From 1e3cc5a3c18a42ecfdb25bad2fdf000354d033e3 Mon Sep 17 00:00:00 2001 From: ZZZJ <3056485198@qq.com> Date: Tue, 9 Dec 2025 15:23:50 +0800 Subject: [PATCH 1/3] fixes AdaptiveAvgPool3d #4 --- S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py | 191 ++++++++++++++++++++++++ S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py | 31 ++++ S1/ZZZJ_#4/prompt.txt | 39 +++++ S1/ZZZJ_#4/run_code.py | 74 +++++++++ 4 files changed, 335 insertions(+) create mode 100644 S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py create mode 100644 S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py create mode 100644 S1/ZZZJ_#4/prompt.txt create mode 100644 S1/ZZZJ_#4/run_code.py diff --git a/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py b/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py new file mode 100644 index 00000000..b0f696b7 --- /dev/null +++ b/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py @@ -0,0 +1,191 @@ +# adaptive_pool3d_cuda.py +import torch +import torch.nn as nn +from torch.utils.cpp_extension import load_inline +import math + +from adaptive_pool3d_torch import BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, D_OUT, H_OUT, W_OUT + + +BLOCK_SIZE = 256 +VEC_SIZE = 4 + +class ModelNew(nn.Module): + + def __init__(self, output_size): + super().__init__() + self.output_size = output_size + self.d_in = D_IN + self.h_in = H_IN + self.w_in = W_IN + self.d_out = output_size[0] + self.h_out = output_size[1] + self.w_out = output_size[2] + self.block_size = BLOCK_SIZE + self._compile_cuda_kernel() + + def _compile_cuda_kernel(self): + + cpp_header = """ + #include + + torch::Tensor adaptive_avgpool3d_forward_cuda( + torch::Tensor input, int D_in, int H_in, int W_in, int D_out, int H_out, int W_out + ); + """ + + cuda_source = f""" + #include + #include + #include + + #define BLOCK_SIZE {BLOCK_SIZE} + #define VEC_SIZE {VEC_SIZE} + + __global__ void adaptive_avgpool3d_kernel( + const float* __restrict__ input_data, + float* __restrict__ output_data, + int N, int C, int D_in, int H_in, int W_in, + int D_out, int H_out, int W_out + ) {{ + const int N_C_D_H_W_out = N * C * D_out * H_out * W_out; + const int tid = blockIdx.x * blockDim.x + threadIdx.x; + const int grid_stride = gridDim.x * blockDim.x; + + const int CHW_in = C * D_in * H_in * W_in; + const int DHW_in = D_in * H_in * W_in; + const int HW_in = H_in * W_in; + + for (int idx = tid; idx < N_C_D_H_W_out; idx += grid_stride) {{ + + const int w_out = idx % W_out; + const int h_w_out = idx / W_out; + + const int h_out = h_w_out % H_out; + const int d_h_w_out = h_w_out / H_out; + + const int d_out = d_h_w_out % D_out; + const int n_c = d_h_w_out / D_out; + + const int n_idx = n_c / C; + const int c_idx = n_c % C; + + + const int d_in_start = (d_out * D_in) / D_out; + const int d_in_end = ((d_out + 1) * D_in) / D_out; + const int d_range = d_in_end - d_in_start; + + const int h_in_start = (h_out * H_in) / H_out; + const int h_in_end = ((h_out + 1) * H_in) / H_out; + const int h_range = h_in_end - h_in_start; + + const int w_in_start = (w_out * W_in) / W_out; + const int w_in_end = ((w_out + 1) * W_in) / W_out; + const int w_range = w_in_end - w_in_start; + + + double thread_sum = 0.0; + const int kernel_size = d_range * h_range * w_range; + + if (kernel_size == 0) {{ + output_data[idx] = 0.0f; + continue; + }} + + const int base_offset = (n_idx * CHW_in) + (c_idx * DHW_in); + + + for (int d = d_in_start; d < d_in_end; d++) {{ + for (int h = h_in_start; h < h_in_end; h++) {{ + + + const int w_start_idx = base_offset + (d * HW_in) + (h * W_in) + w_in_start; + + int w_current = 0; + int w_len = w_in_end - w_in_start; + + + while (w_current < w_len && ((w_in_start + w_current) % VEC_SIZE) != 0) {{ + thread_sum += (double)input_data[w_start_idx + w_current]; + w_current++; + }} + + + const int w_vector_len = w_len - w_current; + const int num_vectors = w_vector_len / VEC_SIZE; + + if (num_vectors > 0) {{ + const float4* vec_ptr = (const float4*)(input_data + w_start_idx + w_current); + + for (int v = 0; v < num_vectors; v++) {{ + float4 val4 = vec_ptr[v]; + thread_sum += (double)val4.x + (double)val4.y + (double)val4.z + (double)val4.w; + }} + w_current += num_vectors * VEC_SIZE; + }} + + + while (w_current < w_len) {{ + thread_sum += (double)input_data[w_start_idx + w_current]; + w_current++; + }} + + + }} + }} + + + output_data[idx] = (float)(thread_sum / (double)kernel_size); + }} + }} + + + torch::Tensor adaptive_avgpool3d_forward_cuda( + torch::Tensor input, int D_in, int H_in, int W_in, int D_out, int H_out, int W_out + ) {{ + TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor"); + TORCH_CHECK(input.is_contiguous(), "input must be contiguous"); + TORCH_CHECK(input.dim() == 5, "input must be 5D (N, C, D_in, H_in, W_in)"); + + const int N = input.size(0); + const int C = input.size(1); + + const int N_elements_out = N * C * D_out * H_out * W_out; + + auto output = torch::empty({{N, C, D_out, H_out, W_out}}, input.options()); + + dim3 block_dim(BLOCK_SIZE); + const int grid_size = (N_elements_out + BLOCK_SIZE - 1) / BLOCK_SIZE; + dim3 grid_dim(grid_size); + + adaptive_avgpool3d_kernel<<>>( + input.data_ptr(), + output.data_ptr(), + N, C, D_in, H_in, W_in, + D_out, H_out, W_out + ); + + return output; + }} + """ + + + self.pad_op = load_inline( + name="adaptive_avgpool3d_op", + cpp_sources=cpp_header, + cuda_sources=cuda_source, + functions=["adaptive_avgpool3d_forward_cuda"], + verbose=False + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + d_out, h_out, w_out = self.output_size + return self.pad_op.adaptive_avgpool3d_forward_cuda( + x.contiguous(), + self.d_in, + self.h_in, + self.w_in, + d_out, + h_out, + w_out + ) \ No newline at end of file diff --git a/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py b/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py new file mode 100644 index 00000000..69784a44 --- /dev/null +++ b/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py @@ -0,0 +1,31 @@ +# adaptive_pool3d_torch.py +import torch +import torch.nn as nn +import torch.nn.functional as F + + +BATCH_SIZE = 4 +CHANNELS = 64 +D_IN, H_IN, W_IN = 32, 32, 32 +D_OUT, H_OUT, W_OUT = 4, 4, 4 + + +class Model(nn.Module): + + + def __init__(self, output_size): + super().__init__() + self.adaptive_pool = nn.AdaptiveAvgPool3d(output_size) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return self.adaptive_pool(x) + + +def get_inputs(): + + x = torch.randn(BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, dtype=torch.float32) + return [x] + + +def get_init_inputs(): + return [(D_OUT, H_OUT, W_OUT)] \ No newline at end of file diff --git a/S1/ZZZJ_#4/prompt.txt b/S1/ZZZJ_#4/prompt.txt new file mode 100644 index 00000000..ebc80150 --- /dev/null +++ b/S1/ZZZJ_#4/prompt.txt @@ -0,0 +1,39 @@ +You write custom CUDA kernels to replace the pytorch operators in the given 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 matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination. + +Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is: + +python +# adaptive_pool3d_torch.py +import torch +import torch.nn as nn +import torch.nn.functional as F + + +BATCH_SIZE = 4 +CHANNELS = 64 +D_IN, H_IN, W_IN = 32, 32, 32 +D_OUT, H_OUT, W_OUT = 4, 4, 4 + + +class Model(nn.Module): + + + def __init__(self, output_size): + super().__init__() + self.adaptive_pool = nn.AdaptiveAvgPool3d(output_size) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return self.adaptive_pool(x) + + +def get_inputs(): + + x = torch.randn(BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, dtype=torch.float32) + return [x] + + +def get_init_inputs(): + return [(D_OUT, H_OUT, W_OUT)] +``` \ No newline at end of file diff --git a/S1/ZZZJ_#4/run_code.py b/S1/ZZZJ_#4/run_code.py new file mode 100644 index 00000000..36fef309 --- /dev/null +++ b/S1/ZZZJ_#4/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 adaptive_avg_pool3d_torch import Model,get_inputs,get_init_inputs +from adaptive_avg_pool3d_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() \ No newline at end of file From d97a4f7b648208c704090d081a031f031b27c63d Mon Sep 17 00:00:00 2001 From: ZZZJ <3056485198@qq.com> Date: Tue, 9 Dec 2025 15:25:00 +0800 Subject: [PATCH 2/3] fixes AdaptiveAvgPool3d #4 --- S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py b/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py index b0f696b7..b2b8a298 100644 --- a/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py +++ b/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py @@ -4,7 +4,7 @@ import torch.nn as nn from torch.utils.cpp_extension import load_inline import math -from adaptive_pool3d_torch import BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, D_OUT, H_OUT, W_OUT +from adaptive_avg_pool3d_torch import BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, D_OUT, H_OUT, W_OUT BLOCK_SIZE = 256 From 43363e641214341659ad282d317bdc81ee692f79 Mon Sep 17 00:00:00 2001 From: ZZZJ <3056485198@qq.com> Date: Tue, 9 Dec 2025 15:41:39 +0800 Subject: [PATCH 3/3] fixes AdaptiveAvgPool3d #4 --- S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py | 386 +++++++++++++----------- S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py | 37 +-- S1/ZZZJ_#4/prompt.txt | 37 +-- 3 files changed, 243 insertions(+), 217 deletions(-) diff --git a/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py b/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py index b2b8a298..b5ba1bde 100644 --- a/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py +++ b/S1/ZZZJ_#4/adaptive_avg_pool3d_cuda.py @@ -1,191 +1,211 @@ -# adaptive_pool3d_cuda.py import torch -import torch.nn as nn from torch.utils.cpp_extension import load_inline -import math -from adaptive_avg_pool3d_torch import BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, D_OUT, H_OUT, W_OUT +adaptive_pool3d_source = """ +#include +#include +// --------------------------------------------------------- +// Helper: Generic Index Calculation +// --------------------------------------------------------- +__device__ __forceinline__ int start_index(int out_idx, int out_len, int in_len) { + return (out_idx * in_len) / out_len; +} -BLOCK_SIZE = 256 -VEC_SIZE = 4 - -class ModelNew(nn.Module): - - def __init__(self, output_size): - super().__init__() - self.output_size = output_size - self.d_in = D_IN - self.h_in = H_IN - self.w_in = W_IN - self.d_out = output_size[0] - self.h_out = output_size[1] - self.w_out = output_size[2] - self.block_size = BLOCK_SIZE - self._compile_cuda_kernel() - - def _compile_cuda_kernel(self): - - cpp_header = """ - #include - - torch::Tensor adaptive_avgpool3d_forward_cuda( - torch::Tensor input, int D_in, int H_in, int W_in, int D_out, int H_out, int W_out - ); - """ - - cuda_source = f""" - #include - #include - #include - - #define BLOCK_SIZE {BLOCK_SIZE} - #define VEC_SIZE {VEC_SIZE} - - __global__ void adaptive_avgpool3d_kernel( - const float* __restrict__ input_data, - float* __restrict__ output_data, - int N, int C, int D_in, int H_in, int W_in, - int D_out, int H_out, int W_out - ) {{ - const int N_C_D_H_W_out = N * C * D_out * H_out * W_out; - const int tid = blockIdx.x * blockDim.x + threadIdx.x; - const int grid_stride = gridDim.x * blockDim.x; - - const int CHW_in = C * D_in * H_in * W_in; - const int DHW_in = D_in * H_in * W_in; - const int HW_in = H_in * W_in; - - for (int idx = tid; idx < N_C_D_H_W_out; idx += grid_stride) {{ - - const int w_out = idx % W_out; - const int h_w_out = idx / W_out; - - const int h_out = h_w_out % H_out; - const int d_h_w_out = h_w_out / H_out; - - const int d_out = d_h_w_out % D_out; - const int n_c = d_h_w_out / D_out; - - const int n_idx = n_c / C; - const int c_idx = n_c % C; - - - const int d_in_start = (d_out * D_in) / D_out; - const int d_in_end = ((d_out + 1) * D_in) / D_out; - const int d_range = d_in_end - d_in_start; - - const int h_in_start = (h_out * H_in) / H_out; - const int h_in_end = ((h_out + 1) * H_in) / H_out; - const int h_range = h_in_end - h_in_start; - - const int w_in_start = (w_out * W_in) / W_out; - const int w_in_end = ((w_out + 1) * W_in) / W_out; - const int w_range = w_in_end - w_in_start; - - - double thread_sum = 0.0; - const int kernel_size = d_range * h_range * w_range; - - if (kernel_size == 0) {{ - output_data[idx] = 0.0f; - continue; - }} - - const int base_offset = (n_idx * CHW_in) + (c_idx * DHW_in); - - - for (int d = d_in_start; d < d_in_end; d++) {{ - for (int h = h_in_start; h < h_in_end; h++) {{ - - - const int w_start_idx = base_offset + (d * HW_in) + (h * W_in) + w_in_start; - - int w_current = 0; - int w_len = w_in_end - w_in_start; - - - while (w_current < w_len && ((w_in_start + w_current) % VEC_SIZE) != 0) {{ - thread_sum += (double)input_data[w_start_idx + w_current]; - w_current++; - }} - - - const int w_vector_len = w_len - w_current; - const int num_vectors = w_vector_len / VEC_SIZE; - - if (num_vectors > 0) {{ - const float4* vec_ptr = (const float4*)(input_data + w_start_idx + w_current); - - for (int v = 0; v < num_vectors; v++) {{ - float4 val4 = vec_ptr[v]; - thread_sum += (double)val4.x + (double)val4.y + (double)val4.z + (double)val4.w; - }} - w_current += num_vectors * VEC_SIZE; - }} - - - while (w_current < w_len) {{ - thread_sum += (double)input_data[w_start_idx + w_current]; - w_current++; - }} - - - }} - }} - - - output_data[idx] = (float)(thread_sum / (double)kernel_size); - }} - }} - - - torch::Tensor adaptive_avgpool3d_forward_cuda( - torch::Tensor input, int D_in, int H_in, int W_in, int D_out, int H_out, int W_out - ) {{ - TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor"); - TORCH_CHECK(input.is_contiguous(), "input must be contiguous"); - TORCH_CHECK(input.dim() == 5, "input must be 5D (N, C, D_in, H_in, W_in)"); - - const int N = input.size(0); - const int C = input.size(1); - - const int N_elements_out = N * C * D_out * H_out * W_out; - - auto output = torch::empty({{N, C, D_out, H_out, W_out}}, input.options()); - - dim3 block_dim(BLOCK_SIZE); - const int grid_size = (N_elements_out + BLOCK_SIZE - 1) / BLOCK_SIZE; - dim3 grid_dim(grid_size); - - adaptive_avgpool3d_kernel<<>>( - input.data_ptr(), - output.data_ptr(), - N, C, D_in, H_in, W_in, - D_out, H_out, W_out - ); - - return output; - }} - """ +__device__ __forceinline__ int end_index(int out_idx, int out_len, int in_len) { + long long tmp = (long long)(out_idx + 1) * in_len; + return (tmp + out_len - 1) / out_len; +} +// --------------------------------------------------------- +// Kernel 1: Fast Path (Integer Scaling) +// Assumes in_len % out_len == 0 for all dims +// --------------------------------------------------------- +__global__ void adaptive_avg_pool3d_fast_kernel( + const float* __restrict__ input, + float* __restrict__ output, + int D_in, int H_in, int W_in, + int D_out, int H_out, int W_out, + int stride_d, int stride_h, int stride_w, + float inv_vol, + long in_stride_nc, // D_in * H_in * W_in + long out_stride_nc // D_out * H_out * W_out +) { + // Grid Mapping: + // Z: Batch * Channel + // Y: Output Depth (D_out) + // X: Output Spatial (H_out * W_out) - self.pad_op = load_inline( - name="adaptive_avgpool3d_op", - cpp_sources=cpp_header, - cuda_sources=cuda_source, - functions=["adaptive_avgpool3d_forward_cuda"], - verbose=False - ) + int nc = blockIdx.z; + int d_out = blockIdx.y; + int spatial_idx = blockIdx.x * blockDim.x + threadIdx.x; + + if (d_out >= D_out || spatial_idx >= H_out * W_out) return; + + int h_out = spatial_idx / W_out; + int w_out = spatial_idx % W_out; + + // Base Pointers + const float* vol_in = input + (long)nc * in_stride_nc; + float* vol_out = output + (long)nc * out_stride_nc; + + // Fixed Window (No div/mod per loop) + int d_start = d_out * stride_d; + int h_start = h_out * stride_h; + int w_start = w_out * stride_w; + + float sum = 0.0f; + + // 3D Loop + #pragma unroll + for (int kz = 0; kz < stride_d; ++kz) { + int d_in = d_start + kz; + long d_offset = (long)d_in * H_in * W_in; + + #pragma unroll + for (int ky = 0; ky < stride_h; ++ky) { + int h_in = h_start + ky; + long h_offset = (long)h_in * W_in; + + #pragma unroll + for (int kx = 0; kx < stride_w; ++kx) { + int w_in = w_start + kx; + + // Use __ldg for read-only cache + sum += __ldg(&vol_in[d_offset + h_offset + w_in]); + } + } + } + + long out_idx = (long)d_out * (H_out * W_out) + spatial_idx; + vol_out[out_idx] = sum * inv_vol; +} - def forward(self, x: torch.Tensor) -> torch.Tensor: - d_out, h_out, w_out = self.output_size - return self.pad_op.adaptive_avgpool3d_forward_cuda( - x.contiguous(), - self.d_in, - self.h_in, - self.w_in, - d_out, - h_out, - w_out - ) \ No newline at end of file +// --------------------------------------------------------- +// Kernel 2: Generic Path +// --------------------------------------------------------- +__global__ void adaptive_avg_pool3d_generic_kernel( + const float* __restrict__ input, + float* __restrict__ output, + int D_in, int H_in, int W_in, + int D_out, int H_out, int W_out, + long in_stride_nc, + long out_stride_nc +) { + int nc = blockIdx.z; + int d_out = blockIdx.y; + int spatial_idx = blockIdx.x * blockDim.x + threadIdx.x; + + if (d_out >= D_out || spatial_idx >= H_out * W_out) return; + + int h_out = spatial_idx / W_out; + int w_out = spatial_idx % W_out; + + const float* vol_in = input + (long)nc * in_stride_nc; + float* vol_out = output + (long)nc * out_stride_nc; + + // Calculate Windows + int d_start = start_index(d_out, D_out, D_in); + int d_end = end_index(d_out, D_out, D_in); + int d_len = d_end - d_start; + + int h_start = start_index(h_out, H_out, H_in); + int h_end = end_index(h_out, H_out, H_in); + int h_len = h_end - h_start; + + int w_start = start_index(w_out, W_out, W_in); + int w_end = end_index(w_out, W_out, W_in); + int w_len = w_end - w_start; + + float sum = 0.0f; + + for (int d = d_start; d < d_end; ++d) { + long d_offset = (long)d * H_in * W_in; + for (int h = h_start; h < h_end; ++h) { + long h_offset = (long)h * W_in; + for (int w = w_start; w < w_end; ++w) { + sum += __ldg(&vol_in[d_offset + h_offset + w]); + } + } + } + + int vol_len = d_len * h_len * w_len; + long out_idx = (long)d_out * (H_out * W_out) + spatial_idx; + vol_out[out_idx] = (vol_len > 0) ? (sum / vol_len) : 0.0f; +} + +torch::Tensor adaptive_avg_pool3d_cuda(torch::Tensor input, torch::Tensor output_size) { + int N = input.size(0); + int C = input.size(1); + int D_in = input.size(2); + int H_in = input.size(3); + int W_in = input.size(4); + + auto size_cpu = output_size.cpu(); + int* dims = size_cpu.data_ptr(); + int D_out = dims[0]; + int H_out = dims[1]; + int W_out = dims[2]; + + auto output = torch::empty({N, C, D_out, H_out, W_out}, input.options()); + + long in_stride_nc = (long)D_in * H_in * W_in; + long out_stride_nc = (long)D_out * H_out * W_out; + int nc = N * C; + + // Check for Integer Scaling (Fast Path) + bool is_fast = (D_in % D_out == 0) && (H_in % H_out == 0) && (W_in % W_out == 0); + + long total_spatial = H_out * W_out; + const int block = 256; + dim3 grid((total_spatial + block - 1) / block, D_out, nc); + + if (is_fast) { + int stride_d = D_in / D_out; + int stride_h = H_in / H_out; + int stride_w = W_in / W_out; + float inv_vol = 1.0f / (float)(stride_d * stride_h * stride_w); + + adaptive_avg_pool3d_fast_kernel<<>>( + input.data_ptr(), + output.data_ptr(), + D_in, H_in, W_in, + D_out, H_out, W_out, + stride_d, stride_h, stride_w, + inv_vol, + in_stride_nc, out_stride_nc + ); + } else { + adaptive_avg_pool3d_generic_kernel<<>>( + input.data_ptr(), + output.data_ptr(), + D_in, H_in, W_in, + D_out, H_out, W_out, + in_stride_nc, out_stride_nc + ); + } + + return output; +} +""" + +cpp_source = "torch::Tensor adaptive_avg_pool3d_cuda(torch::Tensor input, torch::Tensor output_size);" + +adaptive_module = load_inline( + name="adaptive_avg_pool3d_extension", + cpp_sources=cpp_source, + cuda_sources=adaptive_pool3d_source, + functions=["adaptive_avg_pool3d_cuda"], + verbose=True, + with_cuda=True +) + +class ModelNew(torch.nn.Module): + def __init__(self): + super(ModelNew, self).__init__() + # 使用 Tensor 传递 size，保持接口灵活 + self.output_size = torch.tensor([32, 32, 32], dtype=torch.int32) + self.cuda_op = adaptive_module + + def forward(self, x): + return self.cuda_op.adaptive_avg_pool3d_cuda(x.contiguous(), self.output_size) \ No newline at end of file diff --git a/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py b/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py index 69784a44..e4f8d86d 100644 --- a/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py +++ b/S1/ZZZJ_#4/adaptive_avg_pool3d_torch.py @@ -1,31 +1,34 @@ -# adaptive_pool3d_torch.py import torch import torch.nn as nn -import torch.nn.functional as F -BATCH_SIZE = 4 -CHANNELS = 64 -D_IN, H_IN, W_IN = 32, 32, 32 -D_OUT, H_OUT, W_OUT = 4, 4, 4 - +torch.backends.cuda.matmul.allow_tf32 = False class Model(nn.Module): - - - def __init__(self, output_size): - super().__init__() - self.adaptive_pool = nn.AdaptiveAvgPool3d(output_size) + def __init__(self): + super(Model, self).__init__() + + self.output_size = (32, 32, 32) + self.pool = nn.AdaptiveAvgPool3d(self.output_size) def forward(self, x: torch.Tensor) -> torch.Tensor: - return self.adaptive_pool(x) + """ + x: [N, C, D_in, H_in, W_in] + Output: [N, C, D_out, H_out, W_out] + """ + return self.pool(x) +N = 8 +C = 32 +D_in = 64 +H_in = 64 +W_in = 64 + def get_inputs(): - - x = torch.randn(BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, dtype=torch.float32) + + x = torch.randint(0, 16, (N, C, D_in, H_in, W_in), device='cuda').float() return [x] - def get_init_inputs(): - return [(D_OUT, H_OUT, W_OUT)] \ No newline at end of file + return [] \ No newline at end of file diff --git a/S1/ZZZJ_#4/prompt.txt b/S1/ZZZJ_#4/prompt.txt index ebc80150..ef956b07 100644 --- a/S1/ZZZJ_#4/prompt.txt +++ b/S1/ZZZJ_#4/prompt.txt @@ -8,32 +8,35 @@ python # adaptive_pool3d_torch.py import torch import torch.nn as nn -import torch.nn.functional as F -BATCH_SIZE = 4 -CHANNELS = 64 -D_IN, H_IN, W_IN = 32, 32, 32 -D_OUT, H_OUT, W_OUT = 4, 4, 4 - +torch.backends.cuda.matmul.allow_tf32 = False class Model(nn.Module): - - - def __init__(self, output_size): - super().__init__() - self.adaptive_pool = nn.AdaptiveAvgPool3d(output_size) + def __init__(self): + super(Model, self).__init__() + + self.output_size = (32, 32, 32) + self.pool = nn.AdaptiveAvgPool3d(self.output_size) def forward(self, x: torch.Tensor) -> torch.Tensor: - return self.adaptive_pool(x) + """ + x: [N, C, D_in, H_in, W_in] + Output: [N, C, D_out, H_out, W_out] + """ + return self.pool(x) +N = 8 +C = 32 +D_in = 64 +H_in = 64 +W_in = 64 + def get_inputs(): - - x = torch.randn(BATCH_SIZE, CHANNELS, D_IN, H_IN, W_IN, dtype=torch.float32) + + x = torch.randint(0, 16, (N, C, D_in, H_in, W_in), device='cuda').float() return [x] - def get_init_inputs(): - return [(D_OUT, H_OUT, W_OUT)] -``` \ No newline at end of file + return []``` \ No newline at end of file