forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish AngularDistance #62' (#503) from gsd123/GPUCodeForces:gsd62 into main
This commit is contained in:
commit
21a8cf7fda
|
|
@ -0,0 +1,380 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
N, D = 32, 64
|
||||
|
||||
|
||||
class AngularDistanceCUDAOp(torch.autograd.Function):
|
||||
epsilon = 1e-6
|
||||
|
||||
def forward(ctx, input, target, beta, reduction_id, op):
|
||||
if not input.is_cuda: input = input.cuda()
|
||||
if not target.is_cuda: target = target.cuda()
|
||||
|
||||
input = input.contiguous()
|
||||
target = target.contiguous()
|
||||
|
||||
N_batch = input.size(0)
|
||||
|
||||
output, dot, norm_i, norm_t = op.angular_loss_forward_cuda(
|
||||
input,
|
||||
target,
|
||||
reduction_id,
|
||||
N_batch,
|
||||
AngularDistanceCUDAOp.epsilon
|
||||
)
|
||||
|
||||
ctx.save_for_backward(input, target, dot, norm_i, norm_t)
|
||||
ctx.reduction_id = reduction_id
|
||||
ctx.N = N_batch
|
||||
ctx.op = op
|
||||
|
||||
return output
|
||||
|
||||
def backward(ctx, grad_output):
|
||||
input, target, dot, norm_i, norm_t = ctx.saved_tensors
|
||||
|
||||
grad_out_scalar = 0.0
|
||||
grad_output_n = None
|
||||
|
||||
if ctx.reduction_id != 0:
|
||||
grad_out_scalar = grad_output[0]
|
||||
if ctx.reduction_id == 1:
|
||||
grad_out_scalar = grad_out_scalar / ctx.N
|
||||
else:
|
||||
grad_output_n = grad_output.contiguous()
|
||||
|
||||
grad_input = torch.empty_like(input)
|
||||
grad_target = torch.empty_like(target)
|
||||
|
||||
ctx.op.angular_loss_backward_cuda(
|
||||
grad_out_scalar,
|
||||
grad_output_n,
|
||||
input,
|
||||
target,
|
||||
dot,
|
||||
norm_i,
|
||||
norm_t,
|
||||
grad_input,
|
||||
grad_target,
|
||||
ctx.reduction_id,
|
||||
input.size(0),
|
||||
input.size(1),
|
||||
AngularDistanceCUDAOp.epsilon
|
||||
)
|
||||
|
||||
return grad_input, grad_target, None, None, None
|
||||
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self, reduction='mean', beta=1.0):
|
||||
super().__init__()
|
||||
self.beta = float(beta)
|
||||
|
||||
self.red_map = {'none': 0, 'mean': 1, 'sum': 2}
|
||||
if reduction not in self.red_map:
|
||||
raise ValueError("Invalid reduction")
|
||||
self.reduction_id = self.red_map[reduction]
|
||||
|
||||
self._compile_cuda_kernel()
|
||||
|
||||
def _compile_cuda_kernel(self):
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <vector>
|
||||
|
||||
std::vector<torch::Tensor> angular_loss_forward_cuda(
|
||||
torch::Tensor input,
|
||||
torch::Tensor target,
|
||||
int reduction,
|
||||
int N,
|
||||
float epsilon);
|
||||
|
||||
void angular_loss_backward_cuda(
|
||||
float grad_out_scalar,
|
||||
torch::Tensor grad_output_n,
|
||||
torch::Tensor input,
|
||||
torch::Tensor target,
|
||||
torch::Tensor dot,
|
||||
torch::Tensor norm_i,
|
||||
torch::Tensor norm_t,
|
||||
torch::Tensor grad_input,
|
||||
torch::Tensor grad_target,
|
||||
int reduction,
|
||||
int N, int D,
|
||||
float epsilon);
|
||||
"""
|
||||
|
||||
cuda_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <vector>
|
||||
#include <algorithm>
|
||||
#include <cuda_runtime.h>
|
||||
#include <cmath>
|
||||
#include <device_functions.h>
|
||||
#include <math_functions.h>
|
||||
|
||||
#define BLOCK_SIZE 256
|
||||
#define MAX_GRID_SIZE 4096
|
||||
|
||||
__inline__ __device__ double warp_reduce_sum_double(double val) {
|
||||
#pragma unroll
|
||||
for (int offset = 16; offset > 0; offset /= 2) {
|
||||
val += __shfl_down_sync(0xffffffff, val, offset);
|
||||
}
|
||||
return val;
|
||||
}
|
||||
|
||||
__inline__ __device__ double block_reduce_sum_double(double val) {
|
||||
__shared__ double shared[32];
|
||||
int lane = threadIdx.x % 32;
|
||||
int wid = threadIdx.x / 32;
|
||||
|
||||
val = warp_reduce_sum_double(val);
|
||||
if (lane == 0) shared[wid] = val;
|
||||
__syncthreads();
|
||||
|
||||
val = (threadIdx.x < blockDim.x / 32) ? shared[lane] : 0.0;
|
||||
if (wid == 0) val = warp_reduce_sum_double(val);
|
||||
return val;
|
||||
}
|
||||
|
||||
__global__ void reduce_kernel(
|
||||
const float* __restrict__ input,
|
||||
float* __restrict__ output,
|
||||
int n)
|
||||
{
|
||||
double local_sum = 0.0;
|
||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
|
||||
for (int i = idx; i < n; i += blockDim.x * gridDim.x) {
|
||||
local_sum += (double)input[i];
|
||||
}
|
||||
|
||||
local_sum = block_reduce_sum_double(local_sum);
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
atomicAdd(output, (float)local_sum);
|
||||
}
|
||||
}
|
||||
|
||||
__global__ void angular_fwd_kernel(
|
||||
const float* __restrict__ input,
|
||||
const float* __restrict__ target,
|
||||
float* __restrict__ loss_n,
|
||||
float* __restrict__ dot_out,
|
||||
float* __restrict__ norm_i_out,
|
||||
float* __restrict__ norm_t_out,
|
||||
int N, int D,
|
||||
float epsilon)
|
||||
{
|
||||
int n_idx = blockIdx.x;
|
||||
if (n_idx >= N) return;
|
||||
|
||||
const float* in_ptr = input + n_idx * D;
|
||||
const float* tgt_ptr = target + n_idx * D;
|
||||
|
||||
double local_dot = 0.0;
|
||||
double local_norm_in_sq = 0.0;
|
||||
double local_norm_tgt_sq = 0.0;
|
||||
|
||||
for (int i = threadIdx.x; i < D; i += blockDim.x) {
|
||||
double in_i = (double)in_ptr[i];
|
||||
double tgt_i = (double)tgt_ptr[i];
|
||||
|
||||
local_dot += in_i * tgt_i;
|
||||
local_norm_in_sq += in_i * in_i;
|
||||
local_norm_tgt_sq += tgt_i * tgt_i;
|
||||
}
|
||||
|
||||
__shared__ double s_data[3];
|
||||
if (threadIdx.x == 0) s_data[0] = 0.0;
|
||||
if (threadIdx.x == 1) s_data[1] = 0.0;
|
||||
if (threadIdx.x == 2) s_data[2] = 0.0;
|
||||
__syncthreads();
|
||||
|
||||
local_dot = block_reduce_sum_double(local_dot);
|
||||
local_norm_in_sq = block_reduce_sum_double(local_norm_in_sq);
|
||||
local_norm_tgt_sq = block_reduce_sum_double(local_norm_tgt_sq);
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
s_data[0] = local_dot;
|
||||
s_data[1] = local_norm_in_sq;
|
||||
s_data[2] = local_norm_tgt_sq;
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
double dot_d = s_data[0];
|
||||
double norm_in_d = sqrt(s_data[1]);
|
||||
double norm_t_d = sqrt(s_data[2]);
|
||||
double norm_mult_d = norm_in_d * norm_t_d + (double)epsilon;
|
||||
double cos_sim_d = dot_d / norm_mult_d;
|
||||
|
||||
loss_n[n_idx] = (float)(1.0 - cos_sim_d);
|
||||
dot_out[n_idx] = (float)dot_d;
|
||||
norm_i_out[n_idx] = (float)norm_in_d;
|
||||
norm_t_out[n_idx] = (float)norm_t_d;
|
||||
}
|
||||
}
|
||||
|
||||
__global__ void angular_bwd_kernel(
|
||||
const double grad_out_scalar_d,
|
||||
const float* __restrict__ grad_output_n,
|
||||
const float* __restrict__ input,
|
||||
const float* __restrict__ target,
|
||||
const float* __restrict__ dot,
|
||||
const float* __restrict__ norm_i,
|
||||
const float* __restrict__ norm_t,
|
||||
float* __restrict__ grad_input,
|
||||
float* __restrict__ grad_target,
|
||||
const int reduction,
|
||||
const int N, const int D,
|
||||
const double eps_d
|
||||
)
|
||||
{
|
||||
int i = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
int stride = blockDim.x * gridDim.x;
|
||||
|
||||
for (int idx = i; idx < N * D; idx += stride) {
|
||||
int n_idx = idx / D;
|
||||
|
||||
const double dot_d = (double)dot[n_idx];
|
||||
const double norm_i_d = (double)norm_i[n_idx];
|
||||
const double norm_t_d = (double)norm_t[n_idx];
|
||||
|
||||
const double u_i = (double)input[idx];
|
||||
const double v_i = (double)target[idx];
|
||||
|
||||
const double grad_out_d = (reduction == 0) ?
|
||||
(double)grad_output_n[n_idx] :
|
||||
grad_out_scalar_d;
|
||||
|
||||
const double norm_i_sq = norm_i_d * norm_i_d + eps_d;
|
||||
const double norm_t_sq = norm_t_d * norm_t_d + eps_d;
|
||||
const double norm_mult = norm_i_d * norm_t_d + eps_d;
|
||||
|
||||
const double cos_sim = dot_d / norm_mult;
|
||||
|
||||
const double grad_u_i = (cos_sim * u_i / norm_i_sq) - (v_i / norm_mult);
|
||||
grad_input[idx] = (float)(grad_u_i * grad_out_d);
|
||||
|
||||
const double grad_v_i = (cos_sim * v_i / norm_t_sq) - (u_i / norm_mult);
|
||||
grad_target[idx] = (float)(grad_v_i * grad_out_d);
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<torch::Tensor> angular_loss_forward_cuda(
|
||||
torch::Tensor input,
|
||||
torch::Tensor target,
|
||||
int reduction,
|
||||
int N,
|
||||
float epsilon)
|
||||
{
|
||||
int D = input.size(1);
|
||||
auto options = input.options();
|
||||
|
||||
torch::Tensor loss_n = torch::empty({N}, options);
|
||||
torch::Tensor dot_out = torch::empty({N}, options);
|
||||
torch::Tensor norm_i_out = torch::empty({N}, options);
|
||||
torch::Tensor norm_t_out = torch::empty({N}, options);
|
||||
|
||||
const int block_size = BLOCK_SIZE;
|
||||
const int grid_size = N;
|
||||
|
||||
angular_fwd_kernel<<<grid_size, block_size>>>(
|
||||
input.data_ptr<float>(),
|
||||
target.data_ptr<float>(),
|
||||
loss_n.data_ptr<float>(),
|
||||
dot_out.data_ptr<float>(),
|
||||
norm_i_out.data_ptr<float>(),
|
||||
norm_t_out.data_ptr<float>(),
|
||||
N, D, epsilon
|
||||
);
|
||||
|
||||
torch::Tensor output;
|
||||
if (reduction == 0) {
|
||||
output = loss_n;
|
||||
} else {
|
||||
output = torch::zeros({1}, options);
|
||||
const int reduce_grid_size = std::min(
|
||||
(int)((N + BLOCK_SIZE - 1) / BLOCK_SIZE),
|
||||
MAX_GRID_SIZE
|
||||
);
|
||||
reduce_kernel<<<reduce_grid_size, BLOCK_SIZE>>>(
|
||||
loss_n.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
N
|
||||
);
|
||||
if (reduction == 1) {
|
||||
output.div_(N);
|
||||
}
|
||||
}
|
||||
|
||||
return {output, dot_out, norm_i_out, norm_t_out};
|
||||
}
|
||||
|
||||
void angular_loss_backward_cuda(
|
||||
float grad_out_scalar,
|
||||
torch::Tensor grad_output_n,
|
||||
torch::Tensor input,
|
||||
torch::Tensor target,
|
||||
torch::Tensor dot,
|
||||
torch::Tensor norm_i,
|
||||
torch::Tensor norm_t,
|
||||
torch::Tensor grad_input,
|
||||
torch::Tensor grad_target,
|
||||
int reduction,
|
||||
int N, int D,
|
||||
float epsilon)
|
||||
{
|
||||
const int64_t n_total = N * D;
|
||||
const int block_size = BLOCK_SIZE;
|
||||
const int grid_size = std::min(
|
||||
(int)((n_total + block_size - 1) / block_size),
|
||||
MAX_GRID_SIZE
|
||||
);
|
||||
|
||||
const float* grad_output_n_ptr = (reduction == 0) ?
|
||||
grad_output_n.data_ptr<float>() :
|
||||
nullptr;
|
||||
|
||||
angular_bwd_kernel<<<grid_size, block_size>>>(
|
||||
(double)grad_out_scalar,
|
||||
grad_output_n_ptr,
|
||||
input.data_ptr<float>(),
|
||||
target.data_ptr<float>(),
|
||||
dot.data_ptr<float>(),
|
||||
norm_i.data_ptr<float>(),
|
||||
norm_t.data_ptr<float>(),
|
||||
grad_input.data_ptr<float>(),
|
||||
grad_target.data_ptr<float>(),
|
||||
reduction,
|
||||
N, D,
|
||||
(double)epsilon
|
||||
);
|
||||
}
|
||||
"""
|
||||
|
||||
self.op = load_inline(
|
||||
name='angular_loss_cuda_v1_full_double',
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=['angular_loss_forward_cuda', 'angular_loss_backward_cuda'],
|
||||
extra_cuda_cflags=['-O3'],
|
||||
verbose=False
|
||||
)
|
||||
|
||||
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
||||
if isinstance(input, (list, tuple)) and len(input) > 0:
|
||||
input = input[0]
|
||||
target = target[0] if len(target) > 0 else target
|
||||
|
||||
return AngularDistanceCUDAOp.apply(
|
||||
input,
|
||||
target,
|
||||
self.beta,
|
||||
self.reduction_id,
|
||||
self.op
|
||||
)
|
||||
|
|
@ -0,0 +1,55 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
N, D = 32, 64
|
||||
|
||||
|
||||
class AngularDistance(nn.Module):
|
||||
def __init__(self, reduction='mean', beta=1.0):
|
||||
super().__init__()
|
||||
self.reduction = reduction
|
||||
self.epsilon = 1e-6
|
||||
|
||||
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
dot_product = (input * target).sum(dim=1)
|
||||
|
||||
norm_input = input.norm(p=2, dim=1)
|
||||
norm_target = target.norm(p=2, dim=1)
|
||||
|
||||
norm_mult = norm_input * norm_target
|
||||
|
||||
cosine_sim = dot_product / (norm_mult + self.epsilon)
|
||||
|
||||
loss = 1.0 - cosine_sim
|
||||
|
||||
if self.reduction == 'mean':
|
||||
return loss.mean()
|
||||
elif self.reduction == 'sum':
|
||||
return loss.sum()
|
||||
else:
|
||||
return loss
|
||||
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, reduction='mean', beta=1.0):
|
||||
super().__init__()
|
||||
self.op = AngularDistance(reduction, beta)
|
||||
|
||||
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
||||
if isinstance(input, (list, tuple)) and len(input) > 0:
|
||||
input = input[0]
|
||||
target = target[0] if len(target) > 0 else target
|
||||
|
||||
return self.op(input, target)
|
||||
|
||||
|
||||
def get_inputs():
|
||||
input = torch.randn(N, D, dtype=torch.float32)
|
||||
target = torch.randn(N, D, dtype=torch.float32)
|
||||
return [input, target]
|
||||
|
||||
|
||||
def get_init_inputs():
|
||||
return ['mean', 1.0]
|
||||
|
|
@ -0,0 +1,120 @@
|
|||
You write custom CUDA kernels to replace the pytorch operators in the given GeGLU architecture to get speedups.
|
||||
|
||||
You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining chunk+gelu+elementwise_mul), or algorithmic changes (such as optimized memory access patterns). You are only limited by your imagination.
|
||||
|
||||
Core Optimization Techniques:
|
||||
|
||||
Performance Optimizations
|
||||
|
||||
Double Precision Reduction - Uses double for accumulation to maintain numerical accuracy
|
||||
|
||||
Two-Stage Reduction - Warp-level shuffle reduction + block-level shared memory reduction
|
||||
|
||||
Parallel Sample Processing - Each CUDA block processes one sample (N-wise parallelization)
|
||||
|
||||
Grid Size Optimization - Limits grid size to MAX_GRID_SIZE for optimal resource usage
|
||||
|
||||
Memory Optimizations
|
||||
|
||||
Intermediate Storage - Stores dot products and norms for backward pass reuse
|
||||
|
||||
Memory Coalescing - Ensures contiguous memory access patterns
|
||||
|
||||
Shared Memory Reduction - Uses shared memory for efficient block-level reductions
|
||||
|
||||
Numerical Stability
|
||||
|
||||
Double Precision - All intermediate calculations use double to prevent precision loss
|
||||
|
||||
Epsilon Protection - Adds epsilon to denominators to prevent division by zero
|
||||
|
||||
Stable Cosine Similarity - Properly handles norm calculations with epsilon protection
|
||||
|
||||
Mathematical Optimizations
|
||||
|
||||
Efficient Cosine Distance - Computes 1 - cos_sim for angular distance
|
||||
|
||||
Analytical Gradients - Implements exact mathematical derivatives for cosine similarity:
|
||||
|
||||
grad_u_i = (cos_sim * u_i / norm_i_sq) - (v_i / norm_mult)
|
||||
|
||||
grad_v_i = (cos_sim * v_i / norm_t_sq) - (u_i / norm_mult)
|
||||
|
||||
Kernel Design
|
||||
|
||||
Separate Forward/Backward Kernels - Optimized kernels for each pass
|
||||
|
||||
Flexible Reduction Support - Handles 'none', 'mean', and 'sum' reduction types
|
||||
|
||||
Atomic Reduction - Uses atomicAdd for efficient multi-block reduction
|
||||
|
||||
Key Features
|
||||
|
||||
High Precision - Double precision ensures numerical accuracy for angular calculations
|
||||
|
||||
Efficient Gradient Computation - Reuses precomputed terms from forward pass
|
||||
|
||||
Proper Normalization - Correctly handles vector normalization in both forward and backward passes
|
||||
|
||||
Memory Efficient - Stores only necessary intermediate terms for gradients
|
||||
|
||||
This implementation provides highly accurate angular distance computation with proper gradient propagation, essential for metric learning and similarity-based tasks
|
||||
|
||||
|
||||
|
||||
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
N, D = 32, 64
|
||||
|
||||
|
||||
class AngularDistance(nn.Module):
|
||||
def __init__(self, reduction='mean', beta=1.0):
|
||||
super().__init__()
|
||||
self.reduction = reduction
|
||||
self.epsilon = 1e-6
|
||||
|
||||
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
dot_product = (input * target).sum(dim=1)
|
||||
|
||||
norm_input = input.norm(p=2, dim=1)
|
||||
norm_target = target.norm(p=2, dim=1)
|
||||
|
||||
norm_mult = norm_input * norm_target
|
||||
|
||||
cosine_sim = dot_product / (norm_mult + self.epsilon)
|
||||
|
||||
loss = 1.0 - cosine_sim
|
||||
|
||||
if self.reduction == 'mean':
|
||||
return loss.mean()
|
||||
elif self.reduction == 'sum':
|
||||
return loss.sum()
|
||||
else:
|
||||
return loss
|
||||
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, reduction='mean', beta=1.0):
|
||||
super().__init__()
|
||||
self.op = AngularDistance(reduction, beta)
|
||||
|
||||
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
||||
if isinstance(input, (list, tuple)) and len(input) > 0:
|
||||
input = input[0]
|
||||
target = target[0] if len(target) > 0 else target
|
||||
|
||||
return self.op(input, target)
|
||||
|
||||
|
||||
def get_inputs():
|
||||
input = torch.randn(N, D, dtype=torch.float32)
|
||||
target = torch.randn(N, D, dtype=torch.float32)
|
||||
return [input, target]
|
||||
|
||||
|
||||
def get_init_inputs():
|
||||
return ['mean', 1.0]
|
||||
|
|
@ -0,0 +1,72 @@
|
|||
import torch
|
||||
import time
|
||||
from AngularDistance_torch import Model, get_inputs, get_init_inputs
|
||||
from AngularDistance_cuda import ModelNew
|
||||
|
||||
def run_benchmark():
|
||||
# 检查 CUDA 是否可用
|
||||
if not torch.cuda.is_available():
|
||||
print("CUDA 不可用，请确保您有可用的 NVIDIA GPU 并已正确安装 PyTorch CUDA 版本。")
|
||||
return
|
||||
else:
|
||||
device = torch.device("cuda")
|
||||
|
||||
# 初始化模型
|
||||
init_inputs = get_init_inputs()
|
||||
init_inputs = [
|
||||
x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in init_inputs
|
||||
]
|
||||
inputs = get_inputs()
|
||||
inputs = [
|
||||
x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in inputs
|
||||
]
|
||||
|
||||
torch_model = Model(*init_inputs).cuda()
|
||||
cuda_model = ModelNew(*init_inputs).cuda()
|
||||
|
||||
torch_model.eval()
|
||||
cuda_model.eval()
|
||||
|
||||
print("-------------------- 精度对齐验证 --------------------")
|
||||
with torch.no_grad():
|
||||
output_torch = torch_model(*inputs)
|
||||
output_cuda = cuda_model(*inputs)
|
||||
|
||||
precision_flag = torch.allclose(output_torch, output_cuda, rtol=1e-03)
|
||||
if precision_flag:
|
||||
print("✅ 精度对齐：两个模型的输出结果非常接近。")
|
||||
else:
|
||||
print("❌ 精度不一致！")
|
||||
|
||||
print("\n-------------------- 性能加速比测试 --------------------")
|
||||
num_iterations = 100
|
||||
|
||||
# PyTorch 模型计时
|
||||
torch.cuda.synchronize()
|
||||
start_time = time.time()
|
||||
for _ in range(num_iterations):
|
||||
_ = torch_model(*inputs)
|
||||
torch.cuda.synchronize()
|
||||
torch_time = (time.time() - start_time) / num_iterations
|
||||
|
||||
# 自定义 CUDA 内核计时
|
||||
torch.cuda.synchronize()
|
||||
start_time = time.time()
|
||||
for _ in range(num_iterations):
|
||||
_ = cuda_model(*inputs)
|
||||
torch.cuda.synchronize()
|
||||
cuda_time = (time.time() - start_time) / num_iterations
|
||||
|
||||
print(f"PyTorch torch.relu 平均执行时间: {torch_time:.6f} 秒")
|
||||
print(f"自定义 CUDA 内核 平均执行时间: {cuda_time:.6f} 秒")
|
||||
speedup = 0
|
||||
if cuda_time > 0:
|
||||
speedup = torch_time / cuda_time
|
||||
print(f"加速比 (Speedup): {speedup:.2f}x")
|
||||
else:
|
||||
print("CUDA 内核执行时间为0，无法计算加速比。")
|
||||
return precision_flag, speedup
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
precision_flag, speedup = run_benchmark()
|
||||
Loading…
Reference in New Issue