📒CUDA-Learn-Notes: 🎉CUDA/C++ 笔记 / 大模型手撕CUDA / 技术博客，更新随缘: flash_attn、sgemm、sgemv、warp reduce、block reduce、dot、elementwise、softmax、layernorm、rmsnorm、histogram、relu、sigmoid etc. 👉News: Most of my time now is focused on LLM/VLM Inference. Please check 📖Awesome-LLM-Inference 
💡说明: 大佬们写的文章实在是太棒了，学到了很多东西。欢迎大家提PR推荐更多优秀的文章！
- 📖 sgemm_naive_f32_kernel
- 📖 sgemm_block_tile_k_tile_vec4_f32_kernel
- 📖 sgemv_k32_f32_kernel
- 📖 sgemv_k128_f32_kernel
- 📖 sgemv_k16_f32_kernel
- 📖 warp_reduce_sum/max_f32_kernel
- 📖 block_reduce_sum/max_f32_kernel
- 📖 block_all_reduce_f32_kernel
- 📖 block_all_reduce_vec4_f32_kernel
- 📖 dot_product_f32_kernel
- 📖 dot_product_vec4_f32_kernel
- 📖 elementwise_f32_kernel
- 📖 elementwise_vec4_f32_kernel
- 📖 histogram_i32_kernel
- 📖 histogram_vec4_i32_kernel
- 📖 softmax_f32_kernel (grid level memory fence)
- 📖 softmax_vec4_f32_kernel (grid level memory fence)
- 📖 safe_softmax_f32_kernel (per token)
- 📖 sigmoid_f32_kernel
- 📖 sigmoid_vec4_f32_kernel
- 📖 safe_sigmoid_f32_kernel
- 📖 relu_f32_kernel
- 📖 relu_vec4_f32_kernel
- 📖 layer_norm_f32_kernel (per token)
- 📖 layer_norm_vec4_f32_kernel (per token)
- 📖 layer_norm_vec4_f16_kernel (per token)
- 📖 rms_norm_f32_kernel (per token)
- 📖 rms_norm_vec4_f32_kernel (per token)
- 📖 rms_norm_vec4_f16_kernel (per token)
- 📖 flash_attn_1_fwd_f32_kernel
- 📖 flash_attn_2_fwd_f32_kernel
- 📖 flash_attn_2_fwd_f16_kernel
- 📖 flash_attn_2_fwd_b16_kernel
- 📖 flash_attn_2_fwd_f8_kernel
- 📖 flash_attn_2_split_kv_f16_kernel
- 📖 flash_attn_2_split_kv_b16_kernel
- 📖 flash_attn_2_split_kv_f8_kernel
- 📖 online_softmax_f32_kernel
- 📖 online_softmax_f16_kernel
- 📖 online_softmax_b16_kernel
- 📖 hgemm_f16_kernel
- 📖 sgemm_dbuf_f32_kernel
0x02 sgemm naive, sgemm + block-tile + k-tile + vec4 (©️back👆🏻)
#include <stdio.h>
#include <stdlib.h>
#include <float.h>
#include <vector>
#include <algorithm>
#include <cuda_runtime.h>
#define WARP_SIZE 32
#define INT4(value) (reinterpret_cast<int4*>(&(value))[0])
#define FLOAT4(value) (reinterpret_cast<float4*>(&(value))[0])
// SGEMM: Block Tile + K Tile, with smem
// Block Tile (BM, BN) + K Tile (BK=32)
// grid((N + BN - 1) / BN, (M + BM - 1) / BM), block(BN, BM)
// a: MxK, b: KxN, c: MxN, compute: c = a * b, all row major
__global__ void sgemm(float* a, float* b, float* c, int M, int N, int K) {
// [1] Block Tile: 32x32的block处理c上一块32x32的元素计算
// [2] K Tile: 使用共享内存，并将K分块为BK大小的块
constexpr int BM = 32;
constexpr int BN = 32;
constexpr int BK = 32;
__shared__ float s_a[BM][BK], s_b[BK][BN];
int bx = blockIdx.x;
int by = blockIdx.y;
int tx = threadIdx.x;
int ty = threadIdx.y;
int tid = threadIdx.y * blockDim.x + tx; // tid within the block
// load values to shared memory, 32x32 threads working together
// to fetch data along the row direction of a and b both for s_a
// and s_b 32x32x4x2=8KB, we use 32x32 threads within block to
// load 32x32 elements from global memory to shared memory, namely,
// each thread will load 1 element.
int load_smem_a_m = tid / 32; // 0~31, tid / 32, tid / BM, threadIdx.y
int load_smem_a_k = tid % 32; // 0~31, tid % 32, tid % BK, threadIdx.x
int load_smem_b_k = tid / 32; // 0~31, tid / 32, tid / BK, threadIdx.y
int load_smem_b_n = tid % 32; // 0~31, tid % 32, tid % BN, threadIdx.x
int load_gmem_a_m = by * BM + load_smem_a_m; // global row of a and c
int load_gmem_b_n = bx * BN + load_smem_b_n; // global col of b and c
// if (load_gmem_a_m >= M || load_gmem_b_n >= N) return;
float sum = 0.f;
for (int bk = 0; bk < (K + BK - 1) / BK; ++bk) {
int load_gmem_a_k = bk * BK + load_smem_a_k;
int load_gmem_a_addr = load_gmem_a_m * K + load_gmem_a_k;
s_a[load_smem_a_m][load_smem_a_k] = a[load_gmem_a_addr];
int load_gmem_b_k = bk * BK + load_smem_b_k;
int load_gmem_b_addr = load_gmem_b_k * N + load_gmem_b_n;
s_b[load_smem_b_k][load_smem_b_n] = b[load_gmem_b_addr];
__syncthreads();
#pragma unroll
for (int k = 0; k < BK; ++k) {
int comp_smem_a_m = load_smem_a_m;
int comp_smem_b_n = load_smem_b_n;
sum += s_a[comp_smem_a_m][k] * s_b[k][comp_smem_b_n];
}
__syncthreads();
}
int store_gmem_c_m = load_gmem_a_m;
int store_gmem_c_n = load_gmem_b_n;
int store_gmem_c_addr = store_gmem_c_m * N + store_gmem_c_n;
c[store_gmem_c_addr] = sum;
}
// SGEMM: Block Tile + Thread Tile + K Tile + Vec4, with smem
// BK:TILE_K=8 BM=BN=128
// TM=TN=8 增加计算密度 BM/TM=16 BN/TN=16
// dim3 blockDim(BN/TN, BM/TM);
// dim3 gridDim((N + BN - 1) / BN, (M + BM - 1) / BM)
__global__ void sgemm_thread_tile_vec4(
float* a, float* b, float* c, int M, int N, int K) {
// [1] Block Tile: 一个16x16的block处理C上大小为128X128的一个目标块
// [2] Thread Tile: 每个thread负责计算TM*TN(8*8)个元素，增加计算密度
// [3] K Tile: 将K分块，每块BK大小，迭代(K+BK-1/BK)次，
// 每次计算TM*TN个元素各自的部分乘累加
// [4] Vectorize: 减少load和store指令，使用float4
constexpr int BM = 128;
constexpr int BN = 128;
constexpr int BK = 8;
constexpr int TM = 8;
constexpr int TN = 8;
int bx = blockIdx.x;
int by = blockIdx.y;
int tx = threadIdx.x;
int ty = threadIdx.y;
int tid = threadIdx.y * blockDim.x + tx; // tid within the block
__shared__ float s_a[BM][BK], s_b[BK][BN]; // 2*128*8*4=8KB
// 0. 先计算shared memory中的索引
// tid和需要加载的smem s_a[BM][BK] 之间的索引关系 BM=128 BK=8 按行读取 A行主序
// 对于s_a每行8个数据，每个线程读取4个，需要2个线程；总共128行，需要128x2刚好256线程
int load_smem_a_m = tid / 2; // tid/2 (128/8)*(128/8)=256 threads per block, tid/2->[0,128), BM=128 0~127
int load_smem_a_k = (tid % 2 == 0) ? 0 : 4; // (tid%2 == 0) ? 0 : 4, col of s_a 0,4
// tid和需要加载的smem s_b[BK][BN] 之间的索引关系 BK=8 BN=128 按行读取 B行主序
// 对于s_b每行128个数据，每个线程读4个数据，需要32个线程；总共8行，需要32x8=256个线程
int load_smem_b_k = tid / 32; // tid/32, row of s_b 256/32=8 行 0~7
int load_smem_b_n = (tid % 32) * 4; // (tid % 32) * 4, col of s_b 0,4,...,124
// 1. 再计算全局内存中的索引
// 要加载到s_a中的元素对应到A全局内存中的行数 每个block负责出C中大小为BM*BN的块
int load_gmem_a_m = by * BM + load_smem_a_m; // global row of a and c
int load_gmem_b_n = bx * BN + load_smem_b_n; // global col of b and c
float r_c[TM][TN] = {0.0}; // 8x8
// 2. 先对K进行分块，每块BK大小
for (int bk = 0; bk < (K + BK - 1) / BK; ++bk) {
// 加载数据到共享内存smem s_a BM*BK 128*8 vectorize float4
int load_gmem_a_k = bk * BK + load_smem_a_k; // global col of a
int load_gmem_a_addr = load_gmem_a_m * K + load_gmem_a_k;
FLOAT4(s_a[load_smem_a_m][load_smem_a_k]) = FLOAT4(a[load_gmem_a_addr]);
// 加载数据到共享内存smem s_b BK*BN 8*128 vectorize float4
int load_gmem_b_k = bk * BK + load_smem_b_k; // global row of b
int load_gmem_b_addr = load_gmem_b_k * N + load_gmem_b_n;
FLOAT4(s_b[load_smem_b_k][load_smem_b_n]) = FLOAT4(b[load_gmem_b_addr]);
__syncthreads();
#pragma unroll
for (int k = 0; k < BK; k++) {
// 3. 每个线程负责计算BM*BN(12x128)中的TM*TN(8x8)个元素
#pragma unroll
for (int m = 0; m < TM; m++) {
#pragma unroll
for (int n = 0; n < TN; n++) {
// k from 0~7，0 ~ BK, ty and tx range from 0 to 15, 16x8=128
int comp_smem_a_m = ty * TM + m; // 128*8 128/TM(8)=16 M方向 16线程
int comp_smem_b_n = tx * TN + n; // 8*128 128/TN(8)=16 N方向 16线程
r_c[m][n] += s_a[comp_smem_a_m][k] * s_b[k][comp_smem_b_n];
}
}
}
__syncthreads();
}
#pragma unroll
for (int m = 0; m < TM; ++m) {
int store_gmem_c_m = by * BM + ty * TM + m;
#pragma unroll
for (int n = 0; n < TN; n += 4) {
int store_gmem_c_n = bx * BN + tx * TN + n;
int store_gmem_c_addr = store_gmem_c_m * N + store_gmem_c_n;
FLOAT4(c[store_gmem_c_addr]) = FLOAT4(r_c[m][n]);
}
}
}这里gemm的实现比较简单，只使用了CUDA Cores，并且只实现Block Tile + K Tile以及Block Tile + K Tile+Thread Tile+向量化的版本。主要在于如何加载gmem中的数据到smem，也就是把全局内存中的数据索引mapping到共享内存中的。核心思维：把一个block中的线程id按照线性来理解，然后把这个线性的id和全局内存索引以及共享内存索引进行匹配。比如Block Tile + K Tile的实现，block内一共32x32个Threads，需要加载到smem的数据也是32x32，那么，最简单的做法，只需要每个线程加载一个互不重复数据即可。NOTE，本文的gemm kernel修改自：紫气东来：CUDA（三）：通用矩阵乘法：从入门到熟练
0x03 warp/block reduce sum/max (©️back👆🏻)
// Warp Reduce Sum
template<const int kWarpSize = WARP_SIZE>
__device__ __forceinline__ float warp_reduce_sum(float val) {
#pragma unroll
for (int mask = kWarpSize >> 1; mask >= 1; mask >>= 1) {
val += __shfl_xor_sync(0xffffffff, val, mask);
}
return val;
}
// Warp Reduce Max
template<const int kWarpSize = WARP_SIZE>
__device__ __forceinline__ float warp_reduce_max(float val) {
#pragma unroll
for (int mask = kWarpSize >> 1; mask >= 1; mask >>= 1) {
val = fmaxf(val, __shfl_xor_sync(0xffffffff, val, mask));
}
return val;
}
// Block reduce sum/max/min device helper for Layer/RMS Norm/Softmax etc.
// grid 1D block 1D, grid(N/128), block(128)
template<const int NUM_THREADS=128>
__device__ __forceinline__ float block_reduce_sum(float val) {
// always <= 32 warps per block (limited by 1024 threads per block)
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
int warp = threadIdx.x / WARP_SIZE;
int lane = threadIdx.x % WARP_SIZE;
static __shared__ float shared[NUM_WARPS];
val = warp_reduce_sum<WARP_SIZE>(val);
if (lane == 0) shared[warp] = val;
__syncthreads();
val = (lane < NUM_WARPS) ? shared[lane] : 0.0f;
val = warp_reduce_sum<NUM_WARPS>(val);
return val;
}
template<const int NUM_THREADS=128>
__device__ __forceinline__ float block_reduce_max(float val) {
// always <= 32 warps per block (limited by 1024 threads per block)
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
int warp = threadIdx.x / WARP_SIZE;
int lane = threadIdx.x % WARP_SIZE;
static __shared__ float shared[NUM_WARPS];
val = warp_reduce_max<WARP_SIZE>(val);
if (lane == 0) shared[warp] = val;
__syncthreads();
val = (lane < NUM_WARPS) ? shared[lane] : -FLT_MAX;
val = warp_reduce_max<NUM_WARPS>(val);
return val;
}warp reduce几乎已经成为大部分reduce kernel的标准写法了，比如vLLM中，就是这种经典的写法。所以，先搞懂warp reduce（也就是搞懂各种warp functions的用法），再去写其他kernel，思路就会容易很多。需要注意的是，warp函数处理的是寄存器上的数据，也就是说，此时，没必要先加载数据到smem，再进行reduce，直接加载到寄存器即可（以前犯过这个小错误...）。Warp Functions建议参考：jhang：CUDA编程入门之Warp-Level Primitives
0x04 block all reduce + vec4 (©️back👆🏻)
// Block All Reduce Sum
// grid(N/128), block(128)
// a: Nx1, y=sum(a)
template<const int NUM_THREADS = 128>
__global__ void block_all_reduce_sum(float* a, float* y, int N) {
int tid = threadIdx.x;
int idx = blockIdx.x * NUM_THREADS + tid;
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
__shared__ float reduce_smem[NUM_WARPS];
// keep the data in register is enougth for warp operaion.
float sum = (idx < N) ? a[idx] : 0.0f;
int warp = tid / WARP_SIZE;
int lane = tid % WARP_SIZE;
// perform warp sync reduce.
sum = warp_reduce_sum<WARP_SIZE>(sum);
// warp leaders store the data to shared memory.
if (lane == 0) reduce_smem[warp] = sum;
__syncthreads(); // make sure the data is in shared memory.
// the first warp compute the final sum.
sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f;
if (warp == 0) sum = warp_reduce_sum<NUM_WARPS>(sum);
if (tid == 0) atomicAdd(y, sum);
}
// Block All Reduce Sum + float4
// grid(N/128), block(128/4)
// a: Nx1, y=sum(a)
template<const int NUM_THREADS = 128/4>
__global__ void block_all_reduce_sum_vec4(float* a, float* y, int N) {
int tid = threadIdx.x;
int idx = (blockIdx.x * NUM_THREADS + tid) * 4;
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
__shared__ float reduce_smem[NUM_WARPS];
float4 reg_a = FLOAT4(a[idx]);
// keep the data in register is enougth for warp operaion.
float sum = (idx < N) ? (reg_a.x + reg_a.y + reg_a.z + reg_a.w) : 0.0f;
int warp = tid / WARP_SIZE;
int lane = tid % WARP_SIZE;
// perform warp sync reduce.
sum = warp_reduce_sum<WARP_SIZE>(sum);
// warp leaders store the data to shared memory.
if (lane == 0) reduce_smem[warp] = sum;
__syncthreads(); // make sure the data is in shared memory.
// the first warp compute the final sum.
sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f;
if (warp == 0) sum = warp_reduce_sum<NUM_WARPS>(sum);
if (tid == 0) atomicAdd(y, sum);
}block all reduce是在warp reduce的基础上进行的，reduce_smem这部分的共享内存申请无法避免，这是用来同步每个warp之间得到局部结果。注意，最后，还需要atomicAdd做一个block级别的原子操作，以得到全局的和。float4向量化优化访存，可以减缓WarpScheduler发送指令的压力。
0x05 sgemv k32/k128/k16 kernel (©️back👆🏻)
// SGEMV: Warp SGEMV K32
// 假设K为32的倍数，每个warp负责一行
// grid(M/4), block(32,4) blockDim.x=32=K, blockDim.y=4
// a: MxK, x: Kx1, y: Mx1, compute: y = a * x
__global__ void sgemv_k32(float* a, float* x, float* y, int M, int K) {
int tx = threadIdx.x; // 0~31
int ty = threadIdx.y; // 0~4
int bx = blockIdx.x; // 0~M/4
int lane = tx % WARP_SIZE; // 0~31
int m = bx * blockDim.y + ty; // (0~M/4) * 4 + (0~3)
if (m < M) {
float sum = 0.0f;
int NUM_WARPS = (K + WARP_SIZE - 1) / WARP_SIZE;
#pragma unroll
for (int w = 0; w < NUM_WARPS; ++w) {
// 若NUM_WARPS>=2，先将当前行的数据累加到第一个warp中
int k = w * WARP_SIZE + lane;
sum += a[m * K + k] * x[k];
}
sum = warp_reduce_sum<WARP_SIZE>(sum);
if (lane == 0) y[m] = sum;
}
}
// SGEMV: Warp SGEMV K128 + Vec4
// 假设K为128的倍数 float4
// grid(M/4), block(32,4) blockDim.x=32=K, blockDim.y=4
// a: MxK, x: Kx1, y: Mx1, compute: y = a * x
__global__ void sgemv_k128(float* a, float* x, float* y, int M, int K) {
// 每个线程负责4个元素，一个warp覆盖128个元素
int tx = threadIdx.x; // 0~31
int ty = threadIdx.y; // 0~3
int bx = blockIdx.x; // 0~M/4
int lane = tx % WARP_SIZE; // 0~31
int m = blockDim.y * bx + ty; // (0~M/4) * 4 + (0~3)
if (m < M) {
float sum = 0.0f;
// process 4*WARP_SIZE elements per warp.
int NUM_WARPS = (((K + WARP_SIZE - 1) / WARP_SIZE) + 4 - 1) / 4;
#pragma unroll
for (int w = 0; w < NUM_WARPS; ++w) {
int k = (w * WARP_SIZE + lane) * 4;
float4 reg_x = FLOAT4(x[k]);
float4 reg_a = FLOAT4(a[m * K + k]);
sum += (reg_a.x * reg_x.x + reg_a.y * reg_x.y
+ reg_a.z * reg_x.z + reg_a.w * reg_x.w);
}
sum = warp_reduce_sum<WARP_SIZE>(sum);
if(lane == 0) y[m] = sum;
}
}
// SGEMV: Warp SGEMV K16
// 假设K为16 < 32,每个warp负责2行，每行有16个元素
// NUM_THREADS=128, NUM_WARPS=NUM_THREADS/WARP_SIZE;
// NUM_ROWS=NUM_WARPS * ROW_PER_WARP, grid(M/NUM_ROWS), block(32,NUM_WARPS)
// a: MxK, x: Kx1, y: Mx1, compute: y = a * x
template<const int ROW_PER_WARP = 2>
__global__ void sgemv_k16(float* A, float* x, float* y, int M, int K) {
constexpr int K_WARP_SIZE = (WARP_SIZE + ROW_PER_WARP - 1) / ROW_PER_WARP;
int tx = threadIdx.x; // 0~31
int ty = threadIdx.y; // 0~NUM_WARPS
int bx = blockIdx.x; // 0~M/NUM_ROWS (NUM_ROWS=NUM_WARPS * ROW_PER_WARP)
int lane = tx % WARP_SIZE; // 0~31
int k = lane % K_WARP_SIZE; // 0~15
// gloabl row of a: MxK and y:Mx1, blockDim.y=NUM_WARPS
int m = (blockDim.y * bx + ty) * ROW_PER_WARP + lane / K_WARP_SIZE;
if (m < M) {
float sum = A[m * K + k] * x[k];
sum = warp_reduce_sum<K_WARP_SIZE>(sum);
// 注意是k == 0，而不是lane == 0
if(k == 0) y[m] = sum;
}
}估计有些大佬倒立都能写sgemv的各种优化版了，核心思路其实也是基于warp reduce，考虑K的不同情况进行优化。本文的sgemv kernel修改自：有了琦琦的棍子：深入浅出GPU优化系列：gemv优化
0x06 dot product, dot product + vec4 (©️back👆🏻)
// Dot Product
// grid(N/128), block(128)
// a: Nx1, b: Nx1, y=sum(elementwise_mul(a,b))
template<const int NUM_THREADS = 128>
__global__ void dot(float* a, float* b, float* y, int N) {
int tid = threadIdx.x;
int idx = blockIdx.x * NUM_THREADS + tid;
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
__shared__ float reduce_smem[NUM_WARPS];
// keep the data in register is enougth for warp operaion.
float prod = (idx < N) ? a[idx] * b[idx] : 0.0f;
int warp = tid / WARP_SIZE;
int lane = tid % WARP_SIZE;
// perform warp sync reduce.
prod = warp_reduce_sum<WARP_SIZE>(prod);
// warp leaders store the data to shared memory.
if (lane == 0) reduce_smem[warp] = prod;
__syncthreads(); // make sure the data is in shared memory.
// the first warp compute the final sum.
prod = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f;
if (warp == 0) prod = warp_reduce_sum<NUM_WARPS>(prod);
if (tid == 0) atomicAdd(y, prod);
}
// Dot Product + Vec4
// grid(N/128), block(128/4)
// a: Nx1, b: Nx1, y=sum(elementwise_mul(a,b))
template<const int NUM_THREADS = 128/4>
__global__ void dot_vec4(float* a, float* b, float* y, int N) {
int tid = threadIdx.x;
int idx = (blockIdx.x * NUM_THREADS + tid) * 4;
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
__shared__ float reduce_smem[NUM_WARPS];
float4 reg_a = FLOAT4(a[idx]);
float4 reg_b = FLOAT4(b[idx]);
float prod = (idx < N) ? (reg_a.x * reg_b.x + reg_a.y * reg_b.y
+ reg_a.z * reg_b.z + reg_a.w * reg_b.w) : 0.0f;
int warp = tid / WARP_SIZE;
int lane = tid % WARP_SIZE;
// perform warp sync reduce.
prod = warp_reduce_sum<WARP_SIZE>(prod);
// warp leaders store the data to shared memory.
if (lane == 0) reduce_smem[warp] = prod;
__syncthreads(); // make sure the data is in shared memory.
// the first warp compute the final sum.
prod = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f;
if (warp == 0) prod = warp_reduce_sum<NUM_WARPS>(prod);
if (tid == 0) atomicAdd(y, prod);
}dot product kernel的核心就是block reduce，不多说了。
0x07 elementwise, elementwise + vec4 (©️back👆🏻)
// ElementWise Add
// grid(N/128), block(128)
// a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b)
__global__ void elementwise_add(float* a, float* b, float* c, int N) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) c[idx] = a[idx] + b[idx];
}
// ElementWise Add + Vec4
// grid(N/128), block(128/4)
// a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b)
__global__ void elementwise_add_vec4(float* a, float* b, float* c, int N) {
int idx = 4 * (blockIdx.x * blockDim.x + threadIdx.x);
if (idx < N) {
float4 reg_a = FLOAT4(a[idx]);
float4 reg_b = FLOAT4(b[idx]);
float4 reg_c;
reg_c.x = reg_a.x + reg_b.x;
reg_c.y = reg_a.y + reg_b.y;
reg_c.z = reg_a.z + reg_b.z;
reg_c.w = reg_a.w + reg_b.w;
FLOAT4(c[idx]) = reg_c;
}
}elementwise可以考虑加点向量化进行访存优化。
// Histogram
// grid(N/128), block(128)
// a: Nx1, y: count histogram
__global__ void histogram(int* a, int* y, int N) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) atomicAdd(&(y[a[idx]]), 1);
}
// Histogram + Vec4
// grid(N/128), block(128/4)
// a: Nx1, y: count histogram
__global__ void histogram_vec4(int* a, int* y, int N) {
int idx = 4 * (blockIdx.x * blockDim.x + threadIdx.x);
if (idx < N) {
int4 reg_a = INT4(a[idx]);
atomicAdd(&(y[reg_a.x]), 1);
atomicAdd(&(y[reg_a.y]), 1);
atomicAdd(&(y[reg_a.z]), 1);
atomicAdd(&(y[reg_a.w]), 1);
}
}统计频数直方图，很简单，两行代码搞定。
0x09 softmax, softmax + vec4 (grid level memory fence) (©️back👆🏻)
// Softmax x: N, y: N
// grid(N/128), block(K=128)
template<const int NUM_THREADS = 128>
__global__ void softmax(float* x, float* y, float* total, int N) {
const int tid = threadIdx.x;
const int idx = blockIdx.x * blockDim.x + tid;
constexpr int NUM_WARPS = (NUM_THREADS + WARP_SIZE - 1) / WARP_SIZE;
__shared__ float reduce_smem[NUM_WARPS];
float sum = (idx < N) ? expf(x[idx]) : 0.0f;
int warp = tid / WARP_SIZE;
int lane = tid % WARP_SIZE;
sum = warp_reduce_sum<WARP_SIZE>(sum);
if (lane == 0) reduce_smem[warp] = sum;
__syncthreads();
// compute the final sum in each warp
sum = (lane < NUM_WARPS) ? reduce_smem[lane] : 0.0f;
sum = warp_reduce_sum<NUM_WARPS>(sum); // sum(e^x_0,...,e^x_n-1)
// get the total sum of all blocks.
if (tid == 0) atomicAdd(total, sum);
__threadfence(); // grid level memory fence 注意这里需要网格级别的内存同步
// e^x_i/sum(e^x_0,...,e^x_n-1)
if (idx < N) y[idx] = block_smem[tid] / (*total);
}
// Softmax x: N, y: N
// grid(N/128), block(K=128)
template<const int NUM_THREADS = 128>
__global__ void softmax_v2(float* x, float* y, float* total, int N) {
const int tid = threadIdx.x;
const int idx = blockIdx.x * blockDim.x + tid;
float exp_val = (idx < N) ? expf(x[idx]) : 0.0f;
float sum = block_reduce_sum<NUM_THREADS>(exp_val);
// get the total sum of all blocks.
if (tid == 0) atomicAdd(total, sum);
__threadfence(); // grid level memory fence 注意这里需要网格级别的内存同步
// e^x_i/sum(e^x_0,...,e^x_n-1)
if (idx < N) y[idx] = exp_val / (*total);
}
// Softmax Vec4 x: N, y: N
// grid(N/128), block(128/4)
template<const int NUM_THREADS = 128/4>
__global__ void softmax_v2_vec4(float* x, float* y, float* total, int N) {
const int tid = threadIdx.x;
const int idx = (blockIdx.x * blockDim.x + tid) * 4;
float4 reg_x = FLOAT4(x[idx]);
float4 reg_exp;
reg_exp.x = (idx < N) ? expf(reg_x.x) : 0.0f;
reg_exp.y = (idx < N) ? expf(reg_x.y) : 0.0f;
reg_exp.z = (idx < N) ? expf(reg_x.z) : 0.0f;
reg_exp.w = (idx < N) ? expf(reg_x.w) : 0.0f;
float exp_val = (reg_exp.x + reg_exp.y + reg_exp.z + reg_exp.w);
float sum = block_reduce_sum<NUM_THREADS>(exp_val);
// get the total sum of all blocks.
if (tid == 0) atomicAdd(total, sum);
__threadfence(); // grid level memory fence 注意这里需要网格级别的内存同步
// e^x_i/sum(e^x_0,...,e^x_n-1)
if (idx < N) {
float4 reg_y;
reg_y.x = reg_exp.x / (*total);
reg_y.y = reg_exp.y / (*total);
reg_y.z = reg_exp.z / (*total);
reg_y.w = reg_exp.w / (*total);
FLOAT4(y[idx]) = reg_y;
}
}softmax稍微要注意的就是内存同步的问题，这里，你需要做一个网格级别的同步，而不能仅仅是block级别，否则拿不到全局的exp sum作为分母项。因此使用 __threadfence 这个网格及内存同步操作。不过效率我还没测过，实在要高效的话，可能得整成FA2那样的 1-pass + online softmax的实现。不过，如果是面试的话，就不要太为难自己了...，但是FA1/FA2的论文很经典，强烈建议多读几遍。
0x0a sigmoid, sigmoid + vec4 (©️back👆🏻)
// Sigmoid x: N, y: N y=1/(1+exp(-x))
// grid(N/128), block(K=128)
__global__ void sigmoid(float* x, float* y, int N) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) y[idx] = 1.0f / (1.0f + expf(-x[idx]));
}
// Sigmoid x: N, y: N y=1/(1+exp(-x)) Vec4
// grid(N/128), block(128/4)
__global__ void sigmoid_vec4(float* x, float* y, int N) {
int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4;
if (idx < N) {
float4 reg_x = FLOAT4(x[idx]);
float4 reg_y;
reg_y.x = 1.0f / (1.0f + expf(-reg_x.x));
reg_y.y = 1.0f / (1.0f + expf(-reg_x.y));
reg_y.z = 1.0f / (1.0f + expf(-reg_x.z));
reg_y.w = 1.0f / (1.0f + expf(-reg_x.w));
FLOAT4(y[idx]) = reg_y;
}
}0x0b relu, relu + vec4 (©️back👆🏻)
// Relu x: N, y: N y=max(0,x)
// grid(N/128), block(K=128)
__global__ void relu(float* x, float* y, int N) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) y[idx] = fmaxf(0.0f, x[idx]);
}
// Relu x: N, y: N y=max(0,x) Vec4
// grid(N/128/4), block(128/4)
__global__ void relu_vec4(float* x, float* y, int N) {
int idx = (blockIdx.x * blockDim.x + threadIdx.x) * 4;
if (idx < N) {
float4 reg_x = FLOAT4(x[idx]);
float4 reg_y;
reg_y.x = fmaxf(0.0f, reg_x.x);
reg_y.y = fmaxf(0.0f, reg_x.y);
reg_y.z = fmaxf(0.0f, reg_x.z);
reg_y.w = fmaxf(0.0f, reg_x.w);
FLOAT4(y[idx]) = reg_y;
}
}0x0c layer_norm, layer_norm + vec4 (©️back👆🏻)
// Layer Norm: x: NxK(K=128<1024), y': NxK, y'=x-mean(x)/std(x) each row
// mean(x) = sum(x)/K, 1/std(x) = rsqrtf( sum( (x-mean(x))^2 )/K ) each row
// grid(N*K/K), block(K<1024) N=batch_size*seq_len, K=hidden_size
// y=y'*g + b (g: scale, b: bias)
template<const int NUM_THREADS=128>
__global__ void layer_norm(float* x, float* y, float g, float b, int N, int K) {
int tid = threadIdx.x; // 0..K-1
int bid = blockIdx.x; // 0..N-1
int idx = bid * blockDim.x + threadIdx.x;
const float epsilon = 1e-5f;
__shared__ float s_mean; // shared within block
__shared__ float s_variance; // shared within block
float value = (idx < N * K) ? x[idx] : 0.0f; // load once only
float sum = block_reduce_sum<NUM_THREADS>(value);
if (tid == 0) s_mean = sum / (float) K;
// wait for s_mean in shared memory to be ready for all threads
__syncthreads();
float variance = (value - s_mean) * (value - s_mean);
variance = block_reduce_sum<NUM_THREADS>(variance);
if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon);
// wait for s_variance in shared memory to be ready for all threads
__syncthreads();
if (idx < N * K) y[idx] = ((value - s_mean) * s_variance) * g + b;
}
// Layer Norm Vec4: x: NxK(K=128<1024), y': NxK, y'=x-mean(x)/std(x) each row
// mean(x) = sum(x)/K, 1/std(x) = rsqrtf( sum( (x-mean(x))^2 )/K ) each row
// grid(N*K/K), block(K/4<1024) N=batch_size*seq_len, K=hidden_size
// y=y'*g + b (g: scale, b: bias)
template<const int NUM_THREADS=128/4>
__global__ void layer_norm_vec4(float* x, float* y, float g, float b, int N, int K) {
int tid = threadIdx.x; // 0..K-1
int bid = blockIdx.x; // 0..N-1
int idx = (bid * blockDim.x + threadIdx.x) * 4;
const float epsilon = 1e-5f;
__shared__ float s_mean; // shared within block
__shared__ float s_variance; // shared within block
float4 reg_x = FLOAT4(x[idx])
float value = (idx < N * K) ? (reg_x.x + reg_x.y
+ reg_x.z + reg_x.w) : 0.0f;
float sum = block_reduce_sum<NUM_THREADS>(value);
if (tid == 0) s_mean = sum / (float) K;
// wait for s_mean in shared memory to be ready for all threads
__syncthreads();
float4 reg_x_hat;
reg_x_hat.x = reg_x.x - s_mean;
reg_x_hat.y = reg_x.y - s_mean;
reg_x_hat.z = reg_x.z - s_mean;
reg_x_hat.w = reg_x.w - s_mean;
float variance = reg_x_hat.x * reg_x_hat.x + reg_x_hat.y * reg_x_hat.y
+ reg_x_hat.z * reg_x_hat.z + reg_x_hat.w * reg_x_hat.w;
variance = block_reduce_sum<NUM_THREADS>(variance);
if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon);
// wait for s_variance in shared memory to be ready for all threads
__syncthreads();
float4 reg_y;
reg_y.x = reg_x_hat.x * s_variance * g + b;
reg_y.y = reg_x_hat.y * s_variance * g + b;
reg_y.z = reg_x_hat.z * s_variance * g + b;
reg_y.w = reg_x_hat.w * s_variance * g + b;
if (idx < N * K) FLOAT4(y[idx]) = reg_y;
}layer norm实现的核心同样也是block reduce和warp reduce，然后再整点向量化...
0x0d rms_norm, rms_norm + vec4 (©️back👆🏻)
// RMS Norm: x: NxK(K=128<1024), y': NxK, y'=x/rms(x) each row
// 1/rms(x) = rsqrtf( sum(x^2)/K ) each row
// grid(N*K/K), block(K<1024) N=batch_size*seq_len, K=hidden_size
// y=y'*g (g: scale)
template<const int NUM_THREADS=128>
__global__ void rms_norm(float* x, float* y, float g, int N, int K) {
int tid = threadIdx.x; // 0..K-1
int bid = blockIdx.x; // 0..N-1
int idx = bid * blockDim.x + threadIdx.x;
const float epsilon = 1e-5f;
__shared__ float s_variance; // shared within block
float value = (idx < N * K) ? x[idx] : 0.0f; // load once only
float variance = value * value;
variance = block_reduce_sum<NUM_THREADS>(variance);
if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon);
// wait for s_variance in shared memory to be ready for all threads
__syncthreads();
if (idx < N * K) y[idx] = (value * s_variance) * g;
}
// RMS Norm Vec4: x: NxK(K=128<1024), y': NxK, y'=x/rms(x) each row
// 1/rms(x) = rsqrtf( sum(x^2)/K ) each row
// grid(N*K/K), block(K/4<1024) N=batch_size*seq_len, K=hidden_size
// y=y'*g (g: scale)
template<const int NUM_THREADS=128/4>
__global__ void rms_norm_vec4(float* x, float* y, float g, int N, int K) {
int tid = threadIdx.x; // 0..K-1
int bid = blockIdx.x; // 0..N-1
int idx = (bid * blockDim.x + threadIdx.x) * 4;
const float epsilon = 1e-5f;
__shared__ float s_variance; // shared within block
float4 reg_x = FLOAT4(x[idx]);
float variance = (idx < N * K) ? (reg_x.x * reg_x.x + reg_x.y * reg_x.y
+ reg_x.z * reg_x.z + reg_x.w * reg_x.w) : 0.0f;
variance = block_reduce_sum<NUM_THREADS>(variance);
if (tid == 0) s_variance = rsqrtf(variance / (float) K + epsilon);
// wait for s_variance in shared memory to be ready for all threads
__syncthreads();
float4 reg_y;
reg_y.x = reg_x.x * s_variance * g;
reg_y.y = reg_x.y * s_variance * g;
reg_y.z = reg_x.z * s_variance * g;
reg_y.w = reg_x.w * s_variance * g;
if (idx < N * K) FLOAT4(y[idx]) = reg_y;
}rms norm实现的核心同样也是block reduce和warp reduce...，然后再加点float4向量化什么的。
0x0e NMS (©️back👆🏻)
struct Box {
float x1, y1, x2, y2, score;
float area() const {return (std::abs(x2 - x1 + 1)) * std::abs(y2 - y1 + 1); }
float iou_of(const Box& other) const{
float inner_x1 = x1 > other.x1 ? x1 : other.x1;
float inner_y1 = y1 > other.y1 ? y1 : other.y1;
float inner_x2 = x2 < other.x2 ? x2 : other.x2;
float inner_y2 = y2 < other.y2 ? y2 : other.y2;
float inner_h = inner_y2 - inner_y1 + 1.0f;
float inner_w = inner_x2 - inner_x1 + 1.0f;
float inner_area = inner_h * inner_w;
return (inner_area / (area() + tbox.area() - inner_area));
}
}
void hard_nms(std::vector<Box> &input, std::vector<Box> &output, float iou_threshold){
if (input.empty()) return;
std::sort(input.begin(), input.end(),[](Box& a, Box& b) { return a.score > b.score; });
int box_num = input.size();
std::vector<int> merged(box_num, 0);
for (int i = 0; i < box_num; ++i) {
if (merged[i]) continue;
merged[i] = 1;
for (int j = i + 1; j < box_num; ++j) {
if (merged[j]) continue;
float iou = input[i].iou_of(input[j]);
if (iou > iou_threshold) merged[j] = 1;
}
output.push_back(input[i]);
}
}CV相关的经常会要手撕NMS，也记录下。
0x0f 总结 (©️back👆🏻)
可以发现，大部分kernel的基本写法都是依赖warp reduce和block reduce的，基本上只要熟练应用warp functions各种场景的写法，应该问题不大；softmax需要考虑网格级同步的问题，或者online softmax以及FlashAttention；sgemm的优化是个很大的课题，不是案例中写的这么简单，但是入门的话，基本就是tiling的思想以及如何做索引之间的mapping；sgemv的优化则主要考虑K不同的值（因为M为1了），比如K=16,64,128等情况下，如何按照warp来处理；relu、sigmoid等都是elementwise的操作，很好实现，可以再考虑加点向量化优化访存；layer norm和rms norm在数学上其实也是挺清晰简单的，落实到cuda kernel时，只要按照逐个token来处理，headdim没有超过1024的情况下（一个block最多可以放1024个threads），可以放到一个block处理，这样并行化就很好写。当然，核心还是warp reduce和block reduce；NMS是乱入的，没有CUDA版本，别问了...
GNU General Public License v3.0
- flash-attention-minimal: Flash Attention in ~100 lines of CUDA (forward pass only)
🌟如果觉得有用，不妨给个🌟👆🏻Star支持一下吧~