|
- #ifdef __aarch64__
-
- .text
- .align 5
- .global Float16ToFloat32
- #ifndef __APPLE__
- .type Float16ToFloat32, %function
- #endif
-
- // void Float16ToFloat32(const float16_t *input, float *output, int number);
- // x0: input, x1: output, x2: number
- Float16ToFloat32:
- // registers v8 ~ v15 must be preserved by a callee across subroutine calls, according to
- // https://github.com/ARM-software/abi-aa/blob/master/aapcs64/aapcs64.rst#simd-and-floating-point-registers
- // x19 ~ x29 should be also preserved
- // whereas our coding style do not permit such amount of parameters
- cmp x2, #64
- blt Loop
- Loop64:
- ld1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x0], #64
- ld1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x0], #64
- fcvtl v16.4s, v0.4h
- fcvtl2 v17.4s, v0.8h
- fcvtl v18.4s, v1.4h
- fcvtl2 v19.4s, v1.8h
- fcvtl v20.4s, v2.4h
- fcvtl2 v21.4s, v2.8h
- fcvtl v22.4s, v3.4h
- fcvtl2 v23.4s, v3.8h
- fcvtl v24.4s, v4.4h
- fcvtl2 v25.4s, v4.8h
- fcvtl v26.4s, v5.4h
- fcvtl2 v27.4s, v5.8h
- fcvtl v28.4s, v6.4h
- fcvtl2 v29.4s, v6.8h
- fcvtl v30.4s, v7.4h
- fcvtl2 v31.4s, v7.8h
- st1 {v16.4s, v17.4s, v18.4s, v19.4s}, [x1], #64
- st1 {v20.4s, v21.4s, v22.4s, v23.4s}, [x1], #64
- st1 {v24.4s, v25.4s, v26.4s, v27.4s}, [x1], #64
- st1 {v28.4s, v29.4s, v30.4s, v31.4s}, [x1], #64
- subs x2, x2, #64
- ble LoopEnd
- cmp x2, #64
- bge Loop64
- Loop:
- ldr h0, [x0], #2
- fcvt s0, h0
- str s0, [x1], #4
- subs x2, x2, #1
- bgt Loop
- LoopEnd:
- ret
- #endif
|
