Implement sin/cos for GP
GP doesn't have instructions for sin and cos, so we need to implement NIR lowering pass to lower sin and cos into polynomial (Taylor series)
Activity
I implemented lowering pass here: https://gitlab.freedesktop.org/anarsoul/mesa/tree/lima-gp-sincos
But unfortunately our GP compiler can't handle result and fails:
glmark2-es2-drm: ../src/gallium/drivers/lima/ir/gp/scheduler.c:395: gpir_sched_node: Assertion `result' failed.
I need some help here since I don't understand GP compiler internals.
@yuq825 Does this mean that
flog2is not yet supported in Lima? Would that be something difficult to implement? It's required by yuv2rgb shader in Kodi: https://github.com/xbmc/xbmc/blob/master/system/shaders/GLES/2.0/gles_yuv2rgb_basic.frag#L65@kszaq It's not that difficult, and your shader is fragment shader which run on PP. lima PP compiler support flog2.
-
The issue with this shader is black screen unless these 3 lines are removed: https://github.com/xbmc/xbmc/blob/master/system/shaders/GLES/2.0/gles_yuv2rgb_basic.frag#L65-L67 That's why I suspected that
flog2might be an issue here.
I hacked vertex shader to workaround missing flog2:
diff --git a/data/shaders/jellyfish.vert b/data/shaders/jellyfish.vert index c125370..fb80865 100644 --- a/data/shaders/jellyfish.vert +++ b/data/shaders/jellyfish.vert @@ -48,7 +48,7 @@ void main(void) //fresnel vec4 worldPos = uWorld * pos4; vec3 vWorldEyeVec = normalize(worldPos.xyz/worldPos.w); - float fresnelProduct = pow(1.0 - max(abs(dot(vVertexNormal, -vWorldEyeVec)), 0.0), uFresnelPower); + float fresnelProduct = pow(1.0 - max(abs(dot(vVertexNormal, -vWorldEyeVec)), 0.0), 2.0); //uFresnelPower); vFresnel = uFresnelCol.rgb * vec3(uFresnelCol.a * fresnelProduct); // texcoordIt'll change colors a bit, but jellyfish geometry should be the same.
Edited by Vasily KhoruzhickI've reproduced this problem.
Seems it's due to complex1 node handle problem in
gpir_sched_instr_pass. complex1 node will consume 2 slots, so we count it as 2, but we may schedule a move for it. In value reg stage, we use dump_f/m node method: https://gitlab.freedesktop.org/lima/mesa/blob/lima-18.3/src/gallium/drivers/lima/ir/gp/lower.c#L324 which permit this, so count reach to 6 but there's no slot in current instruction when fail case.gpir_sched_instr_passneed to handle this case to schedule a move for 2 slot complex1 node, so that the count can be reduced to 5 again.Here's a patch to fix the compile fail, I can see the jellyfish on screen, but not correct shape and GP error at last. 0001-lima-gpir-fix-compile-fail-when-two-slot-node.patch
mentioned in merge request mesa/mesa!636 (merged)
The jellyfish error seems to be currently not because of sin.
I wrote a simple program to verify sin(). See https://pastebin.aosc.io/paste/JI9dY1gqd2wBhheITIGfmw . It will draw the figure of y=sin(2x), with the left-top of screen is (-1, 1) and right-bottom (1, -1).
Here's simplified shader for jellyfish. GP compiler still can't handle it and fails with:
glmark2-es2-drm: ../src/gallium/drivers/lima/ir/gp/scheduler.c:395: gpir_sched_node: Assertion `result' failed.
So far I think nothing is wrong with sin/cos lowering - NIR representation looks correct. Something is wrong with GP compiler.
Edited by Vasily KhoruzhickAgree.
My previous patch fix compile fail with this shader due to complex1 handling: jellyfish.vert
Your shader compile fail due to a bug in the alu check, here is a patch to fix it: 0001-lima-gpir-fix-alu-check-miss-last-store-slot.patch
@yuq825 it compiles now, but result is not correct. Jellyfish animation is wrong.
-
@anarsoul Interesting. I had thought the broken animation as intended... oops. (My main laptop is really IVB)
But Lima is too broken on jellyfish that even the color gradient of the background is missing.
@icenowy color gradient can be missing due to low precision (PP on mali4x0 can do only fp16). I actually see some gradient (or rather color banding) on pinebook.
@yuq825 GP compiler has an issue with registers. If you look at shader disassembly for jellyfish shader there's a read from uninitialized register (look for $1):
=======disassembly:======= 000: mul ^3 a0.y t[20].w, 001: , add ^6 t[4].y ^3, 002: mul ^15 ^6 t[21].x, 003: , floor ^18 ^15, 004: , add ^24 ^15 -^18, 005: , add ^30 ^24 t[21].y, 006: mov ^38 ^24, mov ^39 ^30, ge ^36 ^24 t[20].w, 007: , lt ^42 ^38 t[21].z, 008: sel ^50 ^36 ^38 ^39, add ^48 ^38 t[20].y, 009: sel ^56 ^42 ^50 ^48, 010: mul ^63 ^56 ^56, 011: mul ^68 ^56 ^63, mul ^69 ^56 t[21].w, mov ^66 ^63, 012: mul ^74 ^68 t[22].x, mul ^75 ^68 ^66, 013: mul ^80 ^75 t[22].y, mul ^81 ^75 ^66, add ^78 ^69 ^74, mov ^79 ^66, 014: mul ^86 ^81 t[22].z, mul ^87 ^81 t[22].w, add ^84 ^78 ^80, 015: mul ^93 ^87 ^79, add ^90 ^84 ^86, 016: , add ^96 ^90 ^93, 017: mul ^104 a2.x t[20].z, mul ^105 a2.y t[20].z, mov ^102 ^96, 018: mul ^110 ^104 ^102, mul ^111 ^105 ^102, 019: mov ^116 ^102, , add ^114/$0.y a0.y ^111, add ^115/$0.x a0.x ^110, 020: , 021: mov ^129 t[0].x, mov ^126 ^116, 022: mul ^134 ^129 $0.x, mul ^135 t[1].x $1.x, <---- $1 has never been initialized!Edited by Vasily Khoruzhick@anarsoul which version of jellyfish.vert? I can't see this result with your simplified shader.
