GPU上如何优化卷积
本文将演示如何在TVM中编写高性能卷积实现。我们以平方大小的输入张量和滤波器为例，假设卷积的输入是大批量的。在本例中，使用不同的布局来存储数据，以实现更好的数据局部性。缓冲区布局为HWCN，代表高度、宽度、通道、批次。
Preparation and Algorithm
对于256个通道和14 x 14维的输入张量，使用固定大小。批量大小是256。卷积滤波器包含512个尺寸为3 x 3的滤波器。使用步幅大小1和填充大小1进行卷积。下面的代码定义了TVM中的卷积算法。
import numpy as np
import tvm
from tvm import te
# The sizes of inputs and filters
batch = 256
in_channel = 256
out_channel = 512
in_size = 14
kernel = 3
pad = 1
stride = 1
# Algorithm
A = te.placeholder((in_size, in_size, in_channel, batch), name="A")
W = te.placeholder((kernel, kernel, in_channel, out_channel), name="W")
out_size = (in_size - kernel + 2 * pad) // stride + 1
# Pad input
Apad = te.compute(
(in_size + 2 * pad, in_size + 2 * pad, in_channel, batch),
lambda yy, xx, cc, nn: tvm.tir.if_then_else(
tvm.tir.all(yy >= pad, yy - pad < in_size, xx >= pad, xx - pad < in_size),
A[yy - pad, xx - pad, cc, nn],
tvm.tir.const(0.0, "float32"),
),
name="Apad",
)
# Create reduction variables
rc = te.reduce_axis((0, in_channel), name="rc")
ry = te.reduce_axis((0, kernel), name="ry")
rx = te.reduce_axis((0, kernel), name="rx")
# Compute the convolution
B = te.compute(
(out_size, out_size, out_channel, batch),
lambda yy, xx, ff, nn: te.sum(
Apad[yy * stride + ry, xx * stride + rx, rc, nn] * W[ry, rx, rc, ff], axis=[ry, rx, rc]
),
name="B",
)
Memory Hierarchy
首先指定缓冲区的内存层次结构。下图显示了GPU内存层次结构。与CPU内存层次结构的一个重要区别是GPU提供了一个称为共享内存的缓存缓冲区，由程序员管理。因此，如何最大限度地利用共享内存中的数据是实现GPU内核高性能的关键。
在本例中，将Apad和W加载到缓冲区AA和WW中，存储在共享内存中。这些缓冲区将由同一线程块内的所有线程共享，以计算卷积。然后每个线程将自己的部分从共享缓冲区加载到本地寄存器AL和WL中。BL是输出B的本地缓存，它也存储在线程本地寄存器中。
# Designate the memory hierarchy
s = te.create_schedule(B.op)
s[Apad].compute_inline() # compute Apad inline
AA = s.cache_read(Apad, "shared", [B])
WW = s.cache_read(W, "shared", [B])
AL = s.cache_read(AA, "local", [B])
WL = s.cache_read(WW, "local", [B])
BL = s.cache_write(B, "local")
Blocking
下面的代码将工作负载分成线程块和单个线程。我们遵循矩阵乘法中的分块方案。如下图所示，给定一个像素坐标（y，x），线程块负责计算输出通道和批处理的块系数x块系数（64x64）的区域。由于共享内存空间的限制，我们每次只从Apad和B加载stepx块系数（8x64）数据到共享内存中的缓冲区。
# tile consts
tile = 8
num_thread = 8
block_factor = tile * num_thread
step = 8
vthread = 2
# Get the GPU thread indices
block_x = te.thread_axis("blockIdx.x")
block_y = te.thread_axis("blockIdx.y")
block_z = te.thread_axis("blockIdx.z")
thread_x = te.thread_axis((0, num_thread), "threadIdx.x")
thread_y = te.thread_axis((0, num_thread), "threadIdx.y")
thread_xz = te.thread_axis((0, vthread), "vthread", name="vx")
thread_yz = te.thread_axis((0, vthread), "vthread", name="vy")
# Split the workloads
hi, wi, fi, ni = s[B].op.axis
bz = s[B].fuse(hi, wi)
by, fi = s[B].split(fi, factor=block_factor)
bx, ni = s[B].split(ni, factor=block_factor)
# Bind the iteration variables to GPU thread indices
s[B].bind(bz, block_z)
s[B].bind(by, block_y)
s[B].bind(bx, block_x)
Virtual Thread Split
进一步将工作负载从一个线程块分割到各个线程。为了避免冲突，将8个线程分成4个部分，然后使用8个线程分成4个部分。因此，如下图所示，每个线程计算4个跨距网格，其中每个网格的大小为4 x 4。
tyz, fi = s[B].split(fi, nparts=vthread) # virtual thread split
txz, ni = s[B].split(ni, nparts=vthread) # virtual thread split
ty, fi = s[B].split(fi, nparts=num_thread)
tx, ni = s[B].split(ni, nparts=num_thread)
s[B].reorder(bz, by, bx, tyz, txz, ty, tx, fi, ni)
s[B].bind(tyz, thread_yz)
s[B].bind(txz, thread_xz)
s[B].bind(ty, thread_y)
s[B].bind(tx, thread_x)
Cooperative Fetching
如前所述，每个时间步都需要将步骤x块因子数据从GPU全局内存传输到共享内存。为了减少每个线程的内存传输，下面的代码允许同一线程块中的线程协同从全局内存中获取相关数据。
# Schedule BL local write
s[BL].compute_at(s[B], tx)
yi, xi, fi, ni = s[BL].op.axis
ry, rx, rc = s[BL].op.reduce_axis
rco, rci = s[BL].split(rc, factor=step)
s[BL].reorder(rco, ry, rx, rci, fi, ni)
# Attach computation to iteration variables
s[AA].compute_at(s[BL], rx)
s[WW].compute_at(s[BL], rx)
s[AL].compute_at(s[BL], rci)
s[WL].compute_at(s[BL], rci)
# Schedule for A's shared memory load
yi, xi, ci, ni = s[AA].op.axis
ty, ci = s[AA].split(ci, nparts=num_thread)
tx, ni = s[AA].split(ni, nparts=num_thread)
_, ni = s[AA].split(ni, factor=4)
s[AA].reorder(ty, tx, yi, xi, ci, ni)
s[AA].bind(ty, thread_y)
s[AA].bind(tx, thread_x)
s[AA].vectorize(ni) # vectorize memory load
# Schedule for W's shared memory load
yi, xi, ci, fi = s[WW].op.axis
ty, ci = s[WW].split(ci, nparts=num_thread)
tx, fi = s[WW].split(fi, nparts=num_thread)
_, fi = s[WW].split(fi, factor=4)
s[WW].reorder(ty, tx, yi, xi, ci, fi)
s[WW].bind(ty, thread_y)
s[WW].bind(tx, thread_x)
s[WW].vectorize(fi) # vectorize memory load
Generate CUDA Kernel
最后利用TVM生成并编译了CUDA内核，并对卷积延迟进行了评估。
func = tvm.build(s, [A, W, B], "cuda")
ctx = tvm.gpu(0)
a_np = np.random.uniform(size=(in_size, in_size, in_channel, batch)).astype(A.dtype)
w_np = np.random.uniform(size=(kernel, kernel, in_channel, out_channel)).astype(W.dtype)
a = tvm.nd.array(a_np, ctx)
w = tvm.nd.array(w_np, ctx)
b = tvm.nd.array(np.zeros((out_size, out_size, out_channel, batch), dtype=B.dtype), ctx)
func(a, w, b)
evaluator = func.time_evaluator(func.entry_name, ctx, number=1)
print("Convolution: %f ms" % (evaluator(a, w, b).mean * 1e3))
Out:
Convolution: 53.197723 ms
https://tvm.apache.org/docs/tutorials/optimize/opt_conv_cuda.html