Potential performance improvements for DSP-centric designs using numeric_std

[malles030]

There is quite some potential to notably speed up NVC for DSP-centric designs. Some of the numeric_std functions have JIT intrinsics (like unsigned comparisons) while many others have not (like signed comparisons). I wonder whether you are open to take pull requests implementing more of these functions as JIT intrinsics or whether there is something that would hold you back from accepting such pull requests? I'm aware that it would add quite a bit of code.

A first step could be to implement signed comparisons and the to_integer functions.

Here is the current status in NVC:

Id	Funktion	master
A.1	"abs"(SIGNED) return SIGNED	✗
A.2	"-"(SIGNED) return SIGNED (unary)	✗
A.3	"+"(UNSIGNED, UNSIGNED) return UNSIGNED	✓
A.4	"+"(SIGNED, SIGNED) return SIGNED	✓
A.5	"+"(UNSIGNED, NATURAL) return UNSIGNED	✓
A.6	"+"(NATURAL, UNSIGNED) return UNSIGNED	✓
A.7	"+"(INTEGER, SIGNED) return SIGNED	✗
A.8	"+"(SIGNED, INTEGER) return SIGNED	✗
A.9	"-"(UNSIGNED, UNSIGNED) return UNSIGNED	✓
A.10	"-"(SIGNED, SIGNED) return SIGNED	✓
A.11	"-"(UNSIGNED, NATURAL) return UNSIGNED	✗
A.12	"-"(NATURAL, UNSIGNED) return UNSIGNED	✗
A.13	"-"(SIGNED, INTEGER) return SIGNED	✗
A.14	"-"(INTEGER, SIGNED) return SIGNED	✗
A.15	"*"(UNSIGNED, UNSIGNED) return UNSIGNED	✓
A.16	"*"(SIGNED, SIGNED) return SIGNED	✓
A.17	"*"(UNSIGNED, NATURAL) return UNSIGNED	✗
A.18	"*"(NATURAL, UNSIGNED) return UNSIGNED	✗
A.19	"*"(SIGNED, INTEGER) return SIGNED	✗
A.20	"*"(INTEGER, SIGNED) return SIGNED	✗
A.21	"/"(UNSIGNED, UNSIGNED) return UNSIGNED	✗
A.22	"/"(SIGNED, SIGNED) return SIGNED	✗
A.23	"/"(UNSIGNED, NATURAL) return UNSIGNED	✗
A.24	"/"(NATURAL, UNSIGNED) return UNSIGNED	✗
A.25	"/"(SIGNED, INTEGER) return SIGNED	✗
A.26	"/"(INTEGER, SIGNED) return SIGNED	✗
A.27	"rem"(UNSIGNED, UNSIGNED) return UNSIGNED	✗
A.28	"rem"(SIGNED, SIGNED) return SIGNED	✗
A.29	"rem"(UNSIGNED, NATURAL) return UNSIGNED	✗
A.30	"rem"(NATURAL, UNSIGNED) return UNSIGNED	✗
A.31	"rem"(SIGNED, INTEGER) return SIGNED	✗
A.32	"rem"(INTEGER, SIGNED) return SIGNED	✗
A.33	"mod"(UNSIGNED, UNSIGNED) return UNSIGNED	✗
A.34	"mod"(SIGNED, SIGNED) return SIGNED	✗
A.35	"mod"(UNSIGNED, NATURAL) return UNSIGNED	✗
A.36	"mod"(NATURAL, UNSIGNED) return UNSIGNED	✗
A.37	"mod"(SIGNED, INTEGER) return SIGNED	✗
A.38	"mod"(INTEGER, SIGNED) return SIGNED	✗
C.1	">"(UNSIGNED, UNSIGNED) return BOOLEAN	✓
C.2	">"(SIGNED, SIGNED) return BOOLEAN	✗
C.3	">"(NATURAL, UNSIGNED) return BOOLEAN	✗
C.4	">"(INTEGER, SIGNED) return BOOLEAN	✗
C.5	">"(UNSIGNED, NATURAL) return BOOLEAN	✗
C.6	">"(SIGNED, INTEGER) return BOOLEAN	✗
C.7	"<"(UNSIGNED, UNSIGNED) return BOOLEAN	✓
C.8	"<"(SIGNED, SIGNED) return BOOLEAN	✗
C.9	"<"(NATURAL, UNSIGNED) return BOOLEAN	✗
C.10	"<"(INTEGER, SIGNED) return BOOLEAN	✗
C.11	"<"(UNSIGNED, NATURAL) return BOOLEAN	✗
C.12	"<"(SIGNED, INTEGER) return BOOLEAN	✗
C.13	"<="(UNSIGNED, UNSIGNED) return BOOLEAN	✓
C.14	"<="(SIGNED, SIGNED) return BOOLEAN	✗
C.15	"<="(NATURAL, UNSIGNED) return BOOLEAN	✗
C.16	"<="(INTEGER, SIGNED) return BOOLEAN	✗
C.17	"<="(UNSIGNED, NATURAL) return BOOLEAN	✗
C.18	"<="(SIGNED, INTEGER) return BOOLEAN	✗
C.19	">="(UNSIGNED, UNSIGNED) return BOOLEAN	✓
C.20	">="(SIGNED, SIGNED) return BOOLEAN	✗
C.21	">="(NATURAL, UNSIGNED) return BOOLEAN	✗
C.22	">="(INTEGER, SIGNED) return BOOLEAN	✗
C.23	">="(UNSIGNED, NATURAL) return BOOLEAN	✗
C.24	">="(SIGNED, INTEGER) return BOOLEAN	✗
C.25	"="(UNSIGNED, UNSIGNED) return BOOLEAN	✗
C.26	"="(SIGNED, SIGNED) return BOOLEAN	✗
C.27	"="(NATURAL, UNSIGNED) return BOOLEAN	✗
C.28	"="(INTEGER, SIGNED) return BOOLEAN	✗
C.29	"="(UNSIGNED, NATURAL) return BOOLEAN	✗
C.30	"="(SIGNED, INTEGER) return BOOLEAN	✗
C.31	"/="(UNSIGNED, UNSIGNED) return BOOLEAN	✗
C.32	"/="(SIGNED, SIGNED) return BOOLEAN	✗
C.33	"/="(NATURAL, UNSIGNED) return BOOLEAN	✗
C.34	"/="(INTEGER, SIGNED) return BOOLEAN	✗
C.35	"/="(UNSIGNED, NATURAL) return BOOLEAN	✗
C.36	"/="(SIGNED, INTEGER) return BOOLEAN	✗
S.1	SHIFT_LEFT(UNSIGNED, NATURAL)	✗
S.2	SHIFT_RIGHT(UNSIGNED, NATURAL)	✗
S.3	SHIFT_LEFT(SIGNED, NATURAL)	✗
S.4	SHIFT_RIGHT(SIGNED, NATURAL)	✗
S.5	ROTATE_LEFT(UNSIGNED, NATURAL)	✗
S.6	ROTATE_RIGHT(UNSIGNED, NATURAL)	✗
S.7	ROTATE_LEFT(SIGNED, NATURAL)	✗
S.8	ROTATE_RIGHT(SIGNED, NATURAL)	✗
S.9	"sll"(UNSIGNED, INTEGER)	✗
S.10	"sll"(SIGNED, INTEGER)	✗
S.11	"srl"(UNSIGNED, INTEGER)	✗
S.12	"srl"(SIGNED, INTEGER)	✗
S.13	"rol"(UNSIGNED, INTEGER)	✗
S.14	"rol"(SIGNED, INTEGER)	✗
S.15	"ror"(UNSIGNED, INTEGER)	✗
S.16	"ror"(SIGNED, INTEGER)	✗
R.1	RESIZE(SIGNED, NATURAL) return SIGNED	✓
R.2	RESIZE(UNSIGNED, NATURAL) return UNSIGNED	✓
D.1	TO_INTEGER(UNSIGNED) return NATURAL	✗
D.2	TO_INTEGER(SIGNED) return INTEGER	✗
D.3	TO_UNSIGNED(NATURAL, NATURAL) return UNSIGNED	✓
D.4	TO_SIGNED(INTEGER, NATURAL) return SIGNED	✓
L.1	"not"(UNSIGNED)	✗
L.2	"and"(UNSIGNED, UNSIGNED)	✗
L.3	"or"(UNSIGNED, UNSIGNED)	✗
L.4	"nand"(UNSIGNED, UNSIGNED)	✗
L.5	"nor"(UNSIGNED, UNSIGNED)	✗
L.6	"xor"(UNSIGNED, UNSIGNED)	✗
L.7	"xnor"(UNSIGNED, UNSIGNED)	✗
L.8	"not"(SIGNED)	✗
L.9	"and"(SIGNED, SIGNED)	✗
L.10	"or"(SIGNED, SIGNED)	✗
L.11	"nand"(SIGNED, SIGNED)	✗
L.12	"nor"(SIGNED, SIGNED)	✗
L.13	"xor"(SIGNED, SIGNED)	✗
L.14	"xnor"(SIGNED, SIGNED)	✗
T.1	TO_01(UNSIGNED, STD_LOGIC) return UNSIGNED	✓
T.2	TO_01(SIGNED, STD_LOGIC) return SIGNED	✓

[nickg]

I prefer not to add too many of these unless it shows a significant speed-up in real benchmarks. They can be tricky to get right with lots of edge cases and they bake-in a lot of details about how the JIT works and its memory layout which makes maintenance difficult.

A lot of these like the relational operators just call through to the std_logic_vector operators which already have intrinsics so I'm not sure there's much to be gained there.

That said signed/unsigned to_integer is definitely worthwhile and I just added that.

[malles030]

Thanks for implementing and for the clarification. I was thinking that A.1 to A.20 were good candidates, but looking into the body of numeric_std I see that the functions that take a bit more effort already have intrinsics. The only exception being A.1 (abs) and A.2 (unary '-'), which should be worth the effort, IMHO.

For the comparisons: All functions that have signed and integer or unsigned and natural as inputs call signed_num_bits and unsigned_num_bits, respectively. Both functions run loops to determine the bit width of the natural/integer parameter. I think in particular these are good candidates, too. Pretty common use-case for these functions are boundary checks or saturation.

Hmm, and the existing unsigned/unsigned comparison intrinsics must have some benefit, too? Most algorithms in signal processing use signed, though. So there is probably also some gain to expect.

I also think that the remaining functions are not that important.

[bpadalino]

I don't want to highjack the discussion, but are the same conversions/intrinsics being used with numeric_bit?

[nickg]

I don't want to highjack the discussion, but are the same conversions/intrinsics being used with numeric_bit?

Not at the moment. There's no reason they couldn't be added but I never see people using numeric_bit.