v3d: refactor ntq_emit_tmu_general() slightly
authorIago Toral Quiroga <itoral@igalia.com>
Tue, 6 Aug 2019 10:18:17 +0000 (12:18 +0200)
committerIago Toral Quiroga <itoral@igalia.com>
Tue, 13 Aug 2019 06:38:19 +0000 (08:38 +0200)
When we implement write masks on store operations we might need to
emit multiple write sequences for a given store intrinsic. To make
that easier, let's split the emission of the tmud instructions to
their own block after we are done with the code that only needs to
run once no matter how many write sequences we need to emit.

Reviewed-by: Eric Anholt <eric@anholt.net>
src/broadcom/compiler/nir_to_vir.c

index 03dac60645dc2739147466afb7772990d9eeb8c6..3857f96a9bd8830717b586557d0df15d4d11050d 100644 (file)
@@ -196,13 +196,20 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
                                      instr->intrinsic == nir_intrinsic_shared_atomic_add) &&
                                     (tmu_op == V3D_TMU_OP_WRITE_AND_READ_INC ||
                                      tmu_op == V3D_TMU_OP_WRITE_OR_READ_DEC));
+
         bool is_store = (instr->intrinsic == nir_intrinsic_store_ssbo ||
                          instr->intrinsic == nir_intrinsic_store_scratch ||
                          instr->intrinsic == nir_intrinsic_store_shared);
+
+        bool is_load = (instr->intrinsic == nir_intrinsic_load_uniform ||
+                        instr->intrinsic == nir_intrinsic_load_ubo ||
+                        instr->intrinsic == nir_intrinsic_load_ssbo ||
+                        instr->intrinsic == nir_intrinsic_load_scratch ||
+                        instr->intrinsic == nir_intrinsic_load_shared);
+
         bool has_index = !is_shared_or_scratch;
 
         int offset_src;
-        int tmu_writes = 1; /* address */
         if (instr->intrinsic == nir_intrinsic_load_uniform) {
                 offset_src = 0;
         } else if (instr->intrinsic == nir_intrinsic_load_ssbo ||
@@ -213,25 +220,8 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
                 offset_src = 0 + has_index;
         } else if (is_store) {
                 offset_src = 1 + has_index;
-                for (int i = 0; i < instr->num_components; i++) {
-                        vir_MOV_dest(c,
-                                     vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
-                                     ntq_get_src(c, instr->src[0], i));
-                        tmu_writes++;
-                }
         } else {
                 offset_src = 0 + has_index;
-                vir_MOV_dest(c,
-                             vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
-                             ntq_get_src(c, instr->src[1 + has_index], 0));
-                tmu_writes++;
-                if (tmu_op == V3D_TMU_OP_WRITE_CMPXCHG_READ_FLUSH) {
-                        vir_MOV_dest(c,
-                                     vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
-                                     ntq_get_src(c, instr->src[2 + has_index],
-                                                 0));
-                        tmu_writes++;
-                }
         }
 
         bool dynamic_src = !nir_src_is_const(instr->src[offset_src]);
@@ -239,12 +229,6 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
         if (!dynamic_src)
                 const_offset = nir_src_as_uint(instr->src[offset_src]);
 
-        /* Make sure we won't exceed the 16-entry TMU fifo if each thread is
-         * storing at the same time.
-         */
-        while (tmu_writes > 16 / c->threads)
-                c->threads /= 2;
-
         struct qreg offset;
         if (instr->intrinsic == nir_intrinsic_load_uniform) {
                 const_offset += nir_intrinsic_base(instr);
@@ -277,6 +261,34 @@ ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
                                                                 1 : 0]));
         }
 
+        int tmu_writes = 1; /* address */
+        if (is_store) {
+                for (int i = 0; i < instr->num_components; i++) {
+                        vir_MOV_dest(c,
+                                     vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
+                                     ntq_get_src(c, instr->src[0], i));
+                        tmu_writes++;
+                }
+        } else if (!is_load && !atomic_add_replaced) {
+                vir_MOV_dest(c,
+                             vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
+                             ntq_get_src(c, instr->src[1 + has_index], 0));
+                tmu_writes++;
+                if (tmu_op == V3D_TMU_OP_WRITE_CMPXCHG_READ_FLUSH) {
+                        vir_MOV_dest(c,
+                                     vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
+                                     ntq_get_src(c, instr->src[2 + has_index],
+                                                 0));
+                        tmu_writes++;
+                }
+        }
+
+        /* Make sure we won't exceed the 16-entry TMU fifo if each thread is
+         * storing at the same time.
+         */
+        while (tmu_writes > 16 / c->threads)
+                c->threads /= 2;
+
         /* The spec says that for atomics, the TYPE field is ignored, but that
          * doesn't seem to be the case for CMPXCHG.  Just use the number of
          * tmud writes we did to decide the type (or choose "32bit" for atomic