cores
/
microwatt
mirror of https://github.com/antonblanchard/microwatt
3
1
Fork 1
Code Issues Projects Releases Wiki Activity

FPU: Avoid doing overflow processing twice in OE=1 case

Browse Source 
Split the ROUND_OFLOW state into two, one which handles the OE=0 case
(disabled overflow exception) and one which handles the OE=1 case
(enabled overflow exception).  This avoids a loop in the state diagram
and prevents us from adding the exponent bias twice.

Also correct a bug in ROUNDING_3 state where for single-precision
operations which yield a result which is denormal in double-precision
format, r.shift was set wrongly.

Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
pull/457/head
Paul Mackerras 2 weeks ago
parent 9f27f60b26
commit 59992eab90
3 changed files with 55 additions and 57 deletions
Show Stats Download Patch File Download Diff File

109
fpu.vhdl
Unescape Escape View File

@ -72,7 +72,7 @@ architecture behaviour of fpu is
INT_SHIFT, INT_ROUND, INT_ISHIFT, INT_SHIFT, INT_ROUND, INT_ISHIFT,
INT_FINAL, INT_CHECK, INT_OFLOW, INT_FINAL, INT_CHECK, INT_OFLOW,
FINISH, NORMALIZE, FINISH, NORMALIZE,
ROUND_UFLOW, NORM_UFLOW, ROUND_OFLOW, ROUND_UFLOW, NORM_UFLOW, ROUND_OFLOW_DIS, ROUND_OFLOW_EN,
ROUNDING, ROUND_INC, ROUNDING_2, ROUNDING_3, ROUNDING, ROUND_INC, ROUNDING_2, ROUNDING_3,
DENORM, DENORM,
RENORM_A, RENORM_B, RENORM_C, RENORM_A, RENORM_B, RENORM_C,
@ -315,6 +315,7 @@ architecture behaviour of fpu is
constant RSCON2_63 : std_ulogic_vector(3 downto 0) := "0111"; constant RSCON2_63 : std_ulogic_vector(3 downto 0) := "0111";
constant RSCON2_64 : std_ulogic_vector(3 downto 0) := "1000"; constant RSCON2_64 : std_ulogic_vector(3 downto 0) := "1000";
constant RSCON2_MINEXP : std_ulogic_vector(3 downto 0) := "1001"; constant RSCON2_MINEXP : std_ulogic_vector(3 downto 0) := "1001";
constant RSCON2_DPMINX : std_ulogic_vector(3 downto 0) := "1010";


signal rs_sel1 : std_ulogic_vector(1 downto 0); signal rs_sel1 : std_ulogic_vector(1 downto 0);
signal rs_sel2 : std_ulogic; signal rs_sel2 : std_ulogic;
@ -1633,10 +1634,10 @@ begin
rs_con2 <= RSCON2_MINEXP; rs_con2 <= RSCON2_MINEXP;
rs_neg2 <= '1'; rs_neg2 <= '1';
set_x := '1'; -- uses r.r and r.shift set_x := '1'; -- uses r.r and r.shift
if r.result_exp < to_signed(-126, EXP_BITS) then if exp_tiny = '1' then
v.state := ROUND_UFLOW; v.state := ROUND_UFLOW;
elsif r.result_exp > to_signed(127, EXP_BITS) then elsif exp_huge = '1' and r.fpscr(FPSCR_OE) = '0' then
v.state := ROUND_OFLOW; v.state := ROUND_OFLOW_DIS;
else else
v.state := ROUNDING; v.state := ROUNDING;
end if; end if;
@ -2406,6 +2407,7 @@ begin
v.state := ROUNDING; v.state := ROUNDING;


when FINISH => when FINISH =>
-- r.shift = 0
if r.is_multiply = '1' and px_nz = '1' then if r.is_multiply = '1' and px_nz = '1' then
v.x := '1'; v.x := '1';
end if; end if;
@ -2420,8 +2422,8 @@ begin
set_x := '1'; set_x := '1';
if exp_tiny = '1' then if exp_tiny = '1' then
v.state := ROUND_UFLOW; v.state := ROUND_UFLOW;
elsif exp_huge = '1' then elsif exp_huge = '1' and r.fpscr(FPSCR_OE) = '0' then
v.state := ROUND_OFLOW; v.state := ROUND_OFLOW_DIS;
else else
v.state := ROUNDING; v.state := ROUNDING;
end if; end if;
@ -2441,8 +2443,8 @@ begin
set_x := '1'; set_x := '1';
if exp_tiny = '1' then if exp_tiny = '1' then
v.state := ROUND_UFLOW; v.state := ROUND_UFLOW;
elsif exp_huge = '1' then elsif exp_huge = '1' and r.fpscr(FPSCR_OE) = '0' then
v.state := ROUND_OFLOW; v.state := ROUND_OFLOW_DIS;
else else
v.state := ROUNDING; v.state := ROUNDING;
end if; end if;
@ -2485,30 +2487,20 @@ begin
set_x := '1'; set_x := '1';
v.state := ROUNDING; v.state := ROUNDING;


when ROUND_OFLOW => when ROUND_OFLOW_DIS =>
-- disabled overflow exception
-- result depends on rounding mode
rcls_op <= RCLS_TINF; rcls_op <= RCLS_TINF;
v.fpscr(FPSCR_OX) := '1'; v.fpscr(FPSCR_OX) := '1';
opsel_r <= RES_MISC; opsel_r <= RES_MISC;
misc_sel <= "010"; misc_sel <= "010";
set_r := '0'; set_r := '1';
if r.fpscr(FPSCR_OE) = '0' then v.fpscr(FPSCR_XX) := '1';
-- disabled overflow exception v.fpscr(FPSCR_FI) := '1';
-- result depends on rounding mode -- construct largest representable number
set_r := '1'; re_con2 <= RECON2_MAX;
v.fpscr(FPSCR_XX) := '1'; re_set_result <= '1';
v.fpscr(FPSCR_FI) := '1'; arith_done := '1';
-- construct largest representable number
re_con2 <= RECON2_MAX;
re_set_result <= '1';
arith_done := '1';
else
-- enabled overflow exception
re_sel1 <= REXP1_R;
re_con2 <= RECON2_BIAS;
re_neg2 <= '1';
re_set_result <= '1';
v.state := ROUNDING;
end if;


when ROUNDING => when ROUNDING =>
opsel_mask <= '1'; opsel_mask <= '1';
@ -2527,6 +2519,8 @@ begin
-- denormalized result that needs to be renormalized -- denormalized result that needs to be renormalized
rs_norm <= '1'; rs_norm <= '1';
v.state := ROUNDING_3; v.state := ROUNDING_3;
elsif r.result_exp > max_exp then
v.state := ROUND_OFLOW_EN;
else else
arith_done := '1'; arith_done := '1';
end if; end if;
@ -2540,49 +2534,40 @@ begin
when ROUND_INC => when ROUND_INC =>
set_r := '1'; set_r := '1';
opsel_a <= AIN_RND; opsel_a <= AIN_RND;
-- set shift to -1
rs_con2 <= RSCON2_1;
rs_neg2 <= '1';
v.state := ROUNDING_2; v.state := ROUNDING_2;


when ROUNDING_2 => when ROUNDING_2 =>
-- Check for overflow during rounding -- Check for overflow during rounding
-- r.shift = -1 -- r.shift = 0
v.x := '0'; if r.r(UNIT_BIT + 1) = '1' or r.r(UNIT_BIT) = '0' then
re_sel2 <= REXP2_NE;
opsel_r <= RES_SHIFT;
set_r := '0';
if r.r(UNIT_BIT + 1) = '1' then
set_r := '1';
re_set_result <= '1';
if exp_huge = '1' then
v.state := ROUND_OFLOW;
else
arith_done := '1';
end if;
elsif r.r(UNIT_BIT) = '0' then
-- Do CLZ so we can renormalize the result -- Do CLZ so we can renormalize the result
rs_norm <= '1'; rs_norm <= '1';
v.state := ROUNDING_3; v.state := ROUNDING_3;
elsif exp_huge = '1' then
v.state := ROUND_OFLOW_EN;
else else
arith_done := '1'; arith_done := '1';
end if; end if;


when ROUNDING_3 => when ROUNDING_3 =>
-- r.shift = clz(r.r) - 9 -- r.shift = clz(r.r) - 7
opsel_r <= RES_SHIFT; opsel_r <= RES_SHIFT;
set_r := '1'; set_r := '1';
re_sel2 <= REXP2_NE; re_sel2 <= REXP2_NE;
-- set shift to new_exp - min_exp (== -1022) -- set shift to new_exp - DP min_exp (== -1022)
rs_sel1 <= RSH1_NE; rs_sel1 <= RSH1_NE;
rs_con2 <= RSCON2_MINEXP; rs_con2 <= RSCON2_DPMINX;
rs_neg2 <= '1'; rs_neg2 <= '1';
rcls_op <= RCLS_TZERO; rcls_op <= RCLS_TZERO;
-- If the result is zero, that's handled below. -- If the result is zero, that's handled below.
-- Renormalize result after rounding -- Renormalize result after rounding
v.denorm := exp_tiny; v.denorm := exp_tiny;
re_set_result <= '1'; re_set_result <= '1';
if new_exp < to_signed(-1022, EXP_BITS) then if exp_huge = '1' and r.fpscr(FPSCR_OE) = '0' then
v.state := ROUND_OFLOW_DIS;
elsif exp_huge = '1' and r.fpscr(FPSCR_OE) = '1' then
v.state := ROUND_OFLOW_EN;
elsif new_exp < to_signed(-1022, EXP_BITS) then
v.state := DENORM; v.state := DENORM;
else else
arith_done := '1'; arith_done := '1';
@ -2596,6 +2581,16 @@ begin
re_set_result <= '1'; re_set_result <= '1';
arith_done := '1'; arith_done := '1';


when ROUND_OFLOW_EN =>
-- enabled overflow exception
-- rounding and normalization has been done
v.fpscr(FPSCR_OX) := '1';
re_sel1 <= REXP1_R;
re_con2 <= RECON2_BIAS;
re_neg2 <= '1';
re_set_result <= '1';
arith_done := '1';

when DO_IDIVMOD => when DO_IDIVMOD =>
opsel_a <= AIN_B; opsel_a <= AIN_B;
opsel_aabs <= '1'; opsel_aabs <= '1';
@ -3201,14 +3196,12 @@ begin
arith_done := '1'; arith_done := '1';
end if; end if;
when RCLS_TINF => when RCLS_TINF =>
if r.fpscr(FPSCR_OE) = '0' then if r.round_mode(1 downto 0) = "00" or
if r.round_mode(1 downto 0) = "00" or (r.round_mode(1) = '1' and r.round_mode(0) = r.result_sign) then
(r.round_mode(1) = '1' and r.round_mode(0) = r.result_sign) then v.result_class := INFINITY;
v.result_class := INFINITY; v.fpscr(FPSCR_FR) := '1';
v.fpscr(FPSCR_FR) := '1'; else
else v.fpscr(FPSCR_FR) := '0';
v.fpscr(FPSCR_FR) := '0';
end if;
end if; end if;
when others => when others =>
end case; end case;
@ -3593,6 +3586,8 @@ begin
rsh_in2 := to_signed(64, EXP_BITS); rsh_in2 := to_signed(64, EXP_BITS);
when RSCON2_MINEXP => when RSCON2_MINEXP =>
rsh_in2 := min_exp; rsh_in2 := min_exp;
when RSCON2_DPMINX =>
rsh_in2 := to_signed(-1022, EXP_BITS);
when others => when others =>
rsh_in2 := to_signed(0, EXP_BITS); rsh_in2 := to_signed(0, EXP_BITS);
end case; end case;

3
tests/fpu/fpu.c
Unescape Escape View File

@ -682,6 +682,9 @@ struct roundvals {
{ FPS_RN_NEAR, 0x37c12345b0000000, 0x37c1234400000000, FPS_FI }, { FPS_RN_NEAR, 0x37c12345b0000000, 0x37c1234400000000, FPS_FI },
{ FPS_RN_NEAR, 0x0000008800000088, 0, FPS_FI }, { FPS_RN_NEAR, 0x0000008800000088, 0, FPS_FI },
{ FPS_RN_NEAR, 0xc2000000c2000000, 0xc2000000c0000000, FPS_FI }, { FPS_RN_NEAR, 0xc2000000c2000000, 0xc2000000c0000000, FPS_FI },
{ FPS_RN_NEAR|FPS_OE, 0xefffffffffffffff, 0xe400000000000000, FPS_FR|FPS_FI },
{ FPS_RN_NEAR|FPS_OE, 0xff0000ff43434343, 0xf30000ff40000000, FPS_FI },
{ FPS_RN_NEAR|FPS_OE, 0xfc00fc0139fffcff, 0xf000fc0140000000, FPS_FR|FPS_FI },
}; };


int test8(long arg) int test8(long arg)

BIN
tests/test_fpu.bin
View File

Binary file not shown.
Write Preview
Loading…
Cancel
Save