Use new checksum function by default

The change required to support third argument 'initial', return value as
host-byte order value and adapt some selftests.

src/lib/newchecksum.dasl → src/arch/checksum.dasl

@@ -35,10 +35,11 @@ local function gen_checksum ()
       | push rbp
       | mov rbp, rsp
       -- Accumulative sum.
-      | xor rax, rax                -- Clear out rax. Stores accumulated sum.
+      | mov rax, rdx                -- Dx (3rd argument: initial).
+      | xchg al, ah                 -- Swap to convert to host-bytes order.
+      | mov rcx, rsi                -- Rsi (2nd argument; size).
       | xor r9, r9                  -- Clear out r9. Stores value of array.
       | xor r8, r8                  -- Clear out r8. Stores array index.
-      | mov rcx, rsi                -- Rsi (2nd argument; size). Assign rsi to rcx.
       | 1:
       | cmp rcx, 32                 -- If index is less than 16.
       | jl >2                       -- Jump to branch '2'.
@@ -99,6 +100,8 @@ local function gen_checksum ()
       -- One's complement.
       | not rax                     -- One-complement of rax.
       | and rax, 0xffff             -- Clear out higher part of rax.
+      -- Swap.
+      | xchg al, ah
       -- Epilogue.
       | mov rsp, rbp
       | pop rbp
@@ -107,8 +110,6 @@ local function gen_checksum ()
    end
 end
 
-local newchecksum = assemble("newchecksum", "uint32_t(*)(uint8_t*, uint32_t)", gen_checksum())
-
 -- Reference implementation in Lua.
 local function checksum_lua (data, size)
    local function r16 (data)
@@ -132,12 +133,10 @@ local function checksum_lua (data, size)
    return bit.band(bit.bnot(csum), 0xffff)
 end
 
+checksum = assemble("checksum", "uint32_t(*)(uint8_t*, uint32_t, uint16_t)", gen_checksum())
+
 function selftest ()
    require("lib.checksum_h")
-   local cpuinfo = lib.readfile("/proc/cpuinfo", "*a")
-   assert(cpuinfo, "failed to read /proc/cpuinfo for hardware check")
-   local have_avx2 = cpuinfo:match("avx2")
-   local have_sse2 = cpuinfo:match("sse2")
    local function create_packet (size)
       local pkt = {
          data = ffi.new("uint8_t[?]", size),
@@ -165,14 +164,8 @@ function selftest ()
       local pkt = create_packet(size)
       print(mpps.."M; "..size.." bytes")
       -- Benchmark for different architectures.
-      print("Gen: ", benchmark(function() return C.cksum_generic(pkt.data, pkt.length, 0), pkt end, times))
-      if have_sse2 then
-         print("SSE2: ", benchmark(function() return C.cksum_sse2(pkt.data, pkt.length, 0), pkt end, times))
-      end
-      if have_avx2 then
-         print("AVX2: ", benchmark(function() return C.cksum_avx2(pkt.data, pkt.length, 0), pkt end, times))
-      end
-      print("New: ", benchmark(function() return newchecksum(pkt.data, pkt.length), pkt end, times))
+      print("C: ", benchmark(function() return C.cksum_generic(pkt.data, pkt.length, 0), pkt end, times))
+      print("ASM: ", benchmark(function() return checksum(pkt.data, pkt.length, 0), pkt end, times))
    end
    local function verify_correctness ()
       local function hex (num)
@@ -181,12 +174,12 @@ function selftest ()
       local ntohs = lib.ntohs
       for size=44,1500 do
          local pkt = create_packet(size)
-         assert(hex(ntohs(newchecksum(pkt.data, pkt.length))) == hex(ntohs(checksum_lua(pkt.data, pkt.length))))
-         assert(hex(ntohs(newchecksum(pkt.data, pkt.length))) == hex(C.cksum_generic(pkt.data, pkt.length, 0)))
+         assert(hex(checksum(pkt.data, pkt.length, 0)) == hex(ntohs(checksum_lua(pkt.data, pkt.length))))
+         assert(hex(checksum(pkt.data, pkt.length, 0)) == hex(C.cksum_generic(pkt.data, pkt.length, 0)))
       end
    end
 
-   print("selftest: newchecksum")
+   print("selftest: checksum")
    verify_correctness()
    benchmark_report(44, 14.4)
    benchmark_report(550, 2)

src/lib/checksum.lua

@@ -10,17 +10,7 @@ local ffi = require("ffi")
 local C = ffi.C
 local band, lshift = bit.band, bit.lshift
 
--- Select ipsum(pointer, len, initial) function based on hardware
--- capability.
-local cpuinfo = lib.readfile("/proc/cpuinfo", "*a")
-assert(cpuinfo, "failed to read /proc/cpuinfo for hardware check")
-local have_avx2 = cpuinfo:match("avx2")
-local have_sse2 = cpuinfo:match("sse2")
-
-if     have_avx2 then ipsum = C.cksum_avx2
-elseif have_sse2 then ipsum = C.cksum_sse2
-else                  ipsum = C.cksum_generic end
-
+ipsum = require("arch.checksum").checksum
 
 function finish_packet (buf, len, offset)
    ffi.cast('uint16_t *', buf+offset)[0] = lib.htons(ipsum(buf, len, 0))
@@ -105,24 +95,13 @@ function selftest ()
    local tests = 1000
    local n = 1000000
    local array = ffi.new("char[?]", n)
-   for i = 0, n-1 do  array[i] = i  end
-   local avx2ok, sse2ok = 0, 0
+   for i = 0, n-1 do array[i] = i end
    for i = 1, tests do
       local initial = math.random(0, 0xFFFF)
-      local ref =   C.cksum_generic(array+i*2, i*10+i, initial)
-      if have_avx2 and C.cksum_avx2(array+i*2, i*10+i, initial) == ref then
-         avx2ok = avx2ok + 1
-      end
-      if have_sse2 and C.cksum_sse2(array+i*2, i*10+i, initial) == ref then
-         sse2ok = sse2ok + 1
-      end
+      local ref = C.cksum_generic(array+i*2, i*10+i, initial)
       assert(ipsum(array+i*2, i*10+i, initial) == ref, "API function check")
    end
-   if have_avx2 then print("avx2: "..avx2ok.."/"..tests) else print("no avx2") end
-   if have_sse2 then print("sse2: "..sse2ok.."/"..tests) else print("no sse2") end
    selftest_ipv4_tcp()
-   assert(not have_avx2 or avx2ok == tests, "AVX2 test failed")
-   assert(not have_sse2 or sse2ok == tests, "SSE2 test failed")
    print("selftest: ok")
 end