Removes col_stride parameter from offset methods

Simplifies mask_offset and bias_offset method signatures by removing the unused col_stride parameter and eliminating col_stride multiplication in offset calculations.

The leftpad_k value is now added directly to the offset instead of being multiplied by col_stride, streamlining the computation logic.

Author: LoserCheems

csrc/src/block_info.h

@@ -36,16 +36,16 @@ struct BlockInfo {
     }
 
     template <typename index_t>
-    __forceinline__ __device__ index_t mask_offset(const index_t batch_stride, const index_t row_stride, const index_t col_stride, const int bidb) const {
+    __forceinline__ __device__ index_t mask_offset(const index_t batch_stride, const index_t row_stride, const int bidb) const {
         index_t offset = sum_s_q == -1 ? bidb * batch_stride : uint32_t(sum_s_q) * row_stride;
-        sum_s_k == -1 ? offset += leftpad_k * col_stride : offset += uint32_t(sum_s_k + leftpad_k) * col_stride;
+        sum_s_k == -1 ? offset += leftpad_k : offset += uint32_t(sum_s_k + leftpad_k);
         return offset;
     }
 
     template <typename index_t>
-    __forceinline__ __device__ index_t bias_offset(const index_t batch_stride, const index_t row_stride, const index_t col_stride, const int bidb) const {
+    __forceinline__ __device__ index_t bias_offset(const index_t batch_stride, const index_t row_stride, const int bidb) const {
         index_t offset = sum_s_q == -1 ? bidb * batch_stride : uint32_t(sum_s_q) * row_stride;
-        sum_s_k == -1 ? offset += leftpad_k * col_stride : offset += uint32_t(sum_s_k + leftpad_k) * col_stride;
+        sum_s_k == -1 ? offset += leftpad_k : offset += uint32_t(sum_s_k + leftpad_k);
         return offset;
     }