Mesa (master): iris: Use copy_region and staging resources to avoid transfer stalls
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Fri Mar 8 21:30:45 UTC 2019
Module: Mesa
Branch: master
Commit: 9d1334d2a0f983e175ffe371bee33f4ce048f910
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=9d1334d2a0f983e175ffe371bee33f4ce048f910
Author: Kenneth Graunke <kenneth at whitecape.org>
Date: Fri Dec 21 03:04:18 2018 -0800
iris: Use copy_region and staging resources to avoid transfer stalls
This is similar to intel_miptree_map_blit and intel_buffer_object.c's
temporary blits in i965.
Improves performance of DiRT Rally by 20-25% by eliminating stalls.
Breaks piglit's spec/arb_shader_image_load_store/host-mem-barrier,
by using the GPU to do uploads, exposing a st/mesa issue where it
doesn't give us memory_barrier() calls. This is a pre-existing issue
and will be fixed by a later patch (currently out for review).
---
src/gallium/drivers/iris/iris_resource.c | 167 ++++++++++++++++++++++++++++---
src/gallium/drivers/iris/iris_resource.h | 5 +
src/gallium/drivers/iris/iris_screen.c | 2 +
src/gallium/drivers/iris/iris_screen.h | 1 +
4 files changed, 161 insertions(+), 14 deletions(-)
diff --git a/src/gallium/drivers/iris/iris_resource.c b/src/gallium/drivers/iris/iris_resource.c
index 03ce35462fa..5040637ee77 100644
--- a/src/gallium/drivers/iris/iris_resource.c
+++ b/src/gallium/drivers/iris/iris_resource.c
@@ -794,6 +794,102 @@ iris_resource_get_handle(struct pipe_screen *pscreen,
}
static void
+iris_unmap_copy_region(struct iris_transfer *map)
+{
+ struct pipe_transfer *xfer = &map->base;
+ struct pipe_box *dst_box = &xfer->box;
+ struct pipe_box src_box = (struct pipe_box) {
+ .x = xfer->resource->target == PIPE_BUFFER ?
+ xfer->box.x % IRIS_MAP_BUFFER_ALIGNMENT : 0,
+ .width = dst_box->width,
+ .height = dst_box->height,
+ .depth = dst_box->depth,
+ };
+
+ if (xfer->usage & PIPE_TRANSFER_WRITE) {
+ iris_copy_region(map->blorp, map->batch, xfer->resource, xfer->level,
+ dst_box->x, dst_box->y, dst_box->z, map->staging, 0,
+ &src_box);
+ }
+
+ iris_resource_destroy(map->staging->screen, map->staging);
+
+ map->ptr = NULL;
+}
+
+static void
+iris_map_copy_region(struct iris_transfer *map)
+{
+ struct pipe_screen *pscreen = &map->batch->screen->base;
+ struct pipe_transfer *xfer = &map->base;
+ struct pipe_box *box = &xfer->box;
+ struct iris_resource *res = (void *) xfer->resource;
+
+ unsigned extra = xfer->resource->target == PIPE_BUFFER ?
+ box->x % IRIS_MAP_BUFFER_ALIGNMENT : 0;
+
+ struct pipe_resource templ = (struct pipe_resource) {
+ .usage = PIPE_USAGE_STAGING,
+ .width0 = box->width + extra,
+ .height0 = box->height,
+ .depth0 = 1,
+ .nr_samples = xfer->resource->nr_samples,
+ .nr_storage_samples = xfer->resource->nr_storage_samples,
+ .array_size = box->depth,
+ };
+
+ if (xfer->resource->target == PIPE_BUFFER)
+ templ.target = PIPE_BUFFER;
+ else if (templ.array_size > 1)
+ templ.target = PIPE_TEXTURE_2D_ARRAY;
+ else
+ templ.target = PIPE_TEXTURE_2D;
+
+ /* Depth, stencil, and ASTC can't be linear surfaces, so we can't use
+ * xfer->resource->format directly. Pick a bpb compatible format so
+ * resource creation will succeed; blorp_copy will override it anyway.
+ */
+ switch (util_format_get_blocksizebits(res->internal_format)) {
+ case 8: templ.format = PIPE_FORMAT_R8_UINT; break;
+ case 16: templ.format = PIPE_FORMAT_R8G8_UINT; break;
+ case 24: templ.format = PIPE_FORMAT_R8G8B8_UINT; break;
+ case 32: templ.format = PIPE_FORMAT_R8G8B8A8_UINT; break;
+ case 48: templ.format = PIPE_FORMAT_R16G16B16_UINT; break;
+ case 64: templ.format = PIPE_FORMAT_R16G16B16A16_UINT; break;
+ case 96: templ.format = PIPE_FORMAT_R32G32B32_UINT; break;
+ case 128: templ.format = PIPE_FORMAT_R32G32B32A32_UINT; break;
+ default: unreachable("Invalid bpb");
+ }
+
+ map->staging = iris_resource_create(pscreen, &templ);
+ assert(map->staging);
+
+ if (templ.target != PIPE_BUFFER) {
+ struct isl_surf *surf = &((struct iris_resource *) map->staging)->surf;
+ xfer->stride = isl_surf_get_row_pitch_B(surf);
+ xfer->layer_stride = isl_surf_get_array_pitch(surf);
+ }
+
+ if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
+ iris_copy_region(map->blorp, map->batch, map->staging, 0, extra, 0, 0,
+ xfer->resource, xfer->level, box);
+ /* Ensure writes to the staging BO land before we map it below. */
+ iris_emit_pipe_control_flush(map->batch,
+ PIPE_CONTROL_RENDER_TARGET_FLUSH |
+ PIPE_CONTROL_CS_STALL);
+ }
+
+ struct iris_bo *staging_bo = iris_resource_bo(map->staging);
+
+ if (iris_batch_references(map->batch, staging_bo))
+ iris_batch_flush(map->batch);
+
+ map->ptr = iris_bo_map(map->dbg, staging_bo, xfer->usage) + extra;
+
+ map->unmap = iris_unmap_copy_region;
+}
+
+static void
get_image_offset_el(struct isl_surf *surf, unsigned level, unsigned z,
unsigned *out_x0_el, unsigned *out_y0_el)
{
@@ -1099,9 +1195,7 @@ iris_transfer_map(struct pipe_context *ctx,
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
usage |= PIPE_TRANSFER_DISCARD_RANGE;
- if (surf->tiling != ISL_TILING_LINEAR &&
- (usage & PIPE_TRANSFER_MAP_DIRECTLY))
- return NULL;
+ bool map_would_stall = false;
if (resource->target != PIPE_BUFFER) {
iris_resource_access_raw(ice, &ice->batches[IRIS_BATCH_RENDER], res,
@@ -1110,13 +1204,18 @@ iris_transfer_map(struct pipe_context *ctx,
}
if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
- for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
- if (iris_batch_references(&ice->batches[i], res->bo))
- iris_batch_flush(&ice->batches[i]);
- }
+ map_would_stall = iris_bo_busy(res->bo);
+
+ for (int i = 0; i < IRIS_BATCH_COUNT; i++)
+ map_would_stall |= iris_batch_references(&ice->batches[i], res->bo);
+
+ if (map_would_stall && (usage & PIPE_TRANSFER_DONTBLOCK) &&
+ (usage & PIPE_TRANSFER_MAP_DIRECTLY))
+ return NULL;
}
- if ((usage & PIPE_TRANSFER_DONTBLOCK) && iris_bo_busy(res->bo))
+ if (surf->tiling != ISL_TILING_LINEAR &&
+ (usage & PIPE_TRANSFER_MAP_DIRECTLY))
return NULL;
struct iris_transfer *map = slab_alloc(&ice->transfer_pool);
@@ -1141,13 +1240,53 @@ iris_transfer_map(struct pipe_context *ctx,
PIPE_TRANSFER_COHERENT |
PIPE_TRANSFER_DISCARD_RANGE);
- if (surf->tiling == ISL_TILING_W) {
- // XXX: just teach iris_map_tiled_memcpy about W tiling...
- iris_map_s8(map);
- } else if (surf->tiling != ISL_TILING_LINEAR) {
- iris_map_tiled_memcpy(map);
+ /* Avoid using GPU copies for persistent/coherent buffers, as the idea
+ * there is to access them simultaneously on the CPU & GPU. This also
+ * avoids trying to use GPU copies for our u_upload_mgr buffers which
+ * contain state we're constructing for a GPU draw call, which would
+ * kill us with infinite stack recursion.
+ */
+ bool no_gpu = usage & (PIPE_TRANSFER_PERSISTENT |
+ PIPE_TRANSFER_COHERENT |
+ PIPE_TRANSFER_MAP_DIRECTLY);
+
+ /* GPU copies are not useful for buffer reads. Instead of stalling to
+ * read from the original buffer, we'd simply copy it to a temporary...
+ * then stall (a bit longer) to read from that buffer.
+ *
+ * Images are less clear-cut. Color resolves are destructive, removing
+ * the underlying compression, so we'd rather blit the data to a linear
+ * temporary and map that, to avoid the resolve. (It might be better to
+ * a tiled temporary and use the tiled_memcpy paths...)
+ */
+ if (!(usage & PIPE_TRANSFER_DISCARD_RANGE) &&
+ res->aux.usage != ISL_AUX_USAGE_CCS_E &&
+ res->aux.usage != ISL_AUX_USAGE_CCS_D) {
+ no_gpu = true;
+ }
+
+ if (map_would_stall && !no_gpu) {
+ /* If we need a synchronous mapping and the resource is busy,
+ * we copy to/from a linear temporary buffer using the GPU.
+ */
+ map->batch = &ice->batches[IRIS_BATCH_RENDER];
+ map->blorp = &ice->blorp;
+ iris_map_copy_region(map);
} else {
- iris_map_direct(map);
+ /* Otherwise we're free to map on the CPU. Flush if needed. */
+ for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
+ if (iris_batch_references(&ice->batches[i], res->bo))
+ iris_batch_flush(&ice->batches[i]);
+ }
+
+ if (surf->tiling == ISL_TILING_W) {
+ /* TODO: Teach iris_map_tiled_memcpy about W-tiling... */
+ iris_map_s8(map);
+ } else if (surf->tiling != ISL_TILING_LINEAR) {
+ iris_map_tiled_memcpy(map);
+ } else {
+ iris_map_direct(map);
+ }
}
return map->ptr;
diff --git a/src/gallium/drivers/iris/iris_resource.h b/src/gallium/drivers/iris/iris_resource.h
index 8b5d176191e..d9cb50de8b6 100644
--- a/src/gallium/drivers/iris/iris_resource.h
+++ b/src/gallium/drivers/iris/iris_resource.h
@@ -180,6 +180,11 @@ struct iris_transfer {
void *buffer;
void *ptr;
+ /** A linear staging resource for GPU-based copy_region transfers. */
+ struct pipe_resource *staging;
+ struct blorp_context *blorp;
+ struct iris_batch *batch;
+
void (*unmap)(struct iris_transfer *);
};
diff --git a/src/gallium/drivers/iris/iris_screen.c b/src/gallium/drivers/iris/iris_screen.c
index 909fb876fbe..f18f986547e 100644
--- a/src/gallium/drivers/iris/iris_screen.c
+++ b/src/gallium/drivers/iris/iris_screen.c
@@ -201,6 +201,8 @@ iris_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
/* 3DSTATE_CONSTANT_XS requires the start of UBOs to be 32B aligned */
return 32;
+ case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT:
+ return IRIS_MAP_BUFFER_ALIGNMENT;
case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
/* Choose a cacheline (64 bytes) so that we can safely have the CPU and
* GPU writing the same SSBO on non-coherent systems (Atom CPUs). With
diff --git a/src/gallium/drivers/iris/iris_screen.h b/src/gallium/drivers/iris/iris_screen.h
index 77fbd906b73..386a5ba6343 100644
--- a/src/gallium/drivers/iris/iris_screen.h
+++ b/src/gallium/drivers/iris/iris_screen.h
@@ -38,6 +38,7 @@ struct iris_bo;
#define IRIS_MAX_TEXTURE_SAMPLERS 32
#define IRIS_MAX_SOL_BUFFERS 4
+#define IRIS_MAP_BUFFER_ALIGNMENT 64
struct iris_screen {
struct pipe_screen base;
More information about the mesa-commit
mailing list