Add build subdirectory

benchmarks/benchmark_utils.hpp

@@ -2,14 +2,13 @@
 
 #include <chrono>
 #include <cmath>
-#include <stdexcept>
 #include <string>
 #include <vector>
 
 #include <benchmark/benchmark.h>
 
-void init_x(float *x, size_t n) {
-  for (size_t i = 0; i < n; ++i) {
+void init_x(std::vector<float> &x) {
+  for (size_t i = 0; i < x.size(); ++i) {
     x[i] = (i % 360) * 0.0174533f; // degrees to radians
   }
 }
@@ -17,31 +16,24 @@ void init_x(float *x, size_t n) {
 template <typename Backend>
 static void benchmark_sinf(benchmark::State &state) {
   const size_t N = static_cast<size_t>(state.range(0));
+  std::vector<float> x(N), s(N);
+  init_x(x);
 
   Backend backend;
 
   auto start = std::chrono::high_resolution_clock::now();
   backend.init(N);
-  float *x =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
-  float *s =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
   auto end = std::chrono::high_resolution_clock::now();
   state.counters["init_ms"] =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start)
           .count() /
       1.e3;
 
-  init_x(x, N);
-
   for (auto _ : state) {
-    backend.compute_sinf(N, x, s);
+    backend.compute_sinf(N, x.data(), s.data());
     benchmark::DoNotOptimize(s);
   }
 
-  backend.free_memory(x);
-  backend.free_memory(s);
-
   state.SetItemsProcessed(static_cast<int64_t>(state.iterations()) *
                           static_cast<int64_t>(N));
 }
@@ -49,35 +41,24 @@ static void benchmark_sinf(benchmark::State &state) {
 template <typename Backend>
 static void benchmark_cosf(benchmark::State &state) {
   const size_t N = static_cast<size_t>(state.range(0));
+  std::vector<float> x(N), c(N);
+  init_x(x);
 
   Backend backend;
 
   auto start = std::chrono::high_resolution_clock::now();
   backend.init(N);
-  float *x =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
-  float *c =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
-
-  if (!x || !c) {
-    throw std::runtime_error("Buffer allocation failed");
-  }
   auto end = std::chrono::high_resolution_clock::now();
   state.counters["init_ms"] =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start)
           .count() /
       1.e3;
 
-  init_x(x, N);
-
   for (auto _ : state) {
-    backend.compute_cosf(N, x, c);
+    backend.compute_cosf(N, x.data(), c.data());
     benchmark::DoNotOptimize(c);
   }
 
-  backend.free_memory(x);
-  backend.free_memory(c);
-
   state.SetItemsProcessed(static_cast<int64_t>(state.iterations()) *
                           static_cast<int64_t>(N));
 }
@@ -85,38 +66,25 @@ static void benchmark_cosf(benchmark::State &state) {
 template <typename Backend>
 static void benchmark_sincosf(benchmark::State &state) {
   const size_t N = static_cast<size_t>(state.range(0));
+  std::vector<float> x(N), s(N), c(N);
+  init_x(x);
 
   Backend backend;
 
   auto start = std::chrono::high_resolution_clock::now();
   backend.init(N);
-  float *x =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
-  float *s =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
-  float *c =
-      reinterpret_cast<float *>(backend.allocate_memory(N * sizeof(float)));
-  if (!x || !s || !c) {
-    throw std::runtime_error("Buffer allocation failed");
-  }
   auto end = std::chrono::high_resolution_clock::now();
   state.counters["init_ms"] =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start)
           .count() /
       1.e3;
 
-  init_x(x, N);
-
   for (auto _ : state) {
-    backend.compute_sincosf(N, x, s, c);
+    backend.compute_sincosf(N, x.data(), s.data(), c.data());
     benchmark::DoNotOptimize(s);
     benchmark::DoNotOptimize(c);
   }
 
-  backend.free_memory(x);
-  backend.free_memory(s);
-  backend.free_memory(c);
-
   state.SetItemsProcessed(static_cast<int64_t>(state.iterations()) *
                           static_cast<int64_t>(N));
 }

include/trigdx/gpu.hpp

@@ -11,8 +11,7 @@ public:
   GPUBackend();
   ~GPUBackend() override;
 
-  void *allocate_memory(size_t bytes) const override;
-  void free_memory(void *ptr) const override;
+  void init(size_t n = 0) override;
   void compute_sinf(size_t n, const float *x, float *s) const override;
   void compute_cosf(size_t n, const float *x, float *c) const override;
   void compute_sincosf(size_t n, const float *x, float *s,

include/trigdx/interface.hpp

@@ -1,8 +1,6 @@
 #pragma once
 
 #include <cstddef>
-#include <cstdint>
-#include <cstdlib>
 
 // Base interface for all math backends
 class Backend {
@@ -12,12 +10,6 @@ public:
   // Optional initialization
   virtual void init(size_t n = 0) {}
 
-  virtual void *allocate_memory(size_t bytes) const {
-    return static_cast<void *>(new uint8_t[bytes]);
-  };
-
-  virtual void free_memory(void *ptr) const { std::free(ptr); };
-
   // Compute sine for n elements
   virtual void compute_sinf(size_t n, const float *x, float *s) const = 0;
 

python/CMakeLists.txt

@@ -1,6 +1,4 @@
-find_package(pybind11 CONFIG QUIET)
-
-if(NOT pybind11_FOUND)
+if(NOT TARGET pybind11)
   FetchContent_Declare(
     pybind11
     GIT_REPOSITORY https://github.com/pybind/pybind11.git

src/gpu.cpp

@@ -10,63 +10,79 @@
 
 struct GPUBackend::Impl {
 
-  void *allocate_memory(size_t bytes) const {
-    void *ptr;
-    cudaMallocHost(&ptr, bytes);
-    return ptr;
+  ~Impl() {
+    if (h_x) {
+      cudaFreeHost(h_x);
+    }
+    if (h_s) {
+      cudaFreeHost(h_s);
+    }
+    if (h_c) {
+      cudaFreeHost(h_c);
+    }
+    if (d_x) {
+      cudaFree(d_x);
+    }
+    if (d_s) {
+      cudaFree(d_s);
+    }
+    if (d_c) {
+      cudaFree(d_c);
+    }
   }
 
-  void free_memory(void *ptr) const { cudaFreeHost(ptr); }
+  void init(size_t n) {
+    const size_t bytes = n * sizeof(float);
+    cudaMallocHost(&h_x, bytes);
+    cudaMallocHost(&h_s, bytes);
+    cudaMallocHost(&h_c, bytes);
+    cudaMalloc(&d_x, bytes);
+    cudaMalloc(&d_s, bytes);
+    cudaMalloc(&d_c, bytes);
+  }
 
   void compute_sinf(size_t n, const float *x, float *s) const {
     const size_t bytes = n * sizeof(float);
-    float *d_x, *d_s;
-    cudaMalloc(&d_x, bytes);
-    cudaMalloc(&d_s, bytes);
-    cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
+    std::memcpy(h_x, x, bytes);
+    cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
     launch_sinf_kernel(d_x, d_s, n);
-    cudaMemcpy(s, d_s, bytes, cudaMemcpyDeviceToHost);
-    cudaFree(d_x);
-    cudaFree(d_s);
+    cudaMemcpy(h_s, d_s, bytes, cudaMemcpyDeviceToHost);
+    std::memcpy(s, h_s, bytes);
   }
 
   void compute_cosf(size_t n, const float *x, float *c) const {
     const size_t bytes = n * sizeof(float);
-    float *d_x, *d_c;
-    cudaMalloc(&d_x, bytes);
-    cudaMalloc(&d_c, bytes);
-    cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
+    std::memcpy(h_x, x, bytes);
+    cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
     launch_cosf_kernel(d_x, d_c, n);
-    cudaMemcpy(c, d_c, bytes, cudaMemcpyDeviceToHost);
-    cudaFree(d_x);
-    cudaFree(d_c);
+    cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
+    std::memcpy(c, h_c, bytes);
   }
 
   void compute_sincosf(size_t n, const float *x, float *s, float *c) const {
     const size_t bytes = n * sizeof(float);
-    float *d_x, *d_s, *d_c;
-    cudaMalloc(&d_x, bytes);
-    cudaMalloc(&d_s, bytes);
-    cudaMalloc(&d_c, bytes);
-    cudaMemcpy(d_x, x, bytes, cudaMemcpyHostToDevice);
+    std::memcpy(h_x, x, bytes);
+    cudaMemcpy(d_x, h_x, bytes, cudaMemcpyHostToDevice);
     launch_sincosf_kernel(d_x, d_s, d_c, n);
-    cudaMemcpy(s, d_s, bytes, cudaMemcpyDeviceToHost);
-    cudaMemcpy(c, d_c, bytes, cudaMemcpyDeviceToHost);
-    cudaFree(d_x);
-    cudaFree(d_s);
-    cudaFree(d_c);
+    cudaMemcpy(h_s, d_s, bytes, cudaMemcpyDeviceToHost);
+    cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
+    std::memcpy(s, h_s, bytes);
+    std::memcpy(c, h_c, bytes);
   }
+
+  float *h_x = nullptr;
+  float *h_s = nullptr;
+  float *h_c = nullptr;
+  float *d_x = nullptr;
+  float *d_s = nullptr;
+  float *d_c = nullptr;
 };
 
 GPUBackend::GPUBackend() : impl(std::make_unique<Impl>()) {}
 
 GPUBackend::~GPUBackend() = default;
 
-void *GPUBackend::allocate_memory(size_t bytes) const {
-  return impl->allocate_memory(bytes);
-}
-
-void GPUBackend::free_memory(void *ptr) const { impl->free_memory(ptr); }
+void GPUBackend::init(size_t n) { impl->init(n); }
 
 void GPUBackend::compute_sinf(size_t n, const float *x, float *s) const {
   impl->compute_sinf(n, x, s);