[Mesa-dev] [PATCH 15/22] i965/vec4: fix SIMD-width lowering for double_to_single operation in IVB/VLV

[Samuel Iglesias Gonsálvez]

From: "Juan A. Suarez Romero" <jasuarez at igalia.com>

When spliting double_to_single() in Ivybridge/Valleyview, the second
part should use a temporal register, and then move the values to the
second half of the original destiny, so we get all the results in the
same register.
---
 src/mesa/drivers/dri/i965/brw_vec4.cpp           | 17 +++++++++++++----
 src/mesa/drivers/dri/i965/brw_vec4_generator.cpp |  1 +
 2 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp
index f533207..afabc22 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
@@ -2199,9 +2199,15 @@ vec4_visitor::lower_simd_width()
          linst->group = channel_offset;
          linst->size_written = size_written;
 
+         /* When spliting double_to_single() in Ivybridge, the second part
+          * should use in a temporal register. Later we will move the values
+          * to the second half of the original destination, so we get all the
+          * results in the same register. We use d2f_pass to detect this case.
+          */
+         bool d2f_pass = (inst->opcode == VEC4_OPCODE_FROM_DOUBLE && n > 0);
          /* Compute split dst region */
          dst_reg dst;
-         if (needs_temp) {
+         if (needs_temp || d2f_pass) {
             unsigned num_regs = DIV_ROUND_UP(size_written, REG_SIZE);
             dst = retype(dst_reg(VGRF, alloc.allocate(num_regs)),
                          inst->dst.type);
@@ -2234,9 +2240,12 @@ vec4_visitor::lower_simd_width()
          /* If we used a temporary to store the result of the split
           * instruction, copy the result to the original destination
           */
-         if (needs_temp) {
-            vec4_instruction *mov =
-               MOV(offset(inst->dst, lowered_width, n), src_reg(dst));
+         if (needs_temp || d2f_pass) {
+            vec4_instruction *mov;
+            if (d2f_pass)
+               mov = MOV(horiz_offset(inst->dst, n * type_sz(inst->dst.type)), src_reg(dst));
+            else
+               mov = MOV(offset(inst->dst, lowered_width, n), src_reg(dst));
             mov->exec_size = lowered_width;
             mov->group = channel_offset;
             mov->size_written = size_written;
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
index a68e14c..9409e9d 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
@@ -1521,6 +1521,7 @@ generate_code(struct brw_codegen *p,
              is_ivb_df);
 
       assert(inst->group % 8 == 0 ||
+             (inst->exec_size == 4 && inst->group % 4 == 0) ||
              inst->dst.type == BRW_REGISTER_TYPE_DF ||
              inst->src[0].type == BRW_REGISTER_TYPE_DF ||
              inst->src[1].type == BRW_REGISTER_TYPE_DF ||
-- 
2.9.3