Lines Matching full:cells
79 // A 12x2 block of 3 4x2 cells Lhs is stored in 16bit in d2--d7
115 // Load 3 Lhs cells of size 4x2 each
120 // Expand Lhs/Rhs cells to 16 bit.
301 // This 12x4 block of global accumulators is laid out as 3 cells of size 4x4
408 // accumulators (3 cells of size 4x4 in diagonal-major order)
413 // A 12x2 block of 3 4x2 cells Lhs is stored in 8bit in d1--d3.
446 /* Load 3 Lhs cells of size 4x2 */ \
699 // A 2x8 block of 2 2x4 cells of Rhs is stored in 16bit in v0--v1.
700 // A 12x2 block of 3 4x2 cells Lhs is stored in 16bit in v2--v4.
736 // Load 3 Lhs cells of size 4x2 each
741 // Expand Lhs/Rhs cells to 16 bit.
989 template <int Cells>
991 typedef KernelFormat<KernelSideFormat<CellFormat<4, 2>, Cells>,
1006 uint32x4_t acc[Cells];
1007 for (int cell = 0; cell < Cells; cell++) {
1012 // Load LHS cells
1013 uint16x8_t lhs[Cells];
1014 for (int cell = 0; cell < Cells; cell++) {
1023 for (int cell = 0; cell < Cells; cell++) {
1027 for (int cell = 0; cell < Cells; cell++) {
1033 for (int cell = 0; cell < Cells; cell++) {
1039 for (int cell = 0; cell < Cells; cell++) {
