optirs_gpu/memory/vendors/

rocm_backend.rs

// ROCm backend for GPU memory management
//
// This module provides AMD ROCm/HIP-specific memory management functionality,
// including device memory allocation, HIP streams, and performance optimization
// features specific to AMD GPUs.
//
// # This is a host-memory simulation, not real ROCm
//
// `optirs-gpu` is Pure Rust with no FFI dependencies by default, and this
// crate ships no ROCm/HIP runtime bindings. There is therefore no real
// `hipMalloc` underneath this module: "device", "host", "coarse-grained",
// "fine-grained" and "host-visible" memory are all the *same* system-heap
// allocation (see `sim_alloc`/`sim_dealloc` below), and `RocmDeviceProperties`
// /`RocmStats` are example numbers, not a query of real hardware. This
// module models the ROCm memory-management *API shape* for testing that
// shape in isolation; treat every allocation as host memory and every
// device number as illustrative.
//
// This extends to data movement: `memcpy` and `memcpy_async` copy **zero
// bytes**. They build a `HipOperation` record, hand it to
// `HipStreamManager::execute_operation` (which returns immediately and never
// dereferences `src_ptr`/`dst_ptr`), and increment
// `RocmStats::memory_transfers` — that counter says "this many `memcpy`
// calls were made," not "this many bytes moved." An earlier revision of
// this module also injected a `std::thread::sleep` here to imitate transfer
// latency by operation kind; that fake timing has been removed, so the
// distinction between `MemcpyHostToDevice`/`MemcpyDeviceToHost`/
// `MemcpyDeviceToDevice`/`MemcpyAsync` no longer affects anything
// observable. `RocmStats::stream_operations` and `RocmStats::kernel_launches`
// are declared for API-shape completeness but nothing in this module ever
// increments them — read a `0` there as "not tracked," not "none occurred."

use std::collections::HashMap;
use std::ffi::c_void;
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};

/// Byte alignment every simulated allocation below uses.
const SIM_ALLOC_ALIGN: usize = 256;

/// Allocate `size` bytes through the system allocator, 256-byte aligned,
/// without the two ways the naive `std::alloc::alloc(Layout::from_size_align_unchecked(size,
/// 256))` this module used to call was undefined behaviour: a zero-size
/// layout is unsound to pass to `GlobalAlloc::alloc`, and a real allocation
/// failure returns null, which must never be treated as valid memory. See
/// `cuda_backend::sim_alloc` for the full rationale (this mirrors it).
///
/// The payload is prefixed with one `SIM_ALLOC_ALIGN`-byte header recording
/// the requested size, so [`sim_dealloc`] can reconstruct the exact `Layout`
/// this function used — the `Layout::from_size_align_unchecked(1, 1)` this
/// module used at free time was a mismatched-layout deallocation, itself
/// unconditionally undefined behaviour.
fn sim_alloc(size: usize) -> Result<*mut c_void, RocmError> {
    if size == 0 {
        return Ok(SIM_ALLOC_ALIGN as *mut c_void);
    }
    let total = SIM_ALLOC_ALIGN.checked_add(size).ok_or_else(|| {
        RocmError::OutOfMemory(format!(
            "{size}-byte request overflows the allocator's size limit"
        ))
    })?;
    let layout = std::alloc::Layout::from_size_align(total, SIM_ALLOC_ALIGN)
        .map_err(|e| RocmError::OutOfMemory(format!("invalid allocation layout: {e}")))?;
    // SAFETY: `layout` has non-zero size (checked above) and a valid
    // (power-of-two) alignment constructed by `Layout::from_size_align`.
    let base = unsafe { std::alloc::alloc(layout) };
    if base.is_null() {
        return Err(RocmError::OutOfMemory(format!(
            "allocator returned null for a {size}-byte request"
        )));
    }
    // SAFETY: `base` is non-null and `layout`'s size is at least
    // `SIM_ALLOC_ALIGN + size >= SIM_ALLOC_ALIGN >= size_of::<usize>()`, so
    // writing one `usize` at the start of the block is in-bounds.
    unsafe { (base as *mut usize).write(size) };
    // SAFETY: `base` was allocated with `total = SIM_ALLOC_ALIGN + size`
    // bytes, so offsetting by `SIM_ALLOC_ALIGN` stays within (or one past)
    // the allocation.
    Ok(unsafe { base.add(SIM_ALLOC_ALIGN) } as *mut c_void)
}

/// Free a pointer returned by [`sim_alloc`]. A no-op for a null pointer or
/// the zero-size sentinel — neither was ever allocated.
///
/// Only ever called from this module (it is not `pub`) with a pointer
/// `sim_alloc` returned that has not already been freed — the unsafety of
/// the pointer arithmetic below is contained to that invariant, matching
/// this module's existing style of confining `unsafe` to the raw
/// `std::alloc` calls rather than marking `free()`'s public wrapper unsafe.
fn sim_dealloc(ptr: *mut c_void) {
    if ptr.is_null() || (ptr as usize) == SIM_ALLOC_ALIGN {
        return;
    }
    // SAFETY: by this function's contract `ptr` came from `sim_alloc`.
    let base = unsafe { (ptr as *mut u8).sub(SIM_ALLOC_ALIGN) };
    // SAFETY: `sim_alloc` wrote a `usize` at `base` before returning.
    let size = unsafe { (base as *const usize).read() };
    if let Ok(layout) = std::alloc::Layout::from_size_align(SIM_ALLOC_ALIGN + size, SIM_ALLOC_ALIGN)
    {
        // SAFETY: `layout` is exactly the layout `sim_alloc` allocated
        // `base` with.
        unsafe { std::alloc::dealloc(base, layout) };
    }
}

/// ROCm memory backend implementation
pub struct RocmMemoryBackend {
    /// Backend configuration
    config: RocmConfig,
    /// Device properties
    device_properties: RocmDeviceProperties,
    /// Active HIP contexts
    contexts: HashMap<u32, HipContext>,
    /// Memory pools
    memory_pools: HashMap<RocmMemoryType, RocmMemoryPool>,
    /// Statistics
    stats: RocmStats,
    /// Stream management
    stream_manager: HipStreamManager,
}

/// ROCm backend configuration
#[derive(Debug, Clone)]
pub struct RocmConfig {
    /// Device ID to use
    pub device_id: u32,
    /// Enable coarse-grained memory
    pub enable_coarse_memory: bool,
    /// Enable fine-grained memory
    pub enable_fine_memory: bool,
    /// Enable memory pools
    pub enable_memory_pools: bool,
    /// Enable async memory operations
    pub enable_async_ops: bool,
    /// Memory pool growth size
    pub pool_growth_size: usize,
    /// Enable host-visible device memory
    pub enable_host_visible: bool,
    /// Enable device coherent memory
    pub enable_device_coherent: bool,
    /// Maximum number of streams
    pub max_streams: u32,
    /// Enable GPU memory profiling
    pub enable_profiling: bool,
}

impl Default for RocmConfig {
    fn default() -> Self {
        Self {
            device_id: 0,
            enable_coarse_memory: true,
            enable_fine_memory: true,
            enable_memory_pools: true,
            enable_async_ops: true,
            pool_growth_size: 64 * 1024 * 1024, // 64MB
            enable_host_visible: true,
            enable_device_coherent: false,
            max_streams: 16,
            enable_profiling: false,
        }
    }
}

/// ROCm device properties
#[derive(Debug, Clone)]
pub struct RocmDeviceProperties {
    pub device_id: u32,
    pub name: String,
    pub arch: String,
    pub gcn_arch_name: String,
    pub total_global_memory: usize,
    pub local_memory_size: usize,
    pub max_work_group_size: u32,
    pub max_work_item_dimensions: u32,
    pub max_work_item_sizes: [u32; 3],
    pub compute_units: u32,
    pub wavefront_size: u32,
    pub memory_clock_frequency: u32,
    pub memory_bus_width: u32,
    pub l2_cache_size: usize,
    pub max_constant_buffer_size: usize,
    pub pci_bus_id: u32,
    pub pci_device_id: u32,
    pub supports_cooperative_launch: bool,
    pub supports_dynamic_parallelism: bool,
}

/// ROCm memory types
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub enum RocmMemoryType {
    Device,
    Host,
    HostVisible,
    DeviceCoherent,
    CoarseGrained,
    FineGrained,
}

/// HIP context for managing device state
pub struct HipContext {
    /// Context handle (simulated)
    pub handle: *mut c_void,
    /// Device ID
    pub device_id: u32,
    /// Context flags
    pub flags: HipContextFlags,
    /// Creation time
    pub created_at: Instant,
    /// Active streams
    pub streams: Vec<HipStream>,
    /// Device memory info
    pub memory_info: HipMemoryInfo,
}

/// HIP context creation flags
#[derive(Debug, Clone)]
pub struct HipContextFlags {
    pub sched_auto: bool,
    pub sched_spin: bool,
    pub sched_yield: bool,
    pub sched_blocking_sync: bool,
    pub map_host: bool,
}

impl Default for HipContextFlags {
    fn default() -> Self {
        Self {
            sched_auto: true,
            sched_spin: false,
            sched_yield: false,
            sched_blocking_sync: false,
            map_host: false,
        }
    }
}

/// HIP memory information
#[derive(Debug, Clone)]
pub struct HipMemoryInfo {
    pub total_memory: usize,
    pub free_memory: usize,
    pub used_memory: usize,
    pub coarse_memory: usize,
    pub fine_memory: usize,
}

/// HIP stream for asynchronous operations
pub struct HipStream {
    /// Stream handle (simulated)
    pub handle: *mut c_void,
    /// Stream ID
    pub id: u32,
    /// Stream priority
    pub priority: i32,
    /// Stream flags
    pub flags: HipStreamFlags,
    /// Creation time
    pub created_at: Instant,
    /// Operations queue
    pub operations: std::collections::VecDeque<HipOperation>,
}

/// HIP stream flags
#[derive(Debug, Clone)]
pub struct HipStreamFlags {
    pub default: bool,
    pub non_blocking: bool,
    pub per_thread: bool,
}

impl Default for HipStreamFlags {
    fn default() -> Self {
        Self {
            default: true,
            non_blocking: false,
            per_thread: false,
        }
    }
}

/// HIP asynchronous operation
#[derive(Debug, Clone)]
pub struct HipOperation {
    pub op_type: HipOperationType,
    pub src_ptr: Option<*mut c_void>,
    pub dst_ptr: Option<*mut c_void>,
    pub size: usize,
    pub timestamp: Instant,
}

/// Types of HIP operations
#[derive(Debug, Clone)]
pub enum HipOperationType {
    MemcpyHostToDevice,
    MemcpyDeviceToHost,
    MemcpyDeviceToDevice,
    MemcpyAsync,
    MemsetAsync,
    KernelLaunch,
    EventRecord,
    EventSynchronize,
}

/// ROCm memory pool
pub struct RocmMemoryPool {
    /// Memory type
    memory_type: RocmMemoryType,
    /// Current size
    current_size: usize,
    /// Maximum size
    max_size: usize,
    /// Used size
    used_size: usize,
    /// Free blocks
    free_blocks: std::collections::VecDeque<RocmMemoryBlock>,
    /// Allocated blocks
    allocated_blocks: HashMap<*mut c_void, RocmMemoryBlock>,
    /// Memory attributes
    attributes: RocmMemoryAttributes,
}

/// ROCm memory block
#[derive(Debug, Clone)]
pub struct RocmMemoryBlock {
    pub ptr: *mut c_void,
    pub size: usize,
    pub memory_type: RocmMemoryType,
    pub allocated_at: Instant,
    pub last_access: Option<Instant>,
    pub ref_count: u32,
    pub agent_accessible: bool,
}

/// ROCm memory attributes
#[derive(Debug, Clone)]
pub struct RocmMemoryAttributes {
    pub is_coarse_grained: bool,
    pub is_fine_grained: bool,
    pub is_host_accessible: bool,
    pub is_device_accessible: bool,
    pub is_coherent: bool,
    pub numa_node: Option<u32>,
}

impl Default for RocmMemoryAttributes {
    fn default() -> Self {
        Self {
            is_coarse_grained: true,
            is_fine_grained: false,
            is_host_accessible: false,
            is_device_accessible: true,
            is_coherent: false,
            numa_node: None,
        }
    }
}

impl RocmMemoryPool {
    pub fn new(memory_type: RocmMemoryType, max_size: usize) -> Self {
        let attributes = match memory_type {
            RocmMemoryType::CoarseGrained => RocmMemoryAttributes {
                is_coarse_grained: true,
                is_fine_grained: false,
                is_host_accessible: false,
                is_device_accessible: true,
                is_coherent: false,
                numa_node: None,
            },
            RocmMemoryType::FineGrained => RocmMemoryAttributes {
                is_coarse_grained: false,
                is_fine_grained: true,
                is_host_accessible: true,
                is_device_accessible: true,
                is_coherent: true,
                numa_node: Some(0),
            },
            RocmMemoryType::HostVisible => RocmMemoryAttributes {
                is_coarse_grained: false,
                is_fine_grained: false,
                is_host_accessible: true,
                is_device_accessible: true,
                is_coherent: false,
                numa_node: None,
            },
            _ => RocmMemoryAttributes::default(),
        };

        Self {
            memory_type,
            current_size: 0,
            max_size,
            used_size: 0,
            free_blocks: std::collections::VecDeque::new(),
            allocated_blocks: HashMap::new(),
            attributes,
        }
    }

    /// Allocate from pool
    pub fn allocate(&mut self, size: usize) -> Result<*mut c_void, RocmError> {
        // Try to find suitable free block
        for i in 0..self.free_blocks.len() {
            if self.free_blocks[i].size >= size {
                let Some(mut block) = self.free_blocks.remove(i) else {
                    continue;
                };

                // Split block if much larger
                if block.size > size * 2 {
                    let remaining_block = RocmMemoryBlock {
                        ptr: unsafe { block.ptr.add(size) },
                        size: block.size - size,
                        memory_type: block.memory_type.clone(),
                        allocated_at: block.allocated_at,
                        last_access: None,
                        ref_count: 0,
                        agent_accessible: block.agent_accessible,
                    };
                    self.free_blocks.push_back(remaining_block);
                    block.size = size;
                }

                block.last_access = Some(Instant::now());
                block.ref_count = 1;

                let ptr = block.ptr;
                self.allocated_blocks.insert(ptr, block);
                self.used_size += size;

                return Ok(ptr);
            }
        }

        // Need to allocate new memory
        if self.current_size + size > self.max_size {
            return Err(RocmError::OutOfMemory(
                "Pool size limit exceeded".to_string(),
            ));
        }

        let ptr = self.hip_malloc(size)?;
        let block = RocmMemoryBlock {
            ptr,
            size,
            memory_type: self.memory_type.clone(),
            allocated_at: Instant::now(),
            last_access: Some(Instant::now()),
            ref_count: 1,
            agent_accessible: self.attributes.is_device_accessible,
        };

        self.allocated_blocks.insert(ptr, block);
        self.current_size += size;
        self.used_size += size;

        Ok(ptr)
    }

    /// Free back to pool
    pub fn free(&mut self, ptr: *mut c_void) -> Result<(), RocmError> {
        if let Some(block) = self.allocated_blocks.remove(&ptr) {
            self.used_size -= block.size;

            // Add to free blocks
            self.free_blocks.push_back(RocmMemoryBlock {
                ptr: block.ptr,
                size: block.size,
                memory_type: block.memory_type,
                allocated_at: block.allocated_at,
                last_access: None,
                ref_count: 0,
                agent_accessible: block.agent_accessible,
            });

            // Try to coalesce adjacent blocks
            self.coalesce_free_blocks();

            Ok(())
        } else {
            Err(RocmError::InvalidPointer(
                "Pointer not found in pool".to_string(),
            ))
        }
    }

    fn coalesce_free_blocks(&mut self) {
        // Sort free blocks by address
        let mut blocks: Vec<RocmMemoryBlock> = self.free_blocks.drain(..).collect();
        blocks.sort_by_key(|block| block.ptr as usize);

        let mut coalesced = Vec::new();
        let mut current_block: Option<RocmMemoryBlock> = None;

        for block in blocks {
            match current_block.take() {
                None => current_block = Some(block),
                Some(mut prev_block) => {
                    let prev_end = prev_block.ptr as usize + prev_block.size;
                    let block_start = block.ptr as usize;

                    if prev_end == block_start && prev_block.memory_type == block.memory_type {
                        // Coalesce blocks
                        prev_block.size += block.size;
                        current_block = Some(prev_block);
                    } else {
                        coalesced.push(prev_block);
                        current_block = Some(block);
                    }
                }
            }
        }

        if let Some(block) = current_block {
            coalesced.push(block);
        }

        self.free_blocks = coalesced.into();
    }

    fn hip_malloc(&self, size: usize) -> Result<*mut c_void, RocmError> {
        // Simulate HIP memory allocation
        match self.memory_type {
            RocmMemoryType::Device => sim_alloc(size), // hipMalloc equivalent
            RocmMemoryType::Host => sim_alloc(size),   // hipMallocHost equivalent
            RocmMemoryType::CoarseGrained => sim_alloc(size), // coarse-grained device memory
            RocmMemoryType::FineGrained => sim_alloc(size), // fine-grained system memory
            RocmMemoryType::HostVisible => sim_alloc(size), // host-visible device memory
            _ => Err(RocmError::UnsupportedOperation(
                "Unsupported memory type for allocation".to_string(),
            )),
        }
    }
}

/// HIP stream manager
pub struct HipStreamManager {
    /// Available streams
    streams: Vec<HipStream>,
    /// Stream pool for reuse
    stream_pool: std::collections::VecDeque<HipStream>,
    /// Next stream ID
    next_stream_id: u32,
    /// Configuration
    config: HipStreamConfig,
}

/// Stream manager configuration
#[derive(Debug, Clone)]
pub struct HipStreamConfig {
    pub default_priority: i32,
    pub enable_priorities: bool,
    pub max_operations_per_stream: usize,
}

impl Default for HipStreamConfig {
    fn default() -> Self {
        Self {
            default_priority: 0,
            enable_priorities: true,
            max_operations_per_stream: 1000,
        }
    }
}

impl HipStreamManager {
    pub fn new(config: HipStreamConfig) -> Self {
        Self {
            streams: Vec::new(),
            stream_pool: std::collections::VecDeque::new(),
            next_stream_id: 0,
            config,
        }
    }

    /// Create new stream
    /// Create new stream
    ///
    /// Reuses a previously [`Self::destroy_stream`]d stream from
    /// `stream_pool` when one is available (its operation queue is cleared
    /// and it is given a fresh ID) instead of always allocating a new one.
    pub fn create_stream(&mut self, priority: Option<i32>) -> Result<u32, RocmError> {
        let stream_id = self.next_stream_id;
        self.next_stream_id += 1;

        let mut stream = self.stream_pool.pop_front().unwrap_or_else(|| HipStream {
            handle: std::ptr::null_mut(), // Would be actual HIP stream
            id: stream_id,
            priority: priority.unwrap_or(self.config.default_priority),
            flags: HipStreamFlags::default(),
            created_at: Instant::now(),
            operations: std::collections::VecDeque::new(),
        });
        stream.id = stream_id;
        stream.priority = priority.unwrap_or(self.config.default_priority);
        stream.created_at = Instant::now();
        stream.operations.clear();

        self.streams.push(stream);
        Ok(stream_id)
    }

    /// Destroy stream
    ///
    /// Returns the stream to `stream_pool` for [`Self::create_stream`] to
    /// reuse instead of dropping it outright.
    pub fn destroy_stream(&mut self, stream_id: u32) -> Result<(), RocmError> {
        if let Some(pos) = self.streams.iter().position(|s| s.id == stream_id) {
            let stream = self.streams.remove(pos);
            self.stream_pool.push_back(stream);
            Ok(())
        } else {
            Err(RocmError::InvalidStream("Stream not found".to_string()))
        }
    }

    /// Add operation to stream
    pub fn add_operation(
        &mut self,
        stream_id: u32,
        operation: HipOperation,
    ) -> Result<(), RocmError> {
        if let Some(stream) = self.streams.iter_mut().find(|s| s.id == stream_id) {
            if stream.operations.len() >= self.config.max_operations_per_stream {
                return Err(RocmError::StreamFull(
                    "Stream operation queue is full".to_string(),
                ));
            }

            stream.operations.push_back(operation);
            Ok(())
        } else {
            Err(RocmError::InvalidStream("Stream not found".to_string()))
        }
    }

    /// Synchronize stream
    pub fn synchronize_stream(&mut self, stream_id: u32) -> Result<(), RocmError> {
        // First, collect all operations from the stream
        let mut operations = Vec::new();
        if let Some(stream) = self.streams.iter_mut().find(|s| s.id == stream_id) {
            while let Some(operation) = stream.operations.pop_front() {
                operations.push(operation);
            }
        } else {
            return Err(RocmError::InvalidStream("Stream not found".to_string()));
        }

        // Now execute all operations
        for operation in operations {
            self.execute_operation(operation)?;
        }

        Ok(())
    }

    fn execute_operation(&self, _operation: HipOperation) -> Result<(), RocmError> {
        // Host-memory simulation (see module docs): there is no real ROCm/HIP
        // device to transfer to or from, so no data movement happens here.
        // This used to also inject an artificial `std::thread::sleep` per
        // operation type to mimic device-transfer latency; that fake timing
        // has been removed rather than left as an undisclosed simulated
        // number, so callers now see the true (near-zero) cost of this
        // simulation instead of a fabricated one.
        Ok(())
    }
}

/// ROCm statistics
#[derive(Debug, Clone, Default)]
pub struct RocmStats {
    pub total_allocations: u64,
    pub total_deallocations: u64,
    pub bytes_allocated: u64,
    pub bytes_deallocated: u64,
    pub device_memory_used: usize,
    pub host_memory_used: usize,
    pub coarse_grained_used: usize,
    pub fine_grained_used: usize,
    pub stream_operations: u64,
    pub kernel_launches: u64,
    pub memory_transfers: u64,
    pub average_allocation_time: Duration,
    pub peak_memory_usage: usize,
}

impl RocmMemoryBackend {
    /// Create new ROCm backend
    pub fn new(config: RocmConfig) -> Result<Self, RocmError> {
        // Initialize ROCm device
        let device_properties = Self::query_device_properties(config.device_id)?;

        // Create memory pools
        let mut memory_pools = HashMap::new();
        if config.enable_memory_pools {
            let pool_size = device_properties.total_global_memory / 4; // Use 1/4 of total memory

            memory_pools.insert(
                RocmMemoryType::Device,
                RocmMemoryPool::new(RocmMemoryType::Device, pool_size),
            );
            memory_pools.insert(
                RocmMemoryType::Host,
                RocmMemoryPool::new(RocmMemoryType::Host, pool_size),
            );

            if config.enable_coarse_memory {
                memory_pools.insert(
                    RocmMemoryType::CoarseGrained,
                    RocmMemoryPool::new(RocmMemoryType::CoarseGrained, pool_size),
                );
            }

            if config.enable_fine_memory {
                memory_pools.insert(
                    RocmMemoryType::FineGrained,
                    RocmMemoryPool::new(RocmMemoryType::FineGrained, pool_size / 2),
                );
            }

            if config.enable_host_visible {
                memory_pools.insert(
                    RocmMemoryType::HostVisible,
                    RocmMemoryPool::new(RocmMemoryType::HostVisible, pool_size / 4),
                );
            }
        }

        let stream_manager = HipStreamManager::new(HipStreamConfig::default());

        Ok(Self {
            config,
            device_properties,
            contexts: HashMap::new(),
            memory_pools,
            stats: RocmStats::default(),
            stream_manager,
        })
    }

    /// Query device properties
    fn query_device_properties(device_id: u32) -> Result<RocmDeviceProperties, RocmError> {
        // Simulate querying ROCm device properties
        Ok(RocmDeviceProperties {
            device_id,
            name: format!("AMD GPU {}", device_id),
            arch: "gfx906".to_string(), // Vega architecture
            gcn_arch_name: "Vega20".to_string(),
            total_global_memory: 16 * 1024 * 1024 * 1024, // 16GB
            local_memory_size: 64 * 1024,                 // 64KB
            max_work_group_size: 1024,
            max_work_item_dimensions: 3,
            max_work_item_sizes: [1024, 1024, 1024],
            compute_units: 64,
            wavefront_size: 64,
            memory_clock_frequency: 1000000, // 1 GHz
            memory_bus_width: 4096,
            l2_cache_size: 4 * 1024 * 1024,      // 4MB
            max_constant_buffer_size: 64 * 1024, // 64KB
            pci_bus_id: 0x03,
            pci_device_id: 0x66AF,
            supports_cooperative_launch: true,
            supports_dynamic_parallelism: false,
        })
    }

    /// Allocate device memory
    pub fn allocate(
        &mut self,
        size: usize,
        memory_type: RocmMemoryType,
    ) -> Result<*mut c_void, RocmError> {
        let start_time = Instant::now();

        let ptr = if self.config.enable_memory_pools {
            if let Some(pool) = self.memory_pools.get_mut(&memory_type) {
                pool.allocate(size)?
            } else {
                return Err(RocmError::UnsupportedMemoryType(
                    "Memory type not supported".to_string(),
                ));
            }
        } else {
            // Direct allocation
            self.direct_allocate(size, memory_type.clone())?
        };

        // Update statistics
        self.stats.total_allocations += 1;
        self.stats.bytes_allocated += size as u64;

        match memory_type {
            RocmMemoryType::Device => self.stats.device_memory_used += size,
            RocmMemoryType::Host => self.stats.host_memory_used += size,
            RocmMemoryType::CoarseGrained => self.stats.coarse_grained_used += size,
            RocmMemoryType::FineGrained => self.stats.fine_grained_used += size,
            _ => {}
        }

        let allocation_time = start_time.elapsed();
        let total_time = self.stats.average_allocation_time.as_nanos() as u64
            * (self.stats.total_allocations - 1)
            + allocation_time.as_nanos() as u64;
        self.stats.average_allocation_time =
            Duration::from_nanos(total_time / self.stats.total_allocations);

        let current_usage = self.stats.device_memory_used
            + self.stats.host_memory_used
            + self.stats.coarse_grained_used
            + self.stats.fine_grained_used;
        if current_usage > self.stats.peak_memory_usage {
            self.stats.peak_memory_usage = current_usage;
        }

        Ok(ptr)
    }

    fn direct_allocate(
        &self,
        size: usize,
        memory_type: RocmMemoryType,
    ) -> Result<*mut c_void, RocmError> {
        // Simulate direct HIP allocation
        match memory_type {
            RocmMemoryType::Device => sim_alloc(size), // hipMalloc
            RocmMemoryType::Host => sim_alloc(size),   // hipMallocHost
            RocmMemoryType::CoarseGrained => sim_alloc(size), // coarse-grained device memory
            RocmMemoryType::FineGrained => sim_alloc(size), // fine-grained system memory
            _ => Err(RocmError::UnsupportedMemoryType(
                "Unsupported memory type".to_string(),
            )),
        }
    }

    /// Free device memory
    pub fn free(&mut self, ptr: *mut c_void, memory_type: RocmMemoryType) -> Result<(), RocmError> {
        if self.config.enable_memory_pools {
            if let Some(pool) = self.memory_pools.get_mut(&memory_type) {
                pool.free(ptr)?;
            } else {
                return Err(RocmError::UnsupportedMemoryType(
                    "Memory type not supported".to_string(),
                ));
            }
        } else {
            // Direct deallocation. `ptr` was returned by `sim_alloc` via
            // `hip_malloc`/`direct_allocate` above, and this is the first
            // time it is freed.
            sim_dealloc(ptr);
        }

        self.stats.total_deallocations += 1;
        Ok(())
    }

    /// Copy memory
    pub fn memcpy(
        &mut self,
        dst: *mut c_void,
        src: *const c_void,
        size: usize,
        kind: RocmMemcpyKind,
    ) -> Result<(), RocmError> {
        let operation = HipOperation {
            op_type: match kind {
                RocmMemcpyKind::HostToDevice => HipOperationType::MemcpyHostToDevice,
                RocmMemcpyKind::DeviceToHost => HipOperationType::MemcpyDeviceToHost,
                RocmMemcpyKind::DeviceToDevice => HipOperationType::MemcpyDeviceToDevice,
                RocmMemcpyKind::HostToHost => HipOperationType::MemcpyAsync,
            },
            src_ptr: Some(src as *mut c_void),
            dst_ptr: Some(dst),
            size,
            timestamp: Instant::now(),
        };

        // Execute synchronously for now
        self.stream_manager.execute_operation(operation)?;
        self.stats.memory_transfers += 1;

        Ok(())
    }

    /// Asynchronous memory copy
    pub fn memcpy_async(
        &mut self,
        dst: *mut c_void,
        src: *const c_void,
        size: usize,
        kind: RocmMemcpyKind,
        stream_id: u32,
    ) -> Result<(), RocmError> {
        let operation = HipOperation {
            // Mirrors the synchronous `memcpy`'s mapping above so the queued
            // operation's recorded direction matches what the caller asked
            // for instead of always reporting a generic `MemcpyAsync`.
            op_type: match kind {
                RocmMemcpyKind::HostToDevice => HipOperationType::MemcpyHostToDevice,
                RocmMemcpyKind::DeviceToHost => HipOperationType::MemcpyDeviceToHost,
                RocmMemcpyKind::DeviceToDevice => HipOperationType::MemcpyDeviceToDevice,
                RocmMemcpyKind::HostToHost => HipOperationType::MemcpyAsync,
            },
            src_ptr: Some(src as *mut c_void),
            dst_ptr: Some(dst),
            size,
            timestamp: Instant::now(),
        };

        self.stream_manager.add_operation(stream_id, operation)?;
        Ok(())
    }

    /// Create HIP context
    pub fn create_context(&mut self, flags: HipContextFlags) -> Result<u32, RocmError> {
        let context_id = self.contexts.len() as u32;

        let memory_info = HipMemoryInfo {
            total_memory: self.device_properties.total_global_memory,
            free_memory: self.device_properties.total_global_memory - self.stats.device_memory_used,
            used_memory: self.stats.device_memory_used,
            coarse_memory: self.stats.coarse_grained_used,
            fine_memory: self.stats.fine_grained_used,
        };

        let context = HipContext {
            handle: std::ptr::null_mut(), // Would be actual HIP context
            device_id: self.config.device_id,
            flags,
            created_at: Instant::now(),
            streams: Vec::new(),
            memory_info,
        };

        self.contexts.insert(context_id, context);
        Ok(context_id)
    }

    /// Get device properties
    pub fn get_device_properties(&self) -> &RocmDeviceProperties {
        &self.device_properties
    }

    /// Get statistics
    pub fn get_stats(&self) -> &RocmStats {
        &self.stats
    }

    /// Synchronize device
    pub fn device_synchronize(&mut self) -> Result<(), RocmError> {
        // Synchronize all streams
        let stream_ids: Vec<u32> = self.stream_manager.streams.iter().map(|s| s.id).collect();
        for stream_id in stream_ids {
            self.stream_manager.synchronize_stream(stream_id)?;
        }
        Ok(())
    }

    /// Create stream
    pub fn create_stream(&mut self, priority: Option<i32>) -> Result<u32, RocmError> {
        self.stream_manager.create_stream(priority)
    }

    /// Destroy stream
    pub fn destroy_stream(&mut self, stream_id: u32) -> Result<(), RocmError> {
        self.stream_manager.destroy_stream(stream_id)
    }

    /// Query memory attributes
    ///
    /// Looks up which pool actually allocated `ptr` and returns that pool's
    /// real attributes (which vary by [`RocmMemoryType`] — see the
    /// attribute construction in [`RocmMemoryPool::new`]) instead of an
    /// unconditional default. A pointer this backend never allocated is an
    /// honest `Err`, not a guess.
    pub fn query_memory_attributes(
        &self,
        ptr: *mut c_void,
    ) -> Result<RocmMemoryAttributes, RocmError> {
        self.memory_pools
            .values()
            .find(|pool| pool.allocated_blocks.contains_key(&ptr))
            .map(|pool| pool.attributes.clone())
            .ok_or_else(|| {
                RocmError::InvalidPointer("pointer was not allocated by this backend".to_string())
            })
    }
}

// Safety: RocmMemoryBackend manages ROCm/HIP GPU memory pointers via *mut c_void.
// While raw pointers are not Send/Sync by default, it's safe to share across threads
// when protected by Arc<Mutex<>> because:
// 1. All pointers point to HIP GPU memory managed by the ROCm driver
// 2. The Mutex provides exclusive access for all mutable operations
// 3. No thread-local state is maintained
unsafe impl Send for RocmMemoryBackend {}
unsafe impl Sync for RocmMemoryBackend {}

/// ROCm memory copy kinds
#[derive(Debug, Clone)]
pub enum RocmMemcpyKind {
    HostToDevice,
    DeviceToHost,
    DeviceToDevice,
    HostToHost,
}

/// ROCm errors
#[derive(Debug, Clone)]
pub enum RocmError {
    DeviceNotFound(String),
    OutOfMemory(String),
    InvalidPointer(String),
    InvalidStream(String),
    StreamFull(String),
    UnsupportedOperation(String),
    UnsupportedMemoryType(String),
    ContextCreationFailed(String),
    KernelLaunchFailed(String),
    SynchronizationFailed(String),
    InternalError(String),
}

impl std::fmt::Display for RocmError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            RocmError::DeviceNotFound(msg) => write!(f, "Device not found: {}", msg),
            RocmError::OutOfMemory(msg) => write!(f, "Out of memory: {}", msg),
            RocmError::InvalidPointer(msg) => write!(f, "Invalid pointer: {}", msg),
            RocmError::InvalidStream(msg) => write!(f, "Invalid stream: {}", msg),
            RocmError::StreamFull(msg) => write!(f, "Stream full: {}", msg),
            RocmError::UnsupportedOperation(msg) => write!(f, "Unsupported operation: {}", msg),
            RocmError::UnsupportedMemoryType(msg) => write!(f, "Unsupported memory type: {}", msg),
            RocmError::ContextCreationFailed(msg) => write!(f, "Context creation failed: {}", msg),
            RocmError::KernelLaunchFailed(msg) => write!(f, "Kernel launch failed: {}", msg),
            RocmError::SynchronizationFailed(msg) => write!(f, "Synchronization failed: {}", msg),
            RocmError::InternalError(msg) => write!(f, "Internal error: {}", msg),
        }
    }
}

impl std::error::Error for RocmError {}

/// Thread-safe ROCm backend wrapper
pub struct ThreadSafeRocmBackend {
    backend: Arc<Mutex<RocmMemoryBackend>>,
}

impl ThreadSafeRocmBackend {
    pub fn new(config: RocmConfig) -> Result<Self, RocmError> {
        let backend = RocmMemoryBackend::new(config)?;
        Ok(Self {
            backend: Arc::new(Mutex::new(backend)),
        })
    }

    pub fn allocate(
        &self,
        size: usize,
        memory_type: RocmMemoryType,
    ) -> Result<*mut c_void, RocmError> {
        let mut backend = self.backend.lock().unwrap_or_else(|e| e.into_inner());
        backend.allocate(size, memory_type)
    }

    pub fn free(&self, ptr: *mut c_void, memory_type: RocmMemoryType) -> Result<(), RocmError> {
        let mut backend = self.backend.lock().unwrap_or_else(|e| e.into_inner());
        backend.free(ptr, memory_type)
    }

    pub fn get_stats(&self) -> RocmStats {
        let backend = self.backend.lock().unwrap_or_else(|e| e.into_inner());
        backend.get_stats().clone()
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    /// Regression test for F26: a zero-size request must not reach
    /// `std::alloc::alloc` (unsound for a zero-size layout).
    #[test]
    fn sim_alloc_zero_size_is_a_safe_sentinel_not_a_ub_call() {
        let ptr = sim_alloc(0).expect("zero-size request must succeed");
        assert!(!ptr.is_null());
        sim_dealloc(ptr);
    }

    /// A real allocation must be readable/writable across its full size and
    /// must free through the same layout it was allocated with.
    #[test]
    fn sim_alloc_real_allocation_round_trips_and_frees_cleanly() {
        for size in [1usize, 7, 256, 4096, 1_000_003] {
            let ptr = sim_alloc(size).expect("allocation must succeed") as *mut u8;
            assert!(!ptr.is_null());
            unsafe {
                for i in 0..size {
                    ptr.add(i).write(0xAB);
                }
                for i in 0..size {
                    assert_eq!(ptr.add(i).read(), 0xAB);
                }
                sim_dealloc(ptr as *mut c_void);
            }
        }
    }

    #[test]
    fn sim_dealloc_null_is_a_no_op() {
        sim_dealloc(std::ptr::null_mut());
    }

    #[test]
    fn test_rocm_backend_creation() {
        let config = RocmConfig::default();
        let backend = RocmMemoryBackend::new(config);
        assert!(backend.is_ok());
    }

    #[test]
    fn test_memory_pool() {
        let mut pool = RocmMemoryPool::new(RocmMemoryType::CoarseGrained, 1024 * 1024);
        let ptr = pool.allocate(1024);
        assert!(ptr.is_ok());

        let ptr = ptr.expect("unwrap failed");
        let result = pool.free(ptr);
        assert!(result.is_ok());
    }

    #[test]
    fn test_hip_stream_manager() {
        let mut manager = HipStreamManager::new(HipStreamConfig::default());
        let stream_id = manager.create_stream(Some(1));
        assert!(stream_id.is_ok());

        let stream_id = stream_id.expect("unwrap failed");
        let result = manager.destroy_stream(stream_id);
        assert!(result.is_ok());
    }

    #[test]
    fn test_thread_safe_backend() {
        let config = RocmConfig::default();
        let backend = ThreadSafeRocmBackend::new(config);
        assert!(backend.is_ok());

        let backend = backend.expect("unwrap failed");
        let stats = backend.get_stats();
        assert_eq!(stats.total_allocations, 0);
    }
}