|
QUDA v0.4.0
A library for QCD on GPUs
|
|
QUDA v0.4.0
A library for QCD on GPUs
|
Go to the source code of this file.
Defines | |
| #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 | 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) |
| #define | SHARED_STRIDE 32 |
Functions | |
| if (aux >=X1 *X4) return | |
| if (x2 >=X2) return | |
| if (x3 >=X3) return | |
Variables | |
| VOLATILE spinorFloat | o00_re = 0 |
| VOLATILE spinorFloat | o00_im = 0 |
| VOLATILE spinorFloat | o01_re = 0 |
| VOLATILE spinorFloat | o01_im = 0 |
| VOLATILE spinorFloat | o02_re = 0 |
| VOLATILE spinorFloat | o02_im = 0 |
| VOLATILE spinorFloat | o10_re = 0 |
| VOLATILE spinorFloat | o10_im = 0 |
| VOLATILE spinorFloat | o11_re = 0 |
| VOLATILE spinorFloat | o11_im = 0 |
| VOLATILE spinorFloat | o12_re = 0 |
| VOLATILE spinorFloat | o12_im = 0 |
| VOLATILE spinorFloat | o20_re = 0 |
| VOLATILE spinorFloat | o20_im = 0 |
| VOLATILE spinorFloat | o21_re = 0 |
| VOLATILE spinorFloat | o21_im = 0 |
| VOLATILE spinorFloat | o22_re = 0 |
| VOLATILE spinorFloat | o22_im = 0 |
| VOLATILE spinorFloat | o30_re = 0 |
| VOLATILE spinorFloat | o30_im = 0 |
| VOLATILE spinorFloat | o31_re = 0 |
| VOLATILE spinorFloat | o31_im = 0 |
| VOLATILE spinorFloat | o32_re = 0 |
| VOLATILE spinorFloat | o32_im = 0 |
| int | x1 = aux - x4*X1 |
| int | x2 = blockIdx.y*blockDim.y + threadIdx.y |
| int | x3 = blockIdx.z*blockDim.z + threadIdx.z |
| int | x4 = aux / X1 |
| int | X = ((x4*X3 + x3)*X2 + x2)*X1 + x1 |
| int | sid = X >> 1 |
| int | xt = blockIdx.x*blockDim.x + threadIdx.x |
| int | aux = xt+xt |
| #define A_im G4.w |
Definition at line 115 of file tm_dslash_fermi_core.h.
| #define A_re G4.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 96 of file tm_dslash_fermi_core.h.
| #define g00_re G0.x |
Definition at line 95 of file tm_dslash_fermi_core.h.
| #define g01_im G0.w |
Definition at line 98 of file tm_dslash_fermi_core.h.
| #define g01_re G0.z |
Definition at line 97 of file tm_dslash_fermi_core.h.
| #define g02_im G1.y |
Definition at line 100 of file tm_dslash_fermi_core.h.
| #define g02_re G1.x |
Definition at line 99 of file tm_dslash_fermi_core.h.
| #define g10_im G1.w |
Definition at line 102 of file tm_dslash_fermi_core.h.
| #define g10_re G1.z |
Definition at line 101 of file tm_dslash_fermi_core.h.
| #define g11_im G2.y |
Definition at line 104 of file tm_dslash_fermi_core.h.
| #define g11_re G2.x |
Definition at line 103 of file tm_dslash_fermi_core.h.
| #define g12_im G2.w |
Definition at line 106 of file tm_dslash_fermi_core.h.
| #define g12_re G2.z |
Definition at line 105 of file tm_dslash_fermi_core.h.
| #define g20_im G3.y |
Definition at line 108 of file tm_dslash_fermi_core.h.
| #define g20_re G3.x |
Definition at line 107 of file tm_dslash_fermi_core.h.
| #define g21_im G3.w |
Definition at line 110 of file tm_dslash_fermi_core.h.
| #define g21_re G3.z |
Definition at line 109 of file tm_dslash_fermi_core.h.
| #define g22_im G4.y |
Definition at line 112 of file tm_dslash_fermi_core.h.
| #define g22_re G4.x |
Definition at line 111 of file tm_dslash_fermi_core.h.
| #define gT00_im (-g00_im) |
Definition at line 121 of file tm_dslash_fermi_core.h.
| #define gT00_re (+g00_re) |
Definition at line 120 of file tm_dslash_fermi_core.h.
| #define gT01_im (-g10_im) |
Definition at line 123 of file tm_dslash_fermi_core.h.
| #define gT01_re (+g10_re) |
Definition at line 122 of file tm_dslash_fermi_core.h.
| #define gT02_im (-g20_im) |
Definition at line 125 of file tm_dslash_fermi_core.h.
| #define gT02_re (+g20_re) |
Definition at line 124 of file tm_dslash_fermi_core.h.
| #define gT10_im (-g01_im) |
Definition at line 127 of file tm_dslash_fermi_core.h.
| #define gT10_re (+g01_re) |
Definition at line 126 of file tm_dslash_fermi_core.h.
| #define gT11_im (-g11_im) |
Definition at line 129 of file tm_dslash_fermi_core.h.
| #define gT11_re (+g11_re) |
Definition at line 128 of file tm_dslash_fermi_core.h.
| #define gT12_im (-g21_im) |
Definition at line 131 of file tm_dslash_fermi_core.h.
| #define gT12_re (+g21_re) |
Definition at line 130 of file tm_dslash_fermi_core.h.
| #define gT20_im (-g02_im) |
Definition at line 133 of file tm_dslash_fermi_core.h.
| #define gT20_re (+g02_re) |
Definition at line 132 of file tm_dslash_fermi_core.h.
| #define gT21_im (-g12_im) |
Definition at line 135 of file tm_dslash_fermi_core.h.
| #define gT21_re (+g12_re) |
Definition at line 134 of file tm_dslash_fermi_core.h.
| #define gT22_im (-g22_im) |
Definition at line 137 of file tm_dslash_fermi_core.h.
| #define gT22_re (+g22_re) |
Definition at line 136 of file tm_dslash_fermi_core.h.
| #define i00_im I0.y |
Definition at line 45 of file tm_dslash_fermi_core.h.
| #define i00_re I0.x |
Definition at line 44 of file tm_dslash_fermi_core.h.
| #define i01_im I0.w |
Definition at line 47 of file tm_dslash_fermi_core.h.
| #define i01_re I0.z |
Definition at line 46 of file tm_dslash_fermi_core.h.
| #define i02_im I1.y |
Definition at line 49 of file tm_dslash_fermi_core.h.
| #define i02_re I1.x |
Definition at line 48 of file tm_dslash_fermi_core.h.
| #define i10_im I1.w |
Definition at line 51 of file tm_dslash_fermi_core.h.
| #define i10_re I1.z |
Definition at line 50 of file tm_dslash_fermi_core.h.
| #define i11_im I2.y |
Definition at line 53 of file tm_dslash_fermi_core.h.
| #define i11_re I2.x |
Definition at line 52 of file tm_dslash_fermi_core.h.
| #define i12_im I2.w |
Definition at line 55 of file tm_dslash_fermi_core.h.
| #define i12_re I2.z |
Definition at line 54 of file tm_dslash_fermi_core.h.
| #define i20_im I3.y |
Definition at line 57 of file tm_dslash_fermi_core.h.
| #define i20_re I3.x |
Definition at line 56 of file tm_dslash_fermi_core.h.
| #define i21_im I3.w |
Definition at line 59 of file tm_dslash_fermi_core.h.
| #define i21_re I3.z |
Definition at line 58 of file tm_dslash_fermi_core.h.
| #define i22_im I4.y |
Definition at line 61 of file tm_dslash_fermi_core.h.
| #define i22_re I4.x |
Definition at line 60 of file tm_dslash_fermi_core.h.
| #define i30_im I4.w |
Definition at line 63 of file tm_dslash_fermi_core.h.
| #define i30_re I4.z |
Definition at line 62 of file tm_dslash_fermi_core.h.
| #define i31_im I5.y |
Definition at line 65 of file tm_dslash_fermi_core.h.
| #define i31_re I5.x |
Definition at line 64 of file tm_dslash_fermi_core.h.
| #define i32_im I5.w |
Definition at line 67 of file tm_dslash_fermi_core.h.
| #define i32_re I5.z |
Definition at line 66 of file tm_dslash_fermi_core.h.
| #define READ_SPINOR_SHARED READ_SPINOR_SHARED_FLOAT4 |
Definition at line 43 of file tm_dslash_fermi_core.h.
| #define SHARED_STRIDE 32 |
Definition at line 168 of file tm_dslash_fermi_core.h.
| #define spinorFloat float |
Definition at line 41 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 42 of file tm_dslash_fermi_core.h.
| if | ( | x3 >= | X3 | ) |
| if | ( | x2 >= | X2 | ) |
Definition at line 193 of file tm_dslash_fermi_core.h.
| o00_im = 0 |
Definition at line 141 of file tm_dslash_fermi_core.h.
| o00_re = 0 |
Definition at line 140 of file tm_dslash_fermi_core.h.
| o01_im = 0 |
Definition at line 143 of file tm_dslash_fermi_core.h.
| o01_re = 0 |
Definition at line 142 of file tm_dslash_fermi_core.h.
| o02_im = 0 |
Definition at line 145 of file tm_dslash_fermi_core.h.
| o02_re = 0 |
Definition at line 144 of file tm_dslash_fermi_core.h.
| o10_im = 0 |
Definition at line 147 of file tm_dslash_fermi_core.h.
| o10_re = 0 |
Definition at line 146 of file tm_dslash_fermi_core.h.
| o11_im = 0 |
Definition at line 149 of file tm_dslash_fermi_core.h.
| o11_re = 0 |
Definition at line 148 of file tm_dslash_fermi_core.h.
| o12_im = 0 |
Definition at line 151 of file tm_dslash_fermi_core.h.
| o12_re = 0 |
Definition at line 150 of file tm_dslash_fermi_core.h.
| o20_im = 0 |
Definition at line 153 of file tm_dslash_fermi_core.h.
| o20_re = 0 |
Definition at line 152 of file tm_dslash_fermi_core.h.
| o21_im = 0 |
Definition at line 155 of file tm_dslash_fermi_core.h.
| o21_re = 0 |
Definition at line 154 of file tm_dslash_fermi_core.h.
| o22_im = 0 |
Definition at line 157 of file tm_dslash_fermi_core.h.
| o22_re = 0 |
Definition at line 156 of file tm_dslash_fermi_core.h.
| o30_im = 0 |
Definition at line 159 of file tm_dslash_fermi_core.h.
| o30_re = 0 |
Definition at line 158 of file tm_dslash_fermi_core.h.
| o31_im = 0 |
Definition at line 161 of file tm_dslash_fermi_core.h.
| o31_re = 0 |
Definition at line 160 of file tm_dslash_fermi_core.h.
| o32_im = 0 |
Definition at line 163 of file tm_dslash_fermi_core.h.
| o32_re = 0 |
Definition at line 162 of file tm_dslash_fermi_core.h.
Definition at line 182 of file tm_dslash_fermi_core.h.
| x2 = blockIdx.y*blockDim.y + threadIdx.y |
Definition at line 175 of file tm_dslash_fermi_core.h.
| x3 = blockIdx.z*blockDim.z + threadIdx.z |
Definition at line 175 of file tm_dslash_fermi_core.h.
Definition at line 175 of file tm_dslash_fermi_core.h.
| int xt = blockIdx.x*blockDim.x + threadIdx.x |
Definition at line 192 of file tm_dslash_fermi_core.h.
1.7.4
