MatMul Tutorial: INT8 Inference

C++ API example demonstrating how one can use MatMul fused with ReLU in INT8 inference.

Concepts:

• Asymmetric quantization
  • Run-time output scales: dnnl::primitive_attr::set_output_scales() and DNNL_RUNTIME_F32_VAL
  • Run-time zero points: dnnl::primitive_attr::set_zero_points() and DNNL_RUNTIME_S32_VAL
• Operation fusion
• Create primitive once, use multiple times
  • Run-time tensor shapes: DNNL_RUNTIME_DIM_VAL
• Weights pre-packing: use dnnl::memory::format_tag::any

Assumptions:

1. The shape of the weights (matrix \(B(K, N)\)) is known in advance, the data type is int8_t and centered around 0 (i.e. the zero point is 0).
2. The shapes of the source matrix \(A\) and destination matrix \(C\) are partially unknown. Both matrices use uint8_t data type and might have arbitrary zero points (specified at execution time only).
3. Scaling (re-quantization) factor specified at run-time only.

Since the shape of weights is known in advance, the MatMul weights can be created with format tag dnnl::memory::format_tag::any to enable the library to choose the most appropriate layout for best performance.

Warning

The format tag dnnl::memory::format_tag::any doesn't work for memory descriptors that have one or more unknown dimensions and/or strides.

/*******************************************************************************
* Copyright 2019-2020 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/


#include <cassert>
#include <cctype>
#include <cmath>
#include <cstdio>
#include <iostream>
#include <random>
#include <stdexcept>
#include <vector>

#include "dnnl.hpp"

#include "example_utils.hpp"

using namespace dnnl;

namespace {

void init_vector(std::vector<float> &v) {
    std::mt19937 gen;
    std::uniform_real_distribution<float> u(0, 1);
    for (auto &e : v)
        e = u(gen);
}

void init_vector(std::vector<uint8_t> &v) {
    std::mt19937 gen;
    std::uniform_int_distribution<unsigned int> u(0, 255);
    for (auto &e : v)
        e = static_cast<uint8_t>(u(gen));
}

} // namespace

int number_of_runs = 1;

// Create a MatMul primitive descriptor for the following op:
// C_u8 = ReLU(scale[:] * (A_u8 - zp_A) * B_s8) + zp_C
//
// Here:
// - Matrices A and C are known to be non-transposed but their M dimension is
//   not known. They can be activation matrices in an MLP topology and the M
//   dimension can be the mini-batch dimension.
// - zp_A and zp_C are zero points for matrices A and C which are stored as
//   uint8_t. These are run-time parameters that are not known at the primitive
//   creation time.
// - The B matrix is stored as int8_t, its zero point is 0, and all its
//   dimensions are known. This matrix can be a matrix of weights in an MLP
//   topology.
// - The scaling values are not known at the primitive creation time.
matmul::primitive_desc matmul_pd_create(
        int64_t K, int64_t N, const engine &eng) {
    const int64_t M = DNNL_RUNTIME_DIM_VAL;

    memory::desc a_md({M, K}, memory::data_type::u8, {K, 1}); // M x K layout
    memory::desc b_md({K, N}, memory::data_type::s8, memory::format_tag::any);
    memory::desc c_md({M, N}, memory::data_type::u8, {N, 1}); // M x N layout

    // Create attributes and indicate that the alpha and zero points are
    // runtime parameters
    primitive_attr attr;
    attr.set_output_scales(/* mask */ (1 << 1), {DNNL_RUNTIME_F32_VAL});
    attr.set_zero_points(DNNL_ARG_SRC, /* mask */ 0, {DNNL_RUNTIME_S32_VAL});
    attr.set_zero_points(DNNL_ARG_DST, /* mask */ 0, {DNNL_RUNTIME_S32_VAL});
    post_ops po;
    po.append_eltwise(1.f, algorithm::eltwise_relu, 0.f, 0.f);
    attr.set_post_ops(po);

    // Create a MatMul primitive descriptor
    matmul::desc matmul_d(a_md, b_md, c_md);
    return matmul::primitive_desc(matmul_d, attr, eng);
}

void prepare_input(memory &A_u8_mem, memory &scale_f32_mem, memory &zp_A_mem,
        memory &zp_C_mem) {
    int64_t M = A_u8_mem.get_desc().dims()[0];
    int64_t N = scale_f32_mem.get_desc().dims()[0];
    int64_t K = A_u8_mem.get_desc().dims()[1];

    std::vector<uint8_t> A_u8(M * K);
    init_vector(A_u8);

    std::vector<float> scales_f32(N);
    init_vector(scales_f32);

    int32_t zp_A = 128, zp_C = 40;

    write_to_dnnl_memory(A_u8.data(), A_u8_mem);
    write_to_dnnl_memory(&zp_A, zp_A_mem);
    write_to_dnnl_memory(&zp_C, zp_C_mem);
    write_to_dnnl_memory(scales_f32.data(), scale_f32_mem);
}

void sanity_check(memory &C_u8_mem, memory &zp_C_mem) {
    int64_t M = C_u8_mem.get_desc().dims()[0];
    int64_t N = C_u8_mem.get_desc().dims()[1];
    int32_t zp_C = 0;
    std::vector<uint8_t> C_u8(M * N);

    read_from_dnnl_memory(C_u8.data(), C_u8_mem);
    read_from_dnnl_memory(&zp_C, zp_C_mem);

    // simple check: C_u8 >= zp_C
    for (int64_t i = 0; i < M * N; ++i)
        if (C_u8[i] < zp_C)
            throw std::logic_error(
                    "Smoke check failed."
                    "\n\tQuantized value is smaller than the zero point,"
                    "\n\twhich should not happen since ReLU was applied.");
}

void infer(const matmul &matmul_p, int64_t M, int64_t N, int64_t K,
        const memory &B_s8_mem, const engine &eng) {
    // inputs of the current layer / operation
    memory A_u8_mem({{M, K}, memory::data_type::u8, {K, 1}}, eng);
    memory zp_A_mem({{1}, memory::data_type::s32, {1}}, eng);
    memory zp_C_mem({{1}, memory::data_type::s32, {1}}, eng);
    memory scale_f32_mem({{N}, memory::data_type::f32, {1}}, eng);

    // the function below fills dnnl::memory with some values
    // these memories, typically, come from the previous layers / operations
    // with meaningful data inside
    prepare_input(A_u8_mem, scale_f32_mem, zp_A_mem, zp_C_mem);

    // output - no initialization required
    memory C_u8_mem({{M, N}, memory::data_type::u8, {N, 1}}, eng);

    stream s(eng);
    for (int run = 0; run < number_of_runs; ++run)
        matmul_p.execute(s,
                {{DNNL_ARG_SRC, A_u8_mem}, {DNNL_ARG_WEIGHTS, B_s8_mem},
                        {DNNL_ARG_DST, C_u8_mem},
                        {DNNL_ARG_ATTR_OUTPUT_SCALES, scale_f32_mem},
                        {DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC, zp_A_mem},
                        {DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_DST, zp_C_mem}});
    s.wait();

    // a sanity check for the correctness of the output
    sanity_check(C_u8_mem, zp_C_mem);
}

void inference_int8_matmul(engine::kind engine_kind) {
    engine eng(engine_kind, 0);

    const int64_t K = 96;
    const int64_t N = 1000;
    auto matmul_pd = matmul_pd_create(K, N, eng);

    // Original weights stored as float in a known format
    std::vector<float> B_f32(K * N);
    init_vector(B_f32);

    // Pre-packed weights stored as int8_t
    memory B_s8_mem(matmul_pd.weights_desc(), eng);
    {
        stream s(eng);
        memory B_f32_mem(
                {{K, N}, memory::data_type::f32, memory::format_tag::ab}, eng);
        write_to_dnnl_memory(B_f32.data(), B_f32_mem);
        reorder(B_f32_mem, B_s8_mem).execute(s, B_f32_mem, B_s8_mem);
        s.wait();
    }

    matmul matmul_p(matmul_pd);

    for (int64_t M : {1, 100})
        infer(matmul_p, M, N, K, B_s8_mem, eng);
}

int main(int argc, char **argv) {
    engine::kind engine_kind = parse_engine_kind(argc, argv);
    return handle_example_errors(inference_int8_matmul, engine_kind);
}