[Mesa-dev] [PATCH 08/13] gk104/ir: Use the new rcp/rsq in library
Karol Herbst
kherbst at redhat.com
Sun Jul 15 18:15:48 UTC 2018
From: Karol Herbst <karolherbst at gmail.com>
Signed-off-by: Karol Herbst <kherbst at redhat.com>
---
.../drivers/nouveau/codegen/lib/gk104.asm | 207 +++++++++++++++++-
.../drivers/nouveau/codegen/lib/gk104.asm.h | 144 +++++++++++-
.../nouveau/codegen/nv50_ir_lowering_nvc0.cpp | 2 +-
3 files changed, 336 insertions(+), 17 deletions(-)
diff --git a/src/gallium/drivers/nouveau/codegen/lib/gk104.asm b/src/gallium/drivers/nouveau/codegen/lib/gk104.asm
index cd65b547279..2a17a557b10 100644
--- a/src/gallium/drivers/nouveau/codegen/lib/gk104.asm
+++ b/src/gallium/drivers/nouveau/codegen/lib/gk104.asm
@@ -543,6 +543,8 @@ $p2 suldgb b32 $r3 cg zero u8 g[$r4d] $r2 $p0
$p1 suldgb b32 $r3 cv zero u8 g[$r4d] $r2 $p0
long mov b32 $r3 0x3f800000
long nop
+sched 0x00 0x00 0x00 0x00 0x00 0x00 0x00
+long nop
long ret
@@ -554,8 +556,145 @@ long ret
// SIZE: 9 * 8 bytes
//
gk104_rcp_f64:
- long nop
- long ret
+ // 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
+ ext u32 $r2 $r1 0xb14
+ add b32 $r3 $r2 0xffffffff
+ joinat #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
+ set $p0 0x1 gt u32 $r3 0x7fd
+ // $r3: 0 for norms, 0x36 for denorms, -1 for others
+ mov b32 $r3 0x0
+ sched 0x2f 0x04 0x2d 0x2b 0x2f 0x28 0x28
+ join (not $p0) nop
+ // Process all special values: NaN, inf, denorm, 0
+ mov b32 $r3 0xffffffff
+ // A number is NaN if its abs value is greater than or unordered with inf
+ set $p0 0x1 gtu f64 abs $r0d 0x7ff0000000000000
+ (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
+ join or b32 $r1 $r1 0x80000
+rcp_inf_or_denorm_or_zero:
+ and b32 $r4 $r1 0x7ff00000
+ // Other values with nonzero in exponent field should be inf
+ set $p0 0x1 eq s32 $r4 0x0
+ sched 0x2b 0x04 0x2f 0x2d 0x2b 0x2f 0x20
+ $p0 bra #rcp_denorm_or_zero
+ // +/-Inf -> +/-0
+ xor b32 $r1 $r1 0x7ff00000
+ join mov b32 $r0 0x0
+rcp_denorm_or_zero:
+ set $p0 0x1 gtu f64 abs $r0d 0x0
+ $p0 bra #rcp_denorm
+ // +/-0 -> +/-Inf
+ join or b32 $r1 $r1 0x7ff00000
+rcp_denorm:
+ // non-0 denorms: multiply with 2^54 (the 0x36 in $r3), join with norms
+ mul rn f64 $r0d $r0d 0x4350000000000000
+ sched 0x2f 0x28 0x2b 0x28 0x28 0x04 0x28
+ join mov b32 $r3 0x36
+rcp_rejoin:
+ // All numbers with -1 in $r3 have their result ready in $r0d, return them
+ // others need further calculation
+ set $p0 0x1 lt s32 $r3 0x0
+ $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.
+ ext u32 $r2 $r1 0xb14
+ and b32 $r7 $r1 0x800fffff
+ add b32 $r7 $r7 0x3ff00000
+ mov b32 $r6 $r0
+ sched 0x2b 0x04 0x28 0x28 0x2a 0x2b 0x2e
+ // 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
+ long rcp f32 $r4 $r5
+ mov b32 $r0 0xbf800000
+ fma rn f32 $r5 $r4 $r5 $r0
+ fma rn f32 $r0 neg $r4 $r5 $r4
+ // Step 4: convert result $r0 back to double, do newton-raphson steps
+ cvt f64 $r0d f32 $r0
+ cvt f64 $r6d neg f64 $r6d
+ sched 0x2e 0x29 0x29 0x29 0x29 0x29 0x29
+ cvt f64 $r8d f32 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}
+ 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
+ sched 0x29 0x20 0x28 0x28 0x28 0x28 0x28
+ fma rn f64 $r4d $r6d $r0d $r8d
+ 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_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
+ 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
+ sched 0x28 0x2b 0x28 0x2b 0x28 0x28 0x2b
+ // (exponent-1) < 0x7fe (unsigned) means the result is in norm range
+ // (same logic as in step 1)
+ set $p0 0x1 lt u32 $r2 0x7fe
+ (not $p0) bra #rcp_result_inf_or_denorm
+ // Norms: convert exponents back and return
+ shl b32 $r4 $r4 clamp 0x14
+ long add b32 $r1 $r4 $r1
+ bra #rcp_end
+rcp_result_inf_or_denorm:
+ // New exponent >= 0x7ff means that result is inf
+ set $p0 0x1 ge s32 $r3 0x7ff
+ (not $p0) bra #rcp_result_denorm
+ sched 0x20 0x25 0x28 0x2b 0x23 0x25 0x2f
+ // Infinity
+ and b32 $r1 $r1 0x80000000
+ long mov b32 $r0 0x0
+ add b32 $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.
+ set $p0 0x1 ne u32 $r3 0x0
+ and b32 $r1 $r1 0x800fffff
+ // 0x3e800000: 1/4
+ $p0 cvt f64 $r6d f32 0x3e800000
+ sched 0x2f 0x28 0x2c 0x2e 0x2a 0x20 0x27
+ // 0x3f000000: 1/2
+ (not $p0) cvt f64 $r6d f32 0x3f000000
+ add b32 $r1 $r1 0x00100000
+ mul rn f64 $r0d $r0d $r6d
+rcp_end:
+ ret
// RSQ F64: Newton Raphson rsqrt(x): r_{i+1} = r_i * (1.5 - 0.5 * x * r_i * r_i)
//
@@ -565,8 +704,68 @@ gk104_rcp_f64:
// SIZE: 14 * 8 bytes
//
gk104_rsq_f64:
- long nop
- long ret
+ // Before getting initial result rsqrt64h, two special cases should be
+ // handled first.
+ // 1. NaN: set the highest bit in mantissa so it'll be surely recognized
+ // as NaN in rsqrt64h
+ set $p0 0x1 gtu f64 abs $r0d 0x7ff0000000000000
+ $p0 or b32 $r1 $r1 0x00080000
+ and b32 $r2 $r1 0x7fffffff
+ sched 0x27 0x20 0x28 0x2c 0x25 0x28 0x28
+ // 2. denorms and small normal values: using their original value will
+ // lose precision either at rsqrt64h or the first step in newton-raphson
+ // steps below. Take 2 as a threshold in exponent field, and multiply
+ // with 2^54 if the exponent is smaller or equal. (will multiply 2^27
+ // to recover in the end)
+ ext u32 $r3 $r1 0xb14
+ set $p1 0x1 le u32 $r3 0x2
+ or b32 $r2 $r0 $r2
+ $p1 mul rn f64 $r0d $r0d 0x4350000000000000
+ rsqrt64h $r5 $r1
+ // rsqrt64h will give correct result for 0/inf/nan, the following logic
+ // checks whether the input is one of those (exponent is 0x7ff or all 0
+ // except for the sign bit)
+ set b32 $r6 ne u32 $r3 0x7ff
+ and b32 $r2 $r2 $r6
+ sched 0x28 0x2b 0x20 0x27 0x28 0x2e 0x28
+ set $p0 0x1 ne u32 $r2 0x0
+ $p0 bra #rsq_norm
+ // For 0/inf/nan, make sure the sign bit agrees with input and return
+ and b32 $r1 $r1 0x80000000
+ long mov b32 $r0 0x0
+ long or b32 $r1 $r1 $r5
+ ret
+rsq_norm:
+ // For others, do 4 Newton-Raphson steps with the formula:
+ // RSQ_{n + 1} = RSQ_{n} * (1.5 - 0.5 * x * RSQ_{n} * RSQ_{n})
+ // In the code below, each step is written as:
+ // tmp1 = 0.5 * x * RSQ_{n}
+ // tmp2 = -RSQ_{n} * tmp1 + 0.5
+ // RSQ_{n + 1} = RSQ_{n} * tmp2 + RSQ_{n}
+ long mov b32 $r4 0x0
+ sched 0x2f 0x29 0x29 0x29 0x29 0x29 0x29
+ // 0x3f000000: 1/2
+ cvt f64 $r8d f32 0x3f000000
+ mul rn f64 $r2d $r0d $r8d
+ mul rn f64 $r0d $r2d $r4d
+ fma rn f64 $r6d neg $r4d $r0d $r8d
+ fma rn f64 $r4d $r4d $r6d $r4d
+ mul rn f64 $r0d $r2d $r4d
+ fma rn f64 $r6d neg $r4d $r0d $r8d
+ sched 0x29 0x29 0x29 0x29 0x29 0x29 0x29
+ fma rn f64 $r4d $r4d $r6d $r4d
+ mul rn f64 $r0d $r2d $r4d
+ fma rn f64 $r6d neg $r4d $r0d $r8d
+ fma rn f64 $r4d $r4d $r6d $r4d
+ mul rn f64 $r0d $r2d $r4d
+ fma rn f64 $r6d neg $r4d $r0d $r8d
+ fma rn f64 $r4d $r4d $r6d $r4d
+ sched 0x29 0x20 0x28 0x2e 0x00 0x00 0x00
+ // Multiply 2^27 to result for small inputs to recover
+ $p1 mul rn f64 $r4d $r4d 0x41a0000000000000
+ long mov b32 $r1 $r5
+ long mov b32 $r0 $r4
+ ret
//
// Trap handler.
diff --git a/src/gallium/drivers/nouveau/codegen/lib/gk104.asm.h b/src/gallium/drivers/nouveau/codegen/lib/gk104.asm.h
index 37998768efe..ed948dee471 100644
--- a/src/gallium/drivers/nouveau/codegen/lib/gk104.asm.h
+++ b/src/gallium/drivers/nouveau/codegen/lib/gk104.asm.h
@@ -481,12 +481,132 @@ uint64_t gk104_builtin_code[] = {
0xd40040000840c785,
0x18fe00000000dde2,
0x4000000000001de4,
- 0x9000000000001de7,
-/* 0x0f08: gk104_rcp_f64 */
+ 0x2000000000000007,
0x4000000000001de4,
0x9000000000001de7,
-/* 0x0f18: gk104_rsq_f64 */
- 0x4000000000001de4,
+/* 0x0f18: gk104_rcp_f64 */
+ 0x7000c02c50109c03,
+ 0x0bfffffffc20dc02,
+ 0x6000000280000007,
+ 0x1a0ec01ff431dc03,
+ 0x180000000000dde2,
+ 0x228282f2b2d042f7,
+ 0x40000000000021f4,
+ 0x1bfffffffc00dde2,
+ 0x1e0edffc0001dc81,
+ 0x40000000200021e7,
+ 0x3800200000105c52,
+/* 0x0f70: rcp_inf_or_denorm_or_zero */
+ 0x39ffc00000111c02,
+ 0x190e0000fc41dc23,
+ 0x2202f2b2d2f042b7,
+ 0x40000000400001e7,
+ 0x39ffc00000105c82,
+ 0x1800000000001df2,
+/* 0x0fa0: rcp_denorm_or_zero */
+ 0x1e0ec0000001dc81,
+ 0x40000000200001e7,
+ 0x39ffc00000105c52,
+/* 0x0fb8: rcp_denorm */
+ 0x5000d0d400001c01,
+ 0x2280428282b282f7,
+ 0x18000000d800ddf2,
+/* 0x0fd0: rcp_rejoin */
+ 0x188e0000fc31dc23,
+ 0x40000006000001e7,
+ 0x7000c02c50109c03,
+ 0x3a003ffffc11dc02,
+ 0x08ffc0000071dc02,
+ 0x2800000000019de4,
+ 0x22e2b2a2828042b7,
+ 0x1006000019a15c04,
+ 0xc800000010511c00,
+ 0x1afe000000001de2,
+ 0x3000000014415c00,
+ 0x3008000014401e00,
+ 0x1000000001301c04,
+ 0x1000000019b19d04,
+ 0x22929292929292e7,
+ 0x1000cfe001321c04,
+ 0x2010000000611c01,
+ 0x2000000010001c01,
+ 0x2010000000611c01,
+ 0x2000000010001c01,
+ 0x2010000000611c01,
+ 0x2000000010001c01,
+ 0x2282828282820297,
+ 0x2010000000611c01,
+ 0x2000000010001c01,
+ 0x0800000ffc209e02,
+ 0x480000000c211c03,
+ 0x7000c02c5010dc03,
+ 0x480000001030dc03,
+ 0x0bfffffffc309c02,
+ 0x22b28282b282b287,
+ 0x188ec01ff821dc03,
+ 0x40000000600021e7,
+ 0x6000c00050411c03,
+ 0x4800000004405c03,
+ 0x40000001c0001de7,
+/* 0x10f0: rcp_result_inf_or_denorm */
+ 0x1b0ec01ffc31dc23,
+ 0x40000000a00021e7,
+ 0x22f25232b2825207,
+ 0x3a00000000105c02,
+ 0x1800000000001de2,
+ 0x09ffc00000105c02,
+ 0x40000000e0001de7,
+/* 0x1128: rcp_result_denorm */
+ 0x1a8e0000fc31dc03,
+ 0x3a003ffffc105c02,
+ 0x1000cfa001318004,
+ 0x227202a2e2c282f7,
+ 0x1000cfc00131a004,
+ 0x0800400000105c02,
+ 0x5000000018001c01,
+/* 0x1160: rcp_end */
+ 0x9000000000001de7,
+/* 0x1168: gk104_rsq_f64 */
+ 0x1e0edffc0001dc81,
+ 0x3800200000104042,
+ 0x39fffffffc109c02,
+ 0x22828252c2820277,
+ 0x7000c02c5010dc03,
+ 0x198ec0000833dc03,
+ 0x6800000008009c43,
+ 0x5000d0d400000401,
+ 0xc80000001c115c00,
+ 0x128ec01ffc319c03,
+ 0x6800000018209c03,
+ 0x2282e2827202b287,
+ 0x1a8e0000fc21dc03,
+ 0x40000000800001e7,
+ 0x3a00000000105c02,
+ 0x1800000000001de2,
+ 0x6800000014105c43,
+ 0x9000000000001de7,
+/* 0x11f8: rsq_norm */
+ 0x1800000000011de2,
+ 0x22929292929292f7,
+ 0x1000cfc001321c04,
+ 0x5000000020009c01,
+ 0x5000000010201c01,
+ 0x2010000000419e01,
+ 0x2008000018411c01,
+ 0x5000000010201c01,
+ 0x2010000000419e01,
+ 0x2292929292929297,
+ 0x2008000018411c01,
+ 0x5000000010201c01,
+ 0x2010000000419e01,
+ 0x2008000018411c01,
+ 0x5000000010201c01,
+ 0x2010000000419e01,
+ 0x2008000018411c01,
+ 0x20000002e2820297,
+ 0x5000d06800410401,
+ 0x2800000014005de4,
+ 0x2800000010001de4,
0x9000000000001de7,
0xc800000003f01cc5,
0x2c00000100005c04,
@@ -495,7 +615,7 @@ uint64_t gk104_builtin_code[] = {
0x680100000c1fdc03,
0x4000000a60001c47,
0x180000004000dde2,
-/* 0x0f60: spill_cfstack */
+/* 0x12e0: spill_cfstack */
0x78000009c0000007,
0x0c0000000430dd02,
0x4003ffffa0001ca7,
@@ -543,14 +663,14 @@ uint64_t gk104_builtin_code[] = {
0x4000000100001ea7,
0x480100000c001c03,
0x0800000000105c42,
-/* 0x10d8: shared_loop */
+/* 0x1458: shared_loop */
0xc100000000309c85,
0x9400000500009c85,
0x0c00000010001d02,
0x0800000000105d42,
0x0c0000001030dd02,
0x4003ffff40001ca7,
-/* 0x1108: shared_done */
+/* 0x1488: shared_done */
0x2800406420001de4,
0x2800406430005de4,
0xe000000000001c45,
@@ -564,7 +684,7 @@ uint64_t gk104_builtin_code[] = {
0x480000000c209c03,
0x4801000008001c03,
0x0800000000105c42,
-/* 0x1170: search_cstack */
+/* 0x14f0: search_cstack */
0x280040646000dde4,
0x8400000020009f05,
0x190ec0002821dc03,
@@ -573,17 +693,17 @@ uint64_t gk104_builtin_code[] = {
0x0800000000105c42,
0x0c0000004030dd02,
0x00029dff0ffc5cbf,
-/* 0x11b0: entry_found */
+/* 0x1530: entry_found */
0x8400000000009f85,
0x2800406400001de4,
0x2800406410005de4,
0x9400000010009c85,
0x4000000000001df4,
-/* 0x11d8: end_exit */
+/* 0x1558: end_exit */
0x9800000003ffdcc5,
0xd000000000008007,
0xa000000000004007,
-/* 0x11f0: end_cont */
+/* 0x1570: end_cont */
0xd000000000008007,
0x3400c3fffc201c04,
0xc000000003f01ec5,
@@ -593,6 +713,6 @@ uint64_t gk104_builtin_code[] = {
uint64_t gk104_builtin_offsets[] = {
0x0000000000000000,
0x00000000000000f0,
- 0x0000000000000f08,
0x0000000000000f18,
+ 0x0000000000001168,
};
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 99758f31e35..5b76872afad 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
@@ -127,7 +127,7 @@ NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
bld.mkSplit(src, 4, i->getSrc(0));
int chip = prog->getTarget()->getChipset();
- if (chip >= NVISA_GK20A_CHIPSET && chip < NVISA_GM107_CHIPSET) {
+ if (chip >= NVISA_GK104_CHIPSET && chip < NVISA_GM107_CHIPSET) {
handleRCPRSQLib(i, src);
return;
}
--
2.17.1
More information about the mesa-dev
mailing list