|
| 1 | +// Copyright © 2025 Apple Inc. |
| 2 | + |
| 3 | +#include "mlx/backend/webgpu/allocator.h" |
| 4 | + |
| 5 | +namespace mlx::core { |
| 6 | + |
| 7 | +namespace allocator { |
| 8 | + |
| 9 | +Allocator& allocator() { |
| 10 | + return webgpu::allocator(); |
| 11 | +} |
| 12 | + |
| 13 | +void* Buffer::raw_ptr() { |
| 14 | + return static_cast<webgpu::DoubleBuffer*>(ptr_)->cpu_data(); |
| 15 | +} |
| 16 | + |
| 17 | +} // namespace allocator |
| 18 | + |
| 19 | +namespace webgpu { |
| 20 | + |
| 21 | +DoubleBuffer::DoubleBuffer(size_t size) |
| 22 | + : cpu_data_(std::malloc(size + sizeof(size_t))) { |
| 23 | + *static_cast<size_t*>(cpu_data_) = size; |
| 24 | +} |
| 25 | + |
| 26 | +DoubleBuffer::DoubleBuffer(betann::Device& device, size_t size) |
| 27 | + : gpu_data_(device.CreateBuffer( |
| 28 | + size, |
| 29 | + betann::BufferUsage::Storage | betann::BufferUsage::CopySrc)) {} |
| 30 | + |
| 31 | +DoubleBuffer::~DoubleBuffer() { |
| 32 | + std::free(cpu_data_); |
| 33 | +} |
| 34 | + |
| 35 | +void DoubleBuffer::copy_to_cpu(const void* data, size_t size) { |
| 36 | + assert(!cpu_data_); |
| 37 | + cpu_data_ = std::malloc(size + sizeof(size_t)); |
| 38 | + *static_cast<size_t*>(cpu_data_) = size; |
| 39 | + std::memcpy(cpu_data(), data, size); |
| 40 | +} |
| 41 | + |
| 42 | +size_t DoubleBuffer::size() const { |
| 43 | + if (cpu_data_) |
| 44 | + return *static_cast<size_t*>(cpu_data_); |
| 45 | + if (gpu_data_) |
| 46 | + return gpu_data_.GetSize(); |
| 47 | + return 0; |
| 48 | +} |
| 49 | + |
| 50 | +WgpuAllocator::WgpuAllocator() : device_(webgpu::device(Device::gpu)) {} |
| 51 | + |
| 52 | +Buffer |
| 53 | +WgpuAllocator::malloc(const Device& device, size_t size, bool allow_swap) { |
| 54 | + if (device.type == Device::gpu) |
| 55 | + return Buffer(new DoubleBuffer(webgpu::device(device), size)); |
| 56 | + else |
| 57 | + return Buffer(new DoubleBuffer(size)); |
| 58 | +} |
| 59 | + |
| 60 | +void WgpuAllocator::free(Buffer buffer) { |
| 61 | + delete static_cast<DoubleBuffer*>(buffer.ptr()); |
| 62 | +} |
| 63 | + |
| 64 | +size_t WgpuAllocator::size(Buffer buffer) const { |
| 65 | + return static_cast<DoubleBuffer*>(buffer.ptr())->size(); |
| 66 | +} |
| 67 | + |
| 68 | +void WgpuAllocator::ensure_gpu_data(Buffer& buffer) { |
| 69 | + auto* dbuf = static_cast<DoubleBuffer*>(buffer.ptr()); |
| 70 | + if (dbuf->gpu_data() || dbuf->size() == 0) |
| 71 | + return; |
| 72 | + dbuf->set_gpu_data(device_.CreateBufferFromData( |
| 73 | + dbuf->cpu_data(), dbuf->size(), betann::BufferUsage::Storage)); |
| 74 | +} |
| 75 | + |
| 76 | +WgpuAllocator& allocator() { |
| 77 | + static WgpuAllocator allocator_; |
| 78 | + return allocator_; |
| 79 | +} |
| 80 | + |
| 81 | +betann::Device& device(mlx::core::Device) { |
| 82 | + static betann::Device device; |
| 83 | + return device; |
| 84 | +} |
| 85 | + |
| 86 | +} // namespace webgpu |
| 87 | + |
| 88 | +namespace metal { |
| 89 | + |
| 90 | +size_t get_active_memory() { |
| 91 | + return 0; |
| 92 | +} |
| 93 | +size_t get_peak_memory() { |
| 94 | + return 0; |
| 95 | +} |
| 96 | +void reset_peak_memory() {} |
| 97 | +size_t get_cache_memory() { |
| 98 | + return 0; |
| 99 | +} |
| 100 | +size_t set_memory_limit(size_t, bool) { |
| 101 | + return 0; |
| 102 | +} |
| 103 | +size_t set_cache_limit(size_t) { |
| 104 | + return 0; |
| 105 | +} |
| 106 | +size_t set_wired_limit(size_t) { |
| 107 | + return 0; |
| 108 | +} |
| 109 | + |
| 110 | +std::unordered_map<std::string, std::variant<std::string, size_t>> |
| 111 | +device_info() { |
| 112 | + throw std::runtime_error("[webgpu::device_info] Not implemented"); |
| 113 | +}; |
| 114 | + |
| 115 | +void clear_cache() {} |
| 116 | + |
| 117 | +} // namespace metal |
| 118 | + |
| 119 | +} // namespace mlx::core |
0 commit comments