CUDA（19）之PTX优化原理

摘要

本文主要讲述CUDA中的PTX的原理实现和分析。

 

1. 不作优化的代码实现

Nvidia GTX 1050, CUDA 8.0测试代码如下：

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

#include 

__global__ void gpu(int *d_ptr, int length) {
	int elemID = blockIdx.x * blockDim.x + threadIdx.x;

	for (int innerloops = 0; innerloops < 100000; innerloops++) {
		if (elemID < length) {
			//unsigned int laneid;
			d_ptr[elemID] = elemID % 32;
		}
	}
}

void valid(int *h_ptr, int length) {
	for (int elemID = 0; elemID> >(d_ptr, N);
	cudaGetLastError();
	cudaDeviceSynchronize();

	// Finish timing
	cudaEventRecord(stop, 0);
	cudaEventSynchronize(start);
	cudaEventSynchronize(stop);
	cudaEventElapsedTime(&time_elapsed, start, stop);

	// Print
	printf("Time Used on GPU:%f(ms)\n", time_elapsed);

	// CPU (results for validate)
	valid(h_ptr, N);

	int *h_d_ptr;
	cudaMallocHost(&h_d_ptr, N * sizeof(int));
	cudaMemcpy(h_d_ptr, d_ptr, N * sizeof(int), cudaMemcpyDeviceToHost);

	bool bValid = true;

	for (int i = 0; i

Nvidia GTX 1050, CUDA 8.0测试结果如下：

Nvidia 780Ti, CUDA 7.5测试代码如下： 

#include 
#include 
#include 
#include 
#include 

__global__ void gpu(int *d_ptr, int length){
    int elemID = blockIdx.x * blockDim.x + threadIdx.x;

	for(int innerloops = 0; innerloops < 100000; innerloops++){
		if (elemID < length){
			//unsigned int laneid;
			d_ptr[elemID] = elemID % 32;
		}
	}
}

void valid(int *h_ptr, int length){
    for (int elemID=0; elemID>>(d_ptr, N);
    checkCudaErrors(cudaGetLastError());
    checkCudaErrors(cudaDeviceSynchronize());

	// Finish timing
	cudaEventRecord(stop,0);    
	cudaEventSynchronize(start);    
	cudaEventSynchronize(stop);   
	cudaEventElapsedTime(&time_elapsed,start,stop);

	// Print
	printf("Time Used on GPU:%f(ms)\n",time_elapsed);

	// CPU (results for validate)
    valid(h_ptr, N);

    int *h_d_ptr;
    checkCudaErrors(cudaMallocHost(&h_d_ptr, N *sizeof(int)));
    checkCudaErrors(cudaMemcpy(h_d_ptr, d_ptr, N *sizeof(int), cudaMemcpyDeviceToHost));

    bool bValid = true;

    for (int i=0; i

Nvidia 780Ti, CUDA 7.5测试结果如下：

2.  PTX

PTX主要特点是预计算的索引值放到L1缓存，实现内存操作的预取优化，达到本案例中性能的大幅提升。

Nvidia 780Ti, CUDA 7.5测试代码如下：

#include 
#include 
#include 
#include 
#include 


__global__ void gpu_ptx(int *d_ptr, int length){
    int elemID = blockIdx.x * blockDim.x + threadIdx.x;

	for(int innerloops = 0; innerloops < 100000; innerloops++){
		if (elemID < length){
			unsigned int laneid;
			asm("mov.u32 %0, %%laneid;" : "=r"(laneid)); // 索引缓存
			d_ptr[elemID] = laneid;
		}
	}
}

void valid(int *h_ptr, int length){
    for (int elemID=0; elemID>>(d_ptr, N);
    checkCudaErrors(cudaGetLastError());
    checkCudaErrors(cudaDeviceSynchronize());

	// Finish timing
	cudaEventRecord(stop,0);    
	cudaEventSynchronize(start);    
	cudaEventSynchronize(stop);   
	cudaEventElapsedTime(&time_elapsed,start,stop);

	// Print
	printf("Time Used on GPU:%f(ms)\n",time_elapsed);

	// CPU (results for validate)
    valid(h_ptr, N);

    int *h_d_ptr;
    checkCudaErrors(cudaMallocHost(&h_d_ptr, N *sizeof(int)));
    checkCudaErrors(cudaMemcpy(h_d_ptr, d_ptr, N *sizeof(int), cudaMemcpyDeviceToHost));

    bool bValid = true;

    for (int i=0; i

Nvidia 780Ti, CUDA 7.5测试结果如下：

Nvidia GTX 1050, CUDA 8.0测试代码如下：

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

#include 

__global__ void gpu_ptx(int *d_ptr, int length) {
	int elemID = blockIdx.x * blockDim.x + threadIdx.x;

	for (int innerloops = 0; innerloops < 100000; innerloops++) {
		if (elemID < length) {
			unsigned int laneid;
			asm("mov.u32 %0, %%laneid;" : "=r"(laneid)); // 索引缓存
			d_ptr[elemID] = laneid;
		}
	}
}

void valid(int *h_ptr, int length) {
	for (int elemID = 0; elemID> >(d_ptr, N);
	cudaGetLastError();
	cudaDeviceSynchronize();

	// Finish timing
	cudaEventRecord(stop, 0);
	cudaEventSynchronize(start);
	cudaEventSynchronize(stop);
	cudaEventElapsedTime(&time_elapsed, start, stop);

	// Print
	printf("Time Used on GPU:%f(ms)\n", time_elapsed);

	// CPU (results for validate)
	valid(h_ptr, N);

	int *h_d_ptr;
	cudaMallocHost(&h_d_ptr, N * sizeof(int));
	cudaMemcpy(h_d_ptr, d_ptr, N * sizeof(int), cudaMemcpyDeviceToHost);

	bool bValid = true;

	for (int i = 0; i

Nvidia GTX 1050, CUDA 8.0测试结果如下：