Re: [PATCH v2 1/2] wined3d: Use 3 component norm for 'nrm' opcode in GLSL backend.

[Henri Verbeet]

On Wed, 12 Jun 2019 at 15:33, Paul Gofman  wrote:
> - shader_addline(buffer, "tmp0.x = dot(%s, %s);\n",
> -
>        src_param.param_str, src_param.param_str);
> - if (mask_size > 3)
> -
>    shader_addline(buffer, "tmp0.x = dot(vec3(%s), vec3(%s));\n",
> -
>            src_param.param_str, src_param.param_str);
> - else
> -
>    shader_addline(buffer, "tmp0.x = dot(%s, %s);\n",
> -
>            src_param.param_str, src_param.param_str);

This is fine.
> - if (mask_size > 1)
> - if (mask_size == 4)
> - {
> -
>    static const float max_float = FLT_MAX;
> -
> -
>    shader_addline(buffer, "tmp0.x == 0.0 ? vec4(vec3(0.0), sign(%s[3]) * ",
> -
>            src_param.param_str);
> -
>    shader_glsl_append_imm_vec(buffer, &max_float, 1, ins->ctx->gl_info);
> -
>    shader_addline(buffer, ") : (%s * inversesqrt(tmp0.x)));\n", src_param.param_str);
> - }
> - else if (mask_size > 1)

This seems like a separate change. I'm also not sure about the FLT_MAX
literal. I'd expect that you could achieve the same test results by
simply multiplying the .w component with the rsq of tmp0.x. (Under
d3d9's "zero wins"-rules at least; there would be a potential NaN
under IEEE rules.)