buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(1+(i-1)*kmax) * sphi_a(1,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(1+(i-1)*kmax) * sphi_a(1,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(1+(i-1)*kmax) * sphi_a(1,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(1+(i-1)*kmax) * sphi_a(1,15+s_offset_a1)
buffer1(i+imax*(17-1)) = buffer1(i+imax*(17-1)) + work(1+(i-1)*kmax) * sphi_a(1,17+s_offset_a1)
buffer1(i+imax*(1-1)) = buffer1(i+imax*(1-1)) + work(2+(i-1)*kmax) * sphi_a(2,1+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(2+(i-1)*kmax) * sphi_a(2,3+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(2+(i-1)*kmax) * sphi_a(2,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(2+(i-1)*kmax) * sphi_a(2,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(3+(i-1)*kmax) * sphi_a(3,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(3+(i-1)*kmax) * sphi_a(3,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(3+(i-1)*kmax) * sphi_a(3,14+s_offset_a1)
buffer1(i+imax*(16-1)) = buffer1(i+imax*(16-1)) + work(3+(i-1)*kmax) * sphi_a(3,16+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(4+(i-1)*kmax) * sphi_a(4,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(4+(i-1)*kmax) * sphi_a(4,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(4+(i-1)*kmax) * sphi_a(4,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(4+(i-1)*kmax) * sphi_a(4,15+s_offset_a1)
buffer1(i+imax*(17-1)) = buffer1(i+imax*(17-1)) + work(4+(i-1)*kmax) * sphi_a(4,17+s_offset_a1)
buffer1(i+imax*(2-1)) = buffer1(i+imax*(2-1)) + work(5+(i-1)*kmax) * sphi_a(5,2+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(5+(i-1)*kmax) * sphi_a(5,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(5+(i-1)*kmax) * sphi_a(5,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(5+(i-1)*kmax) * sphi_a(5,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(6+(i-1)*kmax) * sphi_a(6,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(6+(i-1)*kmax) * sphi_a(6,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(6+(i-1)*kmax) * sphi_a(6,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(6+(i-1)*kmax) * sphi_a(6,15+s_offset_a1)
buffer1(i+imax*(1-1)) = buffer1(i+imax*(1-1)) + work(7+(i-1)*kmax) * sphi_a(7,1+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(7+(i-1)*kmax) * sphi_a(7,3+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(7+(i-1)*kmax) * sphi_a(7,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(7+(i-1)*kmax) * sphi_a(7,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(8+(i-1)*kmax) * sphi_a(8,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(8+(i-1)*kmax) * sphi_a(8,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(8+(i-1)*kmax) * sphi_a(8,14+s_offset_a1)
buffer1(i+imax*(16-1)) = buffer1(i+imax*(16-1)) + work(8+(i-1)*kmax) * sphi_a(8,16+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(9+(i-1)*kmax) * sphi_a(9,3+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(9+(i-1)*kmax) * sphi_a(9,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(9+(i-1)*kmax) * sphi_a(9,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(10+(i-1)*kmax) * sphi_a(10,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(10+(i-1)*kmax) * sphi_a(10,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(10+(i-1)*kmax) * sphi_a(10,14+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(11+(i-1)*kmax) * sphi_a(11,9+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(11+(i-1)*kmax) * sphi_a(11,13+s_offset_a1)
buffer1(i+imax*(17-1)) = buffer1(i+imax*(17-1)) + work(11+(i-1)*kmax) * sphi_a(11,17+s_offset_a1)
buffer1(i+imax*(2-1)) = buffer1(i+imax*(2-1)) + work(12+(i-1)*kmax) * sphi_a(12,2+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(12+(i-1)*kmax) * sphi_a(12,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(12+(i-1)*kmax) * sphi_a(12,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(12+(i-1)*kmax) * sphi_a(12,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(13+(i-1)*kmax) * sphi_a(13,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(13+(i-1)*kmax) * sphi_a(13,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(13+(i-1)*kmax) * sphi_a(13,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(13+(i-1)*kmax) * sphi_a(13,15+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(14+(i-1)*kmax) * sphi_a(14,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(14+(i-1)*kmax) * sphi_a(14,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(14+(i-1)*kmax) * sphi_a(14,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(15+(i-1)*kmax) * sphi_a(15,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(15+(i-1)*kmax) * sphi_a(15,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(15+(i-1)*kmax) * sphi_a(15,13+s_offset_a1)
buffer1(i+imax*(1-1)) = buffer1(i+imax*(1-1)) + work(16+(i-1)*kmax) * sphi_a(16,1+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(16+(i-1)*kmax) * sphi_a(16,3+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(16+(i-1)*kmax) * sphi_a(16,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(16+(i-1)*kmax) * sphi_a(16,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(17+(i-1)*kmax) * sphi_a(17,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(17+(i-1)*kmax) * sphi_a(17,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(17+(i-1)*kmax) * sphi_a(17,14+s_offset_a1)
buffer1(i+imax*(16-1)) = buffer1(i+imax*(16-1)) + work(17+(i-1)*kmax) * sphi_a(17,16+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(18+(i-1)*kmax) * sphi_a(18,3+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(18+(i-1)*kmax) * sphi_a(18,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(19+(i-1)*kmax) * sphi_a(19,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(19+(i-1)*kmax) * sphi_a(19,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(19+(i-1)*kmax) * sphi_a(19,14+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(20+(i-1)*kmax) * sphi_a(20,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(20+(i-1)*kmax) * sphi_a(20,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(21+(i-1)*kmax) * sphi_a(21,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(21+(i-1)*kmax) * sphi_a(21,12+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(22+(i-1)*kmax) * sphi_a(22,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(22+(i-1)*kmax) * sphi_a(22,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(22+(i-1)*kmax) * sphi_a(22,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(22+(i-1)*kmax) * sphi_a(22,15+s_offset_a1)
buffer1(i+imax*(17-1)) = buffer1(i+imax*(17-1)) + work(22+(i-1)*kmax) * sphi_a(22,17+s_offset_a1)
buffer1(i+imax*(2-1)) = buffer1(i+imax*(2-1)) + work(23+(i-1)*kmax) * sphi_a(23,2+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(23+(i-1)*kmax) * sphi_a(23,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(23+(i-1)*kmax) * sphi_a(23,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(23+(i-1)*kmax) * sphi_a(23,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(24+(i-1)*kmax) * sphi_a(24,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(24+(i-1)*kmax) * sphi_a(24,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(24+(i-1)*kmax) * sphi_a(24,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(24+(i-1)*kmax) * sphi_a(24,15+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(25+(i-1)*kmax) * sphi_a(25,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(25+(i-1)*kmax) * sphi_a(25,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(25+(i-1)*kmax) * sphi_a(25,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(26+(i-1)*kmax) * sphi_a(26,9+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(26+(i-1)*kmax) * sphi_a(26,13+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(27+(i-1)*kmax) * sphi_a(27,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(27+(i-1)*kmax) * sphi_a(27,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(28+(i-1)*kmax) * sphi_a(28,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(28+(i-1)*kmax) * sphi_a(28,11+s_offset_a1)
buffer1(i+imax*(1-1)) = buffer1(i+imax*(1-1)) + work(29+(i-1)*kmax) * sphi_a(29,1+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(29+(i-1)*kmax) * sphi_a(29,3+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(29+(i-1)*kmax) * sphi_a(29,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(29+(i-1)*kmax) * sphi_a(29,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(30+(i-1)*kmax) * sphi_a(30,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(30+(i-1)*kmax) * sphi_a(30,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(30+(i-1)*kmax) * sphi_a(30,14+s_offset_a1)
buffer1(i+imax*(16-1)) = buffer1(i+imax*(16-1)) + work(30+(i-1)*kmax) * sphi_a(30,16+s_offset_a1)
buffer1(i+imax*(3-1)) = buffer1(i+imax*(3-1)) + work(31+(i-1)*kmax) * sphi_a(31,3+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(31+(i-1)*kmax) * sphi_a(31,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(31+(i-1)*kmax) * sphi_a(31,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(32+(i-1)*kmax) * sphi_a(32,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(32+(i-1)*kmax) * sphi_a(32,12+s_offset_a1)
buffer1(i+imax*(14-1)) = buffer1(i+imax*(14-1)) + work(32+(i-1)*kmax) * sphi_a(32,14+s_offset_a1)
buffer1(i+imax*(5-1)) = buffer1(i+imax*(5-1)) + work(33+(i-1)*kmax) * sphi_a(33,5+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(33+(i-1)*kmax) * sphi_a(33,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(34+(i-1)*kmax) * sphi_a(34,10+s_offset_a1)
buffer1(i+imax*(12-1)) = buffer1(i+imax*(12-1)) + work(34+(i-1)*kmax) * sphi_a(34,12+s_offset_a1)
buffer1(i+imax*(7-1)) = buffer1(i+imax*(7-1)) + work(35+(i-1)*kmax) * sphi_a(35,7+s_offset_a1)
buffer1(i+imax*(10-1)) = buffer1(i+imax*(10-1)) + work(36+(i-1)*kmax) * sphi_a(36,10+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(37+(i-1)*kmax) * sphi_a(37,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(37+(i-1)*kmax) * sphi_a(37,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(37+(i-1)*kmax) * sphi_a(37,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(37+(i-1)*kmax) * sphi_a(37,15+s_offset_a1)
buffer1(i+imax*(17-1)) = buffer1(i+imax*(17-1)) + work(37+(i-1)*kmax) * sphi_a(37,17+s_offset_a1)
buffer1(i+imax*(2-1)) = buffer1(i+imax*(2-1)) + work(38+(i-1)*kmax) * sphi_a(38,2+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(38+(i-1)*kmax) * sphi_a(38,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(38+(i-1)*kmax) * sphi_a(38,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(38+(i-1)*kmax) * sphi_a(38,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(39+(i-1)*kmax) * sphi_a(39,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(39+(i-1)*kmax) * sphi_a(39,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(39+(i-1)*kmax) * sphi_a(39,13+s_offset_a1)
buffer1(i+imax*(15-1)) = buffer1(i+imax*(15-1)) + work(39+(i-1)*kmax) * sphi_a(39,15+s_offset_a1)
buffer1(i+imax*(4-1)) = buffer1(i+imax*(4-1)) + work(40+(i-1)*kmax) * sphi_a(40,4+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(40+(i-1)*kmax) * sphi_a(40,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(40+(i-1)*kmax) * sphi_a(40,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(41+(i-1)*kmax) * sphi_a(41,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(41+(i-1)*kmax) * sphi_a(41,11+s_offset_a1)
buffer1(i+imax*(13-1)) = buffer1(i+imax*(13-1)) + work(41+(i-1)*kmax) * sphi_a(41,13+s_offset_a1)
buffer1(i+imax*(6-1)) = buffer1(i+imax*(6-1)) + work(42+(i-1)*kmax) * sphi_a(42,6+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(42+(i-1)*kmax) * sphi_a(42,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(43+(i-1)*kmax) * sphi_a(43,9+s_offset_a1)
buffer1(i+imax*(11-1)) = buffer1(i+imax*(11-1)) + work(43+(i-1)*kmax) * sphi_a(43,11+s_offset_a1)
buffer1(i+imax*(8-1)) = buffer1(i+imax*(8-1)) + work(44+(i-1)*kmax) * sphi_a(44,8+s_offset_a1)
buffer1(i+imax*(9-1)) = buffer1(i+imax*(9-1)) + work(45+(i-1)*kmax) * sphi_a(45,9+s_offset_a1)
