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) [] = { 0xf7, 0x33, 0x33, 0x33,
       7  				       0x33, 0x33, 0x33, 0x33 };
       8  VECT_VAR_DECL(expected,int,16,4) [] = { 0xfff3, 0x3333, 0x3333, 0x3333 };
       9  VECT_VAR_DECL(expected,int,32,2) [] = { 0xfffffff1, 0x33333333 };
      10  VECT_VAR_DECL(expected,int,64,1) [] = { 0xfffffffffffffff0 };
      11  VECT_VAR_DECL(expected,uint,8,8) [] = { 0xf6, 0x33, 0x33, 0x33,
      12  					0x33, 0x33, 0x33, 0x33 };
      13  VECT_VAR_DECL(expected,uint,16,4) [] = { 0xfff2, 0x3333, 0x3333, 0x3333 };
      14  VECT_VAR_DECL(expected,uint,32,2) [] = { 0xfffffff0, 0x33333333 };
      15  VECT_VAR_DECL(expected,uint,64,1) [] = { 0xfffffffffffffff0 };
      16  VECT_VAR_DECL(expected,poly,8,8) [] = { 0xf6, 0x33, 0x33, 0x33,
      17  					0x33, 0x33, 0x33, 0x33 };
      18  VECT_VAR_DECL(expected,poly,16,4) [] = { 0xfff2, 0x3333, 0x3333, 0x3333 };
      19  VECT_VAR_DECL(expected,hfloat,16,4) [] = { 0xcb80, 0x3333, 0x3333, 0x3333 };
      20  VECT_VAR_DECL(expected,hfloat,32,2) [] = { 0xc1700000, 0x33333333 };
      21  VECT_VAR_DECL(expected,int,8,16) [] = { 0xff, 0x33, 0x33, 0x33,
      22  					0x33, 0x33, 0x33, 0x33,
      23  					0x33, 0x33, 0x33, 0x33,
      24  					0x33, 0x33, 0x33, 0x33 };
      25  VECT_VAR_DECL(expected,int,16,8) [] = { 0xfff5, 0x3333, 0x3333, 0x3333,
      26  					0x3333, 0x3333, 0x3333, 0x3333 };
      27  VECT_VAR_DECL(expected,int,32,4) [] = { 0xfffffff1, 0x33333333,
      28  					0x33333333, 0x33333333 };
      29  VECT_VAR_DECL(expected,int,64,2) [] = { 0xfffffffffffffff1, 0x3333333333333333 };
      30  VECT_VAR_DECL(expected,uint,8,16) [] = { 0xfa, 0x33, 0x33, 0x33,
      31  					 0x33, 0x33, 0x33, 0x33,
      32  					 0x33, 0x33, 0x33, 0x33,
      33  					 0x33, 0x33, 0x33, 0x33 };
      34  VECT_VAR_DECL(expected,uint,16,8) [] = { 0xfff4, 0x3333, 0x3333, 0x3333,
      35  					 0x3333, 0x3333, 0x3333, 0x3333 };
      36  VECT_VAR_DECL(expected,uint,32,4) [] = { 0xfffffff3, 0x33333333,
      37  					 0x33333333, 0x33333333 };
      38  VECT_VAR_DECL(expected,uint,64,2) [] = { 0xfffffffffffffff0,
      39  					 0x3333333333333333 };
      40  VECT_VAR_DECL(expected,poly,8,16) [] = { 0xfa, 0x33, 0x33, 0x33,
      41  					 0x33, 0x33, 0x33, 0x33,
      42  					 0x33, 0x33, 0x33, 0x33,
      43  					 0x33, 0x33, 0x33, 0x33 };
      44  VECT_VAR_DECL(expected,poly,16,8) [] = { 0xfff4, 0x3333, 0x3333, 0x3333,
      45  					 0x3333, 0x3333, 0x3333, 0x3333 };
      46  VECT_VAR_DECL(expected,hfloat,16,8) [] = { 0xc900, 0x3333, 0x3333, 0x3333,
      47  					   0x3333, 0x3333, 0x3333, 0x3333 };
      48  VECT_VAR_DECL(expected,hfloat,32,4) [] = { 0xc1700000, 0x33333333,
      49  					   0x33333333, 0x33333333 };
      50  
      51  #define TEST_MSG "VST1_LANE/VST1_LANEQ"
      52  void exec_vst1_lane (void)
      53  {
      54  #define TEST_VST1_LANE(Q, T1, T2, W, N, L)		\
      55    VECT_VAR(vector, T1, W, N) =				\
      56      vld1##Q##_##T2##W(VECT_VAR(buffer, T1, W, N));	\
      57    vst1##Q##_lane_##T2##W(VECT_VAR(result, T1, W, N),	\
      58  			 VECT_VAR(vector, T1, W, N), L)
      59  
      60    DECL_VARIABLE_ALL_VARIANTS(vector);
      61  
      62    clean_results ();
      63  
      64    /* Choose lane arbitrarily.  */
      65    TEST_VST1_LANE(, int, s, 8, 8, 7);
      66    TEST_VST1_LANE(, int, s, 16, 4, 3);
      67    TEST_VST1_LANE(, int, s, 32, 2, 1);
      68    TEST_VST1_LANE(, int, s, 64, 1, 0);
      69    TEST_VST1_LANE(, uint, u, 8, 8, 6);
      70    TEST_VST1_LANE(, uint, u, 16, 4, 2);
      71    TEST_VST1_LANE(, uint, u, 32, 2, 0);
      72    TEST_VST1_LANE(, uint, u, 64, 1, 0);
      73    TEST_VST1_LANE(, poly, p, 8, 8, 6);
      74    TEST_VST1_LANE(, poly, p, 16, 4, 2);
      75  #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE)
      76    TEST_VST1_LANE(, float, f, 16, 4, 1);
      77  #endif
      78    TEST_VST1_LANE(, float, f, 32, 2, 1);
      79  
      80    TEST_VST1_LANE(q, int, s, 8, 16, 15);
      81    TEST_VST1_LANE(q, int, s, 16, 8, 5);
      82    TEST_VST1_LANE(q, int, s, 32, 4, 1);
      83    TEST_VST1_LANE(q, int, s, 64, 2, 1);
      84    TEST_VST1_LANE(q, uint, u, 8, 16, 10);
      85    TEST_VST1_LANE(q, uint, u, 16, 8, 4);
      86    TEST_VST1_LANE(q, uint, u, 32, 4, 3);
      87    TEST_VST1_LANE(q, uint, u, 64, 2, 0);
      88    TEST_VST1_LANE(q, poly, p, 8, 16, 10);
      89    TEST_VST1_LANE(q, poly, p, 16, 8, 4);
      90  #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE)
      91    TEST_VST1_LANE(q, float, f, 16, 8, 6);
      92  #endif
      93    TEST_VST1_LANE(q, float, f, 32, 4, 1);
      94  
      95    CHECK_RESULTS(TEST_MSG, "");
      96  }
      97  
      98  int main (void)
      99  {
     100    exec_vst1_lane ();
     101    return 0;
     102  }