--- /dev/null
+ENVYAS ?= envyas
+
+all: gf100.asm.h gk104.asm.h gk110.asm.h
+
+gf100.asm.h: %.asm.h: %.asm
+ $(ENVYAS) -a -W -mnvc0 -Vnvc0 $< -o $@
+gk104.asm.h: %.asm.h: %.asm
+ $(ENVYAS) -a -W -mnvc0 -Vnve4 $< -o $@
+gk110.asm.h: %.asm.h: %.asm
+ $(ENVYAS) -a -W -mgk110 $< -o $@
--- /dev/null
+.section #gf100_builtin_code
+// DIV U32
+//
+// UNR recurrence (q = a / b):
+// look for z such that 2^32 - b <= b * z < 2^32
+// then q - 1 <= (a * z) / 2^32 <= q
+//
+// INPUT: $r0: dividend, $r1: divisor
+// OUTPUT: $r0: result, $r1: modulus
+// CLOBBER: $r2 - $r3, $p0 - $p1
+// SIZE: 22 / 14 * 8 bytes
+//
+gf100_div_u32:
+ bfind u32 $r2 $r1
+ xor b32 $r2 $r2 0x1f
+ mov b32 $r3 0x1
+ shl b32 $r2 $r3 clamp $r2
+ cvt u32 $r1 neg u32 $r1
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mov b32 $r3 $r0
+ mul high $r0 u32 $r0 u32 $r2
+ cvt u32 $r2 neg u32 $r1
+ add $r1 (mul u32 $r1 u32 $r0) $r3
+ set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ $p0 set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ ret
+
+// DIV S32, like DIV U32 after taking ABS(inputs)
+//
+// INPUT: $r0: dividend, $r1: divisor
+// OUTPUT: $r0: result, $r1: modulus
+// CLOBBER: $r2 - $r3, $p0 - $p3
+//
+gf100_div_s32:
+ set $p2 0x1 lt s32 $r0 0x0
+ set $p3 0x1 lt s32 $r1 0x0 xor $p2
+ cvt s32 $r0 abs s32 $r0
+ cvt s32 $r1 abs s32 $r1
+ bfind u32 $r2 $r1
+ xor b32 $r2 $r2 0x1f
+ mov b32 $r3 0x1
+ shl b32 $r2 $r3 clamp $r2
+ cvt u32 $r1 neg u32 $r1
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mov b32 $r3 $r0
+ mul high $r0 u32 $r0 u32 $r2
+ cvt u32 $r2 neg u32 $r1
+ add $r1 (mul u32 $r1 u32 $r0) $r3
+ set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ $p0 set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ $p3 cvt s32 $r0 neg s32 $r0
+ $p2 cvt s32 $r1 neg s32 $r1
+ ret
+
+// RCP F64: Newton Raphson reciprocal(x): r_{i+1} = r_i * (2.0 - x * r_i)
+//
+// INPUT: $r0d (x)
+// OUTPUT: $r0d (rcp(x))
+// CLOBBER: $r2 - $r7
+// SIZE: 9 * 8 bytes
+//
+gf100_rcp_f64:
+ nop
+ ret
+
+// RSQ F64: Newton Raphson rsqrt(x): r_{i+1} = r_i * (1.5 - 0.5 * x * r_i * r_i)
+//
+// INPUT: $r0d (x)
+// OUTPUT: $r0d (rsqrt(x))
+// CLOBBER: $r2 - $r7
+// SIZE: 14 * 8 bytes
+//
+gf100_rsq_f64:
+ nop
+ ret
+
+.section #gf100_builtin_offsets
+.b64 #gf100_div_u32
+.b64 #gf100_div_s32
+.b64 #gf100_rcp_f64
+.b64 #gf100_rsq_f64
--- /dev/null
+uint64_t gf100_builtin_code[] = {
+/* 0x0000: gf100_div_u32 */
+ 0x7800000004009c03,
+ 0x0010dd187c209cdd,
+ 0x6000000008309c03,
+ 0x0810dc2a05605c18,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x280000000000dde4,
+ 0x5000000008001c43,
+ 0x0010430d05609c18,
+ 0x1b0e00000811dc03,
+ 0x4800000008104103,
+ 0x0800000004000002,
+ 0x1b0e00000811c003,
+ 0x4800000008104103,
+ 0x90001dff040000ac,
+/* 0x00b0: gf100_div_s32 */
+ 0x188e0000fc05dc23,
+ 0x18c40000fc17dc23,
+ 0x07305e1803301e18,
+ 0x7800000004009c03,
+ 0x0010dd187c209cdd,
+ 0x6000000008309c03,
+ 0x0810dc2a05605c18,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x280000000000dde4,
+ 0x5000000008001c43,
+ 0x0010430d05609c18,
+ 0x1b0e00000811dc03,
+ 0x4800000008104103,
+ 0x0800000004000002,
+ 0x1b0e00000811c003,
+ 0x4800000008104103,
+ 0x01700e18040000ac,
+ 0x90001dff05704a18,
+/* 0x0180: gf100_rcp_f64 */
+ 0x90001dff00001c08,
+/* 0x0188: gf100_rsq_f64 */
+ 0x90001dff00001c08,
+};
+
+uint64_t gf100_builtin_offsets[] = {
+ 0x0000000000000000,
+ 0x00000000000000b0,
+ 0x0000000000000180,
+ 0x0000000000000188,
+};
-//
+.section #gk104_builtin_code
// DIV U32
//
// UNR recurrence (q = a / b):
// CLOBBER: $r2 - $r3, $p0 - $p1
// SIZE: 22 / 14 * 8 bytes
//
-sched 0x28 0x4 0x28 0x4 0x28 0x28 0x28
-bfind u32 $r2 $r1
-long xor b32 $r2 $r2 0x1f
-long mov b32 $r3 0x1
-shl b32 $r2 $r3 clamp $r2
-long cvt u32 $r1 neg u32 $r1
-long mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-sched 0x28 0x28 0x28 0x28 0x28 0x28 0x28
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-sched 0x4 0x28 0x4 0x28 0x28 0x2c 0x4
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mov b32 $r3 $r0
-mul high $r0 u32 $r0 u32 $r2
-long cvt u32 $r2 neg u32 $r1
-long add $r1 (mul u32 $r1 u32 $r0) $r3
-set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-sched 0x28 0x2c 0x4 0x20 0x2e 0x28 0x20
-$p0 add b32 $r0 $r0 0x1
-$p0 set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-long ret
-//
+gk104_div_u32:
+ sched 0x28 0x4 0x28 0x4 0x28 0x28 0x28
+ bfind u32 $r2 $r1
+ long xor b32 $r2 $r2 0x1f
+ long mov b32 $r3 0x1
+ shl b32 $r2 $r3 clamp $r2
+ long cvt u32 $r1 neg u32 $r1
+ long mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ sched 0x28 0x28 0x28 0x28 0x28 0x28 0x28
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ sched 0x4 0x28 0x4 0x28 0x28 0x2c 0x4
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mov b32 $r3 $r0
+ mul high $r0 u32 $r0 u32 $r2
+ long cvt u32 $r2 neg u32 $r1
+ long add $r1 (mul u32 $r1 u32 $r0) $r3
+ set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ sched 0x28 0x2c 0x4 0x20 0x2e 0x28 0x20
+ $p0 add b32 $r0 $r0 0x1
+ $p0 set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ long ret
+
// DIV S32, like DIV U32 after taking ABS(inputs)
//
// INPUT: $r0: dividend, $r1: divisor
// OUTPUT: $r0: result, $r1: modulus
// CLOBBER: $r2 - $r3, $p0 - $p3
//
-set $p2 0x1 lt s32 $r0 0x0
-set $p3 0x1 lt s32 $r1 0x0 xor $p2
-sched 0x20 0x28 0x28 0x4 0x28 0x04 0x28
-long cvt s32 $r0 abs s32 $r0
-long cvt s32 $r1 abs s32 $r1
-bfind u32 $r2 $r1
-long xor b32 $r2 $r2 0x1f
-long mov b32 $r3 0x1
-shl b32 $r2 $r3 clamp $r2
-cvt u32 $r1 neg u32 $r1
-sched 0x28 0x28 0x28 0x28 0x28 0x28 0x28
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-sched 0x28 0x28 0x4 0x28 0x04 0x28 0x28
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mov b32 $r3 $r0
-mul high $r0 u32 $r0 u32 $r2
-long cvt u32 $r2 neg u32 $r1
-long add $r1 (mul u32 $r1 u32 $r0) $r3
-sched 0x2c 0x04 0x28 0x2c 0x04 0x28 0x20
-set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-$p0 set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-long $p0 add b32 $r0 $r0 0x1
-long $p3 cvt s32 $r0 neg s32 $r0
-sched 0x04 0x2e 0x04 0x28 0x04 0x20 0x2c
-$p2 cvt s32 $r1 neg s32 $r1
-long ret
-//
+gk104_div_s32:
+ set $p2 0x1 lt s32 $r0 0x0
+ set $p3 0x1 lt s32 $r1 0x0 xor $p2
+ sched 0x20 0x28 0x28 0x4 0x28 0x04 0x28
+ long cvt s32 $r0 abs s32 $r0
+ long cvt s32 $r1 abs s32 $r1
+ bfind u32 $r2 $r1
+ long xor b32 $r2 $r2 0x1f
+ long mov b32 $r3 0x1
+ shl b32 $r2 $r3 clamp $r2
+ cvt u32 $r1 neg u32 $r1
+ sched 0x28 0x28 0x28 0x28 0x28 0x28 0x28
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ sched 0x28 0x28 0x4 0x28 0x04 0x28 0x28
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mov b32 $r3 $r0
+ mul high $r0 u32 $r0 u32 $r2
+ long cvt u32 $r2 neg u32 $r1
+ long add $r1 (mul u32 $r1 u32 $r0) $r3
+ sched 0x2c 0x04 0x28 0x2c 0x04 0x28 0x20
+ set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ $p0 set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ long $p0 add b32 $r0 $r0 0x1
+ long $p3 cvt s32 $r0 neg s32 $r0
+ sched 0x04 0x2e 0x04 0x28 0x04 0x20 0x2c
+ $p2 cvt s32 $r1 neg s32 $r1
+ long ret
+
// SULDP [for each format]
// $r4d: address
// $r2: surface info (format)
long mov b32 $r3 0x3f800000
long nop
long ret
-//
+
+
// RCP F64: Newton Raphson reciprocal(x): r_{i+1} = r_i * (2.0 - x * r_i)
//
// INPUT: $r0d (x)
// CLOBBER: $r2 - $r7
// SIZE: 9 * 8 bytes
//
-long nop
-long ret
+gk104_rcp_f64:
+ long nop
+ long ret
+
// RSQ F64: Newton Raphson rsqrt(x): r_{i+1} = r_i * (1.5 - 0.5 * x * r_i * r_i)
//
// INPUT: $r0d (x)
// CLOBBER: $r2 - $r7
// SIZE: 14 * 8 bytes
//
-long nop
-long ret
+gk104_rsq_f64:
+ long nop
+ long ret
+
//
// Trap handler.
// Requires at least 4 GPRs and 32 bytes of l[] memory to temporarily save GPRs.
mov $flags $r2 mask 0xffff
ld b128 $r0q cs l[0x00]
rtt
+
+.section #gk104_builtin_offsets
+.b64 #gk104_div_u32
+.b64 #gk104_div_s32
+.b64 #gk104_rcp_f64
+.b64 #gk104_rsq_f64
--- /dev/null
+uint64_t gk104_builtin_code[] = {
+/* 0x0000: gk104_div_u32 */
+ 0x2282828042804287,
+ 0x7800000004009c03,
+ 0x380000007c209c82,
+ 0x180000000400dde2,
+ 0x6000000008309c03,
+ 0x1c00000005205d04,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x2282828282828287,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x2042c28280428047,
+ 0x200400000c209c43,
+ 0x280000000000dde4,
+ 0x5000000008001c43,
+ 0x1c00000005209d04,
+ 0x2006000000105c03,
+ 0x1b0e00000811dc03,
+ 0x4800000008104103,
+ 0x220282e20042c287,
+ 0x0800000004000002,
+ 0x1b0e00000811c003,
+ 0x4800000008104103,
+ 0x0800000004000002,
+ 0x9000000000001de7,
+/* 0x00f0: gk104_div_s32 */
+ 0x188e0000fc05dc23,
+ 0x18c40000fc17dc23,
+ 0x2280428042828207,
+ 0x1c00000001201ec4,
+ 0x1c00000005205ec4,
+ 0x7800000004009c03,
+ 0x380000007c209c82,
+ 0x180000000400dde2,
+ 0x6000000008309c03,
+ 0x1c00000005205d04,
+ 0x2282828282828287,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x2282804280428287,
+ 0x200400000c209c43,
+ 0x500000000810dc03,
+ 0x200400000c209c43,
+ 0x280000000000dde4,
+ 0x5000000008001c43,
+ 0x1c00000005209d04,
+ 0x2006000000105c03,
+ 0x22028042c28042c7,
+ 0x1b0e00000811dc03,
+ 0x4800000008104103,
+ 0x0800000004000002,
+ 0x1b0e00000811c003,
+ 0x4800000008104103,
+ 0x0800000004000002,
+ 0x1c00000001200f84,
+ 0x22c200428042e047,
+ 0x1c00000005204b84,
+ 0x9000000000001de7,
+ 0xd4004000084004c5,
+ 0x0c5400000013dc04,
+ 0xd4004000084009c5,
+ 0xd4004000084007c5,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd4004000084004c5,
+ 0x0c5400000013dc04,
+ 0xd4004000084009c5,
+ 0xd4004000084007c5,
+ 0x1900000004a0dc04,
+ 0x1800000004a09c04,
+ 0x30de0001d030dc02,
+ 0x2000000000000007,
+ 0x1900000000a05c04,
+ 0x30de0001d0209c02,
+ 0x1800000000a01c04,
+ 0x30de0001d0105c02,
+ 0x30de0001d0001c02,
+ 0x9000000000001de7,
+ 0xd4004000084004a5,
+ 0x2000000000000007,
+ 0x0c5400000013dc04,
+ 0xd4004000084009a5,
+ 0xd4004000084007a5,
+ 0x1900000004a0de04,
+ 0x1800000004a09e04,
+ 0x30e000061c30dc02,
+ 0x1900000000a05e04,
+ 0x2000000000000007,
+ 0x30e000061c209c02,
+ 0x1800000000a01e04,
+ 0x30e000061c105c02,
+ 0x30e000061c001c02,
+ 0x9000000000001de7,
+ 0xd4004000084004a5,
+ 0x0c5400000013dc04,
+ 0x2000000000000007,
+ 0xd4004000084009a5,
+ 0xd4004000084007a5,
+ 0x1d00000004a0de84,
+ 0x1c00000004a09e84,
+ 0x1d00000000a05e84,
+ 0x1c00000000a01e84,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd4004000084004a5,
+ 0x0c5400000013dc04,
+ 0xd4004000084009a5,
+ 0xd4004000084007a5,
+ 0x1d00000004a0dc04,
+ 0x1c00000004a09c04,
+ 0x1d00000000a05c04,
+ 0x2000000000000007,
+ 0x1c00000000a01c04,
+ 0x9000000000001de7,
+ 0xd4004000084004a5,
+ 0x0c5400000013dc04,
+ 0xd4004000084009a5,
+ 0xd4004000084007a5,
+ 0x1100000004a0dc04,
+ 0x2000000000000007,
+ 0x1000000004a09c04,
+ 0x1100000000a05c04,
+ 0x1000000000a01c04,
+ 0x9000000000001de7,
+ 0xd4004000084004a5,
+ 0x0c5400000013dc04,
+ 0xd4004000084009a5,
+ 0x2000000000000007,
+ 0xd4004000084007a5,
+ 0x1800000000009de2,
+ 0x18fe00000000dde2,
+ 0x9000000000001de7,
+ 0xd4004000084004a5,
+ 0x0c5400000013dc04,
+ 0xd4004000084009a5,
+ 0x2000000000000007,
+ 0xd4004000084007a5,
+ 0x1800000000009de2,
+ 0x180000000400dde2,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0x2000000000000007,
+ 0xd400400008400785,
+ 0x7000c02828005c03,
+ 0x18fe00000000dde2,
+ 0x7000c02850009c03,
+ 0x3800000ffc001c02,
+ 0x1800000008a09c04,
+ 0x1800000004a05c04,
+ 0x2000000000000007,
+ 0x30ea00801c209c02,
+ 0x1800000000a01c04,
+ 0x30ea00801c105c02,
+ 0x30ea00801c001c02,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0x2000000000000007,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x7000c02828005c03,
+ 0x180000000400dde2,
+ 0x7000c02850009c03,
+ 0x3800000ffc001c02,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x198000000020dc04,
+ 0x1900000000209c04,
+ 0x30ee02020430dc02,
+ 0x2000000000000007,
+ 0x1880000000205c04,
+ 0x30ee020204209c02,
+ 0x1800000000201c04,
+ 0x30ee020204105c02,
+ 0x30ee020204001c02,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x2000000000000007,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x198000000020de04,
+ 0x1900000000209e04,
+ 0x30f004081030dc02,
+ 0x1880000000205e04,
+ 0x2000000000000007,
+ 0x30f0040810209c02,
+ 0x1800000000201e04,
+ 0x30f0040810105c02,
+ 0x30f0040810001c02,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0x2000000000000007,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x1d8000000020de84,
+ 0x1d00000000209e84,
+ 0x1c80000000205e84,
+ 0x1c00000000201e84,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x1d8000000020dc04,
+ 0x1d00000000209c04,
+ 0x1c80000000205c04,
+ 0x2000000000000007,
+ 0x1c00000000201c04,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x7000c01814005c03,
+ 0x2000000000000007,
+ 0x18fe00000000dde2,
+ 0x7000c0142c009c03,
+ 0x380000007c001c02,
+ 0x1800000008209c04,
+ 0x1800000004205c04,
+ 0x30f4108420209c02,
+ 0x1800000000201c04,
+ 0x2000000000000007,
+ 0x30f2082084105c02,
+ 0x30f4108420001c02,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x2000000000000007,
+ 0x7000c01414005c03,
+ 0x7000c01428009c03,
+ 0x380000007c001c02,
+ 0x18fe00000000dde2,
+ 0x1800000008209c04,
+ 0x1800000004205c04,
+ 0x1800000000201c04,
+ 0x2000000000000007,
+ 0x30f4108420209c02,
+ 0x30f4108420105c02,
+ 0x30f4108420001c02,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0x2000000000000007,
+ 0xd400400008400785,
+ 0x1900000000a05c04,
+ 0x1800000000a01c04,
+ 0x30de0001d0105c02,
+ 0x30de0001d0001c02,
+ 0x1800000000009de2,
+ 0x18fe00000000dde2,
+ 0x2000000000000007,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x18fe00000000dde2,
+ 0x1900000000a05e04,
+ 0x2000000000000007,
+ 0x1800000000009de2,
+ 0x1800000000a01e04,
+ 0x30e000061c105c02,
+ 0x30e000061c001c02,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0x2000000000000007,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x180000000400dde2,
+ 0x1d00000000a05e84,
+ 0x1800000000009de2,
+ 0x1c00000000a01e84,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x180000000400dde2,
+ 0x1d00000000a05c04,
+ 0x1800000000009de2,
+ 0x2000000000000007,
+ 0x1c00000000a01c04,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x18fe00000000dde2,
+ 0x2000000000000007,
+ 0x1100000000a05c04,
+ 0x1800000000009de2,
+ 0x1000000000a01c04,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0xd400400008400985,
+ 0x2000000000000007,
+ 0xd400400008400785,
+ 0x18fe00000000dde2,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0xd400400008400485,
+ 0x0c5400000013dc04,
+ 0x2000000000000007,
+ 0xd400400008400985,
+ 0xd400400008400785,
+ 0x180000000400dde2,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x2000000000000007,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x18fe00000000dde2,
+ 0x1880000000205c04,
+ 0x1800000000009de2,
+ 0x1800000000201c04,
+ 0x2000000000000007,
+ 0x30ee020204105c02,
+ 0x30ee020204001c02,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x2000000000000007,
+ 0x18fe00000000dde2,
+ 0x1880000000205e04,
+ 0x1800000000009de2,
+ 0x1800000000201e04,
+ 0x30f0040810105c02,
+ 0x30f0040810001c02,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x180000000400dde2,
+ 0x1c80000000205c04,
+ 0x1800000000009de2,
+ 0x2000000000000007,
+ 0x1c00000000201c04,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x180000000400dde2,
+ 0x2000000000000007,
+ 0x1c80000000205e84,
+ 0x1800000000009de2,
+ 0x1c00000000201e84,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0x2000000000000007,
+ 0xd400400008400745,
+ 0x18fe00000000dde2,
+ 0x1800000000a01c04,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x30de0001d0001c02,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x18fe00000000dde2,
+ 0x1800000000a01e04,
+ 0x1800000000009de2,
+ 0x2000000000000007,
+ 0x1800000000005de2,
+ 0x30e000061c001c02,
+ 0x9000000000001de7,
+ 0xd400400008400465,
+ 0x0c5400000013dc04,
+ 0xd400400008400965,
+ 0xd400400008400765,
+ 0x2000000000000007,
+ 0x180000000400dde2,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0xd400400008400945,
+ 0x2000000000000007,
+ 0xd400400008400745,
+ 0x180000000400dde2,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0xd400400008400445,
+ 0x0c5400000013dc04,
+ 0x2000000000000007,
+ 0xd400400008400945,
+ 0xd400400008400745,
+ 0x18fe00000000dde2,
+ 0x1800000000009de2,
+ 0x1000000000a01c04,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400405,
+ 0x0c5400000013dc04,
+ 0xd400400008400905,
+ 0xd400400008400705,
+ 0x18fe00000000dde2,
+ 0x1800000000201c04,
+ 0x1800000000009de2,
+ 0x2000000000000007,
+ 0x30ee020204001c02,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0xd400400008400405,
+ 0x0c5400000013dc04,
+ 0xd400400008400905,
+ 0xd400400008400705,
+ 0x2000000000000007,
+ 0x18fe00000000dde2,
+ 0x1800000000201e04,
+ 0x1800000000009de2,
+ 0x30f0040810001c02,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0xd400400008400425,
+ 0x2000000000000007,
+ 0x0c5400000013dc04,
+ 0xd400400008400925,
+ 0xd400400008400725,
+ 0x180000000400dde2,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x9000000000001de7,
+ 0x2000000000000007,
+ 0xd400400008400405,
+ 0x0c5400000013dc04,
+ 0xd400400008400905,
+ 0xd400400008400705,
+ 0x180000000400dde2,
+ 0x1800000000009de2,
+ 0x1800000000005de2,
+ 0x2000000000000007,
+ 0x9000000000001de7,
+ 0xd40040000840c485,
+ 0x0c5400000013dc04,
+ 0xd40040000840c985,
+ 0xd40040000840c785,
+ 0x18fe00000000dde2,
+ 0x4000000000001de4,
+ 0x9000000000001de7,
+/* 0x0f08: gk104_rcp_f64 */
+ 0x4000000000001de4,
+ 0x9000000000001de7,
+/* 0x0f18: gk104_rsq_f64 */
+ 0x4000000000001de4,
+ 0x9000000000001de7,
+ 0xc800000003f01cc5,
+ 0x2c00000100005c04,
+ 0x2c0000010800dc04,
+ 0x3000c3fffff09c04,
+ 0x680100000c1fdc03,
+ 0x4000000a60001c47,
+ 0x180000004000dde2,
+/* 0x0f60: spill_cfstack */
+ 0x78000009c0000007,
+ 0x0c0000000430dd02,
+ 0x4003ffffa0001ca7,
+ 0x2800406400001de4,
+ 0x2800406410005de4,
+ 0x180000000400dde2,
+ 0x547e18000000dd05,
+ 0x60000008e0000007,
+ 0x190ec0000431dc03,
+ 0x40000000000001f4,
+ 0x94000004c0009c85,
+ 0x2c00000100009c04,
+ 0x2c0000010800dc04,
+ 0x9400000020009ca5,
+ 0x9400000100011cc5,
+ 0x9400000140021cc5,
+ 0x9400000180031cc5,
+ 0x94000001c0041cc5,
+ 0x9400000200051cc5,
+ 0x9400000240061cc5,
+ 0x9400000280071cc5,
+ 0x94000002c0081cc5,
+ 0x9400000300091cc5,
+ 0x94000003400a1cc5,
+ 0x94000003800b1cc5,
+ 0x94000003c00c1cc5,
+ 0x94000004000d1cc5,
+ 0x94000004400e1cc5,
+ 0x94000004800f1cc5,
+ 0xc000000003f09ea5,
+ 0x94000000c0009ca5,
+ 0xc000000023f09ea5,
+ 0x94000000e0009ca5,
+ 0x2c00000084009c04,
+ 0x2c0000008800dc04,
+ 0x9400000040009ca5,
+ 0x2c0000008c009c04,
+ 0x2c0000009400dc04,
+ 0x9400000060009ca5,
+ 0x2c00000098009c04,
+ 0x2c0000009c00dc04,
+ 0x9400000080009ca5,
+ 0x2c000000c800dc04,
+ 0x0c0000001030dd02,
+ 0x4000000100001ea7,
+ 0x480100000c001c03,
+ 0x0800000000105c42,
+/* 0x10d8: shared_loop */
+ 0xc100000000309c85,
+ 0x9400000500009c85,
+ 0x0c00000010001d02,
+ 0x0800000000105d42,
+ 0x0c0000001030dd02,
+ 0x4003ffff40001ca7,
+/* 0x1108: shared_done */
+ 0x2800406420001de4,
+ 0x2800406430005de4,
+ 0xe000000000001c45,
+ 0xd000000003ffdcc5,
+ 0x9c000000000fdcc5,
+ 0x2c0000000c009c04,
+ 0x7000c0205020dc03,
+ 0x7000c01820209c03,
+ 0x5000406450209c03,
+ 0x500040644030dc03,
+ 0x480000000c209c03,
+ 0x4801000008001c03,
+ 0x0800000000105c42,
+/* 0x1170: search_cstack */
+ 0x280040646000dde4,
+ 0x8400000020009f05,
+ 0x190ec0002821dc03,
+ 0x40000000800001e7,
+ 0x0c00000040001c02,
+ 0x0800000000105c42,
+ 0x0c0000004030dd02,
+ 0x00029dff0ffc5cbf,
+/* 0x11b0: entry_found */
+ 0x8400000000009f85,
+ 0x2800406400001de4,
+ 0x2800406410005de4,
+ 0x9400000010009c85,
+ 0x4000000000001df4,
+/* 0x11d8: end_exit */
+ 0x9800000003ffdcc5,
+ 0xd000000000008007,
+ 0xa000000000004007,
+/* 0x11f0: end_cont */
+ 0xd000000000008007,
+ 0x3400c3fffc201c04,
+ 0xc000000003f01ec5,
+ 0xa000000000000007,
+};
+
+uint64_t gk104_builtin_offsets[] = {
+ 0x0000000000000000,
+ 0x00000000000000f0,
+ 0x0000000000000f08,
+ 0x0000000000000f18,
+};
--- /dev/null
+.section #gk110_builtin_code
+// DIV U32
+//
+// UNR recurrence (q = a / b):
+// look for z such that 2^32 - b <= b * z < 2^32
+// then q - 1 <= (a * z) / 2^32 <= q
+//
+// INPUT: $r0: dividend, $r1: divisor
+// OUTPUT: $r0: result, $r1: modulus
+// CLOBBER: $r2 - $r3, $p0 - $p1
+// SIZE: 22 / 14 * 8 bytes
+//
+gk110_div_u32:
+ sched 0x28282804280428
+ bfind u32 $r2 $r1
+ xor b32 $r2 $r2 0x1f
+ mov b32 $r3 0x1
+ shl b32 $r2 $r3 clamp $r2
+ cvt u32 $r1 neg u32 $r1
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ sched 0x28282828282828
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ sched 0x042c2828042804
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mov b32 $r3 $r0
+ mul high $r0 u32 $r0 u32 $r2
+ cvt u32 $r2 neg u32 $r1
+ add $r1 (mul u32 $r1 u32 $r0) $r3
+ set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ sched 0x20282e20042c28
+ $p0 add b32 $r0 $r0 0x1
+ $p0 set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ ret
+
+// DIV S32, like DIV U32 after taking ABS(inputs)
+//
+// INPUT: $r0: dividend, $r1: divisor
+// OUTPUT: $r0: result, $r1: modulus
+// CLOBBER: $r2 - $r3, $p0 - $p3
+//
+gk110_div_s32:
+ set $p2 0x1 lt s32 $r0 0x0
+ set $p3 0x1 lt s32 $r1 0x0 xor $p2
+ sched 0x28042804282820
+ cvt s32 $r0 abs s32 $r0
+ cvt s32 $r1 abs s32 $r1
+ bfind u32 $r2 $r1
+ xor b32 $r2 $r2 0x1f
+ mov b32 $r3 0x1
+ shl b32 $r2 $r3 clamp $r2
+ cvt u32 $r1 neg u32 $r1
+ sched 0x28282828282828
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ sched 0x28280428042828
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mul $r3 u32 $r1 u32 $r2
+ add $r2 (mul high u32 $r2 u32 $r3) $r2
+ mov b32 $r3 $r0
+ mul high $r0 u32 $r0 u32 $r2
+ cvt u32 $r2 neg u32 $r1
+ add $r1 (mul u32 $r1 u32 $r0) $r3
+ sched 0x2028042c28042c
+ set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ $p0 set $p0 0x1 ge u32 $r1 $r2
+ $p0 sub b32 $r1 $r1 $r2
+ $p0 add b32 $r0 $r0 0x1
+ $p3 cvt s32 $r0 neg s32 $r0
+ sched 0x2c200428042e04
+ $p2 cvt s32 $r1 neg s32 $r1
+ ret
+
+gk110_rcp_f64:
+gk110_rsq_f64:
+ ret
+
+.section #gk110_builtin_offsets
+.b64 #gk110_div_u32
+.b64 #gk110_div_s32
+.b64 #gk110_rcp_f64
+.b64 #gk110_rsq_f64
--- /dev/null
+uint64_t gk110_builtin_code[] = {
+/* 0x0000: gk110_div_u32 */
+ 0x08a0a0a010a010a0,
+ 0xe1800000009c000a,
+ 0x220000000f9c0808,
+ 0x74000000009fc00e,
+ 0xe2400000011c0c0a,
+ 0xe6010000009c2806,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0x08a0a0a0a0a0a0a0,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0x0810b0a0a010a010,
+ 0xd2000800019c080a,
+ 0xe4c03c00001c000e,
+ 0xe1c00400011c0002,
+ 0xe6010000009c280a,
+ 0xd0000c00001c0406,
+ 0xdb601c00011c041e,
+ 0xe088000001000406,
+ 0x0880a0b88010b0a0,
+ 0x4000000000800001,
+ 0xdb601c000100041e,
+ 0xe088000001000406,
+ 0x4000000000800001,
+ 0x19000000001c003c,
+/* 0x00f0: gk110_div_s32 */
+ 0xdb181c007f9c005e,
+ 0xdb1a08007f9c047e,
+ 0x08a010a010a0a080,
+ 0xe6100000001ce802,
+ 0xe6100000009ce806,
+ 0xe1800000009c000a,
+ 0x220000000f9c0808,
+ 0x74000000009fc00e,
+ 0xe2400000011c0c0a,
+ 0xe6010000009c2806,
+ 0x08a0a0a0a0a0a0a0,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0x08a0a010a010a0a0,
+ 0xd2000800019c080a,
+ 0xe1c00000011c040e,
+ 0xd2000800019c080a,
+ 0xe4c03c00001c000e,
+ 0xe1c00400011c0002,
+ 0xe6010000009c280a,
+ 0xd0000c00001c0406,
+ 0x0880a010b0a010b0,
+ 0xdb601c00011c041e,
+ 0xe088000001000406,
+ 0x4000000000800001,
+ 0xdb601c000100041e,
+ 0xe088000001000406,
+ 0x4000000000800001,
+ 0xe6010000000ce802,
+ 0x08b08010a010b810,
+ 0xe60100000088e806,
+ 0x19000000001c003c,
+/* 0x0218: gk110_rcp_f64 */
+/* 0x0218: gk110_rsq_f64 */
+ 0x19000000001c003c,
+};
+
+uint64_t gk110_builtin_offsets[] = {
+ 0x0000000000000000,
+ 0x00000000000000f0,
+ 0x0000000000000218,
+ 0x0000000000000218,
+};
// lazyness -> will just hardcode everything for the time being
-#include "target_lib_nvc0.asm.h"
-#include "target_lib_nve4.asm.h"
-#include "target_lib_nvf0.asm.h"
+#include "lib/gf100.asm.h"
+#include "lib/gk104.asm.h"
+#include "lib/gk110.asm.h"
void
TargetNVC0::getBuiltinCode(const uint32_t **code, uint32_t *size) const
{
switch (chipset & ~0xf) {
case 0xe0:
- *code = (const uint32_t *)&nve4_builtin_code[0];
- *size = sizeof(nve4_builtin_code);
+ *code = (const uint32_t *)&gk104_builtin_code[0];
+ *size = sizeof(gk104_builtin_code);
break;
case 0xf0:
case 0x100:
- *code = (const uint32_t *)&nvf0_builtin_code[0];
- *size = sizeof(nvf0_builtin_code);
+ *code = (const uint32_t *)&gk110_builtin_code[0];
+ *size = sizeof(gk110_builtin_code);
break;
default:
- *code = (const uint32_t *)&nvc0_builtin_code[0];
- *size = sizeof(nvc0_builtin_code);
+ *code = (const uint32_t *)&gf100_builtin_code[0];
+ *size = sizeof(gf100_builtin_code);
break;
}
}
switch (chipset & ~0xf) {
case 0xe0:
- return nve4_builtin_offsets[builtin];
+ return gk104_builtin_offsets[builtin];
case 0xf0:
case 0x100:
- return nvf0_builtin_offsets[builtin];
+ return gk110_builtin_offsets[builtin];
default:
- return nvc0_builtin_offsets[builtin];
+ return gf100_builtin_offsets[builtin];
}
}
+++ /dev/null
-//
-// DIV U32
-//
-// UNR recurrence (q = a / b):
-// look for z such that 2^32 - b <= b * z < 2^32
-// then q - 1 <= (a * z) / 2^32 <= q
-//
-// INPUT: $r0: dividend, $r1: divisor
-// OUTPUT: $r0: result, $r1: modulus
-// CLOBBER: $r2 - $r3, $p0 - $p1
-// SIZE: 22 / 14 * 8 bytes
-//
-bfind u32 $r2 $r1
-xor b32 $r2 $r2 0x1f
-mov b32 $r3 0x1
-shl b32 $r2 $r3 clamp $r2
-cvt u32 $r1 neg u32 $r1
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mov b32 $r3 $r0
-mul high $r0 u32 $r0 u32 $r2
-cvt u32 $r2 neg u32 $r1
-add $r1 (mul u32 $r1 u32 $r0) $r3
-set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-$p0 set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-ret
-//
-// DIV S32, like DIV U32 after taking ABS(inputs)
-//
-// INPUT: $r0: dividend, $r1: divisor
-// OUTPUT: $r0: result, $r1: modulus
-// CLOBBER: $r2 - $r3, $p0 - $p3
-//
-set $p2 0x1 lt s32 $r0 0x0
-set $p3 0x1 lt s32 $r1 0x0 xor $p2
-cvt s32 $r0 abs s32 $r0
-cvt s32 $r1 abs s32 $r1
-bfind u32 $r2 $r1
-xor b32 $r2 $r2 0x1f
-mov b32 $r3 0x1
-shl b32 $r2 $r3 clamp $r2
-cvt u32 $r1 neg u32 $r1
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mov b32 $r3 $r0
-mul high $r0 u32 $r0 u32 $r2
-cvt u32 $r2 neg u32 $r1
-add $r1 (mul u32 $r1 u32 $r0) $r3
-set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-$p0 set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-$p3 cvt s32 $r0 neg s32 $r0
-$p2 cvt s32 $r1 neg s32 $r1
-ret
-//
-// RCP F64: Newton Raphson reciprocal(x): r_{i+1} = r_i * (2.0 - x * r_i)
-//
-// INPUT: $r0d (x)
-// OUTPUT: $r0d (rcp(x))
-// CLOBBER: $r2 - $r7
-// SIZE: 9 * 8 bytes
-//
-nop
-ret
-// RSQ F64: Newton Raphson rsqrt(x): r_{i+1} = r_i * (1.5 - 0.5 * x * r_i * r_i)
-//
-// INPUT: $r0d (x)
-// OUTPUT: $r0d (rsqrt(x))
-// CLOBBER: $r2 - $r7
-// SIZE: 14 * 8 bytes
-//
-nop
-ret
+++ /dev/null
-
-static const uint32_t nvc0_builtin_code[] =
-{
- 0x04009c03,
- 0x78000000,
- 0x7c209cdd,
- 0x0010dd18,
- 0x08309c03,
- 0x60000000,
- 0x05605c18,
- 0x0810dc2a,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0000dde4,
- 0x28000000,
- 0x08001c43,
- 0x50000000,
- 0x05609c18,
- 0x0010430d,
- 0x0811dc03,
- 0x1b0e0000,
- 0x08104103,
- 0x48000000,
- 0x04000002,
- 0x08000000,
- 0x0811c003,
- 0x1b0e0000,
- 0x08104103,
- 0x48000000,
- 0x040000ac,
- 0x90001dff,
- 0xfc05dc23,
- 0x188e0000,
- 0xfc17dc23,
- 0x18c40000,
- 0x03301e18,
- 0x07305e18,
- 0x04009c03,
- 0x78000000,
- 0x7c209cdd,
- 0x0010dd18,
- 0x08309c03,
- 0x60000000,
- 0x05605c18,
- 0x0810dc2a,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0810dc03,
- 0x50000000,
- 0x0c209c43,
- 0x20040000,
- 0x0000dde4,
- 0x28000000,
- 0x08001c43,
- 0x50000000,
- 0x05609c18,
- 0x0010430d,
- 0x0811dc03,
- 0x1b0e0000,
- 0x08104103,
- 0x48000000,
- 0x04000002,
- 0x08000000,
- 0x0811c003,
- 0x1b0e0000,
- 0x08104103,
- 0x48000000,
- 0x040000ac,
- 0x01700e18,
- 0x05704a18,
- 0x90001dff,
- 0x00001c08,
- 0x90001dff,
- 0x00001c08,
- 0x90001dff,
-};
-
-static const uint16_t nvc0_builtin_offsets[NVC0_BUILTIN_COUNT] =
-{
- 0x0000,
- 0x00b0,
- 0x0180,
- 0x0188
-};
+++ /dev/null
-
-// Assembled from target_lib_nve4.asm by envyas -m nvc0 -V nve4 -W.
-
-static const uint64_t nve4_builtin_code[] =
-{
- 0x2282828042804287ULL,
- 0x7800000004009c03ULL,
- 0x380000007c209c82ULL,
- 0x180000000400dde2ULL,
- 0x6000000008309c03ULL,
- 0x1c00000005205d04ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x2282828282828287ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x2042c28280428047ULL,
- 0x200400000c209c43ULL,
- 0x280000000000dde4ULL,
- 0x5000000008001c43ULL,
- 0x1c00000005209d04ULL,
- 0x2006000000105c03ULL,
- 0x1b0e00000811dc03ULL,
- 0x4800000008104103ULL,
- 0x220282e20042c287ULL,
- 0x0800000004000002ULL,
- 0x1b0e00000811c003ULL,
- 0x4800000008104103ULL,
- 0x0800000004000002ULL,
- 0x9000000000001de7ULL,
- 0x188e0000fc05dc23ULL,
- 0x18c40000fc17dc23ULL,
- 0x2280428042828207ULL,
- 0x1c00000001201ec4ULL,
- 0x1c00000005205ec4ULL,
- 0x7800000004009c03ULL,
- 0x380000007c209c82ULL,
- 0x180000000400dde2ULL,
- 0x6000000008309c03ULL,
- 0x1c00000005205d04ULL,
- 0x2282828282828287ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x2282804280428287ULL,
- 0x200400000c209c43ULL,
- 0x500000000810dc03ULL,
- 0x200400000c209c43ULL,
- 0x280000000000dde4ULL,
- 0x5000000008001c43ULL,
- 0x1c00000005209d04ULL,
- 0x2006000000105c03ULL,
- 0x22028042c28042c7ULL,
- 0x1b0e00000811dc03ULL,
- 0x4800000008104103ULL,
- 0x0800000004000002ULL,
- 0x1b0e00000811c003ULL,
- 0x4800000008104103ULL,
- 0x0800000004000002ULL,
- 0x1c00000001200f84ULL,
- 0x22c200428042e047ULL,
- 0x1c00000005204b84ULL,
- 0x9000000000001de7ULL,
- 0xd4004000084004c5ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009c5ULL,
- 0xd4004000084007c5ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd4004000084004c5ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009c5ULL,
- 0xd4004000084007c5ULL,
- 0x1900000004a0dc04ULL,
- 0x1800000004a09c04ULL,
- 0x30de0001d030dc02ULL,
- 0x2000000000000007ULL,
- 0x1900000000a05c04ULL,
- 0x30de0001d0209c02ULL,
- 0x1800000000a01c04ULL,
- 0x30de0001d0105c02ULL,
- 0x30de0001d0001c02ULL,
- 0x9000000000001de7ULL,
- 0xd4004000084004a5ULL,
- 0x2000000000000007ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009a5ULL,
- 0xd4004000084007a5ULL,
- 0x1900000004a0de04ULL,
- 0x1800000004a09e04ULL,
- 0x30e000061c30dc02ULL,
- 0x1900000000a05e04ULL,
- 0x2000000000000007ULL,
- 0x30e000061c209c02ULL,
- 0x1800000000a01e04ULL,
- 0x30e000061c105c02ULL,
- 0x30e000061c001c02ULL,
- 0x9000000000001de7ULL,
- 0xd4004000084004a5ULL,
- 0x0c5400000013dc04ULL,
- 0x2000000000000007ULL,
- 0xd4004000084009a5ULL,
- 0xd4004000084007a5ULL,
- 0x1d00000004a0de84ULL,
- 0x1c00000004a09e84ULL,
- 0x1d00000000a05e84ULL,
- 0x1c00000000a01e84ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd4004000084004a5ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009a5ULL,
- 0xd4004000084007a5ULL,
- 0x1d00000004a0dc04ULL,
- 0x1c00000004a09c04ULL,
- 0x1d00000000a05c04ULL,
- 0x2000000000000007ULL,
- 0x1c00000000a01c04ULL,
- 0x9000000000001de7ULL,
- 0xd4004000084004a5ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009a5ULL,
- 0xd4004000084007a5ULL,
- 0x1100000004a0dc04ULL,
- 0x2000000000000007ULL,
- 0x1000000004a09c04ULL,
- 0x1100000000a05c04ULL,
- 0x1000000000a01c04ULL,
- 0x9000000000001de7ULL,
- 0xd4004000084004a5ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009a5ULL,
- 0x2000000000000007ULL,
- 0xd4004000084007a5ULL,
- 0x1800000000009de2ULL,
- 0x18fe00000000dde2ULL,
- 0x9000000000001de7ULL,
- 0xd4004000084004a5ULL,
- 0x0c5400000013dc04ULL,
- 0xd4004000084009a5ULL,
- 0x2000000000000007ULL,
- 0xd4004000084007a5ULL,
- 0x1800000000009de2ULL,
- 0x180000000400dde2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0x2000000000000007ULL,
- 0xd400400008400785ULL,
- 0x7000c02828005c03ULL,
- 0x18fe00000000dde2ULL,
- 0x7000c02850009c03ULL,
- 0x3800000ffc001c02ULL,
- 0x1800000008a09c04ULL,
- 0x1800000004a05c04ULL,
- 0x2000000000000007ULL,
- 0x30ea00801c209c02ULL,
- 0x1800000000a01c04ULL,
- 0x30ea00801c105c02ULL,
- 0x30ea00801c001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0x2000000000000007ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x7000c02828005c03ULL,
- 0x180000000400dde2ULL,
- 0x7000c02850009c03ULL,
- 0x3800000ffc001c02ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x198000000020dc04ULL,
- 0x1900000000209c04ULL,
- 0x30ee02020430dc02ULL,
- 0x2000000000000007ULL,
- 0x1880000000205c04ULL,
- 0x30ee020204209c02ULL,
- 0x1800000000201c04ULL,
- 0x30ee020204105c02ULL,
- 0x30ee020204001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x2000000000000007ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x198000000020de04ULL,
- 0x1900000000209e04ULL,
- 0x30f004081030dc02ULL,
- 0x1880000000205e04ULL,
- 0x2000000000000007ULL,
- 0x30f0040810209c02ULL,
- 0x1800000000201e04ULL,
- 0x30f0040810105c02ULL,
- 0x30f0040810001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0x2000000000000007ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x1d8000000020de84ULL,
- 0x1d00000000209e84ULL,
- 0x1c80000000205e84ULL,
- 0x1c00000000201e84ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x1d8000000020dc04ULL,
- 0x1d00000000209c04ULL,
- 0x1c80000000205c04ULL,
- 0x2000000000000007ULL,
- 0x1c00000000201c04ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x7000c01814005c03ULL,
- 0x2000000000000007ULL,
- 0x18fe00000000dde2ULL,
- 0x7000c0142c009c03ULL,
- 0x380000007c001c02ULL,
- 0x1800000008209c04ULL,
- 0x1800000004205c04ULL,
- 0x30f4108420209c02ULL,
- 0x1800000000201c04ULL,
- 0x2000000000000007ULL,
- 0x30f2082084105c02ULL,
- 0x30f4108420001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x2000000000000007ULL,
- 0x7000c01414005c03ULL,
- 0x7000c01428009c03ULL,
- 0x380000007c001c02ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000008209c04ULL,
- 0x1800000004205c04ULL,
- 0x1800000000201c04ULL,
- 0x2000000000000007ULL,
- 0x30f4108420209c02ULL,
- 0x30f4108420105c02ULL,
- 0x30f4108420001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0x2000000000000007ULL,
- 0xd400400008400785ULL,
- 0x1900000000a05c04ULL,
- 0x1800000000a01c04ULL,
- 0x30de0001d0105c02ULL,
- 0x30de0001d0001c02ULL,
- 0x1800000000009de2ULL,
- 0x18fe00000000dde2ULL,
- 0x2000000000000007ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x18fe00000000dde2ULL,
- 0x1900000000a05e04ULL,
- 0x2000000000000007ULL,
- 0x1800000000009de2ULL,
- 0x1800000000a01e04ULL,
- 0x30e000061c105c02ULL,
- 0x30e000061c001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0x2000000000000007ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x180000000400dde2ULL,
- 0x1d00000000a05e84ULL,
- 0x1800000000009de2ULL,
- 0x1c00000000a01e84ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x180000000400dde2ULL,
- 0x1d00000000a05c04ULL,
- 0x1800000000009de2ULL,
- 0x2000000000000007ULL,
- 0x1c00000000a01c04ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x18fe00000000dde2ULL,
- 0x2000000000000007ULL,
- 0x1100000000a05c04ULL,
- 0x1800000000009de2ULL,
- 0x1000000000a01c04ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400985ULL,
- 0x2000000000000007ULL,
- 0xd400400008400785ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400485ULL,
- 0x0c5400000013dc04ULL,
- 0x2000000000000007ULL,
- 0xd400400008400985ULL,
- 0xd400400008400785ULL,
- 0x180000000400dde2ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x2000000000000007ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x18fe00000000dde2ULL,
- 0x1880000000205c04ULL,
- 0x1800000000009de2ULL,
- 0x1800000000201c04ULL,
- 0x2000000000000007ULL,
- 0x30ee020204105c02ULL,
- 0x30ee020204001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x2000000000000007ULL,
- 0x18fe00000000dde2ULL,
- 0x1880000000205e04ULL,
- 0x1800000000009de2ULL,
- 0x1800000000201e04ULL,
- 0x30f0040810105c02ULL,
- 0x30f0040810001c02ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x180000000400dde2ULL,
- 0x1c80000000205c04ULL,
- 0x1800000000009de2ULL,
- 0x2000000000000007ULL,
- 0x1c00000000201c04ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x180000000400dde2ULL,
- 0x2000000000000007ULL,
- 0x1c80000000205e84ULL,
- 0x1800000000009de2ULL,
- 0x1c00000000201e84ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0x2000000000000007ULL,
- 0xd400400008400745ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000000a01c04ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x30de0001d0001c02ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000000a01e04ULL,
- 0x1800000000009de2ULL,
- 0x2000000000000007ULL,
- 0x1800000000005de2ULL,
- 0x30e000061c001c02ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400465ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400965ULL,
- 0xd400400008400765ULL,
- 0x2000000000000007ULL,
- 0x180000000400dde2ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400945ULL,
- 0x2000000000000007ULL,
- 0xd400400008400745ULL,
- 0x180000000400dde2ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400445ULL,
- 0x0c5400000013dc04ULL,
- 0x2000000000000007ULL,
- 0xd400400008400945ULL,
- 0xd400400008400745ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000000009de2ULL,
- 0x1000000000a01c04ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400405ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400905ULL,
- 0xd400400008400705ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000000201c04ULL,
- 0x1800000000009de2ULL,
- 0x2000000000000007ULL,
- 0x30ee020204001c02ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400405ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400905ULL,
- 0xd400400008400705ULL,
- 0x2000000000000007ULL,
- 0x18fe00000000dde2ULL,
- 0x1800000000201e04ULL,
- 0x1800000000009de2ULL,
- 0x30f0040810001c02ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0xd400400008400425ULL,
- 0x2000000000000007ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400925ULL,
- 0xd400400008400725ULL,
- 0x180000000400dde2ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x9000000000001de7ULL,
- 0x2000000000000007ULL,
- 0xd400400008400405ULL,
- 0x0c5400000013dc04ULL,
- 0xd400400008400905ULL,
- 0xd400400008400705ULL,
- 0x180000000400dde2ULL,
- 0x1800000000009de2ULL,
- 0x1800000000005de2ULL,
- 0x2000000000000007ULL,
- 0x9000000000001de7ULL,
- 0xd40040000840c485ULL,
- 0x0c5400000013dc04ULL,
- 0xd40040000840c985ULL,
- 0xd40040000840c785ULL,
- 0x18fe00000000dde2ULL,
- 0x4000000000001de4ULL,
- 0x9000000000001de7ULL,
- 0x4000000000001de4ULL,
- 0x9000000000001de7ULL,
- 0x4000000000001de4ULL,
- 0x9000000000001de7ULL,
- 0xc800000003f01cc5ULL,
- 0x2c00000100005c04ULL,
- 0x2c0000010800dc04ULL,
- 0x3000c3fffff09c04ULL,
- 0x680100000c1fdc03ULL,
- 0x4000000a60001c47ULL,
- 0x180000004000dde2ULL,
- 0x78000009c0000007ULL,
- 0x0c0000000430dd02ULL,
- 0x4003ffffa0001ca7ULL,
- 0x2800406400001de4ULL,
- 0x2800406410005de4ULL,
- 0x180000000400dde2ULL,
- 0x547e18000000dd05ULL,
- 0x60000008e0000007ULL,
- 0x190ec0000431dc03ULL,
- 0x40000000000001f4ULL,
- 0x94000004c0009c85ULL,
- 0x2c00000100009c04ULL,
- 0x2c0000010800dc04ULL,
- 0x9400000020009ca5ULL,
- 0x9400000100011cc5ULL,
- 0x9400000140021cc5ULL,
- 0x9400000180031cc5ULL,
- 0x94000001c0041cc5ULL,
- 0x9400000200051cc5ULL,
- 0x9400000240061cc5ULL,
- 0x9400000280071cc5ULL,
- 0x94000002c0081cc5ULL,
- 0x9400000300091cc5ULL,
- 0x94000003400a1cc5ULL,
- 0x94000003800b1cc5ULL,
- 0x94000003c00c1cc5ULL,
- 0x94000004000d1cc5ULL,
- 0x94000004400e1cc5ULL,
- 0x94000004800f1cc5ULL,
- 0xc000000003f09ea5ULL,
- 0x94000000c0009ca5ULL,
- 0xc000000023f09ea5ULL,
- 0x94000000e0009ca5ULL,
- 0x2c00000084009c04ULL,
- 0x2c0000008800dc04ULL,
- 0x9400000040009ca5ULL,
- 0x2c0000008c009c04ULL,
- 0x2c0000009400dc04ULL,
- 0x9400000060009ca5ULL,
- 0x2c00000098009c04ULL,
- 0x2c0000009c00dc04ULL,
- 0x9400000080009ca5ULL,
- 0x2c000000c800dc04ULL,
- 0x0c0000001030dd02ULL,
- 0x4000000100001ea7ULL,
- 0x480100000c001c03ULL,
- 0x0800000000105c42ULL,
- 0xc100000000309c85ULL,
- 0x9400000500009c85ULL,
- 0x0c00000010001d02ULL,
- 0x0800000000105d42ULL,
- 0x0c0000001030dd02ULL,
- 0x4003ffff40001ca7ULL,
- 0x2800406420001de4ULL,
- 0x2800406430005de4ULL,
- 0xe000000000001c45ULL,
- 0xd000000003ffdcc5ULL,
- 0x9c000000000fdcc5ULL,
- 0x2c0000000c009c04ULL,
- 0x7000c0205020dc03ULL,
- 0x7000c01820209c03ULL,
- 0x5000406450209c03ULL,
- 0x500040644030dc03ULL,
- 0x480000000c209c03ULL,
- 0x4801000008001c03ULL,
- 0x0800000000105c42ULL,
- 0x280040646000dde4ULL,
- 0x8400000020009f05ULL,
- 0x190ec0002821dc03ULL,
- 0x40000000800001e7ULL,
- 0x0c00000040001c02ULL,
- 0x0800000000105c42ULL,
- 0x0c0000004030dd02ULL,
- 0x00029dff0ffc5cbfULL,
- 0x8400000000009f85ULL,
- 0x2800406400001de4ULL,
- 0x2800406410005de4ULL,
- 0x9400000010009c85ULL,
- 0x4000000000001df4ULL,
- 0x9800000003ffdcc5ULL,
- 0xd000000000008007ULL,
- 0xa000000000004007ULL,
- 0xd000000000008007ULL,
- 0x3400c3fffc201c04ULL,
- 0xc000000003f01ec5ULL,
- 0xa000000000000007ULL
-};
-
-static const uint16_t nve4_builtin_offsets[NVC0_BUILTIN_COUNT] =
-{
- 0x0000,
- 0x00f0,
- 0x0f08,
- 0x0f18,
-};
+++ /dev/null
-//
-// DIV U32
-//
-// UNR recurrence (q = a / b):
-// look for z such that 2^32 - b <= b * z < 2^32
-// then q - 1 <= (a * z) / 2^32 <= q
-//
-// INPUT: $r0: dividend, $r1: divisor
-// OUTPUT: $r0: result, $r1: modulus
-// CLOBBER: $r2 - $r3, $p0 - $p1
-// SIZE: 22 / 14 * 8 bytes
-//
-sched 0x28282804280428
-bfind u32 $r2 $r1
-xor b32 $r2 $r2 0x1f
-mov b32 $r3 0x1
-shl b32 $r2 $r3 clamp $r2
-cvt u32 $r1 neg u32 $r1
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-sched 0x28282828282828
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-sched 0x042c2828042804
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mov b32 $r3 $r0
-mul high $r0 u32 $r0 u32 $r2
-cvt u32 $r2 neg u32 $r1
-add $r1 (mul u32 $r1 u32 $r0) $r3
-set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-sched 0x20282e20042c28
-$p0 add b32 $r0 $r0 0x1
-$p0 set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-ret
-//
-// DIV S32, like DIV U32 after taking ABS(inputs)
-//
-// INPUT: $r0: dividend, $r1: divisor
-// OUTPUT: $r0: result, $r1: modulus
-// CLOBBER: $r2 - $r3, $p0 - $p3
-//
-set $p2 0x1 lt s32 $r0 0x0
-set $p3 0x1 lt s32 $r1 0x0 xor $p2
-sched 0x28042804282820
-cvt s32 $r0 abs s32 $r0
-cvt s32 $r1 abs s32 $r1
-bfind u32 $r2 $r1
-xor b32 $r2 $r2 0x1f
-mov b32 $r3 0x1
-shl b32 $r2 $r3 clamp $r2
-cvt u32 $r1 neg u32 $r1
-sched 0x28282828282828
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-sched 0x28280428042828
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mul $r3 u32 $r1 u32 $r2
-add $r2 (mul high u32 $r2 u32 $r3) $r2
-mov b32 $r3 $r0
-mul high $r0 u32 $r0 u32 $r2
-cvt u32 $r2 neg u32 $r1
-add $r1 (mul u32 $r1 u32 $r0) $r3
-sched 0x2028042c28042c
-set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-$p0 set $p0 0x1 ge u32 $r1 $r2
-$p0 sub b32 $r1 $r1 $r2
-$p0 add b32 $r0 $r0 0x1
-$p3 cvt s32 $r0 neg s32 $r0
-sched 0x2c200428042e04
-$p2 cvt s32 $r1 neg s32 $r1
-ret
+++ /dev/null
-
-// Assembled from target_lib_nvf0.asm by envyas -m gk110 -W.
-
-static const uint64_t nvf0_builtin_code[] =
-{
-// DIV U32
-0x08a0a0a010a010a0ULL,
-0xe1800000009c000aULL,
-0x220000000f9c0808ULL,
-0x74000000009fc00eULL,
-0xe2400000011c0c0aULL,
-0xe6010000009c2806ULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0x08a0a0a0a0a0a0a0ULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0x0810b0a0a010a010ULL,
-0xd2000800019c080aULL,
-0xe4c03c00001c000eULL,
-0xe1c00400011c0002ULL,
-0xe6010000009c280aULL,
-0xd0000c00001c0406ULL,
-0xdb601c00011c041eULL,
-0xe088000001000406ULL,
-0x0880a0b88010b0a0ULL,
-0x4000000000800001ULL,
-0xdb601c000100041eULL,
-0xe088000001000406ULL,
-0x4000000000800001ULL,
-0x19000000001c003cULL,
-// DIV S32
-0xdb181c007f9c005eULL,
-0xdb1a08007f9c047eULL,
-0x08a010a010a0a080ULL,
-0xe6100000001ce802ULL,
-0xe6100000009ce806ULL,
-0xe1800000009c000aULL,
-0x220000000f9c0808ULL,
-0x74000000009fc00eULL,
-0xe2400000011c0c0aULL,
-0xe6010000009c2806ULL,
-0x08a0a0a0a0a0a0a0ULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0x08a0a010a010a0a0ULL,
-0xd2000800019c080aULL,
-0xe1c00000011c040eULL,
-0xd2000800019c080aULL,
-0xe4c03c00001c000eULL,
-0xe1c00400011c0002ULL,
-0xe6010000009c280aULL,
-0xd0000c00001c0406ULL,
-0x0880a010b0a010b0ULL,
-0xdb601c00011c041eULL,
-0xe088000001000406ULL,
-0x4000000000800001ULL,
-0xdb601c000100041eULL,
-0xe088000001000406ULL,
-0x4000000000800001ULL,
-0xe6010000000ce802ULL,
-0x08b08010a010b810ULL,
-0xe60100000088e806ULL,
-0x19000000001c003cULL,
-};
-
-static const uint16_t nvf0_builtin_offsets[NVC0_BUILTIN_COUNT] =
-{
- 0x0000,
- 0x00f0,
- /* Just point at a ret instruction for now. */
- 0x00f0 - 8,
- 0x00f0 - 8
-};
projects / platform / upstream / mesa.git / commitdiff