Lines Matching refs:T
54 my @T = map("%ymm$_",(7..15));
55 my ($C14,$C00,$D00,$D14) = @T[5..8];
73 vpxor $A11,$A21,@T[2]
74 vpternlogq \$0x96,$A01,$T[2],$C14 # C[1..4]
77 vpermq \$0b01001110,$C00,@T[0]
79 vpermq \$0b10010011,$C14,@T[4]
80 vprolq \$1,$C14,@T[1] # ROL64(C[1..4],1)
82 vpermq \$0b00111001,@T[1],$D14
83 vpxor @T[4],@T[1],$D00
86 vpternlogq \$0x96,@T[0],$A00,$C00 # C[0..0]
87 vprolq \$1,$C00,@T[1] # ROL64(C[0..0],1)
91 vpblendd \$0b11000000,@T[1],$D14,$D14
92 vpblendd \$0b00000011,$C00,@T[4],@T[0]
98 vpternlogq \$0x96,@T[0],$D14,$A31 # ^= D[1..4] from Theta
101 vpternlogq \$0x96,@T[0],$D14,$A21 # ^= D[1..4] from Theta
104 vpternlogq \$0x96,@T[0],$D14,$A41 # ^= D[1..4] from Theta
107 vpermq \$0b10001101,$A20,@T[3] # $A20 -> future $A31
108 vpermq \$0b10001101,$A31,@T[4] # $A31 -> future $A21
109 vpternlogq \$0x96,@T[0],$D14,$A11 # ^= D[1..4] from Theta
110 vprolvq $R11,$A11,@T[1] # $A11 -> future $A01
112 vpermq \$0b00011011,$A21,@T[5] # $A21 -> future $A41
113 vpermq \$0b01110010,$A41,@T[6] # $A41 -> future $A11
114 vpternlogq \$0x96,@T[0],$D14,$A01 # ^= D[1..4] from Theta
115 vprolvq $R01,$A01,@T[2] # $A01 -> future $A20
118 vpblendd \$0b00001100,@T[6],@T[2],$A31 # [4][4] [2][0]
119 vpblendd \$0b00001100,@T[2],@T[4],@T[8] # [4][0] [2][1]
120 vpblendd \$0b00001100,@T[4],@T[3],$A41 # [4][2] [2][4]
121 vpblendd \$0b00001100,@T[3],@T[2],@T[7] # [4][3] [2][0]
122 vpblendd \$0b00110000,@T[4],$A31,$A31 # [1][3] [4][4] [2][0]
123 vpblendd \$0b00110000,@T[5],@T[8],@T[8] # [1][4] [4][0] [2][1]
124 vpblendd \$0b00110000,@T[2],$A41,$A41 # [1][0] [4][2] [2][4]
125 vpblendd \$0b00110000,@T[6],@T[7],@T[7] # [1][1] [4][3] [2][0]
126 vpblendd \$0b11000000,@T[5],$A31,$A31 # [3][2] [1][3] [4][4] [2][0]
127 vpblendd \$0b11000000,@T[6],@T[8],@T[8] # [3][3] [1][4] [4][0] [2][1]
128 vpblendd \$0b11000000,@T[6],$A41,$A41 # [3][3] [1][0] [4][2] [2][4]
129 vpblendd \$0b11000000,@T[4],@T[7],@T[7] # [3][4] [1][1] [4][3] [2][0]
130 vpternlogq \$0xC6,@T[8],@T[3],$A31 # [3][1] [1][2] [4][3] [2][4]
131 vpternlogq \$0xC6,@T[7],@T[5],$A41 # [3][2] [1][4] [4][1] [2][3]
133 vpsrldq \$8,@T[1],@T[0]
134 vpandn @T[0],@T[1],@T[0] # tgting [0][0] [0][0] [0][0] [0][0]
136 vpblendd \$0b00001100,@T[2],@T[5],$A11 # [4][0] [2][3]
137 vpblendd \$0b00001100,@T[5],@T[3],@T[8] # [4][1] [2][4]
138 vpblendd \$0b00110000,@T[3],$A11,$A11 # [1][2] [4][0] [2][3]
139 vpblendd \$0b00110000,@T[4],@T[8],@T[8] # [1][3] [4][1] [2][4]
140 vpblendd \$0b11000000,@T[4],$A11,$A11 # [3][4] [1][2] [4][0] [2][3]
141 vpblendd \$0b11000000,@T[2],@T[8],@T[8] # [3][0] [1][3] [4][1] [2][4]
142 vpternlogq \$0xC6,@T[8],@T[6],$A11 # [3][3] [1][1] [4][4] [2][2]
144 vpermq \$0b00011110,@T[1],$A21 # [0][1] [0][2] [0][4] [0][3]
145 vpblendd \$0b00110000,$A00,$A21,@T[8] # [0][1] [0][0] [0][4] [0][3]
146 vpermq \$0b00111001,@T[1],$A01 # [0][1] [0][4] [0][3] [0][2]
149 vpblendd \$0b00001100,@T[5],@T[4],$A20 # [4][1] [2][1]
150 vpblendd \$0b00001100,@T[4],@T[6],@T[7] # [4][2] [2][2]
151 vpblendd \$0b00110000,@T[6],$A20,$A20 # [1][1] [4][1] [2][1]
152 vpblendd \$0b00110000,@T[3],@T[7],@T[7] # [1][2] [4][2] [2][2]
153 vpblendd \$0b11000000,@T[3],$A20,$A20 # [3][1] [1][1] [4][1] [2][1]
154 vpblendd \$0b11000000,@T[5],@T[7],@T[7] # [3][2] [1][2] [4][2] [2][2]
155 vpternlogq \$0xC6,@T[7],@T[2],$A20 # [3][0] [1][0] [4][0] [2][0]
157 vpermq \$0b00000000,@T[0],@T[0] # [0][0] [0][0] [0][0] [0][0]
162 vpblendd \$0b00001100,@T[3],@T[6],$A21 # [4][3] [2][2]
163 vpblendd \$0b00001100,@T[6],@T[5],@T[7] # [4][4] [2][3]
164 vpblendd \$0b00110000,@T[5],$A21,$A21 # [1][4] [4][3] [2][2]
165 vpblendd \$0b00110000,@T[2],@T[7],@T[7] # [1][0] [4][4] [2][3]
166 vpblendd \$0b11000000,@T[2],$A21,$A21 # [3][0] [1][4] [4][3] [2][2]
167 vpblendd \$0b11000000,@T[3],@T[7],@T[7] # [3][1] [1][0] [4][4] [2][3]
169 vpternlogq \$0xC6,@T[8],@T[1],$A01 # [0][4] [0][3] [0][2] [0][1]
170 vpternlogq \$0xC6,@T[7],@T[4],$A21 # [3][4] [1][3] [4][2] [2][1]
173 vpternlogq \$0x96,(%r10),@T[0],$A00
217 vpxor @T[0],@T[0],@T[0]
218 vmovdqa @T[0],32*2-96(%r10) # zero transfer area on stack
219 vmovdqa @T[0],32*3-96(%r10)
220 vmovdqa @T[0],32*4-96(%r10)
221 vmovdqa @T[0],32*5-96(%r10)
222 vmovdqa @T[0],32*6-96(%r10)
230 vpbroadcastq 0-96($inp),@T[0]
231 vmovdqu 8-96($inp),@T[1]
246 vpxor @T[0],$A00,$A00
247 vpxor @T[1],$A01,$A01
288 vpxor @T[0],@T[0],@T[0]
