broadcom/vc4: Simplify the relocation handling for index buffers.

Originally there was CL code for handling various relocations back when I
had relocs for the TSDA/TA buffers.  Now that the kernel handles those
entirely on its own, I can inline that code into the one place using it.

diff --git a/src/gallium/drivers/vc4/vc4_cl.h b/src/gallium/drivers/vc4/vc4_cl.h
index 8df9dbfe65a8d8abb1d06f879392374d5e54a282..39d1d347bbad22164a886de97d1fc254b44e7a6b 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_cl.h
+++ b/src/gallium/drivers/vc4/vc4_cl.h
@@ -159,21 +159,6 @@ cl_aligned_f(struct vc4_cl_out **cl, float f)
         cl_aligned_u32(cl, fui(f));
 }
 
-static inline void
-cl_start_reloc(struct vc4_cl *cl, struct vc4_cl_out **out, uint32_t n)
-{
-        assert(n == 1 || n == 2);
-        assert(cl->reloc_count == 0);
-#ifndef NDEBUG
-        cl->reloc_count = n;
-#endif
-
-        cl_u8(out, VC4_PACKET_GEM_HANDLES);
-        cl->reloc_next = *out;
-        cl_u32(out, 0); /* Space where hindex will be written. */
-        cl_u32(out, 0); /* Space where hindex will be written. */
-}
-
 static inline struct vc4_cl_out *
 cl_start_shader_reloc(struct vc4_cl *cl, uint32_t n)
 {

diff --git a/src/gallium/drivers/vc4/vc4_draw.c b/src/gallium/drivers/vc4/vc4_draw.c
index 6fd25b45b989f218e0762d2a4e79ac4e3b312b77..fd80f67b6a17f2d46acddde34b787b440c4ec439 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_draw.c
+++ b/src/gallium/drivers/vc4/vc4_draw.c
@@ -377,7 +377,21 @@ vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                 struct vc4_resource *rsc = vc4_resource(prsc);
 
                 struct vc4_cl_out *bcl = cl_start(&job->bcl);
-                cl_start_reloc(&job->bcl, &bcl, 1);
+
+                /* The original design for the VC4 kernel UABI had multiple
+                 * packets that used relocations in the BCL (some of which
+                 * needed two BOs), but later modifications eliminated all but
+                 * this one usage.  We have an arbitrary 32-bit offset value,
+                 * and need to also supply an arbitrary 32-bit index buffer
+                 * GEM handle, so we have this fake packet we emit in our BCL
+                 * to be validated, which the kernel uses at validation time
+                 * to perform the relocation in the IB packet (without
+                 * emitting to the actual HW).
+                 */
+                cl_u8(&bcl, VC4_PACKET_GEM_HANDLES);
+                cl_u32(&bcl, vc4_gem_hindex(job, rsc->bo));
+                cl_u32(&bcl, 0);
+
                 cl_u8(&bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
                 cl_u8(&bcl,
                       info->mode |
@@ -385,8 +399,9 @@ vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
                        VC4_INDEX_BUFFER_U16:
                        VC4_INDEX_BUFFER_U8));
                 cl_u32(&bcl, info->count);
-                cl_reloc(job, &job->bcl, &bcl, rsc->bo, offset);
+                cl_u32(&bcl, offset);
                 cl_u32(&bcl, vc4->max_index);
+
                 cl_end(&job->bcl, bcl);
                 job->draw_calls_queued++;