i965/cnl: Make URB {VS, GS, HS, DS} sizes non multiple of 3

v1: By Ben Widawsky <[email protected]>
v2: v1 had an assert only for VS. Add the restriction for GS, HS and
    DS as well and make sure the allocated sizes are not multiple of 3.
v3: Move the entry_size checks in to compiler code (Ken)

Signed-off-by: Anuj Phogat <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>

5 files changed, 34 insertions, 4 deletions

diff --git a/src/intel/compiler/brw_shader.cpp b/src/intel/compiler/brw_shader.cpp
index 269b8a099a4..53d0742d2e8 100644
--- a/src/intel/compiler/brw_shader.cpp
+++ b/src/intel/compiler/brw_shader.cpp
@@ -1197,6 +1197,14 @@ brw_compile_tes(const struct brw_compiler *compiler,
 
    /* URB entry sizes are stored as a multiple of 64 bytes. */
    prog_data->base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
+
+   /* On Cannonlake software shall not program an allocation size that
+    * specifies a size that is a multiple of 3 64B (512-bit) cachelines.
+    */
+   if (devinfo->gen == 10 &&
+       prog_data->base.urb_entry_size % 3 == 0)
+      prog_data->base.urb_entry_size++;
+
    prog_data->base.urb_read_length = 0;
 
    STATIC_ASSERT(BRW_TESS_PARTITIONING_INTEGER == TESS_SPACING_EQUAL - 1);
diff --git a/src/intel/compiler/brw_vec4.cpp b/src/intel/compiler/brw_vec4.cpp
index a641ebfbba9..b443effca9a 100644
--- a/src/intel/compiler/brw_vec4.cpp
+++ b/src/intel/compiler/brw_vec4.cpp
@@ -2839,10 +2839,17 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,
    const unsigned vue_entries =
       MAX2(nr_attribute_slots, (unsigned)prog_data->base.vue_map.num_slots);
 
-   if (compiler->devinfo->gen == 6)
+   if (compiler->devinfo->gen == 6) {
       prog_data->base.urb_entry_size = DIV_ROUND_UP(vue_entries, 8);
-   else
+   } else {
       prog_data->base.urb_entry_size = DIV_ROUND_UP(vue_entries, 4);
+      /* On Cannonlake software shall not program an allocation size that
+       * specifies a size that is a multiple of 3 64B (512-bit) cachelines.
+       */
+      if (compiler->devinfo->gen == 10 &&
+          prog_data->base.urb_entry_size % 3 == 0)
+         prog_data->base.urb_entry_size++;
+   }
 
    if (INTEL_DEBUG & DEBUG_VS) {
       fprintf(stderr, "VS Output ");
diff --git a/src/intel/compiler/brw_vec4_gs_visitor.cpp b/src/intel/compiler/brw_vec4_gs_visitor.cpp
index d0236df0912..f763f482365 100644
--- a/src/intel/compiler/brw_vec4_gs_visitor.cpp
+++ b/src/intel/compiler/brw_vec4_gs_visitor.cpp
@@ -817,10 +817,17 @@ brw_compile_gs(const struct brw_compiler *compiler, void *log_data,
    /* URB entry sizes are stored as a multiple of 64 bytes in gen7+ and
     * a multiple of 128 bytes in gen6.
     */
-   if (compiler->devinfo->gen >= 7)
+   if (compiler->devinfo->gen >= 7) {
       prog_data->base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
-   else
+      /* On Cannonlake software shall not program an allocation size that
+       * specifies a size that is a multiple of 3 64B (512-bit) cachelines.
+       */
+      if (compiler->devinfo->gen == 10 &&
+          prog_data->base.urb_entry_size % 3 == 0)
+         prog_data->base.urb_entry_size++;
+   } else {
       prog_data->base.urb_entry_size = ALIGN(output_size_bytes, 128) / 128;
+   }
 
    assert(shader->info.gs.output_primitive < ARRAY_SIZE(gl_prim_to_hw_prim));
    prog_data->output_topology =
diff --git a/src/intel/compiler/brw_vec4_tcs.cpp b/src/intel/compiler/brw_vec4_tcs.cpp
index 96597b8f2ad..c4d9f89a91b 100644
--- a/src/intel/compiler/brw_vec4_tcs.cpp
+++ b/src/intel/compiler/brw_vec4_tcs.cpp
@@ -441,6 +441,13 @@ brw_compile_tcs(const struct brw_compiler *compiler,
    /* URB entry sizes are stored as a multiple of 64 bytes. */
    vue_prog_data->urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
 
+   /* On Cannonlake software shall not program an allocation size that
+    * specifies a size that is a multiple of 3 64B (512-bit) cachelines.
+    */
+   if (devinfo->gen == 10 &&
+       vue_prog_data->urb_entry_size % 3 == 0)
+      vue_prog_data->urb_entry_size++;
+
    /* HS does not use the usual payload pushing from URB to GRFs,
     * because we don't have enough registers for a full-size payload, and
     * the hardware is broken on Haswell anyway.
diff --git a/src/mesa/drivers/dri/i965/gen7_urb.c b/src/mesa/drivers/dri/i965/gen7_urb.c
index 028161df395..525c9c4d3da 100644
--- a/src/mesa/drivers/dri/i965/gen7_urb.c
+++ b/src/mesa/drivers/dri/i965/gen7_urb.c
@@ -224,6 +224,7 @@ gen7_upload_urb(struct brw_context *brw, unsigned vs_size,
 
    BEGIN_BATCH(8);
    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+      assert(brw->gen != 10 || entry_size[i] % 3);
       OUT_BATCH((_3DSTATE_URB_VS + i) << 16 | (2 - 2));
       OUT_BATCH(entries[i] |
                 ((entry_size[i] - 1) << GEN7_URB_ENTRY_SIZE_SHIFT) |