Mesa (master): ac/nir: select v_cvt_pkrtz for all conversions from f32 to f16 for radeonsi

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: bfb95725aaa13a16011f16ad9ec9501b5a373ce6
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=bfb95725aaa13a16011f16ad9ec9501b5a373ce6

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Sat May  9 23:00:44 2020 -0400

ac/nir: select v_cvt_pkrtz for all conversions from f32 to f16 for radeonsi

Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5003>

---

 src/amd/llvm/ac_nir_to_llvm.c | 43 ++++++++++++++++++++++++++++++++++++-------
 1 file changed, 36 insertions(+), 7 deletions(-)

diff --git a/src/amd/llvm/ac_nir_to_llvm.c b/src/amd/llvm/ac_nir_to_llvm.c
index d47624a07e1..8a707c93666 100644
--- a/src/amd/llvm/ac_nir_to_llvm.c
+++ b/src/amd/llvm/ac_nir_to_llvm.c
@@ -943,18 +943,47 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
 		result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
 		break;
 	case nir_op_f2f16_rtz:
+	case nir_op_f2f16:
+	case nir_op_f2fmp:
 		src[0] = ac_to_float(&ctx->ac, src[0]);
-		if (LLVMTypeOf(src[0]) == ctx->ac.f64)
-			src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
-		LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
-		result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
-		result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
+
+		/* For OpenGL, we want fast packing with v_cvt_pkrtz_f16, but if we use it,
+		 * all f32->f16 conversions have to round towards zero, because both scalar
+		 * and vec2 down-conversions have to round equally.
+		 */
+		if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL ||
+		    instr->op == nir_op_f2f16_rtz) {
+			src[0] = ac_to_float(&ctx->ac, src[0]);
+
+			if (LLVMTypeOf(src[0]) == ctx->ac.f64)
+				src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
+
+			/* Fast path conversion. This only works if NIR is vectorized
+			 * to vec2 16.
+			 */
+			if (LLVMTypeOf(src[0]) == ctx->ac.v2f32) {
+				LLVMValueRef args[] = {
+					ac_llvm_extract_elem(&ctx->ac, src[0], 0),
+					ac_llvm_extract_elem(&ctx->ac, src[0], 1),
+				};
+				result = ac_build_cvt_pkrtz_f16(&ctx->ac, args);
+				break;
+			}
+
+			assert(ac_get_llvm_num_components(src[0]) == 1);
+			LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
+			result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
+			result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
+		} else {
+			if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
+				result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
+			else
+				result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
+		}
 		break;
 	case nir_op_f2f16_rtne:
-	case nir_op_f2f16:
 	case nir_op_f2f32:
 	case nir_op_f2f64:
-	case nir_op_f2fmp:
 		src[0] = ac_to_float(&ctx->ac, src[0]);
 		if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
 			result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");