|
QUDA
v0.7.0
A library for QCD on GPUs
|
Go to the source code of this file.
Macros | |
| #define | DSLASH_SHARED_FLOATS_PER_THREAD 24 |
| #define | VOLATILE volatile |
| #define | spinorFloat float |
| #define | WRITE_SPINOR_SHARED WRITE_SPINOR_SHARED_FLOAT4 |
| #define | READ_SPINOR_SHARED READ_SPINOR_SHARED_FLOAT4 |
| #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 | acc00_re accum0.x |
| #define | acc00_im accum0.y |
| #define | acc01_re accum0.z |
| #define | acc01_im accum0.w |
| #define | acc02_re accum1.x |
| #define | acc02_im accum1.y |
| #define | acc10_re accum1.z |
| #define | acc10_im accum1.w |
| #define | acc11_re accum2.x |
| #define | acc11_im accum2.y |
| #define | acc12_re accum2.z |
| #define | acc12_im accum2.w |
| #define | acc20_re accum3.x |
| #define | acc20_im accum3.y |
| #define | acc21_re accum3.z |
| #define | acc21_im accum3.w |
| #define | acc22_re accum4.x |
| #define | acc22_im accum4.y |
| #define | acc30_re accum4.z |
| #define | acc30_im accum4.w |
| #define | acc31_re accum5.x |
| #define | acc31_im accum5.y |
| #define | acc32_re accum5.z |
| #define | acc32_im accum5.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 | 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) |
| #define | SHARED_STRIDE 32 |
| #define acc00_im accum0.y |
Definition at line 93 of file tm_dslash_fermi_core.h.
| #define acc00_re accum0.x |
Definition at line 92 of file tm_dslash_fermi_core.h.
| #define acc01_im accum0.w |
Definition at line 95 of file tm_dslash_fermi_core.h.
| #define acc01_re accum0.z |
Definition at line 94 of file tm_dslash_fermi_core.h.
| #define acc02_im accum1.y |
Definition at line 97 of file tm_dslash_fermi_core.h.
| #define acc02_re accum1.x |
Definition at line 96 of file tm_dslash_fermi_core.h.
| #define acc10_im accum1.w |
Definition at line 99 of file tm_dslash_fermi_core.h.
| #define acc10_re accum1.z |
Definition at line 98 of file tm_dslash_fermi_core.h.
| #define acc11_im accum2.y |
Definition at line 101 of file tm_dslash_fermi_core.h.
| #define acc11_re accum2.x |
Definition at line 100 of file tm_dslash_fermi_core.h.
| #define acc12_im accum2.w |
Definition at line 103 of file tm_dslash_fermi_core.h.
| #define acc12_re accum2.z |
Definition at line 102 of file tm_dslash_fermi_core.h.
| #define acc20_im accum3.y |
Definition at line 105 of file tm_dslash_fermi_core.h.
| #define acc20_re accum3.x |
Definition at line 104 of file tm_dslash_fermi_core.h.
| #define acc21_im accum3.w |
Definition at line 107 of file tm_dslash_fermi_core.h.
| #define acc21_re accum3.z |
Definition at line 106 of file tm_dslash_fermi_core.h.
| #define acc22_im accum4.y |
Definition at line 109 of file tm_dslash_fermi_core.h.
| #define acc22_re accum4.x |
Definition at line 108 of file tm_dslash_fermi_core.h.
| #define acc30_im accum4.w |
Definition at line 111 of file tm_dslash_fermi_core.h.
| #define acc30_re accum4.z |
Definition at line 110 of file tm_dslash_fermi_core.h.
| #define acc31_im accum5.y |
Definition at line 113 of file tm_dslash_fermi_core.h.
| #define acc31_re accum5.x |
Definition at line 112 of file tm_dslash_fermi_core.h.
| #define acc32_im accum5.w |
Definition at line 115 of file tm_dslash_fermi_core.h.
| #define acc32_re accum5.z |
Definition at line 114 of file tm_dslash_fermi_core.h.
| #define DSLASH_SHARED_FLOATS_PER_THREAD 24 |
Definition at line 3 of file tm_dslash_fermi_core.h.
| #define g00_im G0.y |
Definition at line 141 of file tm_dslash_fermi_core.h.
| #define g00_re G0.x |
Definition at line 140 of file tm_dslash_fermi_core.h.
| #define g01_im G0.w |
Definition at line 143 of file tm_dslash_fermi_core.h.
| #define g01_re G0.z |
Definition at line 142 of file tm_dslash_fermi_core.h.
| #define g02_im G1.y |
Definition at line 145 of file tm_dslash_fermi_core.h.
| #define g02_re G1.x |
Definition at line 144 of file tm_dslash_fermi_core.h.
| #define g10_im G1.w |
Definition at line 147 of file tm_dslash_fermi_core.h.
| #define g10_re G1.z |
Definition at line 146 of file tm_dslash_fermi_core.h.
| #define g11_im G2.y |
Definition at line 149 of file tm_dslash_fermi_core.h.
| #define g11_re G2.x |
Definition at line 148 of file tm_dslash_fermi_core.h.
| #define g12_im G2.w |
Definition at line 151 of file tm_dslash_fermi_core.h.
| #define g12_re G2.z |
Definition at line 150 of file tm_dslash_fermi_core.h.
| #define g20_im G3.y |
Definition at line 153 of file tm_dslash_fermi_core.h.
| #define g20_re G3.x |
Definition at line 152 of file tm_dslash_fermi_core.h.
| #define g21_im G3.w |
Definition at line 155 of file tm_dslash_fermi_core.h.
| #define g21_re G3.z |
Definition at line 154 of file tm_dslash_fermi_core.h.
| #define g22_im G4.y |
Definition at line 157 of file tm_dslash_fermi_core.h.
| #define g22_re G4.x |
Definition at line 156 of file tm_dslash_fermi_core.h.
| #define gT00_im (-g00_im) |
Definition at line 163 of file tm_dslash_fermi_core.h.
| #define gT00_re (+g00_re) |
Definition at line 162 of file tm_dslash_fermi_core.h.
| #define gT01_im (-g10_im) |
Definition at line 165 of file tm_dslash_fermi_core.h.
| #define gT01_re (+g10_re) |
Definition at line 164 of file tm_dslash_fermi_core.h.
| #define gT02_im (-g20_im) |
Definition at line 167 of file tm_dslash_fermi_core.h.
| #define gT02_re (+g20_re) |
Definition at line 166 of file tm_dslash_fermi_core.h.
| #define gT10_im (-g01_im) |
Definition at line 169 of file tm_dslash_fermi_core.h.
| #define gT10_re (+g01_re) |
Definition at line 168 of file tm_dslash_fermi_core.h.
| #define gT11_im (-g11_im) |
Definition at line 171 of file tm_dslash_fermi_core.h.
| #define gT11_re (+g11_re) |
Definition at line 170 of file tm_dslash_fermi_core.h.
| #define gT12_im (-g21_im) |
Definition at line 173 of file tm_dslash_fermi_core.h.
| #define gT12_re (+g21_re) |
Definition at line 172 of file tm_dslash_fermi_core.h.
| #define gT20_im (-g02_im) |
Definition at line 175 of file tm_dslash_fermi_core.h.
| #define gT20_re (+g02_re) |
Definition at line 174 of file tm_dslash_fermi_core.h.
| #define gT21_im (-g12_im) |
Definition at line 177 of file tm_dslash_fermi_core.h.
| #define gT21_re (+g12_re) |
Definition at line 176 of file tm_dslash_fermi_core.h.
| #define gT22_im (-g22_im) |
Definition at line 179 of file tm_dslash_fermi_core.h.
| #define gT22_re (+g22_re) |
Definition at line 178 of file tm_dslash_fermi_core.h.
| #define i00_im I0.y |
Definition at line 69 of file tm_dslash_fermi_core.h.
| #define i00_re I0.x |
Definition at line 68 of file tm_dslash_fermi_core.h.
| #define i01_im I0.w |
Definition at line 71 of file tm_dslash_fermi_core.h.
| #define i01_re I0.z |
Definition at line 70 of file tm_dslash_fermi_core.h.
| #define i02_im I1.y |
Definition at line 73 of file tm_dslash_fermi_core.h.
| #define i02_re I1.x |
Definition at line 72 of file tm_dslash_fermi_core.h.
| #define i10_im I1.w |
Definition at line 75 of file tm_dslash_fermi_core.h.
| #define i10_re I1.z |
Definition at line 74 of file tm_dslash_fermi_core.h.
| #define i11_im I2.y |
Definition at line 77 of file tm_dslash_fermi_core.h.
| #define i11_re I2.x |
Definition at line 76 of file tm_dslash_fermi_core.h.
| #define i12_im I2.w |
Definition at line 79 of file tm_dslash_fermi_core.h.
| #define i12_re I2.z |
Definition at line 78 of file tm_dslash_fermi_core.h.
| #define i20_im I3.y |
Definition at line 81 of file tm_dslash_fermi_core.h.
| #define i20_re I3.x |
Definition at line 80 of file tm_dslash_fermi_core.h.
| #define i21_im I3.w |
Definition at line 83 of file tm_dslash_fermi_core.h.
| #define i21_re I3.z |
Definition at line 82 of file tm_dslash_fermi_core.h.
| #define i22_im I4.y |
Definition at line 85 of file tm_dslash_fermi_core.h.
| #define i22_re I4.x |
Definition at line 84 of file tm_dslash_fermi_core.h.
| #define i30_im I4.w |
Definition at line 87 of file tm_dslash_fermi_core.h.
| #define i30_re I4.z |
Definition at line 86 of file tm_dslash_fermi_core.h.
| #define i31_im I5.y |
Definition at line 89 of file tm_dslash_fermi_core.h.
| #define i31_re I5.x |
Definition at line 88 of file tm_dslash_fermi_core.h.
| #define i32_im I5.w |
Definition at line 91 of file tm_dslash_fermi_core.h.
| #define i32_re I5.z |
Definition at line 90 of file tm_dslash_fermi_core.h.
| #define READ_SPINOR_SHARED READ_SPINOR_SHARED_FLOAT4 |
Definition at line 67 of file tm_dslash_fermi_core.h.
| #define SHARED_STRIDE 32 |
Definition at line 210 of file tm_dslash_fermi_core.h.
| #define spinorFloat float |
Definition at line 65 of file tm_dslash_fermi_core.h.
| #define VOLATILE volatile |
Definition at line 9 of file tm_dslash_fermi_core.h.
| #define WRITE_SPINOR_SHARED WRITE_SPINOR_SHARED_FLOAT4 |
Definition at line 66 of file tm_dslash_fermi_core.h.
| __syncthreads | ( | ) |
| if | ( | x2 >= | X2 | ) |
| if | ( | x3 >= | X3 | ) |
| if | ( | ) |
Definition at line 716 of file tm_dslash_fermi_core.h.
| READ_SPINOR_SHARED | ( | tx | , |
| threadIdx. | y, | ||
| threadIdx. | z | ||
| ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 0 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 1 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 2 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 3 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 4 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 5 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 6 | ) |
| RECONSTRUCT_GAUGE_MATRIX | ( | 7 | ) |
| WRITE_SPINOR | ( | param. | sp_stride | ) |
| WRITE_SPINOR_SHARED | ( | threadIdx. | x, |
| threadIdx. | y, | ||
| threadIdx. | z, | ||
| i | |||
| ) |
Definition at line 311 of file tm_dslash_fermi_core.h.
| A0_im = 0 |
Definition at line 377 of file tm_dslash_fermi_core.h.
Definition at line 311 of file tm_dslash_fermi_core.h.
| A0_re = 0 |
Definition at line 370 of file tm_dslash_fermi_core.h.
Definition at line 312 of file tm_dslash_fermi_core.h.
| A1_im = 0 |
Definition at line 407 of file tm_dslash_fermi_core.h.
Definition at line 312 of file tm_dslash_fermi_core.h.
| A1_re = 0 |
Definition at line 400 of file tm_dslash_fermi_core.h.
Definition at line 313 of file tm_dslash_fermi_core.h.
| A2_im = 0 |
Definition at line 437 of file tm_dslash_fermi_core.h.
Definition at line 313 of file tm_dslash_fermi_core.h.
| A2_re = 0 |
Definition at line 430 of file tm_dslash_fermi_core.h.
Definition at line 314 of file tm_dslash_fermi_core.h.
| B0_im = 0 |
Definition at line 391 of file tm_dslash_fermi_core.h.
Definition at line 314 of file tm_dslash_fermi_core.h.
| B0_re = 0 |
Definition at line 384 of file tm_dslash_fermi_core.h.
Definition at line 315 of file tm_dslash_fermi_core.h.
| B1_im = 0 |
Definition at line 421 of file tm_dslash_fermi_core.h.
Definition at line 315 of file tm_dslash_fermi_core.h.
| B1_re = 0 |
Definition at line 414 of file tm_dslash_fermi_core.h.
Definition at line 316 of file tm_dslash_fermi_core.h.
| B2_im = 0 |
Definition at line 451 of file tm_dslash_fermi_core.h.
Definition at line 316 of file tm_dslash_fermi_core.h.
| B2_re = 0 |
Definition at line 444 of file tm_dslash_fermi_core.h.
| else |
Definition at line 736 of file tm_dslash_fermi_core.h.
| const int ga_idx = sid |
Definition at line 309 of file tm_dslash_fermi_core.h.
| o00_im = 0 |
Definition at line 183 of file tm_dslash_fermi_core.h.
| o00_re = 0 |
Definition at line 182 of file tm_dslash_fermi_core.h.
| o01_im = 0 |
Definition at line 185 of file tm_dslash_fermi_core.h.
| o01_re = 0 |
Definition at line 184 of file tm_dslash_fermi_core.h.
| o02_im = 0 |
Definition at line 187 of file tm_dslash_fermi_core.h.
| o02_re = 0 |
Definition at line 186 of file tm_dslash_fermi_core.h.
| o10_im = 0 |
Definition at line 189 of file tm_dslash_fermi_core.h.
| o10_re = 0 |
Definition at line 188 of file tm_dslash_fermi_core.h.
| o11_im = 0 |
Definition at line 191 of file tm_dslash_fermi_core.h.
| o11_re = 0 |
Definition at line 190 of file tm_dslash_fermi_core.h.
| o12_im = 0 |
Definition at line 193 of file tm_dslash_fermi_core.h.
| o12_re = 0 |
Definition at line 192 of file tm_dslash_fermi_core.h.
| o20_im = 0 |
Definition at line 195 of file tm_dslash_fermi_core.h.
| o20_re = 0 |
Definition at line 194 of file tm_dslash_fermi_core.h.
| o21_im = 0 |
Definition at line 197 of file tm_dslash_fermi_core.h.
| o21_re = 0 |
Definition at line 196 of file tm_dslash_fermi_core.h.
| o22_im = 0 |
Definition at line 199 of file tm_dslash_fermi_core.h.
| o22_re = 0 |
Definition at line 198 of file tm_dslash_fermi_core.h.
| o30_im = 0 |
Definition at line 201 of file tm_dslash_fermi_core.h.
| o30_re = 0 |
Definition at line 200 of file tm_dslash_fermi_core.h.
| o31_im = 0 |
Definition at line 203 of file tm_dslash_fermi_core.h.
| o31_re = 0 |
Definition at line 202 of file tm_dslash_fermi_core.h.
| o32_im = 0 |
Definition at line 205 of file tm_dslash_fermi_core.h.
| o32_re = 0 |
Definition at line 204 of file tm_dslash_fermi_core.h.
| int sid |
Definition at line 223 of file tm_dslash_fermi_core.h.
| int tx = (threadIdx.x > 0) ? threadIdx.x-1 : blockDim.x-1 |
Definition at line 524 of file tm_dslash_fermi_core.h.
| int ty = (threadIdx.y < blockDim.y - 1) ? threadIdx.y + 1 : 0 |
Definition at line 739 of file tm_dslash_fermi_core.h.
| int tz = (threadIdx.z < blockDim.z - 1) ? threadIdx.z + 1 : 0 |
Definition at line 1173 of file tm_dslash_fermi_core.h.
| int X |
Definition at line 217 of file tm_dslash_fermi_core.h.
| int x1 |
Definition at line 216 of file tm_dslash_fermi_core.h.
| int x2 |
Definition at line 216 of file tm_dslash_fermi_core.h.
| int x3 |
Definition at line 216 of file tm_dslash_fermi_core.h.
| int x4 |
Definition at line 216 of file tm_dslash_fermi_core.h.
1.8.6