radv: enable FP16/FP64 denormals earlier and only for LLVM

ACO sets this itself and will have to set it differently in the future to
support shaderDenormFlushToZeroFloat64.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index c2562c028c4e00ba4b42d32f4c7a6fab2a551f22..1e550526f56cb70a049d9719e3a29c14823373ac 100644 (file)
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -700,20 +700,6 @@ static void radv_postprocess_config(const struct radv_physical_device *pdevice,
        config_out->num_sgprs = num_sgprs;
        config_out->num_shared_vgprs = num_shared_vgprs;
 
-       /* Enable 64-bit and 16-bit denormals, because there is no performance
-        * cost.
-        *
-        * If denormals are enabled, all floating-point output modifiers are
-        * ignored.
-        *
-        * Don't enable denormals for 32-bit floats, because:
-        * - Floating-point output modifiers would be ignored by the hw.
-        * - Some opcodes don't support denormals, such as v_mad_f32. We would
-        *   have to stop using those.
-        * - GFX6 & GFX7 would be very slow.
-        */
-       config_out->float_mode |= V_00B028_FP_64_DENORMS;
-
        config_out->rsrc2 = S_00B12C_USER_SGPR(info->num_user_sgprs) |
                            S_00B12C_SCRATCH_EN(scratch_enabled);
 
@@ -971,6 +957,20 @@ radv_shader_variant_create(struct radv_device *device,
                        return NULL;
                }
 
+               /* Enable 64-bit and 16-bit denormals, because there is no performance
+                * cost.
+                *
+                * If denormals are enabled, all floating-point output modifiers are
+                * ignored.
+                *
+                * Don't enable denormals for 32-bit floats, because:
+                * - Floating-point output modifiers would be ignored by the hw.
+                * - Some opcodes don't support denormals, such as v_mad_f32. We would
+                *   have to stop using those.
+                * - GFX6 & GFX7 would be very slow.
+                */
+               config.float_mode |= V_00B028_FP_64_DENORMS;
+
                if (rtld_binary.lds_size > 0) {
                        unsigned alloc_granularity = device->physical_device->rad_info.chip_class >= GFX7 ? 512 : 256;
                        config.lds_size = align(rtld_binary.lds_size, alloc_granularity) / alloc_granularity;