Mesa (master): ac: add helpers for fast integer division by a constant

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: bfc795670ec82af5767cf360806e32322664604e
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=bfc795670ec82af5767cf360806e32322664604e

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Sat Sep 22 21:17:52 2018 -0400

ac: add helpers for fast integer division by a constant

---

 src/amd/common/ac_llvm_build.c | 61 ++++++++++++++++++++++++++++++++++++++++++
 src/amd/common/ac_llvm_build.h | 17 ++++++++++++
 2 files changed, 78 insertions(+)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index d0bcfe2dfb..2d78ca1b52 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -594,6 +594,67 @@ ac_build_fdiv(struct ac_llvm_context *ctx,
 	return ret;
 }
 
+/* See fast_idiv_by_const.h. */
+/* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
+LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
+				LLVMValueRef num,
+				LLVMValueRef multiplier,
+				LLVMValueRef pre_shift,
+				LLVMValueRef post_shift,
+				LLVMValueRef increment)
+{
+	LLVMBuilderRef builder = ctx->builder;
+
+	num = LLVMBuildLShr(builder, num, pre_shift, "");
+	num = LLVMBuildMul(builder,
+			   LLVMBuildZExt(builder, num, ctx->i64, ""),
+			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+	num = LLVMBuildAdd(builder, num,
+			   LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
+	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+	return LLVMBuildLShr(builder, num, post_shift, "");
+}
+
+/* See fast_idiv_by_const.h. */
+/* If num != UINT_MAX, this more efficient version can be used. */
+/* Set: increment = util_fast_udiv_info::increment; */
+LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
+				    LLVMValueRef num,
+				    LLVMValueRef multiplier,
+				    LLVMValueRef pre_shift,
+				    LLVMValueRef post_shift,
+				    LLVMValueRef increment)
+{
+	LLVMBuilderRef builder = ctx->builder;
+
+	num = LLVMBuildLShr(builder, num, pre_shift, "");
+	num = LLVMBuildNUWAdd(builder, num, increment, "");
+	num = LLVMBuildMul(builder,
+			   LLVMBuildZExt(builder, num, ctx->i64, ""),
+			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+	return LLVMBuildLShr(builder, num, post_shift, "");
+}
+
+/* See fast_idiv_by_const.h. */
+/* Both operands must fit in 31 bits and the divisor must not be 1. */
+LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
+					      LLVMValueRef num,
+					      LLVMValueRef multiplier,
+					      LLVMValueRef post_shift)
+{
+	LLVMBuilderRef builder = ctx->builder;
+
+	num = LLVMBuildMul(builder,
+			   LLVMBuildZExt(builder, num, ctx->i64, ""),
+			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+	return LLVMBuildLShr(builder, num, post_shift, "");
+}
+
 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
  * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
  * already multiplied by two. id is the cube face number.
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 5c3e221b2c..f68efbc49f 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -182,6 +182,23 @@ ac_build_fdiv(struct ac_llvm_context *ctx,
 	      LLVMValueRef num,
 	      LLVMValueRef den);
 
+LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
+				LLVMValueRef num,
+				LLVMValueRef multiplier,
+				LLVMValueRef pre_shift,
+				LLVMValueRef post_shift,
+				LLVMValueRef increment);
+LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
+				    LLVMValueRef num,
+				    LLVMValueRef multiplier,
+				    LLVMValueRef pre_shift,
+				    LLVMValueRef post_shift,
+				    LLVMValueRef increment);
+LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
+					      LLVMValueRef num,
+					      LLVMValueRef multiplier,
+					      LLVMValueRef post_shift);
+
 void
 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 		       bool is_deriv, bool is_array, bool is_lod,