Optimizing Parallel Reduction in CUDA : NOTES

Optimizing Parallel Reduction in CUDA
Mark Harris
NVIDIA Developer Technology

2
Parallel Reduction
Common and important data parallel primitive
Easy to implement in CUDA
Harder to get it right
Serves as a great optimization example
We’ll walk step by step through 7 different versions
Demonstrates several important optimization strategies

3
Parallel Reduction
Tree-based approach used within each thread block
Need to be able to use multiple thread blocks
To process very large arrays
To keep all multiprocessors on the GPU busy
Each thread block reduces a portion of the array
But how do we communicate partial results between
thread blocks?
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3

4
Problem: Global Synchronization
If we could synchronize across all thread blocks, could easily
reduce very large arrays, right?
Global sync after each block produces its result
Once all blocks reach sync, continue recursively
But CUDA has no global synchronization. Why?
Expensive to build in hardware for GPUs with high processor
count
Would force programmer to run fewer blocks (no more than #
multiprocessors * # resident blocks / multiprocessor) to avoid
deadlock, which may reduce overall efficiency
Solution: decompose into multiple kernels
Kernel launch serves as a global synchronization point
Kernel launch has negligible HW overhead, low SW overhead

5
Solution: Kernel Decomposition
Avoid global sync by decomposing computation
into multiple kernel invocations
In the case of reductions, code for all levels is the
same
Recursive kernel invocation
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
4 7 5 9
11 14
25
3 1 7 0 4 1 6 3
Level 0:
8 blocks
Level 1:
1 block
// Already tried this

6
What is Our Optimization Goal?
We should strive to reach GPU peak performance
Choose the right metric:
GFLOP/s: for compute-bound kernels
Bandwidth: for memory-bound kernels
Reductions have very low arithmetic intensity
1 flop per element loaded (bandwidth-optimal)
Therefore we should strive for peak bandwidth
Will use G80 GPU for this example
384-bit memory interface, 900 MHz DDR
384 * 1800 / 8 = 86.4 GB/s

7
Reduction #1: Interleaved Addressing
__global__ void reduce0(int *g_idata, int *g_odata) {
extern __shared__ int sdata[];
// each thread loads one element from global to shared mem
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*blockDim.x + threadIdx.x;
sdata[tid] = g_idata[i];
__syncthreads();
// do reduction in shared mem
for(unsigned int s=1; s < blockDim.x; s *= 2) {
if (tid % (2*s) == 0) {
sdata[tid] += sdata[tid + s];
}
__syncthreads();
}
// write result for this block to global mem
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
}

8
Parallel Reduction: Interleaved Addressing
10 1 8 -1 0 -2 3 5 -2 -3 2 7 0 11 0 2
Values (shared memory)
0 2 4 6 8 10 12 14
11 1 7 -1 -2 -2 8 5 -5 -3 9 7 11 11 2 2
Values
0 4 8 12
18 1 7 -1 6 -2 8 5 4 -3 9 7 13 11 2 2
Values
0 8
24 1 7 -1 6 -2 8 5 17 -3 9 7 13 11 2 2
Values
0
41 1 7 -1 6 -2 8 5 17 -3 9 7 13 11 2 2
Values
Thread
IDs
Step 1
Stride 1
Step 2
Stride 2
Step 3
Stride 4
Step 4
Stride 8
Thread
IDs
Thread
IDs
Thread
IDs

9
__global__ void reduce1(int *g_idata, int *g_odata) {
__syncthreads();
// do reduction in shared mem
for (unsigned int s=1; s < blockDim.x; s *= 2) {
if (tid % (2*s) == 0) {
}
__syncthreads();
}
// write result for this block to global mem
}
Problem: highly divergent
warps are very inefficient, and
% operator is very slow

10
Performance for 4M element reduction
Kernel 1:
interleaved addressing
with divergent branching
8.054 ms 2.083 GB/s
Note: Block Size = 128 threads for all tests
Bandwidth
Time (222 ints)

11
if (tid % (2*s) == 0) {
}
__syncthreads();
}
int index = 2 * s * tid;
if (index < blockDim.x) {
sdata[index] += sdata[index + s];
}
__syncthreads();
}
Just replace divergent branch in inner loop:
With strided index and non-divergent branch:

12
Parallel Reduction: Interleaved Addressing
10 1 8 -1 0 -2 3 5 -2 -3 2 7 0 11 0 2
0 1 2 3 4 5 6 7
11 1 7 -1 -2 -2 8 5 -5 -3 9 7 11 11 2 2
Values
0 1 2 3
18 1 7 -1 6 -2 8 5 4 -3 9 7 13 11 2 2
Values
0 1
24 1 7 -1 6 -2 8 5 17 -3 9 7 13 11 2 2
Values
0
41 1 7 -1 6 -2 8 5 17 -3 9 7 13 11 2 2
Values
Thread
IDs
Step 1
Stride 1
Step 2
Stride 2
Step 3
Stride 4
Step 4
Stride 8
Thread
IDs
Thread
IDs
Thread
IDs
New Problem: Shared Memory Bank Conflicts

13
Kernel 1:
8.054 ms 2.083 GB/s
Kernel 2:
with bank conflicts
3.456 ms 4.854 GB/s 2.33x 2.33x
Step
Speedup
Bandwidth
Time (222 ints)
Cumulative
Speedup

14
Parallel Reduction: Sequential Addressing
10 1 8 -1 0 -2 3 5 -2 -3 2 7 0 11 0 2
0 1 2 3 4 5 6 7
8 -2 10 6 0 9 3 7 -2 -3 2 7 0 11 0 2
Values
0 1 2 3
8 7 13 13 0 9 3 7 -2 -3 2 7 0 11 0 2
Values
0 1
21 20 13 13 0 9 3 7 -2 -3 2 7 0 11 0 2
Values
0
41 20 13 13 0 9 3 7 -2 -3 2 7 0 11 0 2
Values
Thread
IDs
Step 1
Stride 8
Step 2
Stride 4
Step 3
Stride 2
Step 4
Stride 1
Thread
IDs
Thread
IDs
Thread
IDs
Sequential addressing is conflict free

15
int index = 2 * s * tid;
if (index < blockDim.x) {
sdata[index] += sdata[index + s];
}
__syncthreads();
}
for (unsigned int s=blockDim.x/2; s>0; s>>=1) {
if (tid < s) {
}
__syncthreads();
}
Reduction #3: Sequential Addressing
Just replace strided indexing in inner loop:
With reversed loop and threadID-based indexing:
// Already using this

16
Kernel 1:
8.054 ms 2.083 GB/s
Kernel 2:
with bank conflicts
3.456 ms 4.854 GB/s 2.33x 2.33x
Kernel 3:
sequential addressing
1.722 ms 9.741 GB/s 2.01x 4.68x
Step
Speedup
Bandwidth
Time (222 ints)
Cumulative
Speedup
Insert text here

17
if (tid < s) {
}
__syncthreads();
}
Idle Threads
Problem:
Half of the threads are idle on first loop iteration!
This is wasteful…

18
__syncthreads();
// perform first level of reduction,
// reading from global memory, writing to shared memory
unsigned int i = blockIdx.x*(blockDim.x*2) + threadIdx.x;
sdata[tid] = g_idata[i] + g_idata[i+blockDim.x];
__syncthreads();
Reduction #4: First Add During Load
Halve the number of blocks, and replace single load:
With two loads and first add of the reduction:

19
Kernel 1:
8.054 ms 2.083 GB/s
Kernel 2:
with bank conflicts
3.456 ms 4.854 GB/s 2.33x 2.33x
Kernel 3:
1.722 ms 9.741 GB/s 2.01x 4.68x
Kernel 4:
first add during global load
0.965 ms 17.377 GB/s 1.78x 8.34x
Step
Speedup
Bandwidth
Time (222 ints)
Cumulative
Speedup

20
Instruction Bottleneck
At 17 GB/s, we’re far from bandwidth bound
And we know reduction has low arithmetic intensity
Therefore a likely bottleneck is instruction overhead
Ancillary instructions that are not loads, stores, or
arithmetic for the core computation
In other words: address arithmetic and loop overhead
Strategy: unroll loops

21
Unrolling the Last Warp
As reduction proceeds, # “active” threads decreases
When s <= 32, we have only one warp left
Instructions are SIMD synchronous within a warp
That means when s <= 32:
We don’t need to __syncthreads()
We don’t need “if (tid < s)” because it doesn’t save any
work
Let’s unroll the last 6 iterations of the inner loop

__device__ void warpReduce(volatile int* sdata, int tid) {
sdata[tid] += sdata[tid + 32];
}
// later…
if (tid < s)
__syncthreads();
}
if (tid < 32) warpReduce(sdata, tid);
22
Reduction #5: Unroll the Last Warp
Note: This saves useless work in all warps, not just the last one!
Without unrolling, all warps execute every iteration of the for loop and if statement
IMPORTANT:
For this to be correct,
we must use the
“volatile” keyword!

23
Kernel 1:
8.054 ms 2.083 GB/s
Kernel 2:
with bank conflicts
3.456 ms 4.854 GB/s 2.33x 2.33x
Kernel 3:
1.722 ms 9.741 GB/s 2.01x 4.68x
Kernel 4:
0.965 ms 17.377 GB/s 1.78x 8.34x
Kernel 5:
unroll last warp
0.536 ms 31.289 GB/s 1.8x 15.01x
Step
Speedup
Bandwidth
Time (222 ints)
Cumulative
Speedup

24
Complete Unrolling
If we knew the number of iterations at compile time,
we could completely unroll the reduction
Luckily, the block size is limited by the GPU to 512 threads
Also, we are sticking to power-of-2 block sizes
So we can easily unroll for a fixed block size
But we need to be generic – how can we unroll for block
sizes that we don’t know at compile time?
Templates to the rescue!
CUDA supports C++ template parameters on device and
host functions

25
Unrolling with Templates
Specify block size as a function template parameter:
template <unsigned int blockSize>
__global__ void reduce5(int *g_idata, int *g_odata)

26
Reduction #6: Completely Unrolled
if (blockSize >= 512) {
if (tid < 256) { sdata[tid] += sdata[tid + 256]; } __syncthreads(); }
if (tid < 32) warpReduce<blockSize>(sdata, tid);
Note: all code in RED will be evaluated at compile time.
Results in a very efficient inner loop!
Template <unsigned int blockSize>
__device__ void warpReduce(volatile int* sdata, int tid) {
if (blockSize >= 64) sdata[tid] += sdata[tid + 32];
}

27
Invoking Template Kernels
Don’t we still need block size at compile time?
Nope, just a switch statement for 10 possible block sizes:
switch (threads)
{
case 512:
reduce5<512><<< dimGrid, dimBlock, smemSize >>>(d_idata, d_odata); break;
case 256:
case 128:
case 64:
reduce5< 64><<< dimGrid, dimBlock, smemSize >>>(d_idata, d_odata); break;
case 32:
case 16:
case 8:
case 4:
case 2:
case 1:
}

28
Kernel 1:
8.054 ms 2.083 GB/s
Kernel 2:
with bank conflicts
3.456 ms 4.854 GB/s 2.33x 2.33x
Kernel 3:
1.722 ms 9.741 GB/s 2.01x 4.68x
Kernel 4:
0.965 ms 17.377 GB/s 1.78x 8.34x
Kernel 5:
unroll last warp
0.536 ms 31.289 GB/s 1.8x 15.01x
Kernel 6:
completely unrolled
0.381 ms 43.996 GB/s 1.41x 21.16x
Step
Speedup
Bandwidth
Time (222 ints)
Cumulative
Speedup

29
Parallel Reduction Complexity
Log(N) parallel steps, each step S does N/2S
independent ops
Step Complexity is O(log N)
For N=2D, performs S[1..D]2D-S = N-1 operations
Work Complexity is O(N) – It is work-efficient
i.e. does not perform more operations than a sequential
algorithm
With P threads physically in parallel (P processors),
time complexity is O(N/P + log N)
Compare to O(N) for sequential reduction
In a thread block, N=P, so O(log N)

30
What About Cost?
Cost of a parallel algorithm is processors time
complexity
Allocate threads instead of processors: O(N) threads
Time complexity is O(log N), so cost is O(N log N) : not
cost efficient!
Brent’s theorem suggests O(N/log N) threads
Each thread does O(log N) sequential work
Then all O(N/log N) threads cooperate for O(log N) steps
Cost = O((N/log N) * log N) = O(N)  cost efficient
Sometimes called algorithm cascading
Can lead to significant speedups in practice

31
Algorithm Cascading
Combine sequential and parallel reduction
Each thread loads and sums multiple elements into
shared memory
Tree-based reduction in shared memory
Brent’s theorem says each thread should sum
O(log n) elements
i.e. 1024 or 2048 elements per block vs. 256
In my experience, beneficial to push it even further
Possibly better latency hiding with more work per thread
More threads per block reduces levels in tree of recursive
kernel invocations
High kernel launch overhead in last levels with few blocks
On G80, best perf with 64-256 blocks of 128 threads
1024-4096 elements per thread

32
__syncthreads();
Reduction #7: Multiple Adds / Thread
Replace load and add of two elements:
With a while loop to add as many as necessary:
unsigned int i = blockIdx.x*(blockSize*2) + threadIdx.x;
unsigned int gridSize = blockSize*2*gridDim.x;
sdata[tid] = 0;
while (i < n) {
sdata[tid] += g_idata[i] + g_idata[i+blockSize];
i += gridSize;
}
__syncthreads();

33
__syncthreads();
Reduction #7: Multiple Adds / Thread
Replace load and add of two elements:
With a while loop to add as many as necessary:
unsigned int i = blockIdx.x*(blockSize*2) + threadIdx.x;
sdata[tid] = 0;
while (i < n) {
sdata[tid] += g_idata[i] + g_idata[i+blockSize];
i += gridSize;
}
__syncthreads();
Note: gridSize loop stride
to maintain coalescing!

34
Kernel 1:
8.054 ms 2.083 GB/s
Kernel 2:
with bank conflicts
3.456 ms 4.854 GB/s 2.33x 2.33x
Kernel 3:
1.722 ms 9.741 GB/s 2.01x 4.68x
Kernel 4:
0.965 ms 17.377 GB/s 1.78x 8.34x
Kernel 5:
unroll last warp
0.536 ms 31.289 GB/s 1.8x 15.01x
Kernel 6:
completely unrolled
0.381 ms 43.996 GB/s 1.41x 21.16x
Kernel 7:
multiple elements per thread
0.268 ms 62.671 GB/s 1.42x 30.04x
Kernel 7 on 32M elements: 73 GB/s!
Step
Speedup
Bandwidth
Time (222 ints)
Cumulative
Speedup

35
__device__ void warpReduce(volatile int *sdata, unsigned int tid) {
}
__global__ void reduce6(int *g_idata, int *g_odata, unsigned int n) {
unsigned int i = blockIdx.x*(blockSize*2) + tid;
sdata[tid] = 0;
while (i < n) { sdata[tid] += g_idata[i] + g_idata[i+blockSize]; i += gridSize; }
__syncthreads();
if (blockSize >= 512) { if (tid < 256) { sdata[tid] += sdata[tid + 256]; } __syncthreads(); }
if (tid < 32) warpReduce(sdata, tid);
}
Final Optimized Kernel
// I guess for global memory, 2 loads in 1 loop good enough
// For shared memory, better load as mush as possible at once (near instruction bottleneck)

36
Performance Comparison
0.01
0.1
1
10
131072
262144
524288
1048576
2097152
4194304
8388608
16777216
33554432
# Elements
Time
(ms)
1: Interleaved Addressing:
Divergent Branches
2: Interleaved Addressing:
Bank Conflicts
3: Sequential Addressing
4: First add during global
load
5: Unroll last warp
6: Completely unroll
7: Multiple elements per
thread (max 64 blocks)

37
Types of optimization
Interesting observation:
Algorithmic optimizations
Changes to addressing, algorithm cascading
11.84x speedup, combined!
Code optimizations
Loop unrolling
2.54x speedup, combined

38
Conclusion
Understand CUDA performance characteristics
Memory coalescing
Divergent branching
Bank conflicts
Latency hiding
Use peak performance metrics to guide optimization
Understand parallel algorithm complexity theory
Know how to identify type of bottleneck
e.g. memory, core computation, or instruction overhead
Optimize your algorithm, then unroll loops
Use template parameters to generate optimal code
Questions: mharris@nvidia.com

Optimizing Parallel Reduction in CUDA : NOTES

More Related Content

Similar to Optimizing Parallel Reduction in CUDA : NOTES

More from Subhajit Sahu

Recently uploaded

Optimizing Parallel Reduction in CUDA : NOTES