new method

Author: xiaolil1

csrc/xpu_cutlass_fusion.cpp

@@ -287,7 +287,7 @@ class gemm_4bit_cutlass_kernel {
     static constexpr auto K = decltype(size(out))::value / N; // 128 / 2 = 64
 
     using compress_type = uint8_t;
-    using vec_type = uint8_t;
+    using vec_type = uint32_t;
 
     static constexpr auto compress_size = sizeof_bits_v<compress_type> / sizeof_bits_v<SrcType>;
     static_assert((compress_size % N) == 0);
@@ -307,12 +307,18 @@ class gemm_4bit_cutlass_kernel {
       #pragma unroll
       for (int s = 0; s < K / compress_size; s++) {
 
+#if 0
         #pragma unroll
         for(int i = 0; i < compress_size / 2; i++) {
           int dst_offset = s * compress_size + i * 2;
           dst[dst_offset] = static_cast<DstType>(quant_map[(src[s] >> (4 * (i * 2 + 1))) & 0xf] * ts);
           dst[dst_offset + 1] = static_cast<DstType>(quant_map[(src[s] >> (4 * (i * 2))) & 0xf] * ts);
         }
+#else
+        int dst_offset = s * compress_size;
+        dst[dst_offset] = static_cast<DstType>(quant_map[src[s] >> 4] * ts);
+        dst[dst_offset + 1] = static_cast<DstType>(quant_map[src[s] & 0xf] * ts);
+#endif        
       }
     }
 #else