Merge pull request #12 from trusted-programming/mem_alloc

feat: implement vadd in npu functions

Author: lucarlig

descend_derive/build.rs

@@ -4,7 +4,7 @@ use std::path::Path;
 fn main() {
     // Tell Cargo to re-run this build script if any .desc files change
     let examples_dir = "examples/core";
-    
+
     if Path::new(examples_dir).exists() {
         // Walk through the directory and tell Cargo to re-run if any .desc files change
         if let Ok(entries) = fs::read_dir(examples_dir) {
@@ -17,7 +17,7 @@ fn main() {
             }
         }
     }
-    
+
     // Also watch the entire directory for new files
     println!("cargo:rerun-if-changed={}", examples_dir);
 }

examples/core/assign.desc

@@ -0,0 +1,7 @@
+fn assign<n: nat, r: prv>(
+    a: &r shrd gpu.global [i16; 16],
+    b: &r uniq gpu.global [i16; 16]
+) -[grid: gpu.grid<X<1>, X<16>>]-> () {
+    b = a;
+    ()
+}

examples/core/vec_add_3.desc.off renamed to examples/core/vadd.desc

@@ -0,0 +1,7 @@
+fn div<n: nat, r: prv>(
+    a: &r shrd gpu.global [i16; 16],
+    b: &r shrd gpu.global [i16; 16]
+) -[grid: gpu.grid<X<1>, X<16>>]-> () {
+    a / b;
+    ()
+}

examples/core/vdiv.desc

@@ -0,0 +1,48 @@
+// Vector addition kernel demonstrating Descend's safe GPU programming model
+// This function showcases extended borrow checking, memory safety, and execution context tracking
+
+// Generic function with type parameters:
+// - n: nat - Natural number parameter (for array size, though not used in this specific function)
+// - r: prv - Provenance parameter tracking memory region/lifetime for all references
+fn add<n: nat, r: prv>(
+    // Shared reference to first input vector - multiple threads can read simultaneously
+    // Memory space: gpu.global (GPU global memory)
+    // Ownership: shrd (shared) - prevents write-after-read data races
+    // Type: 16-element array of 16-bit signed integers
+    a: &r shrd gpu.global [i16; 16],
+    
+    // Shared reference to second input vector - multiple threads can read simultaneously  
+    // Same memory space and ownership constraints as 'a'
+    b: &r shrd gpu.global [i16; 16],
+    
+    // Unique reference to output vector - only one thread can write at a time
+    // Ownership: uniq (unique) - prevents write-after-write data races
+    // The compiler statically ensures no conflicting borrows exist
+    c: &r uniq gpu.global [i16; 16]
+    
+// Execution context specification - defines how this function runs on GPU hardware
+// - grid: gpu.grid<X<1>, X<16>> - GPU execution grid with 1 block containing 16 threads
+// - The type system ensures GPU memory is only accessed in GPU execution contexts
+// - Prevents invalid cross-device memory accesses (CPU accessing GPU memory)
+) -[grid: gpu.grid<X<1>, X<16>>]-> () {
+    
+    // Vector addition operation - element-wise addition of arrays
+    // The compiler generates safe parallel code that:
+    // 1. Loads data from global memory to local memory for each thread
+    // 2. Performs vectorized addition using HIVM dialect operations
+    // 3. Stores results back to global memory safely
+    // The ownership system ensures this operation is race-free
+    //
+    // LAZY LOADING: Descend's compiler implements lazy loading strategies:
+    // - Memory loads are deferred until actually needed by computation
+    // - The HIVM dialect generates 'hivm.hir.load' operations that load from
+    //   global memory (gm) to local memory (ub) only when data is accessed
+    // - This minimizes memory bandwidth usage and improves cache efficiency
+    // - The type system ensures loads happen in the correct execution context
+    // - Shared references enable read-only access without unnecessary copies
+    c = a + b;
+    
+    // Unit return value - indicates successful completion
+    // In MLIR, this becomes a 'return' operation
+    ()
+}

examples/core/vec_add.desc

@@ -0,0 +1,7 @@
+fn mul<n: nat, r: prv>(
+    a: &r shrd gpu.global [i16; 16],
+    b: &r shrd gpu.global [i16; 16]
+) -[grid: gpu.grid<X<1>, X<16>>]-> () {
+    a * b;
+    ()
+}

examples/core/vec_add.desc.off

@@ -0,0 +1,7 @@
+fn assign_to_shared_ref<n: nat, r: prv>(
+    a: &r shrd gpu.global [i16; 16],
+    b: &r shrd gpu.global [i16; 16]
+) -[grid: gpu.grid<X<1>, X<16>>]-> () {
+    b = a;  // This should fail - cannot assign to shared reference
+    ()
+}