Ignore:
Timestamp:
Mar 13, 2017, 10:38:00 PM (2 years ago)
Author:
cameron
Message:

Adam's changes to add swizzling option to DeleteByPEXT kernel

File:
1 edited

Legend:

Unmodified
Added
Removed
  • icGREP/icgrep-devel/icgrep/kernels/alignedprint.cpp

    r5356 r5362  
    1212namespace kernel {
    1313
    14     void p2s_step(IDISA::IDISA_Builder * iBuilder, Value * p0, Value * p1, Value * hi_mask, unsigned shift, Value * &s1, Value * &s0) {
     14    void ap_p2s_step(IDISA::IDISA_Builder * iBuilder, Value * p0, Value * p1, Value * hi_mask, unsigned shift, Value * &s1, Value * &s0) {
    1515    Value * t0 = iBuilder->simd_if(1, hi_mask, p0, iBuilder->simd_srli(16, p1, shift));
    1616    Value * t1 = iBuilder->simd_if(1, hi_mask, iBuilder->simd_slli(16, p0, shift), p1);
     
    2424    Value * bit11115555[2];
    2525    Value * bit33337777[2];
    26     p2s_step(iBuilder, p[0], p[4], iBuilder->simd_himask(8), 4, bit00004444[1], bit00004444[0]);
    27     p2s_step(iBuilder, p[1], p[5], iBuilder->simd_himask(8), 4, bit11115555[1], bit11115555[0]);
    28     p2s_step(iBuilder, p[2], p[6], iBuilder->simd_himask(8), 4, bit22226666[1], bit22226666[0]);
    29     p2s_step(iBuilder, p[3], p[7], iBuilder->simd_himask(8), 4, bit33337777[1], bit33337777[0]);
     26    ap_p2s_step(iBuilder, p[0], p[4], iBuilder->simd_himask(8), 4, bit00004444[1], bit00004444[0]);
     27    ap_p2s_step(iBuilder, p[1], p[5], iBuilder->simd_himask(8), 4, bit11115555[1], bit11115555[0]);
     28    ap_p2s_step(iBuilder, p[2], p[6], iBuilder->simd_himask(8), 4, bit22226666[1], bit22226666[0]);
     29    ap_p2s_step(iBuilder, p[3], p[7], iBuilder->simd_himask(8), 4, bit33337777[1], bit33337777[0]);
    3030    Value * bit00224466[4];
    3131    Value * bit11335577[4];
    3232    for (unsigned j = 0; j<2; j++) {
    33         p2s_step(iBuilder, bit00004444[j], bit22226666[j],iBuilder->simd_himask(4), 2, bit00224466[2*j+1], bit00224466[2*j]);
    34         p2s_step(iBuilder, bit11115555[j], bit33337777[j],iBuilder->simd_himask(4), 2, bit11335577[2*j+1], bit11335577[2*j]);
     33        ap_p2s_step(iBuilder, bit00004444[j], bit22226666[j],iBuilder->simd_himask(4), 2, bit00224466[2*j+1], bit00224466[2*j]);
     34        ap_p2s_step(iBuilder, bit11115555[j], bit33337777[j],iBuilder->simd_himask(4), 2, bit11335577[2*j+1], bit11335577[2*j]);
    3535    }
    3636    for (unsigned j = 0; j<4; j++) {
    37         p2s_step(iBuilder, bit00224466[j], bit11335577[j], iBuilder->simd_himask(2), 1, s[2*j+1], s[2*j]);
     37        ap_p2s_step(iBuilder, bit00224466[j], bit11335577[j], iBuilder->simd_himask(2), 1, s[2*j+1], s[2*j]);
    3838    }
    3939}
Note: See TracChangeset for help on using the changeset viewer.