Merge pull request 'finish Polar-affine #95' (#496) from gsd123/GPUCodeForces:gsd95 into main

S1/gsd123_#95/polaraffine_cuda.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+
+class ModelNew(nn.Module):
+    def __init__(self, num_features=512):
+        super().__init__()
+        assert num_features % 2 == 0
+        self.num_features = num_features
+        self.weight_r = nn.Parameter(torch.ones(num_features // 2))
+        self.bias_r = nn.Parameter(torch.zeros(num_features // 2))
+        self.weight_theta = nn.Parameter(torch.ones(num_features // 2))
+        self.bias_theta = nn.Parameter(torch.zeros(num_features // 2))
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        torch::Tensor polaraffine_cuda(
+            torch::Tensor x,
+            torch::Tensor weight_r,
+            torch::Tensor bias_r,
+            torch::Tensor weight_theta,
+            torch::Tensor bias_theta);
+        """
+
+        cuda_source = """
+        #include <torch/extension.h>
+        #include <cuda_runtime.h>
+        #include <math.h>
+
+        __global__ void polaraffine_kernel(
+            const float* __restrict__ x,
+            const float* __restrict__ weight_r,
+            const float* __restrict__ bias_r,
+            const float* __restrict__ weight_theta,
+            const float* __restrict__ bias_theta,
+            float* __restrict__ output,
+            const int rows,
+            const int cols)
+        {
+            const int n_pairs = cols / 2;
+            const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+            const int stride = blockDim.x * gridDim.x;
+            const int total_pairs = rows * n_pairs;
+
+            for (int i = tid; i < total_pairs; i += stride) {
+                const int r_idx = i / n_pairs;
+                const int p_idx = i % n_pairs;
+
+                const int idx_even = r_idx * cols + 2 * p_idx;
+                const int idx_odd  = r_idx * cols + 2 * p_idx + 1;
+
+                float x_ev = x[idx_even];
+                float x_od = x[idx_odd];
+
+                float r = hypotf(x_ev, x_od);
+                float theta = atan2f(x_od, x_ev);
+
+                float w_r = weight_r[p_idx];
+                float b_r = bias_r[p_idx];
+                float w_th = weight_theta[p_idx];
+                float b_th = bias_theta[p_idx];
+
+                float r_new = __fadd_rn(__fmul_rn(r, w_r), b_r);
+                float theta_new = __fadd_rn(__fmul_rn(theta, w_th), b_th);
+
+                float c = cosf(theta_new);
+                float s = sinf(theta_new);
+
+                output[idx_even] = __fmul_rn(r_new, c);
+                output[idx_odd]  = __fmul_rn(r_new, s);
+            }
+        }
+
+        torch::Tensor polaraffine_cuda(
+            torch::Tensor x,
+            torch::Tensor weight_r,
+            torch::Tensor bias_r,
+            torch::Tensor weight_theta,
+            torch::Tensor bias_theta) 
+        {
+            auto x_c = x.contiguous();
+            auto wr_c = weight_r.contiguous();
+            auto br_c = bias_r.contiguous();
+            auto wth_c = weight_theta.contiguous();
+            auto bth_c = bias_theta.contiguous();
+
+            const int rows = x_c.size(0);
+            const int cols = x_c.size(1);
+
+            auto output = torch::empty_like(x_c);
+
+            const int total_pairs = rows * (cols / 2);
+            const int threads = 256;
+            const int blocks = min((total_pairs + threads - 1) / threads, 65535);
+
+            polaraffine_kernel<<<blocks, threads>>>(
+                x_c.data_ptr<float>(),
+                wr_c.data_ptr<float>(),
+                br_c.data_ptr<float>(),
+                wth_c.data_ptr<float>(),
+                bth_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                rows,
+                cols
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="polaraffine_op",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["polaraffine_cuda"],
+            extra_cuda_cflags=["-O3", "-fmad=false"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.polaraffine_cuda(
+            x, self.weight_r, self.bias_r, self.weight_theta, self.bias_theta
+        )

S1/gsd123_#95/polaraffine_torch.py

@@ -0,0 +1,44 @@
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self, num_features=512):
+        super().__init__()
+        assert num_features % 2 == 0
+        self.num_features = num_features
+        self.weight_r = nn.Parameter(torch.ones(num_features // 2))
+        self.bias_r = nn.Parameter(torch.zeros(num_features // 2))
+        self.weight_theta = nn.Parameter(torch.ones(num_features // 2))
+        self.bias_theta = nn.Parameter(torch.zeros(num_features // 2))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x_even = x[:, 0::2]
+        x_odd = x[:, 1::2]
+
+        r = torch.hypot(x_even, x_odd)
+        theta = torch.atan2(x_odd, x_even)
+
+        r_new = r * self.weight_r + self.bias_r
+        theta_new = theta * self.weight_theta + self.bias_theta
+
+        x_even_new = r_new * torch.cos(theta_new)
+        x_odd_new = r_new * torch.sin(theta_new)
+
+        y = torch.empty_like(x)
+        y[:, 0::2] = x_even_new
+        y[:, 1::2] = x_odd_new
+        return y
+
+
+batch_size = 128
+feature_dim = 512
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return []

S1/gsd123_#95/prompt.txt

@@ -0,0 +1,71 @@
+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.
+Custom CUDA kernel extension via torch.utils.cpp_extension.load_inline
+
+Polar coordinate transformation: (x,y) ↔ (r,θ) with learnable affine transforms
+
+Element-wise parallelization using CUDA grid-stride loops over feature pairs
+
+Per-feature-pair learnable parameters for radius and angle (weight_r, bias_r, weight_θ, bias_θ)
+
+Contiguous tensor handling for all input tensors
+
+Memory-efficient in-place-like computation with torch.empty_like
+
+Mathematical operations: hypotf, atan2f, cosf, sinf
+
+Fused multiply-add with __fmul_rn and __fadd_rn for precision control
+
+Auto-tuning block/grid size based on number of feature pairs
+
+Even-odd feature pairing for polar coordinate processing
+
+
+
+
+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
+
+
+class Model(nn.Module):
+    def __init__(self, num_features=512):
+        super().__init__()
+        assert num_features % 2 == 0
+        self.num_features = num_features
+        self.weight_r = nn.Parameter(torch.ones(num_features // 2))
+        self.bias_r = nn.Parameter(torch.zeros(num_features // 2))
+        self.weight_theta = nn.Parameter(torch.ones(num_features // 2))
+        self.bias_theta = nn.Parameter(torch.zeros(num_features // 2))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x_even = x[:, 0::2]
+        x_odd = x[:, 1::2]
+
+        r = torch.hypot(x_even, x_odd)
+        theta = torch.atan2(x_odd, x_even)
+
+        r_new = r * self.weight_r + self.bias_r
+        theta_new = theta * self.weight_theta + self.bias_theta
+
+        x_even_new = r_new * torch.cos(theta_new)
+        x_odd_new = r_new * torch.sin(theta_new)
+
+        y = torch.empty_like(x)
+        y[:, 0::2] = x_even_new
+        y[:, 1::2] = x_odd_new
+        return y
+
+
+batch_size = 128
+feature_dim = 512
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return []

S1/gsd123_#95/run_code.py

@@ -0,0 +1,77 @@
+###########################################################
+# 性能和精度验证程序
+###########################################################
+import torch
+import torch.nn as nn
+import time
+from polaraffine_torch import Model, get_inputs, get_init_inputs
+from polaraffine_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()