不同大小的阵列和写入方法的 RAM 写入速度不一致。 填充数组的最快方法？

Question

I've encountered some inconsistencies in writing speed across different array sizes and methods, so i made a separate program to investigate, its a simple c# console application which creates a byte, int and float array of a specified size, then fills it with a defined value using different methods and measures write speed, in between each method call it also clears the cpu cache for unadulterated results when running the same method n times.

Here is the pastebin dump of the program.cs the only file in a simple program, just copy paste it into a new empty console application in visual studio

The byte/int arrays are meant to be bitmap data, and the float array a depthbuffer, so the byteframe is width * height * 4 for bgra, and the int/float frame is just width * height in size.

This is the results i get on my computer:

500*500
ClearByteFrameAsSpanFill:              MS:0,05 GB/s:18,79 MB/s:18794,50
ClearByteFrameArrayFill:               MS:0,05 GB/s:18,54 MB/s:18541,56
ClearByteFrameAsByteAVX:               MS:0,05 GB/s:19,57 MB/s:19575,94
ClearByteFrameAsByteAVXUnRolled:       MS:0,05 GB/s:18,14 MB/s:18148,42
ClearByteFrameAsByteAVXThreaded(2):    MS:0,08 GB/s:11,96 MB/s:11968,89
ClearByteFrameAsByteAVXThreaded(4):    MS:0,08 GB/s:12,19 MB/s:12192,34
ClearByteFrameAsByteAVXThreaded(6):    MS:0,08 GB/s:12,53 MB/s:12537,63
ClearByteFrameAsByteAVXThreaded(8):    MS:0,08 GB/s:12,37 MB/s:12377,25
ClearByteFrameAsByteAVXThreaded(12):   MS:0,08 GB/s:12,13 MB/s:12137,08
ClearByteFrameMarshalCopy:             MS:0,09 GB/s:10,35 MB/s:10357,00
ClearByteFrameMarshalCopy4:            MS:0,11 GB/s:9,22 MB/s:9220,86
ClearByteFrameMarshalThreaded(2):      MS:0,14 GB/s:7,17 MB/s:7171,80
ClearByteFrameMarshalThreaded(4):      MS:0,12 GB/s:7,89 MB/s:7892,86
ClearByteFrameMarshalThreaded(6):      MS:0,12 GB/s:7,99 MB/s:7993,24
ClearByteFrameMarshalThreaded(8):      MS:0,13 GB/s:7,28 MB/s:7287,14
ClearByteFrameMarshalThreaded(12):     MS:0,13 GB/s:7,48 MB/s:7485,39
ClearByteFrameNaive:                   MS:0,30 GB/s:3,30 MB/s:3308,06
ClearByteFrameBytePointer:             MS:0,25 GB/s:3,97 MB/s:3970,49
ClearByteFrameInt32Pointer:            MS:0,07 GB/s:13,98 MB/s:13981,94
ClearByteFrameInt32PointerThreads(2):  MS:0,07 GB/s:12,58 MB/s:12582,46
ClearByteFrameInt32PointerThreads(4):  MS:0,08 GB/s:11,93 MB/s:11938,29
ClearByteFrameInt32PointerThreads(6):  MS:0,08 GB/s:12,14 MB/s:12144,36
ClearByteFrameInt32PointerThreads(8):  MS:0,08 GB/s:11,62 MB/s:11629,97
ClearByteFrameInt32PointerThreads(12): MS:0,09 GB/s:11,38 MB/s:11388,57
ClearByteFrameInt64Pointer:            MS:0,05 GB/s:18,69 MB/s:18692,43
ClearByteFrameInt64PointerThreads(2):  MS:0,07 GB/s:12,91 MB/s:12917,26
ClearByteFrameInt64PointerThreads(4):  MS:0,08 GB/s:12,47 MB/s:12477,97
ClearByteFrameInt64PointerThreads(6):  MS:0,10 GB/s:11,43 MB/s:11439,57
ClearByteFrameInt64PointerThreads(8):  MS:0,08 GB/s:11,50 MB/s:11504,56
ClearByteFrameInt64PointerThreads(12): MS:0,09 GB/s:11,05 MB/s:11052,96

ClearInt32FrameAsSpanFill:             MS:0,05 GB/s:19,25 MB/s:19256,44
ClearInt32FrameArrayFill:              MS:0,05 GB/s:18,80 MB/s:18803,56

ClearFloatFrameAsSpanFill:             MS:0,05 GB/s:19,24 MB/s:19245,53
ClearFloatFrameArrayFill:              MS:0,05 GB/s:19,29 MB/s:19296,62

--

1500*1500
ClearByteFrameAsSpanFill:              MS:0,47 GB/s:18,96 MB/s:18963,17
ClearByteFrameArrayFill:               MS:0,48 GB/s:18,63 MB/s:18638,61
ClearByteFrameAsByteAVX:               MS:0,46 GB/s:19,54 MB/s:19547,95
ClearByteFrameAsByteAVXUnRolled:       MS:0,47 GB/s:18,98 MB/s:18987,69
ClearByteFrameAsByteAVXThreaded(2):    MS:0,55 GB/s:16,71 MB/s:16715,60
ClearByteFrameAsByteAVXThreaded(4):    MS:0,53 GB/s:17,09 MB/s:17097,25
ClearByteFrameAsByteAVXThreaded(6):    MS:0,53 GB/s:17,10 MB/s:17106,20
ClearByteFrameAsByteAVXThreaded(8):    MS:0,48 GB/s:19,06 MB/s:19066,60
ClearByteFrameAsByteAVXThreaded(12):   MS:0,51 GB/s:17,75 MB/s:17759,48
ClearByteFrameMarshalCopy:             MS:0,81 GB/s:11,05 MB/s:11059,42
ClearByteFrameMarshalCopy4:            MS:0,80 GB/s:11,20 MB/s:11207,69
ClearByteFrameMarshalThreaded(2):      MS:0,83 GB/s:11,14 MB/s:11142,98
ClearByteFrameMarshalThreaded(4):      MS:0,79 GB/s:11,49 MB/s:11494,85
ClearByteFrameMarshalThreaded(6):      MS:0,87 GB/s:10,37 MB/s:10372,56
ClearByteFrameMarshalThreaded(8):      MS:0,85 GB/s:10,69 MB/s:10692,81
ClearByteFrameMarshalThreaded(12):     MS:0,86 GB/s:10,49 MB/s:10497,28
ClearByteFrameNaive:                   MS:2,67 GB/s:3,35 MB/s:3359,32
ClearByteFrameBytePointer:             MS:2,24 GB/s:4,01 MB/s:4015,63
ClearByteFrameInt32Pointer:            MS:0,62 GB/s:14,34 MB/s:14340,04
ClearByteFrameInt32PointerThreads(2):  MS:0,50 GB/s:18,05 MB/s:18052,87
ClearByteFrameInt32PointerThreads(4):  MS:0,49 GB/s:18,30 MB/s:18306,56
ClearByteFrameInt32PointerThreads(6):  MS:0,46 GB/s:19,37 MB/s:19378,58
ClearByteFrameInt32PointerThreads(8):  MS:0,48 GB/s:18,48 MB/s:18483,02
ClearByteFrameInt32PointerThreads(12): MS:0,49 GB/s:18,26 MB/s:18266,97
ClearByteFrameInt64Pointer:            MS:0,46 GB/s:19,26 MB/s:19264,51
ClearByteFrameInt64PointerThreads(2):  MS:0,51 GB/s:17,59 MB/s:17599,42
ClearByteFrameInt64PointerThreads(4):  MS:0,50 GB/s:18,07 MB/s:18075,17
ClearByteFrameInt64PointerThreads(6):  MS:0,47 GB/s:19,19 MB/s:19196,42
ClearByteFrameInt64PointerThreads(8):  MS:0,49 GB/s:18,27 MB/s:18273,37
ClearByteFrameInt64PointerThreads(12): MS:0,48 GB/s:18,49 MB/s:18499,04

ClearInt32FrameAsSpanFill:             MS:0,47 GB/s:18,98 MB/s:18987,70
ClearInt32FrameArrayFill:              MS:0,48 GB/s:18,70 MB/s:18702,73

ClearFloatFrameAsSpanFill:             MS:0,48 GB/s:18,60 MB/s:18609,78
ClearFloatFrameArrayFill:              MS:0,47 GB/s:19,07 MB/s:19077,63

--

4500*4500
ClearByteFrameAsSpanFill:              MS:3,49 GB/s:23,25 MB/s:23251,93
ClearByteFrameArrayFill:               MS:3,45 GB/s:23,47 MB/s:23473,07
ClearByteFrameAsByteAVX:               MS:7,24 GB/s:11,20 MB/s:11200,27
ClearByteFrameAsByteAVXUnRolled:       MS:7,32 GB/s:11,08 MB/s:11081,40
ClearByteFrameAsByteAVXThreaded(2):    MS:6,93 GB/s:11,70 MB/s:11702,41
ClearByteFrameAsByteAVXThreaded(4):    MS:6,44 GB/s:12,58 MB/s:12588,30
ClearByteFrameAsByteAVXThreaded(6):    MS:6,48 GB/s:12,53 MB/s:12536,29
ClearByteFrameAsByteAVXThreaded(8):    MS:6,49 GB/s:12,47 MB/s:12479,53
ClearByteFrameAsByteAVXThreaded(12):   MS:6,59 GB/s:12,28 MB/s:12286,35
ClearByteFrameMarshalCopy:             MS:7,19 GB/s:11,27 MB/s:11270,57
ClearByteFrameMarshalCopy4:            MS:7,25 GB/s:11,17 MB/s:11179,60
ClearByteFrameMarshalThreaded(2):      MS:7,15 GB/s:11,33 MB/s:11337,41
ClearByteFrameMarshalThreaded(4):      MS:7,38 GB/s:10,97 MB/s:10970,25
ClearByteFrameMarshalThreaded(6):      MS:7,58 GB/s:10,68 MB/s:10682,29
ClearByteFrameMarshalThreaded(8):      MS:7,64 GB/s:10,60 MB/s:10601,81
ClearByteFrameMarshalThreaded(12):     MS:7,78 GB/s:10,40 MB/s:10400,19
ClearByteFrameNaive:                   MS:24,18 GB/s:3,34 MB/s:3348,89
ClearByteFrameBytePointer:             MS:20,35 GB/s:3,97 MB/s:3979,91
ClearByteFrameInt32Pointer:            MS:8,10 GB/s:10,00 MB/s:10009,30
ClearByteFrameInt32PointerThreads(2):  MS:7,00 GB/s:11,57 MB/s:11575,67
ClearByteFrameInt32PointerThreads(4):  MS:6,29 GB/s:12,86 MB/s:12868,42
ClearByteFrameInt32PointerThreads(6):  MS:6,50 GB/s:12,48 MB/s:12485,34
ClearByteFrameInt32PointerThreads(8):  MS:6,49 GB/s:12,49 MB/s:12496,20
ClearByteFrameInt32PointerThreads(12): MS:6,64 GB/s:12,19 MB/s:12194,29
ClearByteFrameInt64Pointer:            MS:7,16 GB/s:11,33 MB/s:11331,90
ClearByteFrameInt64PointerThreads(2):  MS:6,80 GB/s:11,92 MB/s:11926,33
ClearByteFrameInt64PointerThreads(4):  MS:6,40 GB/s:12,67 MB/s:12670,50
ClearByteFrameInt64PointerThreads(6):  MS:6,51 GB/s:12,49 MB/s:12490,41
ClearByteFrameInt64PointerThreads(8):  MS:6,48 GB/s:12,51 MB/s:12511,71
ClearByteFrameInt64PointerThreads(12): MS:6,61 GB/s:12,25 MB/s:12256,23

ClearInt32FrameAsSpanFill:             MS:7,41 GB/s:10,94 MB/s:10948,01
ClearInt32FrameArrayFill:              MS:7,28 GB/s:11,13 MB/s:11133,96

ClearFloatFrameAsSpanFill:             MS:7,23 GB/s:11,20 MB/s:11202,94
ClearFloatFrameArrayFill:              MS:7,19 GB/s:11,26 MB/s:11265,38

For reference my computer is an amd ryzen 3600 and independent benchmarking software puts my max write speed at about 20 GB/s

My question is as follows:

1. Why does ClearByteFrameAsSpanFill give good performance on all 3 sizes but int32frame and floatframe using same AsSpan fill method caps at 10GB/s on the large size
2. Why does ClearByteFrameAsByteAVX, ClearByteFrameInt32Pointer and ClearByteFrameInt64Pointer do well on the medium and small size, but not on large? What could be causing this? It's literally the simplest straight forward functions you can imagine, write to a pointer, increment pointer and repeat, array size should have no impact at all.
3. Why does the different methods that come close to the 20GBs mark on the medium test all cap together at around 10GBs mark when slow, is there some secret behind the scene mechanism that caps writing speeds at 50% under certain circumstances?

The only winner is ClearByteFrameAsSpanFill and ClearByteFrameArrayFill, but it is useless since i must reset the buffer with different 4 values per pixel. I thought i could use an int32 array instead of bytes, and just bitshift my rgba values into a single int and clear with that, but it fails on the large size. ClearByteFrameAsByteAVX works small and medium, but fails on large.

I need a singular method that performs well across all sizes, and i am morbidly curious why it behaves like this, you would think something as basic as this would be well optimized in the .net framework.

Any help would be appreciated. Please make a new console application and run the program, see if you get the same problem.

Answer 1

You can use AsSpan approach in combination with MemoryMarshal.Cast to use int or long as fill values:

static double ClearByteFrameAsIntSpanFill(int clearValue) 
{
    var asSpan = ByteFrame.AsSpan();
    var cast = MemoryMarshal.Cast<byte, int>(asSpan);
    cast.Fill(clearValue);
    return (float)ByteFrame.Length / 1000000;
}

static double ClearByteFrameAsLongSpanFill(long clearValue) 
{
    var asSpan = ByteFrame.AsSpan();
    var cast = MemoryMarshal.Cast<byte, long>(asSpan);
    cast.Fill(clearValue);
    return (float)ByteFrame.Length / 1000000;
}

I've run few tests using your code and it seems to have performance similar to ClearByteFrameAsSpanFill / ClearByteFrameArrayFill on my machine.

Though in general I recommend using BenchmarkDotNet for performance testing/investigation.