|
QUDA v0.4.0
A library for QCD on GPUs
|
Go to the source code of this file.
Defines | |
| #define | SHARED_BYTES_DOUBLE (BLOCK_DIM*SHARED_FLOATS_PER_THREAD*sizeof(double)) |
| #define | SHARED_BYTES_SINGLE (BLOCK_DIM*SHARED_FLOATS_PER_THREAD*sizeof(float)) |
| #define | spinorFloat float |
| #define | i00_re I0.x |
| #define | i00_im I0.y |
| #define | i01_re I0.z |
| #define | i01_im I0.w |
| #define | i02_re I1.x |
| #define | i02_im I1.y |
| #define | i10_re I1.z |
| #define | i10_im I1.w |
| #define | i11_re I2.x |
| #define | i11_im I2.y |
| #define | i12_re I2.z |
| #define | i12_im I2.w |
| #define | i20_re I3.x |
| #define | i20_im I3.y |
| #define | i21_re I3.z |
| #define | i21_im I3.w |
| #define | i22_re I4.x |
| #define | i22_im I4.y |
| #define | i30_re I4.z |
| #define | i30_im I4.w |
| #define | i31_re I5.x |
| #define | i31_im I5.y |
| #define | i32_re I5.z |
| #define | i32_im I5.w |
| #define | g00_re G0.x |
| #define | g00_im G0.y |
| #define | g01_re G0.z |
| #define | g01_im G0.w |
| #define | g02_re G1.x |
| #define | g02_im G1.y |
| #define | g10_re G1.z |
| #define | g10_im G1.w |
| #define | g11_re G2.x |
| #define | g11_im G2.y |
| #define | g12_re G2.z |
| #define | g12_im G2.w |
| #define | g20_re G3.x |
| #define | g20_im G3.y |
| #define | g21_re G3.z |
| #define | g21_im G3.w |
| #define | g22_re G4.x |
| #define | g22_im G4.y |
| #define | A_re G4.z |
| #define | A_im G4.w |
| #define | gT00_re (+g00_re) |
| #define | gT00_im (-g00_im) |
| #define | gT01_re (+g10_re) |
| #define | gT01_im (-g10_im) |
| #define | gT02_re (+g20_re) |
| #define | gT02_im (-g20_im) |
| #define | gT10_re (+g01_re) |
| #define | gT10_im (-g01_im) |
| #define | gT11_re (+g11_re) |
| #define | gT11_im (-g11_im) |
| #define | gT12_re (+g21_re) |
| #define | gT12_im (-g21_im) |
| #define | gT20_re (+g02_re) |
| #define | gT20_im (-g02_im) |
| #define | gT21_re (+g12_re) |
| #define | gT21_im (-g12_im) |
| #define | gT22_re (+g22_re) |
| #define | gT22_im (-g22_im) |
Functions | |
| if (sid >=param.threads) return | |
Variables | |
| volatile spinorFloat | o00_re = o00_im = 0 |
| volatile spinorFloat | o00_im |
| volatile spinorFloat | o01_re = o01_im = 0 |
| volatile spinorFloat | o01_im |
| volatile spinorFloat | o02_re = o02_im = 0 |
| volatile spinorFloat | o02_im |
| volatile spinorFloat | o10_re = o10_im = 0 |
| volatile spinorFloat | o10_im |
| volatile spinorFloat | o11_re = o11_im = 0 |
| volatile spinorFloat | o11_im |
| volatile spinorFloat | o12_re = o12_im = 0 |
| volatile spinorFloat | o12_im |
| volatile spinorFloat | o20_re = o20_im = 0 |
| volatile spinorFloat | o20_im |
| volatile spinorFloat | o21_re = o21_im = 0 |
| volatile spinorFloat | o21_im |
| volatile spinorFloat | o22_re = o22_im = 0 |
| volatile spinorFloat | o22_im |
| volatile spinorFloat | o30_re = o30_im = 0 |
| volatile spinorFloat | o30_im |
| volatile spinorFloat | o31_re = o31_im = 0 |
| volatile spinorFloat | o31_im |
| volatile spinorFloat | o32_re = o32_im = 0 |
| volatile spinorFloat | o32_im |
| int | sid = blockIdx.x*blockDim.x + threadIdx.x |
| int | boundaryCrossings = sid/X1h + sid/(X2*X1h) + sid/(X3*X2*X1h) + sid/(X4*X3*X2*X1h) |
| int | boundaryCrossings4d = sid/X1h + sid/(X2*X1h) + sid/(X3*X2*X1h) |
| int | X = 2*sid + (boundaryCrossings + param.parity) % 2 |
| int | xs = X/(X4*X3*X2*X1) |
| int | x4 = (X/(X3*X2*X1)) % X4 |
| int | x3 = (X/(X2*X1)) % X3 |
| int | x2 = (X/X1) % X2 |
| int | x1 = X % X1 |
| #define A_im G4.w |
Definition at line 133 of file dw_dslash_core.h.
| #define A_re G4.z |
Definition at line 132 of file dw_dslash_core.h.
| #define g00_im G0.y |
Definition at line 110 of file dw_dslash_core.h.
| #define g00_re G0.x |
Definition at line 109 of file dw_dslash_core.h.
| #define g01_im G0.w |
Definition at line 112 of file dw_dslash_core.h.
| #define g01_re G0.z |
Definition at line 111 of file dw_dslash_core.h.
| #define g02_im G1.y |
Definition at line 114 of file dw_dslash_core.h.
| #define g02_re G1.x |
Definition at line 113 of file dw_dslash_core.h.
| #define g10_im G1.w |
Definition at line 117 of file dw_dslash_core.h.
| #define g10_re G1.z |
Definition at line 116 of file dw_dslash_core.h.
| #define g11_im G2.y |
Definition at line 119 of file dw_dslash_core.h.
| #define g11_re G2.x |
Definition at line 118 of file dw_dslash_core.h.
| #define g12_im G2.w |
Definition at line 121 of file dw_dslash_core.h.
| #define g12_re G2.z |
Definition at line 120 of file dw_dslash_core.h.
| #define g20_im G3.y |
Definition at line 125 of file dw_dslash_core.h.
| #define g20_re G3.x |
Definition at line 124 of file dw_dslash_core.h.
| #define g21_im G3.w |
Definition at line 127 of file dw_dslash_core.h.
| #define g21_re G3.z |
Definition at line 126 of file dw_dslash_core.h.
| #define g22_im G4.y |
Definition at line 129 of file dw_dslash_core.h.
| #define g22_re G4.x |
Definition at line 128 of file dw_dslash_core.h.
| #define gT00_im (-g00_im) |
Definition at line 139 of file dw_dslash_core.h.
| #define gT00_re (+g00_re) |
Definition at line 138 of file dw_dslash_core.h.
| #define gT01_im (-g10_im) |
Definition at line 141 of file dw_dslash_core.h.
| #define gT01_re (+g10_re) |
Definition at line 140 of file dw_dslash_core.h.
| #define gT02_im (-g20_im) |
Definition at line 143 of file dw_dslash_core.h.
| #define gT02_re (+g20_re) |
Definition at line 142 of file dw_dslash_core.h.
| #define gT10_im (-g01_im) |
Definition at line 145 of file dw_dslash_core.h.
| #define gT10_re (+g01_re) |
Definition at line 144 of file dw_dslash_core.h.
| #define gT11_im (-g11_im) |
Definition at line 147 of file dw_dslash_core.h.
| #define gT11_re (+g11_re) |
Definition at line 146 of file dw_dslash_core.h.
| #define gT12_im (-g21_im) |
Definition at line 149 of file dw_dslash_core.h.
| #define gT12_re (+g21_re) |
Definition at line 148 of file dw_dslash_core.h.
| #define gT20_im (-g02_im) |
Definition at line 151 of file dw_dslash_core.h.
| #define gT20_re (+g02_re) |
Definition at line 150 of file dw_dslash_core.h.
| #define gT21_im (-g12_im) |
Definition at line 153 of file dw_dslash_core.h.
| #define gT21_re (+g12_re) |
Definition at line 152 of file dw_dslash_core.h.
| #define gT22_im (-g22_im) |
Definition at line 155 of file dw_dslash_core.h.
| #define gT22_re (+g22_re) |
Definition at line 154 of file dw_dslash_core.h.
| #define i00_im I0.y |
Definition at line 58 of file dw_dslash_core.h.
| #define i00_re I0.x |
Definition at line 57 of file dw_dslash_core.h.
| #define i01_im I0.w |
Definition at line 60 of file dw_dslash_core.h.
| #define i01_re I0.z |
Definition at line 59 of file dw_dslash_core.h.
| #define i02_im I1.y |
Definition at line 62 of file dw_dslash_core.h.
| #define i02_re I1.x |
Definition at line 61 of file dw_dslash_core.h.
| #define i10_im I1.w |
Definition at line 64 of file dw_dslash_core.h.
| #define i10_re I1.z |
Definition at line 63 of file dw_dslash_core.h.
| #define i11_im I2.y |
Definition at line 66 of file dw_dslash_core.h.
| #define i11_re I2.x |
Definition at line 65 of file dw_dslash_core.h.
| #define i12_im I2.w |
Definition at line 68 of file dw_dslash_core.h.
| #define i12_re I2.z |
Definition at line 67 of file dw_dslash_core.h.
| #define i20_im I3.y |
Definition at line 70 of file dw_dslash_core.h.
| #define i20_re I3.x |
Definition at line 69 of file dw_dslash_core.h.
| #define i21_im I3.w |
Definition at line 72 of file dw_dslash_core.h.
| #define i21_re I3.z |
Definition at line 71 of file dw_dslash_core.h.
| #define i22_im I4.y |
Definition at line 74 of file dw_dslash_core.h.
| #define i22_re I4.x |
Definition at line 73 of file dw_dslash_core.h.
| #define i30_im I4.w |
Definition at line 76 of file dw_dslash_core.h.
| #define i30_re I4.z |
Definition at line 75 of file dw_dslash_core.h.
| #define i31_im I5.y |
Definition at line 78 of file dw_dslash_core.h.
| #define i31_re I5.x |
Definition at line 77 of file dw_dslash_core.h.
| #define i32_im I5.w |
Definition at line 80 of file dw_dslash_core.h.
| #define i32_re I5.z |
Definition at line 79 of file dw_dslash_core.h.
| #define SHARED_BYTES_DOUBLE (BLOCK_DIM*SHARED_FLOATS_PER_THREAD*sizeof(double)) |
Definition at line 16 of file dw_dslash_core.h.
| #define SHARED_BYTES_SINGLE (BLOCK_DIM*SHARED_FLOATS_PER_THREAD*sizeof(float)) |
Definition at line 18 of file dw_dslash_core.h.
| #define spinorFloat float |
Definition at line 56 of file dw_dslash_core.h.
| if | ( | sid >=param. | threads | ) |
Definition at line 216 of file dw_dslash_core.h.
Definition at line 219 of file dw_dslash_core.h.
| volatile spinorFloat o00_im |
Definition at line 164 of file dw_dslash_core.h.
Definition at line 163 of file dw_dslash_core.h.
| volatile spinorFloat o01_im |
Definition at line 166 of file dw_dslash_core.h.
Definition at line 165 of file dw_dslash_core.h.
| volatile spinorFloat o02_im |
Definition at line 168 of file dw_dslash_core.h.
Definition at line 167 of file dw_dslash_core.h.
| volatile spinorFloat o10_im |
Definition at line 170 of file dw_dslash_core.h.
Definition at line 169 of file dw_dslash_core.h.
| volatile spinorFloat o11_im |
Definition at line 172 of file dw_dslash_core.h.
Definition at line 171 of file dw_dslash_core.h.
| volatile spinorFloat o12_im |
Definition at line 174 of file dw_dslash_core.h.
Definition at line 173 of file dw_dslash_core.h.
| volatile spinorFloat o20_im |
Definition at line 176 of file dw_dslash_core.h.
Definition at line 175 of file dw_dslash_core.h.
| volatile spinorFloat o21_im |
Definition at line 178 of file dw_dslash_core.h.
Definition at line 177 of file dw_dslash_core.h.
| volatile spinorFloat o22_im |
Definition at line 180 of file dw_dslash_core.h.
Definition at line 179 of file dw_dslash_core.h.
| volatile spinorFloat o30_im |
Definition at line 182 of file dw_dslash_core.h.
Definition at line 181 of file dw_dslash_core.h.
| volatile spinorFloat o31_im |
Definition at line 184 of file dw_dslash_core.h.
Definition at line 183 of file dw_dslash_core.h.
| volatile spinorFloat o32_im |
Definition at line 186 of file dw_dslash_core.h.
Definition at line 185 of file dw_dslash_core.h.
| int sid = blockIdx.x*blockDim.x + threadIdx.x |
Definition at line 203 of file dw_dslash_core.h.
| int X = 2*sid + (boundaryCrossings + param.parity) % 2 |
Definition at line 224 of file dw_dslash_core.h.
Definition at line 231 of file dw_dslash_core.h.
1.7.4