x64: matmul: Weight decompression fix group blocking

Author: inteldimitrius

src/cpu/x64/matmul/brgemm_matmul.cpp

@@ -1304,19 +1304,19 @@ void brgemm_matmul_t<isa>::copy_b_chunk_in_buffer(
 
         // Handle first block
         if (k_start % adj_k_blk > 0) {
-            call_copy_kernel(k_start, k_start % adj_k_blk, 0);
-            k += k_start % adj_k_blk;
+            const auto first_blk_size = adj_k_blk - (k_start % adj_k_blk);
+            call_copy_kernel(k_start, first_blk_size, 0);
+            k += first_blk_size;
         }
         // Handle full blocks
-        for (; k < k_end; k += adj_k_blk) {
+        for (; (k + adj_k_blk) <= k_end; k += adj_k_blk) {
             const auto gb = (k - k_start) / bgmmc.K_blk;
             call_copy_kernel(k, adj_k_blk, gb);
         }
         // Handle last block
-        if (k_end % adj_k_blk > 0) {
-            k -= adj_k_blk;
+        if (k_end > k) {
             const auto gb = (k - k_start) / bgmmc.K_blk;
-            call_copy_kernel(k, k_end % adj_k_blk, gb);
+            call_copy_kernel(k, k_end - k, gb);
         }
     } else { // Default case with k_blk blocking
         for (int gb = 0; gb < gemm_batch; ++gb) {

tests/benchdnn/inputs/matmul/harness_matmul_decompression

@@ -330,3 +330,29 @@
 --attr-zero-points=,wei:common:2,wei:per_ocic:s8:32x1
 --attr-fpmath=strict:true
 1x4096:4096x4096
+
+
+## Additional grouped scales/ZP testing
+--reset
+--dt=f16:s4:f16
+--wtag=abc,acb
+--attr-scales=,wei:common:2,wei:per_oc:f16,wei:per_ocic:f16:192x1
+--attr-zero-points=,wei:common:2,wei:per_oc:s4,wei:per_ocic:s4:192x1
+--attr-fpmath=f16:true
+12x4x576:12x576x192
+
+--reset
+--dt=bf16:s4:bf16
+--wtag=abc,acb
+--attr-scales=,wei:common:2,wei:per_oc:f16,wei:per_ocic:f16:192x1
+--attr-zero-points=,wei:common:2,wei:per_oc:s4,wei:per_ocic:s4:192x1
+--attr-fpmath=bf16:true
+12x4x576:12x576x192
+
+--reset
+--dt=f32:s4:f32
+--wtag=abc,acb
+--attr-scales=,wei:common:2,wei:per_oc:f16,wei:per_ocic:f16:192x1
+--attr-zero-points=,wei:common:2,wei:per_oc:s4,wei:per_ocic:s4:192x1
+--attr-fpmath=strict:true
+12x4x576:12x576x192