[PATCH 8/8] wined3d: Implement applying state for compute pipeline.

[Józef Kucia]

On Thu, Feb 2, 2017 at 3:07 PM, Henri Verbeet <hverbeet at gmail.com> wrote:
> On 2 February 2017 at 12:43, Józef Kucia <jkucia at codeweavers.com> wrote:
>> @@ -3520,19 +3546,19 @@ BOOL context_apply_draw_state(struct wined3d_context *context,
>>              wined3d_buffer_load_sysmem(state->index_buffer, context);
>>      }
>>
>> -    for (i = 0; i < context->numDirtyEntries; ++i)
>> +    for (i = 0, dirty_idx = 0; i < context->numDirtyEntries; ++i)
>>      {
>>          DWORD rep = context->dirtyArray[i];
>> -        DWORD idx = rep / (sizeof(*context->isStateDirty) * CHAR_BIT);
>> -        BYTE shift = rep & ((sizeof(*context->isStateDirty) * CHAR_BIT) - 1);
>> -        context->isStateDirty[idx] &= ~(1u << shift);
>> -        state_table[rep].apply(context, state, rep);
>> +        if (!is_compute_state(rep))
>> +            apply_state(context, state_table, state, rep);
>> +        else
>> +            context->dirtyArray[dirty_idx++] = rep;
> I'd rather not do this. How important is it to avoid applying compute
> dispatch states for draws and vice-versa?

I think it would complicate things a lot. Applying all states for
dispatches would mean that we have to acquire, prepare and load all
resources potentially used by draws, e.g. render targets and constant
buffers. This would be required mainly because of the way how
context_state_fb(), state_cb() and similar are implemented. It could
be possibly mitigated to some extent by refactoring how the states are
applied, e.g. using a similar mechanism to
"context->update_shader_resource_bindings", but this approach doesn't
appeal to me.

Applying just necessary states may also avoid unnecessary work, but
this shouldn't matter much in practice.