h, v7.8h, v0.h[1]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr x12, [x17, #0xa8]\n"
+ "ldr x21, [x17, #0xa8]\n"
"fmla v19.8h, v7.8h, v2.h[1]\n"
"fmla v23.8h, v7.8h, v3.h[1]\n"
"fmla v27.8h, v7.8h, v4.h[1]\n"
"fmla v31.8h, v7.8h, v5.h[1]\n"
"ldr d7, [x17, #0x90]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[2]\n"
"fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xb8]\n"
+ "ldr x20, [x17, #0xb8]\n"
"fmla v16.8h, v6.8h, v2.h[2]\n"
"fmla v20.8h, v6.8h, v3.h[2]\n"
"fmla v24.8h, v6.8h, v4.h[2]\n"
"fmla v28.8h, v6.8h, v5.h[2]\n"
"ldr d6, [x17, #0xa0]\n"
"fmla v9.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xc8]\n"
+ "ldr x21, [x17, #0xc8]\n"
"fmla v17.8h, v7.8h, v2.h[2]\n"
"fmla v21.8h, v7.8h, v3.h[2]\n"
"fmla v25.8h, v7.8h, v4.h[2]\n"
"fmla v29.8h, v7.8h, v5.h[2]\n"
"ldr d7, [x17, #0xb0]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[2]\n"
"fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr x11, [x17, #0xd8]\n"
+ "ldr x20, [x17, #0xd8]\n"
"fmla v18.8h, v6.8h, v2.h[2]\n"
"fmla v22.8h, v6.8h, v3.h[2]\n"
"fmla v26.8h, v6.8h, v4.h[2]\n"
"fmla v30.8h, v6.8h, v5.h[2]\n"
"ldr d6, [x17, #0xc0]\n"
"fmla v11.8h, v7.8h, v0.h[2]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr x12, [x17, #0xe8]\n"
+ "ldr x21, [x17, #0xe8]\n"
"fmla v19.8h, v7.8h, v2.h[2]\n"
"fmla v23.8h, v7.8h, v3.h[2]\n"
"fmla v27.8h, v7.8h, v4.h[2]\n"
"fmla v31.8h, v7.8h, v5.h[2]\n"
"ldr d7, [x17, #0xd0]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[3]\n"
"fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0xf8]\n"
+ "ldr x20, [x17, #0xf8]\n"
"fmla v16.8h, v6.8h, v2.h[3]\n"
"fmla v20.8h, v6.8h, v3.h[3]\n"
"fmla v24.8h, v6.8h, v4.h[3]\n"
"fmla v28.8h, v6.8h, v5.h[3]\n"
"ldr d6, [x17, #0xe0]\n"
"fmla v9.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x108]\n"
+ "ldr x21, [x17, #0x108]\n"
"fmla v17.8h, v7.8h, v2.h[3]\n"
"fmla v21.8h, v7.8h, v3.h[3]\n"
"fmla v25.8h, v7.8h, v4.h[3]\n"
"fmla v29.8h, v7.8h, v5.h[3]\n"
"ldr d7, [x17, #0xf0]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[3]\n"
"fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr x11, [x17, #0x118]\n"
+ "ldr x20, [x17, #0x118]\n"
"fmla v18.8h, v6.8h, v2.h[3]\n"
"fmla v22.8h, v6.8h, v3.h[3]\n"
"fmla v26.8h, v6.8h, v4.h[3]\n"
"fmla v30.8h, v6.8h, v5.h[3]\n"
"ldr d6, [x17, #0x100]\n"
"fmla v11.8h, v7.8h, v0.h[3]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr x12, [x17, #0x128]\n"
+ "ldr x21, [x17, #0x128]\n"
"fmla v19.8h, v7.8h, v2.h[3]\n"
"fmla v23.8h, v7.8h, v3.h[3]\n"
"fmla v27.8h, v7.8h, v4.h[3]\n"
"fmla v31.8h, v7.8h, v5.h[3]\n"
"ldr d7, [x17, #0x110]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[4]\n"
"fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x138]\n"
+ "ldr x20, [x17, #0x138]\n"
"fmla v16.8h, v6.8h, v2.h[4]\n"
"fmla v20.8h, v6.8h, v3.h[4]\n"
"fmla v24.8h, v6.8h, v4.h[4]\n"
"fmla v28.8h, v6.8h, v5.h[4]\n"
"ldr d6, [x17, #0x120]\n"
"fmla v9.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x148]\n"
+ "ldr x21, [x17, #0x148]\n"
"fmla v17.8h, v7.8h, v2.h[4]\n"
"fmla v21.8h, v7.8h, v3.h[4]\n"
"fmla v25.8h, v7.8h, v4.h[4]\n"
"fmla v29.8h, v7.8h, v5.h[4]\n"
"ldr d7, [x17, #0x130]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[4]\n"
"fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr x11, [x17, #0x158]\n"
+ "ldr x20, [x17, #0x158]\n"
"fmla v18.8h, v6.8h, v2.h[4]\n"
"fmla v22.8h, v6.8h, v3.h[4]\n"
"fmla v26.8h, v6.8h, v4.h[4]\n"
"fmla v30.8h, v6.8h, v5.h[4]\n"
"ldr d6, [x17, #0x140]\n"
"fmla v11.8h, v7.8h, v0.h[4]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr x12, [x17, #0x168]\n"
+ "ldr x21, [x17, #0x168]\n"
"fmla v19.8h, v7.8h, v2.h[4]\n"
"fmla v23.8h, v7.8h, v3.h[4]\n"
"fmla v27.8h, v7.8h, v4.h[4]\n"
"fmla v31.8h, v7.8h, v5.h[4]\n"
"ldr d7, [x17, #0x150]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[5]\n"
"fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x178]\n"
+ "ldr x20, [x17, #0x178]\n"
"fmla v16.8h, v6.8h, v2.h[5]\n"
"fmla v20.8h, v6.8h, v3.h[5]\n"
"fmla v24.8h, v6.8h, v4.h[5]\n"
"fmla v28.8h, v6.8h, v5.h[5]\n"
"ldr d6, [x17, #0x160]\n"
"fmla v9.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x188]\n"
+ "ldr x21, [x17, #0x188]\n"
"fmla v17.8h, v7.8h, v2.h[5]\n"
"fmla v21.8h, v7.8h, v3.h[5]\n"
"fmla v25.8h, v7.8h, v4.h[5]\n"
"fmla v29.8h, v7.8h, v5.h[5]\n"
"ldr d7, [x17, #0x170]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[5]\n"
"fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr x11, [x17, #0x198]\n"
+ "ldr x20, [x17, #0x198]\n"
"fmla v18.8h, v6.8h, v2.h[5]\n"
"fmla v22.8h, v6.8h, v3.h[5]\n"
"fmla v26.8h, v6.8h, v4.h[5]\n"
"fmla v30.8h, v6.8h, v5.h[5]\n"
"ldr d6, [x17, #0x180]\n"
"fmla v11.8h, v7.8h, v0.h[5]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr x12, [x17, #0x1a8]\n"
+ "ldr x21, [x17, #0x1a8]\n"
"fmla v19.8h, v7.8h, v2.h[5]\n"
"fmla v23.8h, v7.8h, v3.h[5]\n"
"fmla v27.8h, v7.8h, v4.h[5]\n"
"fmla v31.8h, v7.8h, v5.h[5]\n"
"ldr d7, [x17, #0x190]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[6]\n"
"fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1b8]\n"
+ "ldr x20, [x17, #0x1b8]\n"
"fmla v16.8h, v6.8h, v2.h[6]\n"
"fmla v20.8h, v6.8h, v3.h[6]\n"
"fmla v24.8h, v6.8h, v4.h[6]\n"
"fmla v28.8h, v6.8h, v5.h[6]\n"
"ldr d6, [x17, #0x1a0]\n"
"fmla v9.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1c8]\n"
+ "ldr x21, [x17, #0x1c8]\n"
"fmla v17.8h, v7.8h, v2.h[6]\n"
"fmla v21.8h, v7.8h, v3.h[6]\n"
"fmla v25.8h, v7.8h, v4.h[6]\n"
"fmla v29.8h, v7.8h, v5.h[6]\n"
"ldr d7, [x17, #0x1b0]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v10.8h, v6.8h, v0.h[6]\n"
"fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr x11, [x17, #0x1d8]\n"
+ "ldr x20, [x17, #0x1d8]\n"
"fmla v18.8h, v6.8h, v2.h[6]\n"
"fmla v22.8h, v6.8h, v3.h[6]\n"
"fmla v26.8h, v6.8h, v4.h[6]\n"
"fmla v30.8h, v6.8h, v5.h[6]\n"
"ldr d6, [x17, #0x1c0]\n"
"fmla v11.8h, v7.8h, v0.h[6]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr x12, [x17, #0x1e8]\n"
+ "ldr x21, [x17, #0x1e8]\n"
"fmla v19.8h, v7.8h, v2.h[6]\n"
"fmla v23.8h, v7.8h, v3.h[6]\n"
"fmla v27.8h, v7.8h, v4.h[6]\n"
"fmla v31.8h, v7.8h, v5.h[6]\n"
"ldr d7, [x17, #0x1d0]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"fmla v8.8h, v6.8h, v0.h[7]\n"
"fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x1f8]\n"
+ "ldr x20, [x17, #0x1f8]\n"
"fmla v16.8h, v6.8h, v2.h[7]\n"
"fmla v20.8h, v6.8h, v3.h[7]\n"
"fmla v24.8h, v6.8h, v4.h[7]\n"
"fmla v28.8h, v6.8h, v5.h[7]\n"
"ldr d6, [x17, #0x1e0]\n"
"fmla v9.8h, v7.8h, v0.h[7]\n"
- "mov v6.d[1], x12\n"
+ "mov v6.d[1], x21\n"
"fmla v13.8h, v7.8h, v1.h[7]\n"
"fmla v17.8h, v7.8h, v2.h[7]\n"
"fmla v21.8h, v7.8h, v3.h[7]\n"
"fmla v25.8h, v7.8h, v4.h[7]\n"
"fmla v29.8h, v7.8h, v5.h[7]\n"
"ldr d7, [x17, #0x1f0]\n"
- "mov v7.d[1], x11\n"
+ "mov v7.d[1], x20\n"
"add x17, x17, #0x200\n"
"fmla v10.8h, v6.8h, v0.h[7]\n"
- "ldr x12, [x17, #0x8]\n"
+ "ldr x21, [x17, #0x8]\n"
"fmla v14.8h, v6.8h, v1.h[7]\n"
- "ldr x11, [x17, #0x18]\n"
+ "ldr x20, [x17, #0x18]\n"
"fmla v18.8h, v6.8h, v2.h[7]\n"
"fmla v22.8h, v6.8h, v3.h[7]\n"
"fmla v26.8h, v6.8h, v4.h[7]\n"
@@ -5079,56 +5079,56 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"fmla v11.8h, v7.8h, v0.h[7]\n"
"ldr d0, [x13, #0x0]\n"
"fmla v15.8h, v7.8h, v1.h[7]\n"
- "ldr d1, [x9, #0x0]\n"
+ "ldr d1, [x12, #0x0]\n"
"fmla v19.8h, v7.8h, v2.h[7]\n"
- "ldr d2, [x27, #0x0]\n"
+ "ldr d2, [x11, #0x0]\n"
"fmla v23.8h, v7.8h, v3.h[7]\n"
- "ldr d3, [x25, #0x0]\n"
+ "ldr d3, [x10, #0x0]\n"
"fmla v27.8h, v7.8h, v4.h[7]\n"
- "ldr d4, [x23, #0x0]\n"
+ "ldr d4, [x9, #0x0]\n"
"fmla v31.8h, v7.8h, v5.h[7]\n"
- "ldr d5, [x21, #0x0]\n"
+ "ldr d5, [x28, #0x0]\n"
"ldr d7, [x17, #0x10]\n"
- "mov v6.d[1], x12\n"
- "mov v0.d[1], x10\n"
- "mov v1.d[1], x28\n"
- "mov v2.d[1], x26\n"
+ "mov v6.d[1], x21\n"
+ "mov v0.d[1], x27\n"
+ "mov v1.d[1], x26\n"
+ "mov v2.d[1], x25\n"
"mov v3.d[1], x24\n"
- "mov v4.d[1], x22\n"
- "mov v5.d[1], x20\n"
- "mov v7.d[1], x11\n"
+ "mov v4.d[1], x23\n"
+ "mov v5.d[1], x22\n"
+ "mov v7.d[1], x20\n"
"bge 271b\n"
"272:" // Height 6: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
"add x13, x13, #0x10\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "add x9, x9, #0x10\n"
+ "add x12, x12, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "add x27, x27, #0x10\n"
+ "add x11, x11, #0x10\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "add x25, x25, #0x10\n"
+ "add x10, x10, #0x10\n"
"fmla v24.8h, v6.8h, v4.h[0]\n"
- "add x23, x23, #0x10\n"
+ "add x9, x9, #0x10\n"
"fmla v28.8h, v6.8h, v5.h[0]\n"
"ldr q6, [x17, #0x20]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "add x21, x21, #0x10\n"
+ "add x28, x28, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
"sub x14, x14, #0x8\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
"prfm pldl1keep, [x13, #0x80]\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "prfm pldl1keep, [x9, #0x80]\n"
+ "prfm pldl1keep, [x12, #0x80]\n"
"fmla v25.8h, v7.8h, v4.h[0]\n"
- "prfm pldl1keep, [x27, #0x80]\n"
+ "prfm pldl1keep, [x11, #0x80]\n"
"fmla v29.8h, v7.8h, v5.h[0]\n"
"ldr q7, [x17, #0x30]\n"
"fmla v10.8h, v6.8h, v0.h[0]\n"
- "prfm pldl1keep, [x25, #0x80]\n"
+ "prfm pldl1keep, [x10, #0x80]\n"
"fmla v14.8h, v6.8h, v1.h[0]\n"
- "prfm pldl1keep, [x23, #0x80]\n"
+ "prfm pldl1keep, [x9, #0x80]\n"
"fmla v18.8h, v6.8h, v2.h[0]\n"
- "prfm pldl1keep, [x21, #0x80]\n"
+ "prfm pldl1keep, [x28, #0x80]\n"
"fmla v22.8h, v6.8h, v3.h[0]\n"
"fmla v26.8h, v6.8h, v4.h[0]\n"
"fmla v30.8h, v6.8h, v5.h[0]\n"
@@ -5338,42 +5338,42 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"273:" // Height 6: Multiply loop: Main loop skip
"cbz x14, 275f\n"
"274:" // Height 6: Multiply loop: Odd block loop
- "ldr h0, [x13], #0x2\n"
+ "ldr h7, [x13], #0x2\n"
"sub x14, x14, #0x1\n"
- "ldr h1, [x9], #0x2\n"
- "ldr h2, [x27], #0x2\n"
- "ldr h3, [x25], #0x2\n"
- "ldr h4, [x23], #0x2\n"
- "ldr h5, [x21], #0x2\n"
- "ldr q6, [x17, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x17, #0x10]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "fmla v16.8h, v6.8h, v2.h[0]\n"
- "fmla v20.8h, v6.8h, v3.h[0]\n"
- "fmla v24.8h, v6.8h, v4.h[0]\n"
- "fmla v28.8h, v6.8h, v5.h[0]\n"
- "ldr q6, [x17, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "fmla v17.8h, v7.8h, v2.h[0]\n"
- "fmla v21.8h, v7.8h, v3.h[0]\n"
- "fmla v25.8h, v7.8h, v4.h[0]\n"
- "fmla v29.8h, v7.8h, v5.h[0]\n"
- "ldr q7, [x17, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
+ "ldr h6, [x12], #0x2\n"
+ "ldr h5, [x11], #0x2\n"
+ "ldr h4, [x10], #0x2\n"
+ "ldr h3, [x9], #0x2\n"
+ "ldr h2, [x28], #0x2\n"
+ "ldr q1, [x17, #0x0]\n"
+ "fmla v8.8h, v1.8h, v7.h[0]\n"
+ "ldr q0, [x17, #0x10]\n"
+ "fmla v12.8h, v1.8h, v6.h[0]\n"
+ "fmla v16.8h, v1.8h, v5.h[0]\n"
+ "fmla v20.8h, v1.8h, v4.h[0]\n"
+ "fmla v24.8h, v1.8h, v3.h[0]\n"
+ "fmla v28.8h, v1.8h, v2.h[0]\n"
+ "ldr q1, [x17, #0x20]\n"
+ "fmla v9.8h, v0.8h, v7.h[0]\n"
+ "fmla v13.8h, v0.8h, v6.h[0]\n"
+ "fmla v17.8h, v0.8h, v5.h[0]\n"
+ "fmla v21.8h, v0.8h, v4.h[0]\n"
+ "fmla v25.8h, v0.8h, v3.h[0]\n"
+ "fmla v29.8h, v0.8h, v2.h[0]\n"
+ "ldr q0, [x17, #0x30]\n"
+ "fmla v10.8h, v1.8h, v7.h[0]\n"
"add x17, x17, #0x40\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "fmla v22.8h, v6.8h, v3.h[0]\n"
- "fmla v26.8h, v6.8h, v4.h[0]\n"
- "fmla v30.8h, v6.8h, v5.h[0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
- "fmla v23.8h, v7.8h, v3.h[0]\n"
- "fmla v27.8h, v7.8h, v4.h[0]\n"
- "fmla v31.8h, v7.8h, v5.h[0]\n"
+ "fmla v14.8h, v1.8h, v6.h[0]\n"
+ "fmla v18.8h, v1.8h, v5.h[0]\n"
+ "fmla v22.8h, v1.8h, v4.h[0]\n"
+ "fmla v26.8h, v1.8h, v3.h[0]\n"
+ "fmla v30.8h, v1.8h, v2.h[0]\n"
+ "fmla v11.8h, v0.8h, v7.h[0]\n"
+ "fmla v15.8h, v0.8h, v6.h[0]\n"
+ "fmla v19.8h, v0.8h, v5.h[0]\n"
+ "fmla v23.8h, v0.8h, v4.h[0]\n"
+ "fmla v27.8h, v0.8h, v3.h[0]\n"
+ "fmla v31.8h, v0.8h, v2.h[0]\n"
"cbnz x14, 274b\n"
"275:" // Height 6: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -5743,7 +5743,6 @@ void a64_hybrid_fp16_mla_6x32_a55 (
"madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
"b 1b\n"
"296:" // Exit
-
: [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
: [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"