forked from ccf-ai-infra/GPUCodeForces
finish fake_quantize_per_tensor_affine #38
This commit is contained in:
commit
53313e7a91
|
|
@ -0,0 +1,148 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
|
||||
torch::Tensor fake_quantize_cuda_forward(
|
||||
const torch::Tensor& input,
|
||||
const torch::Tensor& scale,
|
||||
const torch::Tensor& zero_point,
|
||||
int64_t quant_min,
|
||||
int64_t quant_max);
|
||||
"""
|
||||
|
||||
cuda_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <math.h>
|
||||
|
||||
struct __align__(16) Float4 {
|
||||
float x, y, z, w;
|
||||
};
|
||||
|
||||
// Core logic based on user request:
|
||||
// q = clamp(round(x / s) + z, qmin, qmax)
|
||||
// out = (q - z) * s
|
||||
__device__ __forceinline__ float fake_quant_op(
|
||||
float x, float s, int z, int qmin, int qmax)
|
||||
{
|
||||
// 1. Division
|
||||
float val = x / s;
|
||||
|
||||
// 2. Rounding (rintf rounds to nearest integer, ties to even)
|
||||
// Formula correction: round(x/s) then add z
|
||||
val = rintf(val);
|
||||
|
||||
// 3. Add Zero Point
|
||||
val += (float)z;
|
||||
|
||||
// 4. Clamp
|
||||
val = fmaxf(val, (float)qmin);
|
||||
val = fminf(val, (float)qmax);
|
||||
|
||||
// 5. Dequantize
|
||||
return (val - (float)z) * s;
|
||||
}
|
||||
|
||||
__global__ void fake_quantize_kernel(
|
||||
const float* __restrict__ input,
|
||||
float* __restrict__ output,
|
||||
const float* __restrict__ scale_ptr,
|
||||
const int* __restrict__ zero_point_ptr,
|
||||
const int qmin,
|
||||
const int qmax,
|
||||
const int n_elements)
|
||||
{
|
||||
// Load scalars once per thread/block (cached)
|
||||
float s = *scale_ptr;
|
||||
int z = *zero_point_ptr;
|
||||
|
||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
int stride = blockDim.x * gridDim.x;
|
||||
|
||||
// 1. Vectorized Loop
|
||||
int vec_loops = n_elements / 4;
|
||||
const Float4* vec_input = reinterpret_cast<const Float4*>(input);
|
||||
Float4* vec_output = reinterpret_cast<Float4*>(output);
|
||||
|
||||
for (int i = idx; i < vec_loops; i += stride) {
|
||||
Float4 in_val = vec_input[i];
|
||||
Float4 out_val;
|
||||
|
||||
out_val.x = fake_quant_op(in_val.x, s, z, qmin, qmax);
|
||||
out_val.y = fake_quant_op(in_val.y, s, z, qmin, qmax);
|
||||
out_val.z = fake_quant_op(in_val.z, s, z, qmin, qmax);
|
||||
out_val.w = fake_quant_op(in_val.w, s, z, qmin, qmax);
|
||||
|
||||
vec_output[i] = out_val;
|
||||
}
|
||||
|
||||
// 2. Tail Loop
|
||||
int tail_start = vec_loops * 4;
|
||||
for (int i = tail_start + idx; i < n_elements; i += stride) {
|
||||
output[i] = fake_quant_op(input[i], s, z, qmin, qmax);
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor fake_quantize_cuda_forward(
|
||||
const torch::Tensor& input,
|
||||
const torch::Tensor& scale,
|
||||
const torch::Tensor& zero_point,
|
||||
int64_t quant_min,
|
||||
int64_t quant_max)
|
||||
{
|
||||
TORCH_CHECK(input.is_cuda(), "Input must be CUDA");
|
||||
TORCH_CHECK(input.is_contiguous(), "Input must be contiguous");
|
||||
TORCH_CHECK(scale.is_cuda(), "Scale must be CUDA");
|
||||
TORCH_CHECK(zero_point.is_cuda(), "Zero Point must be CUDA");
|
||||
|
||||
auto output = torch::empty_like(input);
|
||||
int n_elements = input.numel();
|
||||
|
||||
const int block_size = 256;
|
||||
int grid_size = (n_elements + block_size * 4 - 1) / (block_size * 4);
|
||||
if (grid_size > 65535) grid_size = 65535;
|
||||
|
||||
fake_quantize_kernel<<<grid_size, block_size>>>(
|
||||
input.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
scale.data_ptr<float>(),
|
||||
zero_point.data_ptr<int>(),
|
||||
(int)quant_min,
|
||||
(int)quant_max,
|
||||
n_elements
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
fake_quant_op = load_inline(
|
||||
name='fake_quant_op',
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=['fake_quantize_cuda_forward'],
|
||||
verbose=False,
|
||||
extra_cuda_cflags=['-O3', '--use_fast_math']
|
||||
)
|
||||
|
||||
class FakeQuantizeNew(nn.Module):
|
||||
def __init__(self, quant_min, quant_max):
|
||||
super(FakeQuantizeNew, self).__init__()
|
||||
self.quant_min = quant_min
|
||||
self.quant_max = quant_max
|
||||
|
||||
def forward(self, x, scale, zero_point):
|
||||
return fake_quant_op.fake_quantize_cuda_forward(
|
||||
x, scale, zero_point, self.quant_min, self.quant_max
|
||||
)
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self):
|
||||
super(ModelNew, self).__init__()
|
||||
self.fq = FakeQuantizeNew(-128, 127)
|
||||
|
||||
def forward(self, x, scale, zero_point):
|
||||
return self.fq(x, scale, zero_point)
|
||||
|
|
@ -0,0 +1,41 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
BATCH_SIZE = 2048
|
||||
DIM = 2048
|
||||
SHAPE = (BATCH_SIZE, DIM)
|
||||
|
||||
QMIN = -128
|
||||
QMAX = 127
|
||||
|
||||
class FakeQuantize(nn.Module):
|
||||
"""
|
||||
Fake Quantize Per Tensor Affine
|
||||
"""
|
||||
def __init__(self, quant_min=QMIN, quant_max=QMAX):
|
||||
super(FakeQuantize, self).__init__()
|
||||
self.quant_min = quant_min
|
||||
self.quant_max = quant_max
|
||||
|
||||
def forward(self, x, scale, zero_point):
|
||||
return torch.fake_quantize_per_tensor_affine(
|
||||
x, scale, zero_point, self.quant_min, self.quant_max
|
||||
)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super(Model, self).__init__()
|
||||
self.fq = FakeQuantize()
|
||||
|
||||
def forward(self, x, scale, zero_point):
|
||||
return self.fq(x, scale, zero_point)
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(SHAPE, dtype=torch.float32)
|
||||
scale = torch.tensor([0.05], dtype=torch.float32)
|
||||
zero_point = torch.tensor([10], dtype=torch.int32)
|
||||
|
||||
return [x.contiguous(), scale, zero_point]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
|
|
@ -0,0 +1,64 @@
|
|||
Write a custom CUDA kernel to optimize the Fake Quantize operation based on a specific formula.
|
||||
|
||||
The mathematical definition provided is:
|
||||
output = (clamp(round(input / scale) + zero_point, quant_min, quant_max) - zero_point) * scale
|
||||
|
||||
Inputs:
|
||||
- input: Float32 Tensor.
|
||||
- scale: Scalar Float32.
|
||||
- zero_point: Scalar Int32.
|
||||
- quant_min, quant_max: Scalar Int32.
|
||||
|
||||
Problem Analysis:
|
||||
This operation simulates quantization error. It involves element-wise division, rounding, addition, clamping, subtraction, and multiplication. Doing this naively involves high memory bandwidth usage.
|
||||
|
||||
Optimization Strategy:
|
||||
1. **Fused Kernel**: Perform the entire logic in a single CUDA kernel pass.
|
||||
2. **Specific Rounding Logic**: Implement `round(input/scale) + zero_point` (rounding before adding zero_point) as requested. Use `rintf` for "round to nearest even".
|
||||
3. **Vectorized Access**: Use `float4` loads/stores to process 4 elements per thread, maximizing memory bandwidth.
|
||||
4. **Scalar Optimization**: Load `scale` and `zero_point` once per thread/block from global memory and keep them in registers.
|
||||
|
||||
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
|
||||
|
||||
```python
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
BATCH_SIZE = 2048
|
||||
DIM = 2048
|
||||
SHAPE = (BATCH_SIZE, DIM)
|
||||
|
||||
QMIN = -128
|
||||
QMAX = 127
|
||||
|
||||
class FakeQuantize(nn.Module):
|
||||
"""
|
||||
Fake Quantize Per Tensor Affine
|
||||
"""
|
||||
def __init__(self, quant_min=QMIN, quant_max=QMAX):
|
||||
super(FakeQuantize, self).__init__()
|
||||
self.quant_min = quant_min
|
||||
self.quant_max = quant_max
|
||||
|
||||
def forward(self, x, scale, zero_point):
|
||||
return torch.fake_quantize_per_tensor_affine(
|
||||
x, scale, zero_point, self.quant_min, self.quant_max
|
||||
)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super(Model, self).__init__()
|
||||
self.fq = FakeQuantize()
|
||||
|
||||
def forward(self, x, scale, zero_point):
|
||||
return self.fq(x, scale, zero_point)
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(SHAPE, dtype=torch.float32)
|
||||
scale = torch.tensor([0.05], dtype=torch.float32)
|
||||
zero_point = torch.tensor([10], dtype=torch.int32)
|
||||
|
||||
return [x.contiguous(), scale, zero_point]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
|
|
@ -0,0 +1,74 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from fake_quantize_per_tensor_affine_torch import Model,get_inputs,get_init_inputs
|
||||
from fake_quantize_per_tensor_affine_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