Packing BCD to DPD: How to improve this amd64 assembly routine?

TYVM for commenting the code clearly and well, BTW. It made is super easy to figure out what was going on, and where the bits were going. I'd never heard of DPD before, so puzzling it out from uncommented code and the wikipedia article would have sucked.


The relevant gotchas are:

  • Avoid 16bit operand size for instructions with immediate constants, on Intel CPUs. (LCP stalls)
  • avoid reading the full 32 or 64bit register after writing only the low 8 or 16, on Intel pre-IvyBridge. (partial-register extra uop). (IvB still has that slowdown if you modify an upper8 reg like AH, but Haswell removes that too). It's not just an extra uop: the penalty on Core2 is 2 to 3 cycles, according to Agner Fog. I might be measuring it wrong, but it seems a lot less bad on SnB.

See http://agner.org/optimize/ for full details.

Other than that, there's no general problem with mixing in some instructions using the operand-size prefix to make them 16-bit.


You should maybe write this as inline asm, rather than as a called function. You only use a couple registers, and the fast-path case is very few instructions.


I had a look at the code. I didn't look into achieving the same result with significantly different logic, just at optimizing the logic you do have.


Possible code suggestions: Switch the branching so the fast-path has the not-taken branch. Actually, it might make no diff either way in this case, or might improve the alignment of the slow-path code.

.p2align 4,,10   # align to 16, unless we're already in the first 6 bytes of a block of 16
bcd2dpd_mul:
        mov %edi,%eax           #   = 0000 abcd efgh iklm
        shl %al                 #   = 0000 abcd fghi klm0
        shr %eax                #   = 0000 0abc dfgh iklm
        test $0x880,%edi        # fast path for a = e = 0
        jnz .Lslow_path
        ret

.p2align 4    # Maybe fine-tune this alignment based on how the rest of the code assembles.    
.Lslow_path:

        ...
        ret

It's sometimes better to duplicate return instructions than to absolutely minimize code-size. The compare-and-branch in this case is the 4th uop of the function, though, so a taken branch wouldn't have prevented 4 uops from issuing in the first clock cycle, and a correctly-predicted branch would still issue the return on the 2nd clock cycle.


You should use a 32bit imul for the one with the table source. (see next section about aligning the table so reading an extra 2B is ok). 32bit imul is one uop instead of two on Intel SnB-family microarches. The result in the low16 should be the same, since the sign bit can't be set. The upper16 gets zeroed by the final and before ret, and doesn't get used in any way where garbage in the upper16 matters while it's there.

Your imul with an immediate operand is problematic, though.

It causes an LCP stall when decoding on Intel, and it writes the the low16 of a register that is later read at full width. Its upper16 would be a problem if not masked off (since it's used as a table index). Its operands are large enough that they will put garbage into the upper16, so it does need to be discarded.

I thought your way of doing it would be optimal for some architectures, but it turns out imul r16,r16,imm16 itself is slower than imul r32,r32,imm32 on every architecture except VIA Nano, AMD K7 (where it's faster than imul32), and Intel P6 (where using it from 32bit / 64bit mode will LCP-stall, and where partial-reg slowdowns are a problem).

On Intel SnB-family CPUs, where imul r16,r16,imm16 is two uops, imul32/movzx would be strictly better, with no downside except code size. On P6-family CPUs (i.e. PPro to Nehalem), imul r16,r16,imm16 is one uop, but those CPUs don't have a uop cache, so the LCP stall is probably critical (except maybe Nehalem calling this in a tight loop, fitting in the 28 uop loop buffer). And for those CPUs, the explicit movzx is probably better from the perspective of the partial-reg stall. Agner Fog says something about there being an extra cycle while the CPU inserts the merging uop, which might mean a cycle where that extra uop is issued alone.

On AMD K8-Steamroller, imul imm16 is 2 m-ops instead of 1 for imul imm32, so imul32/movzx is about equal to imul16 there. They don't suffer from LCP stalls, or from partial-reg problems.

On Intel Silvermont, imul imm16 is 2 uops (with one per 4 clocks throughput), vs. imul imm32 being 1 uops (with one per 1 clock throughput). Same thing on Atom (the in-order predecessor to Silvermont): imul16 is an extra uop and much slower. On most other microarchitectures, throughput isn't worse, just latency.

So if you're willing to increase the code-size in bytes where it will give a speedup, you should use a 32bit imul and a movzwl %di, %edi. On some architectures, this will be about the same speed as the imul imm16, while on others it will be much faster. It might be slightly worse on AMD bulldozer-family, which isn't very good at using both integer execution units at once, apparently, so a 2 m-op instruction for EX1 might be better than two 1 m-op instructions where one of them is still an EX1-only instruction. Benchmark this if you care.


Align tab to at least a 32B boundary, so your 32bit imul and or can do a 4B load from any 2B-aligned entry in it without crossing a cache-line boundary. Unaligned accesses have no penalty on all recent CPUs (Nehalem and later, and recent AMD), as long as they don't span two cache lines.

Making the operations that read from the table 32bit avoids the partial-register penalty that Intel CPUs have. AMD CPUs, and Silvermont, don't track partial-registers separately, so even instructions that write-only to the low16 have to wait for the result in the rest of the reg. This stops 16bit insns from breaking dependency chains. Intel P6 and SnB microarch families track partial regs. Haswell does full dual bookkeeping or something, because there's no penalty when merging is needed, like after you shift al, then shift eax. SnB will insert an extra uop there, and there may be a penalty of a cycle or two while it does this. I'm not sure, and haven't tested. However, I don't see a nice way to avoid this.

The shl %al could be replaced with a add %al, %al. That can run on more ports. Probably no difference, since port0/5 (or port0/6 on Haswell and later) probably aren't saturated. They have the same effect on the bits, but set flags differently. Otherwise they could be decoded to the same uop.


changes: split the pext/pdep / vectorize version into a separate answer, partly so it can have its own comment thread.


(I split the BMI2 version into a separate answer, since it could end up totally different)


After seeing what you're doing with that imul/shr to get a table index, I can see where you could use BMI2 pextr to replace and/imul/shr, or BMI1 bextr to replace just the shr (allowing use of imul32, instead of imul16, since you'd just extract the bits you want, rather than needing to shift zeros from the upper16). There are AMD CPUs with BMI1, but even steamroller lacks BMI2. Intel introduced BMI1 and BMI2 at the same time with Haswell.

You could maybe process two or four 16bit words at once, with 64bit pextr. But not for the whole algorithm: you can't do 4 parallel table lookups. (AVX2 VPGATHERDD is not worth using here.) Actually, you can use pshufb to implement a LUT with indices up to 4bits, see below.

Minor improvement version:

.section .rodata
  # This won't won't assemble, written this way for humans to line up with comments.

extmask_lobits:     .long           0b0000 0111 0111 0111
extmask_hibits:     .long           0b0000 1000 1000 1000

# pext doesn't have an immediate-operand form, but it can take the mask from a memory operand.
# Load these into regs if running in a tight loop.

#### TOTALLY UNTESTED #####
.text
.p2align 4,,10
bcd2dpd_bmi2:
#       mov   %edi,%eax           #   = 0000 abcd efgh iklm
#       shl   %al                 #   = 0000 abcd fghi klm0
#       shr   %eax                #   = 0000 0abc dfgh iklm

        pext  extmask_lobits, %edi, %eax
                                #   = 0000 0abc dfgh iklm
        mov   %eax, %esi        # insn scheduling for 4-issue front-end: Fast-path is 4 fused-domain uops
          # And doesn't waste issue capacity when we're taking the slow path.  CPUs with mov-elimination won't waste execution units from issuing an extra mov
        test  $0x880, %edi        # fast path for a = e = 0
        jnz .Lslow_path
        ret

.p2align 4
.Lslow_path:
        # 8 uops, including the `ret`: can issue in 2 clocks.

        # replaces and/imul/shr
        pext  extmask_hibits, %edi, %edi #u= 0000 0000 0000 0aei
        and   $0x66, %esi                # q = 0000 0000 0fg0 0kl0
        imul  tab-8(,%rdi,4), %esi       # v = q * tab[u-2][0]
        and   $0x397, %eax               # r = 0000 00bc d00h 0klm
        xor   %esi, %eax                 # w = r ^ v
        or    tab-6(,%rdi,4), %eax       # x = w | tab[u-2][1]
        and   $0x3ff, %eax               #   = 0000 00xx xxxx xxxx
        ret

Of course, if making this an inline-asm, rather than a stand-alone function, you'd change back to the fast path branching to the end, and the slow-path falling through. And you wouldn't waste space with alignment padding mid-function either.

There might be more scope for using pextr and/or pdep for more of the rest of the function.


I was thinking about how to do even better with BMI2. I think we could get multiple aei selectors from four shorts packed into 64b, then use pdep to deposit them in the low bits of different bytes. Then movq that to a vector register, where you use it as a shuffle control-mask for pshufb to do multiple 4bit LUT lookups.

So we could go from 60 BCD bits to 50 DPD bits at a time. (Use shrd to shift bits between registers to handle loads/stores to byte-addressable memory.)

Actually, 48 BCD bits (4 groups of 12bits each) -> 40 DPD bits is probably a lot easier, because you can unpack that to 4 groups of 16bits in a 64b integer register, using pdep. Dealing with the selectors for 5 groups is fine, you can unpack with pmovzx, but dealing with the rest of the data would require bit-shuffling in vector registers. Not even the slow AVX2 variable-shift insns would make that easy to do. (Although it might be interesting to consider how to implement this with BMI2 at all, for big speedups on CPUs with just SSSE3 (i.e. every relevant CPU) or maybe SSE4.1.)

This also means we can put two clusters of 4 groups into the low and high halves of a 128b register, to get even more parallelism.

As a bonus, 48bits is a whole number of bytes, so reading from a buffer of BCD digits wouldn't require any shrd insns to get the leftover 4 bits from the last 64b into the low 4 for the next. Or two offset pextr masks to work when the 4 ignored bits were the low or high 4 of the 64b.... Anyway, I think doing 5 groups at once just isn't worth considering.

Full BMI2 / AVX pshufb LUT version (vectorizable)

The data movement could be:

ignored | group 3        | group 2        | group 1        |  group 0
16bits  | abcd efgh iklm | abcd efgh iklm | abcd efgh iklm | abcd efgh iklm

         3   2   1 | 0
pext -> aei|aei|aei|aei  # packed together in the low bits

          2  |      1            |        0
pdep ->  ... |0000 0000 0000 0aei|0000 0000 0000 0aei  # each in a separate 16b word

movq -> xmm vector register.
 (Then pinsrq another group of 4 selectors into the upper 64b of the vector reg).  So the vector part can handle 2 (or AVX2: 4) of this at once

vpshufb xmm2 -> map each byte to another byte (IMUL table)
vpshufb xmm3 -> map each byte to another byte (OR table)


Get the bits other than `aei` from each group of 3 BCD digits unpacked from 48b to 64b, into separate 16b words:

                  group 3       | group 2             | group 1             |  group 0
pdep(src)-> 0000 abcd efgh iklm | 0000 abcd efgh iklm | 0000 abcd efgh iklm | 0000 abcd efgh iklm

 movq this into a vector reg (xmm1).  (And repeat for the next 48b and pinsrq that to the upper64)

VPAND  xmm1, mask  (to zero aei in each group)

 Then use the vector-LUT results:
VPMULLW xmm1, xmm2 -> packed 16b multiply, keeping only the low16 of the result

VPAND   xmm1,  mask
VPXOR   xmm1, something
VPOR    xmm1, xmm3

movq / pextrq back to integer regs

pext to pack the bits back together
  You don't need the AND 0x3ff or equivalent:
  Those bits go away when you pext to pack each 16b down to 10b

shrd or something to pack the 40b results of this into 64b chunks for store to memory.
  Or: 32b store, then shift and store the last 8b, but that seems lame
  Or: just do 64b stores, overlapping with the previous.  So you write 24b of garbage every time.  Take care at the very end of the buffer.

Use AVX 3-operand versions of the 128b SSE instructions to avoid needing movdqa to not overwrite the table for pshufb. As long as you never run a 256b AVX instruction, you don't need to mess with vzeroupper. You might as well use the v (VEX) versions of all vector instructions, though, if you use any. Inside a VM, you might be running on a virtual CPU with BMI2 but not AVX support, so it's prob. still a good idea to check both CPU feature flags, rather than assuming AVX if you see BMI2 (even though that's safe for all physical hardware that currently exists).


This is starting to look really efficient. It might be worth doing the mul/xor/and stuff in vector regs, even if you don't have BMI2 pext/pdep to do the bit packing/unpacking. I guess you could use code like the existing non-BMI scalar routing to get selectors, and mask/shift/or could build up the non-selector data into 16b chunks. Or maybe shrd for shifting data from one reg into another?