1 commit e0107a6ca637eb3997131e966e19fcd6001b37ad
2 Author: Jan Wassenberg <janwas@google.com>
3 Date: Fri Nov 12 09:12:35 2021 +0100
5 Avoid deprecated Highway functions
7 Instead use the newer overloads with the extra d arg required to
10 diff --git a/lib/jxl/dec_reconstruct.cc b/lib/jxl/dec_reconstruct.cc
11 index 9d1eb61..69a4361 100644
12 --- a/lib/jxl/dec_reconstruct.cc
13 +++ b/lib/jxl/dec_reconstruct.cc
14 @@ -357,8 +357,8 @@ void DoYCbCrUpsampling(size_t hs, size_t vs, ImageF* plane_in, const Rect& rect,
15 Store(left, d, out + x);
16 Store(right, d, out + x + 1);
18 - Store(InterleaveLower(left, right), d, out + x);
19 - Store(InterleaveUpper(left, right), d, out + x + Lanes(d));
20 + Store(InterleaveLower(d, left, right), d, out + x);
21 + Store(InterleaveUpper(d, left, right), d, out + x + Lanes(d));
25 diff --git a/lib/jxl/dec_upsample.cc b/lib/jxl/dec_upsample.cc
26 index 7277e4f..3cb3f36 100644
27 --- a/lib/jxl/dec_upsample.cc
28 +++ b/lib/jxl/dec_upsample.cc
29 @@ -176,8 +176,8 @@ void Upsample(const ImageF& src, const Rect& src_rect, ImageF* dst,
30 min = Min(LoadU(df, raw_min_row + sx + fx), min);
31 max = Max(LoadU(df, raw_max_row + sx + fx), max);
33 - min = MinOfLanes(min);
34 - max = MaxOfLanes(max);
35 + min = MinOfLanes(df, min);
36 + max = MaxOfLanes(df, max);
37 for (size_t lx = 0; lx < N; lx += V) {
38 StoreU(min, df, min_row + N * sx + lx);
39 StoreU(max, df, max_row + N * sx + lx);
40 diff --git a/lib/jxl/enc_ac_strategy.cc b/lib/jxl/enc_ac_strategy.cc
41 index bc50465..c0ed68f 100644
42 --- a/lib/jxl/enc_ac_strategy.cc
43 +++ b/lib/jxl/enc_ac_strategy.cc
44 @@ -429,8 +429,8 @@ float EstimateEntropy(const AcStrategy& acs, size_t x, size_t y,
46 entropy_v += nzeros_v * cost1;
48 - entropy += GetLane(SumOfLanes(entropy_v));
49 - size_t num_nzeros = GetLane(SumOfLanes(nzeros_v));
50 + entropy += GetLane(SumOfLanes(df, entropy_v));
51 + size_t num_nzeros = GetLane(SumOfLanes(df, nzeros_v));
52 // Add #bit of num_nonzeros, as an estimate of the cost for encoding the
53 // number of non-zeros of the block.
54 size_t nbits = CeilLog2Nonzero(num_nzeros + 1) + 1;
55 @@ -441,9 +441,9 @@ float EstimateEntropy(const AcStrategy& acs, size_t x, size_t y,
59 - ((config.info_loss_multiplier * GetLane(SumOfLanes(info_loss))) +
60 + ((config.info_loss_multiplier * GetLane(SumOfLanes(df, info_loss))) +
61 (config.info_loss_multiplier2 *
62 - sqrt(num_blocks * GetLane(SumOfLanes(info_loss2)))));
63 + sqrt(num_blocks * GetLane(SumOfLanes(df, info_loss2)))));
67 diff --git a/lib/jxl/enc_adaptive_quantization.cc b/lib/jxl/enc_adaptive_quantization.cc
68 index 7a6c772..51cea65 100644
69 --- a/lib/jxl/enc_adaptive_quantization.cc
70 +++ b/lib/jxl/enc_adaptive_quantization.cc
71 @@ -189,7 +189,7 @@ V GammaModulation(const D d, const size_t x, const size_t y,
72 overall_ratio += avg_ratio;
75 - overall_ratio = SumOfLanes(overall_ratio);
76 + overall_ratio = SumOfLanes(d, overall_ratio);
77 overall_ratio *= Set(d, 1.0f / 64);
78 // ideally -1.0, but likely optimal correction adds some entropy, so slightly
80 @@ -246,12 +246,12 @@ V ColorModulation(const D d, const size_t x, const size_t y,
81 // blue we consider as if it was fully red or blue.
82 static const float ratio = 30.610615782142737f; // out of 64 pixels.
84 - auto overall_red_coverage = SumOfLanes(red_coverage);
85 + auto overall_red_coverage = SumOfLanes(d, red_coverage);
86 overall_red_coverage =
87 Min(overall_red_coverage, Set(d, ratio * kRedRampLength));
88 overall_red_coverage *= Set(d, red_strength / ratio);
90 - auto overall_blue_coverage = SumOfLanes(blue_coverage);
91 + auto overall_blue_coverage = SumOfLanes(d, blue_coverage);
92 overall_blue_coverage =
93 Min(overall_blue_coverage, Set(d, ratio * kBlueRampLength));
94 overall_blue_coverage *= Set(d, blue_strength / ratio);
95 @@ -295,7 +295,7 @@ V HfModulation(const D d, const size_t x, const size_t y, const ImageF& xyb,
99 - sum = SumOfLanes(sum);
100 + sum = SumOfLanes(d, sum);
101 return MulAdd(sum, Set(d, -2.0052193233688884f / 112), out_val);
104 diff --git a/lib/jxl/enc_ar_control_field.cc b/lib/jxl/enc_ar_control_field.cc
105 index f43340e..f8025ac 100644
106 --- a/lib/jxl/enc_ar_control_field.cc
107 +++ b/lib/jxl/enc_ar_control_field.cc
108 @@ -157,7 +157,7 @@ void ProcessTile(const Image3F& opsin, PassesEncoderState* enc_state,
109 sum += LoadU(df4, rows_in[iy] + x * 4 + ix + 2);
112 - row_out[x] = GetLane(Sqrt(SumOfLanes(sum))) * (1.0f / 4.0f);
113 + row_out[x] = GetLane(Sqrt(SumOfLanes(df4, sum))) * (1.0f / 4.0f);
116 // Indexing iy and ix is a bit tricky as we include a 2 pixel border
117 @@ -193,7 +193,7 @@ void ProcessTile(const Image3F& opsin, PassesEncoderState* enc_state,
118 sum += Load(df4, rows_in[iy] + sx + ix);
121 - row_out[x] = GetLane(Sqrt(SumOfLanes(sum))) * (1.0f / 4.0f);
122 + row_out[x] = GetLane(Sqrt(SumOfLanes(df4, sum))) * (1.0f / 4.0f);
125 for (size_t iy = sy; iy < ey; iy++) {
126 diff --git a/lib/jxl/enc_butteraugli_pnorm.cc b/lib/jxl/enc_butteraugli_pnorm.cc
127 index 32623b8..b7feac0 100644
128 --- a/lib/jxl/enc_butteraugli_pnorm.cc
129 +++ b/lib/jxl/enc_butteraugli_pnorm.cc
130 @@ -87,13 +87,13 @@ double ComputeDistanceP(const ImageF& distmap, const ButteraugliParams& params,
134 - onePerPixels * (sum1[0] + GetLane(SumOfLanes(Load(d, sum_totals0)))),
135 + onePerPixels * (sum1[0] + GetLane(SumOfLanes(d, Load(d, sum_totals0)))),
138 - onePerPixels * (sum1[1] + GetLane(SumOfLanes(Load(d, sum_totals1)))),
139 + onePerPixels * (sum1[1] + GetLane(SumOfLanes(d, Load(d, sum_totals1)))),
142 - onePerPixels * (sum1[2] + GetLane(SumOfLanes(Load(d, sum_totals2)))),
143 + onePerPixels * (sum1[2] + GetLane(SumOfLanes(d, Load(d, sum_totals2)))),
147 diff --git a/lib/jxl/enc_chroma_from_luma.cc b/lib/jxl/enc_chroma_from_luma.cc
148 index e5c3f38..370595c 100644
149 --- a/lib/jxl/enc_chroma_from_luma.cc
150 +++ b/lib/jxl/enc_chroma_from_luma.cc
151 @@ -91,9 +91,9 @@ struct CFLFunction {
152 fdme_v += IfThenElse(av >= thres, zero, dme);
155 - *fpeps = first_derivative_peps + GetLane(SumOfLanes(fdpe_v));
156 - *fmeps = first_derivative_meps + GetLane(SumOfLanes(fdme_v));
157 - return first_derivative + GetLane(SumOfLanes(fd_v));
158 + *fpeps = first_derivative_peps + GetLane(SumOfLanes(df, fdpe_v));
159 + *fmeps = first_derivative_meps + GetLane(SumOfLanes(df, fdme_v));
160 + return first_derivative + GetLane(SumOfLanes(df, fd_v));
163 const float* JXL_RESTRICT values_m;
164 @@ -124,8 +124,8 @@ int32_t FindBestMultiplier(const float* values_m, const float* values_s,
165 cb = MulAdd(a, b, cb);
167 // + distance_mul * x^2 * num
168 - x = -GetLane(SumOfLanes(cb)) /
169 - (GetLane(SumOfLanes(ca)) + num * distance_mul * 0.5f);
170 + x = -GetLane(SumOfLanes(df, cb)) /
171 + (GetLane(SumOfLanes(df, ca)) + num * distance_mul * 0.5f);
173 constexpr float eps = 1;
174 constexpr float kClamp = 20.0f;
175 diff --git a/lib/jxl/enc_cluster.cc b/lib/jxl/enc_cluster.cc
176 index 1f12a29..8ae863c 100644
177 --- a/lib/jxl/enc_cluster.cc
178 +++ b/lib/jxl/enc_cluster.cc
179 @@ -49,7 +49,7 @@ void HistogramEntropy(const Histogram& a) {
180 const auto counts = LoadU(di, &a.data_[i]);
181 entropy_lanes += Entropy(ConvertTo(df, counts), inv_tot, total);
183 - a.entropy_ += GetLane(SumOfLanes(entropy_lanes));
184 + a.entropy_ += GetLane(SumOfLanes(df, entropy_lanes));
187 float HistogramDistance(const Histogram& a, const Histogram& b) {
188 @@ -71,7 +71,7 @@ float HistogramDistance(const Histogram& a, const Histogram& b) {
189 const auto counts = ConvertTo(df, a_counts + b_counts);
190 distance_lanes += Entropy(counts, inv_tot, total);
192 - const float total_distance = GetLane(SumOfLanes(distance_lanes));
193 + const float total_distance = GetLane(SumOfLanes(df, distance_lanes));
194 return total_distance - a.entropy_ - b.entropy_;
197 diff --git a/lib/jxl/enc_entropy_coder.cc b/lib/jxl/enc_entropy_coder.cc
198 index 0946300..07fe5a0 100644
199 --- a/lib/jxl/enc_entropy_coder.cc
200 +++ b/lib/jxl/enc_entropy_coder.cc
201 @@ -86,7 +86,7 @@ int32_t NumNonZeroExceptLLF(const size_t cx, const size_t cy,
203 // We want area - sum_zero, add because neg_sum_zero is already negated.
204 const int32_t nzeros =
205 - int32_t(cx * cy * kDCTBlockSize) + GetLane(SumOfLanes(neg_sum_zero));
206 + int32_t(cx * cy * kDCTBlockSize) + GetLane(SumOfLanes(di, neg_sum_zero));
208 const int32_t shifted_nzeros = static_cast<int32_t>(
209 (nzeros + covered_blocks - 1) >> log2_covered_blocks);
210 @@ -135,7 +135,7 @@ int32_t NumNonZero8x8ExceptDC(const int32_t* JXL_RESTRICT block,
212 // We want 64 - sum_zero, add because neg_sum_zero is already negated.
213 const int32_t nzeros =
214 - int32_t(kDCTBlockSize) + GetLane(SumOfLanes(neg_sum_zero));
215 + int32_t(kDCTBlockSize) + GetLane(SumOfLanes(di, neg_sum_zero));
217 *nzeros_pos = nzeros;
219 diff --git a/lib/jxl/enc_fast_heuristics.cc b/lib/jxl/enc_fast_heuristics.cc
220 index 16f7670..0551782 100644
221 --- a/lib/jxl/enc_fast_heuristics.cc
222 +++ b/lib/jxl/enc_fast_heuristics.cc
223 @@ -94,8 +94,8 @@ Status Heuristics(PassesEncoderState* enc_state,
224 cb = MulAdd(a, b, cb);
228 - -GetLane(SumOfLanes(cb)) / (GetLane(SumOfLanes(ca)) + 1e-9f);
229 + float best = -GetLane(SumOfLanes(df, cb)) /
230 + (GetLane(SumOfLanes(df, ca)) + 1e-9f);
231 int8_t& res = (c == 0 ? shared.cmap.ytox_map : shared.cmap.ytob_map)
233 res = std::max(-128.0f, std::min(127.0f, roundf(best)));
234 @@ -124,8 +124,8 @@ Status Heuristics(PassesEncoderState* enc_state,
235 max = IfThenElse(max > nn, max, nn);
238 - row_out_avg[x] = GetLane(SumOfLanes(sum));
239 - row_out[x] = GetLane(MaxOfLanes(max));
240 + row_out_avg[x] = GetLane(SumOfLanes(df4, sum));
241 + row_out[x] = GetLane(MaxOfLanes(df4, max));
245 diff --git a/lib/jxl/gauss_blur.cc b/lib/jxl/gauss_blur.cc
246 index f9babe7..f24a74c 100644
247 --- a/lib/jxl/gauss_blur.cc
248 +++ b/lib/jxl/gauss_blur.cc
249 @@ -421,7 +421,7 @@ ImageF ConvolveXSampleAndTranspose(const ImageF& in,
250 for (int i = -r; i <= r; i += Lanes(df)) {
251 sum = MulAdd(LoadU(df, rowp + x + i), LoadU(df, kernelp + i), sum);
253 - out.Row(ox)[y] = GetLane(SumOfLanes(sum));
254 + out.Row(ox)[y] = GetLane(SumOfLanes(df, sum));
256 for (; x < in.xsize(); x += res, ++ox) {
258 diff --git a/lib/jxl/modular/encoding/enc_ma.cc b/lib/jxl/modular/encoding/enc_ma.cc
259 index f847db6..7700ecc 100644
260 --- a/lib/jxl/modular/encoding/enc_ma.cc
261 +++ b/lib/jxl/modular/encoding/enc_ma.cc
262 @@ -60,7 +60,7 @@ float EstimateBits(const int32_t *counts, int32_t *rounded_counts,
264 IfThenElse(counts_v == zero, zero, counts_v * BitCast(df, nbps));
266 - return GetLane(SumOfLanes(bits_lanes));
267 + return GetLane(SumOfLanes(df, bits_lanes));
270 void MakeSplitNode(size_t pos, int property, int splitval, Predictor lpred,
271 diff --git a/lib/jxl/rational_polynomial_test.cc b/lib/jxl/rational_polynomial_test.cc
272 index d0e628d..e0d5a6e 100644
273 --- a/lib/jxl/rational_polynomial_test.cc
274 +++ b/lib/jxl/rational_polynomial_test.cc
275 @@ -50,7 +50,7 @@ struct EvalLog2 {
276 const HWY_FULL(int32_t) di;
277 const auto x_bits = BitCast(di, vx);
278 // Cannot handle negative numbers / NaN.
279 - JXL_DASSERT(AllTrue(Abs(x_bits) == x_bits));
280 + JXL_DASSERT(AllTrue(di, Abs(x_bits) == x_bits));
282 // Range reduction to [-1/3, 1/3] - 3 integer, 2 float ops
283 const auto exp_bits = x_bits - Set(di, 0x3f2aaaab); // = 2/3
284 diff --git a/lib/jxl/splines.cc b/lib/jxl/splines.cc
285 index 58ebfd6..0026fa9 100644
286 --- a/lib/jxl/splines.cc
287 +++ b/lib/jxl/splines.cc
288 @@ -53,7 +53,7 @@ float ContinuousIDCT(const float dct[32], float t) {
289 auto local_res = LoadU(df, dct + i) * cos;
290 result = MulAdd(Set(df, square_root<2>::value), local_res, result);
292 - return GetLane(SumOfLanes(result));
293 + return GetLane(SumOfLanes(df, result));
296 template <typename DF>
297 diff --git a/lib/jxl/transpose-inl.h b/lib/jxl/transpose-inl.h
298 index d12b129..e89e8af 100644
299 --- a/lib/jxl/transpose-inl.h
300 +++ b/lib/jxl/transpose-inl.h
301 @@ -137,25 +137,26 @@ JXL_INLINE_TRANSPOSE void GenericTransposeBlock(TransposeSimdTag<true>,
302 static_assert(COLS_or_0 % 4 == 0, "Invalid number of columns");
303 for (size_t n = 0; n < ROWS; n += 4) {
304 for (size_t m = 0; m < COLS; m += 4) {
305 - const auto p0 = from.LoadPart(BlockDesc<4>(), n + 0, m + 0);
306 - const auto p1 = from.LoadPart(BlockDesc<4>(), n + 1, m + 0);
307 - const auto p2 = from.LoadPart(BlockDesc<4>(), n + 2, m + 0);
308 - const auto p3 = from.LoadPart(BlockDesc<4>(), n + 3, m + 0);
310 - const auto q0 = InterleaveLower(p0, p2);
311 - const auto q1 = InterleaveLower(p1, p3);
312 - const auto q2 = InterleaveUpper(p0, p2);
313 - const auto q3 = InterleaveUpper(p1, p3);
315 - const auto r0 = InterleaveLower(q0, q1);
316 - const auto r1 = InterleaveUpper(q0, q1);
317 - const auto r2 = InterleaveLower(q2, q3);
318 - const auto r3 = InterleaveUpper(q2, q3);
320 - to.StorePart(BlockDesc<4>(), r0, m + 0, n + 0);
321 - to.StorePart(BlockDesc<4>(), r1, m + 1, n + 0);
322 - to.StorePart(BlockDesc<4>(), r2, m + 2, n + 0);
323 - to.StorePart(BlockDesc<4>(), r3, m + 3, n + 0);
324 + const BlockDesc<4> d;
325 + const auto p0 = from.LoadPart(d, n + 0, m + 0);
326 + const auto p1 = from.LoadPart(d, n + 1, m + 0);
327 + const auto p2 = from.LoadPart(d, n + 2, m + 0);
328 + const auto p3 = from.LoadPart(d, n + 3, m + 0);
330 + const auto q0 = InterleaveLower(d, p0, p2);
331 + const auto q1 = InterleaveLower(d, p1, p3);
332 + const auto q2 = InterleaveUpper(d, p0, p2);
333 + const auto q3 = InterleaveUpper(d, p1, p3);
335 + const auto r0 = InterleaveLower(d, q0, q1);
336 + const auto r1 = InterleaveUpper(d, q0, q1);
337 + const auto r2 = InterleaveLower(d, q2, q3);
338 + const auto r3 = InterleaveUpper(d, q2, q3);
340 + to.StorePart(d, r0, m + 0, n + 0);
341 + to.StorePart(d, r1, m + 1, n + 0);
342 + to.StorePart(d, r2, m + 2, n + 0);
343 + to.StorePart(d, r3, m + 3, n + 0);
347 From 0902f85ca6e9e305338baf7974192acab8c53ac5 Mon Sep 17 00:00:00 2001
348 From: Jan Wassenberg <janwas@google.com>
349 Date: Thu, 6 Jan 2022 09:36:20 +0100
350 Subject: [PATCH] Avoid deprecated Hwy functions - add d arg. Also remove
354 .../test_render_pipeline_stages.h | 1 -
355 lib/jxl/transpose-inl.h | 85 ++++++++++---------
356 2 files changed, 43 insertions(+), 43 deletions(-)
358 #diff --git a/lib/jxl/render_pipeline/test_render_pipeline_stages.h b/lib/jxl/render_pipeline/test_render_pipeline_stages.h
359 #index 6b33b677a..fc6b38a83 100644
360 #--- a/lib/jxl/render_pipeline/test_render_pipeline_stages.h
361 #+++ b/lib/jxl/render_pipeline/test_render_pipeline_stages.h
366 #-#include "gtest/gtest.h"
367 # #include "lib/jxl/render_pipeline/render_pipeline_stage.h"
370 diff --git a/lib/jxl/transpose-inl.h b/lib/jxl/transpose-inl.h
371 index e89e8af0a..467442073 100644
372 --- a/lib/jxl/transpose-inl.h
373 +++ b/lib/jxl/transpose-inl.h
374 @@ -74,50 +74,51 @@ JXL_INLINE_TRANSPOSE void GenericTransposeBlock(TransposeSimdTag<true>,
375 static_assert(COLS_or_0 % 8 == 0, "Invalid number of columns");
376 for (size_t n = 0; n < ROWS; n += 8) {
377 for (size_t m = 0; m < COLS; m += 8) {
378 - auto i0 = from.LoadPart(BlockDesc<8>(), n + 0, m + 0);
379 - auto i1 = from.LoadPart(BlockDesc<8>(), n + 1, m + 0);
380 - auto i2 = from.LoadPart(BlockDesc<8>(), n + 2, m + 0);
381 - auto i3 = from.LoadPart(BlockDesc<8>(), n + 3, m + 0);
382 - auto i4 = from.LoadPart(BlockDesc<8>(), n + 4, m + 0);
383 - auto i5 = from.LoadPart(BlockDesc<8>(), n + 5, m + 0);
384 - auto i6 = from.LoadPart(BlockDesc<8>(), n + 6, m + 0);
385 - auto i7 = from.LoadPart(BlockDesc<8>(), n + 7, m + 0);
386 + const BlockDesc<8> d;
387 + auto i0 = from.LoadPart(d, n + 0, m + 0);
388 + auto i1 = from.LoadPart(d, n + 1, m + 0);
389 + auto i2 = from.LoadPart(d, n + 2, m + 0);
390 + auto i3 = from.LoadPart(d, n + 3, m + 0);
391 + auto i4 = from.LoadPart(d, n + 4, m + 0);
392 + auto i5 = from.LoadPart(d, n + 5, m + 0);
393 + auto i6 = from.LoadPart(d, n + 6, m + 0);
394 + auto i7 = from.LoadPart(d, n + 7, m + 0);
395 // Surprisingly, this straightforward implementation (24 cycles on port5)
396 // is faster than load128+insert and LoadDup128+ConcatUpperLower+blend.
397 - const auto q0 = InterleaveLower(i0, i2);
398 - const auto q1 = InterleaveLower(i1, i3);
399 - const auto q2 = InterleaveUpper(i0, i2);
400 - const auto q3 = InterleaveUpper(i1, i3);
401 - const auto q4 = InterleaveLower(i4, i6);
402 - const auto q5 = InterleaveLower(i5, i7);
403 - const auto q6 = InterleaveUpper(i4, i6);
404 - const auto q7 = InterleaveUpper(i5, i7);
406 - const auto r0 = InterleaveLower(q0, q1);
407 - const auto r1 = InterleaveUpper(q0, q1);
408 - const auto r2 = InterleaveLower(q2, q3);
409 - const auto r3 = InterleaveUpper(q2, q3);
410 - const auto r4 = InterleaveLower(q4, q5);
411 - const auto r5 = InterleaveUpper(q4, q5);
412 - const auto r6 = InterleaveLower(q6, q7);
413 - const auto r7 = InterleaveUpper(q6, q7);
415 - i0 = ConcatLowerLower(r4, r0);
416 - i1 = ConcatLowerLower(r5, r1);
417 - i2 = ConcatLowerLower(r6, r2);
418 - i3 = ConcatLowerLower(r7, r3);
419 - i4 = ConcatUpperUpper(r4, r0);
420 - i5 = ConcatUpperUpper(r5, r1);
421 - i6 = ConcatUpperUpper(r6, r2);
422 - i7 = ConcatUpperUpper(r7, r3);
423 - to.StorePart(BlockDesc<8>(), i0, m + 0, n + 0);
424 - to.StorePart(BlockDesc<8>(), i1, m + 1, n + 0);
425 - to.StorePart(BlockDesc<8>(), i2, m + 2, n + 0);
426 - to.StorePart(BlockDesc<8>(), i3, m + 3, n + 0);
427 - to.StorePart(BlockDesc<8>(), i4, m + 4, n + 0);
428 - to.StorePart(BlockDesc<8>(), i5, m + 5, n + 0);
429 - to.StorePart(BlockDesc<8>(), i6, m + 6, n + 0);
430 - to.StorePart(BlockDesc<8>(), i7, m + 7, n + 0);
431 + const auto q0 = InterleaveLower(d, i0, i2);
432 + const auto q1 = InterleaveLower(d, i1, i3);
433 + const auto q2 = InterleaveUpper(d, i0, i2);
434 + const auto q3 = InterleaveUpper(d, i1, i3);
435 + const auto q4 = InterleaveLower(d, i4, i6);
436 + const auto q5 = InterleaveLower(d, i5, i7);
437 + const auto q6 = InterleaveUpper(d, i4, i6);
438 + const auto q7 = InterleaveUpper(d, i5, i7);
440 + const auto r0 = InterleaveLower(d, q0, q1);
441 + const auto r1 = InterleaveUpper(d, q0, q1);
442 + const auto r2 = InterleaveLower(d, q2, q3);
443 + const auto r3 = InterleaveUpper(d, q2, q3);
444 + const auto r4 = InterleaveLower(d, q4, q5);
445 + const auto r5 = InterleaveUpper(d, q4, q5);
446 + const auto r6 = InterleaveLower(d, q6, q7);
447 + const auto r7 = InterleaveUpper(d, q6, q7);
449 + i0 = ConcatLowerLower(d, r4, r0);
450 + i1 = ConcatLowerLower(d, r5, r1);
451 + i2 = ConcatLowerLower(d, r6, r2);
452 + i3 = ConcatLowerLower(d, r7, r3);
453 + i4 = ConcatUpperUpper(d, r4, r0);
454 + i5 = ConcatUpperUpper(d, r5, r1);
455 + i6 = ConcatUpperUpper(d, r6, r2);
456 + i7 = ConcatUpperUpper(d, r7, r3);
457 + to.StorePart(d, i0, m + 0, n + 0);
458 + to.StorePart(d, i1, m + 1, n + 0);
459 + to.StorePart(d, i2, m + 2, n + 0);
460 + to.StorePart(d, i3, m + 3, n + 0);
461 + to.StorePart(d, i4, m + 4, n + 0);
462 + to.StorePart(d, i5, m + 5, n + 0);
463 + to.StorePart(d, i6, m + 6, n + 0);
464 + to.StorePart(d, i7, m + 7, n + 0);
468 From cacad76ac565a9e008409207de5b2197f5128f1f Mon Sep 17 00:00:00 2001
469 From: Evgenii Kliuchnikov <eustas@google.com>
470 Date: Tue, 1 Feb 2022 15:09:13 +0000
471 Subject: [PATCH] Ramp-up EMCC and V8 versions for CI
473 Drive-by: ramp-up HWY version to 0.16.0rc
475 .github/workflows/build_test.yml | 6 +++---
477 lib/jxl/quant_weights.cc | 4 ++--
478 third_party/CMakeLists.txt | 1 +
479 third_party/highway | 2 +-
480 5 files changed, 8 insertions(+), 7 deletions(-)
482 #diff --git a/.github/workflows/build_test.yml b/.github/workflows/build_test.yml
483 #index 7c1e9fd72..2c99b0ab7 100644
484 #--- a/.github/workflows/build_test.yml
485 #+++ b/.github/workflows/build_test.yml
486 #@@ -461,8 +461,8 @@ jobs:
487 # runs-on: ubuntu-latest
489 # CCACHE_DIR: ${{ github.workspace }}/.ccache
490 #- EM_VERSION: 2.0.23
491 #- V8_VERSION: 9.3.22
493 #+ V8_VERSION: 9.8.177
494 # V8: ${{ github.workspace }}/.jsvu/v8
495 # BUILD_TARGET: wasm32
497 #@@ -506,7 +506,7 @@ jobs:
498 # ${{ runner.os }}-${{ steps.git-env.outputs.parent }}-${{ matrix.variant }}
500 # - name: Install emsdk
501 #- uses: mymindstorm/setup-emsdk@v10
502 #+ uses: mymindstorm/setup-emsdk@v11
503 # # TODO(deymo): We could cache this action but it doesn't work when running
506 #diff --git a/deps.sh b/deps.sh
507 #index 1abf18742..e2bbd755a 100755
510 #@@ -14,7 +14,7 @@ MYDIR=$(dirname $(realpath "$0"))
511 # # Git revisions we use for the given submodules. Update these whenever you
512 # # update a git submodule.
513 # THIRD_PARTY_GFLAGS="827c769e5fc98e0f2a34c47cef953cc6328abced"
514 #-THIRD_PARTY_HIGHWAY="e69083a12a05caf037cabecdf1b248b7579705a5"
515 #+THIRD_PARTY_HIGHWAY="f13e3b956eb226561ac79427893ec0afd66f91a8"
516 # THIRD_PARTY_SKCMS="64374756e03700d649f897dbd98c95e78c30c7da"
517 # THIRD_PARTY_SJPEG="868ab558fad70fcbe8863ba4e85179eeb81cc840"
518 # THIRD_PARTY_ZLIB="cacf7f1d4e3d44d871b605da3b647f07d718623f"
519 #diff --git a/lib/jxl/quant_weights.cc b/lib/jxl/quant_weights.cc
520 #index 2acd639b0..e8d9a10ed 100644
521 #--- a/lib/jxl/quant_weights.cc
522 #+++ b/lib/jxl/quant_weights.cc
523 #@@ -318,8 +318,8 @@ Status ComputeQuantTable(const QuantEncoding& encoding,
524 # HWY_CAPPED(float, 64) d;
525 # for (size_t i = 0; i < num * 3; i += Lanes(d)) {
526 # auto inv_val = LoadU(d, weights.data() + i);
527 #- if (JXL_UNLIKELY(!AllFalse(inv_val >= Set(d, 1.0f / kAlmostZero)) |
528 #- !AllFalse(inv_val < Set(d, kAlmostZero)))) {
529 #+ if (JXL_UNLIKELY(!AllFalse(d, inv_val >= Set(d, 1.0f / kAlmostZero)) ||
530 #+ !AllFalse(d, inv_val < Set(d, kAlmostZero)))) {
531 # return JXL_FAILURE("Invalid quantization table");
533 # auto val = Set(d, 1.0f) / inv_val;
534 #diff --git a/third_party/CMakeLists.txt b/third_party/CMakeLists.txt
535 #index 82d55d47a..afefbaa80 100644
536 #--- a/third_party/CMakeLists.txt
537 #+++ b/third_party/CMakeLists.txt
538 #@@ -82,6 +82,7 @@ endif() # BUILD_TESTING
541 # set(HWY_SYSTEM_GTEST ON CACHE INTERNAL "")
542 #+set(HWY_FORCE_STATIC_LIBS ON CACHE INTERNAL "")
543 # if((SANITIZER STREQUAL "asan") OR (SANITIZER STREQUAL "msan"))
544 # set(HWY_EXAMPLES_TESTS_INSTALL OFF CACHE INTERNAL "")
546 #diff --git a/third_party/highway b/third_party/highway
547 #index e69083a12..f13e3b956 160000
548 #--- a/third_party/highway
549 #+++ b/third_party/highway
551 #-Subproject commit e69083a12a05caf037cabecdf1b248b7579705a5
552 #+Subproject commit f13e3b956eb226561ac79427893ec0afd66f91a8
553 commit 600b591538827247f1b2ad6ee6ae627a2173b7ec
554 Author: Jan Wassenberg <janwas@google.com>
555 Date: Fri Feb 4 11:37:43 2022 +0100
557 Rename deprecated StoreFence to FlushStream
559 diff --git a/lib/profiler/profiler.cc b/lib/profiler/profiler.cc
560 index 63d8569..c72656e 100644
561 --- a/lib/profiler/profiler.cc
562 +++ b/lib/profiler/profiler.cc
563 @@ -138,7 +138,7 @@ class Results {
564 void AnalyzePackets(const Packet* HWY_RESTRICT packets,
565 const size_t num_packets) {
566 // Ensures prior weakly-ordered streaming stores are globally visible.
568 + hwy::FlushStream();
570 const uint64_t t0 = TicksBefore();
572 @@ -448,12 +448,12 @@ void ThreadSpecific::ComputeOverhead() {
573 const size_t kReps = 10000;
574 // Analysis time should not be included => must fit within buffer.
575 HWY_ASSERT(kReps * 2 < max_packets_);
577 + hwy::FlushStream();
578 const uint64_t t0 = TicksBefore();
579 for (size_t i = 0; i < kReps; ++i) {
580 PROFILER_ZONE("Dummy");
583 + hwy::FlushStream();
584 const uint64_t t1 = TicksAfter();
585 HWY_ASSERT(num_packets_ + buffer_size_ == kReps * 2);
587 #diff --git a/lib/profiler/tsc_timer.h b/lib/profiler/tsc_timer.h
588 #index d3c1bee..802e40d 100644
589 #--- a/lib/profiler/tsc_timer.h
590 #+++ b/lib/profiler/tsc_timer.h
592 # #include <windows.h>
593 # // Undef macros to avoid collisions
598 # #if defined(__MACH__)