chore: switch to buffer

candle-metal-kernels/src/lib.rs

@@ -1,6 +1,7 @@
 use metal::{
     Buffer, CommandBufferRef, CompileOptions, ComputeCommandEncoderRef, ComputePipelineState,
-    Device, Function, FunctionConstantValues, Library, MTLDataType, MTLSize, NSUInteger,
+    Device, Function, FunctionConstantValues, Library, MTLDataType, MTLResourceOptions, MTLSize,
+    NSUInteger,
 };
 use std::collections::HashMap;
 use std::ffi::c_void;
@@ -1359,17 +1360,20 @@ pub fn call_gemm(
         // TODO byte_stride_d
         let byte_stride_d = 0;
 
-        let buffer: Vec<u64> = vec![
-            byte_stride_a as _,
-            byte_stride_b as _,
-            byte_stride_c as _,
-            byte_stride_d as _,
-        ];
-        encoder.set_bytes(
-            10,
-            (buffer.len() * core::mem::size_of::<u64>()) as NSUInteger,
+        let mut buffer: Vec<u64> = Vec::with_capacity(b * 4);
+        for i in 0..b {
+            buffer.push((i * byte_stride_a) as u64);
+            buffer.push((i * byte_stride_b) as u64);
+            buffer.push((i * byte_stride_c) as u64);
+            buffer.push((i * byte_stride_d) as u64);
+        }
+
+        let matrix_offsets = device.new_buffer_with_data(
             buffer.as_ptr() as *const NSUInteger as *const c_void,
+            (buffer.len() * core::mem::size_of::<u64>()) as NSUInteger,
+            MTLResourceOptions::StorageModePrivate,
         );
+        encoder.set_buffer(10, Some(&matrix_offsets), 0);
     }
 
     let grid_size = MTLSize {