From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-io1-xd34.google.com (mail-io1-xd34.google.com [IPv6:2607:f8b0:4864:20::d34]) by sourceware.org (Postfix) with ESMTPS id B9E5838576A1 for ; Fri, 4 Nov 2022 08:31:22 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org B9E5838576A1 Authentication-Results: sourceware.org; dmarc=pass (p=none dis=none) header.from=linaro.org Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=linaro.org Received: by mail-io1-xd34.google.com with SMTP id s10so1415850ioa.5 for ; Fri, 04 Nov 2022 01:31:22 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; h=to:subject:message-id:date:from:in-reply-to:references:mime-version :from:to:cc:subject:date:message-id:reply-to; bh=kseMlN+UvwKE3tuXrarxtIyRl9pXXtv/pn4ke2NP6ms=; b=nsw/I4LJGfqOhZ1dZZHtllAQPi0nPBygoATCbR9xt4OAUBEGJcrKEqGeXZs/lTxv1L QoPiGWD9/GpShKNrEHRXpOF1VpTfsHMlxEKEKOl+6Xdqe9KeAkPcuAU/6mH+/FLKDO0t 1cQ3EL51T4ShkMlwHe0gOKMcl62PuA74STGsKJvUK5tNYIV+BxzrthrqXV67N0sL6i37 AyCNa13Nn+PnOFPi6SmDmj8ZRO72pMTLY5uAh9Bhd7eojifyCcI/kPO90+PKhUNQvE+j RpcSa5KFETnI4maq/hMgqFkSugRGI/SnMqohaX3R0hwCHdTUjDbVwSM45QGyxK0wFUzy 1t6Q== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=to:subject:message-id:date:from:in-reply-to:references:mime-version :x-gm-message-state:from:to:cc:subject:date:message-id:reply-to; bh=kseMlN+UvwKE3tuXrarxtIyRl9pXXtv/pn4ke2NP6ms=; b=HXIZ+2kfhIbgKEyZBIEj1+Soyk4msRJS9PGGcpUMKnlqyoHZ+bhX0efSKvRmfphk9X ZjMPIDaIdD/4kcSl1Qr+vdBnP9H1Zh+N0+d7fGOOjOWDF4tmTpkhBurFNdA+4+/A05BJ YdNda6vtze8Kzunnw2zmYbfCNd93r2IfJ023q57QHKbSyOcPDtkm0VrcSF1merpC/FQU kcFDfbVEBbacExFqW8rzFHQyTcuStqYaAMXTX3YO/Xqj+dtv6X+JDwaEPAWdfFFYsVB3 Y5R07aRWCatDYuuoKA4YgPd4w1KwmPebfkmVtMPuDjn0wgkC5qkbgW9reaymdfrHjUMk FhZA== X-Gm-Message-State: ACrzQf2im4AmKzAc1oMSAGHExhRNURHwbAqq127a/01Ooafa+KG2bBzE MZHD2ktoxXDCJpPSZS3AE3mtGqxQkoBS/R+nqitBGQ== X-Google-Smtp-Source: AMsMyM5hCIca6jib1XgDSvnd3m1ntODpK/HyYonNPY+zHLQZTp/6gWgMzxZCHnRDWYMBahP4gx65u8eb3eLqdzAKM8k= X-Received: by 2002:a05:6638:1186:b0:36f:ae2e:a396 with SMTP id f6-20020a056638118600b0036fae2ea396mr20794676jas.89.1667550681733; Fri, 04 Nov 2022 01:31:21 -0700 (PDT) MIME-Version: 1.0 References: In-Reply-To: From: Prathamesh Kulkarni Date: Fri, 4 Nov 2022 14:00:45 +0530 Message-ID: Subject: Re: Extend fold_vec_perm to fold VEC_PERM_EXPR in VLA manner To: Prathamesh Kulkarni , gcc Patches , Richard Biener , richard.sandiford@arm.com Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-2.5 required=5.0 tests=BAYES_00,DKIM_SIGNED,DKIM_VALID,DKIM_VALID_AU,DKIM_VALID_EF,RCVD_IN_DNSWL_NONE,SCC_5_SHORT_WORD_LINES,SPF_HELO_NONE,SPF_PASS,TXREP autolearn=no autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org List-Id: On Mon, 31 Oct 2022 at 15:27, Richard Sandiford wrote: > > Prathamesh Kulkarni writes: > > On Wed, 26 Oct 2022 at 21:07, Richard Sandiford > > wrote: > >> > >> Sorry for the slow response. I wanted to find some time to think > >> about this a bit more. > >> > >> Prathamesh Kulkarni writes: > >> > On Fri, 30 Sept 2022 at 21:38, Richard Sandiford > >> > wrote: > >> >> > >> >> Richard Sandiford via Gcc-patches writes: > >> >> > Prathamesh Kulkarni writes: > >> >> >> Sorry to ask a silly question but in which case shall we select 2nd vector ? > >> >> >> For num_poly_int_coeffs == 2, > >> >> >> a1 /trunc n1 == (a1 + 0x) / (n1.coeffs[0] + n1.coeffs[1]*x) > >> >> >> If a1/trunc n1 succeeds, > >> >> >> 0 / n1.coeffs[1] == a1/n1.coeffs[0] == 0. > >> >> >> So, a1 has to be < n1.coeffs[0] ? > >> >> > > >> >> > Remember that a1 is itself a poly_int. It's not necessarily a constant. > >> >> > > >> >> > E.g. the TRN1 .D instruction maps to a VEC_PERM_EXPR with the selector: > >> >> > > >> >> > { 0, 2 + 2x, 1, 4 + 2x, 2, 6 + 2x, ... } > >> >> > >> >> Sorry, should have been: > >> >> > >> >> { 0, 2 + 2x, 2, 4 + 2x, 4, 6 + 2x, ... } > >> > Hi Richard, > >> > Thanks for the clarifications, and sorry for late reply. > >> > I have attached POC patch that tries to implement the above approach. > >> > Passes bootstrap+test on x86_64-linux-gnu and aarch64-linux-gnu for VLS vectors. > >> > > >> > For VLA vectors, I have only done limited testing so far. > >> > It seems to pass couple of tests written in the patch for > >> > nelts_per_pattern == 3, > >> > and folds the following svld1rq test: > >> > int32x4_t v = {1, 2, 3, 4}; > >> > return svld1rq_s32 (svptrue_b8 (), &v[0]) > >> > into: > >> > return {1, 2, 3, 4, ...}; > >> > I will try to bootstrap+test it on SVE machine to test further for VLA folding. > >> > > >> > I have a couple of questions: > >> > 1] When mask selects elements from same vector but from different patterns: > >> > For eg: > >> > arg0 = {1, 11, 2, 12, 3, 13, ...}, > >> > arg1 = {21, 31, 22, 32, 23, 33, ...}, > >> > mask = {0, 0, 0, 1, 0, 2, ... }, > >> > All have npatterns = 2, nelts_per_pattern = 3. > >> > > >> > With above mask, > >> > Pattern {0, ...} selects arg0[0], ie {1, ...} > >> > Pattern {0, 1, 2, ...} selects arg0[0], arg0[1], arg0[2], ie {1, 11, 2, ...} > >> > While arg0[0] and arg0[2] belong to same pattern, arg0[1] belongs to different > >> > pattern in arg0. > >> > The result is: > >> > res = {1, 1, 1, 11, 1, 2, ...} > >> > In this case, res's 2nd pattern {1, 11, 2, ...} is encoded with: > >> > with a0 = 1, a1 = 11, S = -9. > >> > Is that expected tho ? It seems to create a new encoding which > >> > wasn't present in the input vector. For instance, the next elem in > >> > sequence would be -7, > >> > which is not present originally in arg0. > >> > >> Yeah, you're right, sorry. Going back to: > >> > >> (2) The explicit encoding can be used to produce a sequence of N*Ex*Px > >> elements for any integer N. This extended sequence can be reencoded > >> as having N*Px patterns, with Ex staying the same. > >> > >> I guess we need to pick an N for the selector such that each new > >> selector pattern (each one out of the N*Px patterns) selects from > >> the *same pattern* of the same data input. > >> > >> So if a particular pattern in the selector has a step S, and the data > >> input it selects from has Pi patterns, N*S must be a multiple of Pi. > >> N must be a multiple of least_common_multiple(S,Pi)/S. > >> > >> I think that means that the total number of patterns in the result > >> (Pr from previous messages) can safely be: > >> > >> Ps * least_common_multiple( > >> least_common_multiple(S[1], P[input(1)]) / S[1], > >> ... > >> least_common_multiple(S[Ps], P[input(Ps)]) / S[Ps] > >> ) > >> > >> where: > >> > >> Ps = the number of patterns in the selector > >> S[I] = the step for selector pattern I (I being 1-based) > >> input(I) = the data input selected by selector pattern I (I being 1-based) > >> P[I] = the number of patterns in data input I > >> > >> That's getting quite complicated :-) If we allow arbitrary P[...] > >> and S[...] then it could also get large. Perhaps we should finally > >> give up on the general case and limit this to power-of-2 patterns and > >> power-of-2 steps, so that least_common_multiple becomes MAX. Maybe that > >> simplifies other things as well. > >> > >> What do you think? > > Hi Richard, > > Thanks for the suggestions. Yeah I suppose we can initially add support for > > power-of-2 patterns and power-of-2 steps and try to generalize it in > > follow up patches if possible. > > > > Sorry if this sounds like a silly ques -- if we are going to have > > pattern in selector, select *same pattern from same input vector*, > > instead of re-encoding the selector to have N * Ps patterns, would it > > make sense for elements in selector to denote pattern number itself > > instead of element index > > if input vectors are VLA ? > > > > For eg: > > op0 = {1, 2, 3, 4, 1, 2, 3, 5, 1, 2, 3, 6, ...} > > op1 = {...} > > with npatterns == 4, nelts_per_pattern == 3, > > sel = {0, 3} should pick pattern 0 and pattern 3 from op0, > > so, res = {1, 4, 1, 5, 1, 6, ...} > > Not sure if this is correct tho. > > This wouldn't allow us to represent things like a "duplicate one > element", or "copy the leading N elements from the first input and > the other elements from elements N+ of the second input", which we > can with the current scheme. > > The restriction about each (unwound) selector pattern selecting from the > same input pattern only applies to case where the selector pattern is > stepped (and only applies to the stepped part of the pattern, not the > leading element). The restriction is also local to this code; it > doesn't make other VEC_PERM_EXPRs invalid. Hi Richard, Thanks for the clarifications. Just to clarify your approach with an eg: Let selected input vector be: arg0: {a0, b0, c0, d0, a0 + S, b0 + S, c0 + S, d0 + S, a0 + 2S, b0 + 2S, c0 + 2S, dd + 2S, ...} where arg0 has npatterns = 4, and nelts_per_pattern = 3. Let sel = {0, 0, 1, 2, 2, 4, ...} where sel_npatterns = 2 and sel_nelts_per_pattern = 3 So, the first pattern in sel: p1: {0, 1, 2, ...} which will select {a0, b0, c0, ...} which would be incorrect, since they belong to different patterns in arg0. So to select elements from same pattern in arg0, we need to divide p1 into at least N1 = P_arg0 / S0 = 4 distinct patterns. Similarly for second pattern in sel: p2: {0, 2, 4, ...}, we need to divide it into at least N2 = P_arg0 / S1 = 2 distinct patterns. Select N = max(N1, N2) = 4 So, the selector will be extended to N * Ps * Es = 4 * 2 * 3 == 24 elements, and will be re-encoded with N*Ps = 8 patterns: re-encoded sel: {a0, b0, c0, d0, a0 + S, b0 + S, c0 + S, d0 + S, a0 + 2S, b0 + 2S, c0 + 2S, d0 + 2S, a0 + 3S, b0 + 3S, c0 + 3S, d0 + 3S, a0 + 4S, b0 + 4S, c0 + 4s, d0 + 4S, a0 + 5S, b0 + 5S, c0 + 5S, d0 + 5S, ...} with 8 patterns, p1: {a0, a0 + 2S, a0 + 4S, ...} p2: {b0, b0 + 2S, b0 + 4S, ...} ... which select elements from same pattern from same input vector. Does this look correct ? For feasibility, we can check initially that sel_npatterns, arg0_npatterns, arg1_npatterns are powers of 2 and for each stepped pattern, it's stepped size S is a power of 2. I suppose this will be sufficient to ensure that sel can be re-encoded with N*Ps npatterns such that each new pattern selects elements from same pattern of the input vector ? Then compute N: N = 1; for (every pattern p in sel) { op = corresponding input vector for pattern; S = step_size (p); N_pattern = max (S, npatterns (op)) / S; N = max(N, N_pattern) } and re-encode selector with N*Ps patterns. I guess rest of the patch will mostly stay the same. Thanks, Prathamesh > > Thanks, > Richard > > > > > Thanks, > > Prathamesh > >> > >> > I suppose it's fine since if the user defines mask to have pattern {0, > >> > 1, 2, ...} > >> > they intended result to have pattern with above encoding. > >> > Just wanted to confirm if this is correct ? > >> > > >> > 2] Could you please suggest a test-case for S < 0 ? > >> > I am not able to come up with one :/ > >> > >> svrev is one way of creating negative steps. > >> > >> Thanks, > >> Richard > >> > >> > > >> > Thanks, > >> > Prathamesh > >> >> > >> >> > which is an interleaving of the two patterns: > >> >> > > >> >> > { 0, 2, 4, ... } a0 = 0, a1 = 2, S = 2 > >> >> > { 2 + 2x, 4 + 2x, 6 + 2x } a0 = 2 + 2x, a1 = 4 + 2x, S = 2