shader: support D16 XY and XYZ formatted buffer loads and stores

This commit is contained in:
nmzik
2026-10-22 05:13:26 +02:00
parent 753f285198
commit c5016d9455
7 changed files with 76 additions and 11 deletions
@@ -74,8 +74,12 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
{0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
{0x72u, Opcode::BUFFER_GL1_INV, 0, 0},
{0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
{0x81u, Opcode::BUFFER_LOAD_FORMAT_D16_XY, 1, 16, false, false, true, 2},
{0x82u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZ, 2, 16, false, false, true, 3},
{0x83u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZW, 2, 16, false, false, true, 4},
{0x84u, Opcode::BUFFER_STORE_FORMAT_D16_X, 1, 16, false, false, true},
{0x85u, Opcode::BUFFER_STORE_FORMAT_D16_XY, 1, 16, false, false, true, 2},
{0x86u, Opcode::BUFFER_STORE_FORMAT_D16_XYZ, 2, 16, false, false, true, 3},
{0x87u, Opcode::BUFFER_STORE_FORMAT_D16_XYZW, 2, 16, false, false, true, 4},
};
@@ -636,6 +636,8 @@ std::string InstructionToString(const Instruction& inst) {
case Opcode::BUFFER_LOAD_FORMAT_XYZ:
case Opcode::BUFFER_LOAD_FORMAT_XYZW:
case Opcode::BUFFER_LOAD_FORMAT_D16_X:
case Opcode::BUFFER_LOAD_FORMAT_D16_XY:
case Opcode::BUFFER_LOAD_FORMAT_D16_XYZ:
case Opcode::BUFFER_LOAD_FORMAT_D16_XYZW:
case Opcode::BUFFER_STORE_FORMAT_X:
case Opcode::BUFFER_STORE_FORMAT_XY:
@@ -643,6 +645,8 @@ std::string InstructionToString(const Instruction& inst) {
case Opcode::BUFFER_STORE_FORMAT_XYZW:
case Opcode::BUFFER_STORE_FORMAT_D16_X:
case Opcode::BUFFER_STORE_FORMAT_D16_HI_X:
case Opcode::BUFFER_STORE_FORMAT_D16_XY:
case Opcode::BUFFER_STORE_FORMAT_D16_XYZ:
case Opcode::BUFFER_STORE_FORMAT_D16_XYZW:
case Opcode::BUFFER_LOAD_UBYTE:
case Opcode::BUFFER_LOAD_UBYTE_D16:
@@ -547,6 +547,8 @@ enum class Opcode {
BUFFER_LOAD_FORMAT_XYZ,
BUFFER_LOAD_FORMAT_XYZW,
BUFFER_LOAD_FORMAT_D16_X,
BUFFER_LOAD_FORMAT_D16_XY,
BUFFER_LOAD_FORMAT_D16_XYZ,
BUFFER_LOAD_FORMAT_D16_XYZW,
BUFFER_STORE_FORMAT_X,
BUFFER_STORE_FORMAT_XY,
@@ -554,6 +556,8 @@ enum class Opcode {
BUFFER_STORE_FORMAT_XYZW,
BUFFER_STORE_FORMAT_D16_X,
BUFFER_STORE_FORMAT_D16_HI_X,
BUFFER_STORE_FORMAT_D16_XY,
BUFFER_STORE_FORMAT_D16_XYZ,
BUFFER_STORE_FORMAT_D16_XYZW,
BUFFER_LOAD_UBYTE,
BUFFER_LOAD_UBYTE_D16,
@@ -997,6 +997,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
case Decoder::Opcode::BUFFER_LOAD_FORMAT_XYZ:
case Decoder::Opcode::BUFFER_LOAD_FORMAT_XYZW:
case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_X:
case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_XY:
case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_XYZ:
case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_XYZW:
case Decoder::Opcode::TBUFFER_LOAD_FORMAT_X:
case Decoder::Opcode::TBUFFER_LOAD_FORMAT_XY:
@@ -1017,6 +1019,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
case Decoder::Opcode::BUFFER_STORE_FORMAT_XYZW:
case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_X:
case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_HI_X:
case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XY:
case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XYZ:
case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XYZW:
case Decoder::Opcode::TBUFFER_STORE_FORMAT_X:
case Decoder::Opcode::TBUFFER_STORE_FORMAT_XY:
@@ -487,7 +487,7 @@ void ValidateProgram(const Program& program, bool require_ssa) {
const auto access = BufferAccessOf(inst.GetOpcode());
const bool packed_formatted =
(access == BufferAccess::Read || access == BufferAccess::Write) && memory.formatted &&
memory.data_bits == 16u && memory.component_count == 4u && buffer_components == 2u;
memory.data_bits == 16u && (memory.component_count + 1u) / 2u == buffer_components;
if (buffer_components > 1u &&
(!vector_buffer || (memory.data_bits != 32u && !packed_formatted) ||
memory.data_dwords != buffer_components || memory.component_index != 0u)) {