r600g: always map uninitialized buffer range as unsynchronized

Any driver can implement this simple and efficient optimization.
Team Fortress 2 hits it always. The DISCARD_RANGE codepath is not even used
with TF2 anymore, so we avoid a ton of useless buffer copies.

Tested-by: Andreas Boll <andreas.boll.dev@gmail.com>
NOTE: This is a candidate for the 9.1 branch.

diff --git a/src/gallium/drivers/r600/evergreen_hw_context.c b/src/gallium/drivers/r600/evergreen_hw_context.c
index b0f64b4ee53e20de9d34430bacb3d43a4ae887a0..f81d7f3216ba4d912b90c9d24d04607bc4cb07da 100644 (file)
--- a/src/gallium/drivers/r600/evergreen_hw_context.c
+++ b/src/gallium/drivers/r600/evergreen_hw_context.c
@@ -241,4 +241,7 @@ void evergreen_dma_copy(struct r600_context *rctx,
                src_offset += csize << shift;
                size -= csize;
        }
+
+       util_range_add(&rdst->valid_buffer_range, dst_offset,
+                      dst_offset + size);
 }

diff --git a/src/gallium/drivers/r600/evergreen_state.c b/src/gallium/drivers/r600/evergreen_state.c
index 41dd70e5ab1f73f5cee2427e7e97b7e7d3571a27..97f91df374ca21d72ab2399f271814b0e408c286 100644 (file)
--- a/src/gallium/drivers/r600/evergreen_state.c
+++ b/src/gallium/drivers/r600/evergreen_state.c
@@ -1377,6 +1377,10 @@ void evergreen_init_color_surface_rat(struct r600_context *rctx,
         * elements. */
        surf->cb_color_dim = pipe_buffer->width0;
 
+       /* Set the buffer range the GPU will have access to: */
+       util_range_add(&r600_resource(pipe_buffer)->valid_buffer_range,
+                      0, pipe_buffer->width0);
+
        surf->cb_color_cmask = surf->cb_color_base;
        surf->cb_color_cmask_slice = 0;
        surf->cb_color_fmask = surf->cb_color_base;

diff --git a/src/gallium/drivers/r600/r600.h b/src/gallium/drivers/r600/r600.h
index d018ebb0298b73300e6626086ab3c29601fccda3..15196f7579fad74902cd62df11f187603cd120e3 100644 (file)
--- a/src/gallium/drivers/r600/r600.h
+++ b/src/gallium/drivers/r600/r600.h
@@ -28,6 +28,7 @@
 
 #include "../../winsys/radeon/drm/radeon_winsys.h"
 #include "util/u_double_list.h"
+#include "util/u_range.h"
 #include "util/u_transfer.h"
 
 #define R600_ERR(fmt, args...) \
@@ -50,6 +51,16 @@ struct r600_resource {
 
        /* Resource state. */
        unsigned                        domains;
+
+       /* The buffer range which is initialized (with a write transfer,
+        * streamout, DMA, or as a random access target). The rest of
+        * the buffer is considered invalid and can be mapped unsynchronized.
+        *
+        * This allows unsychronized mapping of a buffer range which hasn't
+        * been used yet. It's for applications which forget to use
+        * the unsynchronized map flag and expect the driver to figure it out.
+         */
+       struct util_range               valid_buffer_range;
 };
 
 #define R600_BLOCK_MAX_BO              32

diff --git a/src/gallium/drivers/r600/r600_buffer.c b/src/gallium/drivers/r600/r600_buffer.c
index 3267c37cc6d50f49cf42fb9e186fcf5bc586eae3..7574cd6c88978ae2171ba52990bbe60cf410d5bf 100644 (file)
--- a/src/gallium/drivers/r600/r600_buffer.c
+++ b/src/gallium/drivers/r600/r600_buffer.c
@@ -34,6 +34,7 @@ static void r600_buffer_destroy(struct pipe_screen *screen,
 {
        struct r600_resource *rbuffer = r600_resource(buf);
 
+       util_range_destroy(&rbuffer->valid_buffer_range);
        pb_reference(&rbuffer->buf, NULL);
        FREE(rbuffer);
 }
@@ -98,6 +99,14 @@ static void *r600_buffer_transfer_map(struct pipe_context *ctx,
 
        assert(box->x + box->width <= resource->width0);
 
+       /* See if the buffer range being mapped has never been initialized,
+        * in which case it can be mapped unsynchronized. */
+       if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
+           usage & PIPE_TRANSFER_WRITE &&
+           !util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
+               usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
+       }
+
        if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
            !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
                assert(usage & PIPE_TRANSFER_WRITE);
@@ -180,6 +189,7 @@ static void r600_buffer_transfer_unmap(struct pipe_context *pipe,
 {
        struct r600_context *rctx = (struct r600_context*)pipe;
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
+       struct r600_resource *rbuffer = r600_resource(transfer->resource);
 
        if (rtransfer->staging) {
                struct pipe_resource *dst, *src;
@@ -205,6 +215,11 @@ static void r600_buffer_transfer_unmap(struct pipe_context *pipe,
                }
                pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
        }
+
+       if (transfer->usage & PIPE_TRANSFER_WRITE) {
+               util_range_add(&rbuffer->valid_buffer_range, transfer->box.x,
+                              transfer->box.x + transfer->box.width);
+       }
        util_slab_free(&rctx->pool_transfers, transfer);
 }
 
@@ -261,6 +276,7 @@ bool r600_init_resource(struct r600_screen *rscreen,
 
        res->cs_buf = rscreen->ws->buffer_get_cs_handle(res->buf);
        res->domains = domains;
+       util_range_set_empty(&res->valid_buffer_range);
        return true;
 }
 
@@ -277,6 +293,7 @@ struct pipe_resource *r600_buffer_create(struct pipe_screen *screen,
        pipe_reference_init(&rbuffer->b.b.reference, 1);
        rbuffer->b.b.screen = screen;
        rbuffer->b.vtbl = &r600_buffer_vtbl;
+       util_range_init(&rbuffer->valid_buffer_range);
 
        if (!r600_init_resource(rscreen, rbuffer, templ->width0, alignment, TRUE, templ->usage)) {
                FREE(rbuffer);

diff --git a/src/gallium/drivers/r600/r600_hw_context.c b/src/gallium/drivers/r600/r600_hw_context.c
index 0ff34353c4a02933fbcb892ddc1ef50522d5d060..c2f3aab781343f6d80a6e2319e3ef241aef156dd 100644 (file)
--- a/src/gallium/drivers/r600/r600_hw_context.c
+++ b/src/gallium/drivers/r600/r600_hw_context.c
@@ -1148,6 +1148,9 @@ void r600_cp_dma_copy_buffer(struct r600_context *rctx,
 
        /* Invalidate the read caches. */
        rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
+
+       util_range_add(&r600_resource(dst)->valid_buffer_range, dst_offset,
+                      dst_offset + size);
 }
 
 void r600_need_dma_space(struct r600_context *ctx, unsigned num_dw)
@@ -1194,4 +1197,7 @@ void r600_dma_copy(struct r600_context *rctx,
                src_offset += csize << shift;
                size -= csize;
        }
+
+       util_range_add(&rdst->valid_buffer_range, dst_offset,
+                      dst_offset + size);
 }

diff --git a/src/gallium/drivers/r600/r600_state_common.c b/src/gallium/drivers/r600/r600_state_common.c
index b067e50746ce27d50b995379541a9a72a840433a..22ac8466e4c6ba9b9887ebb9d68d43828d6eb2fc 100644 (file)
--- a/src/gallium/drivers/r600/r600_state_common.c
+++ b/src/gallium/drivers/r600/r600_state_common.c
@@ -977,6 +977,7 @@ r600_create_so_target(struct pipe_context *ctx,
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
        struct r600_so_target *t;
+       struct r600_resource *rbuffer = (struct r600_resource*)buffer;
 
        t = CALLOC_STRUCT(r600_so_target);
        if (!t) {
@@ -996,6 +997,9 @@ r600_create_so_target(struct pipe_context *ctx,
        pipe_resource_reference(&t->b.buffer, buffer);
        t->b.buffer_offset = buffer_offset;
        t->b.buffer_size = buffer_size;
+
+       util_range_add(&rbuffer->valid_buffer_range, buffer_offset,
+                      buffer_offset + buffer_size);
        return &t->b;
 }