Jeff Bolz | 4605e2e | 2019-02-19 13:10:32 -0600 | [diff] [blame^] | 1 | #version 450 core |
| 2 | #extension GL_KHR_memory_scope_semantics : enable
|
| 3 | #extension GL_NV_cooperative_matrix : enable
|
| 4 | #extension GL_EXT_shader_explicit_arithmetic_types_float16 : enable
|
| 5 |
|
| 6 | layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in; |
| 7 |
|
| 8 | float<16> ftemplate16;
|
| 9 |
|
| 10 | fcoopmatNV fnoparams;
|
| 11 |
|
| 12 | fcoopmatNV<8, gl_ScopeSubgroup, 8, 8> fbadbits;
|
| 13 |
|
| 14 | fcoopmatNV<16, gl_ScopeSubgroup, 8> fbadnumparams;
|
| 15 |
|
| 16 | int X = 8;
|
| 17 |
|
| 18 | fcoopmatNV<16, gl_ScopeSubgroup, 8, X> fbadparam;
|
| 19 |
|
| 20 | layout(constant_id = 0) int Y = 1;
|
| 21 |
|
| 22 | shared fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sharedmat;
|
| 23 |
|
| 24 | layout(set = 0, binding = 0) buffer InvBlock {
|
| 25 | fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> bufmat;
|
| 26 | } invblock;
|
| 27 |
|
| 28 | void main() |
| 29 | {
|
| 30 | fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> f32_16_8;
|
| 31 | fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> f16_16_8;
|
| 32 |
|
| 33 | // invalid implicit conversions
|
| 34 | f32_16_8 = f16_16_8;
|
| 35 | f32_16_8 = f16_16_8 + f16_16_8;
|
| 36 |
|
| 37 | fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> f16_8_8;
|
| 38 |
|
| 39 | // mismatching dimensions
|
| 40 | f16_16_8 = f16_8_8;
|
| 41 |
|
| 42 | fcoopmatNV<16, gl_ScopeSubgroup, 8, Y> f16_8_Y;
|
| 43 | fcoopmatNV<16, gl_ScopeSubgroup, 8, (Y+1)> f16_8_Y1;
|
| 44 |
|
| 45 | // mismatching dimensions with specialization constants
|
| 46 | f16_8_Y = f16_8_Y1;
|
| 47 |
|
| 48 | // wrong arguments for constructor
|
| 49 | f16_8_8 = fcoopmatNV<16, gl_ScopeSubgroup, 8, 8>(1, 1);
|
| 50 |
|
| 51 | // can't construct from a builtin type
|
| 52 | mat4 m4;
|
| 53 | fcoopmatNV<32, gl_ScopeSubgroup, 4, 4> f32_4_4 = fcoopmatNV<32, gl_ScopeSubgroup, 4, 4>(m4);
|
| 54 |
|
| 55 | // only support a single array subscript
|
| 56 | f16_16_8[0][0];
|
| 57 |
|
| 58 | // don't support scalar component selection
|
| 59 | f16_16_8.x;
|
| 60 |
|
| 61 | f16_16_8 * f16_16_8;
|
| 62 |
|
| 63 | f16_16_8 + 1.0;
|
| 64 | f16_16_8 - 1.0;
|
| 65 | f16_16_8 / 1.0;
|
| 66 | f16_16_8 += 1.0;
|
| 67 | f16_16_8 -= 1.0;
|
| 68 | f16_16_8 /= 1.0;
|
| 69 |
|
| 70 | f16_16_8*2.0;
|
| 71 | 2.0*f16_16_8;
|
| 72 | f32_16_8*float16_t(2.0);
|
| 73 | float16_t(2.0)*f32_16_8;
|
| 74 |
|
| 75 | transpose(f16_8_8);
|
| 76 | } |