TVM优化Deep Learning GPU算子
TVM优化Deep Learning GPU算子
高效的深度学习算子是深度学习系统的核心。通常,这些算子很难优化,需要HPC专家付出巨大的努力。 端到端张量IR / DSL堆栈TVM使这一过程变得更加容易。
如何在TVM的帮助下编写高性能GPU运算符内核。本文以深度卷积(即topi.nn.depthwise_conv2d_nchw)为例,并演示如何在tensorflow中改进已经手工优化的CUDA内核。在不同的工作负载下,最终版本比tf-1.2中优化的内核快2到4倍,在启用了算子融合的情况下,最终版本快3到7倍。以下是在GTX1080上测试的结果,filter size = [1, 256, 3, 3], stride = [1, 1], padding = ‘SAME’:
深度卷积简介
深度卷积是现代体系结构的重要组成部分,例如Xception和MobileNet。这是降低深度神经网络计算复杂度的有效方法。
在TVM中,深度卷积可以声明为:
# padding stage
PaddedInput = tvm.compute(
(batch, in_channel, height_after_pad, width_after_pad),
lambda b, c, i, j: tvm.select(
tvm.all(i >= pad_top, i - pad_top < in_height, j >= pad_left, j - pad_left < in_width),
Input[b, c, i - pad_top, j - pad_left], tvm.const(0.0)),
name="PaddedInput")
# depthconv stage
di = tvm.reduce_axis((0, filter_height), name='di')
dj = tvm.reduce_axis((0, filter_width), name='dj')
Output = tvm.compute(
(batch, out_channel, out_height, out_width),
lambda b, c, i, j: tvm.sum(
PaddedInput[b, c/channel_multiplier, i*stride_h + di, j*stride_w + dj] * Filter[c/channel_multiplier, c%channel_multiplier, di, dj],
axis=[di, dj]),
name='DepthwiseConv2d')
通用GPU优化准则
本部分简要讨论了优化CUDA代码时应了解的三个概念:数据重用,共享内存和存储冲突。
数据重用
在现代计算体系结构中,从内存中加载数据的成本要比进行单个浮点计算高得多。因此,始终希望在将输入数据加载到寄存器或共享内存(高速缓存)后重用。
深度卷积有两种形式的数据重用:filter过滤器重用和输入重用。当filter过滤器在输入通道上滑动并多次计算时,会发生filter过滤器重用。输入重用是通过平铺实现的,以3x3深度转换为例:
General GPU Optimization Guidelines
在不进行平铺的情况下,每个线程计算1个输出元素并加载3x3输入数据。16个线程加在一起有9x16的负载。
通过平铺,每个线程计算2x2输出元素并加载4x4输入数据。4个线程加在一起有16x4的负载。
共享内存和bank冲突
共享内存可以看作是GPU中的缓存。这是片上的,并且比全局存储快得多。
共享内存按block块分配。通常的做法是将数据从全局内存加载到共享内存中, block块中的所有线程都从共享内存中读取数据。
共享内存的大小是有限的(通常为48K),必须谨慎对待共享内存的溢出。此外,分配给一个block块的共享内存过多,限制了每个多处理器的活动块数。
共享内存的另一个性能问题是存储区冲突。共享内存分为大小相等的内存模块(可同时访问),但是,如果多个线程访问同一内存库(导致内存库冲突),则访问将被序列化,从而降低了有效带宽。
共享存储体的组织方式是将连续的地址分配给连续的存储体。为避免存储区冲突,最好连续的线程,访问连续的内存地址,如下图所示(每种颜色代表一个共享存储区):
开始优化TVM中的深度卷积。
调度优化
计算PaddedInput内联以节省内存分配
从第1部分中可以看到,padding填充被显式声明为一个单独的阶段。内联计算以避免冗余的内存分配:
s=
tvm.create_schedule(Output.op)
s[PaddedInput].compute_inline()
将一个大通道划分为较小的块
深度卷积的一个简单明了的调度表是,一个cuda块负责一个输入通道和相应的filter过滤器,加载到共享内存中,然后进行计算:
IS = s.cache_read(PaddedInput, "shared", [DepthwiseConv2d])
FS = s.cache_read(Filter, "shared", [DepthwiseConv2d])
block_y = tvm.thread_axis("blockIdx.y")
block_x = tvm.thread_axis("blockIdx.x")
# bind the dimension of batch (N in NCHW) with block_y
s[Output].bind(Output.op.axis[0], block_y)
# bind the dimension of channel (C in NCHW) with block_x
s[Output].bind(Output.op.axis[1], block_x)
Here is the result: 测试了在GTX 1080上运行1000次的平均时间成本,并与tensorflow中的depthwise_conv2d进行了比较。结果如下:
Input |
Filter |
stride |
tf-1.2 SAME pad (us) |
TVM SAME pad (us) |
[1, 256, 21, 21] |
[256, 1, 3, 3] |
[1, 1] |
16.1 |
9.1 |
[1, 256, 32, 32] |
[256, 1, 3, 3] |
[1, 1] |
34.8 |
14.5 |
[1, 256, 64, 64] |
[256, 1, 3, 3] |
[1, 1] |
130.9 |
98.9 |
[1, 256, 96, 96] |
[256, 1, 3, 3] |
[1, 1] |
251.6 |
387.4 |
As we can see, this schedule performs well with small channel size like 21 x 21 or 32 x 32, however, its performance drops seriously as the channel size increases to larger than 64 x 64. One main reason is that too much shared memory allocated to one block limits the number of active blocks per multiprocessor.
此调度在较小的通道大小(例如21 x 21或32 x 32)下表现良好,但是,当通道大小增加到大于64 x 64时,其性能会严重下降。一个主要原因是分配的共享内存过多分配到一块,限制每个多处理器的活动块数。
修改了调度表,将一个大频道划分为多个较小的块。例如,一个通道(64 x 64或96 x 96)被分成32 x 32的块,而一个cuda块负责一个32 x 32的块:
blocking_h = 32
blocking_w = 32
# split the dimension of height (H in NCHW)
bx1, _ = s[Output].split(Output.op.axis[2], factor=blocking_h)
# split the dimension of width (W in NCHW)
bx2, _ = s[Output].split(Output.op.axis[3], factor=blocking_w)
# assign one 32 x 32 block to one cuda block
by = s[Output].fuse(Output.op.axis[0], Output.op.axis[1])
s[Output].bind(by, block_y)
bx = s[Output].fuse(bx1, bx2)
s[Output].bind(bx, block_x)
结果如下:
Input |
[blocking_h, blocking_w] |
tf-1.2 SAME pad (us) |
TVM SAME pad (us) |
[1, 256, 64, 64] |
[32, 32] |
130.9 |
63.4 |
[1, 256, 96, 96] |
[32, 32] |
251.6 |
132.5 |
封锁策略有效!对于64 x 64通道大小,带来1.6倍加速(98.9us-> 63.4us);对于96 x 96通道大小,带来2.9倍加速(387.4us-> 132.5us)。
线程的调整参数
如何在一个cuda块的线程之间调度工作负载(例如32x32)?直观地,应该是这样的:
num_thread_y = 8
num_thread_x = 8
thread_y = tvm.thread_axis((0, num_thread_y), "threadIdx.y")
thread_x = tvm.thread_axis((0, num_thread_x), "threadIdx.x")
ty, yi = s[Output].split(h_dim, nparts=num_thread_y)
tx, xi = s[Output].split(w_dim, nparts=num_thread_x)
s[Output].reorder(ty, tx, yi, xi)
s[Output].bind(ty, thread_y)
s[Output].bind(tx, thread_x)
调度表中有两个参数:num_thread_y
和num_thread_x
。如何确定最佳组合?先做一些实验。以下是Filter = [256,1,3,3]和stride = [1,1]的结果:
Case |
Input |
num_thread_y |
num_thread_x |
TVM SAME pad (us) |
1 |
[1, 256, 32, 32] |
8 |
32 |
9.7 |
2 |
[1, 256, 32, 32] |
4 |
32 |
8.8 |
3 |
[1, 256, 32, 32] |
1 |
32 |
17.7 |
4 |
[1, 256, 32, 32] |
32 |
1 |
32.5 |
从以上结果中可以得到:
- 情况2比情况1快。在情况2中,每个线程在输出中计算一个8x1的图块,对应于输入中的10x3的图块。比情况1的4x1 tile具有更好的数据重用性。
- 情况3比情况2慢。这是因为在情况3中,每个线程的工作量太大,导致读取本地内存的成本较高。
- 情况4比情况3慢。这是因为num_thread_x = 32确保没有bank冲突,而num_thread_y = 32没有。
总结一下:
- 大图块有利于数据重用,但不利于本地内存读取。
- num_thread_y和num_thread_x对bank冲突的影响是不对称的。
- 为了找到num_thread_y和num_thread_x的最佳组合,实现高效共享存储器访问(避免组冲突),数据复用,本地存储器read的平衡。
如何才能找到最佳组合呢?答案是蛮力搜索。可以将num_thread_y和num_thread_x作为参数传递给schedule函数,并尝试所有可能的组合以找到最佳组合。这可以在TVM中轻松完成:
def schedule_depthwise_conv2d(..., num_thread_y=8, num_thread_x=8):
num_thread_y = num_thread_y
num_thread_x = num_thread_x
do_schedule_as_usual
return schedule
min_time_cost = inf
for num_thread_y, num_thread_x in all_possible_combinations:
schedule = schedule_depthwise_conv2d(..., num_thread_y=num_thread_y, num_thread_x=num_thread_x)
time_cost = test_depthwise_conv2d(..., schedule)
if time_cost < min_time_cost:
min_time_cost = time_cost
optimal_combination = [num_thread_y, num_thread_x]
实际上,可以看作是一个简单的自动调度程序。
Vthread和交叉模式
引入TVM中的Vthread(虚拟线程),支持跨步模式。可以这样使用:
num_vthread_y = 2
num_vthread_x = 2
num_thread_y = 8
num_thread_x = 8
thread_vy = tvm.thread_axis((0, num_vthread_y), "vthread", name="vy")
thread_vx = tvm.thread_axis((0, num_vthread_x), "vthread", name="vx")
thread_y = tvm.thread_axis((0, num_thread_y), "threadIdx.y")
thread_x = tvm.thread_axis((0, num_thread_x), "threadIdx.x")
# split the dimension of height (H in NCHW) twice
tvy, vyi = s[Output].split(h_dim, nparts=num_vthread_y)
ty, yi = s[Output].split(vyi, nparts=num_thread_y)
# split the dimension of width (W in NCHW) twice
tvx, vxi = s[Output].split(w_dim, nparts=num_vthread_x)
tx, xi = s[Output].split(vxi, nparts=num_thread_x)
# bind thread and vthread respectively
s[Output].bind(tvy, thread_vy)
s[Output].bind(tvx, thread_vx)
s[Output].bind(ty, thread_y)
s[Output].bind(tx, thread_x)
s[Output].reorder(tvy, tvx, ty, tx, yi, xi)
Let’s print the IR to see what vthread does:
/* Input = [1, 1, 32, 32], Filter = [1, 1, 3, 3], stride = [1, 1], padding = 'SAME' */
produce DepthwiseConv2d {
// attr [iter_var(blockIdx.y, , blockIdx.y)] thread_extent = 1
// attr [iter_var(blockIdx.x, , blockIdx.x)] thread_extent = 1
// attr [iter_var(threadIdx.y, Range(min=0, extent=8), threadIdx.y)] thread_extent = 8
// attr [iter_var(threadIdx.x, Range(min=0, extent=8), threadIdx.x)] thread_extent = 8
for (i.inner.inner.inner, 0, 2) {
for (j.inner.inner.inner, 0, 2) {
DepthwiseConv2d[((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner)] = 0.000000f
DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 512)] = 0.000000f
DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 16)] = 0.000000f
DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 528)] = 0.000000f
for (di, 0, 3) {
for (dj, 0, 3) {
DepthwiseConv2d[((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner)] = (DepthwiseConv2d[((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner)] + (tvm_if_then_else(((((((1 - di) - i.inner.inner.inner) <= (((blockIdx.x*16) + threadIdx.y)*2)) && ((((blockIdx.x*16) + threadIdx.y)*2) < ((33 - di) - i.inner.inner.inner))) && (((1 - dj) - j.inner.inner.inner) <= (threadIdx.x*2))) && ((threadIdx.x*2) < ((33 - dj) - j.inner.inner.inner))), Input[(((((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + (di*32)) + dj) + -33)], 0.000000f)*Filter[((di*3) + dj)]))
DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 512)] = (DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 512)] + (tvm_if_then_else(((((((-15 - di) - i.inner.inner.inner) <= (((blockIdx.x*16) + threadIdx.y)*2)) && ((((blockIdx.x*16) + threadIdx.y)*2) < ((17 - di) - i.inner.inner.inner))) && (((1 - dj) - j.inner.inner.inner) <= (threadIdx.x*2))) && ((threadIdx.x*2) < ((33 - dj) - j.inner.inner.inner))), Input[(((((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + (di*32)) + dj) + 479)], 0.000000f)*Filter[((di*3) + dj)]))
DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 16)] = (DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 16)] + (tvm_if_then_else(((((((1 - di) - i.inner.inner.inner) <= (((blockIdx.x*16) + threadIdx.y)*2)) && ((((blockIdx.x*16) + threadIdx.y)*2) < ((33 - di) - i.inner.inner.inner))) && (((-15 - dj) - j.inner.inner.inner) <= (threadIdx.x*2))) && ((threadIdx.x*2) < ((17 - dj) - j.inner.inner.inner))), Input[(((((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + (di*32)) + dj) + -17)], 0.000000f)*Filter[((di*3) + dj)]))
DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 528)] = (DepthwiseConv2d[(((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + 528)] + (tvm_if_then_else(((((((-15 - di) - i.inner.inner.inner) <= (((blockIdx.x*16) + threadIdx.y)*2)) && ((((blockIdx.x*16) + threadIdx.y)*2) < ((17 - di) - i.inner.inner.inner))) && (((-15 - dj) - j.inner.inner.inner) <= (threadIdx.x*2))) && ((threadIdx.x*2) < ((17 - dj) - j.inner.inner.inner))), Input[(((((((((((blockIdx.y + blockIdx.x)*16) + threadIdx.y)*32) + threadIdx.x)*2) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + (di*32)) + dj) + 495)], 0.000000f)*Filter[((di*3) + dj)]))
}
}
}
}
}
Without vthread (just set to 1), the IR is:
/* Input = [1, 1, 32, 32], Filter = [1, 1, 3, 3], stride = [1, 1], padding = 'SAME' */
produce DepthwiseConv2d {
// attr [iter_var(blockIdx.y, , blockIdx.y)] thread_extent = 1
// attr [iter_var(blockIdx.x, , blockIdx.x)] thread_extent = 1
// attr [iter_var(threadIdx.y, Range(min=0, extent=8), threadIdx.y)] thread_extent = 8
// attr [iter_var(threadIdx.x, Range(min=0, extent=8), threadIdx.x)] thread_extent = 8
for (i.inner.inner.inner, 0, 4) {
for (j.inner.inner.inner, 0, 4) {
DepthwiseConv2d[((((((((blockIdx.y + blockIdx.x)*8) + threadIdx.y)*32) + threadIdx.x)*4) + (i.inner.inner.inner*32)) + j.inner.inner.inner)] = 0.000000f
for (di, 0, 3) {
for (dj, 0, 3) {
DepthwiseConv2d[((((((((blockIdx.y + blockIdx.x)*8) + threadIdx.y)*32) + threadIdx.x)*4) + (i.inner.inner.inner*32)) + j.inner.inner.inner)] = (DepthwiseConv2d[((((((((blockIdx.y + blockIdx.x)*8) + threadIdx.y)*32) + threadIdx.x)*4) + (i.inner.inner.inner*32)) + j.inner.inner.inner)] + (tvm_if_then_else(((((((1 - di) - i.inner.inner.inner) <= (((blockIdx.x*8) + threadIdx.y)*4)) && ((((blockIdx.x*8) + threadIdx.y)*4) < ((33 - di) - i.inner.inner.inner))) && (((1 - dj) - j.inner.inner.inner) <= (threadIdx.x*4))) && ((threadIdx.x*4) < ((33 - dj) - j.inner.inner.inner))), Input[(((((((((((blockIdx.y + blockIdx.x)*8) + threadIdx.y)*32) + threadIdx.x)*4) + (i.inner.inner.inner*32)) + j.inner.inner.inner) + (di*32)) + dj) + -33)], 0.000000f)*Filter[((di*3) + dj)]))
}
}
}
}
}
可以看到,当num_vthread_y = 2和时num_vthread_x = 2,将32 x 32通道分为四个16 x 16子通道。每个线程一次计算四个输出元素,一个子通道中一个元素。
以下是Filter = [256,1,3,3],stride = [1,1],blocking_h = 32,blocking_w = 32的结果:
Case |
Input |
num_thread_y, num_thread_x |
num_vthread_y, num_vthread_x |
TVM SAME pad (us) |
1 |
[1, 256, 96, 96] |
8, 8 |
1, 1 |
132.5 |
2 |
[1, 256, 96, 96] |
8, 8 |
1, 4 |
103.1 |
3 |
[1, 256, 96, 96] |
4, 32 |
1, 1 |
95.9 |
4 |
[1, 256, 96, 96] |
8, 16 |
1, 2 |
90.9 |
Case 2比Case 1快。在Case 2中,num_thread_x=8
并且num_vthread_x=4
一起确保连续的线程访问连续的内存地址,从而避免了存储区冲突,如下所示(每种颜色代表一个线程的工作量):
从理论上讲,case 3和case 4应该同样很快,每个线程的工作量相同,并且都享有有效的共享内存访问。不管怎样,case 4快了一点。
还记得tensorflow的速度吗?现在是251.6us,现在TVM快了2.8倍。387.4-> 132.5-> 95.9-> 90.9,封锁最有帮助;调整线程数可节省37us;vthread可以节省额外的5us。
实际上,TVM可以比具有大内核大小或channel_multiplier的tensorflow快得多(因为更多的filter过滤器重用):
Input |
Filter |
stride |
tf-1.2 SAME pad (us) |
TVM SAME pad (us) |
How faster is TVM |
[1, 256, 96, 96] |
[256, 1, 3, 3] |
[1, 1] |
251.6 |
90.9 |
2.8x |
[1, 256, 96, 96] |
[256, 1, 5, 5] |
[1, 1] |
597.6 |
128.9 |
4.6x |
[1, 256, 96, 96] |
[256, 2, 3, 3] |
[1, 1] |
659.9 |
143.7 |
4.6x |
[1, 256, 96, 96] |
[256, 2, 5, 5] |
[1, 1] |
1203.9 |
170.5 |
7.1x |
Consider a common pattern in neural networks: depthwise_conv2d + scale_shift + relu. We can fuse the three operators into one, by slightly modifying the original schedule:
算子融合
算子融合是可以在深度学习中进行的一种典型优化,可以在单个内核中一起计算多个算子,无需将中间结果保存回全局内存中。TVM对此提供了开箱即用的支持。
神经网络中的一个常见模式:depthwise_conv2d
+ scale_shift
+ relu
。稍微修改原始调度表,可以将三个算子融合为一个:
DepthwiseConv2d = topi.nn.depthwise_conv2d(Input, Filter, stride, padding)
ScaleShift = topi.nn.scale_shift(DepthwiseConv2d, Scale, Shift)
Relu = topi.nn.relu(ScaleShift)
Output = Relu # is no longer DepthwiseConv2d
s[ScaleShift].compute_inline() # this line fuses ScaleShift, explicitly
s[DepthwiseConv2d].set_scope("local") # this line fuses DepthwiseConv2d, implicitly
schedule(Output) # schedule for Output the same way we schedule for DepthwiseConv2d as discussed above
s[DepthwiseConv2d].compute_at(s[Output], tx) # tx is the inner most axis, bound to threadIdx.x
生成IR,如下所示:
/* Input = [1, 1, 32, 32], Filter = [1, 1, 3, 3], stride = [1, 1], padding = 'SAME' */
produce Relu {
// attr [iter_var(blockIdx.y, , blockIdx.y)] thread_extent = 1
// attr [DepthwiseConv2d] storage_scope = "local"
allocate DepthwiseConv2d[float32 * 1 * 1 * 4 * 4]
// attr [iter_var(blockIdx.x, , blockIdx.x)] thread_extent = 1
// attr [iter_var(threadIdx.y, Range(min=0, extent=8), threadIdx.y)] thread_extent = 8
// attr [iter_var(threadIdx.x, Range(min=0, extent=8), threadIdx.x)] thread_extent = 8
produce DepthwiseConv2d {
for (i, 0, 4) {
for (j, 0, 4) {
DepthwiseConv2d[((i*4) + j)] = 0.000000f
for (di, 0, 3) {
for (dj, 0, 3) {
DepthwiseConv2d[((i*4) + j)] = (DepthwiseConv2d[((i*4) + j)] + (tvm_if_then_else(((((((1 - di) - i) <= (((blockIdx.x*8) + threadIdx.y)*4)) && ((((blockIdx.x*8) + threadIdx.y)*4) < ((33 - di) - i))) && (((1 - dj) - j) <= (threadIdx.x*4))) && ((threadIdx.x*4) < ((33 - dj) - j))), Input[(((((((((((blockIdx.y + blockIdx.x)*8) + threadIdx.y)*32) + threadIdx.x)*4) + (i*32)) + j) + (di*32)) + dj) + -33)], 0.000000f)*Filter[((di*3) + dj)]))
}
}
}
}
}
for (i2.inner.inner.inner, 0, 4) {
for (i3.inner.inner.inner, 0, 4) {
Relu[((((((((blockIdx.y + blockIdx.x)*8) + threadIdx.y)*32) + threadIdx.x)*4) + (i2.inner.inner.inner*32)) + i3.inner.inner.inner)] = max(((DepthwiseConv2d[((i2.inner.inner.inner*4) + i3.inner.inner.inner)]*Scale[0]) + Shift[0]), 0.000000f)
}
}
}
写入depthwise_conv2d
全局内存的结果之前,每个线程计算scale_shift
和relu
。融合算子的速度与single depthwise_conv2d
一样快。以下是输入= [1、256、96、96],filter过滤器= [256、1、3、3],stride步幅= [1、1],padding填充='SAME'的结果:
- tf-1.2 depthwise_conv2d: 251.6 us
- tf-1.2 depthwise_conv2d + scale_shift + relu (separate): 419.9 us
- TVM depthwise_conv2d: 90.9 us
- TVM depthwise_conv2d + scale_shift + relu (fused): 91.5 us
The advantage of operator fusion is obvious.
This is not the end, TVM can do operator fusion in a smarter way. You may refer to this and read the source code provided below.
Show me the code算子融合的优势显而易见的。
这不是终点,TVM可以以更智能的方式进行算子融合。参考链接:
- Declare: https://github.com/apache/incubator-tvm/blob/main/topi/python/topi/nn/depthwise_conv2d.py
- Schedule: https://github.com/apache/incubator-tvm/blob/main/topi/python/topi/cuda/depthwise_conv2d.py
- Test: https://github.com/apache/incubator-tvm/blob/main/topi/recipe/conv/depthwise_conv2d_test.py