[Mesa-dev] [PATCH 1/7] radv: add 16bit conversion operations

[Daniel Schürmann]

Signed-off-by: Daniel Schürmann <daniel.schuermann at campus.tu-berlin.de>
---
 src/amd/common/ac_llvm_build.c  |  5 +++++
 src/amd/common/ac_nir_to_llvm.c | 25 ++++++++++++++++++++++---
 2 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index a86ba962fa..d0b288f51d 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -136,6 +136,8 @@ ac_get_type_size(LLVMTypeRef type)
 	switch (kind) {
 	case LLVMIntegerTypeKind:
 		return LLVMGetIntTypeWidth(type) / 8;
+	case LLVMHalfTypeKind:
+		return 2;
 	case LLVMFloatTypeKind:
 		return 4;
 	case LLVMDoubleTypeKind:
@@ -278,6 +280,9 @@ void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
 	case LLVMIntegerTypeKind:
 		snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
 		break;
+	case LLVMHalfTypeKind:
+		snprintf(buf, bufsize, "f16");
+		break;
 	case LLVMFloatTypeKind:
 		snprintf(buf, bufsize, "f32");
 		break;
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index e44afbb9b2..827ba0b9c9 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1926,33 +1926,47 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
 			src[i] = ac_to_integer(&ctx->ac, src[i]);
 		result = ac_build_gather_values(&ctx->ac, src, num_components);
 		break;
+	case nir_op_f2i16:
 	case nir_op_f2i32:
 	case nir_op_f2i64:
 		src[0] = ac_to_float(&ctx->ac, src[0]);
 		result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
 		break;
+	case nir_op_f2u16:
 	case nir_op_f2u32:
 	case nir_op_f2u64:
 		src[0] = ac_to_float(&ctx->ac, src[0]);
 		result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
 		break;
+	case nir_op_i2f16:
 	case nir_op_i2f32:
 	case nir_op_i2f64:
 		src[0] = ac_to_integer(&ctx->ac, src[0]);
 		result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
 		break;
+	case nir_op_u2f16:
 	case nir_op_u2f32:
 	case nir_op_u2f64:
 		src[0] = ac_to_integer(&ctx->ac, src[0]);
 		result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
 		break;
-	case nir_op_f2f64:
+	case nir_op_f2f16_rtz:
 		src[0] = ac_to_float(&ctx->ac, src[0]);
-		result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
+		LLVMValueRef param[2] = { src[0], src[0] };
+		result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
+		result = LLVMBuildTrunc(ctx->ac.builder, result, def_type, "");
 		break;
+	case nir_op_f2f16_rtne:
+	case nir_op_f2f16_undef:
 	case nir_op_f2f32:
-		result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
+	case nir_op_f2f64:
+		src[0] = ac_to_float(&ctx->ac, src[0]);
+		if (get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < 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_u2u16:
 	case nir_op_u2u32:
 	case nir_op_u2u64:
 		src[0] = ac_to_integer(&ctx->ac, src[0]);
@@ -1961,6 +1975,7 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
 		else
 			result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
 		break;
+	case nir_op_i2i16:
 	case nir_op_i2i32:
 	case nir_op_i2i64:
 		src[0] = ac_to_integer(&ctx->ac, src[0]);
@@ -2089,6 +2104,10 @@ static void visit_load_const(struct ac_nir_context *ctx,
 
 	for (unsigned i = 0; i < instr->def.num_components; ++i) {
 		switch (instr->def.bit_size) {
+		case 16:
+			values[i] = LLVMConstInt(element_type,
+			                         instr->value.u16[i], false);
+			break;
 		case 32:
 			values[i] = LLVMConstInt(element_type,
 			                         instr->value.u32[i], false);
-- 
2.14.1