nir: add options::vectorize_vec2_16bit to limit vectorization to vec2 16

for hardware that is scalar but can do 2 16-bit operations on low and high
16 bits of registers at once.

Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5002>

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 8b43c98754eec6db9302d74417c3b86e8056389a..9bf399698951b8767ffc83cbba5bb6319da73eb8 100644 (file)
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -3086,6 +3086,11 @@ typedef struct nir_shader_compiler_options {
    bool vectorize_io;
    bool lower_to_scalar;
 
+   /**
+    * Whether nir_opt_vectorize should only create 16-bit 2D vectors.
+    */
+   bool vectorize_vec2_16bit;
+
    /**
     * Should the linker unify inputs_read/outputs_written between adjacent
     * shader stages which are linked into a single program?

diff --git a/src/compiler/nir/nir_opt_vectorize.c b/src/compiler/nir/nir_opt_vectorize.c
index aee902c8a9407beb08067fee7dec3560b34e0c67..3bf65151e0d978ad70ae4e8b63765a9c598c1cda 100644 (file)
--- a/src/compiler/nir/nir_opt_vectorize.c
+++ b/src/compiler/nir/nir_opt_vectorize.c
@@ -162,7 +162,7 @@ instr_can_rewrite(nir_instr *instr)
  */
 
 static nir_instr *
-instr_try_combine(nir_instr *instr1, nir_instr *instr2)
+instr_try_combine(struct nir_shader *nir, nir_instr *instr1, nir_instr *instr2)
 {
    assert(instr1->type == nir_instr_type_alu);
    assert(instr2->type == nir_instr_type_alu);
@@ -177,6 +177,10 @@ instr_try_combine(nir_instr *instr1, nir_instr *instr2)
    if (total_components > 4)
       return NULL;
 
+   if (nir->options->vectorize_vec2_16bit &&
+       (total_components > 2 || alu1->dest.dest.ssa.bit_size != 16))
+      return NULL;
+
    nir_builder b;
    nir_builder_init(&b, nir_cf_node_get_function(&instr1->block->cf_node));
    b.cursor = nir_after_instr(instr1);
@@ -315,13 +319,14 @@ vec_instr_stack_create(void *mem_ctx)
 /* returns true if we were able to successfully replace the instruction */
 
 static bool
-vec_instr_stack_push(struct util_dynarray *stack, nir_instr *instr)
+vec_instr_stack_push(struct nir_shader *nir, struct util_dynarray *stack,
+                     nir_instr *instr)
 {
    /* Walk the stack from child to parent to make live ranges shorter by
     * matching the closest thing we can
     */
    util_dynarray_foreach_reverse(stack, nir_instr *, stack_instr) {
-      nir_instr *new_instr = instr_try_combine(*stack_instr, instr);
+      nir_instr *new_instr = instr_try_combine(nir, *stack_instr, instr);
       if (new_instr) {
          *stack_instr = new_instr;
          return true;
@@ -372,20 +377,21 @@ vec_instr_set_destroy(struct set *instr_set)
 }
 
 static bool
-vec_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr)
+vec_instr_set_add_or_rewrite(struct nir_shader *nir, struct set *instr_set,
+                             nir_instr *instr)
 {
    if (!instr_can_rewrite(instr))
       return false;
 
    struct util_dynarray *new_stack = vec_instr_stack_create(instr_set);
-   vec_instr_stack_push(new_stack, instr);
+   vec_instr_stack_push(nir, new_stack, instr);
 
    struct set_entry *entry = _mesa_set_search(instr_set, new_stack);
 
    if (entry) {
       ralloc_free(new_stack);
       struct util_dynarray *stack = (struct util_dynarray *) entry->key;
-      return vec_instr_stack_push(stack, instr);
+      return vec_instr_stack_push(nir, stack, instr);
    }
 
    _mesa_set_add(instr_set, new_stack);
@@ -393,7 +399,8 @@ vec_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr)
 }
 
 static void
-vec_instr_set_remove(struct set *instr_set, nir_instr *instr)
+vec_instr_set_remove(struct nir_shader *nir, struct set *instr_set,
+                     nir_instr *instr)
 {
    if (!instr_can_rewrite(instr))
       return;
@@ -410,7 +417,7 @@ vec_instr_set_remove(struct set *instr_set, nir_instr *instr)
     * comparison function as well.
     */
    struct util_dynarray *temp = vec_instr_stack_create(instr_set);
-   vec_instr_stack_push(temp, instr);
+   vec_instr_stack_push(nir, temp, instr);
    struct set_entry *entry = _mesa_set_search(instr_set, temp);
    ralloc_free(temp);
 
@@ -425,34 +432,35 @@ vec_instr_set_remove(struct set *instr_set, nir_instr *instr)
 }
 
 static bool
-vectorize_block(nir_block *block, struct set *instr_set)
+vectorize_block(struct nir_shader *nir, nir_block *block,
+                struct set *instr_set)
 {
    bool progress = false;
 
    nir_foreach_instr_safe(instr, block) {
-      if (vec_instr_set_add_or_rewrite(instr_set, instr))
+      if (vec_instr_set_add_or_rewrite(nir, instr_set, instr))
          progress = true;
    }
 
    for (unsigned i = 0; i < block->num_dom_children; i++) {
       nir_block *child = block->dom_children[i];
-      progress |= vectorize_block(child, instr_set);
+      progress |= vectorize_block(nir, child, instr_set);
    }
 
    nir_foreach_instr_reverse(instr, block)
-      vec_instr_set_remove(instr_set, instr);
+      vec_instr_set_remove(nir, instr_set, instr);
 
    return progress;
 }
 
 static bool
-nir_opt_vectorize_impl(nir_function_impl *impl)
+nir_opt_vectorize_impl(struct nir_shader *nir, nir_function_impl *impl)
 {
    struct set *instr_set = vec_instr_set_create();
 
    nir_metadata_require(impl, nir_metadata_dominance);
 
-   bool progress = vectorize_block(nir_start_block(impl), instr_set);
+   bool progress = vectorize_block(nir, nir_start_block(impl), instr_set);
 
    if (progress)
       nir_metadata_preserve(impl, nir_metadata_block_index |
@@ -469,7 +477,7 @@ nir_opt_vectorize(nir_shader *shader)
 
    nir_foreach_function(function, shader) {
       if (function->impl)
-         progress |= nir_opt_vectorize_impl(function->impl);
+         progress |= nir_opt_vectorize_impl(shader, function->impl);
    }
 
    return progress;