added optimized dtrsm_LT kernel for POWER8

This commit is contained in:
Werner Saar
2016-05-22 13:09:05 +02:00
parent 8a149e6294
commit 8fb5a1aaff
4 changed files with 5633 additions and 1 deletions
+716
View File
@@ -0,0 +1,716 @@
srawi. J, N, 2
ble DSTRM_LT_L4_END
DSTRM_LT_L4_BEGIN:
mr CO, C
mr AO, A
slwi T1, LDC , 2
add C, C, T1
mr KK, OFFSET
srawi. I, M, 4
ble DSTRM_LT_L4x16_END
DSTRM_LT_L4x16_BEGIN:
mr BO, B
DSTRM_LT_L4x16_LOOP_START:
INIT_16x4
addic. L, KK, 0
ble DSTRM_LT_L4x16_SAVE
DSTRM_LT_L4x16_LOOP:
KERNEL_16x4
addic. L, L, -1
bgt DSTRM_LT_L4x16_LOOP
DSTRM_LT_L4x16_SAVE:
SOLVE_LT_16x4
addi CO, CO, 16*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 4+BASE_SHIFT
slwi T4, T4, 2+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 16
addic. I, I, -1
bgt DSTRM_LT_L4x16_BEGIN
DSTRM_LT_L4x16_END:
DSTRM_LT_L4x8_BEGIN:
andi. T2, M, 15
ble DSTRM_LT_L4x1_END
andi. T1, M, 8
ble DSTRM_LT_L4x8_END
mr BO, B
DSTRM_LT_L4x8_LOOP_START:
INIT_8x4
addic. L, KK, 0
ble DSTRM_LT_L4x8_SAVE
DSTRM_LT_L4x8_LOOP:
KERNEL_8x4
addic. L, L, -1
bgt DSTRM_LT_L4x8_LOOP
DSTRM_LT_L4x8_SAVE:
SOLVE_LT_8x4
addi CO, CO, 8*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 3+BASE_SHIFT
slwi T4, T4, 2+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 8
DSTRM_LT_L4x8_END:
DSTRM_LT_L4x4_BEGIN:
andi. T1, M, 4
ble DSTRM_LT_L4x4_END
mr BO, B
DSTRM_LT_L4x4_LOOP_START:
INIT_4x4
addic. L, KK, 0
ble DSTRM_LT_L4x4_SAVE
DSTRM_LT_L4x4_LOOP:
KERNEL_4x4
addic. L, L, -1
bgt DSTRM_LT_L4x4_LOOP
DSTRM_LT_L4x4_SAVE:
SOLVE_LT_4x4
addi CO, CO, 4*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 2+BASE_SHIFT
slwi T4, T4, 2+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 4
DSTRM_LT_L4x4_END:
DSTRM_LT_L4x2_BEGIN:
andi. T1, M, 2
ble DSTRM_LT_L4x2_END
mr BO, B
DSTRM_LT_L4x2_LOOP_START:
INIT_2x4
addic. L, KK, 0
ble DSTRM_LT_L4x2_SAVE
DSTRM_LT_L4x2_LOOP:
KERNEL_2x4
addic. L, L, -1
bgt DSTRM_LT_L4x2_LOOP
DSTRM_LT_L4x2_SAVE:
SOLVE_LT_2x4
addi CO, CO, 2*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 1+BASE_SHIFT
slwi T4, T4, 2+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 2
DSTRM_LT_L4x2_END:
DSTRM_LT_L4x1_BEGIN:
andi. T1, M, 1
ble DSTRM_LT_L4x1_END
mr BO, B
DSTRM_LT_L4x1_LOOP_START:
INIT_1x4
addic. L, KK, 0
ble DSTRM_LT_L4x1_SAVE
DSTRM_LT_L4x1_LOOP:
KERNEL_1x4
addic. L, L, -1
bgt DSTRM_LT_L4x1_LOOP
DSTRM_LT_L4x1_SAVE:
SOLVE_LT_1x4
addi CO, CO, 1*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 0+BASE_SHIFT
slwi T4, T4, 2+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 1
DSTRM_LT_L4x1_END:
slwi T1, K, 2+BASE_SHIFT
add B, B, T1
addic. J, J, -1
bgt DSTRM_LT_L4_BEGIN
andi. T2, N, 3
ble L999
DSTRM_LT_L4_END:
b DSTRM_LT_L2_BEGIN
L999_H1:
b L999
DSTRM_LT_L2_BEGIN:
andi. T1, N, 2
ble DSTRM_LT_L2_END
mr CO, C
mr AO, A
slwi T1, LDC , 1
add C, C, T1
mr KK, OFFSET
srawi. I, M, 4
ble DSTRM_LT_L2x16_END
DSTRM_LT_L2x16_BEGIN:
mr BO, B
DSTRM_LT_L2x16_LOOP_START:
INIT_16x2
addic. L, KK, 0
ble DSTRM_LT_L2x16_SAVE
DSTRM_LT_L2x16_LOOP:
KERNEL_16x2
addic. L, L, -1
bgt DSTRM_LT_L2x16_LOOP
DSTRM_LT_L2x16_SAVE:
SOLVE_LT_16x2
addi CO, CO, 16*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 4+BASE_SHIFT
slwi T4, T4, 1+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 16
addic. I, I, -1
bgt DSTRM_LT_L2x16_BEGIN
DSTRM_LT_L2x16_END:
DSTRM_LT_L2x8_BEGIN:
andi. T2, M, 15
ble DSTRM_LT_L2x1_END
andi. T1, M, 8
ble DSTRM_LT_L2x8_END
mr BO, B
DSTRM_LT_L2x8_LOOP_START:
INIT_8x2
addic. L, KK, 0
ble DSTRM_LT_L2x8_SAVE
DSTRM_LT_L2x8_LOOP:
KERNEL_8x2
addic. L, L, -1
bgt DSTRM_LT_L2x8_LOOP
DSTRM_LT_L2x8_SAVE:
SOLVE_LT_8x2
addi CO, CO, 8*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 3+BASE_SHIFT
slwi T4, T4, 1+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 8
DSTRM_LT_L2x8_END:
DSTRM_LT_L2x4_BEGIN:
andi. T1, M, 4
ble DSTRM_LT_L2x4_END
mr BO, B
DSTRM_LT_L2x4_LOOP_START:
INIT_4x2
addic. L, KK, 0
ble DSTRM_LT_L2x4_SAVE
DSTRM_LT_L2x4_LOOP:
KERNEL_4x2
addic. L, L, -1
bgt DSTRM_LT_L2x4_LOOP
DSTRM_LT_L2x4_SAVE:
SOLVE_LT_4x2
addi CO, CO, 4*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 2+BASE_SHIFT
slwi T4, T4, 1+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 4
DSTRM_LT_L2x4_END:
DSTRM_LT_L2x2_BEGIN:
andi. T1, M, 2
ble DSTRM_LT_L2x2_END
mr BO, B
DSTRM_LT_L2x2_LOOP_START:
INIT_2x2
addic. L, KK, 0
ble DSTRM_LT_L2x2_SAVE
DSTRM_LT_L2x2_LOOP:
KERNEL_2x2
addic. L, L, -1
bgt DSTRM_LT_L2x2_LOOP
DSTRM_LT_L2x2_SAVE:
SOLVE_LT_2x2
addi CO, CO, 2*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 1+BASE_SHIFT
slwi T4, T4, 1+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 2
DSTRM_LT_L2x2_END:
DSTRM_LT_L2x1_BEGIN:
andi. T1, M, 1
ble DSTRM_LT_L2x1_END
mr BO, B
DSTRM_LT_L2x1_LOOP_START:
INIT_1x2
addic. L, KK, 0
ble DSTRM_LT_L2x1_SAVE
DSTRM_LT_L2x1_LOOP:
KERNEL_1x2
addic. L, L, -1
bgt DSTRM_LT_L2x1_LOOP
DSTRM_LT_L2x1_SAVE:
SOLVE_LT_1x2
addi CO, CO, 1*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 0+BASE_SHIFT
slwi T4, T4, 1+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 1
DSTRM_LT_L2x1_END:
slwi T1, K, 1+BASE_SHIFT
add B, B, T1
DSTRM_LT_L2_END:
DSTRM_LT_L1_BEGIN:
andi. T1, N, 1
ble DSTRM_LT_L1_END
mr CO, C
mr AO, A
mr KK, OFFSET
srawi. I, M, 4
ble DSTRM_LT_L1x16_END
DSTRM_LT_L1x16_BEGIN:
mr BO, B
DSTRM_LT_L1x16_LOOP_START:
INIT_16x1
addic. L, KK, 0
ble DSTRM_LT_L1x16_SAVE
DSTRM_LT_L1x16_LOOP:
KERNEL_16x1
addic. L, L, -1
bgt DSTRM_LT_L1x16_LOOP
DSTRM_LT_L1x16_SAVE:
SOLVE_LT_16x1
addi CO, CO, 16*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 4+BASE_SHIFT
slwi T4, T4, 0+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 16
addic. I, I, -1
bgt DSTRM_LT_L1x16_BEGIN
DSTRM_LT_L1x16_END:
DSTRM_LT_L1x8_BEGIN:
andi. T1, M, 8
ble DSTRM_LT_L1x8_END
mr BO, B
DSTRM_LT_L1x8_LOOP_START:
INIT_8x1
addic. L, KK, 0
ble DSTRM_LT_L1x8_SAVE
DSTRM_LT_L1x8_LOOP:
KERNEL_8x1
addic. L, L, -1
bgt DSTRM_LT_L1x8_LOOP
DSTRM_LT_L1x8_SAVE:
SOLVE_LT_8x1
addi CO, CO, 8*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 3+BASE_SHIFT
slwi T4, T4, 0+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 8
DSTRM_LT_L1x8_END:
DSTRM_LT_L1x4_BEGIN:
andi. T1, M, 4
ble DSTRM_LT_L1x4_END
mr BO, B
DSTRM_LT_L1x4_LOOP_START:
INIT_4x1
addic. L, KK, 0
ble DSTRM_LT_L1x4_SAVE
DSTRM_LT_L1x4_LOOP:
KERNEL_4x1
addic. L, L, -1
bgt DSTRM_LT_L1x4_LOOP
DSTRM_LT_L1x4_SAVE:
SOLVE_LT_4x1
addi CO, CO, 4*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 2+BASE_SHIFT
slwi T4, T4, 0+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 4
DSTRM_LT_L1x4_END:
DSTRM_LT_L1x2_BEGIN:
andi. T1, M, 2
ble DSTRM_LT_L1x2_END
mr BO, B
DSTRM_LT_L1x2_LOOP_START:
INIT_2x1
addic. L, KK, 0
ble DSTRM_LT_L1x2_SAVE
DSTRM_LT_L1x2_LOOP:
KERNEL_2x1
addic. L, L, -1
bgt DSTRM_LT_L1x2_LOOP
DSTRM_LT_L1x2_SAVE:
SOLVE_LT_2x1
addi CO, CO, 2*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 1+BASE_SHIFT
slwi T4, T4, 0+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 2
DSTRM_LT_L1x2_END:
DSTRM_LT_L1x1_BEGIN:
andi. T1, M, 1
ble DSTRM_LT_L1x1_END
mr BO, B
DSTRM_LT_L1x1_LOOP_START:
INIT_1x1
addic. L, KK, 0
ble DSTRM_LT_L1x1_SAVE
DSTRM_LT_L1x1_LOOP:
KERNEL_1x1
addic. L, L, -1
bgt DSTRM_LT_L1x1_LOOP
DSTRM_LT_L1x1_SAVE:
SOLVE_LT_1x1
addi CO, CO, 1*SIZE
sub T3, K, KK
sub T4, K, KK
slwi T3, T3, 0+BASE_SHIFT
slwi T4, T4, 0+BASE_SHIFT
add AO, AO, T3
add BO, BO, T4
addi KK, KK, 1
DSTRM_LT_L1x1_END:
DSTRM_LT_L1_END: