ispc: A SPMD Compiler for High-Performance CPU Programming
ispc：用于高性能 CPU 编程的 SPMD 编译器

Categories and Subject Descriptors
类别和主题描述符

Keywords 关键词

1. INTRODUCTION 1.引言

• Explicit language support for both scalar and SIMD operations.
  为标量和 SIMD 运算提供明确的语言支持。
• Support for structure-of-arrays data structures, including for converting previously-declared data types into structure of arrays layout.
  支持数组结构数据结构，包括将先前声明的数据类型转换为数组结构布局。
• Access to the full flexibility of the underlying CPU hardware, including the ability to launch asynchronous tasks and to perform fast cross-lane SIMD operations.
  访问底层 CPU 硬件的全部灵活性，包括启动异步任务和执行快速跨行 SIMD 操作的能力。

• Support for tight coupling between and ispc, including the ability to directly call ispc routines from and to also call routines from ispc.
  支持 和 ispc 之间的紧密耦合，包括从 直接调用 ispc 例程，以及从 ispc 调用 例程。
• Coherent shared memory between C++ and ispc.
  C++ 和 ispc 之间的连贯共享内存
• Familiar syntax and language features due to its basis in C.
  以 C 语言为基础，语法和语言特点非常熟悉。

2. DESIGN GOALS 2.设计目标

2.1 Goals 2.1 目标

2.2 Non-goals 2.2 非目标

2.3 Target Hardware 2.3 目标硬件

3. PARALLELISM MODEL: SPMD ON SIMD
3.并行模型：SPMD 上的 SIMD

3.1 Why SPMD? 3.1 为什么选择 SPMD？

3.2 Basic Execution Model
3.2 基本执行模式

3.3 Mapping SPMD To Hardware: Control
3.3 将 SPMD 映射到硬件：控制

3.4 SPMD and Synchronization
3.4 SPMD 和同步

3.5 Mapping SPMD To Hardware: Memory
3.5 将 SPMD 映射到硬件：内存

4. LANGUAGE OVERVIEW 4.语言概述

float *values = new float[1024];
int *iterations = new int[1024];
// ... initialize values[], iterations[] ...
update(values, iterations, 1024);

export void update(uniform float values[],
                    uniform int iterations[],
                    uniform int num_values) {
    for (int i = programIndex; i < num_values;
        i += programCount) {
        int iters = iterations[i];
        while (iters-- > 0)
            values[i] *= values[i];
    }
}

4.1 Mapping Computation to Data
4.1 将计算映射到数据

4.2 Implementation 4.2 实施

5. DESIGN FEATURES 5.设计特点

5.1 "Uniform" Datatypes 5.1 "统一 "数据类型

• Interoperability with C/C++ data structures: uniform is necessary to explicitly declare in-memory variables of just a single element, as is common in .
  与 C/C++ 数据结构的互操作性：有必要像 中常见的那样，明确声明仅包含单个元素的内存变量。
• Performance transparency: Treating uniform as an optimization rather than an explicit type property would make it difficult for the programmer to reason about performance. A small change to a program could inadvertently inhibit the optimization elsewhere resulting in significant and difficult-to-understand performance regressions.
  性能透明：将统一作为一种优化而非明确的类型属性，会使程序员难以推理性能。对程序稍作改动，就可能无意中抑制了其他地方的优化，导致难以理解的重大性能倒退。
• Support for separate compilation: Optimizations cannot cross separate-compilation boundaries, so at a minimum it must be possible to define a formal function parameter as uniform. But to guarantee that a call to such a function with a variable as an actual parameter is legal, uniform must be an explicit part of the type system. Otherwise, the legality of the function call would depend on the optimizer's behavior for the variable.
  支持单独编译：优化不能跨越单独编译的界限，因此至少要能将形式化函数参数定义为 uniform。但是，为了保证以变量作为实际参数的函数调用是合法的，"统一 "必须是类型系统的明确组成部分。否则，函数调用的合法性将取决于优化器对变量的处理。

5.2 Support For SOA Layout
5.2 支持 SOA 布局

struct Foo { float x, y, z; };
uniform Foo a[...] = { ... };
int index = ...;
float x = a[index].x;

struct Foo { float x, y, z; };
soa<4> struct Foo a[...] = { ... };
int index = ...;
float x = a[index].x;

uniform struct Foo a[\ldots...] = {...};
int index = ...;
varying Foo fv = a[index];
// now e.g. fv.x is contiguous in memory
fv.x = fv.y + fv.z; // looks scalar

5.3 Full C Pointer Model
5.3 完整的 C 指针模型

uniform float * uniform x;
varying float * uniform x;
uniform float * varying x;
varying float * varying x;

5.4 Task Launch 5.4 任务启动

5.5 Cross-lane operations
5.5 跨线运行

5.6 Coherent Control Flow Hints
5.6 相干控制流提示

float x = ...;
cif (x < 0) {
    // handle negative x
}

5.7 Native Object Files and Function Calls
5.7 本地对象文件和函数调用

6. EFFICIENT SPMD-ON-SIMD
6.高效的模拟数据管理

6.1 Benefits of "Uniform"
6.1 "统一 "的好处

• It reduces the total amount of in-memory storage used for data, which in turn can lead to better cache performance.
  它减少了用于数据的内存存储总量，从而提高了缓存性能。
• Less bandwidth is consumed when reading and writing scalar values to memory.
  将标量值读写到内存时消耗的带宽更少。
• CPUs have separate register sets for scalar and vector values; storing values in scalar registers when possible reduces pressure on vector registers.
  CPU 有独立的标量和矢量寄存器集；尽可能将数值存储在标量寄存器中，以减少对矢量寄存器的压力。
• CPUs can co-issue scalar and vector instructions, so that scalar and vector computations can happen concurrently.
  CPU 可以同时发出标量和矢量指令，因此标量和矢量计算可以同时进行。
• In the usual case of using 64-bit pointers, pointer arithmetic (e.g. for addressing calculations) is more efficient for scalar pointers than for vector pointers.
  在通常使用 64 位指针的情况下，标量指针的指针运算（例如寻址计算）比向量指针更有效率。
• Dereferencing a uniform pointer (or using a uniform value to index into an array) corresponds to a single scalar or vector memory access, rather than a general gather or scatter.
  取消引用统一指针（或使用统一值索引数组）对应于单个标量或矢量内存访问，而不是一般的聚集或分散。
• Code for control flow based on uniform quantities can be more efficient than code for control flow based on non-uniform quantities (Section 6.2).
  与基于非均匀量的控制流代码相比，基于均匀量的控制流代码效率更高（第 6.2 节）。

6.2 Uniform Control Flow 6.2 统一控制流

6.3 Benefits of SOA 6.3 SOA 的优势

6.4 Coherent Memory Access
6.4 相干内存访问

6.5 Dynamic Control Flow Coherence
6.5 动态控制流一致性

• It makes it possible to avoid executing instructions when the mask is "all off" and to jump over them.
  这样就可以避免在掩码 "全部关闭 "时执行指令，并跳过这些指令。
• It gives an opportunity for dynamically reestablishing that the mask is "all on" and then taking a specialized code path for that case; the advantages of doing so are discussed in the following subsection.
  它提供了一个动态重新确定掩码 "全开 "的机会，然后针对这种情况采取专门的代码路径；这样做的好处将在下一小节中讨论。

6.6 All On Mask 6.6 全靠面具

7. RESULTS 7.结果

7.1 Speedup Compared to Serial Code
7.1 与串行代码相比的速度提升

7.2 Speedup Versus Intrinsics
7.2 加速与内嵌字

7.3 Speedup with wider vectors
7.3 更宽矢量的提速

7.4 Scalability on Larger Systems
7.4 大型系统的可扩展性

7.5 Users 7.5 用户

8. RELATED WORK 8.相关工作

8.1 Historical systems 8.1 历史系统

8.2 Contemporary systems 8.2 现代系统

8.3 Concurrently-developed systems
8.3 同时开发的系统

9. CONCLUSION 9.结论

Acknowledgments 致谢

10. REFERENCES 10.参考文献