Get Started with VTA

Author: Thierry Moreau

This is an introduction tutorial on how to use TVM to program the VTA design.

In this tutorial, we will demonstrate the basic TVM workflow to implement a vector addition on the VTA design’s vector ALU. This process includes specific scheduling transformations necessary to lower computation down to low-level accelerator operations.

To begin, we need to import TVM which is our deep learning optimizing compiler. We also need to import the VTA python package which contains VTA specific extensions for TVM to target the VTA design.

from __future__ import absolute_import, print_function

import os
import tvm
import vta
import numpy as np

Loading in VTA Parameters

VTA is a modular and customizable design. Consequently, the user is free to modify high-level hardware parameters that affect the hardware design layout. These parameters are specified in the vta_config.json file by their log2 values. These VTA parameters can be loaded with the vta.get_env function.

Finally, the TVM target is also specified in the vta_config.json file. When set to sim, execution will take place inside of a behavioral VTA simulator. If you want to run this tutorial on the Pynq FPGA development platform, follow the VTA Pynq-Based Testing Setup guide.

env = vta.get_env()

FPGA Programming

When targeting the Pynq FPGA development board, we need to configure the board with a VTA bitstream.

# We'll need the TVM RPC module and the VTA simulator module
from tvm import rpc
from tvm.contrib import util
from vta.testing import simulator

# We read the Pynq RPC host IP address and port number from the OS environment
host = os.environ.get("VTA_PYNQ_RPC_HOST", "192.168.2.99")
port = int(os.environ.get("VTA_PYNQ_RPC_PORT", "9091"))

# We configure both the bitstream and the runtime system on the Pynq
# to match the VTA configuration specified by the vta_config.json file.
if env.TARGET == "pynq":

    # Make sure that TVM was compiled with RPC=1
    assert tvm.module.enabled("rpc")
    remote = rpc.connect(host, port)

    # Reconfigure the JIT runtime
    vta.reconfig_runtime(remote)

    # Program the FPGA with a pre-compiled VTA bitstream.
    # You can program the FPGA with your own custom bitstream
    # by passing the path to the bitstream file instead of None.
    vta.program_fpga(remote, bitstream=None)

# In simulation mode, host the RPC server locally.
elif env.TARGET == "sim":
    remote = rpc.LocalSession()

Computation Declaration

As a first step, we need to describe our computation. TVM adopts tensor semantics, with each intermediate result represented as multi-dimensional array. The user needs to describe the computation rule that generates the output tensors.

In this example we describe a vector addition, which requires multiple computation stages, as shown in the dataflow diagram below. First we describe the input tensors A and B that are living in main memory. Second, we need to declare intermediate tensors A_buf and B_buf, which will live in VTA’s on-chip buffers. Having this extra computational stage allows us to explicitly stage cached reads and writes. Third, we describe the vector addition computation which will add A_buf to B_buf to produce C_buf. The last operation is a cast and copy back to DRAM, into results tensor C.

https://raw.githubusercontent.com/uwsaml/web-data/master/vta/tutorial/vadd_dataflow.png

Input Placeholders

We describe the placeholder tensors A, and B in a tiled data format to match the data layout requirements imposed by the VTA vector ALU.

For VTA’s general purpose operations such as vector adds, the tile size is (env.BATCH, env.BLOCK_OUT). The dimensions are specified in the vta_config.json configuration file and are set by default to a (1, 16) vector.

In addition, A and B’s data types also needs to match the env.acc_dtype which is set by the vta_config.json file to be a 32-bit integer.

# Output channel factor m - total 64 x 16 = 1024 output channels
m = 64
# Batch factor o - total 1 x 1 = 1
o = 1
# A placeholder tensor in tiled data format
A = tvm.placeholder((o, m, env.BATCH, env.BLOCK_OUT), name="A", dtype=env.acc_dtype)
# B placeholder tensor in tiled data format
B = tvm.placeholder((o, m, env.BATCH, env.BLOCK_OUT), name="B", dtype=env.acc_dtype)

Copy Buffers

One specificity of hardware accelerators, is that on-chip memory has to be explicitly managed. This means that we’ll need to describe intermediate tensors A_buf and B_buf that can have a different memory scope than the original placeholder tensors A and B.

Later in the scheduling phase, we can tell the compiler that A_buf and B_buf will live in the VTA’s on-chip buffers (SRAM), while A and B will live in main memory (DRAM). We describe A_buf and B_buf as the result of a compute operation that is the identity function. This can later be interpreted by the compiler as a cached read operation.

# A copy buffer
A_buf = tvm.compute((o, m, env.BATCH, env.BLOCK_OUT), lambda *i: A(*i), "A_buf")
# B copy buffer
B_buf = tvm.compute((o, m, env.BATCH, env.BLOCK_OUT), lambda *i: B(*i), "B_buf")

Vector Addition

Now we’re ready to describe the vector addition result tensor C, with another compute operation. The compute function takes the shape of the tensor, as well as a lambda function that describes the computation rule for each position of the tensor.

No computation happens during this phase, as we are only declaring how the computation should be done.

# Describe the in-VTA vector addition
C_buf = tvm.compute(
    (o, m, env.BATCH, env.BLOCK_OUT),
    lambda *i: A_buf(*i).astype(env.acc_dtype) + B_buf(*i).astype(env.acc_dtype),
    name="C_buf")

Casting the Results

After the computation is done, we’ll need to send the results computed by VTA back to main memory.

Note

Memory Store Restrictions

One specificity of VTA is that it only supports DRAM stores in the narrow env.inp_dtype data type format. This lets us reduce the data footprint for memory transfers (more on this in the basic matrix multiply example).

We perform one last typecast operation to the narrow input activation data format.

# Cast to output type, and send to main memory
C = tvm.compute(
    (o, m, env.BATCH, env.BLOCK_OUT),
    lambda *i: C_buf(*i).astype(env.inp_dtype),
    name="C")

This concludes the computation declaration part of this tutorial.

Scheduling the Computation

While the above lines describes the computation rule, we can obtain C in many ways. TVM asks the user to provide an implementation of the computation called schedule.

A schedule is a set of transformations to an original computation that transforms the implementation of the computation without affecting correctness. This simple VTA programming tutorial aims to demonstrate basic schedule transformations that will map the original schedule down to VTA hardware primitives.

Default Schedule

After we construct the schedule, by default the schedule computes C in the following way:

# Let's take a look at the generated schedule
s = tvm.create_schedule(C.op)

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

Out:

// attr [A_buf] storage_scope = "global"
allocate A_buf[int32 * 1024]
// attr [B_buf] storage_scope = "global"
allocate B_buf[int32 * 1024]
produce A_buf {
  for (i1, 0, 64) {
    for (i3, 0, 16) {
      A_buf[((i1*16) + i3)] = A[((i1*16) + i3)]
    }
  }
}
produce B_buf {
  for (i1, 0, 64) {
    for (i3, 0, 16) {
      B_buf[((i1*16) + i3)] = B[((i1*16) + i3)]
    }
  }
}
produce C_buf {
  for (i1, 0, 64) {
    for (i3, 0, 16) {
      A_buf[((i1*16) + i3)] = (A_buf[((i1*16) + i3)] + B_buf[((i1*16) + i3)])
    }
  }
}
produce C {
  for (i1, 0, 64) {
    for (i3, 0, 16) {
      C[((i1*16) + i3)] = int8(A_buf[((i1*16) + i3)])
    }
  }
}

Although this schedule makes sense, it won’t compile to VTA. In order to obtain correct code generation, we need to apply scheduling primitives and code annotation that will transform the schedule into one that can be directly lowered onto VTA hardware intrinsics. Those include:

  • DMA copy operations which will take globally-scoped tensors and copy those into locally-scoped tensors.
  • Vector ALU operations that will perform the vector add.

Buffer Scopes

First, we set the scope of the copy buffers to indicate to TVM that these intermediate tensors will be stored in the VTA’s on-chip SRAM buffers. Below, we tell TVM that A_buf, B_buf, C_buf will live in VTA’s on-chip accumulator buffer which serves as VTA’s general purpose register file.

Set the intermediate tensors’ scope to VTA’s on-chip accumulator buffer

s[A_buf].set_scope(env.acc_scope)
s[B_buf].set_scope(env.acc_scope)
s[C_buf].set_scope(env.acc_scope)

DMA Transfers

We need to schedule DMA transfers to move data living in DRAM to and from the VTA on-chip buffers. We insert dma_copy pragmas to indicate to the compiler that the copy operations will be performed in bulk via DMA, which is common in hardware accelerators.

# Tag the buffer copies with the DMA pragma to map a copy loop to a
# DMA transfer operation
s[A_buf].pragma(s[A_buf].op.axis[0], env.dma_copy)
s[B_buf].pragma(s[B_buf].op.axis[0], env.dma_copy)
s[C].pragma(s[C].op.axis[0], env.dma_copy)

ALU Operations

VTA has a vector ALU that can perform vector operations on tensors in the accumulator buffer. In order to tell TVM that a given operation needs to be mapped to the VTA’s vector ALU, we need to explicitly tag the vector addition loop with an env.alu pragma.

# Tell TVM that the computation needs to be performed
# on VTA's vector ALU
s[C_buf].pragma(C_buf.op.axis[0], env.alu)

# Let's take a look at the finalized schedule
print(vta.lower(s, [A, B, C], simple_mode=True))

Out:

// attr [A_buf] storage_scope = "local.acc_buffer"
// attr [iter_var(vta, , vta)] coproc_scope = 2
produce A_buf {
  VTALoadBuffer2D(tvm_thread_context(VTATLSCommandHandle()), A, 0, 64, 1, 64, 0, 0, 0, 0, 0, 3)
}
produce B_buf {
  VTALoadBuffer2D(tvm_thread_context(VTATLSCommandHandle()), B, 0, 64, 1, 64, 0, 0, 0, 0, 64, 3)
}
// attr [iter_var(vta, , vta)] coproc_uop_scope = "VTAPushALUOp"
produce C_buf {
  VTAUopLoopBegin(64, 1, 1, 0)
  VTAUopPush(1, 0, 0, 64, 0, 2, 0, 0)
  VTAUopLoopEnd()
}
vta.coproc_dep_push(2, 3)
// attr [iter_var(vta, , vta)] coproc_scope = 3
vta.coproc_dep_pop(2, 3)
produce C {
  VTAStoreBuffer2D(tvm_thread_context(VTATLSCommandHandle()), 0, 4, C, 0, 64, 1, 64)
}
vta.coproc_sync()

This concludes the scheduling portion of this tutorial.

TVM Compilation

After we have finished specifying the schedule, we can compile it into a TVM function. By default TVM compiles into a type-erased function that can be directly called from python side.

In the following line, we use tvm.build to create a function. The build function takes the schedule, the desired signature of the function(including the inputs and outputs) as well as target language we want to compile to.

my_vadd = vta.build(s, [A, B, C], "ext_dev", env.target_host, name="my_vadd")

Saving the Module

TVM lets us save our module into a file so it can loaded back later. This is called ahead-of-time compilation and allows us to save some compilation time. More importantly, this allows us to cross-compile the executable on our development machine and send it over to the Pynq FPGA board over RPC for execution.

# Write the compiled module into an object file.
temp = util.tempdir()
my_vadd.save(temp.relpath("vadd.o"))

# Send the executable over RPC
remote.upload(temp.relpath("vadd.o"))

Loading the Module

We can load the compiled module from the file system to run the code.

f = remote.load_module("vadd.o")

Running the Function

The compiled TVM function uses a concise C API and can be invoked from any language.

TVM provides an array API in python to aid quick testing and prototyping. The array API is based on DLPack standard.

  • We first create a remote context (for remote execution on the Pynq).
  • Then tvm.nd.array formats the data accordingly.
  • f() runs the actual computation.
  • asnumpy() copies the result array back in a format that can be interpreted.
# Get the remote device context
ctx = remote.ext_dev(0)

# Initialize the A and B arrays randomly in the int range of (-128, 128]
A_orig = np.random.randint(
    -128, 128, size=(o * env.BATCH, m * env.BLOCK_OUT)).astype(A.dtype)
B_orig = np.random.randint(
    -128, 128, size=(o * env.BATCH, m * env.BLOCK_OUT)).astype(B.dtype)

# Apply packing to the A and B arrays from a 2D to a 4D packed layout
A_packed = A_orig.reshape(
    o, env.BATCH, m, env.BLOCK_OUT).transpose((0, 2, 1, 3))
B_packed = B_orig.reshape(
    o, env.BATCH, m, env.BLOCK_OUT).transpose((0, 2, 1, 3))

# Format the input/output arrays with tvm.nd.array to the DLPack standard
A_nd = tvm.nd.array(A_packed, ctx)
B_nd = tvm.nd.array(B_packed, ctx)
C_nd = tvm.nd.array(np.zeros((o, m, env.BATCH, env.BLOCK_OUT)).astype(C.dtype), ctx)

# Invoke the module to perform the computation
f(A_nd, B_nd, C_nd)

Verifying Correctness

Compute the reference result with numpy and assert that the output of the matrix multiplication indeed is correct

# Compute reference result with numpy
C_ref = (A_orig.astype(env.acc_dtype) + B_orig.astype(env.acc_dtype)).astype(C.dtype)
C_ref = C_ref.reshape(
    o, env.BATCH, m, env.BLOCK_OUT).transpose((0, 2, 1, 3))
np.testing.assert_equal(C_ref, C_nd.asnumpy())
print("Successful vector add test!")

Out:

Successful vector add test!

Summary

This tutorial provides a walk-through of TVM for programming the deep learning accelerator VTA with a simple vector addition example. The general workflow includes:

  • Programming the FPGA with the VTA bitstream over RPC.
  • Describing the vector add computation via a series of computations.
  • Describing how we want to perform the computation using schedule primitives.
  • Compiling the function to the VTA target.
  • Running the compiled module and verifying it against a numpy implementation.

You are more than welcome to check other examples out and tutorials to learn more about the supported operations, schedule primitives and other features supported by TVM to program VTA.

Total running time of the script: ( 0 minutes 0.229 seconds)

Gallery generated by Sphinx-Gallery