fix: shaders.

Author: b4rtaz

src/nn/vulkan/cast-forward-f32-f32.comp

@@ -15,15 +15,13 @@ layout(binding = 0) readonly buffer inputBuffer { float x[]; };
 layout(binding = 1) writeonly buffer outputBuffer { float y[]; };
 layout(binding = 2) readonly buffer batchInfosBuffer { BatchInfo infos[]; };
 
-shared uint sharedPosition;
 shared BatchInfo sharedInfo;
 
 void main() {
     const uint threadIndex = gl_LocalInvocationID.x;
     const uint batchIndex = gl_GlobalInvocationID.y;
 
     if (threadIndex == 0) {
-        sharedPosition = uint(x[batchIndex]);
         sharedInfo = infos[batchIndex];
     }
 

src/nn/vulkan/mul-forward-f32-f32.comp

@@ -31,10 +31,11 @@ void main() {
     memoryBarrierShared();
     barrier();
 
-    const uint start = sharedInfo.inputOffset + threadIndex;
-    const uint end = sharedInfo.inputOffset + sharedInfo.inputSizeX;
+    const uint inputSizeX = sharedInfo.inputSizeX;
+    const uint xyOffset = sharedInfo.inputOffset;
+    const uint mOffset = inputSizeX * batchIndex;
 
-    for (uint i = threadIndex; i < end; i += N_THREADS) {
-        y[i] = x[i] * m[i];
+    for (uint i = threadIndex; i < inputSizeX; i += N_THREADS) {
+        y[xyOffset + i] = x[xyOffset + i] * m[mOffset + i];
     }
 }