Re: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support
Hi, Richard. After several tries with your testcases (I already added into V15 patch). I think "using a new IV" would be better than "multiplication" Now: loop_len_34 = MIN_EXPR ; _74 = MIN_EXPR ; --> multiplication approach will changed into _74 = loop_len_34 * 2; loop_len_48 = MIN_EXPR <_74, 4>; _77 = _74 - loop_len_48; loop_len_49 = MIN_EXPR <_77, 4>; _78 = _77 - loop_len_49; loop_len_50 = MIN_EXPR <_78, 4>; loop_len_51 = _78 - loop_len_50; I prefer "new IV" since it looks more reasonable and better codegen. Could you take a look at it: V15 patch: https://gcc.gnu.org/pipermail/gcc-patches/2023-May/619534.html Thanks. juzhe.zh...@rivai.ai From: Richard Sandiford Date: 2023-05-25 04:05 To: 钟居哲 CC: gcc-patches; rguenther Subject: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support I'll look at the samples tomorrow, but just to address one thing: 钟居哲 writes: >>> What gives the best code in these cases? Is emitting a multiplication >>> better? Or is using a new IV better? > Could you give me more detail information about "new refresh IV" approach. > I'd like to try that. By “using a new IV” I meant calling vect_set_loop_controls_directly for every rgroup, not just the first. So in the earlier example, there would be one decrementing IV for x and one decrementing IV for y. Thanks, Richard
Re: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support
Hi, Richard. After I fix codes, now IR is correct I think: loop_len_34 = MIN_EXPR ; _74 = loop_len_34 * 2; loop_len_48 = MIN_EXPR <_74, 4>; _75 = _74 - loop_len_48; loop_len_49 = MIN_EXPR <_75, 4>; _76 = _75 - loop_len_49; loop_len_50 = MIN_EXPR <_76, 4>; loop_len_51 = _76 - loop_len_50; ... vect__1.8_33 = .LEN_LOAD (_17, 16B, loop_len_34, 0); ... .LEN_STORE (_17, 16B, loop_len_34, vect__4.11_21, 0); ... vect__10.16_52 = .LEN_LOAD (_31, 32B, loop_len_48, 0); ... vect__10.17_54 = .LEN_LOAD (_29, 32B, loop_len_49, 0); ... vect__10.18_56 = .LEN_LOAD (_25, 32B, loop_len_50, 0); ... vect__10.19_58 = .LEN_LOAD (_80, 32B, loop_len_51, 0); For this case: uint64_t x2[100]; uint16_t y2[200]; void f2(int n) { for (int i = 0, j = 0; i < n; i += 2, j += 4) { x2[i + 0] += 1; x2[i + 1] += 2; y2[j + 0] += 1; y2[j + 1] += 2; y2[j + 2] += 3; y2[j + 3] += 4; } } The IR is like this: loop_len_56 = MIN_EXPR ; _66 = loop_len_56 * 4; loop_len_43 = _66 + 18446744073709551614; ... vect__1.44_44 = .LEN_LOAD (_6, 64B, 2, 0); ... vect__1.45_46 = .LEN_LOAD (_14, 64B, loop_len_43, 0); vect__2.46_47 = vect__1.44_44 + { 1, 2 }; vect__2.46_48 = vect__1.45_46 + { 1, 2 }; .LEN_STORE (_6, 64B, 2, vect__2.46_47, 0); .LEN_STORE (_14, 64B, loop_len_43, vect__2.46_48, 0); ... vect__6.51_57 = .LEN_LOAD (_10, 16B, loop_len_56, 0); vect__7.52_58 = vect__6.51_57 + { 1, 2, 3, 4, 1, 2, 3, 4 }; .LEN_STORE (_10, 16B, loop_len_56, vect__7.52_58, 0); It seems correct too ? >> What gives the best code in these cases? Is emitting a multiplication >> better? Or is using a new IV better? Could you give me more detail information about "new refresh IV" approach. I'd like to try that. Thanks. juzhe.zh...@rivai.ai From: Richard Sandiford Date: 2023-05-25 00:00 To: 钟居哲 CC: gcc-patches; rguenther Subject: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support 钟居哲 writes: > Oh. I see. Thank you so much for pointing this. > Could you tell me what I should do in the codes? > It seems that I should adjust it in > vect_adjust_loop_lens_control > > muliply by some factor ? Is this correct multiply by max_nscalars_per_iter > ? max_nscalars_per_iter * factor rather than just max_nscalars_per_iter Note that it's possible for later max_nscalars_per_iter * factor to be smaller, so a division might be needed in rare cases. E.g.: uint64_t x[100]; uint16_t y[200]; void f() { for (int i = 0, j = 0; i < 100; i += 2, j += 4) { x[i + 0] += 1; x[i + 1] += 2; y[j + 0] += 1; y[j + 1] += 2; y[j + 2] += 3; y[j + 3] += 4; } } where y has a single-control rgroup with max_nscalars_per_iter == 4 and x has a 2-control rgroup with max_nscalars_per_iter == 2 What gives the best code in these cases? Is emitting a multiplication better? Or is using a new IV better? Thanks, Richard
Re: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support
Hi, For the first piece of code ,I tried: unsigned int nitems_per_iter = dest_rgm->max_nscalars_per_iter * dest_rgm->factor; step = gimple_build (seq, MULT_EXPR, iv_type, step, build_int_cst (iv_type, nitems_per_iter)); Then optimized IR: loop_len_34 = MIN_EXPR ; _74 = loop_len_34 * 4; loop_len_51 = _74 + 18446744073709551604; _16 = (void *) ivtmp.27_41; _17 = &MEM [(short int *)_16]; vect__1.7_33 = .LEN_LOAD (_17, 16B, loop_len_34, 0); vect__2.8_23 = VIEW_CONVERT_EXPR(vect__1.7_33); vect__3.9_22 = vect__2.8_23 + { 1, 2, 1, 2, 1, 2, 1, 2 }; vect__4.10_21 = VIEW_CONVERT_EXPR(vect__3.9_22); .LEN_STORE (_17, 16B, loop_len_34, vect__4.10_21, 0); _20 = (void *) ivtmp.28_1; _31 = &MEM [(int *)_20]; vect__10.15_52 = .LEN_LOAD (_31, 32B, 4, 0); _30 = (void *) ivtmp.31_4; _29 = &MEM [(int *)_30]; vect__10.16_54 = .LEN_LOAD (_29, 32B, 4, 0); _26 = (void *) ivtmp.32_8; _25 = &MEM [(int *)_26]; vect__10.17_56 = .LEN_LOAD (_25, 32B, 4, 0); _79 = (void *) ivtmp.33_12; _80 = &MEM [(int *)_79]; vect__10.18_58 = .LEN_LOAD (_80, 32B, loop_len_51, 0); Is it correct ? It looks wierd ? juzhe.zh...@rivai.ai From: Richard Sandiford Date: 2023-05-25 00:00 To: 钟居哲 CC: gcc-patches; rguenther Subject: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support 钟居哲 writes: > Oh. I see. Thank you so much for pointing this. > Could you tell me what I should do in the codes? > It seems that I should adjust it in > vect_adjust_loop_lens_control > > muliply by some factor ? Is this correct multiply by max_nscalars_per_iter > ? max_nscalars_per_iter * factor rather than just max_nscalars_per_iter Note that it's possible for later max_nscalars_per_iter * factor to be smaller, so a division might be needed in rare cases. E.g.: uint64_t x[100]; uint16_t y[200]; void f() { for (int i = 0, j = 0; i < 100; i += 2, j += 4) { x[i + 0] += 1; x[i + 1] += 2; y[j + 0] += 1; y[j + 1] += 2; y[j + 2] += 3; y[j + 3] += 4; } } where y has a single-control rgroup with max_nscalars_per_iter == 4 and x has a 2-control rgroup with max_nscalars_per_iter == 2 What gives the best code in these cases? Is emitting a multiplication better? Or is using a new IV better? Thanks, Richard
Re: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support
Oh. I see. Thank you so much for pointing this. Could you tell me what I should do in the codes? It seems that I should adjust it in vect_adjust_loop_lens_control muliply by some factor ? Is this correct multiply by max_nscalars_per_iter ? Thanks. juzhe.zh...@rivai.ai From: Richard Sandiford Date: 2023-05-24 23:47 To: 钟居哲 CC: gcc-patches; rguenther Subject: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support 钟居哲 writes: > Hi, Richard. I still don't understand it. Sorry about that. > >>> loop_len_48 = MIN_EXPR ; > >> _74 = loop_len_34 * 2 - loop_len_48; > > I have the tests already tested. > We have a MIN_EXPR to calculate the total elements: > loop_len_34 = MIN_EXPR ; > I think "8" is already multiplied by 2? > > Why do we need loop_len_34 * 2 ? > Could you give me more informations, The similiar tests you present we > already have > execution check and passed. I am not sure whether this patch has the issue > that I didn't notice. Think about the maximum values of each SSA name: loop_len_34 = MIN_EXPR ; // MAX 8 loop_len_48 = MIN_EXPR ;// MAX 4 _74 = loop_len_34 - loop_len_48;// MAX 4 loop_len_49 = MIN_EXPR <_74, 4>;// MAX 4 (always == _74) _75 = _74 - loop_len_49;// 0 loop_len_50 = MIN_EXPR <_75, 4>;// 0 loop_len_51 = _75 - loop_len_50;// 0 So the final two y vectors will always have 0 controls. Thanks, Richard
Re: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support
Hi, Richard. I still don't understand it. Sorry about that. >> loop_len_48 = MIN_EXPR ; >> _74 = loop_len_34 * 2 - loop_len_48; I have the tests already tested. We have a MIN_EXPR to calculate the total elements: loop_len_34 = MIN_EXPR ; I think "8" is already multiplied by 2? Why do we need loop_len_34 * 2 ? Could you give me more informations, The similiar tests you present we already have execution check and passed. I am not sure whether this patch has the issue that I didn't notice. Thanks. juzhe.zh...@rivai.ai From: Richard Sandiford Date: 2023-05-24 23:31 To: 钟居哲 CC: gcc-patches; rguenther Subject: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support 钟居哲 writes: > Hi, the .optimized dump is like this: > >[local count: 21045336]: > ivtmp.26_36 = (unsigned long) &x; > ivtmp.27_3 = (unsigned long) &y; > ivtmp.30_6 = (unsigned long) &MEM [(void *)&y + 16B]; > ivtmp.31_10 = (unsigned long) &MEM [(void *)&y + 32B]; > ivtmp.32_14 = (unsigned long) &MEM [(void *)&y + 48B]; > >[local count: 273589366]: > # ivtmp_72 = PHI > # ivtmp.26_41 = PHI > # ivtmp.27_1 = PHI > # ivtmp.30_4 = PHI > # ivtmp.31_8 = PHI > # ivtmp.32_12 = PHI > loop_len_34 = MIN_EXPR ; > loop_len_48 = MIN_EXPR ; > _74 = loop_len_34 - loop_len_48; Yeah, I think this needs to be: loop_len_48 = MIN_EXPR ; _74 = loop_len_34 * 2 - loop_len_48; (as valid gimple). The point is that... > loop_len_49 = MIN_EXPR <_74, 4>; > _75 = _74 - loop_len_49; > loop_len_50 = MIN_EXPR <_75, 4>; > loop_len_51 = _75 - loop_len_50; ...there are 4 lengths capped to 4, for a total element count of 16. But loop_len_34 is never greater than 8. So for this case we either need to multiply, or we need to create a fresh IV for the second rgroup. Both approaches are fine. Thanks, Richard
Re: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support
Hi, the .optimized dump is like this: [local count: 21045336]: ivtmp.26_36 = (unsigned long) &x; ivtmp.27_3 = (unsigned long) &y; ivtmp.30_6 = (unsigned long) &MEM [(void *)&y + 16B]; ivtmp.31_10 = (unsigned long) &MEM [(void *)&y + 32B]; ivtmp.32_14 = (unsigned long) &MEM [(void *)&y + 48B]; [local count: 273589366]: # ivtmp_72 = PHI # ivtmp.26_41 = PHI # ivtmp.27_1 = PHI # ivtmp.30_4 = PHI # ivtmp.31_8 = PHI # ivtmp.32_12 = PHI loop_len_34 = MIN_EXPR ; loop_len_48 = MIN_EXPR ; _74 = loop_len_34 - loop_len_48; loop_len_49 = MIN_EXPR <_74, 4>; _75 = _74 - loop_len_49; loop_len_50 = MIN_EXPR <_75, 4>; loop_len_51 = _75 - loop_len_50; _16 = (void *) ivtmp.26_41; _17 = &MEM [(short int *)_16]; vect__1.6_33 = .LEN_LOAD (_17, 16B, loop_len_34, 0); vect__2.7_23 = VIEW_CONVERT_EXPR(vect__1.6_33); vect__3.8_22 = vect__2.7_23 + { 1, 2, 1, 2, 1, 2, 1, 2 }; vect__4.9_21 = VIEW_CONVERT_EXPR(vect__3.8_22); .LEN_STORE (_17, 16B, loop_len_34, vect__4.9_21, 0); _20 = (void *) ivtmp.27_1; _31 = &MEM [(int *)_20]; vect__10.14_52 = .LEN_LOAD (_31, 32B, loop_len_48, 0); _30 = (void *) ivtmp.30_4; _29 = &MEM [(int *)_30]; vect__10.15_54 = .LEN_LOAD (_29, 32B, loop_len_49, 0); _26 = (void *) ivtmp.31_8; _25 = &MEM [(int *)_26]; vect__10.16_56 = .LEN_LOAD (_25, 32B, loop_len_50, 0); _78 = (void *) ivtmp.32_12; _79 = &MEM [(int *)_78]; vect__10.17_58 = .LEN_LOAD (_79, 32B, loop_len_51, 0); vect__11.18_59 = vect__10.14_52 + { 1, 2, 3, 4 }; vect__11.18_60 = vect__10.15_54 + { 1, 2, 3, 4 }; vect__11.18_61 = vect__10.16_56 + { 1, 2, 3, 4 }; vect__11.18_62 = vect__10.17_58 + { 1, 2, 3, 4 }; .LEN_STORE (_31, 32B, loop_len_48, vect__11.18_59, 0); .LEN_STORE (_29, 32B, loop_len_49, vect__11.18_60, 0); .LEN_STORE (_25, 32B, loop_len_50, vect__11.18_61, 0); .LEN_STORE (_79, 32B, loop_len_51, vect__11.18_62, 0); ivtmp_73 = ivtmp_72 - loop_len_34; ivtmp.26_37 = ivtmp.26_41 + 16; ivtmp.27_2 = ivtmp.27_1 + 64; ivtmp.30_5 = ivtmp.30_4 + 64; ivtmp.31_9 = ivtmp.31_8 + 64; ivtmp.32_13 = ivtmp.32_12 + 64; if (ivtmp_73 != 0) goto ; [92.31%] else goto ; [7.69%] I am still check about it but I send it to you earlier. Thanks. juzhe.zh...@rivai.ai From: Richard Sandiford Date: 2023-05-24 23:07 To: juzhe.zhong CC: gcc-patches; rguenther Subject: Re: [PATCH V14] VECT: Add decrement IV iteration loop control by variable amount support Thanks for trying it. I'm still surprised that no multiplication is needed though. Does the patch work for: short x[100]; int y[200]; void f() { for (int i = 0, j = 0; i < 100; i += 2, j += 4) { x[i + 0] += 1; x[i + 1] += 2; y[j + 0] += 1; y[j + 1] += 2; y[j + 2] += 3; y[j + 3] += 4; } } ? Here, there should be a single-control rgroup for x, counting 2 units per scalar iteration. I'd expect the IV to use this scale. There should also be a 4-control rgroup for y, counting 4 units per scalar iteration. So I think the IV would need to be multiplied by 2 before being used for the y rgroup. Thanks, Richard juzhe.zh...@rivai.ai writes: > From: Ju-Zhe Zhong > > This patch is supporting decrement IV by following the flow designed by > Richard: > > (1) In vect_set_loop_condition_partial_vectors, for the first iteration of: > call vect_set_loop_controls_directly. > > (2) vect_set_loop_controls_directly calculates "step" as in your patch. > If rgc has 1 control, this step is the SSA name created for that control. > Otherwise the step is a fresh SSA name, as in your patch. > > (3) vect_set_loop_controls_directly stores this step somewhere for later > use, probably in LOOP_VINFO. Let's use "S" to refer to this stored step. > > (4) After the vect_set_loop_controls_directly call above, and outside > the "if" statement that now contains vect_set_loop_controls_directly, > check whether rgc->controls.length () > 1. If so, use > vect_adjust_loop_lens_control to set the controls based on S. > > Then the only caller of vect_adjust_loop_lens_control is > vect_set_loop_condition_partial_vectors. And the starting > step for vect_adjust_loop_lens_control is always S. > > This patch has well tested for single-rgroup and multiple-rgroup (SLP) and > passed all testcase in RISC-V port. > > Also, pass tests for multiple-rgroup (non-SLP) tested on vec_pack_trunk. > > --- > gcc/tree-vect-loop-manip.cc | 178 +--- > gcc/tree-vect-loop.cc | 13 +++ > gcc/tree-vectorizer.h | 12 +++ > 3 files changed, 192 insertions(+), 11 deletions(-) > > diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc > index ff6159e08d5..578ac5b783e 100644 > --- a/gcc/tree-vect-loop-manip.cc > +++ b/gcc/tree-vect-loop-manip.cc > @@ -468,6 +468,38 @@ vect_set_loop_controls_directly (class loop *loop, > loop_vec_info loop_vinfo, >gimple_stmt_iterator incr_gsi; >bool insert_after; >standard_iv_increment_position (loop,