gm107/ir: add fp64 rcp

Acked-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: 19.0 <mesa-stable@lists.freedesktop.org>

3 files changed, 270 insertions, 4 deletions

diff --git a/src/gallium/drivers/nouveau/codegen/lib/gm107.asm b/src/gallium/drivers/nouveau/codegen/lib/gm107.asm
index 7ee5f8fc65b..595d9dc5d41 100644
--- a/src/gallium/drivers/nouveau/codegen/lib/gm107.asm
+++ b/src/gallium/drivers/nouveau/codegen/lib/gm107.asm
@@ -100,8 +100,175 @@ gm107_div_s32:
    ret
    nop 0
 
-// STUB
+// 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).
+//
 gm107_rcp_f64:
+   // Step 1: classify input according to exponent and value, and calculate
+   // result for 0/inf/nan. $r2 holds the exponent value, which starts at
+   // bit 52 (bit 20 of the upper half) and is 11 bits in length
+   sched (st 0x0) (st 0x0) (st 0x0)
+   bfe u32 $r2 $r1 0xb14
+   iadd32i $r3 $r2 -1
+   ssy #rcp_rejoin
+   // We want to check whether the exponent is 0 or 0x7ff (i.e. NaN, inf,
+   // denorm, or 0). Do this by substracting 1 from the exponent, which will
+   // mean that it's > 0x7fd in those cases when doing unsigned comparison
+   sched (st 0x0) (st 0x0) (st 0x0)
+   isetp gt u32 and $p0 1 $r3 0x7fd 1
+   // $r3: 0 for norms, 0x36 for denorms, -1 for others
+   mov $r3 0x0 0xf
+   not $p0 sync
+   // Process all special values: NaN, inf, denorm, 0
+   sched (st 0x0) (st 0x0) (st 0x0)
+   mov32i $r3 0xffffffff 0xf
+   // A number is NaN if its abs value is greater than or unordered with inf
+   dsetp gtu and $p0 1 abs $r0 0x7ff0000000000000 1
+   not $p0 bra #rcp_inf_or_denorm_or_zero
+   // NaN -> NaN, the next line sets the "quiet" bit of the result. This
+   // behavior is both seen on the CPU and the blob
+   sched (st 0x0) (st 0x0) (st 0x0)
+   lop32i or $r1 $r1 0x80000
+   sync
+rcp_inf_or_denorm_or_zero:
+   lop32i and $r4 $r1 0x7ff00000
+   sched (st 0x0) (st 0x0) (st 0x0)
+   // Other values with nonzero in exponent field should be inf
+   isetp eq and $p0 1 $r4 0x0 1
+   $p0 bra #rcp_denorm_or_zero
+   // +/-Inf -> +/-0
+   lop32i xor $r1 $r1 0x7ff00000
+   sched (st 0x0) (st 0x0) (st 0x0)
+   mov $r0 0x0 0xf
+   sync
+rcp_denorm_or_zero:
+   dsetp gtu and $p0 1 abs $r0 0x0 1
+   sched (st 0x0) (st 0x0) (st 0x0)
+   $p0 bra #rcp_denorm
+   // +/-0 -> +/-Inf
+   lop32i or $r1 $r1 0x7ff00000
+   sync
+rcp_denorm:
+   // non-0 denorms: multiply with 2^54 (the 0x36 in $r3), join with norms
+   sched (st 0x0) (st 0x0) (st 0x0)
+   dmul $r0 $r0 0x4350000000000000
+   mov $r3 0x36 0xf
+   sync
+rcp_rejoin:
+   // All numbers with -1 in $r3 have their result ready in $r0d, return them
+   // others need further calculation
+   sched (st 0x0) (st 0x0) (st 0x0)
+   isetp lt and $p0 1 $r3 0x0 1
+   $p0 bra #rcp_end
+   // 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.
+   bfe u32 $r2 $r1 0xb14
+   sched (st 0x0) (st 0x0) (st 0x0)
+   lop32i and $r7 $r1 0x800fffff
+   iadd32i $r7 $r7 0x3ff00000
+   mov $r6 $r0 0xf
+   // Step 3: Convert new value to float (no overflow will occur due to step
+   // 2), calculate rcp and do newton-raphson step once
+   sched (st 0x0) (st 0x0) (st 0x0)
+   f2f ftz f64 f32 $r5 $r6
+   mufu rcp $r4 $r5
+   mov32i $r0 0xbf800000 0xf
+   sched (st 0x0) (st 0x0) (st 0x0)
+   ffma $r5 $r4 $r5 $r0
+   ffma $r0 $r5 neg $r4 $r4
+   // Step 4: convert result $r0 back to double, do newton-raphson steps
+   f2f f32 f64 $r0 $r0
+   sched (st 0x0) (st 0x0) (st 0x0)
+   f2f f64 f64 $r6 neg $r6
+   f2f f32 f64 $r8 0x3f800000
+   // 4 Newton-Raphson Steps, tmp in $r4d, result in $r0d
+   // The formula used here (and above) 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
+   // looks like:
+   //     tmp = -src * RCP_{n} + 1
+   //     RCP_{n + 1} = RCP_{n} * tmp + RCP_{n}
+   dfma $r4 $r6 $r0 $r8
+   sched (st 0x0) (st 0x0) (st 0x0)
+   dfma $r0 $r0 $r4 $r0
+   dfma $r4 $r6 $r0 $r8
+   dfma $r0 $r0 $r4 $r0
+   sched (st 0x0) (st 0x0) (st 0x0)
+   dfma $r4 $r6 $r0 $r8
+   dfma $r0 $r0 $r4 $r0
+   dfma $r4 $r6 $r0 $r8
+   sched (st 0x0) (st 0x0) (st 0x0)
+   dfma $r0 $r0 $r4 $r0
+   // 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_denorm, 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
+   iadd $r2 neg $r2 0x3ff
+   iadd $r4 $r2 $r3
+   sched (st 0x0) (st 0x0) (st 0x0)
+   bfe u32 $r3 $r1 0xb14
+   // New exponent in $r3
+   iadd $r3 $r3 $r4
+   iadd32i $r2 $r3 -1
+   // (exponent-1) < 0x7fe (unsigned) means the result is in norm range
+   // (same logic as in step 1)
+   sched (st 0x0) (st 0x0) (st 0x0)
+   isetp lt u32 and $p0 1 $r2 0x7fe 1
+   not $p0 bra #rcp_result_inf_or_denorm
+   // Norms: convert exponents back and return
+   shl $r4 $r4 0x14
+   sched (st 0x0) (st 0x0) (st 0x0)
+   iadd $r1 $r4 $r1
+   bra #rcp_end
+rcp_result_inf_or_denorm:
+   // New exponent >= 0x7ff means that result is inf
+   isetp ge and $p0 1 $r3 0x7ff 1
+   sched (st 0x0) (st 0x0) (st 0x0)
+   not $p0 bra #rcp_result_denorm
+   // Infinity
+   lop32i and $r1 $r1 0x80000000
+   mov $r0 0x0 0xf
+   sched (st 0x0) (st 0x0) (st 0x0)
+   iadd32i $r1 $r1 0x7ff00000
+   bra #rcp_end
+rcp_result_denorm:
+   // Denorm result comes from huge input. The greatest possible fp64, i.e.
+   // 0x7fefffffffffffff's rcp is 0x0004000000000000, 1/4 of the smallest
+   // normal value. Other rcp result should be greater than that. If we
+   // set the exponent field to 1, we can recover the result by multiplying
+   // it with 1/2 or 1/4. 1/2 is used if the "exponent" $r3 is 0, otherwise
+   // 1/4 ($r3 should be -1 then). This is quite tricky but greatly simplifies
+   // the logic here.
+   isetp ne u32 and $p0 1 $r3 0x0 1
+   sched (st 0x0) (st 0x0) (st 0x0)
+   lop32i and $r1 $r1 0x800fffff
+   // 0x3e800000: 1/4
+   $p0 f2f f32 f64 $r6 0x3e800000
+   // 0x3f000000: 1/2
+   not $p0 f2f f32 f64 $r6 0x3f000000
+   sched (st 0x0) (st 0x0) (st 0x0)
+   iadd32i $r1 $r1 0x00100000
+   dmul $r0 $r0 $r6
+rcp_end:
+   ret
+
 gm107_rsq_f64:
    sched (st 0x0) (st 0x0) (st 0x0)
    ret
diff --git a/src/gallium/drivers/nouveau/codegen/lib/gm107.asm.h b/src/gallium/drivers/nouveau/codegen/lib/gm107.asm.h
index 65c93f7ae89..206d01bde83 100644
--- a/src/gallium/drivers/nouveau/codegen/lib/gm107.asm.h
+++ b/src/gallium/drivers/nouveau/codegen/lib/gm107.asm.h
@@ -82,7 +82,106 @@ uint64_t gm107_builtin_code[] = {
 	0xe32000000007000f,
 	0x50b0000000070f00,
 /* 0x0280: gm107_rcp_f64 */
-/* 0x0280: gm107_rsq_f64 */
+	0x001f8000fc0007e0,
+	0x38000000b1470102,
+	0x1c0ffffffff70203,
+	0xe29000000e000000,
+	0x001f8000fc0007e0,
+	0x366803807fd70307,
+	0x5c9807800ff70003,
+	0xf0f800000008000f,
+	0x001f8000fc0007e0,
+	0x010ffffffff7f003,
+	0x368c03fff0070087,
+	0xe24000000188000f,
+	0x001f8000fc0007e0,
+	0x0420008000070101,
+	0xf0f800000007000f,
+/* 0x02f8: rcp_inf_or_denorm_or_zero */
+	0x0407ff0000070104,
+	0x001f8000fc0007e0,
+	0x5b6503800ff70407,
+	0xe24000000200000f,
+	0x0447ff0000070101,
+	0x001f8000fc0007e0,
+	0x5c9807800ff70000,
+	0xf0f800000007000f,
+/* 0x0338: rcp_denorm_or_zero */
+	0x5b8c03800ff70087,
+	0x001f8000fc0007e0,
+	0xe24000000100000f,
+	0x0427ff0000070101,
+	0xf0f800000007000f,
+/* 0x0360: rcp_denorm */
+	0x001f8000fc0007e0,
+	0x3880004350070000,
+	0x3898078003670003,
+	0xf0f800000007000f,
+/* 0x0380: rcp_rejoin */
+	0x001f8000fc0007e0,
+	0x5b6303800ff70307,
+	0xe24000001c00000f,
+	0x38000000b1470102,
+	0x001f8000fc0007e0,
+	0x040800fffff70107,
+	0x1c03ff0000070707,
+	0x5c98078000070006,
+	0x001f8000fc0007e0,
+	0x5ca8100000670e05,
+	0x5080000000470504,
+	0x010bf8000007f000,
+	0x001f8000fc0007e0,
+	0x5980000000570405,
+	0x5981020000470500,
+	0x5ca8000000070b00,
+	0x001f8000fc0007e0,
+	0x5ca8200000670f06,
+	0x38a8003f80070b08,
+	0x5b70040000070604,
+	0x001f8000fc0007e0,
+	0x5b70000000470000,
+	0x5b70040000070604,
+	0x5b70000000470000,
+	0x001f8000fc0007e0,
+	0x5b70040000070604,
+	0x5b70000000470000,
+	0x5b70040000070604,
+	0x001f8000fc0007e0,
+	0x5b70000000470000,
+	0x381200003ff70202,
+	0x5c10000000370204,
+	0x001f8000fc0007e0,
+	0x38000000b1470103,
+	0x5c10000000470303,
+	0x1c0ffffffff70302,
+	0x001f8000fc0007e0,
+	0x366203807fe70207,
+	0xe24000000208000f,
+	0x3848000001470404,
+	0x001f8000fc0007e0,
+	0x5c10000000170401,
+	0xe24000000807000f,
+/* 0x04d8: rcp_result_inf_or_denorm */
+	0x366d03807ff70307,
+	0x001f8000fc0007e0,
+	0xe24000000288000f,
+	0x0408000000070101,
+	0x5c9807800ff70000,
+	0x001f8000fc0007e0,
+	0x1c07ff0000070101,
+	0xe24000000407000f,
+/* 0x0518: rcp_result_denorm */
+	0x5b6a03800ff70307,
+	0x001f8000fc0007e0,
+	0x040800fffff70101,
+	0x38a8003e80000b06,
+	0x38a8003f00080b06,
+	0x001f8000fc0007e0,
+	0x1c00010000070101,
+	0x5c80000000670000,
+/* 0x0558: rcp_end */
+	0xe32000000007000f,
+/* 0x0560: gm107_rsq_f64 */
 	0x001f8000fc0007e0,
 	0xe32000000007000f,
 	0x50b0000000070f00,
@@ -93,5 +192,5 @@ uint64_t gm107_builtin_offsets[] = {
 	0x0000000000000000,
 	0x0000000000000120,
 	0x0000000000000280,
-	0x0000000000000280,
+	0x0000000000000560,
 };
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
index 7d28c5f6e52..5d69ce2ab7c 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
@@ -129,7 +129,7 @@ NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
    bld.mkSplit(src, 4, i->getSrc(0));
 
    int chip = prog->getTarget()->getChipset();
-   if (chip >= NVISA_GK104_CHIPSET && chip < NVISA_GM107_CHIPSET) {
+   if (chip >= NVISA_GK104_CHIPSET && (i->op == OP_RCP || chip < NVISA_GM107_CHIPSET)) {
       handleRCPRSQLib(i, src);
       return;
    }