update

include/experimental/kernel/int4_dequantize/int4_dequantize_xe_impl.hpp

@@ -241,7 +241,6 @@ struct int4_dequantize_t<
 
     mat_qweight_t mat_qweight;
     mat_dequant_weight_t mat_dequant_weight;
-    mat_dequant_weight_t trans_mat_dequant_weight;
     scale_t scale;
     zp_t zp;
 
@@ -279,12 +278,8 @@ struct int4_dequantize_t<
       }
       SW_BARRIER();
       dequantize(mat_dequant_weight, mat_qweight, scale, zp, dequantize_args);
-      //   for (uint32_t j = 0; j < sg_tile_n; j++) {
-      //     trans_mat_dequant_weight.reg.xetla_select<k_stride, sg_tile_n>(j) =
-      //         mat_dequant_weight.reg.xetla_select<k_stride, 1>(j * k_stride);
-      //   }
+      tile_transpose(mat_dequant_weight);
       subgroup::tile_store(mat_dequant_weight, mat_dequant_weight_payload);
-      //   trans_mat_dequant_weight, mat_dequant_weight_payload);
       mat_dequant_weight_payload.template update_tdesc<tdesc_update_dir::y_dir>(
           mat_dequant_weight_t::tile_size_y);
       SW_BARRIER();

tests/integration/int4_dequantize/main.cpp

@@ -130,8 +130,9 @@ int int4_dequantize_result_validate(T* gold, T* out, size_t k, size_t n) {
   for (uint32_t i = 0; i < k; i++) {
     for (uint32_t j = 0; j < n; j++) {
       if (gold[i * n + j] != out[i * n + j]) {
-        std::cout << i * n + j << " " << gold[i * n + j] << " "
-                  << out[i * n + j] << std::endl;
+        if (err_num < 10)
+          std::cout << i * n + j << " " << gold[i * n + j] << " "
+                    << out[i * n + j] << std::endl;
         err_num++;
       }
     }