v3d: Use combined input/output segments.

The HW apparently has some issues (or at least a much more complicated VCM
calculation) with non-combined segments, and the closed source driver also
uses combined I/O.  Until I get the last CTS failure resolved (which does
look plausibly like some VPM stomping), let's use combined I/O too.

4 files changed, 41 insertions, 5 deletions

diff --git a/src/broadcom/compiler/qpu_schedule.c b/src/broadcom/compiler/qpu_schedule.c
index 4f3b621fd29..944059c9778 100644
--- a/src/broadcom/compiler/qpu_schedule.c
+++ b/src/broadcom/compiler/qpu_schedule.c
@@ -280,6 +280,11 @@ calculate_deps(struct schedule_state *state, struct schedule_node *n)
         const struct v3d_device_info *devinfo = state->devinfo;
         struct qinst *qinst = n->inst;
         struct v3d_qpu_instr *inst = &qinst->qpu;
+        /* If the input and output segments are shared, then all VPM reads to
+         * a location need to happen before all writes.  We handle this by
+         * serializing all VPM operations for now.
+         */
+        bool separate_vpm_segment = false;
 
         if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
                 if (inst->branch.cond != V3D_QPU_BRANCH_COND_ALWAYS)
@@ -321,6 +326,14 @@ calculate_deps(struct schedule_state *state, struct schedule_node *n)
                 add_write_dep(state, &state->last_vpm, n);
                 break;
 
+        case V3D_QPU_A_LDVPMV_IN:
+        case V3D_QPU_A_LDVPMD_IN:
+        case V3D_QPU_A_LDVPMG_IN:
+        case V3D_QPU_A_LDVPMP:
+                if (!separate_vpm_segment)
+                        add_write_dep(state, &state->last_vpm, n);
+                break;
+
         case V3D_QPU_A_VPMWT:
                 add_read_dep(state, state->last_vpm, n);
                 break;
@@ -414,9 +427,16 @@ calculate_deps(struct schedule_state *state, struct schedule_node *n)
         if (inst->sig.ldtlb | inst->sig.ldtlbu)
                 add_read_dep(state, state->last_tlb, n);
 
-        if (inst->sig.ldvpm)
+        if (inst->sig.ldvpm) {
                 add_write_dep(state, &state->last_vpm_read, n);
 
+                /* At least for now, we're doing shared I/O segments, so queue
+                 * all writes after all reads.
+                 */
+                if (!separate_vpm_segment)
+                        add_write_dep(state, &state->last_vpm, n);
+        }
+
         /* inst->sig.ldunif or sideband uniform read */
         if (qinst->uniform != ~0)
                 add_write_dep(state, &state->last_unif, n);
diff --git a/src/broadcom/compiler/v3d_compiler.h b/src/broadcom/compiler/v3d_compiler.h
index 070e6a3aa59..1e899393441 100644
--- a/src/broadcom/compiler/v3d_compiler.h
+++ b/src/broadcom/compiler/v3d_compiler.h
@@ -649,6 +649,11 @@ struct v3d_vs_prog_data {
         /* Total number of components written, for the shader state record. */
         uint32_t vpm_output_size;
 
+        /* Set if there should be separate VPM segments for input and output.
+         * If unset, vpm_input_size will be 0.
+         */
+        bool separate_segments;
+
         /* Value to be programmed in VCM_CACHE_SIZE. */
         uint8_t vcm_cache_size;
 };
diff --git a/src/broadcom/compiler/vir.c b/src/broadcom/compiler/vir.c
index a52205b1a2a..518aace991a 100644
--- a/src/broadcom/compiler/vir.c
+++ b/src/broadcom/compiler/vir.c
@@ -789,6 +789,14 @@ uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
         prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
         prog_data->vpm_output_size = align(c->num_vpm_writes, 8) / 8;
 
+        /* Set us up for shared input/output segments.  This is apparently
+         * necessary for our VCM setup to avoid varying corruption.
+         */
+        prog_data->separate_segments = false;
+        prog_data->vpm_output_size = MAX2(prog_data->vpm_output_size,
+                                          prog_data->vpm_input_size);
+        prog_data->vpm_input_size = 0;
+
         /* Compute VCM cache size.  We set up our program to take up less than
          * half of the VPM, so that any set of bin and render programs won't
          * run out of space.  We need space for at least one input segment,
diff --git a/src/gallium/drivers/v3d/v3dx_draw.c b/src/gallium/drivers/v3d/v3dx_draw.c
index bfb4af13ceb..7db1285f11d 100644
--- a/src/gallium/drivers/v3d/v3dx_draw.c
+++ b/src/gallium/drivers/v3d/v3dx_draw.c
@@ -201,12 +201,15 @@ v3d_emit_gl_shader_state(struct v3d_context *v3d,
                 /* XXX: Use combined input/output size flag in the common
                  * case.
                  */
-                shader.coordinate_shader_has_separate_input_and_output_vpm_blocks = true;
-                shader.vertex_shader_has_separate_input_and_output_vpm_blocks = true;
+                shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
+                        v3d->prog.cs->prog_data.vs->separate_segments;
+                shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
+                        v3d->prog.vs->prog_data.vs->separate_segments;
+
                 shader.coordinate_shader_input_vpm_segment_size =
-                        MAX2(v3d->prog.cs->prog_data.vs->vpm_input_size, 1);
+                        v3d->prog.cs->prog_data.vs->vpm_input_size;
                 shader.vertex_shader_input_vpm_segment_size =
-                        MAX2(v3d->prog.vs->prog_data.vs->vpm_input_size, 1);
+                        v3d->prog.vs->prog_data.vs->vpm_input_size;
 
                 shader.coordinate_shader_output_vpm_segment_size =
                         v3d->prog.cs->prog_data.vs->vpm_output_size;