Add snabbmark checksum subprogram

snabbco · Aug 23, 2018 · 4ab4551 · 4ab4551
1 parent 24ee9fb
commit 4ab4551
Show file tree

Hide file tree

Showing 3 changed files with 51 additions and 35 deletions.
diff --git a/src/arch/checksum.dasl b/src/arch/checksum.dasl
@@ -136,7 +136,11 @@ end
 checksum = assemble("checksum", "uint32_t(*)(uint8_t*, uint32_t, uint16_t)", gen_checksum())
 
 function selftest ()
+   print("selftest: checksum")
    require("lib.checksum_h")
+   local function hex (num)
+      return ("0x%.2x"):format(num)
+   end
    local function create_packet (size)
       local pkt = {
          data = ffi.new("uint8_t[?]", size),
@@ -147,41 +151,10 @@ function selftest ()
       end
       return pkt
    end
-   local function benchmark (fn, times)
-      local now = os.clock()
-      local csum
-      for i=1,times do
-         csum, pkt = fn()
-      end
-      local elapse = os.clock() - now
-      local ns_per_csum = elapse * 10e9 / times
-      local ns_per_byte = ns_per_csum / pkt.length
-      local ret = {elapse = elapse, ns_per_csum = ns_per_csum, ns_per_byte = ns_per_byte, csum = csum}
-      return ("elapse: %.6f; ns_per_csum: %.2f; ns_per_byte: %.2f"):format(ret.elapse, ret.ns_per_csum, ret.ns_per_byte)
-   end
-   local function benchmark_report (size, mpps)
-      local times = mpps*10^6
+   local ntohs = lib.ntohs
+   for size=44,1500 do
       local pkt = create_packet(size)
-      print(mpps.."M; "..size.." bytes")
-      -- Benchmark for different architectures.
-      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))
+      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
-   local function verify_correctness ()
-      local function hex (num)
-         return ("0x%.2x"):format(num)
-      end
-      local ntohs = lib.ntohs
-      for size=44,1500 do
-         local pkt = create_packet(size)
-         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: checksum")
-   verify_correctness()
-   benchmark_report(44, 14.4)
-   benchmark_report(550, 2)
-   benchmark_report(1500, 1)
 end
diff --git a/src/program/snabbmark/README b/src/program/snabbmark/README
@@ -50,3 +50,6 @@ Usage:
 
   snabbmark ctable
     Benchmark insertion and lookup for the "ctable" data structure.
+
+  snabbmark basic1    <npackets>
+    Benchmark checksum computation implementations in C and DynASM.
diff --git a/src/program/snabbmark/snabbmark.lua b/src/program/snabbmark/snabbmark.lua
@@ -27,6 +27,8 @@ function run (args)
       hash(unpack(args))
    elseif command == 'ctable' and #args == 0 then
       ctable(unpack(args))
+   elseif command == 'checksum' and #args == 0 then
+      checksum_bench(unpack(args))
    else
       print(usage) 
       main.exit(1)
@@ -571,3 +573,41 @@ function ctable ()
       stride = stride * 2
    until stride > 256
 end
+
+function checksum_bench ()
+   require("lib.checksum_h")
+   local checksum = require('arch.checksum').checksum
+   local function create_packet (size)
+      local pkt = {
+         data = ffi.new("uint8_t[?]", size),
+         length = size
+      }
+      for i=0,size-1 do
+         pkt.data[i] = math.random(255)
+      end
+      return pkt
+   end
+   local function benchmark (fn, times)
+      local now = os.clock()
+      local csum
+      for i=1,times do
+         csum, pkt = fn()
+      end
+      local elapse = os.clock() - now
+      local ns_per_csum = elapse * 10e9 / times
+      local ns_per_byte = ns_per_csum / pkt.length
+      local ret = {elapse = elapse, ns_per_csum = ns_per_csum, ns_per_byte = ns_per_byte, csum = csum}
+      return ("%.2f ns per checksum; %.2f ns per byte"):format(ret.ns_per_csum, ret.ns_per_byte)
+   end
+   local function benchmark_report (size, mpps)
+      local times = mpps*10^6
+      local pkt = create_packet(size)
+      io.stdout:write("C:   ", "Size="..size.." bytes".."; MPPS="..mpps.."M; ")
+      print(benchmark(function() return C.cksum_generic(pkt.data, pkt.length, 0), pkt end, times))
+      io.stdout:write("ASM: ", "Size="..size.." bytes".."; MPPS="..mpps.."M; ")
+      print(benchmark(function() return checksum(pkt.data, pkt.length, 0), pkt end, times))
+   end
+   benchmark_report(44, 14.4)
+   benchmark_report(550, 2)
+   benchmark_report(1516, 1)
+end