summaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/drivers')
-rw-r--r--src/gallium/drivers/radeonsi/r600_buffer.c11
-rw-r--r--src/gallium/drivers/radeonsi/si_descriptors.c123
-rw-r--r--src/gallium/drivers/radeonsi/si_state.h1
3 files changed, 135 insertions, 0 deletions
diff --git a/src/gallium/drivers/radeonsi/r600_buffer.c b/src/gallium/drivers/radeonsi/r600_buffer.c
index 4c95130c352..560e7daf3cc 100644
--- a/src/gallium/drivers/radeonsi/r600_buffer.c
+++ b/src/gallium/drivers/radeonsi/r600_buffer.c
@@ -56,6 +56,17 @@ static void *r600_buffer_transfer_map(struct pipe_context *ctx,
struct r600_resource *rbuffer = r600_resource(resource);
uint8_t *data;
+ if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
+ !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
+ assert(usage & PIPE_TRANSFER_WRITE);
+
+ /* Check if mapping this buffer would cause waiting for the GPU. */
+ if (r600_rings_is_buffer_referenced(&rctx->b, rbuffer->cs_buf, RADEON_USAGE_READWRITE) ||
+ rctx->b.ws->buffer_is_busy(rbuffer->buf, RADEON_USAGE_READWRITE)) {
+ si_invalidate_buffer(&rctx->b.b, &rbuffer->b.b);
+ }
+ }
+
data = rctx->b.ws->buffer_map(rbuffer->cs_buf, rctx->b.rings.gfx.cs, usage);
if (!data) {
return NULL;
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index c591352bf38..62c354fffcf 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -523,6 +523,129 @@ static void si_set_streamout_targets(struct pipe_context *ctx,
si_update_descriptors(rctx, &buffers->desc);
}
+static void si_desc_reset_buffer_offset(struct pipe_context *ctx,
+ uint32_t *desc, uint64_t old_buf_va,
+ struct pipe_resource *new_buf)
+{
+ /* Retrieve the buffer offset from the descriptor. */
+ uint64_t old_desc_va =
+ desc[0] | ((uint64_t)G_008F04_BASE_ADDRESS_HI(desc[1]) << 32);
+
+ assert(old_buf_va <= old_desc_va);
+ uint64_t offset_within_buffer = old_desc_va - old_buf_va;
+
+ /* Update the descriptor. */
+ uint64_t va = r600_resource_va(ctx->screen, new_buf) + offset_within_buffer;
+
+ desc[0] = va;
+ desc[1] = (desc[1] & C_008F04_BASE_ADDRESS_HI) |
+ S_008F04_BASE_ADDRESS_HI(va >> 32);
+}
+
+/* BUFFER DISCARD/INVALIDATION */
+
+/* Reallocate a buffer a update all resource bindings where the buffer is
+ * bound.
+ *
+ * This is used to avoid CPU-GPU synchronizations, because it makes the buffer
+ * idle by discarding its contents. Apps usually tell us when to do this using
+ * map_buffer flags, for example.
+ */
+void si_invalidate_buffer(struct pipe_context *ctx, struct pipe_resource *buf)
+{
+ struct r600_context *rctx = (struct r600_context*)ctx;
+ struct r600_resource *rbuffer = r600_resource(buf);
+ unsigned i, shader, alignment = rbuffer->buf->alignment;
+ uint64_t old_va = r600_resource_va(ctx->screen, buf);
+
+ /* Discard the buffer. */
+ pb_reference(&rbuffer->buf, NULL);
+
+ /* Create a new one in the same pipe_resource. */
+ r600_init_resource(&rctx->screen->b, rbuffer, rbuffer->b.b.width0, alignment,
+ TRUE, rbuffer->b.b.usage);
+
+ /* We changed the buffer, now we need to bind it where the old one
+ * was bound. This consists of 2 things:
+ * 1) Updating the resource descriptor and dirtying it.
+ * 2) Adding a relocation to the CS, so that it's usable.
+ */
+
+ /* Vertex buffers. */
+ /* Nothing to do. Vertex buffer bindings are updated before every draw call. */
+
+ /* Streamout buffers. */
+ for (i = 0; i < rctx->streamout_buffers.num_buffers; i++) {
+ if (rctx->streamout_buffers.buffers[i] == buf) {
+ /* Update the descriptor. */
+ si_desc_reset_buffer_offset(ctx, rctx->streamout_buffers.desc_data[i],
+ old_va, buf);
+
+ r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
+ (struct r600_resource*)buf,
+ rctx->streamout_buffers.shader_usage);
+ rctx->streamout_buffers.desc.dirty_mask |= 1 << i;
+ si_update_descriptors(rctx, &rctx->streamout_buffers.desc);
+
+ /* Update the streamout state. */
+ if (rctx->b.streamout.begin_emitted) {
+ r600_emit_streamout_end(&rctx->b);
+ }
+ rctx->b.streamout.append_bitmask = rctx->b.streamout.enabled_mask;
+ r600_streamout_buffers_dirty(&rctx->b);
+ }
+ }
+
+ /* Constant buffers. */
+ for (shader = 0; shader < SI_NUM_SHADERS; shader++) {
+ struct si_buffer_resources *buffers = &rctx->const_buffers[shader];
+ bool found = false;
+ uint32_t mask = buffers->desc.enabled_mask;
+
+ while (mask) {
+ unsigned i = u_bit_scan(&mask);
+ if (buffers->buffers[i] == buf) {
+ si_desc_reset_buffer_offset(ctx, buffers->desc_data[i],
+ old_va, buf);
+
+ r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
+ rbuffer, buffers->shader_usage);
+
+ buffers->desc.dirty_mask |= 1 << i;
+ found = true;
+ }
+ }
+ if (found) {
+ si_update_descriptors(rctx, &buffers->desc);
+ }
+ }
+
+ /* Texture buffers. */
+ for (shader = 0; shader < SI_NUM_SHADERS; shader++) {
+ struct si_sampler_views *views = &rctx->samplers[shader].views;
+ bool found = false;
+ uint32_t mask = views->desc.enabled_mask;
+
+ while (mask) {
+ unsigned i = u_bit_scan(&mask);
+ if (views->views[i]->texture == buf) {
+ /* This updates the sampler view directly. */
+ si_desc_reset_buffer_offset(ctx, views->desc_data[i],
+ old_va, buf);
+
+ r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
+ rbuffer, RADEON_USAGE_READ);
+
+ views->desc.dirty_mask |= 1 << i;
+ found = true;
+ }
+ }
+ if (found) {
+ si_update_descriptors(rctx, &views->desc);
+ }
+ }
+}
+
/* CP DMA */
/* The max number of bytes to copy per packet. */
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 6774e57e21f..7e2cb81eaaf 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -200,6 +200,7 @@ void si_all_descriptors_begin_new_cs(struct r600_context *rctx);
void si_copy_buffer(struct r600_context *rctx,
struct pipe_resource *dst, struct pipe_resource *src,
uint64_t dst_offset, uint64_t src_offset, unsigned size);
+void si_invalidate_buffer(struct pipe_context *ctx, struct pipe_resource *buf);
/* si_state.c */
struct si_pipe_shader_selector;