[libc][math][c23] Optimize fabsf16 on x86 with Clang by overmighty · Pull Request #104869 · llvm/llvm-project (original) (raw)
Minimum latencies:
Before:
- Clang 18, libgcc 14, Intel Core i7-13700H:
fabsf16 from 0x3c00 to 0x4000: 8.01073 ns
fabsf16 from 0 to 0x400: 66.8712 ns Disassembly:
0000000000002290 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2290: 55 push rbp
2291: 48 89 e5 mov rbp,rsp
2294: e8 07 01 00 00 call 23a0 <__extendhfsf2@plt>
2299: c4 e2 79 18 0d 92 e7 vbroadcastss xmm1,DWORD PTR [rip+0xffffffffffffe792] # a34 <_IO_stdin_used+0x4>
22a0: ff ff
22a2: c5 f8 54 c1 vandps xmm0,xmm0,xmm1
22a6: e8 05 01 00 00 call 23b0 <__truncsfhf2@plt>
22ab: 5d pop rbp
22ac: c3 ret
- Clang 18, libgcc 14, Intel Core i7-13700H,
-march=native:
fabsf16 from 0x3c00 to 0x4000: 6.66634 ns
fabsf16 from 0 to 0x400: 6.70341 ns Disassembly:
0000000000002210 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2210: 55 push rbp
2211: 48 89 e5 mov rbp,rsp
2214: c5 f9 c5 c0 00 vpextrw eax,xmm0,0x0
2219: 0f b7 c0 movzx eax,ax
221c: c5 f9 6e c0 vmovd xmm0,eax
2220: c4 e2 79 13 c0 vcvtph2ps xmm0,xmm0
2225: c4 e2 79 18 0d 76 e7 vbroadcastss xmm1,DWORD PTR [rip+0xffffffffffffe776] # 9a4 <_IO_stdin_used+0x4>
222c: ff ff
222e: c5 f8 54 c1 vandps xmm0,xmm0,xmm1
2232: c4 e3 79 1d c0 04 vcvtps2ph xmm0,xmm0,0x4
2238: c5 f9 7e c0 vmovd eax,xmm0
223c: c5 f9 c4 c0 00 vpinsrw xmm0,xmm0,eax,0x0
2241: 5d pop rbp
2242: c3 ret
After:
- Clang 18, libgcc 14, Intel Core i7-13700H:
fabsf16 from 0x3c00 to 0x4000: 2.24488 ns
fabsf16 from 0 to 0x400: 2.23415 ns Disassembly:
0000000000002200 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2200: 55 push rbp
2201: 48 89 e5 mov rbp,rsp
2204: c5 f9 c5 c0 00 vpextrw eax,xmm0,0x0
2209: 66 23 05 ee e7 ff ff and ax,WORD PTR [rip+0xffffffffffffe7ee] # 9fe <_ZZN22__llvm_libc_20_0_0_git6fp>
2210: c5 f9 c4 c0 00 vpinsrw xmm0,xmm0,eax,0x0
2215: 5d pop rbp
2216: c3 ret
- Clang 18, libgcc 14, Intel Core i7-13700H,
-march=native:
fabsf16 from 0x3c00 to 0x4000: 2.2878 ns
fabsf16 from 0 to 0x400: 2.26927 ns Disassembly:
0000000000002210 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2210: 55 push rbp
2211: 48 89 e5 mov rbp,rsp
2214: c5 f9 c5 c0 00 vpextrw eax,xmm0,0x0
2219: 66 23 05 de e7 ff ff and ax,WORD PTR [rip+0xffffffffffffe7de] # 9fe <_ZZN22__llvm_libc_20_0_0_git6fp>
2220: c5 f9 c4 c0 00 vpinsrw xmm0,xmm0,eax,0x0
2225: 5d pop rbp
2226: c3 ret
If we enabled the same code path for other compilers/targets:
Before:
- GCC 14, Intel Core i7-13700H:
fabsf16 from 0x3c00 to 0x4000: 1.05366 ns
fabsf16 from 0 to 0x400: 1.05366 ns Disassembly:
0000000000002200 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2200: c5 f9 7e c0 vmovd eax,xmm0
2204: 66 25 ff 7f and ax,0x7fff
2208: c5 f9 6e c0 vmovd xmm0,eax
220c: c3 ret
- GCC 14, Intel Core i7-13700H,
-march=native:
fabsf16 from 0x3c00 to 0x4000: 1.05268 ns
fabsf16 from 0 to 0x400: 1.05171 ns Disassembly:
0000000000002210 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2210: c5 f9 7e c0 vmovd eax,xmm0
2214: 66 25 ff 7f and ax,0x7fff
2218: c5 f9 6e c0 vmovd xmm0,eax
221c: c3 ret
- Clang 18, Google Tensor G3,
-Xclang -target-feature -Xclang +fullfp16:
fabsf16 from 0x3c00 to 0x4000: 0.873171 ns
fabsf16 from 0 to 0x400: 0.675122 ns Disassembly:
0000000000024cdc <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
24cdc: 1ee0c000 fabs h0, h0
24ce0: d65f03c0 ret
After:
- GCC 14, Intel Core i7-13700H:
fabsf16 from 0x3c00 to 0x4000: 1.05366 ns
fabsf16 from 0 to 0x400: 1.05366 ns Disassembly:
0000000000002210 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2210: 0f b7 05 77 e8 ff ff movzx eax,WORD PTR [rip+0xffffffffffffe877] # a8e <__llvm_libc_20_0_0_git::fputil>
2217: c5 f9 7e c2 vmovd edx,xmm0
221b: 21 c2 and edx,eax
221d: c5 f9 6e c2 vmovd xmm0,edx
2221: c3 ret
- GCC 14, Intel Core i7-13700H,
-march=native:
fabsf16 from 0x3c00 to 0x4000: 1.05366 ns
fabsf16 from 0 to 0x400: 1.05366 ns Disassembly:
0000000000002220 <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
2220: 0f b7 05 67 e8 ff ff movzx eax,WORD PTR [rip+0xffffffffffffe867] # a8e <__llvm_libc_20_0_0_git::fputil>
2227: c5 f9 7e c2 vmovd edx,xmm0
222b: 21 c2 and edx,eax
222d: c5 f9 6e c2 vmovd xmm0,edx
2231: c3 ret
- Clang 18, Google Tensor G3,
-Xclang -target-feature -Xclang +fullfp16:
fabsf16 from 0x3c00 to 0x4000: 0.873171 ns
fabsf16 from 0 to 0x400: 0.674146 ns Disassembly:
0000000000024cdc <__llvm_libc_20_0_0_git::fabsf16(_Float16)>:
24cdc: d0ffff68 adrp x8, 0x12000 <wmemchr+0x12000>
24ce0: 1e260009 fmov w9, s0
24ce4: 794ad508 ldrh w8, [x8, #0x56a]
24ce8: 0a090108 and w8, w8, w9
24cec: 1e270100 fmov s0, w8
24cf0: d65f03c0 ret