ARCH/X86: Use UCS function to count leading zeros

Author: tvegas1

contrib/test_jenkins.sh

@@ -1128,12 +1128,9 @@ run_release_mode_tests() {
 # Run nt_buffer_transfer tests
 #
 run_nt_buffer_transfer_tests() {
-    if lscpu | grep -q 'AuthenticAMD'
-    then
-	    build release --enable-gtest --enable-optimizations
-	    echo "==== Running nt_buffer_transfer tests ===="
-	    ./test/gtest/gtest --gtest_filter="test_arch.nt_buffer_transfer_*"
-    fi
+    build release --enable-gtest --enable-optimizations
+    echo "==== Running test_arch tests with optimizations ===="
+    ./test/gtest/gtest --gtest_filter="test_arch.*"
 }
 
 set_ucx_common_test_env() {

src/ucs/arch/x86_64/cpu.c

@@ -1061,7 +1061,7 @@ size_t ucs_x86_nt_src_buffer_transfer(void *dst, const void *src, size_t len)
 }
 
 static UCS_F_ALWAYS_INLINE
-void ucs_x86_copy_bytes_le_128(void *dst, const void *src, size_t len)
+void ucs_x86_copy_bytes_le_128(void *dst, const void *src, uint32_t len)
 {
     __m256i y0, y1, y2, y3;
     /* Handle lengths that fall usually within eager short range */

test/gtest/ucs/arch/test_x86_64.cc

@@ -76,6 +76,12 @@ class test_arch : public ucs::test {
         test_window_size = 8 * 1024;
         hole_size        = 2 * align;
 
+        auto msg = [&]() {
+            std::stringstream ss;
+            ss << "using length=" << len << " src_align=" << i << " dst_align=" << j;
+            return ss.str();
+        };
+
         /*
          * Allocate a hole above and below the test_window_size
          * to check for writes beyond the designated area.
@@ -113,14 +119,20 @@ class test_arch : public ucs::test {
                     /* Perform the transfer */
                     ucs_x86_nt_buffer_transfer(dst + i, src + j, len, hint, len);
                     result = memcmp(src + j, dst + i, len);
-                    EXPECT_EQ(0, result);
+                    EXPECT_EQ(0, result) << msg();
+                    if (result) {
+                        goto terminate;
+                    }
 
                     /* reset the copied region back to zero */
                     memset(dst + i, 0x0, len);
 
                     /* check for any modifications in the holes */
                     result = memcmp(test_window_dst, dup, total_size);
-                    EXPECT_EQ(0, result);
+                    EXPECT_EQ(0, result) << msg();
+                    if (result) {
+                        goto terminate;
+                    }
                 }
             }
             /* Check for each len for less than 1k sizes
@@ -133,6 +145,7 @@ class test_arch : public ucs::test {
             }
         }
 
+    terminate:
         free(dup);
 
     dup_fail: