diff options
| author | Cedric Nugteren <web@cedricnugteren.nl> | 2017-12-09 14:09:13 +0100 |
|---|---|---|
| committer | Cedric Nugteren <web@cedricnugteren.nl> | 2017-12-09 14:09:13 +0100 |
| commit | 23e3a85f2c328d4a23db2fca5d1d89d78163711f (patch) | |
| tree | 02b8dd5364d958184c45c9bfdb2c28e38d72b24e /src/kernels/level3/xgemm_part2.opencl | |
| parent | d9df62b7942bb8af5fd385b8545aceb1d8b578f3 (diff) | |
Reformatted GEMM kernel to support array-to-register promotion
Diffstat (limited to 'src/kernels/level3/xgemm_part2.opencl')
| -rw-r--r-- | src/kernels/level3/xgemm_part2.opencl | 48 |
1 files changed, 0 insertions, 48 deletions
diff --git a/src/kernels/level3/xgemm_part2.opencl b/src/kernels/level3/xgemm_part2.opencl index 88100e96..14a0493a 100644 --- a/src/kernels/level3/xgemm_part2.opencl +++ b/src/kernels/level3/xgemm_part2.opencl @@ -63,54 +63,6 @@ INLINE_FUNC realM MultiplyAddVector(realM cvec, const realM avec, const real bva return cvec; } -// Performs the actual computation: Cpm += Apm * Bpm -INLINE_FUNC void MultiplyAccumulate(realM cpm[NWI*MWI/VWM], realM apm[MWI/VWM], realN bpm[NWI/VWN]) { - #pragma unroll - for (int _ni = 0; _ni < NWI/VWN; _ni += 1) { - #pragma unroll - for (int _mi = 0; _mi < MWI/VWM; _mi += 1) { - const realM aval = apm[_mi]; - #if VWN == 1 - cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni]); - #elif VWN == 2 - cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni].x); - cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi], aval, bpm[_ni].y); - #elif VWN == 4 - cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni].x); - cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi], aval, bpm[_ni].y); - cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi], aval, bpm[_ni].z); - cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi], aval, bpm[_ni].w); - #elif VWN == 8 - cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni].s0); - cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi], aval, bpm[_ni].s1); - cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi], aval, bpm[_ni].s2); - cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi], aval, bpm[_ni].s3); - cpm[(_ni*VWN + 4)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 4)*(MWI/VWM) + _mi], aval, bpm[_ni].s4); - cpm[(_ni*VWN + 5)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 5)*(MWI/VWM) + _mi], aval, bpm[_ni].s5); - cpm[(_ni*VWN + 6)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 6)*(MWI/VWM) + _mi], aval, bpm[_ni].s6); - cpm[(_ni*VWN + 7)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 7)*(MWI/VWM) + _mi], aval, bpm[_ni].s7); - #elif VWN == 16 - cpm[(_ni*VWN + 0 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0 )*(MWI/VWM) + _mi], aval, bpm[_ni].s0); - cpm[(_ni*VWN + 1 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1 )*(MWI/VWM) + _mi], aval, bpm[_ni].s1); - cpm[(_ni*VWN + 2 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 2 )*(MWI/VWM) + _mi], aval, bpm[_ni].s2); - cpm[(_ni*VWN + 3 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 3 )*(MWI/VWM) + _mi], aval, bpm[_ni].s3); - cpm[(_ni*VWN + 4 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 4 )*(MWI/VWM) + _mi], aval, bpm[_ni].s4); - cpm[(_ni*VWN + 5 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 5 )*(MWI/VWM) + _mi], aval, bpm[_ni].s5); - cpm[(_ni*VWN + 6 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 6 )*(MWI/VWM) + _mi], aval, bpm[_ni].s6); - cpm[(_ni*VWN + 7 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 7 )*(MWI/VWM) + _mi], aval, bpm[_ni].s7); - cpm[(_ni*VWN + 8 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 8 )*(MWI/VWM) + _mi], aval, bpm[_ni].s8); - cpm[(_ni*VWN + 9 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 9 )*(MWI/VWM) + _mi], aval, bpm[_ni].s9); - cpm[(_ni*VWN + 10)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 10)*(MWI/VWM) + _mi], aval, bpm[_ni].sA); - cpm[(_ni*VWN + 11)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 11)*(MWI/VWM) + _mi], aval, bpm[_ni].sB); - cpm[(_ni*VWN + 12)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 12)*(MWI/VWM) + _mi], aval, bpm[_ni].sC); - cpm[(_ni*VWN + 13)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 13)*(MWI/VWM) + _mi], aval, bpm[_ni].sD); - cpm[(_ni*VWN + 14)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 14)*(MWI/VWM) + _mi], aval, bpm[_ni].sE); - cpm[(_ni*VWN + 15)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 15)*(MWI/VWM) + _mi], aval, bpm[_ni].sF); - #endif - } - } -} - // ================================================================================================= // Merges the results in Cpm with the global array in Cgm. This also performs the multiplication |