finish conj_physical #36

S1/hli28146_#36/conjphysical_cuda.py

@@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+ 
+cpp_source = """
+#include <torch/extension.h>
+
+torch::Tensor conj_physical_cuda_forward(const torch::Tensor& input);
+"""
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+
+struct __align__(16) Float4 {
+    float x, y, z, w; // R1, I1, R2, I2
+};
+
+struct __align__(8) Float2 {
+    float x, y; // R, I
+};
+
+// Core logic: z = x + iy -> z* = x - iy
+// We manipulate raw floats to avoid complex class overhead
+__global__ void conj_physical_kernel(
+    const float* __restrict__ input, 
+    float* __restrict__ output,
+    const int n_complex_elements) 
+{
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int stride = blockDim.x * gridDim.x;
+    
+    int vec_loops = n_complex_elements / 2;
+
+    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 val = vec_input[i];
+        
+        // Logical layout: x=Real1, y=Imag1, z=Real2, w=Imag2
+        // Operation: Negate Imag parts
+        val.y = -val.y;
+        val.w = -val.w;
+        
+        vec_output[i] = val;
+    }
+
+    // Only happens if n_complex_elements is odd
+    int tail_idx = vec_loops * 2; 
+    
+    // We switch to Float2 pointer to access single complex elements
+    const Float2* scalar_input = reinterpret_cast<const Float2*>(input);
+    Float2* scalar_output = reinterpret_cast<Float2*>(output);
+    
+    // Standard grid stride logic applied to the tail part
+    // Though usually this loop runs at most once per thread if aligned
+    for (int i = tail_idx + idx; i < n_complex_elements; i += stride) {
+        Float2 val = scalar_input[i];
+        val.y = -val.y; // Negate Imag
+        scalar_output[i] = val;
+    }
+}
+
+torch::Tensor conj_physical_cuda_forward(const torch::Tensor& input) {
+    TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
+    TORCH_CHECK(input.is_contiguous(), "Input must be contiguous");
+    TORCH_CHECK(input.scalar_type() == torch::kComplexFloat, "Input must be ComplexFloat (complex64)");
+
+    int n_elements = input.numel();
+    auto output = torch::empty_like(input);
+
+    const int block_size = 256;
+    // Each thread handles 2 elements ideally
+    int num_vectors = (n_elements + 1) / 2; 
+    int grid_size = (num_vectors + block_size - 1) / block_size;
+    if (grid_size > 65535) grid_size = 65535;
+
+    conj_physical_kernel<<<grid_size, block_size>>>(
+        reinterpret_cast<float*>(input.data_ptr<c10::complex<float>>()),
+        reinterpret_cast<float*>(output.data_ptr<c10::complex<float>>()),
+        n_elements
+    );
+
+    return output;
+}
+"""
+
+conj_op = load_inline(
+    name='conj_physical_op',
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=['conj_physical_cuda_forward'],
+    verbose=False,
+    extra_cuda_cflags=['-O3']
+)
+
+class ConjNew(nn.Module):
+    def __init__(self):
+        super(ConjNew, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return conj_op.conj_physical_cuda_forward(x)
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+        self.act = ConjNew()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)

S1/hli28146_#36/conjphysical_torch.py

@@ -0,0 +1,28 @@
+import torch
+import torch.nn as nn
+
+BATCH_SIZE = 4096
+DIM = 4096
+SHAPE = (BATCH_SIZE, DIM)
+
+class ConjPhysicalModel(nn.Module):
+    def __init__(self):
+        super(ConjPhysicalModel, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.conj_physical(x)
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.act = ConjPhysicalModel()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)
+
+def get_inputs():
+    x = torch.randn(SHAPE, dtype=torch.complex64)
+    return [x.contiguous()]
+
+def get_init_inputs():
+    return []

S1/hli28146_#36/prompt.txt

@@ -0,0 +1,54 @@
+Write a custom CUDA kernel to optimize `torch.conj_physical` for complex tensors.
+
+The operation computes the element-wise conjugate of a complex tensor. For z = x + iy, conj_physical(z) = x - iy. It explicitly materializes the result in memory.
+
+Problem Analysis:
+This is a strictly memory-bound operation.
+1. Data Layout: `complex64` stores data as contiguous pairs of floats [Real, Imag].
+2. Computation: The only arithmetic operation is negating the imaginary part.
+3. Bottleneck: The performance is strictly limited by Global Memory bandwidth.
+
+Optimization Strategy: Vectorized Access (2x Complex Elements per Thread)
+
+1. Vectorized I/O (Float4): 
+   - A single `complex64` is 8 bytes (2 floats).
+   - Using `float4` (16 bytes) allows a single thread to load/store **two** complex numbers at once.
+   - Layout loaded into registers: `x`=Real1, `y`=Imag1, `z`=Real2, `w`=Imag2.
+
+2. In-Register Computation:
+   - Negate the `y` and `w` components (the imaginary parts).
+   - Store the modified `float4` back to global memory.
+
+3. Grid-Stride Loop: Implement a robust grid-stride loop to handle arbitrary tensor sizes, processing 2 complex elements per iteration in the vectorized loop, and handling remainders with a scalar loop.
+  
+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 = 4096
+DIM = 4096
+SHAPE = (BATCH_SIZE, DIM)
+
+class ConjPhysicalModel(nn.Module):
+    def __init__(self):
+        super(ConjPhysicalModel, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.conj_physical(x)
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.act = ConjPhysicalModel()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)
+
+def get_inputs():
+    x = torch.randn(SHAPE, dtype=torch.complex64)
+    return [x.contiguous()]
+
+def get_init_inputs():
+    return []

S1/hli28146_#36/run_code.py

@@ -0,0 +1,74 @@
+###########################################################
+# 性能和精度验证程序
+###########################################################
+import torch
+import torch.nn as nn
+import time
+from conjphysical_torch import Model,get_inputs,get_init_inputs
+from conjphysical_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()