nouveau/mme: add a macro exit helper

At the moment it's turing+ only.

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

src/nouveau/mme/mme_builder.h

@@ -633,6 +633,39 @@ MME_DEF_END_WHILE(ine,   EQ,  false)
    for (bool run = (mme_start_while(b), true); run; \
         run = false, mme_end_while_##cmp((b), x, y))
 
+#define MME_DEF_EXIT(op, OP, if_true)                             \
+static inline void                                                \
+mme_exit_if_##op(struct mme_builder *b,                           \
+                 struct mme_value x, struct mme_value y)          \
+{                                                                 \
+   if (b->devinfo->cls_eng3d >= MME_CLS_TURING)                   \
+      mme_tu104_exit_if(b, MME_CMP_OP_##OP, if_true, x, y);       \
+   else                                                           \
+      unreachable("Unsupported GPU class");                       \
+}
+
+MME_DEF_EXIT(ilt,   LT,  true)
+MME_DEF_EXIT(ult,   LTU, true)
+MME_DEF_EXIT(ile,   LE,  true)
+MME_DEF_EXIT(ule,   LEU, true)
+MME_DEF_EXIT(ieq,   EQ,  true)
+MME_DEF_EXIT(ige,   LT,  false)
+MME_DEF_EXIT(uge,   LTU, false)
+MME_DEF_EXIT(igt,   LE,  false)
+MME_DEF_EXIT(ugt,   LEU, false)
+MME_DEF_EXIT(ine,   EQ,  false)
+
+#undef MME_DEF_EXIT
+
+#define mme_exit_if(b, cmp, x, y) \
+   mme_exit_if_##cmp(b, x, y)
+
+static inline void
+mme_exit(struct mme_builder *b)
+{
+   mme_exit_if_ieq(b, mme_zero(), mme_zero());
+}
+
 #ifdef __cplusplus
 }
 #endif

src/nouveau/mme/mme_tu104_builder.c

@@ -688,6 +688,24 @@ mme_tu104_end_while(struct mme_builder *b,
    mme_tu104_new_inst(tb);
 }
 
+void mme_tu104_exit_if(struct mme_builder *b,
+                       enum mme_cmp_op op,
+                       bool if_true,
+                       struct mme_value x,
+                       struct mme_value y)
+{
+   struct mme_tu104_builder *tb = &b->tu104;
+
+   /* we reverse it as we want to take the branch if the condition is true */
+   uint16_t control = if_true ? BITFIELD_BIT(15) : 0;
+   /* magic offset to exit the macro */
+   control |= 0x1000;
+   build_alu_to(b, mme_zero(), mme_cmp_to_tu104_branch_op(op), x, y, control,
+                true);
+
+   mme_tu104_new_inst(tb);
+}
+
 uint32_t *
 mme_tu104_builder_finish(struct mme_tu104_builder *tb, size_t *size_out)
 {

src/nouveau/mme/mme_tu104_builder.h

@@ -85,6 +85,12 @@ void mme_tu104_end_while(struct mme_builder *b,
                          struct mme_value x,
                          struct mme_value y);
 
+void mme_tu104_exit_if(struct mme_builder *b,
+                       enum mme_cmp_op op,
+                       bool if_true,
+                       struct mme_value x,
+                       struct mme_value y);
+
 uint32_t *mme_tu104_builder_finish(struct mme_tu104_builder *b,
                                    size_t *size_out);
 

src/nouveau/mme/tests/mme_tu104_sim_hw_test.cpp

@@ -1200,6 +1200,73 @@ TEST_F(mme_tu104_sim_test, bxx_exit)
       ASSERT_EQ(data[i], 0);
 }
 
+TEST_F(mme_tu104_sim_test, mme_exit)
+{
+   mme_builder b;
+   mme_builder_init(&b, devinfo);
+
+   mme_value vals[10];
+   for (uint32_t i = 0; i < 10; i++)
+      vals[i] = mme_mov(&b, mme_zero());
+
+   for (uint32_t i = 0; i < 10; i++)
+      mme_store_imm_addr(&b, data_addr + i * 4, mme_imm(0));
+
+   /* abort */
+   mme_exit(&b);
+
+   /* those writes won't be visible */
+   for (uint32_t i = 0; i < 10; i++)
+      vals[i] = mme_mov(&b, mme_imm(i));
+
+   for (uint32_t i = 0; i < 10; i++) {
+      mme_store_imm_addr(&b, data_addr + i * 4, vals[i]);
+   }
+
+   std::vector<uint32_t> params;
+
+   auto macro = mme_builder_finish_vec(&b);
+   test_macro(&b, macro, params);
+
+   uint32_t i;
+   for (i = 0; i < 10; i++)
+      ASSERT_EQ(data[i], 0);
+}
+
+TEST_F(mme_tu104_sim_test, mme_exit_if)
+{
+   mme_builder b;
+   mme_builder_init(&b, devinfo);
+
+   mme_value vals[10];
+   for (uint32_t i = 0; i < 10; i++)
+      vals[i] = mme_mov(&b, mme_zero());
+
+   for (uint32_t i = 0; i < 10; i++)
+      mme_store_imm_addr(&b, data_addr + i * 4, mme_imm(0));
+
+   /* shouldn't do anything */
+   mme_exit_if(&b, ieq, mme_zero(), mme_imm(1));
+
+   for (uint32_t i = 0; i < 10; i++)
+      vals[i] = mme_mov(&b, mme_imm(i));
+
+   for (uint32_t i = 0; i < 10; i++) {
+      /* abort on reaching 5 */
+      mme_exit_if(&b, ile, mme_imm(5), vals[i]);
+      mme_store_imm_addr(&b, data_addr + i * 4, vals[i]);
+   }
+
+   std::vector<uint32_t> params;
+
+   auto macro = mme_builder_finish_vec(&b);
+   test_macro(&b, macro, params);
+
+   uint32_t i;
+   for (i = 0; i < 10; i++)
+      ASSERT_EQ(data[i], i < 5 ? i : 0);
+}
+
 static bool c_ilt(int32_t x, int32_t y) { return x < y; };
 static bool c_ult(uint32_t x, uint32_t y) { return x < y; };
 static bool c_ile(int32_t x, int32_t y) { return x <= y; };