jbarczak

Results for Sebastian Aaltonen's buffer tester https://github.com/sebbbi/perftest 
From Intel Haswell GT2 (i3 4010-U).   It was necessary to change the threadgroup count to 64x64 down from 1024, or the test would TDR.

Load R8 invariant: 2.106ms
Load R8 linear: 13.438ms
Load R8 random: 6.053ms
Load RG8 invariant: 2.105ms
Load RG8 linear: 12.763ms
Load RG8 random: 6.229ms
Load RGBA8 invariant: 2.105ms

jbarczak

#include <string>
#include <fstream>
#include <istream>
#include <sstream>
#include <boost/tokenizer.hpp>
#include <boost/timer/timer.hpp>

using namespace std;

jbarczak

    static const __m128i SHUFFLE_TABLE[16] = {
         _mm_setr_epi8(12,13,14,15, 8, 9,10,11, 4, 5, 6, 7, 0, 1, 2, 3),
         _mm_setr_epi8( 0, 1, 2, 3,12,13,14,15, 8, 9,10,11, 4, 5, 6, 7),
         _mm_setr_epi8( 4, 5, 6, 7,12,13,14,15, 8, 9,10,11, 0, 1, 2, 3),
         _mm_setr_epi8( 0, 1, 2, 3, 4, 5, 6, 7,12,13,14,15, 8, 9,10,11),
         _mm_setr_epi8( 8, 9,10,11,12,13,14,15, 4, 5, 6, 7, 0, 1, 2, 3),
         _mm_setr_epi8( 0, 1, 2, 3, 8, 9,10,11,12,13,14,15, 4, 5, 6, 7),
         _mm_setr_epi8( 4, 5, 6, 7, 8, 9,10,11,12,13,14,15, 0, 1, 2, 3),

jbarczak

// Tried this, and it was marginally slower
//
//  Some notes about this:
//  1.  Seperate hit/miss arrays force me to use a lot more stack than I did before, and 
//        probably doesn't use the cache quite as well.
//  2.  The prefetching of the rays doesn't fit in quite as neatly, and doesn't help anymore if I stick it in there
//        it might make more sense to move that elsewhere anyway
//  3. LUT is 256 bytes.  Not too bad, but it's probably knocking a few rays out of the cache
//  4.  Reordering can produce at least one packet that is partially miss and partially hit.

jbarczak

  static void __fastcall ReorderRays( StackFrame& frame, size_t nGroups )
    {        
        RayPacket** pPackets = frame.pActivePackets;

        uint32 pIDs[MAX_TRACER_SIZE];
   
        size_t nHitLoc  = 0;
        size_t nMissLoc = 8*nGroups;

        const char* pRays = (const char*) frame.pRays;

jbarczak

Microsoft compiler appears to ignore prefetches inside a loop.  

Tested this on MSVC 2013 express edition. Microsoft's connect site says I am not authorized to submit feedback for who knows what reason, or else I'd send it there directly......

Code I used:

void Foo( char* p, int* q )
{
    for( size_t i=0; i<8; i++ )