vc4: Lower uniform loads to scalar in NIR.

This also moves the vec4-to-byte-addressing math into NIR, so that
algebraic has a chance at it.

2 files changed, 81 insertions, 31 deletions

diff --git a/src/gallium/drivers/vc4/vc4_nir_lower_io.c b/src/gallium/drivers/vc4/vc4_nir_lower_io.c
index fa06c893cfb..ffc120e8865 100644
--- a/src/gallium/drivers/vc4/vc4_nir_lower_io.c
+++ b/src/gallium/drivers/vc4/vc4_nir_lower_io.c
@@ -29,11 +29,30 @@
  * Walks the NIR generated by TGSI-to-NIR to lower its io intrinsics into
  * something amenable to the VC4 architecture.
  *
- * Currently, it split inputs and outputs into scalars, and drops any
- * non-position outputs in coordinate shaders.
+ * Currently, it split inputs, outputs, and uniforms into scalars, drops any
+ * non-position outputs in coordinate shaders, and fixes up the addressing on
+ * indirect uniform loads.
  */
 
 static void
+replace_intrinsic_with_vec4(nir_builder *b, nir_intrinsic_instr *intr,
+                            nir_ssa_def **comps)
+{
+
+        /* Batch things back together into a vec4.  This will get split by the
+         * later ALU scalarization pass.
+         */
+        nir_ssa_def *vec = nir_vec4(b, comps[0], comps[1], comps[2], comps[3]);
+
+        /* Replace the old intrinsic with a reference to our reconstructed
+         * vec4.
+         */
+        nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(vec),
+                                 ralloc_parent(b->impl));
+        nir_instr_remove(&intr->instr);
+}
+
+static void
 vc4_nir_lower_input(struct vc4_compile *c, nir_builder *b,
                     nir_intrinsic_instr *intr)
 {
@@ -102,18 +121,7 @@ vc4_nir_lower_input(struct vc4_compile *c, nir_builder *b,
                 break;
         }
 
-        /* Batch things back together into a vec4.  This will get split by the
-         * later ALU scalarization pass.
-         */
-        nir_ssa_def *vec_instr = nir_vec4(b, dests[0], dests[1],
-                                          dests[2], dests[3]);
-
-        /* Replace the old intrinsic with a reference to our reconstructed
-         * vec4.
-         */
-        nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(vec_instr),
-                                 ralloc_parent(b->impl));
-        nir_instr_remove(&intr->instr);
+        replace_intrinsic_with_vec4(b, intr, dests);
 }
 
 static void
@@ -159,6 +167,51 @@ vc4_nir_lower_output(struct vc4_compile *c, nir_builder *b,
 }
 
 static void
+vc4_nir_lower_uniform(struct vc4_compile *c, nir_builder *b,
+                      nir_intrinsic_instr *intr)
+{
+        /* All TGSI-to-NIR uniform loads are vec4. */
+        assert(intr->num_components == 4);
+
+        nir_builder_insert_before_instr(b, &intr->instr);
+
+        /* Generate scalar loads equivalent to the original VEC4. */
+        nir_ssa_def *dests[4];
+        for (unsigned i = 0; i < intr->num_components; i++) {
+                nir_intrinsic_instr *intr_comp =
+                        nir_intrinsic_instr_create(c->s, intr->intrinsic);
+                intr_comp->num_components = 1;
+                nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1, NULL);
+
+                if (intr->intrinsic == nir_intrinsic_load_uniform_indirect) {
+                        /* Convert the variable TGSI register index to a byte
+                         * offset.
+                         */
+                        intr_comp->src[0] =
+                                nir_src_for_ssa(nir_ishl(b,
+                                                         intr->src[0].ssa,
+                                                         nir_imm_int(b, 4)));
+
+                        /* Convert the offset to be a byte index, too. */
+                        intr_comp->const_index[0] = (intr->const_index[0] * 16 +
+                                                     i * 4);
+                } else {
+                        /* We want a dword index for non-indirect uniform
+                         * loads.
+                         */
+                        intr_comp->const_index[0] = (intr->const_index[0] * 4 +
+                                                     i);
+                }
+
+                dests[i] = &intr_comp->dest.ssa;
+
+                nir_builder_instr_insert(b, &intr_comp->instr);
+        }
+
+        replace_intrinsic_with_vec4(b, intr, dests);
+}
+
+static void
 vc4_nir_lower_io_instr(struct vc4_compile *c, nir_builder *b,
                        struct nir_instr *instr)
 {
@@ -175,6 +228,11 @@ vc4_nir_lower_io_instr(struct vc4_compile *c, nir_builder *b,
                 vc4_nir_lower_output(c, b, intr);
                 break;
 
+        case nir_intrinsic_load_uniform:
+        case nir_intrinsic_load_uniform_indirect:
+                vc4_nir_lower_uniform(c, b, intr);
+                break;
+
         default:
                 break;
         }
diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
index ddc997003b2..f2742986beb 100644
--- a/src/gallium/drivers/vc4/vc4_program.c
+++ b/src/gallium/drivers/vc4/vc4_program.c
@@ -42,6 +42,9 @@
 #include "simpenrose/simpenrose.h"
 #endif
 
+static struct qreg
+ntq_get_src(struct vc4_compile *c, nir_src src, int i);
+
 static void
 resize_qreg_array(struct vc4_compile *c,
                   struct qreg **regs,
@@ -64,10 +67,10 @@ resize_qreg_array(struct vc4_compile *c,
 }
 
 static struct qreg
-indirect_uniform_load(struct vc4_compile *c,
-                      struct qreg indirect_offset,
-                      unsigned offset)
+indirect_uniform_load(struct vc4_compile *c, nir_intrinsic_instr *intr)
 {
+        struct qreg indirect_offset = ntq_get_src(c, intr->src[0], 0);
+        uint32_t offset = intr->const_index[0];
         struct vc4_compiler_ubo_range *range = NULL;
         unsigned i;
         for (i = 0; i < c->num_uniform_ranges; i++) {
@@ -89,10 +92,6 @@ indirect_uniform_load(struct vc4_compile *c,
         };
 
         offset -= range->src_offset;
-        /* Translate the user's TGSI register index from the TGSI register
-         * base to a byte offset.
-         */
-        indirect_offset = qir_SHL(c, indirect_offset, qir_uniform_ui(c, 4));
 
         /* Adjust for where we stored the TGSI register base. */
         indirect_offset = qir_ADD(c, indirect_offset,
@@ -1793,19 +1792,12 @@ ntq_emit_intrinsic(struct vc4_compile *c, nir_intrinsic_instr *instr)
 
         switch (instr->intrinsic) {
         case nir_intrinsic_load_uniform:
-                for (int i = 0; i < instr->num_components; i++) {
-                        dest[i] = qir_uniform(c, QUNIFORM_UNIFORM,
-                                              instr->const_index[0] * 4 + i);
-                }
+                assert(instr->num_components == 1);
+                *dest = qir_uniform(c, QUNIFORM_UNIFORM, instr->const_index[0]);
                 break;
 
         case nir_intrinsic_load_uniform_indirect:
-                for (int i = 0; i < instr->num_components; i++) {
-                        dest[i] = indirect_uniform_load(c,
-                                                        ntq_get_src(c, instr->src[0], 0),
-                                                        (instr->const_index[0] *
-                                                         4 + i) * sizeof(float));
-                }
+                *dest = indirect_uniform_load(c, instr);
 
                 break;