When the following code path happens:
lima_context_create
|-> lima_submit_create
    |-> drmSyncobjCreate: can fail here
    |Thus ctx->pp_submit and ctx->gp_submit are NULL
    |-> lima_context_destroy
        |-> lima_submit_free: will access submit->in_sync_fd

It triggers a segfault in mesa because we are trying to derefence
both ctx->pp_submit and ctx->gp_submit which are NULL.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

This can free user space to record the submit bos
before task finish.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

On GCC8 egl fails to compile with this error:
  error: 'NULL' undeclared (first use in this function)

As NULL is declared in stddef.h add it as an include to fix the build
failure.

Signed-off-by: Alistair Francis <alistair@alistair23.me>

Arno Messiaen authored Dec 19, 2018 and Qiang Yu committed Jan 08, 2019

Signed-off-by: Arno Messiaen <arnomessiaen@gmail.com>

Arno Messiaen authored Dec 18, 2018 and Qiang Yu committed Jan 08, 2019

Signed-off-by: Arno Messiaen <arnomessiaen@gmail.com>

Arno Messiaen authored Dec 18, 2018 and Qiang Yu committed Jan 08, 2019

Signed-off-by: Arno Messiaen <arnomessiaen@gmail.com>

As already mentioned in the lima project utgard documentation, for sin
and cos, the input has to multiplied by the constant 1/(2*pi),
presumably to simplify the hardware.
The binary compiler outputs a multiplication in the fragment shader to
compensate for this, do the same in ppir.

Signed-off-by: Erico Nunes <nunes.erico@gmail.com>

condition instruction must be in fmul slot.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Arno Messiaen authored Dec 16, 2018 and Qiang Yu committed Dec 20, 2018

Signed-off-by: Arno Messiaen <arnomessiaen@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Some lima_ctx_buff is allocated conditionally for
each draw like gp_uniform, pp_uniform and tex_desc.
mmu fault will happen when:

draw 1: tex_desc (i.e.) is allocated and added to
  PP submit, draw is OK
draw 2: no LIMA_CONTEXT_DIRTY_TEXTURES, so tex_desc
  is unchanged, but not added to PP submit
draw 3: LIMA_CONTEXT_DIRTY_TEXTURES, so original
  tex_desc is freed and allocate a new one

So draw 2 may get mmu fault because the tex_desc
buffer is freed when its execution.

Move submit add to lima_ctx_buff_va so that the
buffer will be added whenever get used by GPU
command stream.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Erico Nunes authored Dec 09, 2018 and Qiang Yu committed Dec 10, 2018

Without this fix, ppir codegen outputs instructions that always write to
the x component of the output register, so a simple shader like this
won't work:
  gl_FragColor = vec4(color.rg, sqrt(color.b*color.b), color.a);
It produces
  sqrt.s2 $0.x  $1.x
instead of
  sqrt.s2 $0.z  $1.x

Signed-off-by: Erico Nunes <nunes.erico@gmail.com>

This help we know the last error draw and
get its command stream dump.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

fragment shader from glamor:

precision mediump float;

varying vec2 glyph_pos;
uniform sampler2D sampler;
uniform sampler2D atlas;
void main() {
       vec4 mask = texture2D(atlas, glyph_pos);
       vec4 source = texture2D(sampler, vec2(0.5));
       gl_FragColor = source * mask.a;
}

v2:
  Solve in lower pass instead of node-to-instr pass.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

RGBX will always set alpha channel to 1 while
RGBA will pass orignal alpha channel.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

When running glxgears, the nir uniform num is 11.
Mesa align it to 12 with some pad bytes when set
constant buffer, but gpir lower pending uniforms
(viewport transform) at offset 11.

So lima_update_gp_uniform should copy pending
uniforms to 11 instead of constant buffer size
12.

Now glxgears works!

Signed-off-by: Qiang Yu <yuq825@gmail.com>

glxgears will pass in some invalid draw like
2 vertex GL_QUARD_STRIP draw which causes GP
hang.

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>

Signed-off-by: Qiang Yu <yuq825@gmail.com>