From 8cfbda97eb2e1548d9f55f2b3145727a7e0302e1 Mon Sep 17 00:00:00 2001 From: notaz Date: Tue, 22 Feb 2011 14:04:47 +0200 Subject: [PATCH] gte: starting some NEON code --- Makefile | 3 + libpcsxcore/gte_neon.h | 1 + libpcsxcore/gte_neon.s | 343 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 347 insertions(+) create mode 100644 libpcsxcore/gte_neon.h create mode 100644 libpcsxcore/gte_neon.s diff --git a/Makefile b/Makefile index 0a99925b..95bc0f29 100644 --- a/Makefile +++ b/Makefile @@ -33,6 +33,9 @@ OBJS += libpcsxcore/cdriso.o libpcsxcore/cdrom.o libpcsxcore/cheat.o libpcsxcore libpcsxcore/psxcommon.o libpcsxcore/psxcounters.o libpcsxcore/psxdma.o libpcsxcore/psxhle.o \ libpcsxcore/psxhw.o libpcsxcore/psxinterpreter.o libpcsxcore/psxmem.o libpcsxcore/r3000a.o \ libpcsxcore/sio.o libpcsxcore/socket.o libpcsxcore/spu.o +ifeq "$(ARCH)" "arm" +OBJS += libpcsxcore/gte_neon.o +endif # dynarec ifndef NO_NEW_DRC OBJS += libpcsxcore/new_dynarec/new_dynarec.o libpcsxcore/new_dynarec/linkage_arm.o diff --git a/libpcsxcore/gte_neon.h b/libpcsxcore/gte_neon.h new file mode 100644 index 00000000..54a7b44c --- /dev/null +++ b/libpcsxcore/gte_neon.h @@ -0,0 +1 @@ +void gteRTPT_neon(void *cp2_regs); diff --git a/libpcsxcore/gte_neon.s b/libpcsxcore/gte_neon.s new file mode 100644 index 00000000..023c7c6a --- /dev/null +++ b/libpcsxcore/gte_neon.s @@ -0,0 +1,343 @@ +/* + * (C) Gražvydas "notaz" Ignotas, 2011 + * + * This work is licensed under the terms of any of these licenses + * (at your option): + * - GNU GPL, version 2 or later. + * - GNU LGPL, version 2.1 or later. + * See the COPYING file in the top-level directory. + */ + + +.bss +.align 6 @ cacheline + +scratch: +.rept 8*8*2/4 + .word 0 +.endr + +.text +.align 2 + +@ approximate signed gteIR|123 flags +@ in: rl/rh with packed gteIR|123 +@ trash: r2,r3,r4 +.macro do_irs_flags rl rh + mov r4, \rl, ror #16 + adds r2, \rl, #1<<16 + subvcs r3, \rl, #1<<16 + orrvs lr, #(1<<31)|(1<<24) @ IR1/limB1 + adds r2, r4, #1<<16 + subvcs r3, r4, #1<<16 + mov r4, \rh, lsl #16 + orrvs lr, #(1<<31) + orrvs lr, #(1<<23) @ IR2/limB2 + adds r2, r4, #1<<16 + subvcs r3, r4, #1<<16 + orrvs lr, #(1<<22) @ IR3/limB3 +.endm + + +/* + * q | d | c code / phase 1 phase 2 scratch + * 0 0 gteR1* [s16] gteMAC3 = gteMAC3 \ v=0 * + * 1 gteR2* gteIR1-3 = gteIR1-3 / * + * 1 2 gteR3* gteMAC3 = gteMAC3 \ v=1 + * 3 * gteIR1-3 = gteIR1-3 / + * 2 4 gteTRX<<12 [s64] gteOFX [s64] gteMAC3 \ v=2 + * 5 gteTRY<<12 gteOFY [s64] gteIR1-3 / + * 3 6 gteTRZ<<12 gteDQA [s64] min gteMAC|12 v=012 + * 7 0 gteDQB [s64] max gteMAC|12 + * 4 8 VXYZ(v) / gteMAC1,2 [s32] min gteIR|123 + * 9 * / gteMAC3 max gteIR|123 + * 5 10 gteIR1-3 [s16] gteIR1-3 v=2 quotients 12 + * 11 0 quotient 3 + * 6 12 gteH (adj. for cmp) + * 13 gteH (float for div) + * ... + * 15 30 0 + * 31 0 + */ +.global gteRTPT_neon @ r0=CP2 (d,c), +gteRTPT_neon: + push {r4-r11,lr} + +@ fmrx r4, fpscr @ vmrs? + movw r1, #:lower16:scratch + movt r1, #:upper16:scratch + mov r12, #0 + veor q15, q15 + + add r3, r0, #4*32 + vldmia r3, {d0-d2} @ gteR* [16*9] + add r3, r0, #4*(32+5) + vldmia r3, {d4-d5} @ gteTR* + vshl.i64 d2, d2, #32 @ | + add r3, r0, #4*(32+26) + vld1.32 d11[0], [r3] @ gteH + vsri.u64 d2, d1, #32 @ | + add r3, r0, #4*19 + vld1.32 d14[0], [r3] @ gteSZ3 + vshll.s32 q3, d5, #12 + vshll.s32 q2, d4, #12 @ gteTRX + vshl.i64 d1, d1, #16 @ | + add r3, r0, #4*16 + vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3 + vmovl.s16 q6, d11 @ gteH + vsri.u64 d1, d0, #48 @ | + + vmov.i32 d22, #0x7fffffff + vmov.i32 d23, #0x80000000 + mov r3, #3 @ counter + mov r2, r0 @ VXYZ(0) +0: + vldmia r2!, {d8} @ VXYZ(v) + vmov.16 d8[3], r12 @ kill unused upper vector + + vmull.s16 q8, d0, d8 + vmull.s16 q9, d1, d8 + vmull.s16 q10, d2, d8 + vpaddl.s32 q8, q8 + vpaddl.s32 q9, q9 + vpaddl.s32 q10, q10 + vadd.s64 d16, d17 @ d16=d0.16[2]*d8.16[2], as + vadd.s64 d18, d19 @ d8[3]==0, so won't affect + vadd.s64 d20, d21 @ QC + vadd.s64 d16, d4 + vadd.s64 d18, d5 + vadd.s64 d20, d6 + vqshrn.s64 d8, q8, #12 @ gteMAC1 + vqshrn.s64 d18, q9, #12 @ gteMAC2 + vqshrn.s64 d9, q10, #12 @ gteMAC3 + vsli.u64 d8, d18, #32 @ gteMAC|12 + vmov.32 d9[1], r12 + vqmovn.s32 d10, q4 @ gteIR1-3; losing 2 cycles? + vmin.s32 d22, d8 @ min gteMAC|12 + vmax.s32 d23, d8 @ max gteMAC|12 + subs r3, #1 + vst1.32 {d9,d10}, [r1, :64]! + bgt 0b + + vst1.32 {d22,d23}, [r1, :64]! @ min/max gteMAC|12 (for flags) + + @ - phase2 - + sub r1, r1, #8*2*4 + vldmia r1, {d0-d3} @ note: d4,d5 is for gteOF|XY + + vmov d20, d0 @ gteMAC3 v=0 + vmin.s16 d24, d1, d3 @ | find min IR + vshr.s32 d22, d12, #1 @ || gteH (adjust for cmp) + vmax.s16 d25, d1, d3 @ | .. also max, for flag gen + vsli.u64 d20, d2, #32 @ gteMAC3 v=1 + vmov d21, d9 @ ... v=2 + + vmov.i32 q14, #0xffff @ 0xffff[32] + vmax.s32 q10, q15 + vmov.i32 q13, #1 + vdup.32 q11, d22[0] @ gteH/2 + vmin.u32 q10, q14 @ saturate to 0..0xffff - fSZ(v) + vmin.s16 d24, d10 @ | find min/max IR + vmax.s16 d25, d10 @ | + + vclt.u32 q11, q11, q10 @ gteH/2 < fSZ(v)? + add r3, r0, #4*17 + vst1.32 d20, [r3]! @ | writeback fSZ(v) + vand q11, q10, q11 + vst1.32 d21[0], [r3] @ | + vmax.u32 q10, q11, q13 @ make divisor 1 if not + add r3, r1, #8*8 + vstmia r3, {q12} @ min/max IR for flags + vcvt.f32.u32 q10, q10 + vshl.u32 d13, d12, #16 @ | preparing gteH + + @ while NEON's busy we calculate some flags on ARM + add r2, r1, #8*2*3 + mov lr, #0 @ gteFLAG + ldmia r2, {r4-r7} @ min/max gteMAC|12 + subs r2, r4, #1 + orrvs lr, #(1<<31)|(1<<27) + subs r3, r5, #1 + orrvs lr, #(1<<31)|(1<<26) + adds r2, r6, #1 + orrvs lr, #(1<<30) + adds r3, r7, #1 + orrvs lr, #(1<<29) + ldr r4, [r1, #0] @ gteMAC3 v=0 + ldr r5, [r1, #8*2] @ ... v=1 + ldr r6, [r1, #8*4] @ ... v=2 + + add r3, r0, #4*(32+24) + vld1.32 d4, [r3] @ || gteOF|XY + add r3, r0, #4*(32+27) + vld1.32 d6, [r3] @ || gteDQAB + + @ divide +.if 1 + vrecpe.f32 q11, q10 @ inv + vmovl.s32 q2, d4 @ || gteOFXY [64] + vmovl.s32 q3, d6 @ || gteDQAB [64] + vrecps.f32 q12, q10, q11 @ step + vcvt.f32.u32 d13, d13 @ | gteH (float for div) + vmul.f32 q11, q12, q11 @ better inv + vdup.32 q13, d13[0] @ | +@ vrecps.f32 q12, q10, q11 @ step +@ vmul.f32 q11, q12, q11 @ better inv + vmul.f32 q10, q13, q11 @ result +.else + vmovl.s32 q2, d4 @ || gteOFXY [64] + vmovl.s32 q3, d6 @ || gteDQAB [64] + vcvt.f32.u32 d13, d13 @ | gteH (float for div) + vdup.32 q13, d13[0] @ | + + vpush {q0} + vmov q0, q10 @ to test against C code + vdiv.f32 s0, s26, s0 + vdiv.f32 s1, s26, s1 + vdiv.f32 s2, s26, s2 + vmov q10, q0 + vpop {q0} +.endif + +@ approximate gteMACx flags +@ in: rr 123 as gteMAC 123, *flags +@ trash: r2,r3 +.macro do_mac_flags rr1 rr2 rr3 nflags pflags + subs r2, \rr1, #1 + subvcs r3, \rr2, #1 + subvcs r2, \rr3, #1 + orrvs lr, #\nflags + adds r3, \rr1, #1 + addvcs r2, \rr2, #1 + addvcs r3, \rr3, #1 + orrvs lr, #\pflags +.endm + + do_mac_flags r4, r5, r6, (1<<31)|(1<<25), (1<<27) @ MAC3 + orr r7, r4, r5 + add r4, r1, #8*8 + orr r3, r7, r6 + ldmia r4, {r7,r8,r10,r11} @ min/max IR + + movs r3, r3, lsr #16 + orrne lr, #(1<<31) + orrne lr, #(1<<18) @ fSZ (limD) + +@ vadd.f32 q10, q @ adjust for vcvt rounding mode + vcvt.u32.f32 q8, q10 + vmovl.s16 q9, d1 @ expand gteIR|12 v=0 + vmovl.s16 q10, d3 @ expand gteIR|12 v=1 + add r6, r1, #8*10 + vstmia r6, {q8} @ wb quotients for flags (pre-limE) + vqshl.u32 q8, #15 + vmovl.s16 q11, d10 @ expand gteIR|12 v=2 + vshr.u32 q8, #15 @ quotients (limE) + vdup.32 d24, d16[0] + vdup.32 d25, d16[1] + vdup.32 d26, d17[0] @ quotient (dup) + + mov r4, r7, ror #16 + mov r5, r10, ror #16 + subs r2, r7, #1<<16 + addvcs r3, r10, #1<<16 + orrvs lr, #(1<<31) + orrvs lr, #(1<<23) @ IR2/limB2 + subs r2, r4, #1<<16 + addvcs r3, r5, #1<<16 + mov r4, r8, lsl #16 + mov r5, r11, lsl #16 + orrvs lr, #(1<<31)|(1<<24) @ IR1/limB1 + subs r2, r4, #1<<16 + addvcs r3, r5, #1<<16 + orrvs lr, #(1<<22) @ IR3/limB3 + + vmull.s32 q9, d18, d24 @ gteIR|12 * quotient v=0 + vmull.s32 q10, d20, d25 @ ... v=1 + vmull.s32 q11, d22, d26 @ ... v=2 + vadd.s64 q9, q2 @ gteOF|XY + gteIR|12 * quotient + vadd.s64 q10, q2 @ ... v=1 + vadd.s64 q11, q2 @ ... v=2 + vqmovn.s64 d18, q9 @ saturate to 32 v=0 + vqmovn.s64 d19, q10 @ ... v=1 + vqmovn.s64 d20, q11 @ ... v=2 + vmin.s32 d14, d18, d19 @ || find min/max fS|XY(v) [32] + vmax.s32 d15, d18, d19 @ || for flags + vmin.s32 d14, d20 + vmax.s32 d15, d20 + vqshl.s32 q11, q9, #5 @ 11bit precision, v=0,1 + vqshl.s32 d24, d20, #5 @ ... v=2 + vmull.s32 q13, d6, d17 @ | gteDQA * quotient v=2 + vpmin.s32 d16, d14, d15 @ || also find min/max in pair + vpmax.s32 d17, d14, d15 @ || + vshr.s32 q11, #16+5 @ can't vqshrn because of insn + vshr.s32 d24, #16+5 @ encoding doesn't allow 21 :( + vqshl.s32 q7, #5 @ || min/max pairs shifted + vsli.u64 d16, d17, #32 @ || pack in-pair min/max + vadd.s64 d26, d7 @ | gteDQB + gteDQA * quotient + vmovn.s32 d12, q11 @ fS|XY(v) [s16] v=0,1 + vmovn.s32 d13, q12 @ 3 + vstmia r1, {d14-d16} @ || other cacheline than quotients + add r3, r0, #4*12 + vst1.32 d12, [r3]! @ writeback fS|XY v=0,1 + vst1.32 d13[0], [r3] + + vqshrn.s64 d26, q13, #12 @ | gteMAC0 + vmovl.u16 q5, d10 @ expand gteIR|123 v=2 + + vmov.i32 d13, #0x1000 + vmax.s32 d12, d26, d30 + + add r3, r0, #4*24 + vst1.32 d26[0], [r3]! @ gteMAC0 + vst1.32 d8, [r3]! @ gteMAC123 (last iteration) + vst1.32 d9[0], [r3] + + vmin.s32 d12, d13 @ | gteIR0 + + @ ~6 cycles + ldmia r6, {r4-r6} @ quotients + orr r4, r5 + orr r4, r6 + add r3, r0, #4*12 + movs r4, r4, lsr #17 + orrne lr, #(1<<31) @ limE + orrne lr, #(1<<17) @ limE + + add r3, r0, #4*8 + vst1.32 d12[0], [r3]! @ gteIR0 + vst1.32 d10, [r3]! @ gteIR12 + vst1.32 d11[0], [r3] @ ..3 + + @ ~19 cycles + ldmia r1, {r4-r9} + subs r2, r4, #1<<21 @ min fSX + addvcs r3, r6, #1<<21 @ max fSX + orrvs lr, #(1<<31) @ limG1 + orrvs lr, #(1<<14) + subs r2, r5, #1<<21 @ min fSY + addvcs r3, r7, #1<<21 @ max fSY + orrvs lr, #(1<<31) @ limG2 + orrvs lr, #(1<<13) + adds r2, r9, #1 + orrvs lr, #(1<<31) @ F + orrvs lr, #(1<<16) + subs r3, r8, #1 + orrvs lr, #(1<<31) @ F + + ldr r4, [r0, #4*24] @ gteMAC0 + orrvs lr, #(1<<15) + + adds r3, r4, #1 + orrvs lr, #(1<<16) + orrvs lr, #(1<<31) @ F + subs r2, r4, #1 + orrvs lr, #(1<<15) + orrvs lr, #(1<<31) @ F + cmp r4, #0x1000 + orrhi lr, #(1<<12) + + str lr, [r0, #4*(32+31)] @ gteFLAG + + pop {r4-r11,pc} + +@ vim:filetype=armasm -- 2.39.5