Skip to content

Commit

Permalink
Merge pull request #1266 from ashwinyes/develop_thunderx2t99_fix_clan…
Browse files Browse the repository at this point in the history 
…g_compilation

THUDNERX2T99: Fix clang compilation
  • Loading branch information
@martin-frbg
martin-frbg authored Aug 3, 2017
2 parents ebb04e3 + 4899d67 commit 0bfe5c6
Show file tree
Hide file tree
Showing 11 changed files with 244 additions and 244 deletions.
42 changes: 21 additions & 21 deletions kernel/arm64/casum_thunderx2t99.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -147,57 +147,57 @@ static FLOAT casum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
" fmov s6, "REG0" \n"
" fmov s7, "REG0" \n"
" cmp "N", xzr \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne .Lasum_kernel_S_BEGIN \n"
" bne 5f //asum_kernel_S_BEGIN \n"

".Lasum_kernel_F_BEGIN: \n"
"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #5 \n"
" cmp "J", xzr \n"
" beq .Lasum_kernel_F1 \n"
" beq 3f //asum_kernel_F1 \n"

".Lasum_kernel_F32: \n"
"2: //asum_kernel_F32: \n"
" "KERNEL_F32" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_F32 \n"
" bne 2b //asum_kernel_F32 \n"
" "KERNEL_F32_FINALIZE" \n"

".Lasum_kernel_F1: \n"
"3: //asum_kernel_F1: \n"
" ands "J", "N", #31 \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"

".Lasum_kernel_F10: \n"
"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_F10 \n"
" b .Lasum_kernel_L999 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"

".Lasum_kernel_S_BEGIN: \n"
"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble .Lasum_kernel_S1 \n"
" ble 7f //asum_kernel_S1 \n"

".Lasum_kernel_S4: \n"
"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_S4 \n"
" bne 6b //asum_kernel_S4 \n"

".Lasum_kernel_S1: \n"
"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"

".Lasum_kernel_S10: \n"
"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_S10 \n"
" bne 8b //asum_kernel_S10 \n"

".Lasum_kernel_L999: \n"
"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMFD" \n"

: [ASUM_] "=r" (asum) //%0
Expand Down
44 changes: 22 additions & 22 deletions kernel/arm64/copy_thunderx2t99.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -90,62 +90,62 @@ static int do_copy(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_
" mov "Y", %[Y_] \n"
" mov "INC_Y", %[INCY_] \n"
" cmp "N", xzr \n"
" ble .Lcopy_kernel_L999 \n"
" ble 8f //copy_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne .Lcopy_kernel_S_BEGIN \n"
" bne 4f //copy_kernel_S_BEGIN \n"
" cmp "INC_Y", #1 \n"
" bne .Lcopy_kernel_S_BEGIN \n"
" bne 4f //copy_kernel_S_BEGIN \n"

".Lcopy_kernel_F_BEGIN: \n"
"// .Lcopy_kernel_F_BEGIN: \n"
" "INIT" \n"
" asr "J", "N", #"N_DIV_SHIFT" \n"
" cmp "J", xzr \n"
" beq .Lcopy_kernel_F1 \n"
" beq 2f //copy_kernel_F1 \n"
" .align 5 \n"

".Lcopy_kernel_F: \n"
"1: //copy_kernel_F: \n"
" "KERNEL_F" \n"
" subs "J", "J", #1 \n"
" bne .Lcopy_kernel_F \n"
" bne 1b //copy_kernel_F \n"

".Lcopy_kernel_F1: \n"
"2: //copy_kernel_F1: \n"
#if defined(COMPLEX) && defined(DOUBLE)
" b .Lcopy_kernel_L999 \n"
" b 8f //copy_kernel_L999 \n"
#else
" ands "J", "N", #"N_REM_MASK" \n"
" ble .Lcopy_kernel_L999 \n"
" ble 8f //copy_kernel_L999 \n"
#endif

".Lcopy_kernel_F10: \n"
"3: //copy_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Lcopy_kernel_F10 \n"
" b .Lcopy_kernel_L999 \n"
" bne 3b //copy_kernel_F10 \n"
" b 8f //copy_kernel_L999 \n"

".Lcopy_kernel_S_BEGIN: \n"
"4: //copy_kernel_S_BEGIN: \n"
" "INIT" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble .Lcopy_kernel_S1 \n"
" ble 6f //copy_kernel_S1 \n"

".Lcopy_kernel_S4: \n"
"5: //copy_kernel_S4: \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Lcopy_kernel_S4 \n"
" bne 5b //copy_kernel_S4 \n"

".Lcopy_kernel_S1: \n"
"6: //copy_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble .Lcopy_kernel_L999 \n"
" ble 8f //copy_kernel_L999 \n"

".Lcopy_kernel_S10: \n"
"7: //copy_kernel_S10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Lcopy_kernel_S10 \n"
" bne 7b //copy_kernel_S10 \n"

".Lcopy_kernel_L999: \n"
"8: //copy_kernel_L999: \n"

:
: [N_] "r" (n), //%1
Expand Down
42 changes: 21 additions & 21 deletions kernel/arm64/dasum_thunderx2t99.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -141,58 +141,58 @@ static FLOAT dasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
" fmov d6, "REG0" \n"
" fmov d7, "REG0" \n"
" cmp "N", xzr \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne .Lasum_kernel_S_BEGIN \n"
" bne 5f //asum_kernel_S_BEGIN \n"

".Lasum_kernel_F_BEGIN: \n"
"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #5 \n"
" cmp "J", xzr \n"
" beq .Lasum_kernel_F1 \n"
" beq 3f //asum_kernel_F1 \n"

".align 5 \n"
".Lasum_kernel_F32: \n"
"2: //asum_kernel_F32: \n"
" "KERNEL_F32" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_F32 \n"
" bne 2b //asum_kernel_F32 \n"
" "KERNEL_F32_FINALIZE" \n"

".Lasum_kernel_F1: \n"
"3: //asum_kernel_F1: \n"
" ands "J", "N", #31 \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"

".Lasum_kernel_F10: \n"
"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_F10 \n"
" b .Lasum_kernel_L999 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"

".Lasum_kernel_S_BEGIN: \n"
"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble .Lasum_kernel_S1 \n"
" ble 7f //asum_kernel_S1 \n"

".Lasum_kernel_S4: \n"
"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_S4 \n"
" bne 6b //asum_kernel_S4 \n"

".Lasum_kernel_S1: \n"
"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble .Lasum_kernel_L999 \n"
" ble 9f //asum_kernel_L999 \n"

".Lasum_kernel_S10: \n"
"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne .Lasum_kernel_S10 \n"
" bne 8b //asum_kernel_S10 \n"

".Lasum_kernel_L999: \n"
"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMF" \n"

: [ASUM_] "=r" (asum) //%0
Expand Down
42 changes: 21 additions & 21 deletions kernel/arm64/dot_thunderx2t99.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -291,61 +291,61 @@ static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, B
" fmov d6, xzr \n"
" fmov d7, xzr \n"
" cmp "N", xzr \n"
" ble .Ldot_kernel_L999 \n"
" ble 9f //dot_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne .Ldot_kernel_S_BEGIN \n"
" bne 5f //dot_kernel_S_BEGIN \n"
" cmp "INC_Y", #1 \n"
" bne .Ldot_kernel_S_BEGIN \n"
" bne 5f //dot_kernel_S_BEGIN \n"

".Ldot_kernel_F_BEGIN: \n"
"1: //dot_kernel_F_BEGIN: \n"
" lsl "INC_X", "INC_X", "INC_SHIFT" \n"
" lsl "INC_Y", "INC_Y", "INC_SHIFT" \n"
" asr "J", "N", #"N_DIV_SHIFT" \n"
" cmp "J", xzr \n"
" beq .Ldot_kernel_F1 \n"
" beq 3f //dot_kernel_F1 \n"

" .align 5 \n"
".Ldot_kernel_F: \n"
"2: //dot_kernel_F: \n"
" "KERNEL_F" \n"
" subs "J", "J", #1 \n"
" bne .Ldot_kernel_F \n"
" bne 2b //dot_kernel_F \n"
" "KERNEL_F_FINALIZE" \n"

".Ldot_kernel_F1: \n"
"3: //dot_kernel_F1: \n"
" ands "J", "N", #"N_REM_MASK" \n"
" ble .Ldot_kernel_L999 \n"
" ble 9f //dot_kernel_L999 \n"

".Ldot_kernel_F10: \n"
"4: //dot_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Ldot_kernel_F10 \n"
" b .Ldot_kernel_L999 \n"
" bne 4b //dot_kernel_F10 \n"
" b 9f //dot_kernel_L999 \n"

".Ldot_kernel_S_BEGIN: \n"
"5: //dot_kernel_S_BEGIN: \n"
" lsl "INC_X", "INC_X", "INC_SHIFT" \n"
" lsl "INC_Y", "INC_Y", "INC_SHIFT" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble .Ldot_kernel_S1 \n"
" ble 7f //dot_kernel_S1 \n"

".Ldot_kernel_S4: \n"
"6: //dot_kernel_S4: \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Ldot_kernel_S4 \n"
" bne 6b //dot_kernel_S4 \n"

".Ldot_kernel_S1: \n"
"7: //dot_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble .Ldot_kernel_L999 \n"
" ble 9f //dot_kernel_L999 \n"

".Ldot_kernel_S10: \n"
"8: //dot_kernel_S10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne .Ldot_kernel_S10 \n"
" bne 8b //dot_kernel_S10 \n"

".Ldot_kernel_L999: \n"
"9: //dot_kernel_L999: \n"
" str "DOTF", [%[DOT_]] \n"

:
Expand Down
Loading

0 comments on commit 0bfe5c6

Please sign in to comment.