gk104/ir: Use the new rcp/rsq in library

[imirkin: add a few more "long" prefixes to safen things up]
Acked-by: Ilia Mirkin <[email protected]>
Cc: 19.0 <[email protected]>

1 files changed, 132 insertions, 12 deletions

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,
 };