tests: Add DXIL test for FP32 -> FP16 conversions.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
This commit is contained in:
parent
9162e82fb3
commit
db943f2341
|
@ -190,6 +190,431 @@ void test_fragment_coords(void)
|
|||
destroy_test_context(&context);
|
||||
}
|
||||
|
||||
void test_shader_instructions_dxil(void)
|
||||
{
|
||||
D3D12_FEATURE_DATA_D3D12_OPTIONS4 options4;
|
||||
D3D12_ROOT_PARAMETER root_parameters[2];
|
||||
ID3D12RootSignature *root_signature;
|
||||
D3D12_ROOT_SIGNATURE_DESC rs_desc;
|
||||
ID3D12Resource *output_buffer;
|
||||
ID3D12Resource *input_buffer;
|
||||
struct test_context context;
|
||||
struct resource_readback rb;
|
||||
ID3D12PipelineState *pso;
|
||||
unsigned int i, j;
|
||||
|
||||
struct test
|
||||
{
|
||||
const D3D12_SHADER_BYTECODE *cs;
|
||||
float input_data[4];
|
||||
uint32_t output_data[4];
|
||||
bool native_fp16;
|
||||
bool is_todo;
|
||||
uint32_t output_data_alt[4];
|
||||
bool allows_alternative_result;
|
||||
};
|
||||
|
||||
static const BYTE legacy_f32_to_f16_code[] =
|
||||
{
|
||||
#if 0
|
||||
StructuredBuffer<float> Buf : register(t0);
|
||||
RWStructuredBuffer<uint> RWBuf : register(u0);
|
||||
|
||||
[numthreads(4, 1, 1)]
|
||||
void main(uint thr : SV_DispatchThreadID)
|
||||
{
|
||||
RWBuf[thr] = f32tof16(Buf[thr]);
|
||||
}
|
||||
#endif
|
||||
0x44, 0x58, 0x42, 0x43, 0x02, 0xfc, 0x26, 0xf2, 0xd6, 0xa4, 0x1f, 0xdd, 0xc5, 0xb6, 0x7b, 0x11, 0xd8, 0xcc, 0xdc, 0xe5, 0x01, 0x00, 0x00, 0x00, 0xd8, 0x06, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00,
|
||||
0x38, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, 0xe8, 0x00, 0x00, 0x00, 0x04, 0x01, 0x00, 0x00, 0x53, 0x46, 0x49, 0x30, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x49, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x4f, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x50, 0x53, 0x56, 0x30, 0x78, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
|
||||
0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x48, 0x41, 0x53, 0x48, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xbe, 0x01, 0x25, 0x60, 0x73, 0xdd, 0xac, 0xf7, 0xf7, 0xc8, 0xb6, 0x07,
|
||||
0x78, 0x1d, 0xab, 0x36, 0x44, 0x58, 0x49, 0x4c, 0xcc, 0x05, 0x00, 0x00, 0x62, 0x00, 0x05, 0x00, 0x73, 0x01, 0x00, 0x00, 0x44, 0x58, 0x49, 0x4c, 0x02, 0x01, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
|
||||
0xb4, 0x05, 0x00, 0x00, 0x42, 0x43, 0xc0, 0xde, 0x21, 0x0c, 0x00, 0x00, 0x6a, 0x01, 0x00, 0x00, 0x0b, 0x82, 0x20, 0x00, 0x02, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, 0x07, 0x81, 0x23, 0x91,
|
||||
0x41, 0xc8, 0x04, 0x49, 0x06, 0x10, 0x32, 0x39, 0x92, 0x01, 0x84, 0x0c, 0x25, 0x05, 0x08, 0x19, 0x1e, 0x04, 0x8b, 0x62, 0x80, 0x14, 0x45, 0x02, 0x42, 0x92, 0x0b, 0x42, 0xa4, 0x10, 0x32, 0x14,
|
||||
0x38, 0x08, 0x18, 0x4b, 0x0a, 0x32, 0x52, 0x88, 0x48, 0x90, 0x14, 0x20, 0x43, 0x46, 0x88, 0xa5, 0x00, 0x19, 0x32, 0x42, 0xe4, 0x48, 0x0e, 0x90, 0x91, 0x22, 0xc4, 0x50, 0x41, 0x51, 0x81, 0x8c,
|
||||
0xe1, 0x83, 0xe5, 0x8a, 0x04, 0x29, 0x46, 0x06, 0x51, 0x18, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x1b, 0x8c, 0xe0, 0xff, 0xff, 0xff, 0xff, 0x07, 0x40, 0x02, 0xa8, 0x0d, 0x86, 0xf0, 0xff, 0xff,
|
||||
0xff, 0xff, 0x03, 0x20, 0x01, 0xd5, 0x06, 0x62, 0xf8, 0xff, 0xff, 0xff, 0xff, 0x01, 0x90, 0x00, 0x49, 0x18, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x13, 0x82, 0x60, 0x42, 0x20, 0x4c, 0x08, 0x06,
|
||||
0x00, 0x00, 0x00, 0x00, 0x89, 0x20, 0x00, 0x00, 0x3a, 0x00, 0x00, 0x00, 0x32, 0x22, 0x48, 0x09, 0x20, 0x64, 0x85, 0x04, 0x93, 0x22, 0xa4, 0x84, 0x04, 0x93, 0x22, 0xe3, 0x84, 0xa1, 0x90, 0x14,
|
||||
0x12, 0x4c, 0x8a, 0x8c, 0x0b, 0x84, 0xa4, 0x4c, 0x10, 0x68, 0x23, 0x00, 0x25, 0x00, 0x14, 0xe6, 0x08, 0xc0, 0xa0, 0x0c, 0x63, 0x0c, 0x22, 0x33, 0x00, 0x65, 0x18, 0xc3, 0xd0, 0xb9, 0x69, 0xb8,
|
||||
0xfc, 0x09, 0x7b, 0x08, 0xc9, 0x5f, 0x09, 0x69, 0x25, 0x26, 0xbf, 0xb8, 0x6d, 0x54, 0x18, 0x63, 0xcc, 0x98, 0x23, 0x40, 0x48, 0xdd, 0x33, 0x5c, 0xfe, 0x84, 0x3d, 0x84, 0xe4, 0x87, 0x40, 0x33,
|
||||
0x2c, 0x04, 0x0a, 0x56, 0x39, 0xd2, 0x60, 0x63, 0xa8, 0x41, 0xad, 0x2c, 0x60, 0xb0, 0x31, 0xc6, 0x18, 0x43, 0x0d, 0x7a, 0x73, 0x04, 0x41, 0x31, 0xd8, 0x50, 0x63, 0x44, 0x92, 0x03, 0x01, 0x33,
|
||||
0x75, 0xe3, 0xc0, 0x0e, 0xe1, 0x30, 0x0f, 0xf3, 0xe0, 0x06, 0xb3, 0x40, 0x0f, 0xf2, 0x50, 0x0f, 0xe3, 0x40, 0x0f, 0xf5, 0x20, 0x0f, 0xe5, 0x40, 0x0e, 0xa2, 0x50, 0x0f, 0xe6, 0x60, 0x0e, 0xe5,
|
||||
0x20, 0x0f, 0x7c, 0x60, 0x0e, 0xec, 0xf0, 0x0e, 0xe1, 0x40, 0x0f, 0x7e, 0x80, 0x02, 0x43, 0x75, 0x26, 0x33, 0x18, 0x07, 0x76, 0x08, 0x87, 0x79, 0x98, 0x07, 0x37, 0x90, 0x85, 0x5b, 0x98, 0x05,
|
||||
0x7a, 0x90, 0x87, 0x7a, 0x18, 0x07, 0x7a, 0xa8, 0x07, 0x79, 0x28, 0x07, 0x72, 0x10, 0x85, 0x7a, 0x30, 0x07, 0x73, 0x28, 0x07, 0x79, 0xe0, 0x83, 0x7a, 0x70, 0x87, 0x79, 0x48, 0x87, 0x73, 0x70,
|
||||
0x87, 0x72, 0x20, 0x07, 0x30, 0x48, 0x07, 0x77, 0xa0, 0x07, 0x3f, 0x40, 0xc1, 0xa0, 0x3b, 0x47, 0x00, 0x0a, 0x00, 0x00, 0x13, 0x14, 0x72, 0xc0, 0x87, 0x74, 0x60, 0x87, 0x36, 0x68, 0x87, 0x79,
|
||||
0x68, 0x03, 0x72, 0xc0, 0x87, 0x0d, 0xaf, 0x50, 0x0e, 0x6d, 0xd0, 0x0e, 0x7a, 0x50, 0x0e, 0x6d, 0x00, 0x0f, 0x7a, 0x30, 0x07, 0x72, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x71, 0xa0,
|
||||
0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x78, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x71, 0x60, 0x07, 0x7a, 0x30, 0x07, 0x72, 0xd0, 0x06, 0xe9, 0x30, 0x07, 0x72, 0xa0, 0x07, 0x73,
|
||||
0x20, 0x07, 0x6d, 0x90, 0x0e, 0x76, 0x40, 0x07, 0x7a, 0x60, 0x07, 0x74, 0xd0, 0x06, 0xe6, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x60, 0x0e, 0x73, 0x20, 0x07, 0x7a, 0x30, 0x07,
|
||||
0x72, 0xd0, 0x06, 0xe6, 0x60, 0x07, 0x74, 0xa0, 0x07, 0x76, 0x40, 0x07, 0x6d, 0xe0, 0x0e, 0x78, 0xa0, 0x07, 0x71, 0x60, 0x07, 0x7a, 0x30, 0x07, 0x72, 0xa0, 0x07, 0x76, 0x40, 0x07, 0x43, 0x9e,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x86, 0x3c, 0x04, 0x10, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x79, 0x0e, 0x20, 0x00, 0x02, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x18, 0xf2, 0x34, 0x40, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x30, 0xe4, 0x79, 0x80, 0x00, 0x18, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x60,
|
||||
0xc8, 0x23, 0x01, 0x01, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x40, 0x16, 0x08, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x32, 0x1e, 0x98, 0x14, 0x19, 0x11, 0x4c, 0x90, 0x8c, 0x09, 0x26, 0x47,
|
||||
0xc6, 0x04, 0x43, 0x1a, 0x25, 0x50, 0x04, 0xc5, 0x30, 0x02, 0x50, 0x18, 0x85, 0x40, 0x71, 0x04, 0x80, 0x72, 0x81, 0x10, 0x9e, 0x01, 0x20, 0x3b, 0x03, 0x00, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00,
|
||||
0x3f, 0x00, 0x00, 0x00, 0x1a, 0x03, 0x4c, 0x90, 0x46, 0x02, 0x13, 0x44, 0x35, 0x18, 0x63, 0x0b, 0x73, 0x3b, 0x03, 0xb1, 0x2b, 0x93, 0x9b, 0x4b, 0x7b, 0x73, 0x03, 0x99, 0x71, 0xb9, 0x01, 0x41,
|
||||
0xa1, 0x0b, 0x3b, 0x9b, 0x7b, 0x91, 0x2a, 0x62, 0x2a, 0x0a, 0x9a, 0x2a, 0xfa, 0x9a, 0xb9, 0x81, 0x79, 0x31, 0x4b, 0x73, 0x0b, 0x63, 0x4b, 0xd9, 0x10, 0x04, 0x13, 0x84, 0xc1, 0x98, 0x20, 0x0c,
|
||||
0xc7, 0x06, 0x61, 0x20, 0x26, 0x08, 0x03, 0xb2, 0x41, 0x18, 0x0c, 0x0a, 0x63, 0x73, 0x1b, 0x06, 0xc4, 0x20, 0x26, 0x08, 0x43, 0x32, 0x41, 0xb0, 0x1e, 0x02, 0x13, 0x84, 0x41, 0x99, 0x20, 0x0c,
|
||||
0xcb, 0x06, 0x61, 0x70, 0x36, 0x24, 0xca, 0xc2, 0x28, 0xca, 0xd0, 0x28, 0xcf, 0x86, 0x00, 0x9a, 0x20, 0x60, 0xce, 0x04, 0x21, 0x62, 0x36, 0x2c, 0x8a, 0xc4, 0x28, 0xca, 0xd0, 0x4c, 0xd3, 0xf4,
|
||||
0x6c, 0x08, 0xa8, 0x0d, 0x44, 0x54, 0x01, 0xc0, 0x04, 0x41, 0x00, 0x48, 0xb4, 0x85, 0xa5, 0xb9, 0x4d, 0x10, 0xb2, 0x66, 0xc3, 0xe0, 0x0c, 0xc3, 0x06, 0x42, 0xc9, 0x1c, 0x6d, 0x43, 0x71, 0x61,
|
||||
0x80, 0xb5, 0x55, 0x61, 0x63, 0xb3, 0x6b, 0x73, 0x49, 0x23, 0x2b, 0x73, 0xa3, 0x9b, 0x12, 0x04, 0x55, 0xc8, 0xf0, 0x5c, 0xec, 0xca, 0xe4, 0xe6, 0xd2, 0xde, 0xdc, 0xa6, 0x04, 0x44, 0x13, 0x32,
|
||||
0x3c, 0x17, 0xbb, 0x30, 0x36, 0xbb, 0x32, 0xb9, 0x29, 0x81, 0x51, 0x87, 0x0c, 0xcf, 0x65, 0x0e, 0x2d, 0x8c, 0xac, 0x4c, 0xae, 0xe9, 0x8d, 0xac, 0x8c, 0x6d, 0x4a, 0x80, 0x94, 0x21, 0xc3, 0x73,
|
||||
0x91, 0x2b, 0x9b, 0x7b, 0xab, 0x93, 0x1b, 0x2b, 0x9b, 0x9b, 0x12, 0x54, 0x75, 0xc8, 0xf0, 0x5c, 0xca, 0xdc, 0xe8, 0xe4, 0xf2, 0xa0, 0xde, 0xd2, 0xdc, 0xe8, 0xe6, 0xa6, 0x04, 0x1b, 0x00, 0x00,
|
||||
0x79, 0x18, 0x00, 0x00, 0x4c, 0x00, 0x00, 0x00, 0x33, 0x08, 0x80, 0x1c, 0xc4, 0xe1, 0x1c, 0x66, 0x14, 0x01, 0x3d, 0x88, 0x43, 0x38, 0x84, 0xc3, 0x8c, 0x42, 0x80, 0x07, 0x79, 0x78, 0x07, 0x73,
|
||||
0x98, 0x71, 0x0c, 0xe6, 0x00, 0x0f, 0xed, 0x10, 0x0e, 0xf4, 0x80, 0x0e, 0x33, 0x0c, 0x42, 0x1e, 0xc2, 0xc1, 0x1d, 0xce, 0xa1, 0x1c, 0x66, 0x30, 0x05, 0x3d, 0x88, 0x43, 0x38, 0x84, 0x83, 0x1b,
|
||||
0xcc, 0x03, 0x3d, 0xc8, 0x43, 0x3d, 0x8c, 0x03, 0x3d, 0xcc, 0x78, 0x8c, 0x74, 0x70, 0x07, 0x7b, 0x08, 0x07, 0x79, 0x48, 0x87, 0x70, 0x70, 0x07, 0x7a, 0x70, 0x03, 0x76, 0x78, 0x87, 0x70, 0x20,
|
||||
0x87, 0x19, 0xcc, 0x11, 0x0e, 0xec, 0x90, 0x0e, 0xe1, 0x30, 0x0f, 0x6e, 0x30, 0x0f, 0xe3, 0xf0, 0x0e, 0xf0, 0x50, 0x0e, 0x33, 0x10, 0xc4, 0x1d, 0xde, 0x21, 0x1c, 0xd8, 0x21, 0x1d, 0xc2, 0x61,
|
||||
0x1e, 0x66, 0x30, 0x89, 0x3b, 0xbc, 0x83, 0x3b, 0xd0, 0x43, 0x39, 0xb4, 0x03, 0x3c, 0xbc, 0x83, 0x3c, 0x84, 0x03, 0x3b, 0xcc, 0xf0, 0x14, 0x76, 0x60, 0x07, 0x7b, 0x68, 0x07, 0x37, 0x68, 0x87,
|
||||
0x72, 0x68, 0x07, 0x37, 0x80, 0x87, 0x70, 0x90, 0x87, 0x70, 0x60, 0x07, 0x76, 0x28, 0x07, 0x76, 0xf8, 0x05, 0x76, 0x78, 0x87, 0x77, 0x80, 0x87, 0x5f, 0x08, 0x87, 0x71, 0x18, 0x87, 0x72, 0x98,
|
||||
0x87, 0x79, 0x98, 0x81, 0x2c, 0xee, 0xf0, 0x0e, 0xee, 0xe0, 0x0e, 0xf5, 0xc0, 0x0e, 0xec, 0x30, 0x03, 0x62, 0xc8, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xdc, 0x61,
|
||||
0x1c, 0xca, 0x21, 0x1c, 0xc4, 0x81, 0x1d, 0xca, 0x61, 0x06, 0xd6, 0x90, 0x43, 0x39, 0xc8, 0x43, 0x39, 0x98, 0x43, 0x39, 0xc8, 0x43, 0x39, 0xb8, 0xc3, 0x38, 0x94, 0x43, 0x38, 0x88, 0x03, 0x3b,
|
||||
0x94, 0xc3, 0x2f, 0xbc, 0x83, 0x3c, 0xfc, 0x82, 0x3b, 0xd4, 0x03, 0x3b, 0xb0, 0xc3, 0x0c, 0xc4, 0x21, 0x07, 0x7c, 0x70, 0x03, 0x7a, 0x28, 0x87, 0x76, 0x80, 0x87, 0x19, 0xd1, 0x43, 0x0e, 0xf8,
|
||||
0xe0, 0x06, 0xe4, 0x20, 0x0e, 0xe7, 0xe0, 0x06, 0xf6, 0x10, 0x0e, 0xf2, 0xc0, 0x0e, 0xe1, 0x90, 0x0f, 0xef, 0x50, 0x0f, 0xf4, 0x00, 0x00, 0x00, 0x71, 0x20, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00,
|
||||
0x56, 0x20, 0x0d, 0x97, 0xef, 0x3c, 0xbe, 0x10, 0x11, 0xc0, 0x44, 0x84, 0x40, 0x33, 0x2c, 0x84, 0x09, 0x50, 0xc3, 0xe5, 0x3b, 0x8f, 0xbf, 0x10, 0x06, 0x20, 0x60, 0xdf, 0x6d, 0xb7, 0xce, 0x57,
|
||||
0xeb, 0x36, 0x70, 0x0d, 0x97, 0xef, 0x3c, 0x7e, 0x04, 0x58, 0x1b, 0x55, 0x14, 0x44, 0x54, 0x3a, 0xc0, 0xe0, 0x17, 0xb7, 0x6d, 0x04, 0xd8, 0x70, 0xf9, 0xce, 0xe3, 0x47, 0x80, 0xb5, 0x51, 0x45,
|
||||
0x41, 0x44, 0xec, 0xe4, 0x44, 0x84, 0x8f, 0xdc, 0xb6, 0x05, 0x48, 0xc3, 0xe5, 0x3b, 0x8f, 0x3f, 0x1d, 0x11, 0x01, 0x0c, 0xe2, 0xe0, 0x23, 0xb7, 0x6d, 0x00, 0x04, 0x03, 0x20, 0x0d, 0x00, 0x00,
|
||||
0x61, 0x20, 0x00, 0x00, 0x1b, 0x00, 0x00, 0x00, 0x13, 0x04, 0x41, 0x2c, 0x10, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x34, 0x66, 0x00, 0x4a, 0xae, 0x74, 0x03, 0xca, 0x52, 0xa0, 0x08, 0x05,
|
||||
0x0a, 0x53, 0x80, 0x54, 0x09, 0x8c, 0x00, 0x00, 0x23, 0x06, 0x09, 0x00, 0x82, 0x60, 0x20, 0x4d, 0x87, 0xf0, 0x3c, 0xcc, 0x88, 0x41, 0x02, 0x80, 0x20, 0x18, 0x48, 0x14, 0x22, 0x40, 0x50, 0x33,
|
||||
0x62, 0x60, 0x00, 0x20, 0x08, 0x06, 0x44, 0x86, 0x44, 0x23, 0x06, 0x0a, 0x00, 0x82, 0x60, 0xd0, 0x60, 0x88, 0x10, 0x48, 0x05, 0x34, 0x9a, 0x10, 0x00, 0x23, 0x06, 0x06, 0x00, 0x82, 0x60, 0x70,
|
||||
0x6c, 0x49, 0x30, 0x62, 0xe0, 0x00, 0x20, 0x08, 0x06, 0x8f, 0x96, 0x18, 0x44, 0x15, 0x38, 0x8e, 0x83, 0x4c, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
};
|
||||
|
||||
static const BYTE fptrunc_f32_to_f16_code[] =
|
||||
{
|
||||
#if 0
|
||||
StructuredBuffer<float> Buf : register(t0);
|
||||
RWStructuredBuffer<uint> RWBuf : register(u0);
|
||||
|
||||
[numthreads(4, 1, 1)]
|
||||
void main(uint thr : SV_DispatchThreadID)
|
||||
{
|
||||
RWBuf[thr] = asuint16(half(Buf[thr]));
|
||||
}
|
||||
#endif
|
||||
0x44, 0x58, 0x42, 0x43, 0x0c, 0xe9, 0x3d, 0x5e, 0x3c, 0xb6, 0x70, 0xf2, 0x4b, 0x1d, 0x8a, 0x9a, 0x5b, 0x8e, 0x3c, 0x17, 0x01, 0x00, 0x00, 0x00, 0xbc, 0x06, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00,
|
||||
0x38, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, 0xe8, 0x00, 0x00, 0x00, 0x04, 0x01, 0x00, 0x00, 0x53, 0x46, 0x49, 0x30, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x49, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x4f, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x50, 0x53, 0x56, 0x30, 0x78, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
|
||||
0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x48, 0x41, 0x53, 0x48, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x76, 0xa9, 0xed, 0x8d, 0xa4, 0x34, 0x22, 0x78, 0xf6, 0x50, 0x75, 0xc7,
|
||||
0xe4, 0x48, 0xa0, 0x6b, 0x44, 0x58, 0x49, 0x4c, 0xb0, 0x05, 0x00, 0x00, 0x62, 0x00, 0x05, 0x00, 0x6c, 0x01, 0x00, 0x00, 0x44, 0x58, 0x49, 0x4c, 0x02, 0x01, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
|
||||
0x98, 0x05, 0x00, 0x00, 0x42, 0x43, 0xc0, 0xde, 0x21, 0x0c, 0x00, 0x00, 0x63, 0x01, 0x00, 0x00, 0x0b, 0x82, 0x20, 0x00, 0x02, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, 0x07, 0x81, 0x23, 0x91,
|
||||
0x41, 0xc8, 0x04, 0x49, 0x06, 0x10, 0x32, 0x39, 0x92, 0x01, 0x84, 0x0c, 0x25, 0x05, 0x08, 0x19, 0x1e, 0x04, 0x8b, 0x62, 0x80, 0x14, 0x45, 0x02, 0x42, 0x92, 0x0b, 0x42, 0xa4, 0x10, 0x32, 0x14,
|
||||
0x38, 0x08, 0x18, 0x4b, 0x0a, 0x32, 0x52, 0x88, 0x48, 0x90, 0x14, 0x20, 0x43, 0x46, 0x88, 0xa5, 0x00, 0x19, 0x32, 0x42, 0xe4, 0x48, 0x0e, 0x90, 0x91, 0x22, 0xc4, 0x50, 0x41, 0x51, 0x81, 0x8c,
|
||||
0xe1, 0x83, 0xe5, 0x8a, 0x04, 0x29, 0x46, 0x06, 0x51, 0x18, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x1b, 0x8c, 0xe0, 0xff, 0xff, 0xff, 0xff, 0x07, 0x40, 0x02, 0xa8, 0x0d, 0x86, 0xf0, 0xff, 0xff,
|
||||
0xff, 0xff, 0x03, 0x20, 0x01, 0xd5, 0x06, 0x62, 0xf8, 0xff, 0xff, 0xff, 0xff, 0x01, 0x90, 0x00, 0x49, 0x18, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x13, 0x82, 0x60, 0x42, 0x20, 0x4c, 0x08, 0x06,
|
||||
0x00, 0x00, 0x00, 0x00, 0x89, 0x20, 0x00, 0x00, 0x3a, 0x00, 0x00, 0x00, 0x32, 0x22, 0x48, 0x09, 0x20, 0x64, 0x85, 0x04, 0x93, 0x22, 0xa4, 0x84, 0x04, 0x93, 0x22, 0xe3, 0x84, 0xa1, 0x90, 0x14,
|
||||
0x12, 0x4c, 0x8a, 0x8c, 0x0b, 0x84, 0xa4, 0x4c, 0x10, 0x68, 0x23, 0x00, 0x25, 0x00, 0x14, 0xe6, 0x08, 0xc0, 0xa0, 0x0c, 0x63, 0x0c, 0x22, 0x33, 0x00, 0x37, 0x0d, 0x97, 0x3f, 0x61, 0x0f, 0x21,
|
||||
0xf9, 0x2b, 0x21, 0xad, 0xc4, 0xe4, 0x17, 0xb7, 0x8d, 0x0a, 0x63, 0x8c, 0x19, 0x73, 0x04, 0x08, 0xa1, 0x7b, 0x86, 0xcb, 0x9f, 0xb0, 0x87, 0x90, 0xfc, 0x10, 0x68, 0x86, 0x85, 0x40, 0x41, 0x2a,
|
||||
0xc7, 0x19, 0x6a, 0x0c, 0x34, 0x68, 0x95, 0x05, 0x0c, 0x35, 0xc6, 0x18, 0x63, 0xa0, 0x41, 0x6d, 0x8e, 0x20, 0x28, 0x86, 0x1a, 0x68, 0x8c, 0x47, 0x70, 0x20, 0x60, 0xa6, 0x6e, 0x1c, 0xd8, 0x21,
|
||||
0x1c, 0xe6, 0x61, 0x1e, 0xdc, 0x60, 0x16, 0xe8, 0x41, 0x1e, 0xea, 0x61, 0x1c, 0xe8, 0xa1, 0x1e, 0xe4, 0xa1, 0x1c, 0xc8, 0x41, 0x14, 0xea, 0xc1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0x81, 0x0f, 0xcc,
|
||||
0x81, 0x1d, 0xde, 0x21, 0x1c, 0xe8, 0xc1, 0x0f, 0x50, 0x60, 0x68, 0xce, 0x64, 0x06, 0xe3, 0xc0, 0x0e, 0xe1, 0x30, 0x0f, 0xf3, 0xe0, 0x06, 0xb2, 0x70, 0x0b, 0xb3, 0x40, 0x0f, 0xf2, 0x50, 0x0f,
|
||||
0xe3, 0x40, 0x0f, 0xf5, 0x20, 0x0f, 0xe5, 0x40, 0x0e, 0xa2, 0x50, 0x0f, 0xe6, 0x60, 0x0e, 0xe5, 0x20, 0x0f, 0x7c, 0x50, 0x0f, 0xee, 0x30, 0x0f, 0xe9, 0x70, 0x0e, 0xee, 0x50, 0x0e, 0xe4, 0x00,
|
||||
0x06, 0xe9, 0xe0, 0x0e, 0xf4, 0xe0, 0x07, 0x28, 0x18, 0x54, 0xe7, 0x08, 0x40, 0x61, 0x14, 0x60, 0x8e, 0x00, 0x02, 0x00, 0x13, 0x14, 0x72, 0xc0, 0x87, 0x74, 0x60, 0x87, 0x36, 0x68, 0x87, 0x79,
|
||||
0x68, 0x03, 0x72, 0xc0, 0x87, 0x0d, 0xae, 0x50, 0x0e, 0x6d, 0xd0, 0x0e, 0x7a, 0x50, 0x0e, 0x6d, 0x00, 0x0f, 0x7a, 0x30, 0x07, 0x72, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x71, 0xa0,
|
||||
0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x78, 0xa0, 0x07, 0x78, 0xd0, 0x06, 0xe9, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x71, 0x60, 0x07, 0x6d, 0x90, 0x0e, 0x73, 0x20, 0x07, 0x7a, 0x30, 0x07, 0x72,
|
||||
0xd0, 0x06, 0xe9, 0x60, 0x07, 0x74, 0xa0, 0x07, 0x76, 0x40, 0x07, 0x6d, 0x60, 0x0e, 0x71, 0x60, 0x07, 0x7a, 0x10, 0x07, 0x76, 0xd0, 0x06, 0xe6, 0x30, 0x07, 0x72, 0xa0, 0x07, 0x73, 0x20, 0x07,
|
||||
0x6d, 0x60, 0x0e, 0x76, 0x40, 0x07, 0x7a, 0x60, 0x07, 0x74, 0xd0, 0x06, 0xee, 0x80, 0x07, 0x7a, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x7a, 0x60, 0x07, 0x74, 0x30, 0xe4, 0x09, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x60, 0xc8, 0x43, 0x00, 0x01, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc0, 0x90, 0x67, 0x01, 0x02, 0x40, 0x00, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x80, 0x21, 0x4f, 0x03, 0x04, 0xc0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x43, 0x1e, 0x08, 0x08, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xb2, 0x40,
|
||||
0x0b, 0x00, 0x00, 0x00, 0x32, 0x1e, 0x98, 0x10, 0x19, 0x11, 0x4c, 0x90, 0x8c, 0x09, 0x26, 0x47, 0xc6, 0x04, 0x43, 0x1a, 0x25, 0x50, 0x04, 0xc5, 0x30, 0x02, 0x50, 0x18, 0x85, 0x40, 0x6f, 0x04,
|
||||
0x80, 0x6e, 0x81, 0x03, 0x02, 0x22, 0x90, 0x9d, 0x01, 0x20, 0x3a, 0x03, 0x00, 0x00, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00, 0x3f, 0x00, 0x00, 0x00, 0x1a, 0x03, 0x4c, 0x90, 0x46, 0x02, 0x13, 0x44,
|
||||
0x35, 0x18, 0x63, 0x0b, 0x73, 0x3b, 0x03, 0xb1, 0x2b, 0x93, 0x9b, 0x4b, 0x7b, 0x73, 0x03, 0x99, 0x71, 0xb9, 0x01, 0x41, 0xa1, 0x0b, 0x3b, 0x9b, 0x7b, 0x91, 0x2a, 0x62, 0x2a, 0x0a, 0x9a, 0x2a,
|
||||
0xfa, 0x9a, 0xb9, 0x81, 0x79, 0x31, 0x4b, 0x73, 0x0b, 0x63, 0x4b, 0xd9, 0x10, 0x04, 0x13, 0x84, 0xa1, 0x98, 0x20, 0x0c, 0xc6, 0x06, 0x61, 0x20, 0x26, 0x08, 0xc3, 0xb1, 0x41, 0x18, 0x0c, 0x0a,
|
||||
0x63, 0x73, 0x1b, 0x06, 0xc4, 0x20, 0x26, 0x08, 0x03, 0x32, 0x41, 0xa0, 0x1c, 0x02, 0x13, 0x84, 0x21, 0x99, 0x20, 0x0c, 0xca, 0x06, 0x61, 0x70, 0x36, 0x24, 0xca, 0xc2, 0x28, 0xca, 0xd0, 0x28,
|
||||
0xcf, 0x86, 0x00, 0x9a, 0x20, 0x58, 0xcd, 0x04, 0xe1, 0x59, 0x36, 0x2c, 0x8a, 0xc4, 0x28, 0xca, 0xd0, 0x4c, 0xd3, 0xf4, 0x6c, 0x08, 0xa8, 0x0d, 0x44, 0x54, 0x01, 0xc0, 0x04, 0x41, 0x00, 0x48,
|
||||
0xb4, 0x85, 0xa5, 0xb9, 0x4d, 0x10, 0x2e, 0x66, 0xc3, 0xe0, 0x0c, 0xc3, 0x06, 0x42, 0xc9, 0x1c, 0x6d, 0x43, 0x71, 0x61, 0x80, 0xb5, 0x55, 0x61, 0x63, 0xb3, 0x6b, 0x73, 0x49, 0x23, 0x2b, 0x73,
|
||||
0xa3, 0x9b, 0x12, 0x04, 0x55, 0xc8, 0xf0, 0x5c, 0xec, 0xca, 0xe4, 0xe6, 0xd2, 0xde, 0xdc, 0xa6, 0x04, 0x44, 0x13, 0x32, 0x3c, 0x17, 0xbb, 0x30, 0x36, 0xbb, 0x32, 0xb9, 0x29, 0x81, 0x51, 0x87,
|
||||
0x0c, 0xcf, 0x65, 0x0e, 0x2d, 0x8c, 0xac, 0x4c, 0xae, 0xe9, 0x8d, 0xac, 0x8c, 0x6d, 0x4a, 0x80, 0x94, 0x21, 0xc3, 0x73, 0x91, 0x2b, 0x9b, 0x7b, 0xab, 0x93, 0x1b, 0x2b, 0x9b, 0x9b, 0x12, 0x54,
|
||||
0x75, 0xc8, 0xf0, 0x5c, 0xca, 0xdc, 0xe8, 0xe4, 0xf2, 0xa0, 0xde, 0xd2, 0xdc, 0xe8, 0xe6, 0xa6, 0x04, 0x1b, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00, 0x4c, 0x00, 0x00, 0x00, 0x33, 0x08, 0x80, 0x1c,
|
||||
0xc4, 0xe1, 0x1c, 0x66, 0x14, 0x01, 0x3d, 0x88, 0x43, 0x38, 0x84, 0xc3, 0x8c, 0x42, 0x80, 0x07, 0x79, 0x78, 0x07, 0x73, 0x98, 0x71, 0x0c, 0xe6, 0x00, 0x0f, 0xed, 0x10, 0x0e, 0xf4, 0x80, 0x0e,
|
||||
0x33, 0x0c, 0x42, 0x1e, 0xc2, 0xc1, 0x1d, 0xce, 0xa1, 0x1c, 0x66, 0x30, 0x05, 0x3d, 0x88, 0x43, 0x38, 0x84, 0x83, 0x1b, 0xcc, 0x03, 0x3d, 0xc8, 0x43, 0x3d, 0x8c, 0x03, 0x3d, 0xcc, 0x78, 0x8c,
|
||||
0x74, 0x70, 0x07, 0x7b, 0x08, 0x07, 0x79, 0x48, 0x87, 0x70, 0x70, 0x07, 0x7a, 0x70, 0x03, 0x76, 0x78, 0x87, 0x70, 0x20, 0x87, 0x19, 0xcc, 0x11, 0x0e, 0xec, 0x90, 0x0e, 0xe1, 0x30, 0x0f, 0x6e,
|
||||
0x30, 0x0f, 0xe3, 0xf0, 0x0e, 0xf0, 0x50, 0x0e, 0x33, 0x10, 0xc4, 0x1d, 0xde, 0x21, 0x1c, 0xd8, 0x21, 0x1d, 0xc2, 0x61, 0x1e, 0x66, 0x30, 0x89, 0x3b, 0xbc, 0x83, 0x3b, 0xd0, 0x43, 0x39, 0xb4,
|
||||
0x03, 0x3c, 0xbc, 0x83, 0x3c, 0x84, 0x03, 0x3b, 0xcc, 0xf0, 0x14, 0x76, 0x60, 0x07, 0x7b, 0x68, 0x07, 0x37, 0x68, 0x87, 0x72, 0x68, 0x07, 0x37, 0x80, 0x87, 0x70, 0x90, 0x87, 0x70, 0x60, 0x07,
|
||||
0x76, 0x28, 0x07, 0x76, 0xf8, 0x05, 0x76, 0x78, 0x87, 0x77, 0x80, 0x87, 0x5f, 0x08, 0x87, 0x71, 0x18, 0x87, 0x72, 0x98, 0x87, 0x79, 0x98, 0x81, 0x2c, 0xee, 0xf0, 0x0e, 0xee, 0xe0, 0x0e, 0xf5,
|
||||
0xc0, 0x0e, 0xec, 0x30, 0x03, 0x62, 0xc8, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xdc, 0x61, 0x1c, 0xca, 0x21, 0x1c, 0xc4, 0x81, 0x1d, 0xca, 0x61, 0x06, 0xd6, 0x90,
|
||||
0x43, 0x39, 0xc8, 0x43, 0x39, 0x98, 0x43, 0x39, 0xc8, 0x43, 0x39, 0xb8, 0xc3, 0x38, 0x94, 0x43, 0x38, 0x88, 0x03, 0x3b, 0x94, 0xc3, 0x2f, 0xbc, 0x83, 0x3c, 0xfc, 0x82, 0x3b, 0xd4, 0x03, 0x3b,
|
||||
0xb0, 0xc3, 0x0c, 0xc4, 0x21, 0x07, 0x7c, 0x70, 0x03, 0x7a, 0x28, 0x87, 0x76, 0x80, 0x87, 0x19, 0xd1, 0x43, 0x0e, 0xf8, 0xe0, 0x06, 0xe4, 0x20, 0x0e, 0xe7, 0xe0, 0x06, 0xf6, 0x10, 0x0e, 0xf2,
|
||||
0xc0, 0x0e, 0xe1, 0x90, 0x0f, 0xef, 0x50, 0x0f, 0xf4, 0x00, 0x00, 0x00, 0x71, 0x20, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x46, 0x20, 0x0d, 0x97, 0xef, 0x3c, 0xbe, 0x10, 0x11, 0xc0, 0x44, 0x84,
|
||||
0x40, 0x33, 0x2c, 0x84, 0x09, 0x5c, 0xc3, 0xe5, 0x3b, 0x8f, 0x1f, 0x01, 0xd6, 0x46, 0x15, 0x05, 0x11, 0x95, 0x0e, 0x30, 0xf8, 0xc5, 0x6d, 0xdb, 0x00, 0x36, 0x5c, 0xbe, 0xf3, 0xf8, 0x11, 0x60,
|
||||
0x6d, 0x54, 0x51, 0x10, 0x11, 0x3b, 0x39, 0x11, 0xe1, 0x23, 0xb7, 0x6d, 0x01, 0xd2, 0x70, 0xf9, 0xce, 0xe3, 0x4f, 0x47, 0x44, 0x00, 0x83, 0x38, 0xf8, 0xc8, 0x6d, 0x1b, 0x00, 0xc1, 0x00, 0x48,
|
||||
0x03, 0x00, 0x00, 0x00, 0x61, 0x20, 0x00, 0x00, 0x1b, 0x00, 0x00, 0x00, 0x13, 0x04, 0x41, 0x2c, 0x10, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x34, 0x66, 0x00, 0x4a, 0xae, 0x74, 0x03, 0xca,
|
||||
0x52, 0xa0, 0x30, 0x05, 0x08, 0x95, 0xc0, 0x08, 0x00, 0x00, 0x00, 0x00, 0x23, 0x06, 0x09, 0x00, 0x82, 0x60, 0x00, 0x49, 0x86, 0xe0, 0x38, 0xcb, 0x88, 0x41, 0x02, 0x80, 0x20, 0x18, 0x40, 0xd3,
|
||||
0x21, 0x3c, 0x0f, 0x33, 0x62, 0x60, 0x00, 0x20, 0x08, 0x06, 0xc4, 0x75, 0x40, 0x23, 0x06, 0x0a, 0x00, 0x82, 0x60, 0xb0, 0x5c, 0x87, 0x10, 0x44, 0xc5, 0x33, 0x9a, 0x10, 0x00, 0x17, 0xe0, 0xbb,
|
||||
0x20, 0xef, 0x05, 0x23, 0x46, 0x0c, 0x1c, 0x00, 0x04, 0xc1, 0xa0, 0xd9, 0x16, 0xc4, 0xb0, 0x82, 0xe7, 0x79, 0x14, 0x0a, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
};
|
||||
|
||||
static const BYTE fp16_add_rounding_mode_code[] =
|
||||
{
|
||||
#if 0
|
||||
StructuredBuffer<float> Buf : register(t0);
|
||||
RWStructuredBuffer<half2> RWBuf : register(u0);
|
||||
|
||||
[numthreads(4, 1, 1)]
|
||||
void main(uint thr : SV_DispatchThreadID)
|
||||
{
|
||||
precise half result = half(Buf[thr]) + half(0.5);
|
||||
RWBuf[thr].x = result;
|
||||
}
|
||||
#endif
|
||||
0x44, 0x58, 0x42, 0x43, 0x80, 0x0b, 0xfa, 0x38, 0x4c, 0x9f, 0x96, 0x0e, 0x2a, 0x53, 0xaa, 0x48, 0x14, 0xe4, 0xf1, 0xad, 0x01, 0x00, 0x00, 0x00, 0xd4, 0x06, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00,
|
||||
0x38, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, 0xe8, 0x00, 0x00, 0x00, 0x04, 0x01, 0x00, 0x00, 0x53, 0x46, 0x49, 0x30, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x49, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x4f, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x50, 0x53, 0x56, 0x30, 0x78, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
|
||||
0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x48, 0x41, 0x53, 0x48, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc1, 0x42, 0x47, 0x10, 0xe6, 0x07, 0x25, 0x9f, 0x02, 0x19, 0x28, 0x76,
|
||||
0x39, 0x75, 0xa7, 0xd5, 0x44, 0x58, 0x49, 0x4c, 0xc8, 0x05, 0x00, 0x00, 0x62, 0x00, 0x05, 0x00, 0x72, 0x01, 0x00, 0x00, 0x44, 0x58, 0x49, 0x4c, 0x02, 0x01, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
|
||||
0xb0, 0x05, 0x00, 0x00, 0x42, 0x43, 0xc0, 0xde, 0x21, 0x0c, 0x00, 0x00, 0x69, 0x01, 0x00, 0x00, 0x0b, 0x82, 0x20, 0x00, 0x02, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, 0x07, 0x81, 0x23, 0x91,
|
||||
0x41, 0xc8, 0x04, 0x49, 0x06, 0x10, 0x32, 0x39, 0x92, 0x01, 0x84, 0x0c, 0x25, 0x05, 0x08, 0x19, 0x1e, 0x04, 0x8b, 0x62, 0x80, 0x14, 0x45, 0x02, 0x42, 0x92, 0x0b, 0x42, 0xa4, 0x10, 0x32, 0x14,
|
||||
0x38, 0x08, 0x18, 0x4b, 0x0a, 0x32, 0x52, 0x88, 0x48, 0x90, 0x14, 0x20, 0x43, 0x46, 0x88, 0xa5, 0x00, 0x19, 0x32, 0x42, 0xe4, 0x48, 0x0e, 0x90, 0x91, 0x22, 0xc4, 0x50, 0x41, 0x51, 0x81, 0x8c,
|
||||
0xe1, 0x83, 0xe5, 0x8a, 0x04, 0x29, 0x46, 0x06, 0x51, 0x18, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x1b, 0x8c, 0xe0, 0xff, 0xff, 0xff, 0xff, 0x07, 0x40, 0x02, 0xa8, 0x0d, 0x86, 0xf0, 0xff, 0xff,
|
||||
0xff, 0xff, 0x03, 0x20, 0x01, 0xd5, 0x06, 0x62, 0xf8, 0xff, 0xff, 0xff, 0xff, 0x01, 0x90, 0x00, 0x49, 0x18, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x13, 0x82, 0x60, 0x42, 0x20, 0x4c, 0x08, 0x06,
|
||||
0x00, 0x00, 0x00, 0x00, 0x89, 0x20, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, 0x32, 0x22, 0x48, 0x09, 0x20, 0x64, 0x85, 0x04, 0x93, 0x22, 0xa4, 0x84, 0x04, 0x93, 0x22, 0xe3, 0x84, 0xa1, 0x90, 0x14,
|
||||
0x12, 0x4c, 0x8a, 0x8c, 0x0b, 0x84, 0xa4, 0x4c, 0x10, 0x68, 0x23, 0x00, 0x25, 0x00, 0x14, 0xe6, 0x08, 0xc0, 0xa0, 0x0c, 0x63, 0x0c, 0x22, 0x33, 0x00, 0x37, 0x0d, 0x97, 0x3f, 0x61, 0x0f, 0x21,
|
||||
0xf9, 0x2b, 0x21, 0xad, 0xc4, 0xe4, 0x17, 0xb7, 0x8d, 0x0a, 0x63, 0x8c, 0x19, 0x73, 0x04, 0x08, 0xa1, 0x7b, 0x86, 0xcb, 0x9f, 0xb0, 0x87, 0x90, 0xfc, 0x10, 0x68, 0x86, 0x85, 0x40, 0x41, 0x2a,
|
||||
0xc7, 0x19, 0x6a, 0x0c, 0x34, 0x68, 0x8d, 0x02, 0x94, 0x05, 0x0c, 0x35, 0x46, 0x6b, 0xad, 0xa1, 0x41, 0x6e, 0x8e, 0x20, 0x28, 0x86, 0x1a, 0x68, 0x0c, 0x48, 0x71, 0x20, 0x60, 0xa6, 0x6e, 0x1c,
|
||||
0xd8, 0x21, 0x1c, 0xe6, 0x61, 0x1e, 0xdc, 0x60, 0x16, 0xe8, 0x41, 0x1e, 0xea, 0x61, 0x1c, 0xe8, 0xa1, 0x1e, 0xe4, 0xa1, 0x1c, 0xc8, 0x41, 0x14, 0xea, 0xc1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0x81,
|
||||
0x0f, 0xcc, 0x81, 0x1d, 0xde, 0x21, 0x1c, 0xe8, 0xc1, 0x0f, 0x50, 0x60, 0x88, 0x0e, 0x23, 0x08, 0xcd, 0x4c, 0x6a, 0x30, 0x0e, 0xec, 0x10, 0x0e, 0xf3, 0x30, 0x0f, 0x6e, 0x20, 0x0b, 0xb7, 0x30,
|
||||
0x0b, 0xf4, 0x20, 0x0f, 0xf5, 0x30, 0x0e, 0xf4, 0x50, 0x0f, 0xf2, 0x50, 0x0e, 0xe4, 0x20, 0x0a, 0xf5, 0x60, 0x0e, 0xe6, 0x50, 0x0e, 0xf2, 0xc0, 0x07, 0xf6, 0x50, 0x0e, 0xe3, 0x40, 0x0f, 0xef,
|
||||
0x20, 0x0f, 0x7c, 0x80, 0x0e, 0xe1, 0xc0, 0x0e, 0xe6, 0xc0, 0x06, 0x60, 0x20, 0x07, 0x7e, 0x00, 0x06, 0x7e, 0x80, 0x02, 0x4b, 0x77, 0x8e, 0x00, 0x14, 0x00, 0x00, 0x00, 0x13, 0x14, 0x72, 0xc0,
|
||||
0x87, 0x74, 0x60, 0x87, 0x36, 0x68, 0x87, 0x79, 0x68, 0x03, 0x72, 0xc0, 0x87, 0x0d, 0xae, 0x50, 0x0e, 0x6d, 0xd0, 0x0e, 0x7a, 0x50, 0x0e, 0x6d, 0x00, 0x0f, 0x7a, 0x30, 0x07, 0x72, 0xa0, 0x07,
|
||||
0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x71, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x78, 0xa0, 0x07, 0x78, 0xd0, 0x06, 0xe9, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x71, 0x60, 0x07, 0x6d, 0x90,
|
||||
0x0e, 0x73, 0x20, 0x07, 0x7a, 0x30, 0x07, 0x72, 0xd0, 0x06, 0xe9, 0x60, 0x07, 0x74, 0xa0, 0x07, 0x76, 0x40, 0x07, 0x6d, 0x60, 0x0e, 0x71, 0x60, 0x07, 0x7a, 0x10, 0x07, 0x76, 0xd0, 0x06, 0xe6,
|
||||
0x30, 0x07, 0x72, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x60, 0x0e, 0x76, 0x40, 0x07, 0x7a, 0x60, 0x07, 0x74, 0xd0, 0x06, 0xee, 0x80, 0x07, 0x7a, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x73, 0x20, 0x07,
|
||||
0x7a, 0x60, 0x07, 0x74, 0x30, 0xe4, 0x09, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x60, 0xc8, 0x43, 0x00, 0x01, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc0, 0x90,
|
||||
0x67, 0x01, 0x02, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x21, 0x8f, 0x03, 0x04, 0xc0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x43, 0x9e, 0x08, 0x08, 0x00, 0x01, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xb2, 0x40, 0x0b, 0x00, 0x00, 0x00, 0x32, 0x1e, 0x98, 0x10, 0x19, 0x11, 0x4c, 0x90, 0x8c, 0x09, 0x26, 0x47, 0xc6, 0x04, 0x43, 0x1a, 0x25, 0x50, 0x04, 0xc5,
|
||||
0x30, 0x02, 0x50, 0x18, 0x85, 0x40, 0x70, 0x04, 0x80, 0x72, 0x81, 0x03, 0x02, 0x22, 0x10, 0x9e, 0x01, 0xa0, 0x3a, 0x03, 0x00, 0x00, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00, 0x3f, 0x00, 0x00, 0x00,
|
||||
0x1a, 0x03, 0x4c, 0x90, 0x46, 0x02, 0x13, 0x44, 0x35, 0x18, 0x63, 0x0b, 0x73, 0x3b, 0x03, 0xb1, 0x2b, 0x93, 0x9b, 0x4b, 0x7b, 0x73, 0x03, 0x99, 0x71, 0xb9, 0x01, 0x41, 0xa1, 0x0b, 0x3b, 0x9b,
|
||||
0x7b, 0x91, 0x2a, 0x62, 0x2a, 0x0a, 0x9a, 0x2a, 0xfa, 0x9a, 0xb9, 0x81, 0x79, 0x31, 0x4b, 0x73, 0x0b, 0x63, 0x4b, 0xd9, 0x10, 0x04, 0x13, 0x84, 0xa1, 0x98, 0x20, 0x0c, 0xc6, 0x06, 0x61, 0x20,
|
||||
0x26, 0x08, 0xc3, 0xb1, 0x41, 0x18, 0x0c, 0x0a, 0x63, 0x73, 0x1b, 0x06, 0xc4, 0x20, 0x26, 0x08, 0x03, 0x32, 0x41, 0xa8, 0x1c, 0x02, 0x13, 0x84, 0x21, 0x99, 0x20, 0x0c, 0xca, 0x06, 0x61, 0x70,
|
||||
0x36, 0x24, 0xca, 0xc2, 0x28, 0xca, 0xd0, 0x28, 0xcf, 0x86, 0x00, 0x9a, 0x20, 0x60, 0xcd, 0x04, 0x01, 0x5a, 0x36, 0x2c, 0x8a, 0xc4, 0x28, 0xca, 0xd0, 0x4c, 0xd3, 0xf4, 0x6c, 0x08, 0xa8, 0x0d,
|
||||
0x44, 0x54, 0x01, 0xc0, 0x04, 0x41, 0x00, 0x48, 0xb4, 0x85, 0xa5, 0xb9, 0x4d, 0x10, 0x32, 0x66, 0xc3, 0xe0, 0x0c, 0xc3, 0x06, 0x42, 0xc9, 0x1c, 0x6d, 0x43, 0x71, 0x61, 0x80, 0xb5, 0x55, 0x61,
|
||||
0x63, 0xb3, 0x6b, 0x73, 0x49, 0x23, 0x2b, 0x73, 0xa3, 0x9b, 0x12, 0x04, 0x55, 0xc8, 0xf0, 0x5c, 0xec, 0xca, 0xe4, 0xe6, 0xd2, 0xde, 0xdc, 0xa6, 0x04, 0x44, 0x13, 0x32, 0x3c, 0x17, 0xbb, 0x30,
|
||||
0x36, 0xbb, 0x32, 0xb9, 0x29, 0x81, 0x51, 0x87, 0x0c, 0xcf, 0x65, 0x0e, 0x2d, 0x8c, 0xac, 0x4c, 0xae, 0xe9, 0x8d, 0xac, 0x8c, 0x6d, 0x4a, 0x80, 0x94, 0x21, 0xc3, 0x73, 0x91, 0x2b, 0x9b, 0x7b,
|
||||
0xab, 0x93, 0x1b, 0x2b, 0x9b, 0x9b, 0x12, 0x54, 0x75, 0xc8, 0xf0, 0x5c, 0xca, 0xdc, 0xe8, 0xe4, 0xf2, 0xa0, 0xde, 0xd2, 0xdc, 0xe8, 0xe6, 0xa6, 0x04, 0x1b, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00,
|
||||
0x50, 0x00, 0x00, 0x00, 0x33, 0x08, 0x80, 0x1c, 0xc4, 0xe1, 0x1c, 0x66, 0x14, 0x01, 0x3d, 0x88, 0x43, 0x38, 0x84, 0xc3, 0x8c, 0x42, 0x80, 0x07, 0x79, 0x78, 0x07, 0x73, 0x98, 0x71, 0x0c, 0xe6,
|
||||
0x00, 0x0f, 0xed, 0x10, 0x0e, 0xf4, 0x80, 0x0e, 0x33, 0x0c, 0x42, 0x1e, 0xc2, 0xc1, 0x1d, 0xce, 0xa1, 0x1c, 0x66, 0x30, 0x05, 0x3d, 0x88, 0x43, 0x38, 0x84, 0x83, 0x1b, 0xcc, 0x03, 0x3d, 0xc8,
|
||||
0x43, 0x3d, 0x8c, 0x03, 0x3d, 0xcc, 0x78, 0x8c, 0x74, 0x70, 0x07, 0x7b, 0x08, 0x07, 0x79, 0x48, 0x87, 0x70, 0x70, 0x07, 0x7a, 0x70, 0x03, 0x76, 0x78, 0x87, 0x70, 0x20, 0x87, 0x19, 0xcc, 0x11,
|
||||
0x0e, 0xec, 0x90, 0x0e, 0xe1, 0x30, 0x0f, 0x6e, 0x30, 0x0f, 0xe3, 0xf0, 0x0e, 0xf0, 0x50, 0x0e, 0x33, 0x10, 0xc4, 0x1d, 0xde, 0x21, 0x1c, 0xd8, 0x21, 0x1d, 0xc2, 0x61, 0x1e, 0x66, 0x30, 0x89,
|
||||
0x3b, 0xbc, 0x83, 0x3b, 0xd0, 0x43, 0x39, 0xb4, 0x03, 0x3c, 0xbc, 0x83, 0x3c, 0x84, 0x03, 0x3b, 0xcc, 0xf0, 0x14, 0x76, 0x60, 0x07, 0x7b, 0x68, 0x07, 0x37, 0x68, 0x87, 0x72, 0x68, 0x07, 0x37,
|
||||
0x80, 0x87, 0x70, 0x90, 0x87, 0x70, 0x60, 0x07, 0x76, 0x28, 0x07, 0x76, 0xf8, 0x05, 0x76, 0x78, 0x87, 0x77, 0x80, 0x87, 0x5f, 0x08, 0x87, 0x71, 0x18, 0x87, 0x72, 0x98, 0x87, 0x79, 0x98, 0x81,
|
||||
0x2c, 0xee, 0xf0, 0x0e, 0xee, 0xe0, 0x0e, 0xf5, 0xc0, 0x0e, 0xec, 0x30, 0x03, 0x62, 0xc8, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xdc, 0x61, 0x1c, 0xca, 0x21, 0x1c,
|
||||
0xc4, 0x81, 0x1d, 0xca, 0x61, 0x06, 0xd6, 0x90, 0x43, 0x39, 0xc8, 0x43, 0x39, 0x98, 0x43, 0x39, 0xc8, 0x43, 0x39, 0xb8, 0xc3, 0x38, 0x94, 0x43, 0x38, 0x88, 0x03, 0x3b, 0x94, 0xc3, 0x2f, 0xbc,
|
||||
0x83, 0x3c, 0xfc, 0x82, 0x3b, 0xd4, 0x03, 0x3b, 0xb0, 0xc3, 0x0c, 0xc4, 0x21, 0x07, 0x7c, 0x70, 0x03, 0x7a, 0x28, 0x87, 0x76, 0x80, 0x87, 0x19, 0xcb, 0x43, 0x0e, 0xf8, 0xe0, 0x06, 0xf0, 0x20,
|
||||
0x0f, 0xe5, 0x30, 0x0e, 0xe9, 0x30, 0x0f, 0xe5, 0x30, 0x23, 0x82, 0xc8, 0x01, 0x1f, 0xdc, 0x80, 0x1c, 0xc4, 0xe1, 0x1c, 0xdc, 0xc0, 0x1e, 0xc2, 0x41, 0x1e, 0xd8, 0x21, 0x1c, 0xf2, 0xe1, 0x1d,
|
||||
0xea, 0x81, 0x1e, 0x00, 0x71, 0x20, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x46, 0x20, 0x0d, 0x97, 0xef, 0x3c, 0xbe, 0x10, 0x11, 0xc0, 0x44, 0x84, 0x40, 0x33, 0x2c, 0x84, 0x09, 0x5c, 0xc3, 0xe5,
|
||||
0x3b, 0x8f, 0x1f, 0x01, 0xd6, 0x46, 0x15, 0x05, 0x11, 0x95, 0x0e, 0x30, 0xf8, 0xc5, 0x6d, 0xdb, 0x00, 0x36, 0x5c, 0xbe, 0xf3, 0xf8, 0x11, 0x60, 0x6d, 0x54, 0x51, 0x10, 0x11, 0x3b, 0x39, 0x11,
|
||||
0xe1, 0x17, 0xb5, 0x6e, 0x01, 0xd2, 0x70, 0xf9, 0xce, 0xe3, 0x4f, 0x47, 0x44, 0x00, 0x83, 0x38, 0xf8, 0xc8, 0x6d, 0x1b, 0x00, 0xc1, 0x00, 0x48, 0x03, 0x00, 0x00, 0x00, 0x61, 0x20, 0x00, 0x00,
|
||||
0x1b, 0x00, 0x00, 0x00, 0x13, 0x04, 0x41, 0x2c, 0x10, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x34, 0x4a, 0xae, 0x74, 0x03, 0xca, 0x52, 0xa0, 0x30, 0x05, 0x08, 0x95, 0xc0, 0x08, 0x00, 0xb5,
|
||||
0x19, 0x80, 0x31, 0x02, 0x10, 0x74, 0x00, 0x00, 0x23, 0x06, 0x09, 0x00, 0x82, 0x60, 0x10, 0x4d, 0x08, 0xf1, 0x3c, 0xcc, 0x88, 0x41, 0x02, 0x80, 0x20, 0x18, 0x44, 0x54, 0x42, 0x40, 0x50, 0x33,
|
||||
0x62, 0x60, 0x00, 0x20, 0x08, 0x06, 0x04, 0x96, 0x44, 0x23, 0x06, 0x0a, 0x00, 0x82, 0x60, 0xb0, 0x60, 0x89, 0x10, 0x48, 0x07, 0x34, 0x9a, 0x10, 0x00, 0x17, 0xb4, 0xab, 0xe0, 0x80, 0x11, 0x03,
|
||||
0x07, 0x00, 0x41, 0x30, 0x70, 0x36, 0xe6, 0x28, 0xac, 0x20, 0x49, 0x92, 0x05, 0x43, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
};
|
||||
|
||||
static const BYTE fp16_arith_denorm_code[] =
|
||||
{
|
||||
#if 0
|
||||
StructuredBuffer<float> Buf : register(t0);
|
||||
RWStructuredBuffer<half2> RWBuf : register(u0);
|
||||
|
||||
[numthreads(4, 1, 1)]
|
||||
void main(uint thr : SV_DispatchThreadID)
|
||||
{
|
||||
precise half result = half(Buf[thr]) - half(Buf[thr ^ 3]);
|
||||
RWBuf[thr].x = result;
|
||||
}
|
||||
#endif
|
||||
0x44, 0x58, 0x42, 0x43, 0x3f, 0x0c, 0x9c, 0xb8, 0xe1, 0xb4, 0x19, 0x4a, 0x0d, 0xa9, 0xd2, 0x4f, 0xdc, 0xa5, 0xee, 0xac, 0x01, 0x00, 0x00, 0x00, 0xe4, 0x06, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00,
|
||||
0x38, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, 0xe8, 0x00, 0x00, 0x00, 0x04, 0x01, 0x00, 0x00, 0x53, 0x46, 0x49, 0x30, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x49, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x4f, 0x53, 0x47, 0x31, 0x08, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x50, 0x53, 0x56, 0x30, 0x78, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
|
||||
0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x48, 0x41, 0x53, 0x48, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe3, 0x01, 0xf8, 0xb0, 0x5f, 0xe7, 0x66, 0x86, 0xb3, 0xc7, 0xdf, 0xa4,
|
||||
0x05, 0x84, 0xe1, 0x4e, 0x44, 0x58, 0x49, 0x4c, 0xd8, 0x05, 0x00, 0x00, 0x62, 0x00, 0x05, 0x00, 0x76, 0x01, 0x00, 0x00, 0x44, 0x58, 0x49, 0x4c, 0x02, 0x01, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
|
||||
0xc0, 0x05, 0x00, 0x00, 0x42, 0x43, 0xc0, 0xde, 0x21, 0x0c, 0x00, 0x00, 0x6d, 0x01, 0x00, 0x00, 0x0b, 0x82, 0x20, 0x00, 0x02, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, 0x07, 0x81, 0x23, 0x91,
|
||||
0x41, 0xc8, 0x04, 0x49, 0x06, 0x10, 0x32, 0x39, 0x92, 0x01, 0x84, 0x0c, 0x25, 0x05, 0x08, 0x19, 0x1e, 0x04, 0x8b, 0x62, 0x80, 0x14, 0x45, 0x02, 0x42, 0x92, 0x0b, 0x42, 0xa4, 0x10, 0x32, 0x14,
|
||||
0x38, 0x08, 0x18, 0x4b, 0x0a, 0x32, 0x52, 0x88, 0x48, 0x90, 0x14, 0x20, 0x43, 0x46, 0x88, 0xa5, 0x00, 0x19, 0x32, 0x42, 0xe4, 0x48, 0x0e, 0x90, 0x91, 0x22, 0xc4, 0x50, 0x41, 0x51, 0x81, 0x8c,
|
||||
0xe1, 0x83, 0xe5, 0x8a, 0x04, 0x29, 0x46, 0x06, 0x51, 0x18, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x1b, 0x8c, 0xe0, 0xff, 0xff, 0xff, 0xff, 0x07, 0x40, 0x02, 0xa8, 0x0d, 0x86, 0xf0, 0xff, 0xff,
|
||||
0xff, 0xff, 0x03, 0x20, 0x01, 0xd5, 0x06, 0x62, 0xf8, 0xff, 0xff, 0xff, 0xff, 0x01, 0x90, 0x00, 0x49, 0x18, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x13, 0x82, 0x60, 0x42, 0x20, 0x4c, 0x08, 0x06,
|
||||
0x00, 0x00, 0x00, 0x00, 0x89, 0x20, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, 0x32, 0x22, 0x48, 0x09, 0x20, 0x64, 0x85, 0x04, 0x93, 0x22, 0xa4, 0x84, 0x04, 0x93, 0x22, 0xe3, 0x84, 0xa1, 0x90, 0x14,
|
||||
0x12, 0x4c, 0x8a, 0x8c, 0x0b, 0x84, 0xa4, 0x4c, 0x10, 0x68, 0x23, 0x00, 0x25, 0x00, 0x14, 0xe6, 0x08, 0xc0, 0xa0, 0x0c, 0x63, 0x0c, 0x22, 0x33, 0x00, 0x37, 0x0d, 0x97, 0x3f, 0x61, 0x0f, 0x21,
|
||||
0xf9, 0x2b, 0x21, 0xad, 0xc4, 0xe4, 0x17, 0xb7, 0x8d, 0x0a, 0x63, 0x8c, 0x19, 0x73, 0x04, 0x08, 0xa1, 0x7b, 0x86, 0xcb, 0x9f, 0xb0, 0x87, 0x90, 0xfc, 0x10, 0x68, 0x86, 0x85, 0x40, 0x41, 0x2a,
|
||||
0xc7, 0x19, 0x6a, 0x0c, 0x34, 0x68, 0x8d, 0x02, 0x94, 0x05, 0x0c, 0x35, 0x46, 0x6b, 0xad, 0xa1, 0x41, 0x6e, 0x8e, 0x20, 0x28, 0x86, 0x1a, 0x68, 0x0c, 0x48, 0x71, 0x20, 0x60, 0xa6, 0x6e, 0x1c,
|
||||
0xd8, 0x21, 0x1c, 0xe6, 0x61, 0x1e, 0xdc, 0x60, 0x16, 0xe8, 0x41, 0x1e, 0xea, 0x61, 0x1c, 0xe8, 0xa1, 0x1e, 0xe4, 0xa1, 0x1c, 0xc8, 0x41, 0x14, 0xea, 0xc1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0x81,
|
||||
0x0f, 0xcc, 0x81, 0x1d, 0xde, 0x21, 0x1c, 0xe8, 0xc1, 0x0f, 0x50, 0x60, 0x88, 0x0e, 0x23, 0x08, 0xcd, 0x4c, 0x6a, 0x30, 0x0e, 0xec, 0x10, 0x0e, 0xf3, 0x30, 0x0f, 0x6e, 0x20, 0x0b, 0xb7, 0x30,
|
||||
0x0b, 0xf4, 0x20, 0x0f, 0xf5, 0x30, 0x0e, 0xf4, 0x50, 0x0f, 0xf2, 0x50, 0x0e, 0xe4, 0x20, 0x0a, 0xf5, 0x60, 0x0e, 0xe6, 0x50, 0x0e, 0xf2, 0xc0, 0x07, 0xf6, 0x50, 0x0e, 0xe3, 0x40, 0x0f, 0xef,
|
||||
0x20, 0x0f, 0x7c, 0x80, 0x0e, 0xe1, 0xc0, 0x0e, 0xe6, 0xc0, 0x06, 0x60, 0x20, 0x07, 0x7e, 0x00, 0x06, 0x7e, 0x80, 0x02, 0x4b, 0x77, 0x8e, 0x00, 0x14, 0x00, 0x00, 0x00, 0x13, 0x14, 0x72, 0xc0,
|
||||
0x87, 0x74, 0x60, 0x87, 0x36, 0x68, 0x87, 0x79, 0x68, 0x03, 0x72, 0xc0, 0x87, 0x0d, 0xae, 0x50, 0x0e, 0x6d, 0xd0, 0x0e, 0x7a, 0x50, 0x0e, 0x6d, 0x00, 0x0f, 0x7a, 0x30, 0x07, 0x72, 0xa0, 0x07,
|
||||
0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x71, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x90, 0x0e, 0x78, 0xa0, 0x07, 0x78, 0xd0, 0x06, 0xe9, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x71, 0x60, 0x07, 0x6d, 0x90,
|
||||
0x0e, 0x73, 0x20, 0x07, 0x7a, 0x30, 0x07, 0x72, 0xd0, 0x06, 0xe9, 0x60, 0x07, 0x74, 0xa0, 0x07, 0x76, 0x40, 0x07, 0x6d, 0x60, 0x0e, 0x71, 0x60, 0x07, 0x7a, 0x10, 0x07, 0x76, 0xd0, 0x06, 0xe6,
|
||||
0x30, 0x07, 0x72, 0xa0, 0x07, 0x73, 0x20, 0x07, 0x6d, 0x60, 0x0e, 0x76, 0x40, 0x07, 0x7a, 0x60, 0x07, 0x74, 0xd0, 0x06, 0xee, 0x80, 0x07, 0x7a, 0x10, 0x07, 0x76, 0xa0, 0x07, 0x73, 0x20, 0x07,
|
||||
0x7a, 0x60, 0x07, 0x74, 0x30, 0xe4, 0x09, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x60, 0xc8, 0x43, 0x00, 0x01, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc0, 0x90,
|
||||
0x67, 0x01, 0x02, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x21, 0x8f, 0x03, 0x04, 0xc0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x43, 0x9e, 0x08, 0x08, 0x00, 0x01, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xb2, 0x40, 0x0b, 0x00, 0x00, 0x00, 0x32, 0x1e, 0x98, 0x10, 0x19, 0x11, 0x4c, 0x90, 0x8c, 0x09, 0x26, 0x47, 0xc6, 0x04, 0x43, 0x1a, 0x25, 0x50, 0x04, 0xc5,
|
||||
0x30, 0x02, 0x50, 0x18, 0x85, 0x40, 0x70, 0x04, 0x80, 0x72, 0x81, 0x03, 0x02, 0x22, 0x10, 0x9e, 0x01, 0xa0, 0x3a, 0x03, 0x00, 0x00, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00, 0x3f, 0x00, 0x00, 0x00,
|
||||
0x1a, 0x03, 0x4c, 0x90, 0x46, 0x02, 0x13, 0x44, 0x35, 0x18, 0x63, 0x0b, 0x73, 0x3b, 0x03, 0xb1, 0x2b, 0x93, 0x9b, 0x4b, 0x7b, 0x73, 0x03, 0x99, 0x71, 0xb9, 0x01, 0x41, 0xa1, 0x0b, 0x3b, 0x9b,
|
||||
0x7b, 0x91, 0x2a, 0x62, 0x2a, 0x0a, 0x9a, 0x2a, 0xfa, 0x9a, 0xb9, 0x81, 0x79, 0x31, 0x4b, 0x73, 0x0b, 0x63, 0x4b, 0xd9, 0x10, 0x04, 0x13, 0x84, 0xa1, 0x98, 0x20, 0x0c, 0xc6, 0x06, 0x61, 0x20,
|
||||
0x26, 0x08, 0xc3, 0xb1, 0x41, 0x18, 0x0c, 0x0a, 0x63, 0x73, 0x1b, 0x06, 0xc4, 0x20, 0x26, 0x08, 0x03, 0x32, 0x41, 0xa8, 0x1c, 0x02, 0x13, 0x84, 0x21, 0x99, 0x20, 0x0c, 0xca, 0x06, 0x61, 0x70,
|
||||
0x36, 0x24, 0xca, 0xc2, 0x28, 0xca, 0xd0, 0x28, 0xcf, 0x86, 0x00, 0x9a, 0x20, 0x60, 0xcd, 0x04, 0x01, 0x5a, 0x36, 0x2c, 0x8a, 0xc4, 0x28, 0xca, 0xd0, 0x4c, 0xd3, 0xf4, 0x6c, 0x08, 0xa8, 0x0d,
|
||||
0x44, 0x54, 0x01, 0xc0, 0x04, 0x41, 0x00, 0x48, 0xb4, 0x85, 0xa5, 0xb9, 0x4d, 0x10, 0x32, 0x66, 0xc3, 0xe0, 0x0c, 0xc3, 0x06, 0x42, 0xc9, 0x1c, 0x6d, 0x43, 0x71, 0x61, 0x80, 0xb5, 0x55, 0x61,
|
||||
0x63, 0xb3, 0x6b, 0x73, 0x49, 0x23, 0x2b, 0x73, 0xa3, 0x9b, 0x12, 0x04, 0x55, 0xc8, 0xf0, 0x5c, 0xec, 0xca, 0xe4, 0xe6, 0xd2, 0xde, 0xdc, 0xa6, 0x04, 0x44, 0x13, 0x32, 0x3c, 0x17, 0xbb, 0x30,
|
||||
0x36, 0xbb, 0x32, 0xb9, 0x29, 0x81, 0x51, 0x87, 0x0c, 0xcf, 0x65, 0x0e, 0x2d, 0x8c, 0xac, 0x4c, 0xae, 0xe9, 0x8d, 0xac, 0x8c, 0x6d, 0x4a, 0x80, 0x94, 0x21, 0xc3, 0x73, 0x91, 0x2b, 0x9b, 0x7b,
|
||||
0xab, 0x93, 0x1b, 0x2b, 0x9b, 0x9b, 0x12, 0x54, 0x75, 0xc8, 0xf0, 0x5c, 0xca, 0xdc, 0xe8, 0xe4, 0xf2, 0xa0, 0xde, 0xd2, 0xdc, 0xe8, 0xe6, 0xa6, 0x04, 0x1b, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00,
|
||||
0x50, 0x00, 0x00, 0x00, 0x33, 0x08, 0x80, 0x1c, 0xc4, 0xe1, 0x1c, 0x66, 0x14, 0x01, 0x3d, 0x88, 0x43, 0x38, 0x84, 0xc3, 0x8c, 0x42, 0x80, 0x07, 0x79, 0x78, 0x07, 0x73, 0x98, 0x71, 0x0c, 0xe6,
|
||||
0x00, 0x0f, 0xed, 0x10, 0x0e, 0xf4, 0x80, 0x0e, 0x33, 0x0c, 0x42, 0x1e, 0xc2, 0xc1, 0x1d, 0xce, 0xa1, 0x1c, 0x66, 0x30, 0x05, 0x3d, 0x88, 0x43, 0x38, 0x84, 0x83, 0x1b, 0xcc, 0x03, 0x3d, 0xc8,
|
||||
0x43, 0x3d, 0x8c, 0x03, 0x3d, 0xcc, 0x78, 0x8c, 0x74, 0x70, 0x07, 0x7b, 0x08, 0x07, 0x79, 0x48, 0x87, 0x70, 0x70, 0x07, 0x7a, 0x70, 0x03, 0x76, 0x78, 0x87, 0x70, 0x20, 0x87, 0x19, 0xcc, 0x11,
|
||||
0x0e, 0xec, 0x90, 0x0e, 0xe1, 0x30, 0x0f, 0x6e, 0x30, 0x0f, 0xe3, 0xf0, 0x0e, 0xf0, 0x50, 0x0e, 0x33, 0x10, 0xc4, 0x1d, 0xde, 0x21, 0x1c, 0xd8, 0x21, 0x1d, 0xc2, 0x61, 0x1e, 0x66, 0x30, 0x89,
|
||||
0x3b, 0xbc, 0x83, 0x3b, 0xd0, 0x43, 0x39, 0xb4, 0x03, 0x3c, 0xbc, 0x83, 0x3c, 0x84, 0x03, 0x3b, 0xcc, 0xf0, 0x14, 0x76, 0x60, 0x07, 0x7b, 0x68, 0x07, 0x37, 0x68, 0x87, 0x72, 0x68, 0x07, 0x37,
|
||||
0x80, 0x87, 0x70, 0x90, 0x87, 0x70, 0x60, 0x07, 0x76, 0x28, 0x07, 0x76, 0xf8, 0x05, 0x76, 0x78, 0x87, 0x77, 0x80, 0x87, 0x5f, 0x08, 0x87, 0x71, 0x18, 0x87, 0x72, 0x98, 0x87, 0x79, 0x98, 0x81,
|
||||
0x2c, 0xee, 0xf0, 0x0e, 0xee, 0xe0, 0x0e, 0xf5, 0xc0, 0x0e, 0xec, 0x30, 0x03, 0x62, 0xc8, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xcc, 0xa1, 0x1c, 0xe4, 0xa1, 0x1c, 0xdc, 0x61, 0x1c, 0xca, 0x21, 0x1c,
|
||||
0xc4, 0x81, 0x1d, 0xca, 0x61, 0x06, 0xd6, 0x90, 0x43, 0x39, 0xc8, 0x43, 0x39, 0x98, 0x43, 0x39, 0xc8, 0x43, 0x39, 0xb8, 0xc3, 0x38, 0x94, 0x43, 0x38, 0x88, 0x03, 0x3b, 0x94, 0xc3, 0x2f, 0xbc,
|
||||
0x83, 0x3c, 0xfc, 0x82, 0x3b, 0xd4, 0x03, 0x3b, 0xb0, 0xc3, 0x0c, 0xc4, 0x21, 0x07, 0x7c, 0x70, 0x03, 0x7a, 0x28, 0x87, 0x76, 0x80, 0x87, 0x19, 0xcb, 0x43, 0x0e, 0xf8, 0xe0, 0x06, 0xf0, 0x20,
|
||||
0x0f, 0xe5, 0x30, 0x0e, 0xe9, 0x30, 0x0f, 0xe5, 0x30, 0x23, 0x82, 0xc8, 0x01, 0x1f, 0xdc, 0x80, 0x1c, 0xc4, 0xe1, 0x1c, 0xdc, 0xc0, 0x1e, 0xc2, 0x41, 0x1e, 0xd8, 0x21, 0x1c, 0xf2, 0xe1, 0x1d,
|
||||
0xea, 0x81, 0x1e, 0x00, 0x71, 0x20, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x46, 0x20, 0x0d, 0x97, 0xef, 0x3c, 0xbe, 0x10, 0x11, 0xc0, 0x44, 0x84, 0x40, 0x33, 0x2c, 0x84, 0x09, 0x5c, 0xc3, 0xe5,
|
||||
0x3b, 0x8f, 0x1f, 0x01, 0xd6, 0x46, 0x15, 0x05, 0x11, 0x95, 0x0e, 0x30, 0xf8, 0xc5, 0x6d, 0xdb, 0x00, 0x36, 0x5c, 0xbe, 0xf3, 0xf8, 0x11, 0x60, 0x6d, 0x54, 0x51, 0x10, 0x11, 0x3b, 0x39, 0x11,
|
||||
0xe1, 0x17, 0xb5, 0x6e, 0x01, 0xd2, 0x70, 0xf9, 0xce, 0xe3, 0x4f, 0x47, 0x44, 0x00, 0x83, 0x38, 0xf8, 0xc8, 0x6d, 0x1b, 0x00, 0xc1, 0x00, 0x48, 0x03, 0x00, 0x00, 0x00, 0x61, 0x20, 0x00, 0x00,
|
||||
0x1f, 0x00, 0x00, 0x00, 0x13, 0x04, 0x41, 0x2c, 0x10, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x34, 0x4a, 0xae, 0x2c, 0x05, 0x4a, 0x37, 0xa0, 0x0c, 0x0a, 0x53, 0x80, 0x50, 0x09, 0x8c, 0x00,
|
||||
0x50, 0x9b, 0x01, 0x00, 0x23, 0x06, 0x09, 0x00, 0x82, 0x60, 0x10, 0x4d, 0xc8, 0xf0, 0x3c, 0xcc, 0x88, 0x41, 0x02, 0x80, 0x20, 0x18, 0x44, 0x54, 0x32, 0x40, 0x50, 0x33, 0x62, 0x60, 0x00, 0x20,
|
||||
0x08, 0x06, 0x04, 0x86, 0x44, 0x23, 0x06, 0x0a, 0x00, 0x82, 0x60, 0xb0, 0x60, 0x8a, 0x10, 0x48, 0x06, 0x34, 0x9a, 0x10, 0x00, 0x17, 0xb4, 0x2b, 0x62, 0xe1, 0x11, 0x03, 0x05, 0x00, 0x41, 0x30,
|
||||
0x58, 0x38, 0xc7, 0x08, 0x2c, 0x85, 0x1a, 0x4d, 0x08, 0x80, 0x0b, 0xda, 0x55, 0x11, 0xc4, 0x88, 0x81, 0x03, 0x80, 0x20, 0x18, 0x38, 0xdf, 0xb3, 0x24, 0x5a, 0xc0, 0x30, 0x8c, 0xc3, 0x21, 0x00,
|
||||
0x00, 0x00, 0x00, 0x00,
|
||||
};
|
||||
|
||||
static const D3D12_SHADER_BYTECODE legacy_f32_to_f16 = SHADER_BYTECODE(legacy_f32_to_f16_code);
|
||||
static const D3D12_SHADER_BYTECODE fptrunc_f32_to_f16 = SHADER_BYTECODE(fptrunc_f32_to_f16_code);
|
||||
static const D3D12_SHADER_BYTECODE fp16_add_rounding_mode = SHADER_BYTECODE(fp16_add_rounding_mode_code);
|
||||
static const D3D12_SHADER_BYTECODE fp16_arith_denorm = SHADER_BYTECODE(fp16_arith_denorm_code);
|
||||
|
||||
union
|
||||
{
|
||||
uint32_t u32[4];
|
||||
float f32[4];
|
||||
} patterns;
|
||||
|
||||
struct test tests[] =
|
||||
{
|
||||
/* LegacyF32toF16 opcode is always RTZ. */
|
||||
{&legacy_f32_to_f16, { 1024.75f, 1025.75f, 1026.75f, 1027.75f }, { 0x6400, 0x6401, 0x6402, 0x6403 }, false, true},
|
||||
{&legacy_f32_to_f16, { -1024.75f, -1025.75f, -1026.75f, -1027.75f }, { 0xe400, 0xe401, 0xe402, 0xe403 }, false, true},
|
||||
{&legacy_f32_to_f16, { 0.75f / 0x1000000, -0.75f / 0x1000000, 1.75f / 0x1000000, -1.75f / 0x1000000 }, { 0x0000, 0x8000, 0x0001, 0x8001 }, false, true},
|
||||
{&legacy_f32_to_f16, { 0.0f /* To be filled in with weird values, +inf, -inf, nan(1), nan(all bits). */ }, {0x7c00, 0xfc00, UINT32_MAX /* NaN, any pattern allowed. */, UINT32_MAX /* NaN, any pattern allowed. */}, false},
|
||||
/* Drivers don't seem to agree if this should be RTZ or RTE ... Accept both results. Unclear if this is even well specified in D3D12. */
|
||||
{&fptrunc_f32_to_f16, { 1024.5f, 1025.5f, 1026.5f, 1027.5f }, { 0x6400, 0x6402, 0x6402, 0x6404 }, true, false, { 0x6400, 0x6401, 0x6402, 0x6403 }, true},
|
||||
{&fptrunc_f32_to_f16, { 0.0f /* To be filled in with weird values, +inf, -inf, nan(1), nan(all bits). */ }, {0x7c00, 0xfc00, UINT32_MAX /* NaN, any pattern allowed. */, UINT32_MAX /* NaN, any pattern allowed. */}, true},
|
||||
/* Arithmetic is RTE. */
|
||||
{&fp16_add_rounding_mode, { 1024.0f, 1025.0f, 1026.0f, 1027.0f }, { 0x6400, 0x6402, 0x6402, 0x6404 }, true},
|
||||
/* Denorm is preserved in arithmetic. */
|
||||
{&fp16_arith_denorm, { 1.0f / 0x1000000, 2.0f / 0x1000000, 3.0f / 0x1000000, 4.0f / 0x1000000 }, { 0x8003, 0x8001, 0x0001, 0x0003 }, true},
|
||||
};
|
||||
|
||||
/* RTZ tests are TODO since we have no direct way of implementing it. */
|
||||
|
||||
patterns.u32[0] = 0x7f800000;
|
||||
patterns.u32[1] = 0xff800000;
|
||||
patterns.u32[2] = 0x7f800001;
|
||||
patterns.u32[3] = 0x7fffffff;
|
||||
memcpy(tests[3].input_data, patterns.f32, sizeof(patterns.f32));
|
||||
memcpy(tests[5].input_data, patterns.f32, sizeof(patterns.f32));
|
||||
|
||||
if (!init_compute_test_context(&context))
|
||||
return;
|
||||
|
||||
if (!context_supports_dxil(&context))
|
||||
{
|
||||
skip("DXIL not supported.\n");
|
||||
destroy_test_context(&context);
|
||||
return;
|
||||
}
|
||||
|
||||
memset(&options4, 0, sizeof(options4));
|
||||
if (FAILED(ID3D12Device_CheckFeatureSupport(context.device, D3D12_FEATURE_D3D12_OPTIONS4, &options4, sizeof(options4))))
|
||||
options4.Native16BitShaderOpsSupported = FALSE;
|
||||
|
||||
memset(&rs_desc, 0, sizeof(rs_desc));
|
||||
memset(root_parameters, 0, sizeof(root_parameters));
|
||||
rs_desc.NumParameters = ARRAY_SIZE(root_parameters);
|
||||
rs_desc.pParameters = root_parameters;
|
||||
root_parameters[0].ParameterType = D3D12_ROOT_PARAMETER_TYPE_SRV;
|
||||
root_parameters[0].ShaderVisibility = D3D12_SHADER_VISIBILITY_ALL;
|
||||
root_parameters[1].ParameterType = D3D12_ROOT_PARAMETER_TYPE_UAV;
|
||||
root_parameters[1].ShaderVisibility = D3D12_SHADER_VISIBILITY_ALL;
|
||||
create_root_signature(context.device, &rs_desc, &root_signature);
|
||||
|
||||
for (i = 0; i < ARRAY_SIZE(tests); i++)
|
||||
{
|
||||
vkd3d_test_set_context("Test %u", i);
|
||||
|
||||
if (!options4.Native16BitShaderOpsSupported && tests[i].native_fp16)
|
||||
{
|
||||
skip("Native FP16 not supported, skipping.\n");
|
||||
continue;
|
||||
}
|
||||
|
||||
pso = create_compute_pipeline_state(context.device, root_signature, *tests[i].cs);
|
||||
input_buffer = create_upload_buffer(context.device, sizeof(tests[i].input_data), tests[i].input_data);
|
||||
output_buffer = create_default_buffer(context.device, sizeof(tests[i].output_data),
|
||||
D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS,
|
||||
D3D12_RESOURCE_STATE_UNORDERED_ACCESS);
|
||||
|
||||
ID3D12GraphicsCommandList_SetComputeRootSignature(context.list, root_signature);
|
||||
ID3D12GraphicsCommandList_SetPipelineState(context.list, pso);
|
||||
ID3D12GraphicsCommandList_SetComputeRootShaderResourceView(context.list, 0,
|
||||
ID3D12Resource_GetGPUVirtualAddress(input_buffer));
|
||||
ID3D12GraphicsCommandList_SetComputeRootUnorderedAccessView(context.list, 1,
|
||||
ID3D12Resource_GetGPUVirtualAddress(output_buffer));
|
||||
ID3D12GraphicsCommandList_Dispatch(context.list, 1, 1, 1);
|
||||
|
||||
transition_resource_state(context.list, output_buffer,
|
||||
D3D12_RESOURCE_STATE_UNORDERED_ACCESS,
|
||||
D3D12_RESOURCE_STATE_COPY_SOURCE);
|
||||
|
||||
get_buffer_readback_with_command_list(output_buffer, DXGI_FORMAT_R32_UINT, &rb, context.queue, context.list);
|
||||
for (j = 0; j < ARRAY_SIZE(tests[i].output_data); j++)
|
||||
{
|
||||
uint32_t value = get_readback_uint(&rb, j, 0, 0);
|
||||
if (tests[i].allows_alternative_result && tests[i].output_data[j] != tests[i].output_data_alt[j])
|
||||
{
|
||||
todo_if(tests[i].is_todo)
|
||||
ok(value == tests[i].output_data[j] || value == tests[i].output_data_alt[j],
|
||||
"Value %u mismatch: %x != (%x or %x)\n",
|
||||
j, value, tests[i].output_data[j], tests[i].output_data_alt[j]);
|
||||
}
|
||||
else if (tests[i].output_data[j] == UINT32_MAX)
|
||||
{
|
||||
todo_if(tests[i].is_todo)
|
||||
ok((value & 0x7fff) > 0x7c00, "Value %u mismatch: Expected NaN, got %x.\n", value);
|
||||
}
|
||||
else
|
||||
{
|
||||
todo_if(tests[i].is_todo)
|
||||
ok(value == tests[i].output_data[j],
|
||||
"Value %u mismatch: %x != %x\n", j, value, tests[i].output_data[j]);
|
||||
}
|
||||
}
|
||||
ID3D12Resource_Release(input_buffer);
|
||||
ID3D12Resource_Release(output_buffer);
|
||||
ID3D12PipelineState_Release(pso);
|
||||
reset_command_list(context.list, context.allocator);
|
||||
release_resource_readback(&rb);
|
||||
}
|
||||
vkd3d_test_set_context(NULL);
|
||||
|
||||
ID3D12RootSignature_Release(root_signature);
|
||||
destroy_test_context(&context);
|
||||
}
|
||||
|
||||
void test_shader_instructions(void)
|
||||
{
|
||||
struct named_shader
|
||||
|
|
|
@ -74,6 +74,7 @@ decl_test(test_map_resource);
|
|||
decl_test(test_map_placed_resources);
|
||||
decl_test(test_bundle_state_inheritance);
|
||||
decl_test(test_shader_instructions);
|
||||
decl_test(test_shader_instructions_dxil);
|
||||
decl_test(test_compute_shader_instructions);
|
||||
decl_test(test_discard_instruction);
|
||||
decl_test(test_shader_interstage_interface);
|
||||
|
|
Loading…
Reference in New Issue