Re: [PATCH 1/2] powerpc: Batch up loads/stores on saving and restoring Altivec
From: Cyril Bur <hidden>
Date: 2016-03-11 04:25:42
On Thu, 10 Mar 2016 16:37:47 +1100 Cyril Bur [off-list ref] wrote:
On Thu, 10 Mar 2016 10:01:07 +1100 Michael Neuling [off-list ref] wrote:quoted
On Tue, 2016-03-01 at 16:55 +1100, Cyril Bur wrote:quoted
Currently the assembly to save and restore Altivec registers boils down to a load immediate of the offset of the specific Altivec register in memory followed by the load/store which repeats in sequence for each Altivec register. This patch attempts to do better by loading up four registers with immediates so that the loads and stores can be batched up and better pipelined by the processor. This patch results in four load/stores in sequence and one add between groups of four. Also, by using a pair of base registers it means that the result of the add is not needed by the following instruction.What the performance improvement?So I have some numbers. This is the same context switch benchmark that was used for my other series, a modified version of: http://www.ozlabs.org/~anton/junkcode/context_switch2.c We pingpong across a pipe and touch FP/VMX/VSX or some combinations of both. The numbers are the number of pingpongs per second, the tests run for 52 seconds with the first two seconds of results being ignored to allow for the test to stabilise. Run 1 Touched Facility Average Stddev Speedup %Speedup Speedup/Stddev None 1845984 8261.28 15017.32 100.8201 1.817793 FPU 1639296 3966.94 54770.04 103.4565 13.80660 FPU + VEC 1555836 3708.59 34533.72 102.2700 9.311813 FPU + VSX 1523202 78984.7 -362.64 99.97619 -0.00459 VEC 1631529 23665.5 -11818.0 99.28085 -0.49937 VEC + VSX 1543007 24614.0 32330.52 102.1401 1.313500 VSX 1554007 28723.0 40296.56 102.6621 1.402932 FPU + VEC + VSX 1546072 17201.1 41687.28 102.7710 2.423519 Run 2 Touched Facility Average Stddev Speedup %Speedup Speedup/Stddev None 1837869 30263.4 -7780.6 99.57843 -0.25709 FPU 1587883 70260.6 -23927 98.51546 -0.34055 FPU + VEC 1552703 13563.6 37243.4 102.4575 2.745831 FPU + VSX 1558519 13706.7 32365.9 102.1207 2.361308 VEC 1651599 1388.83 13474.3 100.8225 9.701918 VEC + VSX 1552319 1752.77 42443.2 102.8110 24.21487 VSX 1559806 7891.66 55542.5 103.6923 7.038124 FPU + VEC + VSX 1549330 22148.1 29010.8 101.9082 1.309849 I can't help but notice these are noisy. These were run on KVM on a fairly busy box. I wonder if the numbers smooth out on an otherwise idle machine. It doesn't look super consistent across two runs.
Did four runs on a OpenPower system, booted into a buildroot ramfs with basically nothing running except ssh daemon and busybox. Run 1 Touched Average Stddev % of Average Speedup %Speedup Speedup/Stddev None 1772303.48 1029.674394 0.05809808564 -2514.2 99.85834038 -2.441742764 FPU 1564827.36 623.4840148 0.03984362945 21815.16 101.4138035 34.98912479 FPU + VEC 1485572.76 865.0766997 0.05823186336 34317.64 102.3646869 39.6700547 FPU + VSX 1485859.72 606.5295067 0.04082010559 47010.68 103.267242 77.50765541 VEC 1571237.88 927.2565997 0.0590143995 8459.76 100.5413283 9.123429268 VEC + VSX 1480972.92 652.0581181 0.04402903722 47769.32 103.3330449 73.25929802 VSX 1468496.12 140955.9898 9.598662731 10461.72 100.7175222 0.0742197619 All 1480732.36 777.801684 0.05252817491 50908.56408 103.5604782 65.45185634 Run 2 Touched Average Stddev % of Average Speedup %Speedup Speedup/Stddev None 1773836.72 868.7485496 0.0489756774 -2374.68 99.86630645 -2.733449168 FPU 1564397.8 688.6031113 0.04401713626 21650.48 101.4033717 31.44115913 FPU + VEC 1484855.48 905.2225623 0.06096368128 33020.52 102.274399 36.47779162 FPU + VSX 1486762.64 933.4386234 0.06278329831 48576.48 103.3776212 52.04035786 VEC 1571551.44 785.1688538 0.04996138426 7980.44 100.5103983 10.16397933 VEC + VSX 1480887.72 574.0000725 0.03876053969 49535.16 103.4607239 86.29817725 VSX 1489902.32 546.0581355 0.03665059972 35301.64 102.4268956 64.64813489 All 1480020.88 733.4377717 0.04955590705 54243.28 103.8044699 73.95757636 Run 3 Touched Average Stddev % of Average Speedup %Speedup Speedup/Stddev None 1766262.8 614.1256934 0.03476978021 -5233.36 99.70457966 -8.521643136 FPU 1565494.04 627.6075064 0.04009006041 22975.6 101.4894862 36.60823009 FPU + VEC 1484286.4 569.3538049 0.03835875643 34148.76 102.3548634 59.97810098 FPU + VSX 1486067.76 847.4846952 0.0570286711 45883.84 103.1859709 54.14120191 VEC 1570977.16 860.8764485 0.05479878832 11356.4 100.7281514 13.1916723 VEC + VSX 1480634.88 912.5305617 0.06163103234 48774.88 103.4064001 53.45013312 VSX 1488969.72 697.1188532 0.04681887374 29984.4 102.0551543 43.01189082 All 1479573.16 681.8953286 0.04608730052 47822.32 103.3401286 70.13146739 Run 4 Touched Average Stddev % of Average Speedup %Speedup Speedup/Stddev None 1773258.28 1230.022665 0.06936511613 -252.88 99.98574128 -0.2055897075 FPU 1564657.04 665.6801021 0.04254479321 20985.88 101.3594787 31.52547287 FPU + VEC 1485552 542.697721 0.03653172161 34167.56 102.3541358 62.95873132 FPU + VSX 1487035.84 659.5565412 0.04435377571 44355.44 103.074516 67.25039815 VEC 1570597.04 726.2013259 0.0462372784 10325.4 100.6617694 14.21837118 VEC + VSX 1480685.84 874.2851391 0.05904595799 49180.12 103.4355518 56.25180825 VSX 1489545.44 718.5865991 0.0482420059 34501.6 102.3711725 48.01314141 All 1480782.08 702.6882445 0.04745385928 49976.52 103.4928939 71.12189565 Compared to the two runs done under KVM, these are so much more consistent. I can't explain Run 1 VSX result, there are a few possibilities. Based on all the other results I'm seeing being so much more consistent, I'm liking the effect bare metal. To add fuel to the green fire, the lack of speedup on the None case is expected, the optimisation would not be used.
quoted
quoted
Signed-off-by: Cyril Bur <redacted> --- These patches need to be applied on to of my rework of FPU/VMX/VSX switching: https://patchwork.ozlabs.org/patch/589703/ I left in some of my comments indicating if functions are called from C or not. Looking at them now, they might be a bit much, let me know what you think.I think that's ok, although they are likely to get stale quickly.quoted
Tested 64 bit BE and LE under KVM, not sure how I can test 32bit. arch/powerpc/include/asm/ppc_asm.h | 63 ++++++++++++++++++++++++++++++-------- arch/powerpc/kernel/tm.S | 6 ++-- arch/powerpc/kernel/vector.S | 20 +++++++++--- 3 files changed, 70 insertions(+), 19 deletions(-)diff --git a/arch/powerpc/include/asm/ppc_asm.h b/arch/powerpc/include/asm/ppc_asm.h index 499d9f8..5ba69ed 100644 --- a/arch/powerpc/include/asm/ppc_asm.h +++ b/arch/powerpc/include/asm/ppc_asm.h@@ -110,18 +110,57 @@ END_FW_FTR_SECTION_IFSET(FW_FEATURE_SPLPAR) #define REST_16FPRS(n, base) REST_8FPRS(n, base); REST_8FPRS(n+8, base) #define REST_32FPRS(n, base) REST_16FPRS(n, base); REST_16FPRS(n+16, base) -#define SAVE_VR(n,b,base) li b,16*(n); stvx n,base,b -#define SAVE_2VRS(n,b,base) SAVE_VR(n,b,base); SAVE_VR(n+1,b,base) -#define SAVE_4VRS(n,b,base) SAVE_2VRS(n,b,base); SAVE_2VRS(n+2,b,base) -#define SAVE_8VRS(n,b,base) SAVE_4VRS(n,b,base); SAVE_4VRS(n+4,b,base) -#define SAVE_16VRS(n,b,base) SAVE_8VRS(n,b,base); SAVE_8VRS(n+8,b,base) -#define SAVE_32VRS(n,b,base) SAVE_16VRS(n,b,base); SAVE_16VRS(n+16,b,base) -#define REST_VR(n,b,base) li b,16*(n); lvx n,base,b -#define REST_2VRS(n,b,base) REST_VR(n,b,base); REST_VR(n+1,b,base) -#define REST_4VRS(n,b,base) REST_2VRS(n,b,base); REST_2VRS(n+2,b,base) -#define REST_8VRS(n,b,base) REST_4VRS(n,b,base); REST_4VRS(n+4,b,base) -#define REST_16VRS(n,b,base) REST_8VRS(n,b,base); REST_8VRS(n+8,b,base) -#define REST_32VRS(n,b,base) REST_16VRS(n,b,base); REST_16VRS(n+16,b,base)Can you use consistent names between off and reg? in the belowquoted
+#define __SAVE_4VRS(n,off0,off1,off2,off3,base) \ + stvx n,base,off0; \ + stvx n+1,base,off1; \ + stvx n+2,base,off2; \ + stvx n+3,base,off3 + +/* Restores the base for the caller */Can you make this: /* Base: non-volatile, reg[0-4]: volatile */quoted
+#define SAVE_32VRS(reg0,reg1,reg2,reg3,reg4,base) \ + addi reg4,base,64; \ + li reg0,0; li reg1,16; li reg2,32; li reg3,48; \ + __SAVE_4VRS(0,reg0,reg1,reg2,reg3,base); \ + addi base,base,128; \ + __SAVE_4VRS(4,reg0,reg1,reg2,reg3,reg4); \ + addi reg4,reg4,128; \ + __SAVE_4VRS(8,reg0,reg1,reg2,reg3,base); \ + addi base,base,128; \ + __SAVE_4VRS(12,reg0,reg1,reg2,reg3,reg4); \ + addi reg4,reg4,128; \ + __SAVE_4VRS(16,reg0,reg1,reg2,reg3,base); \ + addi base,base,128; \ + __SAVE_4VRS(20,reg0,reg1,reg2,reg3,reg4); \ + addi reg4,reg4,128; \ + __SAVE_4VRS(24,reg0,reg1,reg2,reg3,base); \ + __SAVE_4VRS(28,reg0,reg1,reg2,reg3,reg4); \ + subi base,base,384You can swap these last two lines which will make base reuse quicker later. Although that might not be needed.quoted
+#define __REST_4VRS(n,off0,off1,off2,off3,base) \ + lvx n,base,off0; \ + lvx n+1,base,off1; \ + lvx n+2,base,off2; \ + lvx n+3,base,off3 + +/* Restores the base for the caller */ +#define REST_32VRS(reg0,reg1,reg2,reg3,reg4,base) \ + addi reg4,base,64; \ + li reg0,0; li reg1,16; li reg2,32; li reg3,48; \ + __REST_4VRS(0,reg0,reg1,reg2,reg3,base); \ + addi base,base,128; \ + __REST_4VRS(4,reg0,reg1,reg2,reg3,reg4); \ + addi reg4,reg4,128; \ + __REST_4VRS(8,reg0,reg1,reg2,reg3,base); \ + addi base,base,128; \ + __REST_4VRS(12,reg0,reg1,reg2,reg3,reg4); \ + addi reg4,reg4,128; \ + __REST_4VRS(16,reg0,reg1,reg2,reg3,base); \ + addi base,base,128; \ + __REST_4VRS(20,reg0,reg1,reg2,reg3,reg4); \ + addi reg4,reg4,128; \ + __REST_4VRS(24,reg0,reg1,reg2,reg3,base); \ + __REST_4VRS(28,reg0,reg1,reg2,reg3,reg4); \ + subi base,base,384 #ifdef __BIG_ENDIAN__ #define STXVD2X_ROT(n,b,base) > STXVD2X(n,b,base)diff --git a/arch/powerpc/kernel/tm.S b/arch/powerpc/kernel/tm.S index bf8f34a..81e1305 100644 --- a/arch/powerpc/kernel/tm.S +++ b/arch/powerpc/kernel/tm.S@@ -96,6 +96,8 @@ _GLOBAL(tm_abort) * they will abort back to the checkpointed state we save out here. * * Call with IRQs off, stacks get all out of sync for some periods in here! + * + * Is called from C */ _GLOBAL(tm_reclaim) mfcr r6@@ -151,7 +153,7 @@ _GLOBAL(tm_reclaim) beq dont_backup_vec addi r7, r3, THREAD_TRANSACT_VRSTATE - SAVE_32VRS(0, r6, r7) /* r6 scratch, r7 transact vr state */ + SAVE_32VRS(r6,r8,r9,r10,r11,r7) /* r6,r8,r9,r10,r11 scratch, r7 transact vr state */Line wrapping here.quoted
mfvscr v0 li r6, VRSTATE_VSCR stvx v0, r7, r6@@ -361,7 +363,7 @@ _GLOBAL(__tm_recheckpoint) li r5, VRSTATE_VSCR lvx v0, r8, r5 mtvscr v0 - REST_32VRS(0, r5, r8) /* r5 scratch, r8 ptr */ + REST_32VRS(r5,r6,r9,r10,r11,r8) /* r5,r6,r9,r10,r11 scratch, r8 ptr */wrapping here tooquoted
dont_restore_vec: ld r5, THREAD_VRSAVE(r3) mtspr SPRN_VRSAVE, r5diff --git a/arch/powerpc/kernel/vector.S b/arch/powerpc/kernel/vector.S index 1c2e7a3..8d587fb 100644 --- a/arch/powerpc/kernel/vector.S +++ b/arch/powerpc/kernel/vector.S@@ -13,6 +13,8 @@ * This is similar to load_up_altivec but for the transactional version of the * vector regs. It doesn't mess with the task MSR or valid flags. * Furthermore, VEC laziness is not supported with TM currently. + * + * Is called from C */ _GLOBAL(do_load_up_transact_altivec) mfmsr r6@@ -27,7 +29,7 @@ _GLOBAL(do_load_up_transact_altivec) lvx v0,r10,r3 mtvscr v0 addi r10,r3,THREAD_TRANSACT_VRSTATE - REST_32VRS(0,r4,r10) + REST_32VRS(r4,r5,r6,r7,r8,r10) blr #endif@@ -35,20 +37,24 @@ _GLOBAL(do_load_up_transact_altivec) /* * Load state from memory into VMX registers including VSCR. * Assumes the caller has enabled VMX in the MSR. + * + * Is called from C */ _GLOBAL(load_vr_state) li r4,VRSTATE_VSCR lvx v0,r4,r3 mtvscr v0 - REST_32VRS(0,r4,r3) + REST_32VRS(r4,r5,r6,r7,r8,r3) blr /* * Store VMX state into memory, including VSCR. * Assumes the caller has enabled VMX in the MSR. + * + * NOT called from C */ _GLOBAL(store_vr_state) - SAVE_32VRS(0, r4, r3) + SAVE_32VRS(r4,r5,r6,r7,r8,r3) mfvscr v0 li r4, VRSTATE_VSCR stvx v0, r4, r3@@ -63,6 +69,8 @@ _GLOBAL(store_vr_state) * * Note that on 32-bit this can only use registers that will be * restored by fast_exception_return, i.e. r3 - r6, r10 and r11. + * + * NOT called from C */ _GLOBAL(load_up_altivec) mfmsr r5 /* grab the current MSR */@@ -101,13 +109,15 @@ _GLOBAL(load_up_altivec) stw r4,THREAD_USED_VR(r5) lvx v0,r10,r6 mtvscr v0 - REST_32VRS(0,r4,r6) + REST_32VRS(r3,r4,r5,r10,r11,r6) /* restore registers and return */ blr /* * save_altivec(tsk) * Save the vector registers to its thread_struct + * + * Is called from C */ _GLOBAL(save_altivec) addi r3,r3,THREAD > /* want THREAD of task */@@ -116,7 +126,7 @@ _GLOBAL(save_altivec) PPC_LCMPI 0,r7,0 bne 2f addi r7,r3,THREAD_VRSTATE -2: SAVE_32VRS(0,r4,r7) +2: SAVE_32VRS(r4,r5,r6,r8,r9,r7) mfvscr v0 li r4,VRSTATE_VSCR stvx v0,r4,r7