Hans de Goede
2016-Feb-19 10:36 UTC
[Nouveau] Dealing with opencl kernel parameters in nouveau now that RES support is gone
Hi, On 18-02-16 17:39, Ilia Mirkin wrote:> On Thu, Feb 18, 2016 at 9:45 AM, Hans de Goede <hdegoede at redhat.com> wrote: >> But this does not seem to be hooked up yet for nouveau. > > Samuel has patches. See > https://cgit.freedesktop.org/~hakzsam/mesa/log/?h=arb_compute_shader_v3Cool, I will take a look at those.>> So some questions: >> -The commit by Samual says: >> This introduces TGSI_FILE_MEMORY for shared, global and local memory. Only >> shared memory is currently supported. >> >> The commit introduces MEMORY[x] and MEMORY[x],SHARED so in reality it also >> introduces >> a second option next to shared, so what are we going to use plain MEMORY[x] >> for? >> I suggest using it for global memory but we need to be in agreement on >> this. > > That sounds fine to me. However what I had in mind was switching the > SHARED field into a 2-bit field and making it > > 1 = SHARED > 2 = GLOBAL > 3 = LOCAL > > (since for OpenCL you also need to be able to address local or private > memory). I sorta wanted Samuel to do it, but since I had no idea where > you were at, or if you were even still working on this, I figured it > should be fixed up by the first person who needed it.Sounds good, only the naming is somewhat unfortunate since opencl uses different naming. I.e. it has no "shared" OpenCL has: -global: accessible by all worker-groups as well as by the CPU -const: read-only global -local: shared by worker-items in the same worker-group, not shared between worker-groups -private: accessible only to a single worker-item So how do these map to the TGSI: > 1 = SHARED > 2 = GLOBAL > 3 = LOCAL ?> >> >> -What about kernel input parameters, so far these have been using RES[32764] >> I must admit that I do not understand where the file_index of 32764 comes >> from (or where any of the file indexes come from for >> src/gallium/tests/trivial/compute.c ? >> I have the feeling that these are not used at all, and everything simply >> goes >> to a flat (virtual) memory space, with the params at address 0, correct ? > > It was never particularly well-specified, which was one of the reasons > it went away. It also didn't map nicely onto the OpenGL model. There > is a remaining question of how to do addressing in memory... there's > 40 bits of address space. Should these implicitly be U64 > (dual-component in TGSI) addresses that are passed around? Not sure > what the OpenCL position on all this is.So far I've been using U32 for addresses as that is what Francisco's original code was using. And this also is what things like the tgsi LOAD instruction take. If you're doing a LOAD on a 1d buffer then you will use TEMP[#].x to specify the index, and the way how this currently works with OpenCL is that clCreateBuffer() will return a cl_mem type which then gets passed into the kernel as input parameter and gets treated as a pointer by the compiler, so e.g. global mem gets treated as a single address space even if there are multiple global buffers and TEMP[#].x contains the value passed in via cl_mem as start offset for the buffer + the index into the buffer. So this means that currently we are limited to U32 since TEMP[#].x is only 32 bits wide. Internally 40 bits addresses can and should probably be used so that at least the different memory spaces each have the full 32 bits available. Note that we could fix this by adding some sort of LOAD64 opcode, which uses TEMP[#].x and TEMP[#].y as address for 1d buffers, I'm not sure how this would work for 3d buffers though. I foresee the llvm backend eventually getting a 64 bit mode where it will use 64 bits for all pointers and use something like a LOAD64 opcode to indicate that the indexes (which it effectively uses as addresses / pointers) are 64 bit wide.> Also it would be highly preferable to avoid using GLOBAL at all in the > first place, and just use BUFFER[] or IMAGE[] or whatever. However > after having a look at how OpenCL works, I don't think that's > possible. But perhaps I missed something?OpenCL is very C-like and as such uses a flat address space per memory space, getting around that is going to be very tricky, if possible at all.>> This seems like a nvidia implementation detail though, and we should IMHO >> still >> implement a specific TGSI register file for accessing kernel input params, >> agreed ? > > I believe that kernel input params should be passed in via user > uniforms. i.e. CONST. It doesn't have to be a separate UBO, it can use > the user buffer thing, which is basically what it is now. > >> >> If we can agree on the TGSI syntax for all this, then I'm pretty sure I can >> hack >> something together to get things to work with the latest master again, and >> then we >> can use this as a starting point for properly implementing this. >> >> Regards, >> >> Hans >> >> >> p.s. >> >> I must say I was somewhat unpleasantly surprised by the RES breakage, yes I >> know >> some changes were coming but still. In the future when my llvm->tgsi work is >> more >> stable and I'll be pushing it to llvm upstream we really cannot make >> breaking tgsi >> changes like this. > > I thought I had sufficiently explained to you that the resource thing > was going away -- it was never spec'd, and there were no users.Yes I was aware that changes were coming.> Perhaps I didn't make it clear enough though -- sorry. The reality is > that TGSI is a living thing, and things change every so often. I don't > think people would be comfortable with locking down TGSI in such a > way.Hmm, I do think that if we end up using a llvm-ir -> tgsi step in some cases that we do need to lock TGSI down. Taken the RES thing for example, it would have been pretty trivial actually to not nuke it, and instead just add the image support next to it. Note I'm fine with making changes while we are figuring this out, but once I start pushing the llvm bits to llvm upstream (which is still months away) we are going to need some sort of TGSI stability. Regards, Hans
Ilia Mirkin
2016-Feb-19 19:43 UTC
[Nouveau] Dealing with opencl kernel parameters in nouveau now that RES support is gone
On Fri, Feb 19, 2016 at 5:36 AM, Hans de Goede <hdegoede at redhat.com> wrote:> Hi, > > On 18-02-16 17:39, Ilia Mirkin wrote: >> >> On Thu, Feb 18, 2016 at 9:45 AM, Hans de Goede <hdegoede at redhat.com> >> wrote: >>> >>> But this does not seem to be hooked up yet for nouveau. >> >> >> Samuel has patches. See >> https://cgit.freedesktop.org/~hakzsam/mesa/log/?h=arb_compute_shader_v3 > > > Cool, I will take a look at those. > >>> So some questions: >>> -The commit by Samual says: >>> This introduces TGSI_FILE_MEMORY for shared, global and local memory. >>> Only >>> shared memory is currently supported. >>> >>> The commit introduces MEMORY[x] and MEMORY[x],SHARED so in reality it >>> also >>> introduces >>> a second option next to shared, so what are we going to use plain >>> MEMORY[x] >>> for? >>> I suggest using it for global memory but we need to be in agreement on >>> this. >> >> >> That sounds fine to me. However what I had in mind was switching the >> SHARED field into a 2-bit field and making it >> >> 1 = SHARED >> 2 = GLOBAL >> 3 = LOCAL >> >> (since for OpenCL you also need to be able to address local or private >> memory). I sorta wanted Samuel to do it, but since I had no idea where >> you were at, or if you were even still working on this, I figured it >> should be fixed up by the first person who needed it. > > > Sounds good, only the naming is somewhat unfortunate since opencl uses > different > naming. I.e. it has no "shared"Sad. Well, "shared" is what OpenGL compute shaders use, which is why I proposed it.> > OpenCL has: > -global: accessible by all worker-groups as well as by the CPU > -const: read-only global > -local: shared by worker-items in the same worker-group, not shared > between worker-groups > -private: accessible only to a single worker-item > > So how do these map to the TGSI: > >> 1 = SHARED >> 2 = GLOBAL >> 3 = LOCALOpenCL global = TGSI global OpenCL const = TGSI global OpenCL local = TGSI shared OpenCL private = TGSI local Not sure what the distinction is between OpenCL const and global is. If the const stuff is actually just user-supplied uniforms (and doesn't need to be in a particular place in memory), then those should go into TGSI CONST somehow.>> >>> >>> -What about kernel input parameters, so far these have been using >>> RES[32764] >>> I must admit that I do not understand where the file_index of 32764 >>> comes >>> from (or where any of the file indexes come from for >>> src/gallium/tests/trivial/compute.c ? >>> I have the feeling that these are not used at all, and everything >>> simply >>> goes >>> to a flat (virtual) memory space, with the params at address 0, correct >>> ? >> >> >> It was never particularly well-specified, which was one of the reasons >> it went away. It also didn't map nicely onto the OpenGL model. There >> is a remaining question of how to do addressing in memory... there's >> 40 bits of address space. Should these implicitly be U64 >> (dual-component in TGSI) addresses that are passed around? Not sure >> what the OpenCL position on all this is. > > > So far I've been using U32 for addresses as that is what Francisco's > original > code was using. And this also is what things like the tgsi LOAD instruction > take. If you're doing a LOAD on a 1d buffer then you will use TEMP[#].x to > specify the index, and the way how this currently works with OpenCL is that > clCreateBuffer() will return a cl_mem type which then gets passed into > the kernel as input parameter and gets treated as a pointer by the compiler, > so e.g. global mem gets treated as a single address space even if there > are multiple global buffers and TEMP[#].x contains the value passed in > via cl_mem as start offset for the buffer + the index into the buffer. > > So this means that currently we are limited to U32 since TEMP[#].x is > only 32 bits wide. Internally 40 bits addresses can and should probably > be used so that at least the different memory spaces each have the full > 32 bits available. > > Note that we could fix this by adding some sort of LOAD64 opcode, which > uses TEMP[#].x and TEMP[#].y as address for 1d buffers, I'm not sure > how this would work for 3d buffers though. I foresee the llvm backend > eventually getting a 64 bit mode where it will use 64 bits for all > pointers and use something like a LOAD64 opcode to indicate that > the indexes (which it effectively uses as addresses / pointers) > are 64 bit wide.Well, this LOAD64/STORE64 would just only be defined for MEMORY[] src/dest, so you don't need to worry about 3d or anything like that. I believe this is a good solution to the problem.> >> Also it would be highly preferable to avoid using GLOBAL at all in the >> first place, and just use BUFFER[] or IMAGE[] or whatever. However >> after having a look at how OpenCL works, I don't think that's >> possible. But perhaps I missed something? > > > OpenCL is very C-like and as such uses a flat address space per memory > space, getting around that is going to be very tricky, if possible at > all.That was my assessment as well.> >> Perhaps I didn't make it clear enough though -- sorry. The reality is >> that TGSI is a living thing, and things change every so often. I don't >> think people would be comfortable with locking down TGSI in such a >> way. > > > Hmm, I do think that if we end up using a llvm-ir -> tgsi step in some > cases that we do need to lock TGSI down. Taken the RES thing for example, > it would have been pretty trivial actually to not nuke it, and instead just > add the image support next to it. Note I'm fine with making changes while > we are figuring this out, but once I start pushing the llvm bits to llvm > upstream (which is still months away) we are going to need some sort of > TGSI stability.I can't unilaterally make such promises. This would have to be discussed and approved by the other gallium/tgsi stakeholders. -ilia
Hans de Goede
2016-Feb-22 11:26 UTC
[Nouveau] Dealing with opencl kernel parameters in nouveau now that RES support is gone
Hi, On 19-02-16 20:43, Ilia Mirkin wrote:> On Fri, Feb 19, 2016 at 5:36 AM, Hans de Goede <hdegoede at redhat.com> wrote: >> Hi, >> >> On 18-02-16 17:39, Ilia Mirkin wrote: >>> >>> On Thu, Feb 18, 2016 at 9:45 AM, Hans de Goede <hdegoede at redhat.com> >>> wrote: >>>> >>>> But this does not seem to be hooked up yet for nouveau. >>> >>> >>> Samuel has patches. See >>> https://cgit.freedesktop.org/~hakzsam/mesa/log/?h=arb_compute_shader_v3 >> >> >> Cool, I will take a look at those. >> >>>> So some questions: >>>> -The commit by Samual says: >>>> This introduces TGSI_FILE_MEMORY for shared, global and local memory. >>>> Only >>>> shared memory is currently supported. >>>> >>>> The commit introduces MEMORY[x] and MEMORY[x],SHARED so in reality it >>>> also >>>> introduces >>>> a second option next to shared, so what are we going to use plain >>>> MEMORY[x] >>>> for? >>>> I suggest using it for global memory but we need to be in agreement on >>>> this. >>> >>> >>> That sounds fine to me. However what I had in mind was switching the >>> SHARED field into a 2-bit field and making it >>> >>> 1 = SHARED >>> 2 = GLOBAL >>> 3 = LOCAL >>> >>> (since for OpenCL you also need to be able to address local or private >>> memory). I sorta wanted Samuel to do it, but since I had no idea where >>> you were at, or if you were even still working on this, I figured it >>> should be fixed up by the first person who needed it. >> >> >> Sounds good, only the naming is somewhat unfortunate since opencl uses >> different >> naming. I.e. it has no "shared" > > Sad. Well, "shared" is what OpenGL compute shaders use, which is why I > proposed it. > >> >> OpenCL has: >> -global: accessible by all worker-groups as well as by the CPU >> -const: read-only global >> -local: shared by worker-items in the same worker-group, not shared >> between worker-groups >> -private: accessible only to a single worker-item >> >> So how do these map to the TGSI: >> >>> 1 = SHARED >>> 2 = GLOBAL >>> 3 = LOCAL > > OpenCL global = TGSI global > OpenCL const = TGSI global > OpenCL local = TGSI shared > OpenCL private = TGSI local > > Not sure what the distinction is between OpenCL const and global is. > If the const stuff is actually just user-supplied uniforms (and > doesn't need to be in a particular place in memory), then those should > go into TGSI CONST somehow.AFAIK OpenCL const is really read-only global, so the data is filled in by the CPU, then passed to the opencl-kernel running on the GPU where all worker-items have access to it. I think that TGSI CONST might indeed be usable for this, but it is probably easiest to treat it as GLOBAL for now.>>>> -What about kernel input parameters, so far these have been using >>>> RES[32764] >>>> I must admit that I do not understand where the file_index of 32764 >>>> comes >>>> from (or where any of the file indexes come from for >>>> src/gallium/tests/trivial/compute.c ? >>>> I have the feeling that these are not used at all, and everything >>>> simply >>>> goes >>>> to a flat (virtual) memory space, with the params at address 0, correct >>>> ? >>> >>> >>> It was never particularly well-specified, which was one of the reasons >>> it went away. It also didn't map nicely onto the OpenGL model. There >>> is a remaining question of how to do addressing in memory... there's >>> 40 bits of address space. Should these implicitly be U64 >>> (dual-component in TGSI) addresses that are passed around? Not sure >>> what the OpenCL position on all this is. >> >> >> So far I've been using U32 for addresses as that is what Francisco's >> original >> code was using. And this also is what things like the tgsi LOAD instruction >> take. If you're doing a LOAD on a 1d buffer then you will use TEMP[#].x to >> specify the index, and the way how this currently works with OpenCL is that >> clCreateBuffer() will return a cl_mem type which then gets passed into >> the kernel as input parameter and gets treated as a pointer by the compiler, >> so e.g. global mem gets treated as a single address space even if there >> are multiple global buffers and TEMP[#].x contains the value passed in >> via cl_mem as start offset for the buffer + the index into the buffer. >> >> So this means that currently we are limited to U32 since TEMP[#].x is >> only 32 bits wide. Internally 40 bits addresses can and should probably >> be used so that at least the different memory spaces each have the full >> 32 bits available. >> >> Note that we could fix this by adding some sort of LOAD64 opcode, which >> uses TEMP[#].x and TEMP[#].y as address for 1d buffers, I'm not sure >> how this would work for 3d buffers though. I foresee the llvm backend >> eventually getting a 64 bit mode where it will use 64 bits for all >> pointers and use something like a LOAD64 opcode to indicate that >> the indexes (which it effectively uses as addresses / pointers) >> are 64 bit wide. > > Well, this LOAD64/STORE64 would just only be defined for MEMORY[] > src/dest, so you don't need to worry about 3d or anything like that. I > believe this is a good solution to the problem.Right for MEMORY this will work fine. I've no clue yet how images will work with OpenCL though, hopefully we can avoid the one flat address space thing there, but I simply don't know yet.>>> Also it would be highly preferable to avoid using GLOBAL at all in the >>> first place, and just use BUFFER[] or IMAGE[] or whatever. However >>> after having a look at how OpenCL works, I don't think that's >>> possible. But perhaps I missed something? >> >> >> OpenCL is very C-like and as such uses a flat address space per memory >> space, getting around that is going to be very tricky, if possible at >> all. > > That was my assessment as well. > >> >>> Perhaps I didn't make it clear enough though -- sorry. The reality is >>> that TGSI is a living thing, and things change every so often. I don't >>> think people would be comfortable with locking down TGSI in such a >>> way. >> >> >> Hmm, I do think that if we end up using a llvm-ir -> tgsi step in some >> cases that we do need to lock TGSI down. Taken the RES thing for example, >> it would have been pretty trivial actually to not nuke it, and instead just >> add the image support next to it. Note I'm fine with making changes while >> we are figuring this out, but once I start pushing the llvm bits to llvm >> upstream (which is still months away) we are going to need some sort of >> TGSI stability. > > I can't unilaterally make such promises. This would have to be > discussed and approved by the other gallium/tgsi stakeholders.Understood. So back to the problem of getting OpenCL(ish) code to work again with the recent mesa changes. For starters I would like to get: src/gallium/tests/trivial/compute.c and then the test with mask 8, test_input_global() to work again, when that is working I should be able to adjust my llvm work (and if necessary clover) to start to work again. Currently the test_input_global() test uses the following bit of TGSI code: COMP DCL SV[0], THREAD_ID[0] DCL TEMP[0], LOCAL DCL TEMP[1], LOCAL IMM UINT32 { 8, 0, 0, 0 } BGNSUB\n" UMUL TEMP[0], SV[0], IMM[0] LOAD TEMP[1].xy, RINPUT, TEMP[0] LOAD TEMP[0].x, RGLOBAL, TEMP[1].yyyy UADD TEMP[1].x, TEMP[0], -TEMP[1] STORE RGLOBAL.x, TEMP[1].yyyy, TEMP[1] RET ENDSUB Where by RINPUT and RGLOBAL get replaces by processing the code with cpp and the following defines: #define RGLOBAL RES[32767] #define RLOCAL RES[32766] #define RPRIVATE RES[32765] #define RINPUT RES[32764] If I understand how memory is supposed to work, then I would need to change the TGSI as follows: COMP DCL SV[0], THREAD_ID[0] DCL MEMORY[0] DCL TEMP[0], LOCAL DCL TEMP[1], LOCAL IMM UINT32 { 8, 0, 0, 0 } BGNSUB\n" UMUL TEMP[0], SV[0], IMM[0] LOAD TEMP[1].xy, RINPUT, TEMP[0] LOAD TEMP[0].x, MEMORY[0], TEMP[1].yyyy UADD TEMP[1].x, TEMP[0], -TEMP[1] STORE MEMORY[0].x, TEMP[1].yyyy, TEMP[1] RET ENDSUB This assumes, that as discussed declaring memory without a , SHARED or other flag means the memory is global. So 2 questions: 1) Do the above changes for using the new MEMORY keyword look as intended to you? 2) This only solves the accessing of the global memory, it does not solve getting to the kernel input kernel parameters, how would I deal with those ? Currently the kernel input parameters are uploaded by src/gallium/drivers/nouveau/nv50/nv50_compute.c: nv50_compute_upload_input() Are we going to change how these parameters get uploaded ? And if not how do we get to these parameters from TGSI now that we no longer have RES[#] to do so ? I think we should also keep in mind that what works for nouveau may not work for other GPU-s, so we should either introduce a MEMORY[x], INPUT or a new DCL INPUT[0] for this, so that the code for other GPU-s can have its own handling for this. Regards, Hans
Maybe Matching Threads
- Dealing with opencl kernel parameters in nouveau now that RES support is gone
- Dealing with opencl kernel parameters in nouveau now that RES support is gone
- Dealing with opencl kernel parameters in nouveau now that RES support is gone
- Some llvm questions (for tgsi backend)
- Some llvm questions (for tgsi backend)