diff options
Diffstat (limited to 'src/kernels/level3/xgemm_direct_part1.opencl')
| -rw-r--r-- | src/kernels/level3/xgemm_direct_part1.opencl | 262 |
1 files changed, 108 insertions, 154 deletions
diff --git a/src/kernels/level3/xgemm_direct_part1.opencl b/src/kernels/level3/xgemm_direct_part1.opencl index cb407824..2e5addef 100644 --- a/src/kernels/level3/xgemm_direct_part1.opencl +++ b/src/kernels/level3/xgemm_direct_part1.opencl @@ -10,7 +10,7 @@ // This is a generic GEMM kernel that works for all sizes and configurations: it doesn't require any // pre and and post-processing kernels. // -// This kernel is seperated into three files. This is part 1 out of 2. +// This kernel is seperated into three files. This is part 1 out of 3. // // ================================================================================================= @@ -92,196 +92,150 @@ R"( // ================================================================================================= -// Caches global off-chip memory into local (shared) memory on-chip. This function is specific for -// caching the A input matrix. -inline void GlobalToLocalDirectA(const __global realMD* restrict agm, __local real* alm, - const int a_ld, const int a_offset, const int kwg, - const int a_transpose, const int a_conjugate) { - #if MDIMCD == MDIMAD - const int la0 = get_local_id(0); - const int la1 = get_local_id(1); - #else - const int tid = get_local_id(0) + MDIMCD*get_local_id(1); - const int la0 = tid % MDIMAD; - const int la1 = tid / MDIMAD; - #endif +// Initializes the accumulation registers to zero +inline void InitAccRegistersDirect(real cpm[NWID][MWID]) { #pragma unroll - for (int mia=0; mia<MWAD/VWMD; ++mia) { + for (int mi=0; mi<MWID; ++mi) { #pragma unroll - for (int kia=0; kia<KWAD; ++kia) { - - // Computes the indices for the global memory - int mg = mia + la0*(MWAD/VWMD); - int kg = kia + la1*KWAD; - int idm = (a_transpose) ? mg + kwg/VWMD : mg + GetGroupID0()*(WGD/VWMD); - int idk = (a_transpose) ? kg + GetGroupID0()*WGD : kg + kwg; - - // Loads the data from global memory into the local memory - const realMD avec = agm[idk*(a_ld/VWMD) + idm + a_offset]; - #if VWMD == 1 - alm[kg*(WGD + PADA) + mg] = avec; - #elif VWMD == 2 - alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.x; - alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.y; - #elif VWMD == 4 - alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.x; - alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.y; - alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.z; - alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.w; - #elif VWMD == 8 - alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.s0; - alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.s1; - alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.s2; - alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.s3; - alm[kg*(WGD + PADA) + mg*VWMD + 4] = avec.s4; - alm[kg*(WGD + PADA) + mg*VWMD + 5] = avec.s5; - alm[kg*(WGD + PADA) + mg*VWMD + 6] = avec.s6; - alm[kg*(WGD + PADA) + mg*VWMD + 7] = avec.s7; - #elif VWMD == 16 - alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.s0; - alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.s1; - alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.s2; - alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.s3; - alm[kg*(WGD + PADA) + mg*VWMD + 4] = avec.s4; - alm[kg*(WGD + PADA) + mg*VWMD + 5] = avec.s5; - alm[kg*(WGD + PADA) + mg*VWMD + 6] = avec.s6; - alm[kg*(WGD + PADA) + mg*VWMD + 7] = avec.s7; - alm[kg*(WGD + PADA) + mg*VWMD + 8] = avec.s8; - alm[kg*(WGD + PADA) + mg*VWMD + 9] = avec.s9; - alm[kg*(WGD + PADA) + mg*VWMD + 10] = avec.sA; - alm[kg*(WGD + PADA) + mg*VWMD + 11] = avec.sB; - alm[kg*(WGD + PADA) + mg*VWMD + 12] = avec.sC; - alm[kg*(WGD + PADA) + mg*VWMD + 13] = avec.sD; - alm[kg*(WGD + PADA) + mg*VWMD + 14] = avec.sE; - alm[kg*(WGD + PADA) + mg*VWMD + 15] = avec.sF; - #endif - if (a_conjugate) { - for (int vm=0; vm<VWMD; ++vm) { - COMPLEX_CONJUGATE(alm[kg*(WGD + PADA) + mg*VWMD + vm]); - } - } + for (int ni=0; ni<NWID; ++ni) { + SetToZero(cpm[ni][mi]); } } } -// Same as above, but now for the B input matrix -inline void GlobalToLocalDirectB(const __global realND* restrict bgm, __local real* blm, - const int b_ld, const int b_offset, const int kwg, - const int b_transpose, const int b_conjugate) { - #if MDIMCD == NDIMBD - const int lb0 = get_local_id(0); - const int lb1 = get_local_id(1); - #else - const int tid = get_local_id(0) + MDIMCD*get_local_id(1); - const int lb0 = tid % NDIMBD; - const int lb1 = tid / NDIMBD; - #endif +// ================================================================================================= + +// Performs the actual computation: Cpm += Apm * Bpm +inline void MultiplyAccumulateDirect(real cpm[NWID][MWID], real apm[MWID], real bpm[NWID]) { #pragma unroll - for (int kib=0; kib<KWBD; ++kib) { + for (int ni=0; ni<NWID; ++ni) { #pragma unroll - for (int nib=0; nib<NWBD/VWND; ++nib) { - - // Computes the indices for the global memory - int ng = nib + lb0*(NWBD/VWND); - int kg = kib + lb1*KWBD; - int idn = (b_transpose) ? ng + kwg/VWND : ng + GetGroupID1()*(WGD/VWND); - int idk = (b_transpose) ? kg + GetGroupID1()*WGD : kg + kwg; - - // Loads the data from global memory into the local memory - const realND bvec = bgm[idk*(b_ld/VWND) + idn + b_offset]; - #if VWND == 1 - blm[kg*(WGD + PADB) + ng] = bvec; - #elif VWND == 2 - blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.x; - blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.y; - #elif VWND == 4 - blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.x; - blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.y; - blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.z; - blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.w; - #elif VWND == 8 - blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.s0; - blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.s1; - blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.s2; - blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.s3; - blm[kg*(WGD + PADB) + ng*VWND + 4] = bvec.s4; - blm[kg*(WGD + PADB) + ng*VWND + 5] = bvec.s5; - blm[kg*(WGD + PADB) + ng*VWND + 6] = bvec.s6; - blm[kg*(WGD + PADB) + ng*VWND + 7] = bvec.s7; - #elif VWND == 16 - blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.s0; - blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.s1; - blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.s2; - blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.s3; - blm[kg*(WGD + PADB) + ng*VWND + 4] = bvec.s4; - blm[kg*(WGD + PADB) + ng*VWND + 5] = bvec.s5; - blm[kg*(WGD + PADB) + ng*VWND + 6] = bvec.s6; - blm[kg*(WGD + PADB) + ng*VWND + 7] = bvec.s7; - blm[kg*(WGD + PADB) + ng*VWND + 8] = bvec.s8; - blm[kg*(WGD + PADB) + ng*VWND + 9] = bvec.s9; - blm[kg*(WGD + PADB) + ng*VWND + 10] = bvec.sA; - blm[kg*(WGD + PADB) + ng*VWND + 11] = bvec.sB; - blm[kg*(WGD + PADB) + ng*VWND + 12] = bvec.sC; - blm[kg*(WGD + PADB) + ng*VWND + 13] = bvec.sD; - blm[kg*(WGD + PADB) + ng*VWND + 14] = bvec.sE; - blm[kg*(WGD + PADB) + ng*VWND + 15] = bvec.sF; - #endif - if (b_conjugate) { - for (int vn=0; vn<VWND; ++vn) { - COMPLEX_CONJUGATE(blm[kg*(WGD + PADB) + ng*VWND + vn]); - } - } + for (int mi=0; mi<MWID; ++mi) { + MultiplyAdd(cpm[ni][mi], apm[mi], bpm[ni]); } } } // ================================================================================================= -// Caches on-chip local memory into per-thread private memory (registers). This function is specific -// for caching the A input matrix. -inline void LocalToPrivateDirectA(__local real* alm, real apm[MWID], const int kg, - const int a_transpose) { +// Loads global off-chip memory into thread-private register files. This function is specific for +// loading the A input matrix. +inline void GlobalToPrivateDirectA(const __global real* restrict agms, real apm[MWID], + const int a_ld, const int a_offset, const int idm, const int idk, + const int a_transpose, const int a_conjugate) { #pragma unroll for (int mi=0; mi<MWID; ++mi) { - const int mg = mi + get_local_id(0)*MWID; - const int index = (a_transpose) ? mg*(WGD + PADA) + kg : kg*(WGD + PADA) + mg; - apm[mi] = alm[index]; + const int a_index = (a_transpose) ? (idm + mi)*a_ld + idk : idk*a_ld + (idm + mi); + apm[mi] = agms[a_index + a_offset]; + if (a_conjugate) { COMPLEX_CONJUGATE(apm[mi]); } } } // Same as above, but now for the B input matrix -inline void LocalToPrivateDirectB(__local real* blm, real bpm[NWID], const int kg, - const int b_transpose) { +inline void GlobalToPrivateDirectB(const __global real* restrict bgms, real bpm[NWID], + const int b_ld, const int b_offset, const int idn, const int idk, + const int b_transpose, const int b_conjugate) { #pragma unroll for (int ni=0; ni<NWID; ++ni) { - const int ng = ni + get_local_id(1)*NWID; - const int index = (b_transpose) ? ng*(WGD + PADB) + kg : kg*(WGD + PADB) + ng; - bpm[ni] = blm[index]; + const int b_index = (b_transpose) ? (idn + ni)*b_ld + idk : idk*b_ld + (idn + ni); + bpm[ni] = bgms[b_index + b_offset]; + if (b_conjugate) { COMPLEX_CONJUGATE(bpm[ni]); } } } -// ================================================================================================= - -// Initializes the accumulation registers to zero -inline void InitAccRegistersDirect(real cpm[NWID][MWID]) { +// Loads global off-chip memory into thread-private register files. This function is specific for +// loading the A input matrix. This is the same as above but now includes a bounds check. +inline void GlobalToPrivateCheckedA(const __global real* restrict agms, real apm[MWID], + const int a_ld, const int a_offset, const int idm, const int idk, + const int a_transpose, const int a_conjugate, + const int kSizeM) { #pragma unroll for (int mi=0; mi<MWID; ++mi) { - #pragma unroll - for (int ni=0; ni<NWID; ++ni) { - SetToZero(cpm[ni][mi]); + if (idm + mi < kSizeM) { + const int a_index = (a_transpose) ? (idm + mi)*a_ld + idk : idk*a_ld + (idm + mi); + apm[mi] = agms[a_index + a_offset]; + if (a_conjugate) { COMPLEX_CONJUGATE(apm[mi]); } + } + else { + SetToZero(apm[mi]); + } + } +} + +// Same as above, but now for the B input matrix +inline void GlobalToPrivateCheckedB(const __global real* restrict bgms, real bpm[NWID], + const int b_ld, const int b_offset, const int idn, const int idk, + const int b_transpose, const int b_conjugate, + const int kSizeN) { + #pragma unroll + for (int ni=0; ni<NWID; ++ni) { + if (idn + ni < kSizeN) { + const int b_index = (b_transpose) ? (idn + ni)*b_ld + idk : idk*b_ld + (idn + ni); + bpm[ni] = bgms[b_index + b_offset]; + if (b_conjugate) { COMPLEX_CONJUGATE(bpm[ni]); } + } + else { + SetToZero(bpm[ni]); } } } // ================================================================================================= -// Performs the actual computation: Cpm += Apm * Bpm -inline void MultiplyAccumulateDirect(real cpm[NWID][MWID], real apm[MWID], real bpm[NWID]) { +// Merges the results in Cpm with the global array in Cgm. This also performs the multiplication +// with the constants: Cgm = alpha*A*B + beta*Cgm = alpha*Cpm + beta*Cgm +inline void StoreResultsDirect(__global real* cgm, real cpm[NWID][MWID], + const int idm, const int idn, + const real alpha, const real beta, + const int c_ld, const int c_offset, const int c_transpose) { #pragma unroll for (int ni=0; ni<NWID; ++ni) { #pragma unroll for (int mi=0; mi<MWID; ++mi) { - MultiplyAdd(cpm[ni][mi], apm[mi], bpm[ni]); + + // Determines the destination index + int c_index = (c_transpose) ? (idm + mi)*c_ld + (idn + ni) : (idn + ni)*c_ld + (idm + mi); + + // The final multiplication with alpha (in case beta == 0) + real result; + if (IsZero(beta)) { + Multiply(result, alpha, cpm[ni][mi]); + } + // The final multiplication with alpha and the addition with beta*C + else { + AXPBY(result, alpha, cpm[ni][mi], beta, cgm[c_index + c_offset]); + } + cgm[c_index + c_offset] = result; + } + } +} + +// Merges the results in Cpm with the global array in Cgm. This also performs the multiplication +// with the constants: Cgm = alpha*A*B + beta*Cgm = alpha*Cpm + beta*Cgm +inline void StoreResultsChecked(__global real* cgm, real cpm[NWID][MWID], + const int idm, const int idn, const int kSizeM, const int kSizeN, + const real alpha, const real beta, + const int c_ld, const int c_offset, const int c_transpose) { + #pragma unroll + for (int ni=0; ni<NWID; ++ni) { + #pragma unroll + for (int mi=0; mi<MWID; ++mi) { + if ((idm + mi) < kSizeM && (idn + ni) < kSizeN) { + + // Determines the destination index + int c_index = (c_transpose) ? (idm + mi)*c_ld + (idn + ni) : (idn + ni)*c_ld + (idm + mi); + + // The final multiplication with alpha (in case beta == 0) + real result; + if (IsZero(beta)) { + Multiply(result, alpha, cpm[ni][mi]); + } + // The final multiplication with alpha and the addition with beta*C + else { + AXPBY(result, alpha, cpm[ni][mi], beta, cgm[c_index + c_offset]); + } + cgm[c_index + c_offset] = result; + } } } } |