vulkan: add noncontiguous GLU support (#21081)

* vulkan: add noncontiguous GLU support

* fix compile issue

ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl

@@ -16,4 +16,14 @@ layout (push_constant) uniform parameter
     uint mode;
     float alpha;
     float limit;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint ne01;
+    uint ne02;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint ne11;
+    uint ne12;
 } p;

ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl

@@ -8,22 +8,32 @@ void main() {
     const uint row = i / p.ne20;
     const uint col = i - row * p.ne20;
 
+    const uint i3 = row / (p.ne01 * p.ne02);
+    const uint i2 = (row % (p.ne01 * p.ne02)) / p.ne01;
+    const uint i1 = row % p.ne01;
+    const uint src_idx = i3 * p.nb03 + i2 * p.nb02 + i1 * p.nb01 + col;
+
+    const uint dst_i3 = row / (p.ne11 * p.ne12);
+    const uint dst_i2 = (row % (p.ne11 * p.ne12)) / p.ne11;
+    const uint dst_i1 = row % p.ne11;
+    const uint dst_idx = dst_i3 * p.nb13 + dst_i2 * p.nb12 + dst_i1 * p.nb11 + col;
+
     if (p.mode == 0) {
         // Default
         const uint offset = p.ne00 / 2;
-        const uint idx = row * p.ne00 + col;
+        const uint idx = src_idx;
 
-        data_d[row * offset + col] = D_TYPE(op(float(data_a[idx]), float(data_a[idx + offset])));
+        data_d[dst_idx] = D_TYPE(op(float(data_a[idx]), float(data_a[idx + offset])));
     } else if (p.mode == 1) {
         // Swapped
         const uint offset = p.ne00 / 2;
-        const uint idx = row * p.ne00 + col;
+        const uint idx = src_idx;
 
-        data_d[row * offset + col] = D_TYPE(op(float(data_a[idx + offset]), float(data_a[idx])));
+        data_d[dst_idx] = D_TYPE(op(float(data_a[idx + offset]), float(data_a[idx])));
     } else {
         // Split
-        const uint idx = row * p.ne00 + col;
+        const uint idx = src_idx;
 
-        data_d[idx] = D_TYPE(op(float(data_a[idx]), float(data_b[idx])));
+        data_d[dst_idx] = D_TYPE(op(float(data_a[idx]), float(data_b[idx])));
     }
 }