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,16,8) [] = { 0x100, 0xe1, 0xc4, 0xa9,
       7  					0x90, 0x79, 0x64, 0x51 };
       8  VECT_VAR_DECL(expected,int,32,4) [] = { 0x100, 0xe1, 0xc4, 0xa9 };
       9  VECT_VAR_DECL(expected,int,64,2) [] = { 0x100, 0xe1 };
      10  VECT_VAR_DECL(expected,uint,16,8) [] = { 0xe100, 0xe2e1, 0xe4c4, 0xe6a9,
      11  					 0xe890, 0xea79, 0xec64, 0xee51 };
      12  VECT_VAR_DECL(expected,uint,32,4) [] = { 0xffe00100, 0xffe200e1,
      13  					 0xffe400c4, 0xffe600a9 };
      14  VECT_VAR_DECL(expected,uint,64,2) [] = { 0xffffffe000000100,
      15  					 0xffffffe2000000e1 };
      16  VECT_VAR_DECL(expected,poly,16,8) [] = { 0x5500, 0x5501, 0x5504, 0x5505,
      17  					 0x5510, 0x5511, 0x5514, 0x5515 };
      18  
      19  #define TEST_MSG "VMULL"
      20  void exec_vmull (void)
      21  {
      22    /* Basic test: y=vmull(x,x), then store the result.  */
      23  #define TEST_VMULL(T1, T2, W, W2, N)					\
      24    VECT_VAR(vector_res, T1, W2, N) =					\
      25      vmull_##T2##W(VECT_VAR(vector, T1, W, N),				\
      26  		  VECT_VAR(vector, T1, W, N));				\
      27    vst1q_##T2##W2(VECT_VAR(result, T1, W2, N), VECT_VAR(vector_res, T1, W2, N))
      28  
      29    DECL_VARIABLE(vector, int, 8, 8);
      30    DECL_VARIABLE(vector, int, 16, 4);
      31    DECL_VARIABLE(vector, int, 32, 2);
      32    DECL_VARIABLE(vector, uint, 8, 8);
      33    DECL_VARIABLE(vector, uint, 16, 4);
      34    DECL_VARIABLE(vector, uint, 32, 2);
      35    DECL_VARIABLE(vector, poly, 8, 8);
      36    DECL_VARIABLE(vector_res, int, 16, 8);
      37    DECL_VARIABLE(vector_res, int, 32, 4);
      38    DECL_VARIABLE(vector_res, int, 64, 2);
      39    DECL_VARIABLE(vector_res, uint, 16, 8);
      40    DECL_VARIABLE(vector_res, uint, 32, 4);
      41    DECL_VARIABLE(vector_res, uint, 64, 2);
      42    DECL_VARIABLE(vector_res, poly, 16, 8);
      43  
      44    clean_results ();
      45  
      46    VLOAD(vector, buffer, , int, s, 8, 8);
      47    VLOAD(vector, buffer, , int, s, 16, 4);
      48    VLOAD(vector, buffer, , int, s, 32, 2);
      49    VLOAD(vector, buffer, , uint, u, 8, 8);
      50    VLOAD(vector, buffer, , uint, u, 16, 4);
      51    VLOAD(vector, buffer, , uint, u, 32, 2);
      52    VLOAD(vector, buffer, , poly, p, 8, 8);
      53  
      54    TEST_VMULL(int, s, 8, 16, 8);
      55    TEST_VMULL(int, s, 16, 32, 4);
      56    TEST_VMULL(int, s, 32, 64, 2);
      57    TEST_VMULL(uint, u, 8, 16, 8);
      58    TEST_VMULL(uint, u, 16, 32, 4);
      59    TEST_VMULL(uint, u, 32, 64, 2);
      60    TEST_VMULL(poly, p, 8, 16, 8);
      61  
      62    CHECK(TEST_MSG, int, 16, 8, PRIx16, expected, "");
      63    CHECK(TEST_MSG, int, 32, 4, PRIx32, expected, "");
      64    CHECK(TEST_MSG, int, 64, 2, PRIx64, expected, "");
      65    CHECK(TEST_MSG, uint, 16, 8, PRIx16, expected, "");
      66    CHECK(TEST_MSG, uint, 32, 4, PRIx32, expected, "");
      67    CHECK(TEST_MSG, uint, 64, 2, PRIx64, expected, "");
      68    CHECK_POLY(TEST_MSG, poly, 16, 8, PRIx16, expected, "");
      69  }
      70  
      71  int main (void)
      72  {
      73    exec_vmull ();
      74    return 0;
      75  }