___                    ___          ___     
                       /\__\                  /\  \        /\  \   
                      /:/  /       ___        \:\  \       \:\  \   
                     /:/  /       /\__\        \:\  \       \:\  \  
                    /:/  /  ___  /:/__/    _____\:\  \  _____\:\  \ 
                   /:/__/  /\__\/::\  \   /::::::::\__\/::::::::\__\
                   \:\  \ /:/  /\/\:\  \__\:\~~\~~\/__/\:\~~\~~\/__/
                    \:\  /:/  /    \:\/\__\\:\  \       \:\  \      
                     \:\/:/  /      \::/  / \:\  \       \:\  \     
                      \::/  /       /:/  /   \:\__\       \:\__\   
                       \/__/        \/__/     \/__/        \/__/   

Installation Guide | Roadmap

The project CINN is a machine learning compiler and executor for multiple hardware backends. It is designed to provide multiple layers of APIs to make tensor computation easier to define, faster to execute, and more convenient to extend with hardware backends. Currently, it targets x86 CPUs and Nvidia GPUs.

This project is under active development.

Let's take python APIs as examples, the corresponding C++ APIs are available and differ little.

You can load a paddle model directly using CINN.

# Load Model to CINN
computation = Computation.compile_paddle_model(
    target = DefaultHostTarget(), model_dir = "./ResNet50", input_tensors = ['inputs'], input_sapes = [[1, 3, 224, 224]], params_combined = True)

# Get input tensor and set input data
a_t = computation.get_tensor(input_tensor)
a_t.from_numpy(np.random.random(x_shape).astype("float32"), target)

You can build and run a model by using NetBuilder APIs. Each NetBuilder API is a paddle operator.

# Define the NetBuilder.
builder = frontend.NetBuilder(name="network")

# Define the input variables of the model
a = builder.create_input(type=common.Float(32), shape=(1, 3, 224, 224), id_hint="A")
b = builder.create_input(type=common.Float(32), shape=(1, 3, 224, 224), id_hint="B")

# Build the model using NetBuilder API
y = builder.add(a, b)
res = builder.relu(y)

# Specify target and generate the computation
target = common.DefaultHostTarget()
computation = Computation.build_and_compile(target, builder)

The following is a naive matrix-multiplication implementation using the CINN DSL

#include "cinn/cinn.h"
using namespace cinn;

// Declare constants
Expr M(10), N(20), K(30);

// Declare the inputs 
auto A = Placeholder<float>("A", {M, K});
auto B = Placeholder<float>("B", {K, N});

auto k1 = Var(K.as_int32(), "k1");
auto C  = Compute(
    {M, N}, [&](Var i, Var j) { return ReduceSum(A(i, k1) * B(k1, j), {k1}); }, "C");

Target target = common::DefaultHostTarget();

int block_size = 32;

// The stages holds all the schedules for each tensors.
auto stages = CreateStages({C});

// Blocking optimization by loop tiling stragety.
auto [i_outer, i_inner, j_outer, j_inner] = stages[C]->Tile(0, 1, bn, bn);
auto [k_outer, k_inner]                   = stages[C]->Split("k0", 4);
stages[C]->Reorder({i_outer, j_outer, k_outer, k_inner, i_inner, j_inner});

// Generate C source code:
Module::Builder builder("module_block", target);
auto func = Lower("matmul_block", stages, {A, B, C});
builder.AddFunction(func);

CodeGenCX86 compiler(target, CodeGenCX86::Feature::AVX512);
Outputs outputs;
outputs = outputs.c_header("./test02_matmul_block.h").c_source("./test02_matmul_block.cc");
compiler.Compile(builder.Build(), outputs);

This can generate the optimized C source code like

void matmul_block(void* _args, int32_t num_args)
{
  const cinn_buffer_t* _A = cinn_pod_value_to_buffer_p(&(((cinn_pod_value_t*)(_args))[0]));
  const cinn_buffer_t* _B = cinn_pod_value_to_buffer_p(&(((cinn_pod_value_t*)(_args))[1]));
  cinn_buffer_t* _C = cinn_pod_value_to_buffer_p(&(((cinn_pod_value_t*)(_args))[2]));
  cinn_buffer_malloc((void*)(0), _C);
  const float* A = ((const float*)(_A->memory));
  const float* B = ((const float*)(_B->memory));
  float* C = ((float*)(_C->memory));
  float* C__reduce_init = ((float*)(_C->memory));
  for (int32_t i = 0; i < 1024; i += 1) {
    for (int32_t j = 0; j < 1024; j += 1) {
      C__reduce_init[((1024 * i) + j)] = 0;
    };
  };
  for (int32_t i_outer = 0; i_outer < 32; i_outer += 1) {
    for (int32_t j_outer = 0; j_outer < 32; j_outer += 1) {
      for (int32_t k0_outer = 0; k0_outer < 256; k0_outer += 1) {
        for (int32_t k0_inner = 0; k0_inner < 4; k0_inner += 1) {
          for (int32_t i_inner = 0; i_inner < 32; i_inner += 1) {
            for (int32_t j_inner = 0; j_inner < 32; j_inner += 1) {
              C[((1024 * i_inner) + ((32768 * i_outer) + ((32 * j_outer) + j_inner)))] = (C[((1024 * i_inner) + ((32768 * i_outer) + ((32 * j_outer) + j_inner)))] + (A[((1024 * i_inner) + ((32768 * i_outer) + ((4 * k0_outer) + k0_inner)))] * B[((32 * j_outer) + ((1024 * k0_inner) + ((4096 * k0_outer) + j_inner)))]));
            };
          };
        };
      };
    };
  };
  cinn_buffer_free((void*)(0), _C);
}

Change the CodeGenCX86 usage to CodeGenLLVM, it will produce a LLVM JIT-compiled function instead which can invoke realtime.

The CINN lowers a traditional DNN model into a two-level intermediate representation(IR), the high-level IR(HLIR) and CINN IR.

The HLIR helps to define some domain-specific computation and perform some overall optimization on the IR-graph; the CINN IR helps to represent some computation semantic and finally lower to a hardware backend.

Both levels of IR have the similar SSA graph, analysis and optimization facilities.

CINN is based on the polyhedral compilation thus it is easy to extend with more loop optimizations. The schedule transform is applied between the lowering from HLIR to CINN IR.

The overall architecture is as follows

Please refer to Installation Guidance and follow the guidance.

There are two levels of APIs in CINN, the higher level is HLIR and the lower level is CINN IR, both contain some concepts.

In HLIR

• Primitive Emitter(PE), encapsulates the computation of different tensor-based algorithms,
• frontend::Interpreter, the container to execute a model (of PaddlePaddle),
• frontend::Program, the program helps to define a machine learning computation,
• hlir::framework::Tensor, multi-dimensional arrays helps to manage a memory buffer.
• hlir::framework::Program, the final executable program in runtime. It holds many basic executable elements.
• hlir::framework::Graph, the graph that represents the structure of a model. Each node in the graph represents an operator (conv2d, relu, mul, etc.).
• hlir::framework::GraphCompiler, the compiler that transforms the graph representation(hlir::framework::Graph) of a model into an executable program(hlir::framework::Program).

In CINN IR

• Compute, the method to define a computation,
• Lower, the method to lower a computation to the corresponding IR,
• LoweredFunc, the function defined in CINN IR,
• Var, a scalar variable,
• Expr, an expression represents any CINN IR node(no specified Statement node),
• Stage, holds some schedule details of a tensor,

Read the code in the tests

For Python API, reference the code inside python/tests.

The C++ API locates in cinn/*/*_test.cc, the high level API locates in hlir/frontend, the lower level API is in cinn/cinn.h.

CINN is licensed under the Apache 2.0 license.

CINN learned a lot from the following projects:

• Halide: Referenced the design of most IR nodes,
• TVM: We learned many ideas including the semantics of some schedule primitives, TOPI, NNVM, and so on,
• tiramisu: The isl usage, polyhedral compilation, schedule primitive implementation, and so on,
• tensorflow/xla: Referenced the semantics of the primitive operations.