700字范文 > TVM Operator Inventory (TOPI)简介

TVM Operator Inventory (TOPI)简介

时间：2019-10-08 12:13:16

TOPI简介

这是 TVM Operator Inventory (TOPI) 的介绍。TOPI 提供了比 TVM 具有更高抽象的 numpy 风格的，通用操作和调度。TOPI 如何在 TVM 中，编写样板代码。

fromfutureimport absolute_import, print_function

import tvm

import tvm.testing

from tvm import te

from tvm import topi

import numpy as np

基本示例

重新审视行总和操作（相当于B = numpy.sum(A, axis=1)），要计算二维 TVM 张量 A 行总和，应该指定符号操作及调度。

n = te.var(“n”)

m = te.var(“m”)

A = te.placeholder((n, m), name=“A”)

k = te.reduce_axis((0, m), “k”)

B = pute((n,), lambda i: te.sum(A[i, k], axis=k), name=“B”)

s = te.create_schedule(B.op)

以人类可读的格式，检查 IR 代码，可以这样做。

print(tvm.lower(s, [A], simple_mode=True))

输出：

primfn(A_1: handle) -> ()

attr = {“from_legacy_te_schedule”: True, “global_symbol”: “main”, “tir.noalias”: True}

buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type=“auto”)}

buffer_map = {A_1: A} {

allocate(B: Pointer(global float32), float32, [n]), storage_scope = global;

for (i: int32, 0, n) {

B[i] = 0f32

for (k: int32, 0, m) {

B[i] = ((float32*)B[i] + (float32*)A_2[((istride) + (kstride_1))])

}

对于这样一个常见的操作，必须定义 reduce 轴，以及使用 pute进行显式计算。对于更复杂的操作，需要提供多少细节。可以用简单topi.sum的，如numpy.sum，替换这两行。

C = topi.sum(A, axis=1)

ts = te.create_schedule(C.op)

print(tvm.lower(ts, [A], simple_mode=True))

输出：

primfn(A_1: handle) -> ()

attr = {“from_legacy_te_schedule”: True, “global_symbol”: “main”, “tir.noalias”: True}

buffers = {A: Buffer(A_2: Pointer(float32), float32, [n: int32, m: int32], [stride: int32, stride_1: int32], type=“auto”)}

buffer_map = {A_1: A} {

allocate(A_red: Pointer(global float32), float32, [n]), storage_scope = global;

for (ax0: int32, 0, n) {

A_red[ax0] = 0f32

for (k1: int32, 0, m) {

A_red[ax0] = ((float32*)A_red[ax0] + (float32*)A_2[((ax0stride) + (k1stride_1))])

}

Numpy 风格的算子重载

可以使用topi.broadcast_add具有正确（可广播特定）shape的张量，添加两个张量。TOPI 为此类常见操作，提供了算子重载。例如，

x, y = 100, 10

a = te.placeholder((x, y, y), name=“a”)

b = te.placeholder((y, y), name=“b”)

c = a + b # same as topi.broadcast_add

d = a * b # same as topi.broadcast_mul

使用相同的语法重载，TOPI 处理，将原语（int，float）广播到 tensor d - 3.14。

通用调度和融合操作

TOPI 如何免于在较低级别的 API 中，编写显式计算。像以前一样进行调度，TOPI根据给定的上下文，提供更高级别的调度方法。例如，对于 CUDA，可以using only topi.generic.schedule_reduce，调度topi.sum结尾的一系列操作。

e = topi.elemwise_sum([c, d])

f = e / 2.0

g = topi.sum(f)

with tvm.target.cuda():

sg = topi.cuda.schedule_reduce(g)

print(tvm.lower(sg, [a, b], simple_mode=True))

输出：

primfn(a_1: handle, b_1: handle) -> ()

attr = {“from_legacy_te_schedule”: True, “global_symbol”: “main”, “tir.noalias”: True}

buffers = {b: Buffer(b_2: Pointer(float32), float32, [10, 10], []),

a: Buffer(a_2: Pointer(float32), float32, [100, 10, 10], [])}

buffer_map = {a_1: a, b_1: b} {

allocate(T_divide_red: Pointer(global float32), float32, [1]), storage_scope = global;

attr [IterVar(threadIdx.x: int32, [0:1024], “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 1024;

allocate(T_divide_red.rf: Pointer(local float32), float32, [1]), storage_scope = local;

allocate(reduce_temp0: Pointer(local float32), float32, [1]), storage_scope = local {

T_divide_red.rf[0] = 0f32

for (k0.k1.fused.k2.fused.outer: int32, 0, 10) {

if @tir.likely((((((k0.k1.fused.k2.fused.outer1024) + threadIdx.x) < 10000) && (((k0.k1.fused.k2.fused.outer1024) + threadIdx.x) < 10000)) && (((k0.k1.fused.k2.fused.outer1024) + threadIdx.x) < 10000)), dtype=bool) {

T_divide_red.rf[0] = ((float32)T_divide_red.rf[0] + ((((float32*)a_2[((k0.k1.fused.k2.fused.outer1024) + threadIdx.x)] + (float32)b_2[floormod(((k0.k1.fused.k2.fused.outer1024) + threadIdx.x), 100)]) + ((float32)a_2[((k0.k1.fused.k2.fused.outer1024) + threadIdx.x)](float32*)b_2[floormod(((k0.k1.fused.k2.fused.outer1024) + threadIdx.x), 100)]))0.5f32))

}

attr [meta[mReducer][0]] “reduce_scope” = @tir.reinterpret(0u64, dtype=handle);

@tir.tvm_thread_allreduce(1u32, (float32)T_divide_red.rf[0], True, reduce_temp0, threadIdx.x, dtype=handle)

if (threadIdx.x == 0) {

T_divide_red[0] = (float32)reduce_temp0[0]

}

计算的预定阶段已经累积，可以通过以下方式检查。

print(sg.stages)

输出：

[stage(a, placeholder(a, 0xd9c0fa00)), stage(b, placeholder(b, 0xe225cf70)), stage(T_add, compute(T_add, body=[(a[ax0, ax1, ax2] + b[ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min=0, ext=10))], reduce_axis=[], tag=broadcast, attrs={})), stage(T_multiply, compute(T_multiply, body=[(a[ax0, ax1, ax2]b[ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min=0, ext=10))], reduce_axis=[], tag=broadcast, attrs={})), stage(T_elemwise_sum, compute(T_elemwise_sum, body=[(T_add[ax0, ax1, ax2] + T_multiply[ax0, ax1, ax2])], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min=0, ext=10))], reduce_axis=[], tag=elemwise, attrs={})), stage(T_divide, compute(T_divide, body=[(T_elemwise_sum[ax0, ax1, ax2]/2f)], axis=[iter_var(ax0, range(min=0, ext=100)), iter_var(ax1, range(min=0, ext=10)), iter_var(ax2, range(min=0, ext=10))], reduce_axis=[], tag=elemwise, attrs={})), stage(T_divide_red.rf, compute(T_divide_red.rf, body=[reduce(combiner=comm_reducer(result=[(x + y)], lhs=[x], rhs=[y], identity_element=[0f]), source=[T_divide[floordiv(floordiv((k0.k1.fused.k2.fused.inner + (k0.k1.fused.k2.fused.outer1024)), 10), 10), floormod(floordiv((k0.k1.fused.k2.fused.inner + (k0.k1.fused.k2.fused.outer1024)), 10), 10), floormod((k0.k1.fused.k2.fused.inner + (k0.k1.fused.k2.fused.outer1024)), 10)]], init=[], axis=[iter_var(k0.k1.fused.k2.fused.outer, range(min=0, ext=10))], where=tir.likely((((floordiv(floordiv((k0.k1.fused.k2.fused.inner + (k0.k1.fused.k2.fused.outer1024)), 10), 10) < 100) && (floordiv((k0.k1.fused.k2.fused.inner + (k0.k1.fused.k2.fused.outer1024)), 10) < 1000)) && ((k0.k1.fused.k2.fused.inner + (k0.k1.fused.k2.fused.outer1024)) < 10000))), value_index=0)], axis=[iter_var(k0.k1.fused.k2.fused.inner, range(min=0, ext=1024))], reduce_axis=[iter_var(k0.k1.fused.k2.fused.outer, range(min=0, ext=10))], tag=, attrs={})), stage(T_divide_red, compute(T_divide_red.repl, body=[reduce(combiner=comm_reducer(result=[(x + y)], lhs=[x], rhs=[y], identity_element=[0f]), source=[T_divide_red.rf[k0.k1.fused.k2.fused.inner.v]], init=[], axis=[iter_var(k0.k1.fused.k2.fused.inner.v, range(min=0, ext=1024))], where=(bool)1, value_index=0)], axis=[], reduce_axis=[iter_var(k0.k1.fused.k2.fused.inner.v, range(min=0, ext=1024))], tag=, attrs={}))]

可以通过与numpy结果进行比较，测试正确性，如下所示。

func = tvm.build(sg, [a, b, g], “cuda”)

dev = tvm.cuda(0)

a_np = np.random.uniform(size=(x, y, y)).astype(a.dtype)

b_np = np.random.uniform(size=(y, y)).astype(b.dtype)

g_np = np.sum(np.add(a_np + b_np, a_np * b_np) / 2.0)

a_nd = tvm.nd.array(a_np, dev)

b_nd = tvm.nd.array(b_np, dev)

g_nd = tvm.nd.array(np.zeros(g_np.shape, dtype=g_np.dtype), dev)

func(a_nd, b_nd, g_nd)

tvm.testing.assert_allclose(g_nd.numpy(), g_np, rtol=1e-5)

TOPI 提供常用的神经网络操作，如softmax优化调度

tarray = te.placeholder((512, 512), name=“tarray”)

softmax_topi = topi.nn.softmax(tarray)

with tvm.target.Target(“cuda”):

sst = topi.cuda.schedule_softmax(softmax_topi)

print(tvm.lower(sst, [tarray], simple_mode=True))

输出：

primfn(tarray_1: handle) -> ()

attr = {“from_legacy_te_schedule”: True, “global_symbol”: “main”, “tir.noalias”: True}

buffers = {tarray: Buffer(tarray_2: Pointer(float32), float32, [512, 512], [])}

buffer_map = {tarray_1: tarray} {

allocate(T_softmax_norm: Pointer(global float32x4), float32x4, [65536]), storage_scope = global;

attr [IterVar(blockIdx.x: int32, (nullptr), “ThreadIndex”, “blockIdx.x”)] “thread_extent” = 512;

allocate(normal_reduce_temp0: Pointer(local float32), float32, [1]), storage_scope = local;

allocate(reduce_temp0: Pointer(local float32), float32, [1]), storage_scope = local;

allocate(T_softmax_exp: Pointer(warp float32), float32, [512]), storage_scope = warp;

allocate(normal_reduce_temp0_1: Pointer(local float32), float32, [1]), storage_scope = local;

allocate(reduce_temp0_1: Pointer(local float32), float32, [1]), storage_scope = local {

attr [IterVar(threadIdx.x: int32, [0:32], “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 32 {

normal_reduce_temp0[0] = -3.40282e+38f32

for (k.inner: int32, 0, 16) {

normal_reduce_temp0[0] = max((float32)normal_reduce_temp0[0], (float32*)tarray_2[(((blockIdx.x512) + (threadIdx.x16)) + k.inner)])

}

attr [meta[mReducer][0]] “reduce_scope” = @tir.reinterpret(0u64, dtype=handle);

@tir.tvm_thread_allreduce(1u32, (float32*)normal_reduce_temp0[0], True, reduce_temp0, threadIdx.x, dtype=handle)

for (i1.inner.outer: int32, 0, 4) {

T_softmax_exp[ramp(((threadIdx.x16) + (i1.inner.outer4)), 1, 4)] = @tir.exp(((float32x4*)tarray_2[ramp((((blockIdx.x512) + (threadIdx.x16)) + (i1.inner.outer4)), 1, 4)] - broadcast((float32)reduce_temp0[0], 4)), dtype=float32x4)

}

attr [IterVar(threadIdx.x, [0:32], “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 32 {

normal_reduce_temp0_1[0] = 0f32

for (k.inner_1: int32, 0, 16) {

normal_reduce_temp0_1[0] = ((float32*)normal_reduce_temp0_1[0] + (float32*)T_softmax_exp[((threadIdx.x16) + k.inner_1)])

}

attr [meta[mReducer][1]] “reduce_scope” = @tir.reinterpret(0u64, dtype=handle);

@tir.tvm_thread_allreduce(1u32, (float32)normal_reduce_temp0_1[0], True, reduce_temp0_1, threadIdx.x, dtype=handle)

for (i1.inner.outer_1: int32, 0, 4) {

T_softmax_norm[ramp((((blockIdx.x512) + (threadIdx.x16)) + (i1.inner.outer_14)), 1, 4)] = ((float32x4)T_softmax_exp[ramp(((threadIdx.x16) + (i1.inner.outer_14)), 1, 4)] / broadcast((float32*)reduce_temp0_1[0], 4))

}

融合卷积

可以融合topi.nn.conv2d和topi.nn.relu在一起。

TOPI 函数都是通用函数。对不同的后端，有不同的实现优化性能。对于每个后端，有必要在计算声明和调度的目标范围内调用。TVM 将选择正确的函数，调用目标信息。

data = te.placeholder((1, 3, 224, 224))

kernel = te.placeholder((10, 3, 5, 5))

with tvm.target.Target(“cuda”):

conv = topi.cuda.conv2d_nchw(data, kernel, 1, 2, 1)

out = topi.nn.relu(conv)

sconv = topi.cuda.schedule_conv2d_nchw([out])

print(tvm.lower(sconv, [data, kernel], simple_mode=True))

Out:

primfn(placeholder_2: handle, placeholder_3: handle) -> ()

attr = {“from_legacy_te_schedule”: True, “global_symbol”: “main”, “tir.noalias”: True}

buffers = {placeholder_1: Buffer(placeholder_4: Pointer(float32), float32, [10, 3, 5, 5], []),

placeholder: Buffer(placeholder_5: Pointer(float32), float32, [1, 3, 224, 224], [])}

buffer_map = {placeholder_2: placeholder, placeholder_3: placeholder_1} {

allocate(compute: Pointer(global float32), float32, [501760]), storage_scope = global;

attr [IterVar(blockIdx.z: int32, (nullptr), “ThreadIndex”, “blockIdx.z”)] “thread_extent” = 5;

allocate(compute_1: Pointer(local float32), float32, [14]), storage_scope = local;

allocate(pad_temp.shared: Pointer(shared float32), float32, [112]), storage_scope = shared;

allocate(placeholder.shared: Pointer(shared float32), float32, [2]), storage_scope = shared;

attr [IterVar(blockIdx.y: int32, (nullptr), “ThreadIndex”, “blockIdx.y”)] “thread_extent” = 224;

attr [IterVar(blockIdx.x: int32, (nullptr), “ThreadIndex”, “blockIdx.x”)] “thread_extent” = 2;

attr [IterVar(threadIdx.z: int32, (nullptr), “ThreadIndex”, “threadIdx.z”)] “thread_extent” = 1;

attr [IterVar(threadIdx.y: int32, (nullptr), “ThreadIndex”, “threadIdx.y”)] “thread_extent” = 1;

attr [IterVar(threadIdx.x: int32, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 16 {

compute_1[0] = 0f32

compute_1[2] = 0f32

compute_1[4] = 0f32

compute_1[6] = 0f32

compute_1[8] = 0f32

compute_1[10] = 0f32

compute_1[12] = 0f32

compute_1[1] = 0f32

compute_1[3] = 0f32

compute_1[5] = 0f32

compute_1[7] = 0f32

compute_1[9] = 0f32

compute_1[11] = 0f32

compute_1[13] = 0f32

for (rc.outer: int32, 0, 3) {

for (ry.outer: int32, 0, 5) {

attr [IterVar(threadIdx.z_1: int32, (nullptr), “ThreadIndex”, “threadIdx.z”)] “thread_extent” = 1;

attr [IterVar(threadIdx.y_1: int32, (nullptr), “ThreadIndex”, “threadIdx.y”)] “thread_extent” = 1;

attr [IterVar(threadIdx.x_1: int32, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 16 {

pad_temp.shared[(threadIdx.x_17)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (2 <= ((blockIdx.x112) + (threadIdx.x_17)))), (float32)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 450)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 1)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x112) + (threadIdx.x_17)))), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 449)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 2)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 448)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 3)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 447)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 4)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 446)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 5)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 445)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 6)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 444)], 0f32, dtype=float32)

}

attr [IterVar(threadIdx.z_2: int32, (nullptr), “ThreadIndex”, “threadIdx.z”)] “thread_extent” = 1;

attr [IterVar(threadIdx.y_2: int32, (nullptr), “ThreadIndex”, “threadIdx.y”)] “thread_extent” = 1;

attr [IterVar(threadIdx.x_2: int32, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 16;

if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {

placeholder.shared[threadIdx.x_2] = (float32)placeholder_4[((((blockIdx.z150) + (threadIdx.x_275)) + (rc.outer25)) + (ry.outer5))]

}

compute_1[0] = ((float32*)compute_1[0] + ((float32*)pad_temp.shared[threadIdx.x](float32)placeholder.shared[0]))

compute_1[2] = ((float32*)compute_1[2] + ((float32*)pad_temp.shared[(threadIdx.x + 16)](float32)placeholder.shared[0]))

compute_1[4] = ((float32*)compute_1[4] + ((float32*)pad_temp.shared[(threadIdx.x + 32)](float32)placeholder.shared[0]))

compute_1[6] = ((float32*)compute_1[6] + ((float32*)pad_temp.shared[(threadIdx.x + 48)](float32)placeholder.shared[0]))

compute_1[8] = ((float32*)compute_1[8] + ((float32*)pad_temp.shared[(threadIdx.x + 64)](float32)placeholder.shared[0]))

compute_1[10] = ((float32*)compute_1[10] + ((float32*)pad_temp.shared[(threadIdx.x + 80)](float32)placeholder.shared[0]))

compute_1[12] = ((float32*)compute_1[12] + ((float32*)pad_temp.shared[(threadIdx.x + 96)](float32)placeholder.shared[0]))

compute_1[1] = ((float32*)compute_1[1] + ((float32*)pad_temp.shared[threadIdx.x](float32)placeholder.shared[1]))

compute_1[3] = ((float32*)compute_1[3] + ((float32*)pad_temp.shared[(threadIdx.x + 16)](float32)placeholder.shared[1]))

compute_1[5] = ((float32*)compute_1[5] + ((float32*)pad_temp.shared[(threadIdx.x + 32)](float32)placeholder.shared[1]))

compute_1[7] = ((float32*)compute_1[7] + ((float32*)pad_temp.shared[(threadIdx.x + 48)](float32)placeholder.shared[1]))

compute_1[9] = ((float32*)compute_1[9] + ((float32*)pad_temp.shared[(threadIdx.x + 64)](float32)placeholder.shared[1]))

compute_1[11] = ((float32*)compute_1[11] + ((float32*)pad_temp.shared[(threadIdx.x + 80)](float32)placeholder.shared[1]))

compute_1[13] = ((float32*)compute_1[13] + ((float32*)pad_temp.shared[(threadIdx.x + 96)](float32)placeholder.shared[1]))

attr [IterVar(threadIdx.z_1, (nullptr), “ThreadIndex”, “threadIdx.z”)] “thread_extent” = 1;

attr [IterVar(threadIdx.y_1, (nullptr), “ThreadIndex”, “threadIdx.y”)] “thread_extent” = 1;

attr [IterVar(threadIdx.x_1, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 16 {

pad_temp.shared[(threadIdx.x_17)] = @tir.if_then_else((((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)) && (1 <= ((blockIdx.x112) + (threadIdx.x_17)))), (float32)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 449)], 0f32, dtype=float32)

pad_temp.shared[((threadIdx.x_17) + 1)] = @tir.if_then_else(((2 <= (blockIdx.y + ry.outer)) && ((blockIdx.y + ry.outer) < 226)), (float32*)placeholder_5[((((((rc.outer50176) + (blockIdx.y224)) + (ry.outer224)) + (blockIdx.x112)) + (threadIdx.x_17)) - 448)], 0f32, dtype=float32)

}

attr [IterVar(threadIdx.z_2, (nullptr), “ThreadIndex”, “threadIdx.z”)] “thread_extent” = 1;

attr [IterVar(threadIdx.y_2, (nullptr), “ThreadIndex”, “threadIdx.y”)] “thread_extent” = 1;

attr [IterVar(threadIdx.x_2, (nullptr), “ThreadIndex”, “threadIdx.x”)] “thread_extent” = 16;

if @tir.likely((threadIdx.x_2 < 2), dtype=bool) {

placeholder.shared[threadIdx.x_2] = (float32)placeholder_4[(((((blockIdx.z150) + (threadIdx.x_275)) + (rc.outer25)) + (ry.outer5)) + 1)]