diff --git a/tests/d3d12.c b/tests/d3d12.c index 341e1981..1ca063d4 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -53172,6 +53172,224 @@ static void test_shader_fp16(void) destroy_test_context(&context); } +static void test_shader_sm62_denorm(void) +{ + D3D12_FEATURE_DATA_SHADER_MODEL shader_model; + D3D12_ROOT_PARAMETER root_parameters[2]; + D3D12_ROOT_SIGNATURE_DESC rs_desc; + struct test_context context; + struct resource_readback rb; + ID3D12PipelineState *pso; + ID3D12Resource *src; + ID3D12Resource *dst; + uint32_t value; + unsigned int i; + +#if 0 + RWStructuredBuffer RW : register(u0); + StructuredBuffer RO : register(t0); + + [numthreads(1, 1, 1)] + void main(uint3 id : SV_DispatchThreadID) + { + float v = RO[id.x]; + v += v; + RW[id.x] = v; + } +#endif + /* -denorm ftz */ + static const BYTE cs_code_ftz[] = + { + 0x44, 0x58, 0x42, 0x43, 0xd4, 0x0a, 0xa5, 0x22, 0xa2, 0xfb, 0x4b, 0xac, 0x99, 0x15, 0xd4, 0xd6, 0x5a, 0x2d, 0x26, 0x37, 0x01, 0x00, 0x00, 0x00, 0xa0, 0x06, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x34, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, 0x64, 0x00, 0x00, 0x00, 0xc8, 0x00, 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, 0x5c, 0x00, 0x00, 0x00, 0x24, 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, 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x44, 0x58, 0x49, 0x4c, 0xd0, 0x05, 0x00, 0x00, 0x62, 0x00, 0x05, 0x00, 0x74, 0x01, 0x00, 0x00, 0x44, 0x58, 0x49, 0x4c, 0x02, 0x01, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0xb8, 0x05, 0x00, 0x00, 0x42, 0x43, 0xc0, 0xde, 0x21, 0x0c, 0x00, 0x00, 0x6b, 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, 0x12, 0x00, 0x00, 0x00, 0x1b, 0xb0, 0xe0, 0xff, 0xff, 0xff, 0xff, 0x07, 0xc2, 0x1c, 0xe0, 0x61, + 0x0e, 0xe4, 0xa0, 0x0d, 0xc8, 0xa1, 0x1c, 0xdc, 0xe1, 0x1d, 0xe4, 0xa1, 0x1d, 0xda, 0xa0, 0x1d, 0xde, 0x81, 0x1c, 0xca, 0x01, 0x30, 0x07, 0x7a, 0xd0, 0x07, 0x60, 0x83, 0x21, 0xfc, 0xff, 0xff, + 0xff, 0xff, 0x00, 0x48, 0x00, 0xb5, 0x81, 0x18, 0xfe, 0xff, 0xff, 0xff, 0x7f, 0x00, 0xa4, 0x0d, 0x06, 0xf1, 0xff, 0xff, 0xff, 0xff, 0x03, 0x20, 0x01, 0x15, 0x00, 0x00, 0x49, 0x18, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x13, 0x82, 0x60, 0x42, 0x20, 0x4c, 0x08, 0x86, 0x09, 0x01, 0x01, 0x00, 0x89, 0x20, 0x00, 0x00, 0x36, 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, 0x60, 0x23, 0x00, 0x25, 0x00, 0x14, 0xe6, 0x08, 0xc0, 0xa0, 0x0c, + 0x63, 0x0c, 0x22, 0x73, 0x04, 0x08, 0x99, 0x7b, 0x86, 0xcb, 0x9f, 0xb0, 0x87, 0x90, 0xfc, 0x10, 0x68, 0x86, 0x85, 0x40, 0xc1, 0x99, 0x01, 0x28, 0x0b, 0x18, 0x68, 0x8c, 0x94, 0x52, 0x32, 0x83, + 0xd4, 0x4d, 0xc3, 0xe5, 0x4f, 0xd8, 0x43, 0x48, 0xfe, 0x4a, 0x48, 0x2b, 0x31, 0xf9, 0xc5, 0x6d, 0xa3, 0x22, 0xa5, 0x94, 0x46, 0x39, 0xd8, 0x40, 0x63, 0x98, 0x41, 0x6d, 0x8e, 0x20, 0x28, 0x06, + 0x1a, 0x66, 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, 0x90, 0x68, 0xce, 0xf4, 0x8d, 0x03, 0x3b, 0x84, 0xc3, 0x3c, 0xcc, + 0x83, 0x1b, 0xc8, 0xc2, 0x2d, 0xcc, 0x02, 0x3d, 0xc8, 0x43, 0x3d, 0x8c, 0x03, 0x3d, 0xd4, 0x83, 0x3c, 0x94, 0x03, 0x39, 0x88, 0x42, 0x3d, 0x98, 0x83, 0x39, 0x94, 0x83, 0x3c, 0xf0, 0x81, 0x39, + 0xb0, 0xc3, 0x3b, 0x84, 0x03, 0x3d, 0xf8, 0x01, 0x0a, 0x12, 0xd5, 0x39, 0x02, 0x50, 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, 0x80, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x86, 0x3c, 0x04, 0x10, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x79, 0x14, 0x20, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x18, 0xf2, 0x34, 0x40, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x30, 0xe4, 0x81, 0x80, 0x00, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x0b, 0x04, 0x00, 0x00, + 0x0a, 0x00, 0x00, 0x00, 0x32, 0x1e, 0x98, 0x10, 0x19, 0x11, 0x4c, 0x90, 0x8c, 0x09, 0x26, 0x47, 0xc6, 0x04, 0x43, 0x1a, 0x25, 0x50, 0x04, 0x85, 0x50, 0x0c, 0x23, 0x00, 0x85, 0x41, 0x6f, 0x04, + 0x80, 0x6e, 0x81, 0x90, 0x9d, 0x01, 0x20, 0x3a, 0x03, 0x00, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x1a, 0x03, 0x4c, 0x90, 0x46, 0x02, 0x13, 0xc4, 0x88, 0x0c, 0x6f, 0xec, + 0xed, 0x4d, 0x0c, 0x44, 0x06, 0x26, 0x26, 0xc7, 0x05, 0xa6, 0xc6, 0x05, 0x06, 0x66, 0x43, 0x10, 0x4c, 0x10, 0x86, 0x62, 0x82, 0x30, 0x18, 0x1b, 0x84, 0x81, 0x98, 0x20, 0x0c, 0xc7, 0x06, 0x61, + 0x30, 0x28, 0x8c, 0xcd, 0x4d, 0x10, 0x06, 0x64, 0xc3, 0x80, 0x24, 0xc4, 0x04, 0x61, 0x48, 0x26, 0x08, 0x94, 0x43, 0x60, 0x82, 0x30, 0x28, 0x1b, 0x92, 0x85, 0x69, 0x96, 0x65, 0x70, 0x96, 0x63, + 0x43, 0xf0, 0x4c, 0x10, 0xac, 0x66, 0x82, 0xf0, 0x2c, 0x1b, 0x96, 0x25, 0x6a, 0x96, 0x65, 0x70, 0x24, 0x49, 0x3a, 0x36, 0x04, 0xd3, 0x06, 0x02, 0xa2, 0x00, 0x60, 0x82, 0x20, 0x00, 0x1b, 0x80, + 0x0d, 0xc3, 0x72, 0x5d, 0x1b, 0x02, 0x6c, 0xc3, 0x30, 0x58, 0x19, 0x89, 0xb6, 0xb0, 0x34, 0xb7, 0x09, 0xc2, 0xc5, 0x6c, 0x18, 0x86, 0x61, 0xd8, 0x40, 0x2c, 0x9c, 0xd1, 0x6d, 0x28, 0xac, 0x0d, + 0xa8, 0xbc, 0x2a, 0x6c, 0x6c, 0x76, 0x6d, 0x2e, 0x69, 0x64, 0x65, 0x6e, 0x74, 0x53, 0x82, 0xa0, 0x0a, 0x19, 0x9e, 0x8b, 0x5d, 0x99, 0xdc, 0x5c, 0xda, 0x9b, 0xdb, 0x94, 0x80, 0x68, 0x42, 0x86, + 0xe7, 0x62, 0x17, 0xc6, 0x66, 0x57, 0x26, 0x37, 0x25, 0x30, 0xea, 0x90, 0xe1, 0xb9, 0xcc, 0xa1, 0x85, 0x91, 0x95, 0xc9, 0x35, 0xbd, 0x91, 0x95, 0xb1, 0x4d, 0x09, 0x92, 0x32, 0x64, 0x78, 0x2e, + 0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x73, 0x53, 0x02, 0xaa, 0x12, 0x19, 0x9e, 0x0b, 0x5d, 0x1e, 0x5c, 0x59, 0x90, 0x9b, 0xdb, 0x1b, 0x5d, 0x18, 0x5d, 0xda, 0x9b, 0xdb, 0xdc, 0x94, + 0x20, 0xab, 0x43, 0x86, 0xe7, 0x52, 0xe6, 0x46, 0x27, 0x97, 0x07, 0xf5, 0x96, 0xe6, 0x46, 0x37, 0x37, 0x25, 0xf0, 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, 0x8c, 0xcc, 0x21, 0x07, 0x7c, 0x70, 0x03, 0x74, 0x60, 0x07, 0x37, 0x90, 0x87, 0x72, 0x98, 0x87, 0x77, 0xa8, 0x07, 0x79, 0x18, 0x87, 0x72, 0x70, 0x83, 0x70, 0xa0, 0x07, 0x7a, 0x90, + 0x87, 0x74, 0x10, 0x87, 0x7a, 0xa0, 0x87, 0x72, 0x00, 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, 0x0d, 0x5c, 0xc3, 0xe5, 0x3b, 0x8f, 0x1f, 0x01, 0xd6, 0x46, 0x15, 0x05, 0x11, 0x95, 0x0e, 0x30, 0xf8, 0xc5, 0x6d, 0x9b, 0x00, 0x36, 0x5c, 0xbe, 0xf3, 0xf8, 0x11, 0x60, + 0x6d, 0x54, 0x51, 0x10, 0x11, 0x3b, 0x39, 0x11, 0xe1, 0x17, 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, 0x1c, 0x00, 0x00, 0x00, 0x13, 0x04, 0x41, 0x2c, 0x10, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x34, 0x4a, 0xae, 0x74, 0x03, 0xca, 0x52, 0xa0, + 0x30, 0x05, 0xc8, 0x94, 0xc0, 0x08, 0x00, 0xa5, 0x19, 0x80, 0x31, 0x02, 0x10, 0x04, 0x41, 0x10, 0x0c, 0x00, 0x00, 0x00, 0x23, 0x06, 0x09, 0x00, 0x82, 0x60, 0x00, 0x4d, 0x08, 0xe1, 0x38, 0xcc, + 0x88, 0x41, 0x02, 0x80, 0x20, 0x18, 0x40, 0x54, 0x42, 0x3c, 0x4f, 0x33, 0x62, 0x60, 0x00, 0x20, 0x08, 0x06, 0x04, 0x96, 0x40, 0x23, 0x06, 0x0a, 0x00, 0x82, 0x60, 0xd0, 0x5c, 0x89, 0x10, 0x44, + 0xc7, 0x34, 0x9a, 0x10, 0x00, 0x16, 0x18, 0xf2, 0x19, 0x31, 0x70, 0x00, 0x10, 0x04, 0x03, 0x65, 0x5b, 0x0c, 0x82, 0x0a, 0x10, 0x04, 0x51, 0x2c, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + }; + + /* -denorm preserve */ + static const BYTE cs_code_preserve[] = + { + 0x44, 0x58, 0x42, 0x43, 0x7c, 0x5e, 0x67, 0x8e, 0xbc, 0xf4, 0x16, 0xc1, 0xcb, 0x9f, 0xed, 0xc2, 0xa1, 0xce, 0x48, 0x34, 0x01, 0x00, 0x00, 0x00, 0xa8, 0x06, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x34, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, 0x64, 0x00, 0x00, 0x00, 0xc8, 0x00, 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, 0x5c, 0x00, 0x00, 0x00, 0x24, 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, 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 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, 0x14, 0x00, 0x00, 0x00, 0x1b, 0xba, 0xe0, 0xff, 0xff, 0xff, 0xff, 0x07, 0xc2, 0x1c, 0xe0, 0x61, + 0x0e, 0xe4, 0xa0, 0x0d, 0xc8, 0xa1, 0x1c, 0xdc, 0xe1, 0x1d, 0xe4, 0xa1, 0x1d, 0xda, 0xa0, 0x1d, 0xde, 0x81, 0x1c, 0xca, 0x01, 0x80, 0x07, 0x79, 0x28, 0x87, 0x79, 0x28, 0x07, 0x79, 0xb0, 0x87, + 0x72, 0x00, 0x36, 0x18, 0xc2, 0xff, 0xff, 0xff, 0xff, 0x0f, 0x80, 0x04, 0x50, 0x1b, 0x88, 0xe1, 0xff, 0xff, 0xff, 0xff, 0x07, 0x40, 0xda, 0x60, 0x10, 0xff, 0xff, 0xff, 0xff, 0x3f, 0x00, 0x12, + 0x50, 0x01, 0x00, 0x00, 0x49, 0x18, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x13, 0x82, 0x60, 0x42, 0x20, 0x4c, 0x08, 0x86, 0x09, 0x01, 0x01, 0x00, 0x89, 0x20, 0x00, 0x00, 0x36, 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, 0x60, 0x23, 0x00, + 0x25, 0x00, 0x14, 0xe6, 0x08, 0xc0, 0xa0, 0x0c, 0x63, 0x0c, 0x22, 0x73, 0x04, 0x08, 0x99, 0x7b, 0x86, 0xcb, 0x9f, 0xb0, 0x87, 0x90, 0xfc, 0x10, 0x68, 0x86, 0x85, 0x40, 0xc1, 0x99, 0x01, 0x28, + 0x0b, 0x18, 0x68, 0x8c, 0x94, 0x52, 0x32, 0x83, 0xd4, 0x4d, 0xc3, 0xe5, 0x4f, 0xd8, 0x43, 0x48, 0xfe, 0x4a, 0x48, 0x2b, 0x31, 0xf9, 0xc5, 0x6d, 0xa3, 0x22, 0xa5, 0x94, 0x46, 0x39, 0xd8, 0x40, + 0x63, 0x98, 0x41, 0x6d, 0x8e, 0x20, 0x28, 0x06, 0x1a, 0x66, 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, 0x90, 0x68, 0xce, + 0xf4, 0x8d, 0x03, 0x3b, 0x84, 0xc3, 0x3c, 0xcc, 0x83, 0x1b, 0xc8, 0xc2, 0x2d, 0xcc, 0x02, 0x3d, 0xc8, 0x43, 0x3d, 0x8c, 0x03, 0x3d, 0xd4, 0x83, 0x3c, 0x94, 0x03, 0x39, 0x88, 0x42, 0x3d, 0x98, + 0x83, 0x39, 0x94, 0x83, 0x3c, 0xf0, 0x81, 0x39, 0xb0, 0xc3, 0x3b, 0x84, 0x03, 0x3d, 0xf8, 0x01, 0x0a, 0x12, 0xd5, 0x39, 0x02, 0x50, 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, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x86, 0x3c, 0x04, 0x10, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x79, 0x14, 0x20, + 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x18, 0xf2, 0x34, 0x40, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x30, 0xe4, 0x81, 0x80, 0x00, 0x20, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x20, 0x0b, 0x04, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x32, 0x1e, 0x98, 0x10, 0x19, 0x11, 0x4c, 0x90, 0x8c, 0x09, 0x26, 0x47, 0xc6, 0x04, 0x43, 0x1a, 0x25, 0x50, 0x04, 0x85, + 0x50, 0x0c, 0x23, 0x00, 0x85, 0x41, 0x6f, 0x04, 0x80, 0x6e, 0x81, 0x90, 0x9d, 0x01, 0x20, 0x3a, 0x03, 0x00, 0x00, 0x00, 0x79, 0x18, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x1a, 0x03, 0x4c, 0x90, + 0x46, 0x02, 0x13, 0xc4, 0x88, 0x0c, 0x6f, 0xec, 0xed, 0x4d, 0x0c, 0x44, 0x06, 0x26, 0x26, 0xc7, 0x05, 0xa6, 0xc6, 0x05, 0x06, 0x66, 0x43, 0x10, 0x4c, 0x10, 0x86, 0x62, 0x82, 0x30, 0x18, 0x1b, + 0x84, 0x81, 0x98, 0x20, 0x0c, 0xc7, 0x06, 0x61, 0x30, 0x28, 0x8c, 0xcd, 0x4d, 0x10, 0x06, 0x64, 0xc3, 0x80, 0x24, 0xc4, 0x04, 0x61, 0x48, 0x26, 0x08, 0x94, 0x43, 0x60, 0x82, 0x30, 0x28, 0x1b, + 0x92, 0x85, 0x69, 0x96, 0x65, 0x70, 0x96, 0x63, 0x43, 0xf0, 0x4c, 0x10, 0xac, 0x66, 0x82, 0xf0, 0x2c, 0x1b, 0x96, 0x25, 0x6a, 0x96, 0x65, 0x70, 0x24, 0x49, 0x3a, 0x36, 0x04, 0xd3, 0x06, 0x02, + 0xa2, 0x00, 0x60, 0x82, 0x20, 0x00, 0x1b, 0x80, 0x0d, 0xc3, 0x72, 0x5d, 0x1b, 0x02, 0x6c, 0xc3, 0x30, 0x58, 0x19, 0x89, 0xb6, 0xb0, 0x34, 0xb7, 0x09, 0xc2, 0xc5, 0x6c, 0x18, 0x86, 0x61, 0xd8, + 0x40, 0x2c, 0x9c, 0xd1, 0x6d, 0x28, 0xac, 0x0d, 0xa8, 0xbc, 0x2a, 0x6c, 0x6c, 0x76, 0x6d, 0x2e, 0x69, 0x64, 0x65, 0x6e, 0x74, 0x53, 0x82, 0xa0, 0x0a, 0x19, 0x9e, 0x8b, 0x5d, 0x99, 0xdc, 0x5c, + 0xda, 0x9b, 0xdb, 0x94, 0x80, 0x68, 0x42, 0x86, 0xe7, 0x62, 0x17, 0xc6, 0x66, 0x57, 0x26, 0x37, 0x25, 0x30, 0xea, 0x90, 0xe1, 0xb9, 0xcc, 0xa1, 0x85, 0x91, 0x95, 0xc9, 0x35, 0xbd, 0x91, 0x95, + 0xb1, 0x4d, 0x09, 0x92, 0x32, 0x64, 0x78, 0x2e, 0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x73, 0x53, 0x02, 0xaa, 0x12, 0x19, 0x9e, 0x0b, 0x5d, 0x1e, 0x5c, 0x59, 0x90, 0x9b, 0xdb, 0x1b, + 0x5d, 0x18, 0x5d, 0xda, 0x9b, 0xdb, 0xdc, 0x94, 0x20, 0xab, 0x43, 0x86, 0xe7, 0x52, 0xe6, 0x46, 0x27, 0x97, 0x07, 0xf5, 0x96, 0xe6, 0x46, 0x37, 0x37, 0x25, 0xf0, 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, 0x8c, 0xcc, 0x21, 0x07, 0x7c, 0x70, 0x03, 0x74, 0x60, 0x07, 0x37, 0x90, 0x87, 0x72, 0x98, 0x87, 0x77, 0xa8, 0x07, 0x79, 0x18, 0x87, + 0x72, 0x70, 0x83, 0x70, 0xa0, 0x07, 0x7a, 0x90, 0x87, 0x74, 0x10, 0x87, 0x7a, 0xa0, 0x87, 0x72, 0x00, 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, 0x0d, 0x5c, 0xc3, 0xe5, 0x3b, 0x8f, 0x1f, 0x01, 0xd6, 0x46, 0x15, 0x05, 0x11, 0x95, 0x0e, 0x30, 0xf8, 0xc5, 0x6d, 0x9b, + 0x00, 0x36, 0x5c, 0xbe, 0xf3, 0xf8, 0x11, 0x60, 0x6d, 0x54, 0x51, 0x10, 0x11, 0x3b, 0x39, 0x11, 0xe1, 0x17, 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, 0x1c, 0x00, 0x00, 0x00, 0x13, 0x04, 0x41, 0x2c, 0x10, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, + 0x34, 0x4a, 0xae, 0x74, 0x03, 0xca, 0x52, 0xa0, 0x30, 0x05, 0xc8, 0x94, 0xc0, 0x08, 0x00, 0xa5, 0x19, 0x80, 0x31, 0x02, 0x10, 0x04, 0x41, 0x10, 0x0c, 0x00, 0x00, 0x00, 0x23, 0x06, 0x09, 0x00, + 0x82, 0x60, 0x00, 0x4d, 0x08, 0xe1, 0x38, 0xcc, 0x88, 0x41, 0x02, 0x80, 0x20, 0x18, 0x40, 0x54, 0x42, 0x3c, 0x4f, 0x33, 0x62, 0x60, 0x00, 0x20, 0x08, 0x06, 0x04, 0x96, 0x40, 0x23, 0x06, 0x0a, + 0x00, 0x82, 0x60, 0xd0, 0x5c, 0x89, 0x10, 0x44, 0xc7, 0x34, 0x9a, 0x10, 0x00, 0x16, 0x18, 0xf2, 0x19, 0x31, 0x70, 0x00, 0x10, 0x04, 0x03, 0x65, 0x5b, 0x0c, 0x82, 0x0a, 0x10, 0x04, 0x51, 0x2c, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + }; + + static const struct D3D12_SHADER_BYTECODE cs_ftz = { cs_code_ftz, sizeof(cs_code_ftz) }; + static const struct D3D12_SHADER_BYTECODE cs_preserve = { cs_code_preserve, sizeof(cs_code_preserve) }; + + struct test + { + const struct D3D12_SHADER_BYTECODE *cs; + uint32_t input_fp32; + uint32_t output_fp32; + }; + + static const struct test tests[] = + { + /* In denorms, additions work like uint32_t. */ + { &cs_ftz, 10, 0 }, + { &cs_preserve, 10, 20 }, + }; + + if (!init_compute_test_context(&context)) + return; + + if (!context_supports_dxil(&context)) + { + skip("Context does not support DXIL.\n"); + destroy_test_context(&context); + return; + } + + memset(&shader_model, 0, sizeof(shader_model)); + shader_model.HighestShaderModel = D3D_SHADER_MODEL_6_2; + if (FAILED(ID3D12Device_CheckFeatureSupport(context.device, D3D12_FEATURE_SHADER_MODEL, &shader_model, sizeof(shader_model))) || + shader_model.HighestShaderModel < D3D_SHADER_MODEL_6_2) + { + skip("Device does not support SM 6.2.\n"); + destroy_test_context(&context); + return; + } + + memset(&rs_desc, 0, sizeof(rs_desc)); + rs_desc.NumParameters = ARRAY_SIZE(root_parameters); + rs_desc.pParameters = root_parameters; + memset(root_parameters, 0, sizeof(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, &context.root_signature); + + for (i = 0; i < ARRAY_SIZE(tests); i++) + { + vkd3d_test_set_context("Test %u", i); + pso = create_compute_pipeline_state(context.device, context.root_signature, *tests[i].cs); + src = create_upload_buffer(context.device, sizeof(tests[i].input_fp32), &tests[i].input_fp32); + dst = create_default_buffer(context.device, sizeof(tests[i].input_fp32), D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_UNORDERED_ACCESS); + ID3D12GraphicsCommandList_SetComputeRootSignature(context.list, context.root_signature); + ID3D12GraphicsCommandList_SetPipelineState(context.list, pso); + ID3D12GraphicsCommandList_SetComputeRootShaderResourceView(context.list, 0, ID3D12Resource_GetGPUVirtualAddress(src)); + ID3D12GraphicsCommandList_SetComputeRootUnorderedAccessView(context.list, 1, ID3D12Resource_GetGPUVirtualAddress(dst)); + ID3D12GraphicsCommandList_Dispatch(context.list, 1, 1, 1); + + transition_resource_state(context.list, dst, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_COPY_SOURCE); + get_buffer_readback_with_command_list(dst, DXGI_FORMAT_R32_UINT, &rb, context.queue, context.list); + + value = get_readback_uint(&rb, 0, 0, 0); + ok(value == tests[i].output_fp32, "0x%x != 0x%x\n", value, tests[i].output_fp32); + + ID3D12Resource_Release(src); + ID3D12Resource_Release(dst); + ID3D12PipelineState_Release(pso); + release_resource_readback(&rb); + reset_command_list(context.list, context.allocator); + } + + vkd3d_test_set_context(NULL); + destroy_test_context(&context); +} + START_TEST(d3d12) { pfn_D3D12CreateDevice = get_d3d12_pfn(D3D12CreateDevice); @@ -53433,4 +53651,5 @@ START_TEST(d3d12) run_test(test_view_min_lod); run_test(test_sv_barycentric); run_test(test_shader_fp16); + run_test(test_shader_sm62_denorm); }