[Mesa-dev] [RFC 1/3] gk110/ir: Add rcp f64 implementation

Boyan Ding boyan.j.ding at gmail.com
Mon Mar 6 09:18:47 UTC 2017


2017-03-06 0:33 GMT+08:00 Ilia Mirkin <imirkin at alum.mit.edu>:
> On Sun, Mar 5, 2017 at 10:34 AM, Boyan Ding <boyan.j.ding at gmail.com> wrote:
>> Signed-off-by: Boyan Ding <boyan.j.ding at gmail.com>
>> ---
>>  src/gallium/drivers/nouveau/codegen/lib/gk110.asm  | 156 ++++++++++++++++++++-
>>  .../drivers/nouveau/codegen/lib/gk110.asm.h        |  96 ++++++++++++-
>>  2 files changed, 248 insertions(+), 4 deletions(-)
>>
>> diff --git a/src/gallium/drivers/nouveau/codegen/lib/gk110.asm b/src/gallium/drivers/nouveau/codegen/lib/gk110.asm
>> index b9c05a04b9..871571e1c3 100644
>> --- a/src/gallium/drivers/nouveau/codegen/lib/gk110.asm
>> +++ b/src/gallium/drivers/nouveau/codegen/lib/gk110.asm
>> @@ -83,11 +83,165 @@ gk110_div_s32:
>>     $p0 sub b32 $r1 $r1 $r2
>>     $p0 add b32 $r0 $r0 0x1
>>     $p3 cvt s32 $r0 neg s32 $r0
>> -   sched 0x04 0x2e 0x04 0x28 0x04 0x20 0x2c
>> +   sched 0x04 0x2e 0x28 0x04 0x28 0x28 0x28
>>     $p2 cvt s32 $r1 neg s32 $r1
>>     ret
>>
>> +// RCP F64
>> +//
>> +// INPUT:   $r0d
>> +// OUTPUT:  $r0d
>> +// CLOBBER: $r2 - $r9, $p0
>> +//
>> +// The core of RCP and RSQ implementation is Newton-Raphson step, which is
>> +// used to find successively better approximation from an imprecise initial
>> +// value (single precision rcp in RCP and rsqrt64h in RSQ).
>> +//
>> +// The formula of Newton-Raphson step used in RCP(x) is:
>> +//     RCP_{n + 1} = 2 * RCP_{n} - x * RCP_{n} * RCP_{n}
>> +// The following code uses 2 FMAs for each step, and it will basically
>> +// look like:
>> +//     tmp = -src * RCP_{n} + 1
>> +//     RCP_{n + 1} = RCP_{n} * tmp + RCP_{n}
>> +//
>>  gk110_rcp_f64:
>> +   // Step1: classify input according to exponent and value, and calculate
>
> Step 1 (note the space).
>
>> +   // result for 0/inf/nan, $r2 holds the exponent value. The 0xb14 in ext
>> +   // instruction means to extract 0xb bits starting from bit 0x14
>
> You should assume that the reader is familiar with the nvidia ISA.
> (Otherwise there will be a LOT of explaining to do.) However it would
> be more valuable to note that this extracts the exponent bits, which
> may not be apparent to someone who hasn't been staring at fp64 values
> for the past 72 hours :)
>
>> +   ext u32 $r2 $r1 0xb14
>> +   add b32 $r3 $r2 0xffffffff
>> +   joinat #rcp_L3
>> +   // (exponent-1) > 0x7fd (unsigned) means exponent is either 0x7ff of 0.
>> +   // There are three cases: nan, inf, and denorm (including 0)
>> +   set b32 $p0 0x1 gt u32 $r3 0x7fd
>
> This is extremely clever. I'm a big fan. But I didn't get it the first
> time around. Perhaps I'm dumb, but I'd rather the comment be reworded
> for fools like me. Something like:
>
> "We want to check whether the exponent is 0 or 0x7ff (i.e. NaN, Inf,
> denorm, or 0). Do this by subtracting 1 from the exponent, which will
> mean that it's > 0x7fd in those cases."
>
>> +   // $r3: 0 for norms, 0x36 for denorms, -1 for others
>> +   mov b32 $r3 0x0
>> +   sched 0x2b 0x04 0x2d 0x2b 0x04 0x2b 0x28
>> +   (not $p0) bra #rcp_L2
>> +   // Nan/Inf/denorm goes here
>
> And zero too, right? How about "Process all the special values: ... "
>
>> +   mov b32 $r3 0xffffffff
>> +   // A number is NaN if its abs value is greater than inf
>
> Actually note that this is gtu, not gt. That means greater *or* unordered with.
>
>> +   set $p0 0x1 gtu f64 abs $r0d 0x7ff0000000000000
>> +   (not $p0) bra #rcp_L4
>> +   // NaN -> NaN
>
> Perhaps say a few more words here? Not entirely apparent to me how
> setting that bit makes it more of a NaN. (Wasn't it already a NaN in
> the first place if we got here?)
>
>> +   or b32 $r1 $r1 0x80000
>> +   bra #rcp_L2
>> +rcp_L4:
>
> Better labels would be immensely useful. This one might be called
> rcp_inf_or_denorm_or_zero.
>
>> +   and b32 $r4 $r1 0x7ff00000
>> +   sched 0x28 0x2b 0x04 0x28 0x2b 0x2d 0x2b
>> +   // Other values with nonzero in exponent field should be inf
>> +   set b32 $p0 0x1 eq s32 $r4 0x0
>> +   $p0 bra #rcp_L5
>> +   // +/-Inf -> +/-0
>> +   xor b32 $r1 $r1 0x7ff00000
>> +   mov b32 $r0 0x0
>> +   bra #rcp_L2
>> +rcp_L5:
>
> rcp_denorm_or_zero
>
>> +   set $p0 0x1 gtu f64 abs $r0d 0x0
>> +   $p0 bra #rcp_L6
>> +   // +/-0 -> +/-Inf
>> +   sched 0x28 0x2b 0x20 0x28 0x2f 0x28 0x2b
>> +   or b32 $r1 $r1 0x7ff00000
>> +   bra #rcp_L2
>> +rcp_L6:
>
> rcp_denorm
>
>> +   // non-0 denorms: multiply with 2^54 (the 0x36 in $r3), join with norms
>> +   mul rn f64 $r0d $r0d 0x4350000000000000
>> +   mov b32 $r3 0x36
>> +rcp_L2:
>> +   join nop
>
> Why not fold this into the the sources? i.e.
>
> join mov b32 $r3 0x36 right above, etc. (I can see how a structurizing
> compiler might have generated such code, but we can do a bit
> better...)
>
>> +rcp_L3:
>> +   // All numbers with -1 in $r3 have their result ready in $r0d, return them
>> +   // others need further calculation
>> +   set b32 $p0 0x1 lt s32 $r3 0x0
>> +   $p0 bra #rcp_end
>> +   sched 0x28 0x28 0x04 0x28 0x2b 0x04 0x28
>> +   // Step 2: Before the real calculation goes on, renormalize the values to
>> +   // range [1, 2) by setting exponent field to 0x3ff (the exponent of 1)
>> +   // result in $r6d. The exponent will be recovered later.
>> +   ext u32 $r2 $r1 0xb14
>> +   and b32 $r7 $r1 0x800fffff
>> +   add b32 $r7 $r7 0x3ff00000
>> +   mov b32 $r6 $r0
>> +   // Step 3: Convert new value to float (no overflow will occur due to step
>> +   // 2), calculate rcp and do newton-raphson step once
>> +   cvt rz f32 $r5 f64 $r6d
>> +   rcp f32 $r4 $r5
>> +   mov b32 $r0 0xbf800000
>> +   sched 0x28 0x28 0x2a 0x2b 0x2e 0x28 0x2e
>> +   fma rn f32 $r5 $r4 $r5 $r0
>> +   add ftz rn f32 $r5 neg $r5 neg 0x0
>
> Is this different than "cvt ftz f32 $r5 neg $r5"? Also, why is this
> not just folded into the fma? Is it just to flush this one value
> (since denorm flushing == ftz, happens on input rather than on
> output)?
>
>> +   fma rn f32 $r0 $r4 $r5 $r4
>> +   // Step 4: convert result $r0 back to double, do newton-raphson steps
>> +   cvt f64 $r0d f32 $r0
>> +   cvt f64 $r6d f64 neg $r6d
>> +   mov b32 $r9 0x3ff00000
>> +   mov b32 $r8 0x0
>
> This is 1.0 right? I prefer
>
> cvt f64 $r8d f32 0x3f800000
>
>> +   sched 0x29 0x29 0x29 0x29 0x29 0x29 0x29
>> +   // 4 Newton-Raphson Steps, tmp in $r4d, result in $r0d
>
> This would be a good place to remind people of the equation. In fact,
> I'd just move the equation down here.
>
>> +   fma rn f64 $r4d $r6d $r0d $r8d
>> +   fma rn f64 $r0d $r0d $r4d $r0d
>> +   fma rn f64 $r4d $r6d $r0d $r8d
>> +   fma rn f64 $r0d $r0d $r4d $r0d
>> +   fma rn f64 $r4d $r6d $r0d $r8d
>> +   fma rn f64 $r0d $r0d $r4d $r0d
>> +   fma rn f64 $r4d $r6d $r0d $r8d
>> +   sched 0x20 0x28 0x28 0x28 0x28 0x28 0x28
>> +   fma rn f64 $r0d $r0d $r4d $r0d
>> +   // Step 5: Exponent recovery and final processing
>> +   // The exponent is recovered by adding what we added to the exponent.
>> +   // Suppose we want to calculate rcp(x), but we have rcp(cx), then
>> +   //     rcp(x) = c * rcp(cx)
>> +   // The delta in exponent comes from two sources:
>> +   //   1) The renormalization in step 2. The delta is:
>> +   //      0x3ff - $r2
>> +   //   2) (For the denorm input) The 2^54 we multiplied at rcp_L6, stored
>> +   //      in $r3
>> +   // These 2 sources are calculated in the first two lines below, and then
>> +   // added to the exponent extracted from the result above.
>> +   // Note that after processing, the new exponent may >= 0x7ff (inf)
>> +   // or <= 0 (denorm). Those cases will be handled respectively below
>> +   subr b32 $r2 $r2 0x3ff
>> +   add b32 $r4 $r2 $r3
>> +   ext u32 $r3 $r1 0xb14
>> +   // New exponent in $r3
>> +   add b32 $r3 $r3 $r4
>> +   add b32 $r2 $r3 0xffffffff
>> +   // (exponent-1) < 0x7fe (unsigned) means the result is in norm range
>> +   // (same logic as in step 1)
>> +   set b32 $p0 0x1 lt u32 $r2 0x7fe
>> +   sched 0x2b 0x28 0x2b 0x28 0x28 0x2b 0x20
>> +   (not $p0) bra #rcp_L7
>> +   // Norms: convert exponents back and return
>> +   shl b32 $r4 $r4 clamp 0x14
>> +   add b32 $r1 $r4 $r1
>> +   bra #rcp_end
>> +rcp_L7:
>
> rcp_result_infinite_or_denorm
>
>> +   // New exponent >= 0x7ff means that result is inf
>> +   set b32 $p0 0x1 ge s32 $r3 0x7ff
>> +   (not $p0) bra #rcp_L8
>> +   // Infinity
>> +   and b32 $r1 $r1 0x80000000
>> +   sched 0x25 0x28 0x2b 0x23 0x25 0x28 0x23
>> +   mov b32 $r0 0x0
>> +   add b32 $r1 $r1 0x7ff00000
>> +   bra #rcp_end
>> +rcp_L8:
>> +   // denorms, they only fall within a small range, can't be smaller than
>> +   // 0x0004000000000000, which means if we set the exponent field to 1,
>
> Huh?
>
>>>> np.uint64(1).view('float64')
> 4.9406564584124654e-324
>
> Did you mean that they have to be smaller than some value? Or is there
> something else clever going on here?

Here, I mean that rcp of the greatest possible fp64 number
(0x7fefffff_ffffffff) is not less than one quarter of the smallest
normal number, i.e. 0x00040000_00000000. So there are only two cases
below. Discovering this trick needs some calculation, but it greatly
simplifies the logic here.

I agree with all the comments above, and will fix the code according
to your advice these days.

Thanks for the review.

Boyan Ding

>
>> +   // we can get the final result by mutiplying it with 1/2 or 1/4. Decide
>> +   // which one of the two is needed with exponent value, if not 0, 1/4 is
>> +   // used, 1/2 otherwise
>> +   set b32 $p0 0x1 ne u32 $r3 0x0
>> +   and b32 $r1 $r1 0x800fffff
>> +   $p0 mov b32 $r7 0x3fd00000
>> +   (not $p0) mov b32 $r7 0x3fe00000
>> +   sched 0x25 0x28 0x2c 0x2e 0x2e 0x00 0x00
>> +   add b32 $r1 $r1 0x00100000
>> +   mov b32 $r6 0x0
>> +   mul rn f64 $r0d $r0d $r6d
>> +rcp_end:
>> +   ret
>> +
>>  gk110_rsq_f64:
>>     ret
>>
>> diff --git a/src/gallium/drivers/nouveau/codegen/lib/gk110.asm.h b/src/gallium/drivers/nouveau/codegen/lib/gk110.asm.h
>> index 8d00e2a224..ce937a71f9 100644
>> --- a/src/gallium/drivers/nouveau/codegen/lib/gk110.asm.h
>> +++ b/src/gallium/drivers/nouveau/codegen/lib/gk110.asm.h
>> @@ -65,11 +65,101 @@ uint64_t gk110_builtin_code[] = {
>>         0xe088000001000406,
>>         0x4000000000800001,
>>         0xe6010000000ce802,
>> -       0x08b08010a010b810,
>> +       0x08a0a0a010a0b810,
>>         0xe60100000088e806,
>>         0x19000000001c003c,
>>  /* 0x0218: gk110_rcp_f64 */
>> -/* 0x0218: gk110_rsq_f64 */
>> +       0xc00000058a1c0409,
>> +       0x407fffffff9c080d,
>> +       0x1480000060000000,
>> +       0xb3401c03fe9c0c1d,
>> +       0xe4c03c007f9c000e,
>> +       0x08a0ac10acb410ac,
>> +       0x120000004c20003c,
>> +       0x747fffffff9fc00e,
>> +       0xb4601fff801c021d,
>> +       0x120000000820003c,
>> +       0x21000400001c0404,
>> +       0x12000000381c003c,
>> +/* 0x0278: rcp_L4 */
>> +       0x203ff800001c0410,
>> +       0x08acb4aca010aca0,
>> +       0xb3281c00001c101d,
>> +       0x120000000c00003c,
>> +       0x223ff800001c0404,
>> +       0xe4c03c007f9c0002,
>> +       0x120000001c1c003c,
>> +/* 0x02b0: rcp_L5 */
>> +       0xb4601c00001c021d,
>> +       0x120000000c00003c,
>> +       0x08aca0bca080aca0,
>> +       0x213ff800001c0404,
>> +       0x12000000081c003c,
>> +/* 0x02d8: rcp_L6 */
>> +       0xc400021a801c0001,
>> +       0x740000001b1fc00e,
>> +/* 0x02e8: rcp_L2 */
>> +       0x85800000005c3c02,
>> +/* 0x02f0: rcp_L3 */
>> +       0xb3181c00001c0c1d,
>> +       0x12000000d000003c,
>> +       0x08a010aca010a0a0,
>> +       0xc00000058a1c0409,
>> +       0x204007ffff9c041c,
>> +       0x401ff800001c1c1d,
>> +       0xe4c03c00001c001a,
>> +       0xe5400c00031c3816,
>> +       0x84000000021c1412,
>> +       0x745fc000001fc002,
>> +       0x08b8a0b8aca8a0a0,
>> +       0xcc000000029c1016,
>> +       0xcac88000001c1415,
>> +       0xcc001000029c1002,
>> +       0xe5400000001c2c02,
>> +       0xe5410000031c3c1a,
>> +       0x741ff800001fc026,
>> +       0xe4c03c007f9c0022,
>> +       0x08a4a4a4a4a4a4a4,
>> +       0xdb802000001c1812,
>> +       0xdb800000021c0002,
>> +       0xdb802000001c1812,
>> +       0xdb800000021c0002,
>> +       0xdb802000001c1812,
>> +       0xdb800000021c0002,
>> +       0xdb802000001c1812,
>> +       0x08a0a0a0a0a0a080,
>> +       0xdb800000021c0002,
>> +       0x48000001ff9c0809,
>> +       0xe0800000019c0812,
>> +       0xc00000058a1c040d,
>> +       0xe0800000021c0c0e,
>> +       0x407fffffff9c0c09,
>> +       0xb3101c03ff1c081d,
>> +       0x0880aca0a0aca0ac,
>> +       0x120000000c20003c,
>> +       0xc24000000a1c1011,
>> +       0xe0800000009c1006,
>> +       0x120000003c1c003c,
>> +/* 0x0428: rcp_L7 */
>> +       0xb3681c03ff9c0c1d,
>> +       0x120000001420003c,
>> +       0x20400000001c0404,
>> +       0x088ca0948caca094,
>> +       0xe4c03c007f9c0002,
>> +       0x403ff800001c0405,
>> +       0x12000000201c003c,
>> +/* 0x0460: rcp_L8 */
>> +       0xb3501c00001c0c1d,
>> +       0x204007ffff9c0404,
>> +       0x741fe8000003c01e,
>> +       0x741ff0000023c01e,
>> +       0x080000b8b8b0a094,
>> +       0x40000800001c0405,
>> +       0xe4c03c007f9c001a,
>> +       0xe4000000031c0002,
>> +/* 0x04a0: rcp_end */
>> +       0x19000000001c003c,
>> +/* 0x04a8: gk110_rsq_f64 */
>>         0x19000000001c003c,
>>  };
>>
>> @@ -77,5 +167,5 @@ uint64_t gk110_builtin_offsets[] = {
>>         0x0000000000000000,
>>         0x00000000000000f0,
>>         0x0000000000000218,
>> -       0x0000000000000218,
>> +       0x00000000000004a8,
>>  };
>> --
>> 2.12.0
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list