gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright (c) 2002-2012 The ANGLE Project Authors. All rights reserved. |
| 3 | // Use of this source code is governed by a BSD-style license that can be |
| 4 | // found in the LICENSE file. |
| 5 | // |
Geoff Lang | 1773282 | 2013-08-29 13:46:49 -0400 | [diff] [blame^] | 6 | #include "compiler/translator/VariablePacker.h" |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 7 | |
| 8 | #include <algorithm> |
Geoff Lang | 1773282 | 2013-08-29 13:46:49 -0400 | [diff] [blame^] | 9 | #include "compiler/translator/ShHandle.h" |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 10 | |
| 11 | namespace { |
| 12 | int GetSortOrder(ShDataType type) |
| 13 | { |
| 14 | switch (type) { |
| 15 | case SH_FLOAT_MAT4: |
shannonwoods@chromium.org | 8da034c | 2013-05-30 00:19:15 +0000 | [diff] [blame] | 16 | case SH_FLOAT_MAT2x4: |
| 17 | case SH_FLOAT_MAT3x4: |
| 18 | case SH_FLOAT_MAT4x2: |
| 19 | case SH_FLOAT_MAT4x3: |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 20 | return 0; |
| 21 | case SH_FLOAT_MAT2: |
| 22 | return 1; |
| 23 | case SH_FLOAT_VEC4: |
| 24 | case SH_INT_VEC4: |
| 25 | case SH_BOOL_VEC4: |
| 26 | return 2; |
| 27 | case SH_FLOAT_MAT3: |
shannonwoods@chromium.org | 8da034c | 2013-05-30 00:19:15 +0000 | [diff] [blame] | 28 | case SH_FLOAT_MAT2x3: |
| 29 | case SH_FLOAT_MAT3x2: |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 30 | return 3; |
| 31 | case SH_FLOAT_VEC3: |
| 32 | case SH_INT_VEC3: |
| 33 | case SH_BOOL_VEC3: |
| 34 | return 4; |
| 35 | case SH_FLOAT_VEC2: |
| 36 | case SH_INT_VEC2: |
| 37 | case SH_BOOL_VEC2: |
| 38 | return 5; |
| 39 | case SH_FLOAT: |
| 40 | case SH_INT: |
| 41 | case SH_BOOL: |
| 42 | case SH_SAMPLER_2D: |
| 43 | case SH_SAMPLER_CUBE: |
| 44 | case SH_SAMPLER_EXTERNAL_OES: |
| 45 | case SH_SAMPLER_2D_RECT_ARB: |
| 46 | return 6; |
| 47 | default: |
| 48 | ASSERT(false); |
| 49 | return 7; |
| 50 | } |
| 51 | } |
| 52 | } // namespace |
| 53 | |
| 54 | int VariablePacker::GetNumComponentsPerRow(ShDataType type) |
| 55 | { |
| 56 | switch (type) { |
| 57 | case SH_FLOAT_MAT4: |
| 58 | case SH_FLOAT_MAT2: |
shannonwoods@chromium.org | 8da034c | 2013-05-30 00:19:15 +0000 | [diff] [blame] | 59 | case SH_FLOAT_MAT2x4: |
| 60 | case SH_FLOAT_MAT3x4: |
| 61 | case SH_FLOAT_MAT4x2: |
| 62 | case SH_FLOAT_MAT4x3: |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 63 | case SH_FLOAT_VEC4: |
| 64 | case SH_INT_VEC4: |
| 65 | case SH_BOOL_VEC4: |
| 66 | return 4; |
| 67 | case SH_FLOAT_MAT3: |
shannonwoods@chromium.org | 8da034c | 2013-05-30 00:19:15 +0000 | [diff] [blame] | 68 | case SH_FLOAT_MAT2x3: |
| 69 | case SH_FLOAT_MAT3x2: |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 70 | case SH_FLOAT_VEC3: |
| 71 | case SH_INT_VEC3: |
| 72 | case SH_BOOL_VEC3: |
| 73 | return 3; |
| 74 | case SH_FLOAT_VEC2: |
| 75 | case SH_INT_VEC2: |
| 76 | case SH_BOOL_VEC2: |
| 77 | return 2; |
| 78 | case SH_FLOAT: |
| 79 | case SH_INT: |
| 80 | case SH_BOOL: |
| 81 | case SH_SAMPLER_2D: |
| 82 | case SH_SAMPLER_CUBE: |
| 83 | case SH_SAMPLER_EXTERNAL_OES: |
| 84 | case SH_SAMPLER_2D_RECT_ARB: |
| 85 | return 1; |
| 86 | default: |
| 87 | ASSERT(false); |
| 88 | return 5; |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | int VariablePacker::GetNumRows(ShDataType type) |
| 93 | { |
| 94 | switch (type) { |
| 95 | case SH_FLOAT_MAT4: |
shannonwoods@chromium.org | 8da034c | 2013-05-30 00:19:15 +0000 | [diff] [blame] | 96 | case SH_FLOAT_MAT2x4: |
| 97 | case SH_FLOAT_MAT3x4: |
| 98 | case SH_FLOAT_MAT4x3: |
| 99 | case SH_FLOAT_MAT4x2: |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 100 | return 4; |
| 101 | case SH_FLOAT_MAT3: |
shannonwoods@chromium.org | 8da034c | 2013-05-30 00:19:15 +0000 | [diff] [blame] | 102 | case SH_FLOAT_MAT2x3: |
| 103 | case SH_FLOAT_MAT3x2: |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 104 | return 3; |
| 105 | case SH_FLOAT_MAT2: |
daniel@transgaming.com | d5903ca | 2013-02-01 03:20:31 +0000 | [diff] [blame] | 106 | return 2; |
gman@chromium.org | 8d80479 | 2012-10-17 21:33:48 +0000 | [diff] [blame] | 107 | case SH_FLOAT_VEC4: |
| 108 | case SH_INT_VEC4: |
| 109 | case SH_BOOL_VEC4: |
| 110 | case SH_FLOAT_VEC3: |
| 111 | case SH_INT_VEC3: |
| 112 | case SH_BOOL_VEC3: |
| 113 | case SH_FLOAT_VEC2: |
| 114 | case SH_INT_VEC2: |
| 115 | case SH_BOOL_VEC2: |
| 116 | case SH_FLOAT: |
| 117 | case SH_INT: |
| 118 | case SH_BOOL: |
| 119 | case SH_SAMPLER_2D: |
| 120 | case SH_SAMPLER_CUBE: |
| 121 | case SH_SAMPLER_EXTERNAL_OES: |
| 122 | case SH_SAMPLER_2D_RECT_ARB: |
| 123 | return 1; |
| 124 | default: |
| 125 | ASSERT(false); |
| 126 | return 100000; |
| 127 | } |
| 128 | } |
| 129 | |
| 130 | struct TVariableInfoComparer { |
| 131 | bool operator()(const TVariableInfo& lhs, const TVariableInfo& rhs) const |
| 132 | { |
| 133 | int lhsSortOrder = GetSortOrder(lhs.type); |
| 134 | int rhsSortOrder = GetSortOrder(rhs.type); |
| 135 | if (lhsSortOrder != rhsSortOrder) { |
| 136 | return lhsSortOrder < rhsSortOrder; |
| 137 | } |
| 138 | // Sort by largest first. |
| 139 | return lhs.size > rhs.size; |
| 140 | } |
| 141 | }; |
| 142 | |
| 143 | unsigned VariablePacker::makeColumnFlags(int column, int numComponentsPerRow) |
| 144 | { |
| 145 | return ((kColumnMask << (kNumColumns - numComponentsPerRow)) & |
| 146 | kColumnMask) >> column; |
| 147 | } |
| 148 | |
| 149 | void VariablePacker::fillColumns(int topRow, int numRows, int column, int numComponentsPerRow) |
| 150 | { |
| 151 | unsigned columnFlags = makeColumnFlags(column, numComponentsPerRow); |
| 152 | for (int r = 0; r < numRows; ++r) { |
| 153 | int row = topRow + r; |
| 154 | ASSERT((rows_[row] & columnFlags) == 0); |
| 155 | rows_[row] |= columnFlags; |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | bool VariablePacker::searchColumn(int column, int numRows, int* destRow, int* destSize) |
| 160 | { |
| 161 | ASSERT(destRow); |
| 162 | |
| 163 | for (; topNonFullRow_ < maxRows_ && rows_[topNonFullRow_] == kColumnMask; |
| 164 | ++topNonFullRow_) { |
| 165 | } |
| 166 | |
| 167 | for (; bottomNonFullRow_ >= 0 && rows_[bottomNonFullRow_] == kColumnMask; |
| 168 | --bottomNonFullRow_) { |
| 169 | } |
| 170 | |
| 171 | if (bottomNonFullRow_ - topNonFullRow_ + 1 < numRows) { |
| 172 | return false; |
| 173 | } |
| 174 | |
| 175 | unsigned columnFlags = makeColumnFlags(column, 1); |
| 176 | int topGoodRow = 0; |
| 177 | int smallestGoodTop = -1; |
| 178 | int smallestGoodSize = maxRows_ + 1; |
| 179 | int bottomRow = bottomNonFullRow_ + 1; |
| 180 | bool found = false; |
| 181 | for (int row = topNonFullRow_; row <= bottomRow; ++row) { |
| 182 | bool rowEmpty = row < bottomRow ? ((rows_[row] & columnFlags) == 0) : false; |
| 183 | if (rowEmpty) { |
| 184 | if (!found) { |
| 185 | topGoodRow = row; |
| 186 | found = true; |
| 187 | } |
| 188 | } else { |
| 189 | if (found) { |
| 190 | int size = row - topGoodRow; |
| 191 | if (size >= numRows && size < smallestGoodSize) { |
| 192 | smallestGoodSize = size; |
| 193 | smallestGoodTop = topGoodRow; |
| 194 | } |
| 195 | } |
| 196 | found = false; |
| 197 | } |
| 198 | } |
| 199 | if (smallestGoodTop < 0) { |
| 200 | return false; |
| 201 | } |
| 202 | |
| 203 | *destRow = smallestGoodTop; |
| 204 | if (destSize) { |
| 205 | *destSize = smallestGoodSize; |
| 206 | } |
| 207 | return true; |
| 208 | } |
| 209 | |
| 210 | bool VariablePacker::CheckVariablesWithinPackingLimits(int maxVectors, const TVariableInfoList& in_variables) |
| 211 | { |
| 212 | ASSERT(maxVectors > 0); |
| 213 | maxRows_ = maxVectors; |
| 214 | topNonFullRow_ = 0; |
| 215 | bottomNonFullRow_ = maxRows_ - 1; |
| 216 | TVariableInfoList variables(in_variables); |
| 217 | |
| 218 | // As per GLSL 1.017 Appendix A, Section 7 variables are packed in specific |
| 219 | // order by type, then by size of array, largest first. |
| 220 | std::sort(variables.begin(), variables.end(), TVariableInfoComparer()); |
| 221 | rows_.clear(); |
| 222 | rows_.resize(maxVectors, 0); |
| 223 | |
| 224 | // Packs the 4 column variables. |
| 225 | size_t ii = 0; |
| 226 | for (; ii < variables.size(); ++ii) { |
| 227 | const TVariableInfo& variable = variables[ii]; |
| 228 | if (GetNumComponentsPerRow(variable.type) != 4) { |
| 229 | break; |
| 230 | } |
| 231 | topNonFullRow_ += GetNumRows(variable.type) * variable.size; |
| 232 | } |
| 233 | |
| 234 | if (topNonFullRow_ > maxRows_) { |
| 235 | return false; |
| 236 | } |
| 237 | |
| 238 | // Packs the 3 column variables. |
| 239 | int num3ColumnRows = 0; |
| 240 | for (; ii < variables.size(); ++ii) { |
| 241 | const TVariableInfo& variable = variables[ii]; |
| 242 | if (GetNumComponentsPerRow(variable.type) != 3) { |
| 243 | break; |
| 244 | } |
| 245 | num3ColumnRows += GetNumRows(variable.type) * variable.size; |
| 246 | } |
| 247 | |
| 248 | if (topNonFullRow_ + num3ColumnRows > maxRows_) { |
| 249 | return false; |
| 250 | } |
| 251 | |
| 252 | fillColumns(topNonFullRow_, num3ColumnRows, 0, 3); |
| 253 | |
| 254 | // Packs the 2 column variables. |
| 255 | int top2ColumnRow = topNonFullRow_ + num3ColumnRows; |
| 256 | int twoColumnRowsAvailable = maxRows_ - top2ColumnRow; |
| 257 | int rowsAvailableInColumns01 = twoColumnRowsAvailable; |
| 258 | int rowsAvailableInColumns23 = twoColumnRowsAvailable; |
| 259 | for (; ii < variables.size(); ++ii) { |
| 260 | const TVariableInfo& variable = variables[ii]; |
| 261 | if (GetNumComponentsPerRow(variable.type) != 2) { |
| 262 | break; |
| 263 | } |
| 264 | int numRows = GetNumRows(variable.type) * variable.size; |
| 265 | if (numRows <= rowsAvailableInColumns01) { |
| 266 | rowsAvailableInColumns01 -= numRows; |
| 267 | } else if (numRows <= rowsAvailableInColumns23) { |
| 268 | rowsAvailableInColumns23 -= numRows; |
| 269 | } else { |
| 270 | return false; |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | int numRowsUsedInColumns01 = |
| 275 | twoColumnRowsAvailable - rowsAvailableInColumns01; |
| 276 | int numRowsUsedInColumns23 = |
| 277 | twoColumnRowsAvailable - rowsAvailableInColumns23; |
| 278 | fillColumns(top2ColumnRow, numRowsUsedInColumns01, 0, 2); |
| 279 | fillColumns(maxRows_ - numRowsUsedInColumns23, numRowsUsedInColumns23, |
| 280 | 2, 2); |
| 281 | |
| 282 | // Packs the 1 column variables. |
| 283 | for (; ii < variables.size(); ++ii) { |
| 284 | const TVariableInfo& variable = variables[ii]; |
| 285 | ASSERT(1 == GetNumComponentsPerRow(variable.type)); |
| 286 | int numRows = GetNumRows(variable.type) * variable.size; |
| 287 | int smallestColumn = -1; |
| 288 | int smallestSize = maxRows_ + 1; |
| 289 | int topRow = -1; |
| 290 | for (int column = 0; column < kNumColumns; ++column) { |
| 291 | int row = 0; |
| 292 | int size = 0; |
| 293 | if (searchColumn(column, numRows, &row, &size)) { |
| 294 | if (size < smallestSize) { |
| 295 | smallestSize = size; |
| 296 | smallestColumn = column; |
| 297 | topRow = row; |
| 298 | } |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | if (smallestColumn < 0) { |
| 303 | return false; |
| 304 | } |
| 305 | |
| 306 | fillColumns(topRow, numRows, smallestColumn, 1); |
| 307 | } |
| 308 | |
| 309 | ASSERT(variables.size() == ii); |
| 310 | |
| 311 | return true; |
| 312 | } |
| 313 | |
| 314 | |
| 315 | |