aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp3020
1 files changed, 1510 insertions, 1510 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
index 18a2db5069..8038612200 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
@@ -265,11 +265,11 @@ void a64_ffhybrid_fp16_mla_6x32 (
"24:" // Height 1: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 25f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
"cbnz x28, 26f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20, LSL #1\n"
@@ -286,69 +286,69 @@ void a64_ffhybrid_fp16_mla_6x32 (
"blt 28f\n"
"27:" // Height 1: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q17, [x10, #0x0]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "ldr q6, [x12, #0x10]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "ldr q7, [x11, #0x10]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "ldr q6, [x10, #0x10]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "ldr q7, [x9, #0x10]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "ldr q6, [x12, #0x20]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "ldr q7, [x11, #0x20]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "ldr q6, [x10, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "ldr q7, [x9, #0x20]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "ldr q6, [x12, #0x30]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "ldr q7, [x11, #0x30]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "ldr q6, [x10, #0x30]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "ldr q7, [x9, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "ldr q6, [x12, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "ldr q7, [x11, #0x40]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "ldr q6, [x10, #0x40]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "ldr q7, [x9, #0x40]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "ldr q6, [x12, #0x50]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "ldr q7, [x11, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "ldr q6, [x10, #0x50]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "ldr q7, [x9, #0x50]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "ldr q6, [x12, #0x60]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "ldr q7, [x11, #0x60]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "ldr q6, [x10, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "ldr q7, [x9, #0x60]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "ldr q6, [x12, #0x70]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "ldr q7, [x11, #0x70]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "ldr q6, [x10, #0x70]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "ldr q7, [x9, #0x70]\n"
+ "ldr q16, [x9, #0x0]\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "ldr q17, [x12, #0x10]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x11, #0x10]\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "ldr q17, [x10, #0x10]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "ldr q16, [x9, #0x10]\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "ldr q17, [x12, #0x20]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "ldr q16, [x11, #0x20]\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "ldr q17, [x10, #0x20]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "ldr q16, [x9, #0x20]\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "ldr q17, [x12, #0x30]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "ldr q16, [x11, #0x30]\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "ldr q17, [x10, #0x30]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "ldr q16, [x9, #0x30]\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "ldr q17, [x12, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "ldr q16, [x11, #0x40]\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "ldr q17, [x10, #0x40]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "ldr q16, [x9, #0x40]\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "ldr q17, [x12, #0x50]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "ldr q16, [x11, #0x50]\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "ldr q17, [x10, #0x50]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "ldr q16, [x9, #0x50]\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "ldr q17, [x12, #0x60]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "ldr q16, [x11, #0x60]\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "ldr q17, [x10, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "ldr q16, [x9, #0x60]\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "ldr q17, [x12, #0x70]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "ldr q16, [x11, #0x70]\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "ldr q17, [x10, #0x70]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "ldr q16, [x9, #0x70]\n"
"sub x27, x27, #0x8\n"
"cmp x27, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v10.8h, v17.8h, v0.h[7]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
"add x26, x26, #0x10\n"
"ldr q0, [x26, #0x0]\n"
"add x12, x12, #0x80\n"
@@ -360,68 +360,68 @@ void a64_ffhybrid_fp16_mla_6x32 (
"bge 27b\n"
"28:" // Height 1: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q17, [x10, #0x0]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "ldr q6, [x12, #0x10]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "ldr q7, [x11, #0x10]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "ldr q6, [x10, #0x10]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "ldr q7, [x9, #0x10]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "ldr q6, [x12, #0x20]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "ldr q7, [x11, #0x20]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "ldr q6, [x10, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "ldr q7, [x9, #0x20]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "ldr q6, [x12, #0x30]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "ldr q7, [x11, #0x30]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "ldr q6, [x10, #0x30]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "ldr q7, [x9, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "ldr q6, [x12, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "ldr q7, [x11, #0x40]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "ldr q6, [x10, #0x40]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "ldr q7, [x9, #0x40]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "ldr q6, [x12, #0x50]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "ldr q7, [x11, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "ldr q6, [x10, #0x50]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "ldr q7, [x9, #0x50]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "ldr q6, [x12, #0x60]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "ldr q7, [x11, #0x60]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "ldr q6, [x10, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "ldr q7, [x9, #0x60]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "ldr q6, [x12, #0x70]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "ldr q7, [x11, #0x70]\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "ldr q6, [x10, #0x70]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "ldr q7, [x9, #0x70]\n"
+ "ldr q16, [x9, #0x0]\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "ldr q17, [x12, #0x10]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x11, #0x10]\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "ldr q17, [x10, #0x10]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "ldr q16, [x9, #0x10]\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "ldr q17, [x12, #0x20]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "ldr q16, [x11, #0x20]\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "ldr q17, [x10, #0x20]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "ldr q16, [x9, #0x20]\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "ldr q17, [x12, #0x30]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "ldr q16, [x11, #0x30]\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "ldr q17, [x10, #0x30]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "ldr q16, [x9, #0x30]\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "ldr q17, [x12, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "ldr q16, [x11, #0x40]\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "ldr q17, [x10, #0x40]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "ldr q16, [x9, #0x40]\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "ldr q17, [x12, #0x50]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "ldr q16, [x11, #0x50]\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "ldr q17, [x10, #0x50]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "ldr q16, [x9, #0x50]\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "ldr q17, [x12, #0x60]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "ldr q16, [x11, #0x60]\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "ldr q17, [x10, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "ldr q16, [x9, #0x60]\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "ldr q17, [x12, #0x70]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "ldr q16, [x11, #0x70]\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "ldr q17, [x10, #0x70]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "ldr q16, [x9, #0x70]\n"
"sub x27, x27, #0x8\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v10.8h, v17.8h, v0.h[7]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
"add x26, x26, #0x10\n"
"add x12, x12, #0x80\n"
"add x11, x11, #0x80\n"
@@ -431,15 +431,15 @@ void a64_ffhybrid_fp16_mla_6x32 (
"cbz x27, 31f\n"
"30:" // Height 1: Multiply loop: Odd block loop
"ldr h0, [x26], #0x2\n"
- "ldr q6, [x12, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
+ "ldr q16, [x12, #0x0]\n"
+ "fmla v8.8h, v16.8h, v0.h[0]\n"
"sub x27, x27, #0x1\n"
- "ldr q7, [x11, #0x0]\n"
- "ldr q6, [x10, #0x0]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
+ "ldr q17, [x11, #0x0]\n"
+ "ldr q16, [x10, #0x0]\n"
+ "fmla v9.8h, v17.8h, v0.h[0]\n"
+ "fmla v10.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x9, #0x0]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
"add x12, x12, #0x10\n"
"add x11, x11, #0x10\n"
"add x10, x10, #0x10\n"
@@ -452,17 +452,17 @@ void a64_ffhybrid_fp16_mla_6x32 (
"bne 24b\n"
"tbz %x[flags], #1, 32f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v1.8h }, [x20]\n"
+ "ld1r { v17.8h }, [x20]\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v1.8h\n"
- "fmin v9.8h, v9.8h, v1.8h\n"
- "fmin v10.8h, v10.8h, v1.8h\n"
- "fmin v11.8h, v11.8h, v1.8h\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"
+ "fmin v8.8h, v8.8h, v17.8h\n"
+ "fmin v9.8h, v9.8h, v17.8h\n"
+ "fmin v10.8h, v10.8h, v17.8h\n"
+ "fmin v11.8h, v11.8h, v17.8h\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"
"32:" // Height 1: No activation
"cmp x14, #0x20\n"
"bge 49f\n"
@@ -778,12 +778,12 @@ void a64_ffhybrid_fp16_mla_6x32 (
"74:" // Height 2: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 75f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
"cbnz x28, 76f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20, LSL #1\n"
@@ -791,7 +791,7 @@ void a64_ffhybrid_fp16_mla_6x32 (
"b 76f\n"
"75:" // Height 2: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20, LSL #1\n"
+ "add x25, x26, x21, LSL #1\n"
"76:" // Height 2: input setup done
"cmp x27, #0x8\n"
"blt 79f\n"
@@ -804,233 +804,233 @@ void a64_ffhybrid_fp16_mla_6x32 (
"77:" // Height 2: Multiply loop: Main loop head
"fmla v8.8h, v6.8h, v0.h[0]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q17, [x10, #0x0]\n"
"sub x27, x27, #0x8\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "ldr q16, [x9, #0x0]\n"
"cmp x27, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x12, #0x10]\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "fmla v14.8h, v17.8h, v1.h[0]\n"
+ "ldr q17, [x12, #0x10]\n"
"add x26, x26, #0x10\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x11, #0x10]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "fmla v15.8h, v16.8h, v1.h[0]\n"
+ "ldr q16, [x11, #0x10]\n"
"add x25, x25, #0x10\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr q6, [x10, #0x10]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr q7, [x9, #0x10]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr q6, [x12, #0x20]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr q7, [x11, #0x20]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr q6, [x10, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr q7, [x9, #0x20]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr q6, [x12, #0x30]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr q7, [x11, #0x30]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr q6, [x10, #0x30]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr q7, [x9, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr q6, [x12, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr q7, [x11, #0x40]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr q6, [x10, #0x40]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr q7, [x9, #0x40]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr q6, [x12, #0x50]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr q7, [x11, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr q6, [x10, #0x50]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr q7, [x9, #0x50]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr q6, [x12, #0x60]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr q7, [x11, #0x60]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr q6, [x10, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr q7, [x9, #0x60]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr q6, [x12, #0x70]\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "fmla v12.8h, v17.8h, v1.h[1]\n"
+ "ldr q17, [x10, #0x10]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "fmla v13.8h, v16.8h, v1.h[1]\n"
+ "ldr q16, [x9, #0x10]\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "fmla v14.8h, v17.8h, v1.h[1]\n"
+ "ldr q17, [x12, #0x20]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "fmla v15.8h, v16.8h, v1.h[1]\n"
+ "ldr q16, [x11, #0x20]\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "fmla v12.8h, v17.8h, v1.h[2]\n"
+ "ldr q17, [x10, #0x20]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "fmla v13.8h, v16.8h, v1.h[2]\n"
+ "ldr q16, [x9, #0x20]\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "fmla v14.8h, v17.8h, v1.h[2]\n"
+ "ldr q17, [x12, #0x30]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "fmla v15.8h, v16.8h, v1.h[2]\n"
+ "ldr q16, [x11, #0x30]\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "fmla v12.8h, v17.8h, v1.h[3]\n"
+ "ldr q17, [x10, #0x30]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "fmla v13.8h, v16.8h, v1.h[3]\n"
+ "ldr q16, [x9, #0x30]\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "fmla v14.8h, v17.8h, v1.h[3]\n"
+ "ldr q17, [x12, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "fmla v15.8h, v16.8h, v1.h[3]\n"
+ "ldr q16, [x11, #0x40]\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "fmla v12.8h, v17.8h, v1.h[4]\n"
+ "ldr q17, [x10, #0x40]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "fmla v13.8h, v16.8h, v1.h[4]\n"
+ "ldr q16, [x9, #0x40]\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "fmla v14.8h, v17.8h, v1.h[4]\n"
+ "ldr q17, [x12, #0x50]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "fmla v15.8h, v16.8h, v1.h[4]\n"
+ "ldr q16, [x11, #0x50]\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "fmla v12.8h, v17.8h, v1.h[5]\n"
+ "ldr q17, [x10, #0x50]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "fmla v13.8h, v16.8h, v1.h[5]\n"
+ "ldr q16, [x9, #0x50]\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "fmla v14.8h, v17.8h, v1.h[5]\n"
+ "ldr q17, [x12, #0x60]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "fmla v15.8h, v16.8h, v1.h[5]\n"
+ "ldr q16, [x11, #0x60]\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "fmla v12.8h, v17.8h, v1.h[6]\n"
+ "ldr q17, [x10, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "fmla v13.8h, v16.8h, v1.h[6]\n"
+ "ldr q16, [x9, #0x60]\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "fmla v14.8h, v17.8h, v1.h[6]\n"
+ "ldr q17, [x12, #0x70]\n"
"add x12, x12, #0x80\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr q7, [x11, #0x70]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "fmla v15.8h, v16.8h, v1.h[6]\n"
+ "ldr q16, [x11, #0x70]\n"
"add x11, x11, #0x80\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr q6, [x10, #0x70]\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "fmla v12.8h, v17.8h, v1.h[7]\n"
+ "ldr q17, [x10, #0x70]\n"
"add x10, x10, #0x80\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "ldr q7, [x9, #0x70]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "fmla v13.8h, v16.8h, v1.h[7]\n"
+ "ldr q16, [x9, #0x70]\n"
"add x9, x9, #0x80\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 q6, [x12, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v16.8h, v0.h[7]\n"
"ldr q0, [x26, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
+ "fmla v15.8h, v16.8h, v1.h[7]\n"
"ldr q1, [x25, #0x0]\n"
"ldr q7, [x11, #0x0]\n"
"bge 77b\n"
"78:" // Height 2: Multiply loop: Single iteration only
"fmla v8.8h, v6.8h, v0.h[0]\n"
"fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q17, [x10, #0x0]\n"
"sub x27, x27, #0x8\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "ldr q16, [x9, #0x0]\n"
"add x26, x26, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x12, #0x10]\n"
+ "fmla v10.8h, v17.8h, v0.h[0]\n"
+ "fmla v14.8h, v17.8h, v1.h[0]\n"
+ "ldr q17, [x12, #0x10]\n"
"add x25, x25, #0x10\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x11, #0x10]\n"
- "fmla v8.8h, v6.8h, v0.h[1]\n"
- "fmla v12.8h, v6.8h, v1.h[1]\n"
- "ldr q6, [x10, #0x10]\n"
- "fmla v9.8h, v7.8h, v0.h[1]\n"
- "fmla v13.8h, v7.8h, v1.h[1]\n"
- "ldr q7, [x9, #0x10]\n"
- "fmla v10.8h, v6.8h, v0.h[1]\n"
- "fmla v14.8h, v6.8h, v1.h[1]\n"
- "ldr q6, [x12, #0x20]\n"
- "fmla v11.8h, v7.8h, v0.h[1]\n"
- "fmla v15.8h, v7.8h, v1.h[1]\n"
- "ldr q7, [x11, #0x20]\n"
- "fmla v8.8h, v6.8h, v0.h[2]\n"
- "fmla v12.8h, v6.8h, v1.h[2]\n"
- "ldr q6, [x10, #0x20]\n"
- "fmla v9.8h, v7.8h, v0.h[2]\n"
- "fmla v13.8h, v7.8h, v1.h[2]\n"
- "ldr q7, [x9, #0x20]\n"
- "fmla v10.8h, v6.8h, v0.h[2]\n"
- "fmla v14.8h, v6.8h, v1.h[2]\n"
- "ldr q6, [x12, #0x30]\n"
- "fmla v11.8h, v7.8h, v0.h[2]\n"
- "fmla v15.8h, v7.8h, v1.h[2]\n"
- "ldr q7, [x11, #0x30]\n"
- "fmla v8.8h, v6.8h, v0.h[3]\n"
- "fmla v12.8h, v6.8h, v1.h[3]\n"
- "ldr q6, [x10, #0x30]\n"
- "fmla v9.8h, v7.8h, v0.h[3]\n"
- "fmla v13.8h, v7.8h, v1.h[3]\n"
- "ldr q7, [x9, #0x30]\n"
- "fmla v10.8h, v6.8h, v0.h[3]\n"
- "fmla v14.8h, v6.8h, v1.h[3]\n"
- "ldr q6, [x12, #0x40]\n"
- "fmla v11.8h, v7.8h, v0.h[3]\n"
- "fmla v15.8h, v7.8h, v1.h[3]\n"
- "ldr q7, [x11, #0x40]\n"
- "fmla v8.8h, v6.8h, v0.h[4]\n"
- "fmla v12.8h, v6.8h, v1.h[4]\n"
- "ldr q6, [x10, #0x40]\n"
- "fmla v9.8h, v7.8h, v0.h[4]\n"
- "fmla v13.8h, v7.8h, v1.h[4]\n"
- "ldr q7, [x9, #0x40]\n"
- "fmla v10.8h, v6.8h, v0.h[4]\n"
- "fmla v14.8h, v6.8h, v1.h[4]\n"
- "ldr q6, [x12, #0x50]\n"
- "fmla v11.8h, v7.8h, v0.h[4]\n"
- "fmla v15.8h, v7.8h, v1.h[4]\n"
- "ldr q7, [x11, #0x50]\n"
- "fmla v8.8h, v6.8h, v0.h[5]\n"
- "fmla v12.8h, v6.8h, v1.h[5]\n"
- "ldr q6, [x10, #0x50]\n"
- "fmla v9.8h, v7.8h, v0.h[5]\n"
- "fmla v13.8h, v7.8h, v1.h[5]\n"
- "ldr q7, [x9, #0x50]\n"
- "fmla v10.8h, v6.8h, v0.h[5]\n"
- "fmla v14.8h, v6.8h, v1.h[5]\n"
- "ldr q6, [x12, #0x60]\n"
- "fmla v11.8h, v7.8h, v0.h[5]\n"
- "fmla v15.8h, v7.8h, v1.h[5]\n"
- "ldr q7, [x11, #0x60]\n"
- "fmla v8.8h, v6.8h, v0.h[6]\n"
- "fmla v12.8h, v6.8h, v1.h[6]\n"
- "ldr q6, [x10, #0x60]\n"
- "fmla v9.8h, v7.8h, v0.h[6]\n"
- "fmla v13.8h, v7.8h, v1.h[6]\n"
- "ldr q7, [x9, #0x60]\n"
- "fmla v10.8h, v6.8h, v0.h[6]\n"
- "fmla v14.8h, v6.8h, v1.h[6]\n"
- "ldr q6, [x12, #0x70]\n"
+ "fmla v11.8h, v16.8h, v0.h[0]\n"
+ "fmla v15.8h, v16.8h, v1.h[0]\n"
+ "ldr q16, [x11, #0x10]\n"
+ "fmla v8.8h, v17.8h, v0.h[1]\n"
+ "fmla v12.8h, v17.8h, v1.h[1]\n"
+ "ldr q17, [x10, #0x10]\n"
+ "fmla v9.8h, v16.8h, v0.h[1]\n"
+ "fmla v13.8h, v16.8h, v1.h[1]\n"
+ "ldr q16, [x9, #0x10]\n"
+ "fmla v10.8h, v17.8h, v0.h[1]\n"
+ "fmla v14.8h, v17.8h, v1.h[1]\n"
+ "ldr q17, [x12, #0x20]\n"
+ "fmla v11.8h, v16.8h, v0.h[1]\n"
+ "fmla v15.8h, v16.8h, v1.h[1]\n"
+ "ldr q16, [x11, #0x20]\n"
+ "fmla v8.8h, v17.8h, v0.h[2]\n"
+ "fmla v12.8h, v17.8h, v1.h[2]\n"
+ "ldr q17, [x10, #0x20]\n"
+ "fmla v9.8h, v16.8h, v0.h[2]\n"
+ "fmla v13.8h, v16.8h, v1.h[2]\n"
+ "ldr q16, [x9, #0x20]\n"
+ "fmla v10.8h, v17.8h, v0.h[2]\n"
+ "fmla v14.8h, v17.8h, v1.h[2]\n"
+ "ldr q17, [x12, #0x30]\n"
+ "fmla v11.8h, v16.8h, v0.h[2]\n"
+ "fmla v15.8h, v16.8h, v1.h[2]\n"
+ "ldr q16, [x11, #0x30]\n"
+ "fmla v8.8h, v17.8h, v0.h[3]\n"
+ "fmla v12.8h, v17.8h, v1.h[3]\n"
+ "ldr q17, [x10, #0x30]\n"
+ "fmla v9.8h, v16.8h, v0.h[3]\n"
+ "fmla v13.8h, v16.8h, v1.h[3]\n"
+ "ldr q16, [x9, #0x30]\n"
+ "fmla v10.8h, v17.8h, v0.h[3]\n"
+ "fmla v14.8h, v17.8h, v1.h[3]\n"
+ "ldr q17, [x12, #0x40]\n"
+ "fmla v11.8h, v16.8h, v0.h[3]\n"
+ "fmla v15.8h, v16.8h, v1.h[3]\n"
+ "ldr q16, [x11, #0x40]\n"
+ "fmla v8.8h, v17.8h, v0.h[4]\n"
+ "fmla v12.8h, v17.8h, v1.h[4]\n"
+ "ldr q17, [x10, #0x40]\n"
+ "fmla v9.8h, v16.8h, v0.h[4]\n"
+ "fmla v13.8h, v16.8h, v1.h[4]\n"
+ "ldr q16, [x9, #0x40]\n"
+ "fmla v10.8h, v17.8h, v0.h[4]\n"
+ "fmla v14.8h, v17.8h, v1.h[4]\n"
+ "ldr q17, [x12, #0x50]\n"
+ "fmla v11.8h, v16.8h, v0.h[4]\n"
+ "fmla v15.8h, v16.8h, v1.h[4]\n"
+ "ldr q16, [x11, #0x50]\n"
+ "fmla v8.8h, v17.8h, v0.h[5]\n"
+ "fmla v12.8h, v17.8h, v1.h[5]\n"
+ "ldr q17, [x10, #0x50]\n"
+ "fmla v9.8h, v16.8h, v0.h[5]\n"
+ "fmla v13.8h, v16.8h, v1.h[5]\n"
+ "ldr q16, [x9, #0x50]\n"
+ "fmla v10.8h, v17.8h, v0.h[5]\n"
+ "fmla v14.8h, v17.8h, v1.h[5]\n"
+ "ldr q17, [x12, #0x60]\n"
+ "fmla v11.8h, v16.8h, v0.h[5]\n"
+ "fmla v15.8h, v16.8h, v1.h[5]\n"
+ "ldr q16, [x11, #0x60]\n"
+ "fmla v8.8h, v17.8h, v0.h[6]\n"
+ "fmla v12.8h, v17.8h, v1.h[6]\n"
+ "ldr q17, [x10, #0x60]\n"
+ "fmla v9.8h, v16.8h, v0.h[6]\n"
+ "fmla v13.8h, v16.8h, v1.h[6]\n"
+ "ldr q16, [x9, #0x60]\n"
+ "fmla v10.8h, v17.8h, v0.h[6]\n"
+ "fmla v14.8h, v17.8h, v1.h[6]\n"
+ "ldr q17, [x12, #0x70]\n"
"add x12, x12, #0x80\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "ldr q7, [x11, #0x70]\n"
+ "fmla v11.8h, v16.8h, v0.h[6]\n"
+ "fmla v15.8h, v16.8h, v1.h[6]\n"
+ "ldr q16, [x11, #0x70]\n"
"add x11, x11, #0x80\n"
- "fmla v8.8h, v6.8h, v0.h[7]\n"
- "fmla v12.8h, v6.8h, v1.h[7]\n"
- "ldr q6, [x10, #0x70]\n"
+ "fmla v8.8h, v17.8h, v0.h[7]\n"
+ "fmla v12.8h, v17.8h, v1.h[7]\n"
+ "ldr q17, [x10, #0x70]\n"
"add x10, x10, #0x80\n"
- "fmla v9.8h, v7.8h, v0.h[7]\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "ldr q7, [x9, #0x70]\n"
+ "fmla v9.8h, v16.8h, v0.h[7]\n"
+ "fmla v13.8h, v16.8h, v1.h[7]\n"
+ "ldr q16, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\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 v10.8h, v17.8h, v0.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"
"79:" // Height 2: Multiply loop: Main loop skip
"cbz x27, 81f\n"
"80:" // Height 2: Multiply loop: Odd block loop
- "ldr h0, [x26], #0x2\n"
- "ldr h1, [x25], #0x2\n"
+ "ldr h1, [x26], #0x2\n"
+ "ldr h0, [x25], #0x2\n"
"sub x27, x27, #0x1\n"
- "ldr q6, [x12, #0x0]\n"
- "ldr q7, [x11, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
+ "ldr q17, [x12, #0x0]\n"
+ "ldr q16, [x11, #0x0]\n"
+ "fmla v8.8h, v17.8h, v1.h[0]\n"
+ "fmla v12.8h, v17.8h, v0.h[0]\n"
+ "ldr q17, [x10, #0x0]\n"
+ "fmla v9.8h, v16.8h, v1.h[0]\n"
+ "fmla v13.8h, v16.8h, v0.h[0]\n"
+ "ldr q16, [x9, #0x0]\n"
+ "fmla v10.8h, v17.8h, v1.h[0]\n"
+ "fmla v14.8h, v17.8h, v0.h[0]\n"
"add x12, x12, #0x10\n"
"add x11, x11, #0x10\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
+ "fmla v11.8h, v16.8h, v1.h[0]\n"
+ "fmla v15.8h, v16.8h, v0.h[0]\n"
"add x10, x10, #0x10\n"
"add x9, x9, #0x10\n"
"cbnz x27, 80b\n"
@@ -1043,25 +1043,25 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x25, x13, x20, LSL #1\n"
"tbz %x[flags], #1, 82f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v1.8h }, [x20]\n"
+ "ld1r { v17.8h }, [x20]\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v1.8h\n"
- "fmin v9.8h, v9.8h, v1.8h\n"
- "fmin v10.8h, v10.8h, v1.8h\n"
- "fmin v11.8h, v11.8h, v1.8h\n"
- "fmin v12.8h, v12.8h, v1.8h\n"
- "fmin v13.8h, v13.8h, v1.8h\n"
- "fmin v14.8h, v14.8h, v1.8h\n"
- "fmin v15.8h, v15.8h, v1.8h\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"
+ "fmin v8.8h, v8.8h, v17.8h\n"
+ "fmin v9.8h, v9.8h, v17.8h\n"
+ "fmin v10.8h, v10.8h, v17.8h\n"
+ "fmin v11.8h, v11.8h, v17.8h\n"
+ "fmin v12.8h, v12.8h, v17.8h\n"
+ "fmin v13.8h, v13.8h, v17.8h\n"
+ "fmin v14.8h, v14.8h, v17.8h\n"
+ "fmin v15.8h, v15.8h, v17.8h\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"
"82:" // Height 2: No activation
"cmp x14, #0x20\n"
"bge 99f\n"
@@ -1458,13 +1458,13 @@ void a64_ffhybrid_fp16_mla_6x32 (
"124:" // Height 3: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 125f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
"cbnz x28, 126f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20, LSL #1\n"
@@ -1473,8 +1473,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
"b 126f\n"
"125:" // Height 3: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20, LSL #1\n"
- "add x24, x25, x20, LSL #1\n"
+ "add x25, x26, x21, LSL #1\n"
+ "add x24, x25, x21, LSL #1\n"
"126:" // Height 3: input setup done
"cmp x27, #0x8\n"
"blt 129f\n"
@@ -1491,139 +1491,139 @@ void a64_ffhybrid_fp16_mla_6x32 (
"sub x27, x27, #0x8\n"
"cmp x27, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q21, [x10, #0x0]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"add x26, x26, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "ldr q20, [x9, #0x0]\n"
"add x25, x25, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
+ "fmla v10.8h, v21.8h, v0.h[0]\n"
+ "fmla v14.8h, v21.8h, v1.h[0]\n"
"add x24, x24, #0x10\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x12, #0x10]\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"
- "ldr q7, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
+ "fmla v18.8h, v21.8h, v2.h[0]\n"
+ "ldr q21, [x12, #0x10]\n"
+ "fmla v11.8h, v20.8h, v0.h[0]\n"
+ "fmla v15.8h, v20.8h, v1.h[0]\n"
+ "fmla v19.8h, v20.8h, v2.h[0]\n"
+ "ldr q20, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
+ "fmla v11.8h, v20.8h, v0.h[6]\n"
"add x12, x12, #0x80\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "ldr q7, [x11, #0x70]\n"
+ "fmla v15.8h, v20.8h, v1.h[6]\n"
+ "fmla v19.8h, v20.8h, v2.h[6]\n"
+ "ldr q20, [x11, #0x70]\n"
"add x11, x11, #0x80\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, [x10, #0x70]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\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, [x10, #0x70]\n"
+ "fmla v9.8h, v20.8h, v0.h[7]\n"
"add x10, x10, #0x80\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "ldr q7, [x9, #0x70]\n"
+ "fmla v13.8h, v20.8h, v1.h[7]\n"
+ "fmla v17.8h, v20.8h, v2.h[7]\n"
+ "ldr q20, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\n"
- "fmla v14.8h, v6.8h, v1.h[7]\n"
- "fmla v18.8h, v6.8h, v2.h[7]\n"
+ "fmla v10.8h, v21.8h, v0.h[7]\n"
+ "fmla v14.8h, v21.8h, v1.h[7]\n"
+ "fmla v18.8h, v21.8h, v2.h[7]\n"
"ldr q6, [x12, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v20.8h, v0.h[7]\n"
"ldr q0, [x26, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
+ "fmla v15.8h, v20.8h, v1.h[7]\n"
"ldr q1, [x25, #0x0]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
+ "fmla v19.8h, v20.8h, v2.h[7]\n"
"ldr q2, [x24, #0x0]\n"
"ldr q7, [x11, #0x0]\n"
"bge 127b\n"
@@ -1633,162 +1633,162 @@ void a64_ffhybrid_fp16_mla_6x32 (
"sub x27, x27, #0x8\n"
"add x26, x26, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q21, [x10, #0x0]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"add x25, x25, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "ldr q20, [x9, #0x0]\n"
"add x24, x24, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x12, #0x10]\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"
- "ldr q7, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
+ "fmla v10.8h, v21.8h, v0.h[0]\n"
+ "fmla v14.8h, v21.8h, v1.h[0]\n"
+ "fmla v18.8h, v21.8h, v2.h[0]\n"
+ "ldr q21, [x12, #0x10]\n"
+ "fmla v11.8h, v20.8h, v0.h[0]\n"
+ "fmla v15.8h, v20.8h, v1.h[0]\n"
+ "fmla v19.8h, v20.8h, v2.h[0]\n"
+ "ldr q20, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
+ "fmla v11.8h, v20.8h, v0.h[6]\n"
"add x12, x12, #0x80\n"
- "fmla v15.8h, v7.8h, v1.h[6]\n"
- "fmla v19.8h, v7.8h, v2.h[6]\n"
- "ldr q7, [x11, #0x70]\n"
+ "fmla v15.8h, v20.8h, v1.h[6]\n"
+ "fmla v19.8h, v20.8h, v2.h[6]\n"
+ "ldr q20, [x11, #0x70]\n"
"add x11, x11, #0x80\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, [x10, #0x70]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\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, [x10, #0x70]\n"
+ "fmla v9.8h, v20.8h, v0.h[7]\n"
"add x10, x10, #0x80\n"
- "fmla v13.8h, v7.8h, v1.h[7]\n"
- "fmla v17.8h, v7.8h, v2.h[7]\n"
- "ldr q7, [x9, #0x70]\n"
+ "fmla v13.8h, v20.8h, v1.h[7]\n"
+ "fmla v17.8h, v20.8h, v2.h[7]\n"
+ "ldr q20, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\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 v10.8h, v21.8h, v0.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"
"129:" // Height 3: Multiply loop: Main loop skip
"cbz x27, 131f\n"
"130:" // Height 3: Multiply loop: Odd block loop
- "ldr h0, [x26], #0x2\n"
+ "ldr h2, [x26], #0x2\n"
"ldr h1, [x25], #0x2\n"
"sub x27, x27, #0x1\n"
- "ldr h2, [x24], #0x2\n"
- "ldr q6, [x12, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q7, [x11, #0x0]\n"
- "fmla v16.8h, v6.8h, v2.h[0]\n"
- "ldr q6, [x10, #0x0]\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, [x9, #0x0]\n"
+ "ldr h0, [x24], #0x2\n"
+ "ldr q21, [x12, #0x0]\n"
+ "fmla v8.8h, v21.8h, v2.h[0]\n"
+ "fmla v12.8h, v21.8h, v1.h[0]\n"
+ "ldr q20, [x11, #0x0]\n"
+ "fmla v16.8h, v21.8h, v0.h[0]\n"
+ "ldr q21, [x10, #0x0]\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, [x9, #0x0]\n"
"add x12, x12, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\n"
- "fmla v14.8h, v6.8h, v1.h[0]\n"
+ "fmla v10.8h, v21.8h, v2.h[0]\n"
+ "fmla v14.8h, v21.8h, v1.h[0]\n"
"add x11, x11, #0x10\n"
"add x10, x10, #0x10\n"
- "fmla v18.8h, v6.8h, v2.h[0]\n"
- "fmla v11.8h, v7.8h, v0.h[0]\n"
+ "fmla v18.8h, v21.8h, v0.h[0]\n"
+ "fmla v11.8h, v20.8h, v2.h[0]\n"
"add x9, x9, #0x10\n"
- "fmla v15.8h, v7.8h, v1.h[0]\n"
- "fmla v19.8h, v7.8h, v2.h[0]\n"
+ "fmla v15.8h, v20.8h, v1.h[0]\n"
+ "fmla v19.8h, v20.8h, v0.h[0]\n"
"cbnz x27, 130b\n"
"131:" // Height 3: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -1800,33 +1800,33 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x24, x25, x20, LSL #1\n"
"tbz %x[flags], #1, 132f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v1.8h }, [x20]\n"
+ "ld1r { v21.8h }, [x20]\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v1.8h\n"
- "fmin v9.8h, v9.8h, v1.8h\n"
- "fmin v10.8h, v10.8h, v1.8h\n"
- "fmin v11.8h, v11.8h, v1.8h\n"
- "fmin v12.8h, v12.8h, v1.8h\n"
- "fmin v13.8h, v13.8h, v1.8h\n"
- "fmin v14.8h, v14.8h, v1.8h\n"
- "fmin v15.8h, v15.8h, v1.8h\n"
- "fmin v16.8h, v16.8h, v1.8h\n"
- "fmin v17.8h, v17.8h, v1.8h\n"
- "fmin v18.8h, v18.8h, v1.8h\n"
- "fmin v19.8h, v19.8h, v1.8h\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"
+ "fmin v8.8h, v8.8h, v21.8h\n"
+ "fmin v9.8h, v9.8h, v21.8h\n"
+ "fmin v10.8h, v10.8h, v21.8h\n"
+ "fmin v11.8h, v11.8h, v21.8h\n"
+ "fmin v12.8h, v12.8h, v21.8h\n"
+ "fmin v13.8h, v13.8h, v21.8h\n"
+ "fmin v14.8h, v14.8h, v21.8h\n"
+ "fmin v15.8h, v15.8h, v21.8h\n"
+ "fmin v16.8h, v16.8h, v21.8h\n"
+ "fmin v17.8h, v17.8h, v21.8h\n"
+ "fmin v18.8h, v18.8h, v21.8h\n"
+ "fmin v19.8h, v19.8h, v21.8h\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"
"132:" // Height 3: No activation
"cmp x14, #0x20\n"
"bge 149f\n"
@@ -2304,14 +2304,14 @@ void a64_ffhybrid_fp16_mla_6x32 (
"174:" // Height 4: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 175f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
"cbnz x28, 176f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20, LSL #1\n"
@@ -2321,9 +2321,9 @@ void a64_ffhybrid_fp16_mla_6x32 (
"b 176f\n"
"175:" // Height 4: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20, LSL #1\n"
- "add x24, x25, x20, LSL #1\n"
- "add x23, x24, x20, LSL #1\n"
+ "add x25, x26, x21, LSL #1\n"
+ "add x24, x25, x21, LSL #1\n"
+ "add x23, x24, x21, LSL #1\n"
"176:" // Height 4: input setup done
"cmp x27, #0x8\n"
"blt 179f\n"
@@ -2342,7 +2342,7 @@ void a64_ffhybrid_fp16_mla_6x32 (
"cmp x27, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q25, [x10, #0x0]\n"
"add x26, x26, #0x10\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
@@ -2350,164 +2350,164 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x24, x24, #0x10\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "ldr q24, [x9, #0x0]\n"
"add x23, x23, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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"
- "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
+ "fmla v10.8h, v25.8h, v0.h[0]\n"
+ "fmla v14.8h, v25.8h, v1.h[0]\n"
+ "fmla v18.8h, v25.8h, v2.h[0]\n"
+ "fmla v22.8h, v25.8h, v3.h[0]\n"
+ "ldr q25, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
"add x12, x12, #0x80\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, [x11, #0x70]\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, [x11, #0x70]\n"
"add x11, x11, #0x80\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, [x10, #0x70]\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, [x10, #0x70]\n"
"add x10, x10, #0x80\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, [x9, #0x70]\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, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\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 v10.8h, v25.8h, v0.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"
"ldr q6, [x12, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v24.8h, v0.h[7]\n"
"ldr q0, [x26, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
+ "fmla v15.8h, v24.8h, v1.h[7]\n"
"ldr q1, [x25, #0x0]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
+ "fmla v19.8h, v24.8h, v2.h[7]\n"
"ldr q2, [x24, #0x0]\n"
- "fmla v23.8h, v7.8h, v3.h[7]\n"
+ "fmla v23.8h, v24.8h, v3.h[7]\n"
"ldr q3, [x23, #0x0]\n"
"ldr q7, [x11, #0x0]\n"
"bge 177b\n"
@@ -2518,7 +2518,7 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x26, x26, #0x10\n"
"fmla v16.8h, v6.8h, v2.h[0]\n"
"fmla v20.8h, v6.8h, v3.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q25, [x10, #0x0]\n"
"add x25, x25, #0x10\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
@@ -2526,191 +2526,191 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x23, x23, #0x10\n"
"fmla v17.8h, v7.8h, v2.h[0]\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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"
- "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
+ "ldr q24, [x9, #0x0]\n"
+ "fmla v10.8h, v25.8h, v0.h[0]\n"
+ "fmla v14.8h, v25.8h, v1.h[0]\n"
+ "fmla v18.8h, v25.8h, v2.h[0]\n"
+ "fmla v22.8h, v25.8h, v3.h[0]\n"
+ "ldr q25, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
"add x12, x12, #0x80\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, [x11, #0x70]\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, [x11, #0x70]\n"
"add x11, x11, #0x80\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, [x10, #0x70]\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, [x10, #0x70]\n"
"add x10, x10, #0x80\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, [x9, #0x70]\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, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\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 v10.8h, v25.8h, v0.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"
"179:" // Height 4: Multiply loop: Main loop skip
"cbz x27, 181f\n"
"180:" // Height 4: Multiply loop: Odd block loop
- "ldr h0, [x26], #0x2\n"
- "ldr h1, [x25], #0x2\n"
+ "ldr h3, [x26], #0x2\n"
+ "ldr h2, [x25], #0x2\n"
"sub x27, x27, #0x1\n"
- "ldr h2, [x24], #0x2\n"
- "ldr h3, [x23], #0x2\n"
- "ldr q6, [x12, #0x0]\n"
- "ldr q7, [x11, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\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, [x10, #0x0]\n"
+ "ldr h1, [x24], #0x2\n"
+ "ldr h0, [x23], #0x2\n"
+ "ldr q25, [x12, #0x0]\n"
+ "ldr q24, [x11, #0x0]\n"
+ "fmla v8.8h, v25.8h, v3.h[0]\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, [x10, #0x0]\n"
"add x12, x12, #0x10\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
+ "fmla v9.8h, v24.8h, v3.h[0]\n"
+ "fmla v13.8h, v24.8h, v2.h[0]\n"
"add x11, x11, #0x10\n"
"add x10, x10, #0x10\n"
- "fmla v17.8h, v7.8h, v2.h[0]\n"
- "fmla v21.8h, v7.8h, v3.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "fmla v17.8h, v24.8h, v1.h[0]\n"
+ "fmla v21.8h, v24.8h, v0.h[0]\n"
+ "ldr q24, [x9, #0x0]\n"
"add x9, x9, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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 v10.8h, v25.8h, 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 x27, 180b\n"
"181:" // Height 4: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -2723,41 +2723,41 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x23, x24, x20, LSL #1\n"
"tbz %x[flags], #1, 182f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v1.8h }, [x20]\n"
+ "ld1r { v25.8h }, [x20]\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v1.8h\n"
- "fmin v9.8h, v9.8h, v1.8h\n"
- "fmin v10.8h, v10.8h, v1.8h\n"
- "fmin v11.8h, v11.8h, v1.8h\n"
- "fmin v12.8h, v12.8h, v1.8h\n"
- "fmin v13.8h, v13.8h, v1.8h\n"
- "fmin v14.8h, v14.8h, v1.8h\n"
- "fmin v15.8h, v15.8h, v1.8h\n"
- "fmin v16.8h, v16.8h, v1.8h\n"
- "fmin v17.8h, v17.8h, v1.8h\n"
- "fmin v18.8h, v18.8h, v1.8h\n"
- "fmin v19.8h, v19.8h, v1.8h\n"
- "fmin v20.8h, v20.8h, v1.8h\n"
- "fmin v21.8h, v21.8h, v1.8h\n"
- "fmin v22.8h, v22.8h, v1.8h\n"
- "fmin v23.8h, v23.8h, v1.8h\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"
+ "fmin v8.8h, v8.8h, v25.8h\n"
+ "fmin v9.8h, v9.8h, v25.8h\n"
+ "fmin v10.8h, v10.8h, v25.8h\n"
+ "fmin v11.8h, v11.8h, v25.8h\n"
+ "fmin v12.8h, v12.8h, v25.8h\n"
+ "fmin v13.8h, v13.8h, v25.8h\n"
+ "fmin v14.8h, v14.8h, v25.8h\n"
+ "fmin v15.8h, v15.8h, v25.8h\n"
+ "fmin v16.8h, v16.8h, v25.8h\n"
+ "fmin v17.8h, v17.8h, v25.8h\n"
+ "fmin v18.8h, v18.8h, v25.8h\n"
+ "fmin v19.8h, v19.8h, v25.8h\n"
+ "fmin v20.8h, v20.8h, v25.8h\n"
+ "fmin v21.8h, v21.8h, v25.8h\n"
+ "fmin v22.8h, v22.8h, v25.8h\n"
+ "fmin v23.8h, v23.8h, v25.8h\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"
"182:" // Height 4: No activation
"cmp x14, #0x20\n"
"bge 199f\n"
@@ -3316,15 +3316,15 @@ void a64_ffhybrid_fp16_mla_6x32 (
"224:" // Height 5: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 225f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
- "ldr x22, [x21, #0x20]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
"cbnz x28, 226f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20, LSL #1\n"
@@ -3335,10 +3335,10 @@ void a64_ffhybrid_fp16_mla_6x32 (
"b 226f\n"
"225:" // Height 5: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20, LSL #1\n"
- "add x24, x25, x20, LSL #1\n"
- "add x23, x24, x20, LSL #1\n"
- "add x22, x23, x20, LSL #1\n"
+ "add x25, x26, x21, LSL #1\n"
+ "add x24, x25, x21, LSL #1\n"
+ "add x23, x24, x21, LSL #1\n"
+ "add x22, x23, x21, LSL #1\n"
"226:" // Height 5: input setup done
"cmp x27, #0x8\n"
"blt 229f\n"
@@ -3361,7 +3361,7 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x26, x26, #0x10\n"
"add x25, x25, #0x10\n"
"fmla v24.8h, v6.8h, v4.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q29, [x10, #0x0]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"add x24, x24, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
@@ -3370,194 +3370,194 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x22, x22, #0x10\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
"fmla v25.8h, v7.8h, v4.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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"
- "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
+ "ldr q28, [x9, #0x0]\n"
+ "fmla v10.8h, v29.8h, v0.h[0]\n"
+ "fmla v14.8h, v29.8h, v1.h[0]\n"
+ "fmla v18.8h, v29.8h, v2.h[0]\n"
+ "fmla v22.8h, v29.8h, v3.h[0]\n"
+ "fmla v26.8h, v29.8h, v4.h[0]\n"
+ "ldr q29, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
+ "fmla v11.8h, v28.8h, v0.h[6]\n"
"add x12, x12, #0x80\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, [x11, #0x70]\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, [x11, #0x70]\n"
"add x11, x11, #0x80\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, [x10, #0x70]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\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, [x10, #0x70]\n"
+ "fmla v9.8h, v28.8h, v0.h[7]\n"
"add x10, x10, #0x80\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, [x9, #0x70]\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, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\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 v10.8h, v29.8h, v0.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"
"ldr q6, [x12, #0x0]\n"
- "fmla v11.8h, v7.8h, v0.h[7]\n"
+ "fmla v11.8h, v28.8h, v0.h[7]\n"
"ldr q0, [x26, #0x0]\n"
- "fmla v15.8h, v7.8h, v1.h[7]\n"
+ "fmla v15.8h, v28.8h, v1.h[7]\n"
"ldr q1, [x25, #0x0]\n"
- "fmla v19.8h, v7.8h, v2.h[7]\n"
+ "fmla v19.8h, v28.8h, v2.h[7]\n"
"ldr q2, [x24, #0x0]\n"
- "fmla v23.8h, v7.8h, v3.h[7]\n"
+ "fmla v23.8h, v28.8h, v3.h[7]\n"
"ldr q3, [x23, #0x0]\n"
- "fmla v27.8h, v7.8h, v4.h[7]\n"
+ "fmla v27.8h, v28.8h, v4.h[7]\n"
"ldr q4, [x22, #0x0]\n"
"ldr q7, [x11, #0x0]\n"
"bge 227b\n"
@@ -3571,7 +3571,7 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x25, x25, #0x10\n"
"add x24, x24, #0x10\n"
"fmla v24.8h, v6.8h, v4.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "ldr q29, [x10, #0x0]\n"
"fmla v9.8h, v7.8h, v0.h[0]\n"
"add x23, x23, #0x10\n"
"fmla v13.8h, v7.8h, v1.h[0]\n"
@@ -3579,226 +3579,226 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x22, x22, #0x10\n"
"fmla v21.8h, v7.8h, v3.h[0]\n"
"fmla v25.8h, v7.8h, v4.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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"
- "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
- "fmla v11.8h, v7.8h, v0.h[6]\n"
+ "ldr q28, [x9, #0x0]\n"
+ "fmla v10.8h, v29.8h, v0.h[0]\n"
+ "fmla v14.8h, v29.8h, v1.h[0]\n"
+ "fmla v18.8h, v29.8h, v2.h[0]\n"
+ "fmla v22.8h, v29.8h, v3.h[0]\n"
+ "fmla v26.8h, v29.8h, v4.h[0]\n"
+ "ldr q29, [x12, #0x10]\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, [x11, #0x10]\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, [x10, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, [x10, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\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, [x11, #0x30]\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, [x10, #0x30]\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, [x9, #0x30]\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, [x12, #0x40]\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, [x11, #0x40]\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, [x10, #0x40]\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, [x9, #0x40]\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, [x12, #0x50]\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, [x11, #0x50]\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, [x10, #0x50]\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, [x9, #0x50]\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, [x12, #0x60]\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, [x11, #0x60]\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, [x10, #0x60]\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, [x9, #0x60]\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, [x12, #0x70]\n"
+ "fmla v11.8h, v28.8h, v0.h[6]\n"
"add x12, x12, #0x80\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, [x11, #0x70]\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, [x11, #0x70]\n"
"add x11, x11, #0x80\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, [x10, #0x70]\n"
- "fmla v9.8h, v7.8h, v0.h[7]\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, [x10, #0x70]\n"
+ "fmla v9.8h, v28.8h, v0.h[7]\n"
"add x10, x10, #0x80\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, [x9, #0x70]\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, [x9, #0x70]\n"
"add x9, x9, #0x80\n"
- "fmla v10.8h, v6.8h, v0.h[7]\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 v10.8h, v29.8h, v0.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"
"229:" // Height 5: Multiply loop: Main loop skip
"cbz x27, 231f\n"
"230:" // Height 5: Multiply loop: Odd block loop
- "ldr h0, [x26], #0x2\n"
- "ldr h1, [x25], #0x2\n"
+ "ldr h4, [x26], #0x2\n"
+ "ldr h3, [x25], #0x2\n"
"sub x27, x27, #0x1\n"
"ldr h2, [x24], #0x2\n"
- "ldr h3, [x23], #0x2\n"
- "ldr h4, [x22], #0x2\n"
- "ldr q6, [x12, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\n"
- "fmla v12.8h, v6.8h, v1.h[0]\n"
- "ldr q7, [x11, #0x0]\n"
- "fmla v16.8h, v6.8h, v2.h[0]\n"
- "fmla v20.8h, v6.8h, v3.h[0]\n"
+ "ldr h1, [x23], #0x2\n"
+ "ldr h0, [x22], #0x2\n"
+ "ldr q29, [x12, #0x0]\n"
+ "fmla v8.8h, v29.8h, v4.h[0]\n"
+ "fmla v12.8h, v29.8h, v3.h[0]\n"
+ "ldr q28, [x11, #0x0]\n"
+ "fmla v16.8h, v29.8h, v2.h[0]\n"
+ "fmla v20.8h, v29.8h, v1.h[0]\n"
"add x12, x12, #0x10\n"
- "fmla v24.8h, v6.8h, v4.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
- "fmla v9.8h, v7.8h, v0.h[0]\n"
+ "fmla v24.8h, v29.8h, v0.h[0]\n"
+ "ldr q29, [x10, #0x0]\n"
+ "fmla v9.8h, v28.8h, v4.h[0]\n"
"add x11, x11, #0x10\n"
- "fmla v13.8h, v7.8h, v1.h[0]\n"
- "fmla v17.8h, v7.8h, v2.h[0]\n"
+ "fmla v13.8h, v28.8h, v3.h[0]\n"
+ "fmla v17.8h, v28.8h, v2.h[0]\n"
"add x10, x10, #0x10\n"
- "fmla v21.8h, v7.8h, v3.h[0]\n"
- "fmla v25.8h, v7.8h, v4.h[0]\n"
- "ldr q7, [x9, #0x0]\n"
+ "fmla v21.8h, v28.8h, v1.h[0]\n"
+ "fmla v25.8h, v28.8h, v0.h[0]\n"
+ "ldr q28, [x9, #0x0]\n"
"add x9, x9, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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 v10.8h, v29.8h, 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 x27, 230b\n"
"231:" // Height 5: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -3812,49 +3812,49 @@ void a64_ffhybrid_fp16_mla_6x32 (
"add x22, x23, x20, LSL #1\n"
"tbz %x[flags], #1, 232f\n"
"add x20, %x[args_ptr], %[offset_max]\n"
- "ld1r { v1.8h }, [x20]\n"
+ "ld1r { v29.8h }, [x20]\n"
"add x20, %x[args_ptr], %[offset_min]\n"
- "ld1r { v0.8h }, [x20]\n"
- "fmin v8.8h, v8.8h, v1.8h\n"
- "fmin v9.8h, v9.8h, v1.8h\n"
- "fmin v10.8h, v10.8h, v1.8h\n"
- "fmin v11.8h, v11.8h, v1.8h\n"
- "fmin v12.8h, v12.8h, v1.8h\n"
- "fmin v13.8h, v13.8h, v1.8h\n"
- "fmin v14.8h, v14.8h, v1.8h\n"
- "fmin v15.8h, v15.8h, v1.8h\n"
- "fmin v16.8h, v16.8h, v1.8h\n"
- "fmin v17.8h, v17.8h, v1.8h\n"
- "fmin v18.8h, v18.8h, v1.8h\n"
- "fmin v19.8h, v19.8h, v1.8h\n"
- "fmin v20.8h, v20.8h, v1.8h\n"
- "fmin v21.8h, v21.8h, v1.8h\n"
- "fmin v22.8h, v22.8h, v1.8h\n"
- "fmin v23.8h, v23.8h, v1.8h\n"
- "fmin v24.8h, v24.8h, v1.8h\n"
- "fmin v25.8h, v25.8h, v1.8h\n"
- "fmin v26.8h, v26.8h, v1.8h\n"
- "fmin v27.8h, v27.8h, v1.8h\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"
+ "fmin v8.8h, v8.8h, v29.8h\n"
+ "fmin v9.8h, v9.8h, v29.8h\n"
+ "fmin v10.8h, v10.8h, v29.8h\n"
+ "fmin v11.8h, v11.8h, v29.8h\n"
+ "fmin v12.8h, v12.8h, v29.8h\n"
+ "fmin v13.8h, v13.8h, v29.8h\n"
+ "fmin v14.8h, v14.8h, v29.8h\n"
+ "fmin v15.8h, v15.8h, v29.8h\n"
+ "fmin v16.8h, v16.8h, v29.8h\n"
+ "fmin v17.8h, v17.8h, v29.8h\n"
+ "fmin v18.8h, v18.8h, v29.8h\n"
+ "fmin v19.8h, v19.8h, v29.8h\n"
+ "fmin v20.8h, v20.8h, v29.8h\n"
+ "fmin v21.8h, v21.8h, v29.8h\n"
+ "fmin v22.8h, v22.8h, v29.8h\n"
+ "fmin v23.8h, v23.8h, v29.8h\n"
+ "fmin v24.8h, v24.8h, v29.8h\n"
+ "fmin v25.8h, v25.8h, v29.8h\n"
+ "fmin v26.8h, v26.8h, v29.8h\n"
+ "fmin v27.8h, v27.8h, v29.8h\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"
"232:" // Height 5: No activation
"cmp x14, #0x20\n"
"bge 249f\n"
@@ -4497,16 +4497,16 @@ void a64_ffhybrid_fp16_mla_6x32 (
"274:" // Height 6: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 275f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
- "ldr x22, [x21, #0x20]\n"
- "ldr x21, [x21, #0x28]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
+ "ldr x21, [x20, #0x28]\n"
"cbnz x28, 276f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20, LSL #1\n"
@@ -4518,11 +4518,11 @@ void a64_ffhybrid_fp16_mla_6x32 (
"b 276f\n"
"275:" // Height 6: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20, LSL #1\n"
- "add x24, x25, x20, LSL #1\n"
- "add x23, x24, x20, LSL #1\n"
- "add x22, x23, x20, LSL #1\n"
- "add x21, x22, x20, LSL #1\n"
+ "add x25, x26, x21, LSL #1\n"
+ "add x24, x25, x21, LSL #1\n"
+ "add x23, x24, x21, LSL #1\n"
+ "add x22, x23, x21, LSL #1\n"
+ "add x21, x22, x21, LSL #1\n"
"276:" // Height 6: input setup done
"cmp x27, #0x8\n"
"blt 279f\n"
@@ -5017,45 +5017,45 @@ void a64_ffhybrid_fp16_mla_6x32 (
"279:" // Height 6: Multiply loop: Main loop skip
"cbz x27, 281f\n"
"280:" // Height 6: Multiply loop: Odd block loop
- "ldr h0, [x26], #0x2\n"
- "ldr h1, [x25], #0x2\n"
+ "ldr h7, [x26], #0x2\n"
+ "ldr h6, [x25], #0x2\n"
"sub x27, x27, #0x1\n"
- "ldr h2, [x24], #0x2\n"
- "ldr h3, [x23], #0x2\n"
- "ldr h4, [x22], #0x2\n"
- "ldr h5, [x21], #0x2\n"
- "ldr q6, [x12, #0x0]\n"
- "ldr q7, [x11, #0x0]\n"
- "fmla v8.8h, v6.8h, v0.h[0]\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 h5, [x24], #0x2\n"
+ "ldr h4, [x23], #0x2\n"
+ "ldr h3, [x22], #0x2\n"
+ "ldr h2, [x21], #0x2\n"
+ "ldr q1, [x12, #0x0]\n"
+ "ldr q0, [x11, #0x0]\n"
+ "fmla v8.8h, v1.8h, v7.h[0]\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"
"add x12, x12, #0x10\n"
"add x11, x11, #0x10\n"
- "fmla v24.8h, v6.8h, v4.h[0]\n"
- "fmla v28.8h, v6.8h, v5.h[0]\n"
- "ldr q6, [x10, #0x0]\n"
+ "fmla v24.8h, v1.8h, v3.h[0]\n"
+ "fmla v28.8h, v1.8h, v2.h[0]\n"
+ "ldr q1, [x10, #0x0]\n"
"add x10, x10, #0x10\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, [x9, #0x0]\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, [x9, #0x0]\n"
"add x9, x9, #0x10\n"
- "fmla v10.8h, v6.8h, v0.h[0]\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 v10.8h, v1.8h, v7.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 x27, 280b\n"
"281:" // Height 6: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"