GPUCodeForces/S1/uucoco_#21/SoftClip_cuda.py

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline


class ModelNew(nn.Module):
    def __init__(self, min_val=-1.0, max_val=1.0, beta=1.0):
        super().__init__()
        self.min_val = min_val
        self.max_val = max_val
        self.beta = beta
        self._compile_cuda_kernel()

    def _compile_cuda_kernel(self):
        cpp_source = """
        #include <torch/extension.h>
        torch::Tensor softclip_cuda(torch::Tensor x, float min_val, float max_val, float beta);
        """

        cuda_source = """
        #include <cuda_runtime.h>

        __device__ __forceinline__ float softplus_f(float x, float beta, float inv_beta) {
            float bx = x * beta;
            if (bx > 20.0f) return x;
            return log1pf(expf(bx)) * inv_beta;
        }

        __global__ void softclip_kernel_ilp(
            const float* __restrict__ x,
            float* __restrict__ y,
            int n_vec,
            float min_val,
            float max_val,
            float beta,
            float inv_beta) 
        {
            int idx = blockIdx.x * blockDim.x + threadIdx.x;
            int stride = gridDim.x * blockDim.x;

            const float4* x_vec = reinterpret_cast<const float4*>(x);
            float4* y_vec = reinterpret_cast<float4*>(y);

            int i = idx;
            // ILP Loop Unrolling (2x float4)
            for (; i < n_vec - 1; i += stride) {
                float4 v1 = x_vec[i];
                float4 v2 = x_vec[i+1];
                float4 o1, o2;

                o1.x = v1.x + softplus_f(min_val - v1.x, beta, inv_beta) - softplus_f(v1.x - max_val, beta, inv_beta);
                o1.y = v1.y + softplus_f(min_val - v1.y, beta, inv_beta) - softplus_f(v1.y - max_val, beta, inv_beta);
                o1.z = v1.z + softplus_f(min_val - v1.z, beta, inv_beta) - softplus_f(v1.z - max_val, beta, inv_beta);
                o1.w = v1.w + softplus_f(min_val - v1.w, beta, inv_beta) - softplus_f(v1.w - max_val, beta, inv_beta);

                o2.x = v2.x + softplus_f(min_val - v2.x, beta, inv_beta) - softplus_f(v2.x - max_val, beta, inv_beta);
                o2.y = v2.y + softplus_f(min_val - v2.y, beta, inv_beta) - softplus_f(v2.y - max_val, beta, inv_beta);
                o2.z = v2.z + softplus_f(min_val - v2.z, beta, inv_beta) - softplus_f(v2.z - max_val, beta, inv_beta);
                o2.w = v2.w + softplus_f(min_val - v2.w, beta, inv_beta) - softplus_f(v2.w - max_val, beta, inv_beta);

                y_vec[i] = o1;
                y_vec[i+1] = o2;

                i++;
            }

            for (; i < n_vec; i += stride) {
                float4 v = x_vec[i];
                float4 o;
                o.x = v.x + softplus_f(min_val - v.x, beta, inv_beta) - softplus_f(v.x - max_val, beta, inv_beta);
                o.y = v.y + softplus_f(min_val - v.y, beta, inv_beta) - softplus_f(v.y - max_val, beta, inv_beta);
                o.z = v.z + softplus_f(min_val - v.z, beta, inv_beta) - softplus_f(v.z - max_val, beta, inv_beta);
                o.w = v.w + softplus_f(min_val - v.w, beta, inv_beta) - softplus_f(v.w - max_val, beta, inv_beta);
                y_vec[i] = o;
            }
        }

        __global__ void softclip_kernel_scalar(
            const float* __restrict__ x,
            float* __restrict__ y,
            int n,
            float min_val,
            float max_val,
            float beta,
            float inv_beta)
        {
            int idx = blockIdx.x * blockDim.x + threadIdx.x;
            int stride = gridDim.x * blockDim.x;

            for(int i = idx; i < n; i += stride) {
                float v = x[i];
                y[i] = v + softplus_f(min_val - v, beta, inv_beta) - softplus_f(v - max_val, beta, inv_beta);
            }
        }

        torch::Tensor softclip_cuda(torch::Tensor x, float min_val, float max_val, float beta) {
            auto x_c = x.contiguous();
            auto output = torch::empty_like(x_c);

            int n = x_c.numel();
            int threads = 256;
            float inv_beta = 1.0f / beta;

            if (n % 4 == 0) {
                int vec_n = n / 4;
                int blocks = (vec_n + threads - 1) / threads;
                if (blocks > 65535) blocks = 65535;
                if (blocks == 0) blocks = 1;

                softclip_kernel_ilp<<<blocks, threads>>>(
                    x_c.data_ptr<float>(),
                    output.data_ptr<float>(),
                    vec_n, min_val, max_val, beta, inv_beta
                );
            } else {
                int blocks = (n + threads - 1) / threads;
                if (blocks > 65535) blocks = 65535;
                if (blocks == 0) blocks = 1;

                softclip_kernel_scalar<<<blocks, threads>>>(
                    x_c.data_ptr<float>(),
                    output.data_ptr<float>(),
                    n, min_val, max_val, beta, inv_beta
                );
            }

            return output;
        }
        """

        self.op = load_inline(
            name="softclip_opt_v3_fix",
            cpp_sources=cpp_source,
            cuda_sources=cuda_source,
            functions=["softclip_cuda"],
            extra_cuda_cflags=["-O3"],
            verbose=False
        )

    def forward(self, x):
        return self.op.softclip_cuda(x, self.min_val, self.max_val, self.beta)