llvmpipe: Add inlines to quickly generate types matching the native SIMD register bitwidth.

2 files changed, 102 insertions, 12 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.h b/src/gallium/drivers/llvmpipe/lp_bld_type.h
index 46c298fa206..2fb233d335f 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_type.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_type.h
@@ -43,13 +43,18 @@
 
 
 /**
+ * Native SIMD register width.
+ *
+ * 128 for all architectures we care about.
+ */
+#define LP_NATIVE_VECTOR_WIDTH 128
+
+/**
  * Several functions can only cope with vectors of length up to this value.
  * You may need to increase that value if you want to represent bigger vectors.
  */
 #define LP_MAX_VECTOR_LENGTH 16
 
-#define LP_MAX_TYPE_WIDTH 64
-
 
 /**
  * The LLVM type system can't conveniently express all the things we care about
@@ -134,6 +139,91 @@ struct lp_build_context
 };
 
 
+static INLINE struct lp_type
+lp_type_float(unsigned width)
+{
+   struct lp_type res_type;
+
+   memset(&res_type, 0, sizeof res_type);
+   res_type.floating = TRUE;
+   res_type.sign = TRUE;
+   res_type.width = width;
+   res_type.length = LP_NATIVE_VECTOR_WIDTH / width;
+
+   return res_type;
+}
+
+
+static INLINE struct lp_type
+lp_type_int(unsigned width)
+{
+   struct lp_type res_type;
+
+   memset(&res_type, 0, sizeof res_type);
+   res_type.sign = TRUE;
+   res_type.width = width;
+   res_type.length = LP_NATIVE_VECTOR_WIDTH / width;
+
+   return res_type;
+}
+
+
+static INLINE struct lp_type
+lp_type_uint(unsigned width)
+{
+   struct lp_type res_type;
+
+   memset(&res_type, 0, sizeof res_type);
+   res_type.width = width;
+   res_type.length = LP_NATIVE_VECTOR_WIDTH / width;
+
+   return res_type;
+}
+
+
+static INLINE struct lp_type
+lp_type_unorm(unsigned width)
+{
+   struct lp_type res_type;
+
+   memset(&res_type, 0, sizeof res_type);
+   res_type.norm = TRUE;
+   res_type.width = width;
+   res_type.length = LP_NATIVE_VECTOR_WIDTH / width;
+
+   return res_type;
+}
+
+
+static INLINE struct lp_type
+lp_type_fixed(unsigned width)
+{
+   struct lp_type res_type;
+
+   memset(&res_type, 0, sizeof res_type);
+   res_type.sign = TRUE;
+   res_type.fixed = TRUE;
+   res_type.width = width;
+   res_type.length = LP_NATIVE_VECTOR_WIDTH / width;
+
+   return res_type;
+}
+
+
+static INLINE struct lp_type
+lp_type_ufixed(unsigned width)
+{
+   struct lp_type res_type;
+
+   memset(&res_type, 0, sizeof res_type);
+   res_type.fixed = TRUE;
+   res_type.width = width;
+   res_type.length = LP_NATIVE_VECTOR_WIDTH / width;
+
+   return res_type;
+}
+
+
 LLVMTypeRef
 lp_build_elem_type(struct lp_type type);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_test_blend.c b/src/gallium/drivers/llvmpipe/lp_test_blend.c
index e3af81cffb0..149fec1d542 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_blend.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_blend.c
@@ -530,11 +530,11 @@ test_one(unsigned verbose,
    success = TRUE;
    for(i = 0; i < n && success; ++i) {
       if(mode == AoS) {
-         uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t con[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t res[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
+         uint8_t src[LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t dst[LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t con[LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t res[LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t ref[LP_NATIVE_VECTOR_WIDTH/8];
          int64_t start_counter = 0;
          int64_t end_counter = 0;
 
@@ -595,11 +595,11 @@ test_one(unsigned verbose,
 
       if(mode == SoA) {
          const unsigned stride = type.length*type.width/8;
-         uint8_t src[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t dst[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t con[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t res[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
-         uint8_t ref[4*LP_MAX_VECTOR_LENGTH*LP_MAX_TYPE_WIDTH/8];
+         uint8_t src[4*LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t dst[4*LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t con[4*LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t res[4*LP_NATIVE_VECTOR_WIDTH/8];
+         uint8_t ref[4*LP_NATIVE_VECTOR_WIDTH/8];
          int64_t start_counter = 0;
          int64_t end_counter = 0;
          boolean mismatch;