lima/ppir: add rsqrt codegen

Signed-off-by: Qiang Yu <yuq825@gmail.com>

4 files changed, 42 insertions, 1 deletions

diff --git a/src/gallium/drivers/lima/ir/pp/codegen.c b/src/gallium/drivers/lima/ir/pp/codegen.c
index def999c5bd..1d8f73036c 100644
--- a/src/gallium/drivers/lima/ir/pp/codegen.c
+++ b/src/gallium/drivers/lima/ir/pp/codegen.c
@@ -26,6 +26,7 @@
 
 #include "util/ralloc.h"
 #include "util/u_half.h"
+#include "util/bitscan.h"
 
 #include "ppir.h"
 #include "codegen.h"
@@ -199,7 +200,38 @@ static void ppir_codegen_encode_scl_add(ppir_node *node, void *code)
 
 static void ppir_codegen_encode_combine(ppir_node *node, void *code)
 {
-   
+   ppir_codegen_field_combine *f = code;
+   ppir_alu_node *alu = ppir_node_to_alu(node);
+
+   switch (node->op) {
+   case ppir_op_rsqrt:
+   {
+      f->scalar.dest_vec = false;
+      f->scalar.arg1_en = false;
+
+      ppir_dest *dest = &alu->dest;
+      int dest_component = ffs(dest->write_mask) - 1;
+      assert(dest_component >= 0);
+      f->scalar.dest = ppir_target_get_dest_reg_index(dest) + dest_component;
+      f->scalar.dest = dest->modifier;
+
+      ppir_src *src = alu->src;
+      int src_component = src->swizzle[dest_component];
+      f->scalar.arg0_src = ppir_target_get_src_reg_index(src) + src_component;
+      f->scalar.arg0_absolute = src->absolute;
+      f->scalar.arg0_negate = src->negate;
+
+      switch (node->op) {
+      case ppir_op_rsqrt:
+         f->scalar.op = ppir_codegen_combine_scalar_op_rsqrt;
+         break;
+      default:
+         break;
+      }
+   }
+   default:
+      break;
+   }
 }
 
 static void ppir_codegen_encode_store_temp(ppir_node *node, void *code)
diff --git a/src/gallium/drivers/lima/ir/pp/instr.c b/src/gallium/drivers/lima/ir/pp/instr.c
index 42bc5f5b82..db15d99681 100644
--- a/src/gallium/drivers/lima/ir/pp/instr.c
+++ b/src/gallium/drivers/lima/ir/pp/instr.c
@@ -128,6 +128,10 @@ bool ppir_instr_insert_node(ppir_instr *instr, ppir_node *node)
       instr->slots[PPIR_INSTR_SLOT_ALU_VEC_ADD] = node;
       break;
 
+   case ppir_op_rsqrt:
+      instr->slots[PPIR_INSTR_SLOT_ALU_COMBINE] = node;
+      break;
+
    case ppir_op_const:
    {
       int i;
diff --git a/src/gallium/drivers/lima/ir/pp/nir.c b/src/gallium/drivers/lima/ir/pp/nir.c
index 5a0aebc986..433ae2cf00 100644
--- a/src/gallium/drivers/lima/ir/pp/nir.c
+++ b/src/gallium/drivers/lima/ir/pp/nir.c
@@ -130,6 +130,7 @@ static int nir_to_ppir_opcodes[nir_num_opcodes] = {
    [nir_op_fdot2] = ppir_op_dot2,
    [nir_op_fdot3] = ppir_op_dot3,
    [nir_op_fdot4] = ppir_op_dot4,
+   [nir_op_frsq] = ppir_op_rsqrt,
 };
 
 static ppir_node *ppir_emit_alu(ppir_compiler *comp, nir_alu_instr *instr)
diff --git a/src/gallium/drivers/lima/ir/pp/node.c b/src/gallium/drivers/lima/ir/pp/node.c
index 18016a0dd6..fdb4649051 100644
--- a/src/gallium/drivers/lima/ir/pp/node.c
+++ b/src/gallium/drivers/lima/ir/pp/node.c
@@ -68,6 +68,10 @@ const ppir_op_info ppir_op_infos[] = {
       .name = "sum4",
       .type = ppir_node_type_alu,
    },
+   [ppir_op_rsqrt] = {
+      .name = "rsqrt",
+      .type = ppir_node_type_alu,
+   },
    [ppir_op_load_varying] = {
       .name = "ld_var",
       .type = ppir_node_type_load,