vulkan : fix FA mask load with bounds check (coopmat2) (llama/17606)

This commit is contained in:
Acly
2025-12-12 17:53:13 +02:00
committed by Georgi Gerganov
parent a3459484bf
commit 2258930c2e
@@ -156,7 +156,7 @@ void main() {
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv, mvmax;
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));