如何在CPU上优化GEMM矩阵乘法

如何在CPU上优化GEMM矩阵乘法

How to optimize GEMM on CPU

(TL;DR) TVM 提供抽象接口,允许用户分别描述算法和算法的实现组织(所谓的调度)。通常,在高性能调度中编写算法会破坏算法的可读性和模块化。此外,尝试各种看似有希望的调度也很耗时。在 TVM 的帮助下,可以有效地尝试这些调度以提高性能。

在本文中,将演示如何使用 TVM 优化方阵乘法,通过简单地添加 18 行额外代码实现比基线快 200 倍。

在 CPU 上执行的密集计算应用程序有两个重要的优化:

l  提高内存访问的缓存命中率。复杂的数值计算和热点内存访问都可以通过高缓存命中率加速。这需要将原始内存访问模式转换为适合缓存策略的模式。

l  SIMD(单指令多数据),或者称之为向量处理单元。每次都会处理一小批数据,不是单个网格。这需要以统一模式转换循环体中的数据访问模式,以便 LLVM 后端可以将其降低为 SIMD。

实际上,本文中使用的所有方法,都是本repo 中提到的技巧的一个子集 。其中一些已被 TVM 抽象自动应用,但由于 TVM 的限制,其中一些不能简单地应用。

下面提到的所有实验结果,都是在配备 Intel i7-4770HQ CPU 的 2015 年的 15' MacBook 上执行的。对于所有 x86 CPU,缓存行大小应为 64 字节。

准备和基线

在本文中,将演示如何使用 TVM 优化矩阵乘法。在实际演示前,先定义这些变量。然后编写一个基线实现,这是在 TVM 中编写矩阵乘法的最简单方法。

import tvm

import tvm.testing

from tvm import te

import numpy

import timeit

 

# The size of the matrix

# (M, K) x (K, N)

# You are free to try out different shapes, sometimes TVM optimization outperforms numpy with MKL.

M = 1024

K = 1024

N = 1024

# The default tensor type in tvm

dtype = "float32"

# using Intel AVX2(Advanced Vector Extensions) ISA for SIMD

# To get the best performance, please change the following line

# to llvm -mcpu=core-avx2, or specific type of CPU you use

target = "llvm"

dev = tvm.device(target, 0)

# Random generated tensor for testing

a = tvm.nd.array(numpy.random.rand(M, K).astype(dtype), dev)

b = tvm.nd.array(numpy.random.rand(K, N).astype(dtype), dev)

np_repeat = 100

np_runing_time = timeit.timeit(

    setup="import numpy\n"

    "M = " + str(M) + "\n"

    "K = " + str(K) + "\n"

    "N = " + str(N) + "\n"

    'dtype = "float32"\n'

    "a = numpy.random.rand(M, K).astype(dtype)\n"

    "b = numpy.random.rand(K, N).astype(dtype)\n",

    stmt="answer = numpy.dot(a, b)",

    number=np_repeat,

)

print("Numpy running time: %f" % (np_runing_time / np_repeat))

answer = numpy.dot(a.numpy(), b.numpy())

# Algorithm

k = te.reduce_axis((0, K), "k")

A = te.placeholder((M, K), name="A")

B = te.placeholder((K, N), name="B")

C = te.compute((M, N), lambda m, n: te.sum(A[m, k] * B[k, n], axis=k), name="C")

# Default schedule

s = te.create_schedule(C.op)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

 

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

 

evaluator = func.time_evaluator(func.entry_name, dev, number=1)

print("Baseline: %f" % evaluator(a, b, c).mean)

输出:

Numpy running time: 0.009229

Baseline: 3.340634

在 TVM 中,始终可以检查较低级别的 IR 以调试或优化调度。这是使用基线调度生成的 IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  for (m: int32, 0, 1024) {

    for (n: int32, 0, 1024) {

      C_2[((m*1024) + n)] = 0f32

      for (k: int32, 0, 1024) {

        C_2[((m*1024) + n)] = ((float32*)C_2[((m*1024) + n)] + ((float32*)A_2[((m*1024) + k)]*(float32*)B_2[((k*1024) + n)]))

      }

    }

  }

}

阻塞

提高缓存命中率的一个重要技巧是阻塞——数据块将逐块计算。块内部的内存访问是一个具有高内存局部性的小邻域。在本文中,选择了 32 作为分块因子。所以该块将填充32 * 32 * sizeof(float),即总大小为32KB的缓存中的4KB(L1数据缓存)

bn = 32

kfactor = 4

s = te.create_schedule(C.op)

# Blocking by loop tiling

mo, no, mi, ni = s[C].tile(C.op.axis[0], C.op.axis[1], bn, bn)

(kaxis,) = s[C].op.reduce_axis

ko, ki = s[C].split(kaxis, factor=kfactor)

# Hoist reduction domain outside the blocking loop

s[C].reorder(mo, no, ko, ki, mi, ni)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

# By simply tiling the loop 32x32, and hoisting ko, ki outside the blocking loops,

# we can see big speedup compared with the baseline.

evaluator = func.time_evaluator(func.entry_name, dev, number=10)

print("Opt1: %f" % evaluator(a, b, c).mean)

输出:

Opt1: 0.295032

这是阻塞后生成的IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  for (m.outer: int32, 0, 32) {

    for (n.outer: int32, 0, 32) {

      for (m.inner.init: int32, 0, 32) {

        for (n.inner.init: int32, 0, 32) {

          C_2[((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)) + n.inner.init)] = 0f32

        }

      }

      for (k.outer: int32, 0, 256) {

        for (k.inner: int32, 0, 4) {

          for (m.inner: int32, 0, 32) {

            for (n.inner: int32, 0, 32) {

              C_2[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = ((float32*)C_2[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] + ((float32*)A_2[((((m.outer*32768) + (m.inner*1024)) + (k.outer*4)) + k.inner)]*(float32*)B_2[((((k.outer*4096) + (k.inner*1024)) + (n.outer*32)) + n.inner)]))

            }

          }

        }

      }

    }

  }

}

矢量化

另一个重要的技巧是矢量化。当内存访问模式一致时,编译器可以检测到这种模式,将连续内存传递给向量处理器。在 TVM 中,可以使用vectorize接口,提示编译器这种模式,这样就可以大大加速。

在本文中,选择矢量化内循环行数据,因为对缓存友好。

s = te.create_schedule(C.op)

mo, no, mi, ni = s[C].tile(C.op.axis[0], C.op.axis[1], bn, bn)

(kaxis,) = s[C].op.reduce_axis

ko, ki = s[C].split(kaxis, factor=kfactor)

s[C].reorder(mo, no, ko, ki, mi, ni)

# Vectorization

s[C].vectorize(ni)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

evaluator = func.time_evaluator(func.entry_name, dev, number=10)

print("Opt2: %f" % evaluator(a, b, c).mean)

输出:

Opt2: 0.331193

这是矢量化后生成的 IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  for (m.outer: int32, 0, 32) {

    for (n.outer: int32, 0, 32) {

      for (m.inner.init: int32, 0, 32) {

        C_2[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)

      }

      for (k.outer: int32, 0, 256) {

        for (k.inner: int32, 0, 4) {

          for (m.inner: int32, 0, 32) {

            C_2[ramp((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)), 1, 32)] = ((float32x32*)C_2[ramp((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.inner*1024)) + (k.outer*4)) + k.inner)], 32)*(float32x32*)B_2[ramp((((k.outer*4096) + (k.inner*1024)) + (n.outer*32)), 1, 32)]))

          }

        }

      }

    }

  }

}

循环排列

如果查看上面的 IR,可以看到 B 和 C 的内循环行数据,都进行了向量化。接下来将查看 A 的访问模式。在当前调度中,A 是逐列访问的,这对缓存不友好. 如果改变 ki 和内轴 mi 的嵌套循环顺序,A 矩阵的访问模式对缓存更友好。

s = te.create_schedule(C.op)

mo, no, mi, ni = s[C].tile(C.op.axis[0], C.op.axis[1], bn, bn)

(kaxis,) = s[C].op.reduce_axis

ko, ki = s[C].split(kaxis, factor=kfactor)

# re-ordering

s[C].reorder(mo, no, ko, mi, ki, ni)

s[C].vectorize(ni)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

evaluator = func.time_evaluator(func.entry_name, dev, number=10)

print("Opt3: %f" % evaluator(a, b, c).mean)

输出:

Opt3: 0.113024

这是循环排列后生成的 IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  for (m.outer: int32, 0, 32) {

    for (n.outer: int32, 0, 32) {

      for (m.inner.init: int32, 0, 32) {

        C_2[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)

      }

      for (k.outer: int32, 0, 256) {

        for (m.inner: int32, 0, 32) {

          for (k.inner: int32, 0, 4) {

            C_2[ramp((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)), 1, 32)] = ((float32x32*)C_2[ramp((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.inner*1024)) + (k.outer*4)) + k.inner)], 32)*(float32x32*)B_2[ramp((((k.outer*4096) + (k.inner*1024)) + (n.outer*32)), 1, 32)]))

          }

        }

      }

    }

  }

}

数组打包

另一个重要的技巧是数组打包。诀窍是对多维数组的存储进行重新排序,以便在展平,存储在一维内存中后按顺序访问。

 

 

 注意:此图是阵列打包工作原理的一般说明。

可以使用数组打包来解决 B 的访问模式。观察扁平化后 B 的数组访问模式,当在 K 维度上迭代时,这不是连续的。可以用维度 [K][N] 重新排序 B,具有维度 [N/bn][K][bn],其中 bn 是阻塞因子,也是内循环中 B 的向量大小。这种重新排序将 N 分成两个维度 — bigN (N/bn) 和 littleN (bn) — 新维度 [N/bn][K][bn] 匹配 B 从外循环到内循环的索引(no, ko, ki, ni) ,在展平后导致 B 的顺序访问模式。

# We have to re-write the algorithm slightly.

packedB = te.compute(

    (N / bn, K, bn), lambda bigN, k, littleN: B[k, bigN * bn + littleN], name="packedB"

)

C = te.compute(

    (M, N),

    lambda m, n: te.sum(A[m, k] * packedB[n // bn, k, tvm.tir.indexmod(n, bn)], axis=k),

    name="C",

)

s = te.create_schedule(C.op)

mo, no, mi, ni = s[C].tile(C.op.axis[0], C.op.axis[1], bn, bn)

(kaxis,) = s[C].op.reduce_axis

ko, ki = s[C].split(kaxis, factor=kfactor)

s[C].reorder(mo, no, ko, mi, ki, ni)

s[C].vectorize(ni)

bigN, _, littleN = s[packedB].op.axis

s[packedB].vectorize(littleN)

s[packedB].parallel(bigN)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

evaluator = func.time_evaluator(func.entry_name, dev, number=10)

print("Opt4: %f" % evaluator(a, b, c).mean)

输出:

Opt4: 0.232269

这是阵列打包后生成的IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {

    for (bigN: int32, 0, 32) "parallel" {

      for (k: int32, 0, 1024) {

        packedB[ramp(((bigN*32768) + (k*32)), 1, 32)] = (float32x32*)B_2[ramp(((k*1024) + (bigN*32)), 1, 32)]

      }

    }

    for (m.outer: int32, 0, 32) {

      for (n.outer: int32, 0, 32) {

        for (m.inner.init: int32, 0, 32) {

          C_2[ramp((((m.outer*32768) + (m.inner.init*1024)) + (n.outer*32)), 1, 32)] = broadcast(0f32, 32)

        }

        for (k.outer: int32, 0, 256) {

          for (m.inner: int32, 0, 32) {

            for (k.inner: int32, 0, 4) {

              C_2[ramp((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)), 1, 32)] = ((float32x32*)C_2[ramp((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.inner*1024)) + (k.outer*4)) + k.inner)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + (k.inner*32)), 1, 32)]))

            }

          }

        }

      }

    }

  }

}

块的写缓存

阻塞后,程序将结果逐块写入C,访问模式不是顺序的。使用一个顺序缓存数组保存块结果,在所有块结果准备好时写入 C。

s = te.create_schedule(C.op)

# Allocate write cache

CC = s.cache_write(C, "global")

mo, no, mi, ni = s[C].tile(C.op.axis[0], C.op.axis[1], bn, bn)

# Write cache is computed at no

s[CC].compute_at(s[C], no)

# New inner axes

mc, nc = s[CC].op.axis

(kaxis,) = s[CC].op.reduce_axis

ko, ki = s[CC].split(kaxis, factor=kfactor)

s[CC].reorder(ko, mc, ki, nc)

s[CC].vectorize(nc)

# TODO: Add separate optimization step to discuss loop unrolloing

# unrolling is a loop optimization strategy which can reduce branch

# prediction failures and increases the chance of concurrent execution

# unroll kfactor loops

s[CC].unroll(ki)

bigN, _, littleN = s[packedB].op.axis

s[packedB].vectorize(littleN)

s[packedB].parallel(bigN)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

evaluator = func.time_evaluator(func.entry_name, dev, number=10)

print("Opt5: %f" % evaluator(a, b, c).mean)

输出:

Opt5: 0.225938

这是阻塞后生成的IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global;

  allocate(C.global: Pointer(global float32), float32, [1024]), storage_scope = global {

    for (bigN: int32, 0, 32) "parallel" {

      for (k: int32, 0, 1024) {

        packedB[ramp(((bigN*32768) + (k*32)), 1, 32)] = (float32x32*)B_2[ramp(((k*1024) + (bigN*32)), 1, 32)]

      }

    }

    for (m.outer: int32, 0, 32) {

      for (n.outer: int32, 0, 32) {

        for (m.c.init: int32, 0, 32) {

          C.global[ramp((m.c.init*32), 1, 32)] = broadcast(0f32, 32)

        }

        for (k.outer: int32, 0, 256) {

          for (m.c: int32, 0, 32) {

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[(((m.outer*32768) + (m.c*1024)) + (k.outer*4))], 32)*(float32x32*)packedB[ramp(((n.outer*32768) + (k.outer*128)), 1, 32)]))

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.c*1024)) + (k.outer*4)) + 1)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + 32), 1, 32)]))

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.c*1024)) + (k.outer*4)) + 2)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + 64), 1, 32)]))

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.c*1024)) + (k.outer*4)) + 3)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + 96), 1, 32)]))

          }

        }

        for (m.inner: int32, 0, 32) {

          for (n.inner: int32, 0, 32) {

            C_2[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = (float32*)C.global[((m.inner*32) + n.inner)]

          }

        }

      }

    }

  }

}

并行化

此外,还可以利用多核处理器进行线程级并行化。

s = te.create_schedule(C.op)

CC = s.cache_write(C, "global")

mo, no, mi, ni = s[C].tile(C.op.axis[0], C.op.axis[1], bn, bn)

s[CC].compute_at(s[C], no)

mc, nc = s[CC].op.axis

(kaxis,) = s[CC].op.reduce_axis

ko, ki = s[CC].split(kaxis, factor=kfactor)

s[CC].reorder(ko, mc, ki, nc)

s[CC].vectorize(nc)

s[CC].unroll(ki)

# parallel

s[C].parallel(mo)

bigN, _, littleN = s[packedB].op.axis

s[packedB].vectorize(littleN)

s[packedB].parallel(bigN)

func = tvm.build(s, [A, B, C], target=target, name="mmult")

assert func

c = tvm.nd.array(numpy.zeros((M, N), dtype=dtype), dev)

func(a, b, c)

tvm.testing.assert_allclose(c.numpy(), answer, rtol=1e-5)

evaluator = func.time_evaluator(func.entry_name, dev, number=50)

opt6_time = evaluator(a, b, c).mean

print("Opt6: %f" % opt6_time)

输出:

Opt6: 0.068730

这是并行化后生成的IR。

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

输出:

primfn(A_1: handle, B_1: handle, C_1: handle) -> ()

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

  buffers = {A: Buffer(A_2: Pointer(float32), float32, [1024, 1024], []),

             C: Buffer(C_2: Pointer(float32), float32, [1024, 1024], []),

             B: Buffer(B_2: Pointer(float32), float32, [1024, 1024], [])}

  buffer_map = {A_1: A, B_1: B, C_1: C} {

  allocate(packedB: Pointer(global float32x32), float32x32, [32768]), storage_scope = global {

    for (bigN: int32, 0, 32) "parallel" {

      for (k: int32, 0, 1024) {

        packedB[ramp(((bigN*32768) + (k*32)), 1, 32)] = (float32x32*)B_2[ramp(((k*1024) + (bigN*32)), 1, 32)]

      }

    }

    for (m.outer: int32, 0, 32) "parallel" {

      allocate(C.global: Pointer(global float32), float32, [1024]), storage_scope = global;

      for (n.outer: int32, 0, 32) {

        for (m.c.init: int32, 0, 32) {

          C.global[ramp((m.c.init*32), 1, 32)] = broadcast(0f32, 32)

        }

        for (k.outer: int32, 0, 256) {

          for (m.c: int32, 0, 32) {

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[(((m.outer*32768) + (m.c*1024)) + (k.outer*4))], 32)*(float32x32*)packedB[ramp(((n.outer*32768) + (k.outer*128)), 1, 32)]))

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.c*1024)) + (k.outer*4)) + 1)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + 32), 1, 32)]))

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.c*1024)) + (k.outer*4)) + 2)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + 64), 1, 32)]))

            C.global[ramp((m.c*32), 1, 32)] = ((float32x32*)C.global[ramp((m.c*32), 1, 32)] + (broadcast((float32*)A_2[((((m.outer*32768) + (m.c*1024)) + (k.outer*4)) + 3)], 32)*(float32x32*)packedB[ramp((((n.outer*32768) + (k.outer*128)) + 96), 1, 32)]))

          }

        }

        for (m.inner: int32, 0, 32) {

          for (n.inner: int32, 0, 32) {

            C_2[((((m.outer*32768) + (m.inner*1024)) + (n.outer*32)) + n.inner)] = (float32*)C.global[((m.inner*32) + n.inner)]

          }

        }

      }

    }

  }

}

概括

在仅用 18 行代码应用上述简单优化后,生成的代码可以使用 MKL实现numpy性能的60%。网页上的输出反映了非排他性 Docker 容器上的运行时间,因此不可靠。强烈建议运行本文,观察 TVM 实现的性能提升。

 

 

参考链接:

https://tvm.apache.org/docs/how_to/optimize_operators/opt_gemm.html

posted @ 2021-11-16 14:37  吴建明wujianming  阅读(473)  评论(0编辑  收藏  举报