From: Marek Olšák <marek.ol...@amd.com>
---
src/gallium/auxiliary/util/u_vbuf.c | 189 ++++++++++++++++++++++++----
1 file changed, 165 insertions(+), 24 deletions(-)
diff --git a/src/gallium/auxiliary/util/u_vbuf.c
b/src/gallium/auxiliary/util/u_vbuf.c
index 87b159ec1bf..b0b92f7e966 100644
--- a/src/gallium/auxiliary/util/u_vbuf.c
+++ b/src/gallium/auxiliary/util/u_vbuf.c
@@ -1132,20 +1132,45 @@ static void u_vbuf_set_driver_vertex_buffers(struct
u_vbuf *mgr)
unsigned start_slot, count;
start_slot = ffs(mgr->dirty_real_vb_mask) - 1;
count = util_last_bit(mgr->dirty_real_vb_mask >> start_slot);
pipe->set_vertex_buffers(pipe, start_slot, count,
mgr->real_vertex_buffer + start_slot);
mgr->dirty_real_vb_mask = 0;
}
+static void
+u_vbuf_split_indexed_multidraw(struct u_vbuf *mgr, struct pipe_draw_info *info,
+ unsigned *indirect_data, unsigned stride,
+ unsigned draw_count)
+{
+ assert(info->index_size);
+ info->indirect = NULL;
+
+ for (unsigned i = 0; i < draw_count; i++) {
+ unsigned offset = i * stride / 4;
+
+ info->count = indirect_data[offset + 0];
+ info->instance_count = indirect_data[offset + 1];
+
+ if (!info->count || !info->instance_count)
+ continue;
+
+ info->start = indirect_data[offset + 2];
+ info->index_bias = indirect_data[offset + 3];
+ info->start_instance = indirect_data[offset + 4];
+
+ u_vbuf_draw_vbo(mgr, info);
+ }
+}
+
void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct pipe_draw_info *info)
{
struct pipe_context *pipe = mgr->pipe;
int start_vertex, min_index;
unsigned num_vertices;
boolean unroll_indices = FALSE;
const uint32_t used_vb_mask = mgr->ve->used_vb_mask;
uint32_t user_vb_mask = mgr->user_vb_mask & used_vb_mask;
const uint32_t incompatible_vb_mask =
mgr->incompatible_vb_mask & used_vb_mask;
@@ -1160,47 +1185,162 @@ void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct
pipe_draw_info *info)
if (mgr->dirty_real_vb_mask & used_vb_mask) {
u_vbuf_set_driver_vertex_buffers(mgr);
}
pipe->draw_vbo(pipe, info);
return;
}
new_info = *info;
- /* Fallback. We need to know all the parameters. */
+ /* Handle indirect (multi)draws. */
if (new_info.indirect) {
- struct pipe_transfer *transfer = NULL;
- int *data;
-
- if (new_info.index_size) {
- data = pipe_buffer_map_range(pipe, new_info.indirect->buffer,
- new_info.indirect->offset, 20,
- PIPE_TRANSFER_READ, &transfer);
- new_info.index_bias = data[3];
- new_info.start_instance = data[4];
- }
- else {
- data = pipe_buffer_map_range(pipe, new_info.indirect->buffer,
- new_info.indirect->offset, 16,
- PIPE_TRANSFER_READ, &transfer);
- new_info.start_instance = data[3];
+ const struct pipe_draw_indirect_info *indirect = new_info.indirect;
+ unsigned draw_count = 0;
+
+ /* Get the number of draws. */
+ if (indirect->indirect_draw_count) {
+ pipe_buffer_read(pipe, indirect->indirect_draw_count,
+ indirect->indirect_draw_count_offset,
+ 4, &draw_count);
+ } else {
+ draw_count = indirect->draw_count;
}
- new_info.count = data[0];
- new_info.instance_count = data[1];
- new_info.start = data[2];
- pipe_buffer_unmap(pipe, transfer);
- new_info.indirect = NULL;
-
- if (!new_info.count)
+ if (!draw_count)
return;
+
+ unsigned data_size = (draw_count - 1) * indirect->stride +
+ (new_info.index_size ? 20 : 16);
+ unsigned *data = alloca(data_size);
+
+ /* Read the used buffer range only once, because the read can be
+ * uncached.
+ */
+ pipe_buffer_read(pipe, indirect->buffer, indirect->offset, data_size,
+ data);
+
+ if (info->index_size) {
+ /* Indexed multidraw. */
+ unsigned index_bias0 = data[3];
+ bool index_bias_same = true;
+
+ /* If we invoke the translate path, we have to split the multidraw. */
+ if (incompatible_vb_mask ||
+ mgr->ve->incompatible_elem_mask) {
+ u_vbuf_split_indexed_multidraw(mgr, &new_info, data,
+ indirect->stride, draw_count);
+ return;
+ }
+
+ /* See if index_bias is the same for all draws. */
+ for (unsigned i = 1; i < draw_count; i++) {
+ if (data[i * indirect->stride / 4 + 3] != index_bias0) {
+ index_bias_same = false;
+ break;
+ }
+ }
+
+ /* Split the multidraw if index_bias is different. */
+ if (!index_bias_same) {
+ u_vbuf_split_indexed_multidraw(mgr, &new_info, data,
+ indirect->stride, draw_count);
+ return;
+ }
+
+ /* If we don't need to use the translate path and index_bias is
+ * the same, we can process the multidraw with the time complexity
+ * equal to 1 draw call (except for the index range computation).
+ * We only need to compute the index range covering all draw calls
+ * of the multidraw.
+ *
+ * The driver will not look at these values because indirect != NULL.
+ * These values determine the user buffer bounds to upload.
+ */
+ new_info.index_bias = index_bias0;
+ new_info.min_index = ~0u;
+ new_info.max_index = 0;
+ new_info.start_instance = ~0u;
+ new_info.instance_count = 0;
+
+ struct pipe_transfer *transfer = NULL;
+ const uint8_t *indices;
+
+ if (info->has_user_indices) {
+ indices = (uint8_t*)info->index.user;
+ } else {
+ indices = (uint8_t*)pipe_buffer_map(pipe, info->index.resource,
+ PIPE_TRANSFER_READ, &transfer);
+ }
+
+ for (unsigned i = 0; i < draw_count; i++) {
+ unsigned offset = i * indirect->stride / 4;
+ unsigned start = data[offset + 2];
+ unsigned count = data[offset + 0];
+ unsigned instance_count = data[offset + 1];
+
+ if (!count || !instance_count)
+ continue;
+
+ /* Update the ranges of instances. */
+ new_info.start_instance = MIN2(new_info.start_instance,
+ data[offset + 4]);
+ new_info.instance_count = MAX2(new_info.instance_count,
+ instance_count);
+
+ /* Update the index range. */
+ unsigned min, max;
+ new_info.count = count; /* only used by get_minmax_index */
+ u_vbuf_get_minmax_index_mapped(&new_info,
+ indices +
+ new_info.index_size * start,
+ (int*)&min, (int*)&max);
+
+ new_info.min_index = MIN2(new_info.min_index, min);
+ new_info.max_index = MAX2(new_info.max_index, max);
+ }
+
+ if (transfer)
+ pipe_buffer_unmap(pipe, transfer);
+
+ if (!new_info.instance_count)
+ return;
+ } else {
+ /* Non-indexed multidraw.
+ *
+ * Keep the draw call indirect and compute minimums & maximums,
+ * which will determine the user buffer bounds to upload, but
+ * the driver will not look at these values because indirect != NULL.
+ *
+ * This efficiently processes the multidraw with the time complexity
+ * equal to 1 draw call.
+ */
+ new_info.start = ~0u;
+ new_info.count = 0;
+ new_info.start_instance = ~0u;
+ new_info.instance_count = 0;
+
+ for (unsigned i = 0; i < draw_count; i++) {
+ unsigned offset = i * indirect->stride / 4;
+
+ new_info.start = MIN2(new_info.start, data[offset + 2]);
+ new_info.count = MAX2(new_info.count, data[offset + 0]);
+
+ new_info.start_instance = MIN2(new_info.start_instance,
+ data[offset + 3]);
+ new_info.instance_count = MAX2(new_info.instance_count,
+ data[offset + 1]);
+ }
+
+ if (!new_info.count || !new_info.instance_count)
+ return;
+ }
}
if (new_info.index_size) {
/* See if anything needs to be done for per-vertex attribs. */
if (u_vbuf_need_minmax_index(mgr)) {
int max_index;
if (new_info.max_index != ~0u) {
min_index = new_info.min_index;
max_index = new_info.max_index;
@@ -1211,21 +1351,22 @@ void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct
pipe_draw_info *info)
assert(min_index <= max_index);
start_vertex = min_index + new_info.index_bias;
num_vertices = max_index + 1 - min_index;
/* Primitive restart doesn't work when unrolling indices.
* We would have to break this drawing operation into several ones. */
/* Use some heuristic to see if unrolling indices improves
* performance. */
- if (!new_info.primitive_restart &&
+ if (!info->indirect &&
+ !new_info.primitive_restart &&
num_vertices > new_info.count*2 &&
num_vertices - new_info.count > 32 &&
!u_vbuf_mapping_vertex_buffer_blocks(mgr)) {
unroll_indices = TRUE;
user_vb_mask &= ~(mgr->nonzero_stride_vb_mask &
mgr->ve->noninstance_vb_mask_any);
}
} else {
/* Nothing to do for per-vertex attribs. */
start_vertex = 0;
--
2.17.1
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
