radeonsi: remove Constant Engine support

[mesa.git] / src / gallium / drivers / radeon / r600_pipe_common.c

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 960b59c5b29c169eab1e6e5f73c604f7c3411e56..37c12dea3740d781a4959aea248b4d654e5ee3e0 100644 (file)
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -103,7 +103,7 @@ 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)
+                             uint32_t new_fence, unsigned query_type)
 {
        struct radeon_winsys_cs *cs = ctx->gfx.cs;
        unsigned op = EVENT_TYPE(event) |
@@ -111,6 +111,29 @@ void r600_gfx_write_event_eop(struct r600_common_context *ctx,
                      event_flags;
 
        if (ctx->chip_class >= GFX9) {
+               /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
+                * counters) must immediately precede every timestamp event to
+                * prevent a GPU hang on GFX9.
+                *
+                * Occlusion queries don't need to do it here, because they
+                * always do ZPASS_DONE before the timestamp.
+                */
+               if (ctx->chip_class == GFX9 &&
+                   query_type != PIPE_QUERY_OCCLUSION_COUNTER &&
+                   query_type != PIPE_QUERY_OCCLUSION_PREDICATE) {
+                       struct r600_resource *scratch = ctx->eop_bug_scratch;
+
+                       assert(16 * ctx->screen->info.num_render_backends <=
+                              scratch->b.b.width0);
+                       radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+                       radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
+                       radeon_emit(cs, scratch->gpu_address);
+                       radeon_emit(cs, scratch->gpu_address >> 32);
+
+                       radeon_add_to_buffer_list(ctx, &ctx->gfx, scratch,
+                                                 RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+               }
+
                radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, 6, 0));
                radeon_emit(cs, op);
                radeon_emit(cs, EOP_DATA_SEL(data_sel));
@@ -122,6 +145,9 @@ void r600_gfx_write_event_eop(struct r600_common_context *ctx,
        } else {
                if (ctx->chip_class == CIK ||
                    ctx->chip_class == VI) {
+                       struct r600_resource *scratch = ctx->eop_bug_scratch;
+                       uint64_t va = scratch->gpu_address;
+
                        /* Two EOP events are required to make all engines go idle
                         * (and optional cache flushes executed) before the timestamp
                         * is written.
@@ -130,8 +156,11 @@ void r600_gfx_write_event_eop(struct r600_common_context *ctx,
                        radeon_emit(cs, op);
                        radeon_emit(cs, va);
                        radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
-                       radeon_emit(cs, old_fence); /* immediate data */
+                       radeon_emit(cs, 0); /* immediate data */
                        radeon_emit(cs, 0); /* unused */
+
+                       radeon_add_to_buffer_list(ctx, &ctx->gfx, scratch,
+                                                 RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
                }
 
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
@@ -655,6 +684,16 @@ bool r600_common_context_init(struct r600_common_context *rctx,
        r600_query_init(rctx);
        cayman_init_msaa(&rctx->b);
 
+       if (rctx->chip_class == CIK ||
+           rctx->chip_class == VI ||
+           rctx->chip_class == GFX9) {
+               rctx->eop_bug_scratch = (struct r600_resource*)
+                       pipe_buffer_create(&rscreen->b, 0, PIPE_USAGE_DEFAULT,
+                                          16 * rscreen->info.num_render_backends);
+               if (!rctx->eop_bug_scratch)
+                       return false;
+       }
+
        rctx->allocator_zeroed_memory =
                u_suballocator_create(&rctx->b, rscreen->info.gart_page_size,
                                      0, PIPE_USAGE_DEFAULT, 0, true);
@@ -724,6 +763,7 @@ void r600_common_context_cleanup(struct r600_common_context *rctx)
        }
        rctx->ws->fence_reference(&rctx->last_gfx_fence, NULL);
        rctx->ws->fence_reference(&rctx->last_sdma_fence, NULL);
+       r600_resource_reference(&rctx->eop_bug_scratch, NULL);
 }
 
 /*
@@ -775,8 +815,6 @@ static const struct debug_named_value common_debug_options[] = {
        { "norbplus", DBG_NO_RB_PLUS, "Disable RB+." },
        { "sisched", DBG_SI_SCHED, "Enable LLVM SI Machine Instruction Scheduler." },
        { "mono", DBG_MONOLITHIC_SHADERS, "Use old-style monolithic shaders compiled on demand" },
-       { "ce", DBG_CE, "Force enable the constant engine" },
-       { "noce", DBG_NO_CE, "Disable the constant engine"},
        { "unsafemath", DBG_UNSAFE_MATH, "Enable unsafe math shader optimizations" },
        { "nodccfb", DBG_NO_DCC_FB, "Disable separate DCC on the main framebuffer" },