[Mesa-dev] [PATCH v4 07/15] mesa/st/glsl_to_tgsi: Add array merge logic
Gert Wollny
gw.fossdev at gmail.com
Tue Jun 5 20:26:41 UTC 2018
v4: - Update the code to use the new merge logic.
- Use a cleaner, class-based approach for the evaluation of merges.
Signed-off-by: Gert Wollny <gw.fossdev at gmail.com>
---
.../state_tracker/st_glsl_to_tgsi_array_merge.cpp | 383 ++++++++++++++++++++-
.../state_tracker/st_glsl_to_tgsi_array_merge.h | 26 +-
2 files changed, 407 insertions(+), 2 deletions(-)
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.cpp
index a6c54c43df..0bdb42b75b 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.cpp
@@ -21,6 +21,109 @@
* DEALINGS IN THE SOFTWARE.
*/
+/* A short overview on how the array merging works:
+ *
+ * Inputs:
+ * - per array information: live range, access mask, size
+ * - the program
+ *
+ * Output:
+ * - the program with updated array addressing
+ *
+ * Pseudo algorithm:
+ *
+ * repeat
+ * for all pairs of arrays:
+ * if they have non-overlapping live ranges and equal access masks:
+ * - pick shorter array
+ * - merge its live range into the longer array
+ * - set its merge target array to the longer array
+ * - mark the shorter array as processed
+ *
+ * for all pairs of arrays:
+ * if they have overlapping live ranges use in sum at most four components:
+ * - pick shorter array
+ * - evaluate reswizzle map to move its components into the components
+ * that are not used by the longer array
+ * - set its merge target array to the longer array
+ * - mark the shorter array as processed
+ * - bail out loop
+ * until no more successfull merges were found
+ *
+ * for all pairs of arrays:
+ * if they have non-overlapping live ranges:
+ * - pick shorter array
+ * - merge its live range into the longer array
+ * - set its merge target array to the longer array
+ * - mark the shorter array as processed
+ *
+ * Finalize remapping map so that target arrays are always final, i.e. have
+ * themselfes no merge target set.
+ *
+ * Example:
+ * ID | Length | Live range | access mask | target id | reswizzle
+ * ================================================================
+ * 1 3 3-10 x___ 0 ____
+ * 2 4 13-20 x___ 0 ____
+ * 3 8 3-20 x___ 0 ____
+ * 4 6 21-40 xy__ 0 ____
+ * 5 7 12-30 xy__ 0 ____
+ *
+ * 1. merge live ranges 1 and 2
+ *
+ * ID | Length | Live range | access mask | target id | reswizzle
+ * ================================================================
+ * 1 - - x___ 2 ____
+ * 2 4 3-20 x___ 0 ____
+ * 3 8 3-20 x___ 0 ____
+ * 4 6 21-40 xy__ 0 ____
+ * 5 7 12-30 xy__ 0 ____
+ *
+ *
+ * 3. interleave 2 and 3
+ *
+ * ID | Length | Live range | access mask | target id | reswizzle
+ * ================================================================
+ * 1 - - x___ 2 ____
+ * 2 - - x___ 3 _x__
+ * 3 8 3-20 xy__ 0 ____
+ * 4 6 21-40 xy__ 0 ____
+ * 5 7 12-30 xy__ 0 ____
+ *
+ * 3. merge live ranges 3 and 4
+ *
+ * ID | Length | Live range | access mask | target id | reswizzle
+ * ================================================================
+ * 1 - - x___ 2 ____
+ * 2 - - x___ 3 _x__
+ * 3 8 3-40 xy__ 0 ____
+ * 4 - - xy__ 3 ____
+ * 5 7 3-21 xy__ 0 ____
+ *
+ * 4. interleave 3 and 5
+ *
+ * ID | Length | Live range | access mask | target id | reswizzle
+ * ================================================================
+ * 1 - - x___ 2 ____
+ * 2 - - x___ 3 _x__
+ * 3 8 3-40 xy__ 0 ____
+ * 4 - - xy__ 3 ____
+ * 5 - - xy__ 3 __xy
+ *
+ * 5. finalize remapping
+ * (Array 1 has been merged with 2 that was later interleaved, so
+ * the reswizzeling must be propagated.
+ *
+ * ID | Length | Live range | new access mask | target id | reswizzle
+ * ================================================================
+ * 1 - - _y__ 3 _x__
+ * 2 - - _y__ 3 _x__
+ * 3 8 3-40 xy__ 0 ____
+ * 4 - - xy__ 3 ____
+ * 5 - - __zw 3 __xy
+ *
+*/
+
#include "program/prog_instruction.h"
#include "util/u_math.h"
#include <ostream>
@@ -313,5 +416,283 @@ bool operator == (const array_remapping& lhs, const array_remapping& rhs)
return true;
}
-/* end namespace tgsi_array_merge */
+static
+bool sort_by_begin(const array_live_range& lhs, const array_live_range& rhs) {
+ return lhs.begin() < rhs.begin();
+}
+
+/* Helper class to evaluate merging and interleaving of arrays */
+class array_merge_evaluator {
+public:
+ typedef int (*array_merger)(array_live_range& range_1,
+ array_live_range& range_2);
+
+ array_merge_evaluator(int _narrays, array_live_range *_ranges,
+ bool _restart);
+
+ /** Run the merge strategy on all arrays
+ * @returns number of successfull merges
+ */
+ int run();
+
+private:
+ virtual int do_run(array_live_range& range_1, array_live_range& range_2) = 0;
+
+ int narrays;
+ array_live_range *ranges;
+ bool restart;
+};
+
+array_merge_evaluator::array_merge_evaluator(int _narrays,
+ array_live_range *_ranges,
+ bool _restart):
+ narrays(_narrays),
+ ranges(_ranges),
+ restart(_restart)
+{
+}
+
+int array_merge_evaluator::run()
+{
+ int remaps = 0;
+
+ for (int i = 0; i < narrays; ++i) {
+ if (ranges[i].is_mapped())
+ continue;
+
+ for (int j = i + 1; j < narrays; ++j) {
+ if (!ranges[j].is_mapped()) {
+ ARRAY_MERGE_DUMP("try merge " << i << " id:" << ranges[i].array_id()
+ << " and " << j << " id: "<< ranges[j].array_id()
+ << "\n");
+ int n = do_run(ranges[i], ranges[j]);
+ if (restart && n)
+ return n;
+ remaps += n;
+ }
+ }
+ }
+ return remaps;
+}
+
+/* Merge live ranges if possible at all */
+class merge_live_range_always: public array_merge_evaluator {
+public:
+ merge_live_range_always(int _narrays, array_live_range *_ranges):
+ array_merge_evaluator(_narrays, _ranges, false) {
+ }
+protected:
+ int do_run(array_live_range& range_1, array_live_range& range_2){
+ if (range_2.time_doesnt_overlap(range_1)) {
+ ARRAY_MERGE_DUMP("merge " << range_2 << " into " << range_1 << "\n");
+ array_live_range::merge(&range_1,&range_2);
+ return 1;
+ }
+ return 0;
+ }
+};
+
+/* Merge live ranges only if they use the same swizzle */
+class merge_live_range_equal_swizzle: public merge_live_range_always {
+public:
+ merge_live_range_equal_swizzle(int _narrays, array_live_range *_ranges):
+ merge_live_range_always(_narrays, _ranges) {
+ }
+private:
+ int do_run(array_live_range& range_1, array_live_range& range_2){
+ if (range_1.access_mask() == range_2.access_mask()) {
+ return merge_live_range_always::do_run(range_1, range_2);
+ }
+ return 0;
+ }
+};
+
+/* Interleave arrays if possible */
+class interleave_live_range: public array_merge_evaluator {
+public:
+ interleave_live_range(int _narrays, array_live_range *_ranges):
+ array_merge_evaluator(_narrays, _ranges, true) {
+ }
+private:
+ int do_run(array_live_range& range_1, array_live_range& range_2){
+ if ((range_2.used_components() + range_1.used_components() <= 4) &&
+ !range_1.time_doesnt_overlap(range_2)) {
+ ARRAY_MERGE_DUMP("Interleave " << range_2 << " into " << range_1 << "\n");
+ array_live_range::interleave(&range_1, &range_2);
+ return 1;
+ }
+ return 0;
+ }
+};
+
+/* Estimate the array merging: First in a loop, arrays with equal access mask
+ * are merged, then interleave arrays that together use at most four components,
+ * and have overlapping live ranges. Finally arrays are merged regardless of
+ * access mask.
+ * @param[in] narrays number of arrays
+ * @param[in,out] alt array life times, the merge target life time will be
+ * updated with the new life time.
+ * @param[in,out] remapping track the arraay index remapping and reswizzeling.
+ * @returns number of merged arrays
+ */
+bool get_array_remapping(int narrays, array_live_range *ranges,
+ array_remapping *remapping)
+{
+ int total_remapped = 0;
+ int n_remapped;
+
+ /* Sort by "begin of live range" so that we don't have to restart searching
+ * after every merge.
+ */
+ std::sort(ranges, ranges + narrays, sort_by_begin);
+ merge_live_range_equal_swizzle merge_evaluator_es(narrays, ranges);
+ interleave_live_range interleave_lr(narrays, ranges);
+ do {
+
+ n_remapped = merge_evaluator_es.run();
+
+ /* try only one array interleave, if successfull, another
+ * live_range merge is tried. The test MergeAndInterleave5
+ * (mesa/st/tests/test_glsl_to_tgsi_array_merge.cpp)
+ * shows that this can result in more arrays being merged/interleaved.
+ */
+ n_remapped += interleave_lr.run();
+ total_remapped += n_remapped;
+
+ ARRAY_MERGE_DUMP("Remapped " << n_remapped << " arrays\n");
+ } while (n_remapped > 0);
+
+ total_remapped += merge_live_range_always(narrays, ranges).run();
+ ARRAY_MERGE_DUMP("Remapped a total of " << total_remapped << " arrays\n");
+
+ /* Resolve the remapping chain */
+ for (int i = 1; i <= narrays; ++i) {
+ ARRAY_MERGE_DUMP("Map " << i << ":");
+ remapping[ranges[i-1].array_id()].init_from(ranges[i-1]);
+ }
+ return total_remapped > 0;
+}
+
+/* Remap the arrays in a TGSI program according to the given mapping.
+ * @param narrays number of arrays
+ * @param array_sizes array of arrays sizes
+ * @param map the array remapping information
+ * @param instructions TGSI program
+ * @returns number of arrays after remapping
+ */
+int remap_arrays(int narrays, unsigned *array_sizes,
+ exec_list *instructions,
+ array_remapping *map)
+{
+ /* re-calculate arrays */
+#if __cplusplus < 201402L
+ int *idx_map = new int[narrays + 1];
+ unsigned *old_sizes = new unsigned[narrays + 1];
+#else
+ unique_ptr<int[]> idx_map = make_unique<int[]>(narrays + 1);
+ unique_ptr<unsigned[]> old_sizes = make_unique<unsigned[]>(narrays + 1);
+#endif
+
+ memcpy(&old_sizes[0], &array_sizes[0], sizeof(unsigned) * narrays);
+
+ /* Evaluate mapping for the array indices and update array sizes */
+ int new_narrays = 0;
+ for (int i = 1; i <= narrays; ++i) {
+ if (!map[i].is_valid()) {
+ ++new_narrays;
+ idx_map[i] = new_narrays;
+ array_sizes[new_narrays] = old_sizes[i];
+ }
+ }
+
+ /* Map the array ids of merged arrays. */
+ for (int i = 1; i <= narrays; ++i) {
+ if (map[i].is_valid()) {
+ map[i].set_target_id(idx_map[map[i].target_array_id()]);
+ }
+ }
+
+ /* Map the array ids of merge targets that got only renumbered. */
+ for (int i = 1; i <= narrays; ++i) {
+ if (!map[i].is_valid()) {
+ map[i].set_target_id(idx_map[i]);
+ }
+ }
+
+ /* Update the array ids and swizzles in the registers */
+ foreach_in_list(glsl_to_tgsi_instruction, inst, instructions) {
+ for (unsigned j = 0; j < num_inst_src_regs(inst); j++) {
+ st_src_reg& src = inst->src[j];
+ if (src.file == PROGRAM_ARRAY && src.array_id > 0) {
+ array_remapping& m = map[src.array_id];
+ if (m.is_valid()) {
+ src.array_id = m.target_array_id();
+ src.swizzle = m.map_swizzles(src.swizzle);
+ }
+ }
+ }
+ for (unsigned j = 0; j < inst->tex_offset_num_offset; j++) {
+ st_src_reg& src = inst->tex_offsets[j];
+ if (src.file == PROGRAM_ARRAY && src.array_id > 0) {
+ array_remapping& m = map[src.array_id];
+ if (m.is_valid()) {
+ src.array_id = m.target_array_id();
+ src.swizzle = m.map_swizzles(src.swizzle);
+ }
+ }
+ }
+ for (unsigned j = 0; j < num_inst_dst_regs(inst); j++) {
+ st_dst_reg& dst = inst->dst[j];
+ if (dst.file == PROGRAM_ARRAY && dst.array_id > 0) {
+ array_remapping& m = map[dst.array_id];
+ if (m.is_valid()) {
+ assert(j == 0 &&
+ "remapping can only be done for single dest ops");
+ dst.array_id = m.target_array_id();
+ dst.writemask = m.map_writemask(dst.writemask);
+
+ /* If the target component is moved, then the source swizzles
+ * must be moved accordingly.
+ */
+ for (unsigned j = 0; j < num_inst_src_regs(inst); j++) {
+ st_src_reg& src = inst->src[j];
+ src.swizzle = m.move_read_swizzles(src.swizzle);
+ }
+ }
+ }
+ }
+ st_src_reg& res = inst->resource;
+ if (res.file == PROGRAM_ARRAY && res.array_id > 0) {
+ array_remapping& m = map[res.array_id];
+ if (m.is_valid()) {
+ res.array_id = m.target_array_id();
+ res.swizzle = m.map_swizzles(res.swizzle);
+ }
+ }
+ }
+
+#if __cplusplus < 201402L
+ delete[] old_sizes;
+ delete[] idx_map;
+#endif
+
+ return new_narrays;
+}
+
}
+
+using namespace tgsi_array_merge;
+
+int merge_arrays(int narrays,
+ unsigned *array_sizes,
+ exec_list *instructions,
+ struct array_live_range *arr_live_ranges)
+{
+ array_remapping *map= new array_remapping[narrays + 1];
+
+ if (get_array_remapping(narrays, arr_live_ranges, map))
+ narrays = remap_arrays(narrays, array_sizes, instructions, map);
+
+ delete[] map;
+ return narrays;
+}
\ No newline at end of file
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.h b/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.h
index 9915bc2e4b..66da0c379c 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.h
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi_array_merge.h
@@ -157,5 +157,29 @@ std::ostream& operator << (std::ostream& os, const array_remapping& am)
return os;
}
+/* Apply the array remapping (internal use, exposed here for testing) */
+ bool get_array_remapping(int narrays, array_live_range *array_live_ranges,
+ array_remapping *remapping);
+
+/* Apply the array remapping (internal use, exposed here for testing) */
+int remap_arrays(int narrays, unsigned *array_sizes,
+ exec_list *instructions,
+ array_remapping *map);
+
}
-#endif
+
+/** Remap the array access to finalize the array merging and interleaving.
+ * @param[in] narrays number of input arrays,
+ * @param[in,out] array_sizes length array of input arrays, on output the
+ * array sizes will be updated according to the remapping,
+ * @param[in,out] instructions TGSI program, on output the arrays access is
+ * remapped to the new array layout,
+ * @param[in] array_live_ranges live ranges and access information of the
+ * arrays.
+ * @returns number of remaining arrays
+ */
+int merge_arrays(int narrays,
+ unsigned *array_sizes,
+ exec_list *instructions,
+ struct array_live_range *arr_live_ranges);
+#endif
\ No newline at end of file
--
2.16.4
More information about the mesa-dev
mailing list