ser-3bit getwei

intel · Jan 24, 2024 · b322d68 · b322d68
1 parent 86771a5
commit b322d68
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 7 deletions.
diff --git a/bestla/bestla/kernel_jit.h b/bestla/bestla/kernel_jit.h
@@ -293,13 +293,13 @@ class DecompresssS3 {
         mov(reg_bit2ptr, ptr[parambase + OFFSET(bit2ptr)]);
         mov(reg_dst, ptr[parambase + OFFSET(dstptr)]);
         L("loop_label");
-        imul(reg_tmp, reg_iter, 32);
-        kmovq(bit1_mask1, ptr[reg_bit1ptr + reg_tmp]);
-        kmovq(bit1_mask2, ptr[reg_bit1ptr + reg_tmp + 8]);
-        kmovq(bit1_mask3, ptr[reg_bit1ptr + reg_tmp + 16]);
-        kmovq(bit1_mask4, ptr[reg_bit1ptr + reg_tmp + 24]);
+        imul(reg_tmp, reg_iter, 96);
+        kmovq(bit1_mask1, ptr[reg_bit2ptr + reg_tmp+64]);
+        kmovq(bit1_mask2, ptr[reg_bit2ptr + reg_tmp + 72]);
+        kmovq(bit1_mask3, ptr[reg_bit2ptr + reg_tmp + 80]);
+        kmovq(bit1_mask4, ptr[reg_bit2ptr + reg_tmp + 88]);
         Xbyak::Zmm bit2_data_zmm = zmm0;
-        imul(reg_tmp, reg_iter, 64);
+        // imul(reg_tmp, reg_iter, 64);
         vmovups(bit2_data_zmm, ptr[reg_bit2ptr + reg_tmp]);
         vextractf32x8(ymm2, bit2_data_zmm, 0x0);
         vextractf32x8(ymm3, bit2_data_zmm, 0x1);

diff --git a/bestla/bestla/ut/bestla_prologue_b.cpp b/bestla/bestla/ut/bestla_prologue_b.cpp
@@ -323,7 +323,7 @@ class UT_BlockQunatize_S3 {
     // }
   }
 };
-static UT_BlockQunatize_S3 sUT_BlockQunatize_S3;
+// static UT_BlockQunatize_S3 sUT_BlockQunatize_S3;
 #ifdef BTLA_UT_PROLOGUE_B
 static UT_BlockQunatize_S3 sUT_BlockQunatize_S3;
 #endif