喵~不知不觉到了CUDA系列学习第五讲,前几讲中我们主要介绍了基础GPU中的软硬件结构,内存管理,task类型等;这一讲中我们将介绍3个基础的GPU算法:reduce,scan,histogram,它们在并行算法中非常常用,我们在本文中分别就其功能用处,串行与并行实现进行阐述。
———-
1. Task complexity
task complexity包括step complexity(可以并行成几个操作) & work complexity(总共有多少个工作要做)。
e.g. 下面的tree-structure图中每个节点表示一个操作数,每条边表示一个操作,同层edge表示相同操作,问该图表示的task的step complexity & work complexity分别是多少。
Ans:
step complexity: 3;
work complexity: 6。
下面会有更具体的例子。
2. Reduce
引入:我们考虑一个task:1+2+3+4+…
1) 最简单的顺序执行顺序组织为((1+2)+3)+4…
2) 由于operation之间没有依赖关系,我们可以用Reduce简化操作,它可以减少serial implementation的步数。
2.1 what is reduce?
Reduce input:
- set of elements
- reduction operation
- binary: 两个输入一个输出
- 操作满足结合律: (a@b)@c = a@(b@c), 其中@表示operator
e.g +, 按位与 都符合;a^b(expotentiation)和减法都不是
2.1.1 Serial implementation of Reduce:
reduce的每一步操作都依赖于其前一个操作的结果。比如对于前面那个例子,n个数相加,work complexity 和 step complexity都是O(n)(原因不言自明吧~)我们的目标就是并行化操作,降下来step complexity. e.g add serial reduce -> parallel reduce。
2.1.2 Parallel implementation of Reduce:
也就是说,我们把step complexity降到了 log2n
举个栗子,如下图所示:
那么如果对 210 个数做parallel reduce add,其step complexity就是10. 那么在这个parallel reduce的第一步,我们需要做512个加法,这对modern gpu不是啥大问题,但是如果我们要对 220 个数做加法呢?就需要考虑到gpu数量了,如果说gpu最多能并行做512个操作,我们就应将 220 个数分成1024*1024(共1024组),每次做 210 个数的加法。这种考虑task规模和gpu数量关系的做法有个理论叫Brent’s Theory. 下面我们具体来看:
也就是进行两步操作,第一步分成1024个block,每个block做加法;第二步将这1024个结果再用1个1024个thread的block进行求和。kernel code:
<code class="hljs objectivec has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">__global__ <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">void</span> parallel_reduce_kernel(<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">float</span> *d_out, <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">float</span>* d_in){
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> myID = threadIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span> + blockIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span> * blockDim<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span>;
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> tid = threadIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span>;
<span class="hljs-comment" style="color: rgb(136, 0, 0); box-sizing: border-box;">//divide threads into two parts according to threadID, and add the right part to the left one, lead to reducing half elements, called an iteration; iterate until left only one element</span>
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">for</span>(<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">unsigned</span> <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> s = blockDim<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span> / <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">2</span> ; s><span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>; s>>=<span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">1</span>){
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">if</span>(tid<s){
d_in[myID] += d_in[myID + s];
}
__syncthreads(); <span class="hljs-comment" style="color: rgb(136, 0, 0); box-sizing: border-box;">//ensure all adds at one iteration are done</span>
}
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">if</span> (tid == <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>){
d_out[blockIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span>] = d_in[myId];
}
}</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li><li style="box-sizing: border-box; padding: 0px 5px;">2</li><li style="box-sizing: border-box; padding: 0px 5px;">3</li><li style="box-sizing: border-box; padding: 0px 5px;">4</li><li style="box-sizing: border-box; padding: 0px 5px;">5</li><li style="box-sizing: border-box; padding: 0px 5px;">6</li><li style="box-sizing: border-box; padding: 0px 5px;">7</li><li style="box-sizing: border-box; padding: 0px 5px;">8</li><li style="box-sizing: border-box; padding: 0px 5px;">9</li><li style="box-sizing: border-box; padding: 0px 5px;">10</li><li style="box-sizing: border-box; padding: 0px 5px;">11</li><li style="box-sizing: border-box; padding: 0px 5px;">12</li><li style="box-sizing: border-box; padding: 0px 5px;">13</li><li style="box-sizing: border-box; padding: 0px 5px;">14</li><li style="box-sizing: border-box; padding: 0px 5px;">15</li></ul>
Quiz: 看一下上面的code可以从哪里进行优化?
Ans:我们在上一讲中提到了global,shared & local memory的速度,那么这里对于global memory的操作可以更改为shared memory,从而进行提速:
<code class="hljs objectivec has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">__global__ <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">void</span> parallel_shared_reduce_kernel(<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">float</span> *d_out, <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">float</span>* d_in){
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> myID = threadIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span> + blockIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span> * blockDim<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span>;
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> tid = threadIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span>;
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">extern</span> __shared__ <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">float</span> sdata[];
sdata[tid] = d_in[myID];
__syncthreads();
<span class="hljs-comment" style="color: rgb(136, 0, 0); box-sizing: border-box;">//divide threads into two parts according to threadID, and add the right part to the left one, lead to reducing half elements, called an iteration; iterate until left only one element</span>
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">for</span>(<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">unsigned</span> <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> s = blockDim<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span> / <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">2</span> ; s><span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>; s>>=<span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">1</span>){
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">if</span>(tid<s){
sdata[tid] += sdata[tid + s];
}
__syncthreads(); <span class="hljs-comment" style="color: rgb(136, 0, 0); box-sizing: border-box;">//ensure all adds at one iteration are done</span>
}
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">if</span> (tid == <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>){
d_out[blockIdx<span class="hljs-variable" style="color: rgb(102, 0, 102); box-sizing: border-box;">.x</span>] = sdata[myId];
}
}</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li><li style="box-sizing: border-box; padding: 0px 5px;">2</li><li style="box-sizing: border-box; padding: 0px 5px;">3</li><li style="box-sizing: border-box; padding: 0px 5px;">4</li><li style="box-sizing: border-box; padding: 0px 5px;">5</li><li style="box-sizing: border-box; padding: 0px 5px;">6</li><li style="box-sizing: border-box; padding: 0px 5px;">7</li><li style="box-sizing: border-box; padding: 0px 5px;">8</li><li style="box-sizing: border-box; padding: 0px 5px;">9</li><li style="box-sizing: border-box; padding: 0px 5px;">10</li><li style="box-sizing: border-box; padding: 0px 5px;">11</li><li style="box-sizing: border-box; padding: 0px 5px;">12</li><li style="box-sizing: border-box; padding: 0px 5px;">13</li><li style="box-sizing: border-box; padding: 0px 5px;">14</li><li style="box-sizing: border-box; padding: 0px 5px;">15</li><li style="box-sizing: border-box; padding: 0px 5px;">16</li><li style="box-sizing: border-box; padding: 0px 5px;">17</li><li style="box-sizing: border-box; padding: 0px 5px;">18</li></ul>
优化的代码中还有一点要注意,就是声明的时候记得我们第三讲中说过的kernel通用表示形式:
<code class="hljs vhdl has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">kernel<<<grid <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">of</span> blocks, <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">block</span> <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">of</span> threads, shmem>>></code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li></ul>
最后一项要在call kernel的时候声明好,即:
<code class="hljs cs has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">parallel_reduce_kernel<<<blocks, threads, threads*<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">sizeof</span>(<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">float</span>)>>>(data_out, data_in);</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li></ul>
好,那么问题来了,对于这两个版本(parallel_reduce_kernel 和 parallel_shared_reduce_kernel), parallel_reduce_kernel比parallel_shared_reduce_kernel多用了几倍的global memory带宽? Ans: 分别考虑两个版本的读写操作:
<code class="hljs has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">parallel_reduce_kernel</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li></ul>
Times |
Read Ops |
Write Ops |
1 |
1024 |
512 |
2 |
512 |
256 |
3 |
256 |
128 |
… |
|
|
n |
1 |
1 |
<code class="hljs has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">parallel_shared_reduce_kernel</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li></ul>
Times |
Read Ops |
Write Ops |
1 |
1024 |
1 |
所以,parallel_reduce_kernel所需的带宽是parallel_shared_reduce_kernel的3倍。
3. Scan
3.1 what is scan?
-
Example:
- input: 1,2,3,4
- operation: Add
- ouput: 1,3,6,10(out[i]=sum(in[0:i]))
-
目的:解决难以并行的问题
拍拍脑袋想想上面这个问题O(n)的一个解法是out[i] = out[i-1] + in[i].下面我们来引入scan。
Inputs to scan:
- input array
- 操作:binary & 满足结合律(和reduce一样)
- identity element [I op a = a], 其中I 是identity element
quiz: what is the identity for 加法,乘法,逻辑与,逻辑或?
Ans:
op |
Identity |
加法 |
0 |
乘法 |
1 |
逻辑或|| |
False |
逻辑与&& |
True |
3.2 what scan does?
I/O |
content |
|
|
|
|
input |
[ a0 |
a1 |
a2 |
… |
an ] |
output |
[ I |
a0 |
a0⨂a1 |
… |
a0⨂a1⨂ … ⨂an ] |
其中 ⨂ 是scan operator,I 是 ⨂ 的identity element
3.2.1 Serial implementation of Scan
很简单:
<code class="hljs matlab has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">int acc = identity;
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">for</span>(<span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span>=<span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>;<span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span><<span class="hljs-transposed_variable" style="box-sizing: border-box;">elements.</span><span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">length</span>();<span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span>++)<span class="hljs-cell" style="box-sizing: border-box;">{
acc = acc op elements[i];
out[i] = acc;
}</span></code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li><li style="box-sizing: border-box; padding: 0px 5px;">2</li><li style="box-sizing: border-box; padding: 0px 5px;">3</li><li style="box-sizing: border-box; padding: 0px 5px;">4</li><li style="box-sizing: border-box; padding: 0px 5px;">5</li></ul>
work complexity: O(n)
step complexity: O(n)
那么,对于scan问题,我们怎样对其进行并行化呢?
3.2.1 Parallel implementation of Scan
考虑scan的并行化,可以并行计算n个output,每个output元素i相当于 a0⨂a1⨂ … ⨂ai ,是一个reduce operation。
Q: 那么问题的work complexity和step complexity分别变为多少了呢?
Ans:
- step complexity:
取决于n个reduction中耗时最长的,即 O(log2n)
- work complexity:
对于每个output元素进行计算,总计算量为0+1+2+…+(n-1),所以复杂度为 O(n2) .
可见,step complexity降下来了,可惜work complexity上去了,那么怎么解决呢?这里有两种Scan算法:
|
more step efficiency |
more work efficiency |
hillis + steele (1986) |
√ |
|
blelloch (1990) |
|
√ |
- Hillis + Steele
对于Scan加法问题,hillis+steele算法的解决方案如下:
即streaming’s
step 0: out[i] = in[i] + in[i-1];
step 1: out[i] = in[i] + in[i-2];
step 2: out[i] = in[i] + in[i-4];
如果元素不存在(向下越界)就记为0;可见step 2的output就是scan 加法的结果(想想为什么,我们一会再分析)。
那么问题来了。。。
Q: hillis + steele算法的work complexity 和 step complexity分别为多少?
Hillis + steele Algorithm complexity
|
log(n) |
O(n−−√) |
O(n) |
O(nlogn) |
O(n^2) |
work complexity |
|
|
|
√ |
|
step complexity |
√ |
|
|
|
|
解释:
为了不妨碍大家思路,我在表格中将答案设为了白色,选中表格可见答案。
- step complexity:
因为第i个step的结果为上一步输出作为in, out[idx] = in[idx] + in[idx - 2^i], 所以step complexity = O(log(n))
- work complexity:
workload = (n−1)+(n−2)+(n−4)+... ,共有 log(n) 项元素相加,所以可以近似看做一个矩阵,对应上图,长 log(n) , 宽n,所以复杂度为 nlog(n) 。
2 .Blelloch
基本思路:Reduce + downsweep
还是先讲做法。我们来看Blelloch算法的具体流程,分为reduce和downsweep 两部分,如图所示。
-
reduce部分:
每个step对相邻两个元素进行求和,但是每个元素在input中只出现一次,即window size=2, step = 2的求和。
Q: reduce部分的step complexity 和 work complexity?
Ans:
Reduce part in Blelloch
|
log(n) |
O(n−−√) |
O(n) |
O(nlogn) |
O(n^2) |
work complexity |
|
|
√ |
|
|
step complexity |
√ |
|
|
|
|
我们依然将答案用白色标出,请选中看答案。
-
downsweep部分:
简单地说,downsweep部分的输入元素是reduce部分镜面反射的结果,对于每一组输入in1 & in2有两个输出,左边输出out1 = in2,右边输出out2 = in1 op in2 (这里的op就是reduce部分的op),如图:
如上上图中的op为加法,那举个例子就有:in1 = 11, in2 = 10, 可得out1 = in2 = 10, out2 = in1 + in2 = 21。由此可以推出downsweep部分的所有value,如上上图。
这里画圈的元素都是从reduce部分直接“天降”(镜面反射)过来的,注意,每一个元素位置只去reduce出来该位置的最终结果,而且由于是镜面反射,step层数越大的reduce计算结果“天降”越快,即从reduce的“天降”顺序为
Q: downsweep部分的step complexity 和 work complexity?
And:downsweep是reduce部分的mirror,所以当然和reduce部分的complexity都一样啦。
综上,Blelloch方法的work complexity为 O(n) ,step 数为 2⋅log(n) .这里我们可以看出相比于Hillis + Steele方法,Blelloch的总工作量更小。那么问题来了,这两种方法哪个更快呢?
ANS:这取决于所用的GPU,问题规模,以及实现时的优化方法。这一边是一个不断变化的问题:一开始我们有很多data(work > processor), 更适合用work efficient parallel algorithm (e.g Blelloch), 随着程序运行,工作量被减少了(processor > work),适合改用step efficient parallel algorithm,这样而后数据又多起来啦,于是我们又适合用work efficient parallel algorithm…
总结一下,见下表为每种方法的complexity,以及适于解决的问题:
|
serial |
Hillis + Steele |
Blelloch |
work |
O(n) |
O(nlogn) |
O(n) |
step |
n |
log(n) |
2*log(n) |
512个元素的vector 512个processor |
|
√ |
|
一百万的vector 512个processor |
|
|
√ |
128k的vector 1个processor |
√ |
|
|
4. Histogram
4.1. what is histogram?
顾名思义,统计直方图就是将一个统计量在直方图中显示出来。
4.2. Histogram 的 Serial 实现:
分两部分:1. 初始化,2. 统计
<code class="hljs matlab has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">for</span>(<span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span> = <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>; <span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span> < <span class="hljs-transposed_variable" style="box-sizing: border-box;">bin.</span>count; <span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span>++)
res<span class="hljs-matrix" style="box-sizing: border-box;">[i]</span> = <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>;
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">for</span>(<span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span> = <span class="hljs-number" style="color: rgb(0, 102, 102); box-sizing: border-box;">0</span>; <span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span><nElements; <span class="hljs-built_in" style="color: rgb(102, 0, 102); box-sizing: border-box;">i</span>++)
res<span class="hljs-matrix" style="box-sizing: border-box;">[computeBin(i)]</span> ++;</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li><li style="box-sizing: border-box; padding: 0px 5px;">2</li><li style="box-sizing: border-box; padding: 0px 5px;">3</li><li style="box-sizing: border-box; padding: 0px 5px;">4</li></ul>
4.3. Histogram 的 Parallel 实现:
- 直接实现:
kernel:
<code class="hljs cs has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">__global__ <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">void</span> naive_histo(<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span>* d_bins, <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">const</span> <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span>* d_in, <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">const</span> <span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">in</span> BIN_COUNT){
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> myID = threadIdx.x + blockDim.x * blockIdx.x;
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> myItem = d_in[myID];
<span class="hljs-keyword" style="color: rgb(0, 0, 136); box-sizing: border-box;">int</span> myBin = myItem % BIN_COUNT;
d_bins[myBin]++;
}</code><ul class="pre-numbering" style="box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; margin: 0px; padding: 6px 0px 40px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="box-sizing: border-box; padding: 0px 5px;">1</li><li style="box-sizing: border-box; padding: 0px 5px;">2</li><li style="box-sizing: border-box; padding: 0px 5px;">3</li><li style="box-sizing: border-box; padding: 0px 5px;">4</li><li style="box-sizing: border-box; padding: 0px 5px;">5</li><li style="box-sizing: border-box; padding: 0px 5px;">6</li></ul>
来想想这样有什么问题?又是我们上次说的read-modify-write问题,而serial implementation不会有这个问题,那么想实现parallel histogram计算有什么方法呢?
法1. accumulate using atomics
即,将最后一句变成
atomicAdd(&(d_bins[myBin]), 1);
但是对于atomics的方法而言,不管GPU多好,并行线程数都被限制到histogram个数N,也就是最多只有N个线程并行。
法2. local memory + reduce
设置n个并行线程,每个线程都有自己的local histogram(一个长为bin数的vector);即每个local histogram都被一个thread顺序访问,所以这样没有shared memory,即便没有用atomics也不会出现read-modify-write问题。
然后,我们将这n个histogram进行合并(即加和),可以通过reduce实现。
法3. sort then reduce by key
将数据组织成key-value对,key为histogram bin,value为1,即
key |
2 |
1 |
1 |
2 |
1 |
0 |
2 |
2 |
value |
1 |
1 |
1 |
1 |
1 |
1 |
1 |
1 |
将其按key排序,形成:
key |
0 |
1 |
1 |
1 |
2 |
2 |
2 |
2 |
value |
1 |
1 |
1 |
1 |
1 |
1 |
1 |
1 |
然后对相同key进行reduce求和,就可以得到histogram中的每个bin的总数。
综上,有三种实现paralle histogram的方法:
1. atomics
2. per_thread histogram, then reduce
3. sort, then reduce by key
5. 总结:
本文介绍了三个gpu基础算法:reduce,scan和histogram的串行及并行实现,并巩固了之前讲过的gpu memory相关知识加以运用。
from: http://blog.csdn.net/abcjennifer/article/details/43528407