On 4/1/19 4:02 PM, Bas Nieuwenhuizen wrote:
This seems weird, do you know which tests are affected?
All 16-bit fma tests.
(I suspect what we really want to do is split into fmuladd and fma, and only lower fmuladd. If that does not work, something else is definitely going on)
I will investigate more.
On Mon, Mar 25, 2019 at 4:10 PM Samuel Pitoiset <samuel.pitoi...@gmail.com> wrote:The lowering needs to be disabled for sufficient precision to pass deqp-vk's 16-bit fma test on radv. Original patch from Rhys. Signed-off-by: Samuel Pitoiset <samuel.pitoi...@gmail.com> --- src/amd/vulkan/radv_shader.c | 1 - 1 file changed, 1 deletion(-) diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c index 63d4147460c..05041b2eae6 100644 --- a/src/amd/vulkan/radv_shader.c +++ b/src/amd/vulkan/radv_shader.c @@ -70,7 +70,6 @@ static const struct nir_shader_compiler_options nir_options = { .lower_unpack_unorm_4x8 = true, .lower_extract_byte = true, .lower_extract_word = true, - .lower_ffma16 = true, .lower_ffma32 = true, .lower_ffma64 = true, .lower_fpow = true, -- 2.21.0 _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
