* pipe_context
*/
-void r600_gfx_write_fence(struct r600_common_context *ctx, struct r600_resource *buf,
- uint64_t va, uint32_t old_value, uint32_t new_value)
+/**
+ * Write an EOP event.
+ *
+ * \param event EVENT_TYPE_*
+ * \param event_flags Optional cache flush flags (TC)
+ * \param data_sel 1 = fence, 3 = timestamp
+ * \param buf Buffer
+ * \param va GPU address
+ * \param old_value Previous fence value (for a bug workaround)
+ * \param new_value Fence value to write for this event.
+ */
+void r600_gfx_write_event_eop(struct r600_common_context *ctx,
+ unsigned event, unsigned event_flags,
+ unsigned data_sel,
+ struct r600_resource *buf, uint64_t va,
+ uint32_t old_fence, uint32_t new_fence)
{
struct radeon_winsys_cs *cs = ctx->gfx.cs;
+ unsigned op = EVENT_TYPE(event) |
+ EVENT_INDEX(5) |
+ event_flags;
if (ctx->chip_class == CIK) {
/* Two EOP events are required to make all engines go idle
* is written.
*/
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
- radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_BOTTOM_OF_PIPE_TS) |
- EVENT_INDEX(5));
+ radeon_emit(cs, op);
radeon_emit(cs, va);
- radeon_emit(cs, (va >> 32) | EOP_DATA_SEL(1));
- radeon_emit(cs, old_value); /* immediate data */
+ radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
+ radeon_emit(cs, old_fence); /* immediate data */
radeon_emit(cs, 0); /* unused */
}
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
- radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_BOTTOM_OF_PIPE_TS) |
- EVENT_INDEX(5));
+ radeon_emit(cs, op);
radeon_emit(cs, va);
- radeon_emit(cs, (va >> 32) | EOP_DATA_SEL(1));
- radeon_emit(cs, new_value); /* immediate data */
+ radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
+ radeon_emit(cs, new_fence); /* immediate data */
radeon_emit(cs, 0); /* unused */
- r600_emit_reloc(ctx, &ctx->gfx, buf, RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+ if (buf)
+ r600_emit_reloc(ctx, &ctx->gfx, buf, RADEON_USAGE_WRITE,
+ RADEON_PRIO_QUERY);
}
unsigned r600_gfx_write_fence_dwords(struct r600_common_screen *screen)
memset(&rctx->debug, 0, sizeof(rctx->debug));
}
+static void r600_set_device_reset_callback(struct pipe_context *ctx,
+ const struct pipe_device_reset_callback *cb)
+{
+ struct r600_common_context *rctx = (struct r600_common_context *)ctx;
+
+ if (cb)
+ rctx->device_reset_callback = *cb;
+ else
+ memset(&rctx->device_reset_callback, 0,
+ sizeof(rctx->device_reset_callback));
+}
+
+bool r600_check_device_reset(struct r600_common_context *rctx)
+{
+ enum pipe_reset_status status;
+
+ if (!rctx->device_reset_callback.reset)
+ return false;
+
+ if (!rctx->b.get_device_reset_status)
+ return false;
+
+ status = rctx->b.get_device_reset_status(&rctx->b);
+ if (status == PIPE_NO_RESET)
+ return false;
+
+ rctx->device_reset_callback.reset(rctx->device_reset_callback.data, status);
+ return true;
+}
+
bool r600_common_context_init(struct r600_common_context *rctx,
struct r600_common_screen *rscreen,
unsigned context_flags)
{
- slab_create(&rctx->pool_transfers,
- sizeof(struct r600_transfer), 64);
+ slab_create_child(&rctx->pool_transfers, &rscreen->pool_transfers);
rctx->screen = rscreen;
rctx->ws = rscreen->ws;
RADEON_GPU_RESET_COUNTER);
}
- LIST_INITHEAD(&rctx->texture_buffers);
+ rctx->b.set_device_reset_callback = r600_set_device_reset_callback;
r600_init_context_texture_functions(rctx);
r600_init_viewport_functions(rctx);
u_upload_destroy(rctx->uploader);
}
- slab_destroy(&rctx->pool_transfers);
+ slab_destroy_child(&rctx->pool_transfers);
if (rctx->allocator_zeroed_memory) {
u_suballocator_destroy(rctx->allocator_zeroed_memory);
rctx->ws->fence_reference(&rctx->last_sdma_fence, NULL);
}
-void r600_context_add_resource_size(struct pipe_context *ctx, struct pipe_resource *r)
-{
- struct r600_common_context *rctx = (struct r600_common_context *)ctx;
- struct r600_resource *rr = (struct r600_resource *)r;
-
- if (!r) {
- return;
- }
-
- /*
- * The idea is to compute a gross estimate of memory requirement of
- * each draw call. After each draw call, memory will be precisely
- * accounted. So the uncertainty is only on the current draw call.
- * In practice this gave very good estimate (+/- 10% of the target
- * memory limit).
- */
- rctx->vram += rr->vram_usage;
- rctx->gtt += rr->gart_usage;
-}
-
/*
* pipe_screen
*/
{ "notgsi", DBG_NO_TGSI, "Don't print the TGSI"},
{ "noasm", DBG_NO_ASM, "Don't print disassembled shaders"},
{ "preoptir", DBG_PREOPT_IR, "Print the LLVM IR before initial optimizations" },
+ { "checkir", DBG_CHECK_IR, "Enable additional sanity checks on shader IR" },
{ "testdma", DBG_TEST_DMA, "Invoke SDMA tests and exit." },
*subgroup_size = r600_wavefront_size(rscreen->family);
}
return sizeof(uint32_t);
+ case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
+ if (ret) {
+ uint64_t *max_variable_threads_per_block = ret;
+ if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
+ ir_type == PIPE_SHADER_IR_TGSI)
+ *max_variable_threads_per_block = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
+ else
+ *max_variable_threads_per_block = 0;
+ }
+ return sizeof(uint64_t);
}
fprintf(stderr, "unknown PIPE_COMPUTE_CAP %d\n", param);
rscreen->chip_class = rscreen->info.chip_class;
rscreen->debug_flags = debug_get_flags_option("R600_DEBUG", common_debug_options, 0);
+ slab_create_parent(&rscreen->pool_transfers, sizeof(struct r600_transfer), 64);
+
rscreen->force_aniso = MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
if (rscreen->force_aniso >= 0) {
printf("radeon: Forcing anisotropy filter to %ix\n",
pipe_mutex_destroy(rscreen->aux_context_lock);
rscreen->aux_context->destroy(rscreen->aux_context);
+ slab_destroy_parent(&rscreen->pool_transfers);
+
rscreen->ws->destroy(rscreen->ws);
FREE(rscreen);
}
}
}
+bool r600_extra_shader_checks(struct r600_common_screen *rscreen, unsigned processor)
+{
+ return (rscreen->debug_flags & DBG_CHECK_IR) ||
+ r600_can_dump_shader(rscreen, processor);
+}
+
void r600_screen_clear_buffer(struct r600_common_screen *rscreen, struct pipe_resource *dst,
uint64_t offset, uint64_t size, unsigned value,
enum r600_coherency coher)