1  #include <arm_neon.h>
       2  #include "arm-neon-ref.h"
       3  #include "compute-ref-data.h"
       4  
       5  /* Expected results.  */
       6  VECT_VAR_DECL(expected,int,8,8) [] = { 0xaa, 0xaa, 0xaa, 0xaa,
       7  				       0xaa, 0xaa, 0xf0, 0xaa };
       8  VECT_VAR_DECL(expected,int,16,4) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xfff0 };
       9  VECT_VAR_DECL(expected,int,32,2) [] = { 0xaaaaaaaa, 0xfffffff0 };
      10  VECT_VAR_DECL(expected,int,64,1) [] = { 0xfffffffffffffff0 };
      11  VECT_VAR_DECL(expected,uint,8,8) [] = { 0xaa, 0xaa, 0xaa, 0xaa,
      12  					0xaa, 0xaa, 0xaa, 0xf0 };
      13  VECT_VAR_DECL(expected,uint,16,4) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xfff0 };
      14  VECT_VAR_DECL(expected,uint,32,2) [] = { 0xaaaaaaaa, 0xfffffff0 };
      15  VECT_VAR_DECL(expected,uint,64,1) [] = { 0xfffffffffffffff0 };
      16  VECT_VAR_DECL(expected,poly,8,8) [] = { 0xaa, 0xaa, 0xaa, 0xaa,
      17  					0xaa, 0xaa, 0xaa, 0xf0 };
      18  VECT_VAR_DECL(expected,poly,16,4) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xfff0 };
      19  VECT_VAR_DECL(expected,hfloat,16,4) [] = { 0xaaaa, 0xaaaa, 0xcc00, 0xaaaa };
      20  VECT_VAR_DECL(expected,hfloat,32,2) [] = { 0xaaaaaaaa, 0xc1800000 };
      21  VECT_VAR_DECL(expected,int,8,16) [] = { 0xaa, 0xaa, 0xaa, 0xaa,
      22  					0xaa, 0xaa, 0xaa, 0xaa,
      23  					0xaa, 0xaa, 0xaa, 0xaa,
      24  					0xaa, 0xaa, 0xaa, 0xf0 };
      25  VECT_VAR_DECL(expected,int,16,8) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xaaaa,
      26  					0xaaaa, 0xfff0, 0xaaaa, 0xaaaa };
      27  VECT_VAR_DECL(expected,int,32,4) [] = { 0xaaaaaaaa, 0xaaaaaaaa,
      28  					0xfffffff0, 0xaaaaaaaa };
      29  VECT_VAR_DECL(expected,int,64,2) [] = { 0xaaaaaaaaaaaaaaaa,
      30  					0xfffffffffffffff0 };
      31  VECT_VAR_DECL(expected,uint,8,16) [] = { 0xaa, 0xaa, 0xaa, 0xaa,
      32  					 0xaa, 0xaa, 0xaa, 0xaa,
      33  					 0xaa, 0xaa, 0xaa, 0xaa,
      34  					 0xf0, 0xaa, 0xaa, 0xaa };
      35  VECT_VAR_DECL(expected,uint,16,8) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xaaaa,
      36  					 0xaaaa, 0xaaaa, 0xfff0, 0xaaaa };
      37  VECT_VAR_DECL(expected,uint,32,4) [] = { 0xaaaaaaaa, 0xaaaaaaaa,
      38  					 0xfffffff0, 0xaaaaaaaa };
      39  VECT_VAR_DECL(expected,uint,64,2) [] = { 0xfffffffffffffff0,
      40  					 0xaaaaaaaaaaaaaaaa };
      41  VECT_VAR_DECL(expected,poly,8,16) [] = { 0xaa, 0xaa, 0xaa, 0xaa,
      42  					 0xaa, 0xaa, 0xaa, 0xaa,
      43  					 0xaa, 0xaa, 0xaa, 0xaa,
      44  					 0xf0, 0xaa, 0xaa, 0xaa };
      45  VECT_VAR_DECL(expected,poly,16,8) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xaaaa,
      46  					 0xaaaa, 0xaaaa, 0xfff0, 0xaaaa };
      47  VECT_VAR_DECL(expected,hfloat,16,8) [] = { 0xaaaa, 0xaaaa, 0xaaaa, 0xaaaa,
      48  					   0xaaaa, 0xcc00, 0xaaaa, 0xaaaa };
      49  VECT_VAR_DECL(expected,hfloat,32,4) [] = { 0xaaaaaaaa, 0xaaaaaaaa,
      50  					   0xc1800000, 0xaaaaaaaa };
      51  
      52  #define TEST_MSG "VLD1_LANE/VLD1_LANEQ"
      53  void exec_vld1_lane (void)
      54  {
      55    /* Fill vector_src with 0xAA, then load 1 lane.  */
      56  #define TEST_VLD1_LANE(Q, T1, T2, W, N, L)				\
      57    memset (VECT_VAR(buffer_src, T1, W, N), 0xAA, W/8*N);			\
      58    VECT_VAR(vector_src, T1, W, N) =					\
      59      vld1##Q##_##T2##W(VECT_VAR(buffer_src, T1, W, N));			\
      60    VECT_VAR(vector, T1, W, N) =						\
      61      vld1##Q##_lane_##T2##W(VECT_VAR(buffer, T1, W, N),			\
      62  			   VECT_VAR(vector_src, T1, W, N), L);		\
      63    vst1##Q##_##T2##W(VECT_VAR(result, T1, W, N), VECT_VAR(vector, T1, W, N))
      64  
      65    DECL_VARIABLE_ALL_VARIANTS(vector);
      66    DECL_VARIABLE_ALL_VARIANTS(vector_src);
      67  
      68    ARRAY(buffer_src, int, 8, 8);
      69    ARRAY(buffer_src, int, 16, 4);
      70    ARRAY(buffer_src, int, 32, 2);
      71    ARRAY(buffer_src, int, 64, 1);
      72    ARRAY(buffer_src, uint, 8, 8);
      73    ARRAY(buffer_src, uint, 16, 4);
      74    ARRAY(buffer_src, uint, 32, 2);
      75    ARRAY(buffer_src, uint, 64, 1);
      76    ARRAY(buffer_src, poly, 8, 8);
      77    ARRAY(buffer_src, poly, 16, 4);
      78  #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE)
      79    ARRAY(buffer_src, float, 16, 4);
      80  #endif
      81    ARRAY(buffer_src, float, 32, 2);
      82  
      83    ARRAY(buffer_src, int, 8, 16);
      84    ARRAY(buffer_src, int, 16, 8);
      85    ARRAY(buffer_src, int, 32, 4);
      86    ARRAY(buffer_src, int, 64, 2);
      87    ARRAY(buffer_src, uint, 8, 16);
      88    ARRAY(buffer_src, uint, 16, 8);
      89    ARRAY(buffer_src, uint, 32, 4);
      90    ARRAY(buffer_src, uint, 64, 2);
      91    ARRAY(buffer_src, poly, 8, 16);
      92    ARRAY(buffer_src, poly, 16, 8);
      93  #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE)
      94    ARRAY(buffer_src, float, 16, 8);
      95  #endif
      96    ARRAY(buffer_src, float, 32, 4);
      97  
      98    clean_results ();
      99  
     100    /* Choose lane arbitrarily.  */
     101    TEST_VLD1_LANE(, int, s, 8, 8, 6);
     102    TEST_VLD1_LANE(, int, s, 16, 4, 3);
     103    TEST_VLD1_LANE(, int, s, 32, 2, 1);
     104    TEST_VLD1_LANE(, int, s, 64, 1, 0);
     105    TEST_VLD1_LANE(, uint, u, 8, 8, 7);
     106    TEST_VLD1_LANE(, uint, u, 16, 4, 3);
     107    TEST_VLD1_LANE(, uint, u, 32, 2, 1);
     108    TEST_VLD1_LANE(, uint, u, 64, 1, 0);
     109    TEST_VLD1_LANE(, poly, p, 8, 8, 7);
     110    TEST_VLD1_LANE(, poly, p, 16, 4, 3);
     111  #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE)
     112    TEST_VLD1_LANE(, float, f, 16, 4, 2);
     113  #endif
     114    TEST_VLD1_LANE(, float, f, 32, 2, 1);
     115  
     116    TEST_VLD1_LANE(q, int, s, 8, 16, 15);
     117    TEST_VLD1_LANE(q, int, s, 16, 8, 5);
     118    TEST_VLD1_LANE(q, int, s, 32, 4, 2);
     119    TEST_VLD1_LANE(q, int, s, 64, 2, 1);
     120    TEST_VLD1_LANE(q, uint, u, 8, 16, 12);
     121    TEST_VLD1_LANE(q, uint, u, 16, 8, 6);
     122    TEST_VLD1_LANE(q, uint, u, 32, 4, 2);
     123    TEST_VLD1_LANE(q, uint, u, 64, 2, 0);
     124    TEST_VLD1_LANE(q, poly, p, 8, 16, 12);
     125    TEST_VLD1_LANE(q, poly, p, 16, 8, 6);
     126  #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE)
     127    TEST_VLD1_LANE(q, float, f, 16, 8, 5);
     128  #endif
     129    TEST_VLD1_LANE(q, float, f, 32, 4, 2);
     130  
     131    CHECK_RESULTS (TEST_MSG, "");
     132  }
     133  
     134  int main (void)
     135  {
     136    exec_vld1_lane ();
     137    return 0;
     138  }