Re: [Mesa-dev] [PATCH 0/6] Prehash all the things
Thank you Thomas for the update, then. Awaiting your coming work. Regards, Dieter Am 10.10.2017 15:22, schrieb Thomas Helland: Hi! Thanks for keeping up with the long wait =) I revisited this not too long ago, and found that with the new pointer hashing function the benefits are zero to negative from this series. I've reduced it to only the instruction set and the string_to_uint_map patch but it's not convincing. I suspect we are seeing cache miss vs hashing tradeoffs. So I've basically put it to rest for now. Might give it another go sometime, but I think right now the effort is better spent elsewhere in the codebase. Greetings, Thomas 2017-10-09 14:02 GMT+02:00 Dieter Nützel : Hello Thomas, now, that you have write commit 'only' this one is missing. Maybe you have time for this. Latest version do not apply any longer. Wende an: util: Avoid computing hash twice in string_to_uint_map error: src/util/string_to_uint_map.h ist nicht im Index Anwendung des Patches fehlgeschlagen bei 0006 util: Avoid computing hash twice in string_to_uint_map Greetings, Dieter Am 19.06.2017 18:09, schrieb Dieter Nützel: Ping! Any news, reviews --- anyone? I'm running this all day without a hitch. Cheers, Dieter Am 23.05.2017 05:40, schrieb Dieter Nützel: For the series: Tested-by: Dieter Nützel on radeonsi/RX580 Unigine_Heaven-4.0, Unigine_Valley-1.0, Unigine_Superposition-1.0, LS2015 (Wine-staging), Mesa-demos (objviewer) Dieter Am 22.05.2017 20:55, schrieb Thomas Helland: While this doesn't prehash all the things, it does switch quite a lot of places from doing a search and then a subsequent insert to first hash the key, and then use this hash when searching / inserting. While our new pointer hashing function remedied much of our overhead hashing pointers, there is still more to gain here. This cuts executed instructions / task-clock by about 0.5% on a shader-db run on my i965 running machine. While that's not a lot, it is still a nice little improvement on the way to less overhead. The changes should also be fairly trivial, so it's not much of a burden. Thomas Helland (6): glsl: Prehash in refcount hash table to reduce hashing nir: Prehash in instr_set to avoid hashing twice glsl: Prehash in constant propagation glsl: Prehash in constant variable pass to avoid hashing twice glsl: Prehash to avoid computing the hash twice util: Avoid computing hash twice in string_to_uint_map src/compiler/glsl/ir_variable_refcount.cpp | 7 +-- src/compiler/glsl/opt_constant_propagation.cpp | 8 +--- src/compiler/glsl/opt_constant_variable.cpp | 6 -- src/compiler/glsl/opt_copy_propagation_elements.cpp | 7 +-- src/compiler/nir/nir_instr_set.c| 7 +-- src/util/string_to_uint_map.h | 9 ++--- 6 files changed, 30 insertions(+), 14 deletions(-) ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 0/6] Prehash all the things
Hi! Thanks for keeping up with the long wait =) I revisited this not too long ago, and found that with the new pointer hashing function the benefits are zero to negative from this series. I've reduced it to only the instruction set and the string_to_uint_map patch but it's not convincing. I suspect we are seeing cache miss vs hashing tradeoffs. So I've basically put it to rest for now. Might give it another go sometime, but I think right now the effort is better spent elsewhere in the codebase. Greetings, Thomas 2017-10-09 14:02 GMT+02:00 Dieter Nützel : > Hello Thomas, > > now, that you have write commit 'only' this one is missing. > Maybe you have time for this. > > Latest version do not apply any longer. > > Wende an: util: Avoid computing hash twice in string_to_uint_map > error: src/util/string_to_uint_map.h ist nicht im Index > Anwendung des Patches fehlgeschlagen bei 0006 util: Avoid computing hash > twice in string_to_uint_map > > Greetings, > Dieter > > > Am 19.06.2017 18:09, schrieb Dieter Nützel: >> >> Ping! >> >> Any news, reviews --- anyone? >> >> I'm running this all day without a hitch. >> >> Cheers, >> Dieter >> >> Am 23.05.2017 05:40, schrieb Dieter Nützel: >>> >>> For the series: >>> >>> Tested-by: Dieter Nützel >>> >>> on radeonsi/RX580 >>> >>> Unigine_Heaven-4.0, Unigine_Valley-1.0, Unigine_Superposition-1.0, >>> LS2015 (Wine-staging), Mesa-demos (objviewer) >>> >>> Dieter >>> >>> Am 22.05.2017 20:55, schrieb Thomas Helland: While this doesn't prehash all the things, it does switch quite a lot of places from doing a search and then a subsequent insert to first hash the key, and then use this hash when searching / inserting. While our new pointer hashing function remedied much of our overhead hashing pointers, there is still more to gain here. This cuts executed instructions / task-clock by about 0.5% on a shader-db run on my i965 running machine. While that's not a lot, it is still a nice little improvement on the way to less overhead. The changes should also be fairly trivial, so it's not much of a burden. Thomas Helland (6): glsl: Prehash in refcount hash table to reduce hashing nir: Prehash in instr_set to avoid hashing twice glsl: Prehash in constant propagation glsl: Prehash in constant variable pass to avoid hashing twice glsl: Prehash to avoid computing the hash twice util: Avoid computing hash twice in string_to_uint_map src/compiler/glsl/ir_variable_refcount.cpp | 7 +-- src/compiler/glsl/opt_constant_propagation.cpp | 8 +--- src/compiler/glsl/opt_constant_variable.cpp | 6 -- src/compiler/glsl/opt_copy_propagation_elements.cpp | 7 +-- src/compiler/nir/nir_instr_set.c| 7 +-- src/util/string_to_uint_map.h | 9 ++--- 6 files changed, 30 insertions(+), 14 deletions(-) >>> >>> ___ >>> mesa-dev mailing list >>> mesa-dev@lists.freedesktop.org >>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev >> >> ___ >> mesa-dev mailing list >> mesa-dev@lists.freedesktop.org >> https://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 0/6] Prehash all the things
Hello Thomas, now, that you have write commit 'only' this one is missing. Maybe you have time for this. Latest version do not apply any longer. Wende an: util: Avoid computing hash twice in string_to_uint_map error: src/util/string_to_uint_map.h ist nicht im Index Anwendung des Patches fehlgeschlagen bei 0006 util: Avoid computing hash twice in string_to_uint_map Greetings, Dieter Am 19.06.2017 18:09, schrieb Dieter Nützel: Ping! Any news, reviews --- anyone? I'm running this all day without a hitch. Cheers, Dieter Am 23.05.2017 05:40, schrieb Dieter Nützel: For the series: Tested-by: Dieter Nützel on radeonsi/RX580 Unigine_Heaven-4.0, Unigine_Valley-1.0, Unigine_Superposition-1.0, LS2015 (Wine-staging), Mesa-demos (objviewer) Dieter Am 22.05.2017 20:55, schrieb Thomas Helland: While this doesn't prehash all the things, it does switch quite a lot of places from doing a search and then a subsequent insert to first hash the key, and then use this hash when searching / inserting. While our new pointer hashing function remedied much of our overhead hashing pointers, there is still more to gain here. This cuts executed instructions / task-clock by about 0.5% on a shader-db run on my i965 running machine. While that's not a lot, it is still a nice little improvement on the way to less overhead. The changes should also be fairly trivial, so it's not much of a burden. Thomas Helland (6): glsl: Prehash in refcount hash table to reduce hashing nir: Prehash in instr_set to avoid hashing twice glsl: Prehash in constant propagation glsl: Prehash in constant variable pass to avoid hashing twice glsl: Prehash to avoid computing the hash twice util: Avoid computing hash twice in string_to_uint_map src/compiler/glsl/ir_variable_refcount.cpp | 7 +-- src/compiler/glsl/opt_constant_propagation.cpp | 8 +--- src/compiler/glsl/opt_constant_variable.cpp | 6 -- src/compiler/glsl/opt_copy_propagation_elements.cpp | 7 +-- src/compiler/nir/nir_instr_set.c| 7 +-- src/util/string_to_uint_map.h | 9 ++--- 6 files changed, 30 insertions(+), 14 deletions(-) ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 0/6] Prehash all the things
Ping! Any news, reviews --- anyone? I'm running this all day without a hitch. Cheers, Dieter Am 23.05.2017 05:40, schrieb Dieter Nützel: For the series: Tested-by: Dieter Nützel on radeonsi/RX580 Unigine_Heaven-4.0, Unigine_Valley-1.0, Unigine_Superposition-1.0, LS2015 (Wine-staging), Mesa-demos (objviewer) Dieter Am 22.05.2017 20:55, schrieb Thomas Helland: While this doesn't prehash all the things, it does switch quite a lot of places from doing a search and then a subsequent insert to first hash the key, and then use this hash when searching / inserting. While our new pointer hashing function remedied much of our overhead hashing pointers, there is still more to gain here. This cuts executed instructions / task-clock by about 0.5% on a shader-db run on my i965 running machine. While that's not a lot, it is still a nice little improvement on the way to less overhead. The changes should also be fairly trivial, so it's not much of a burden. Thomas Helland (6): glsl: Prehash in refcount hash table to reduce hashing nir: Prehash in instr_set to avoid hashing twice glsl: Prehash in constant propagation glsl: Prehash in constant variable pass to avoid hashing twice glsl: Prehash to avoid computing the hash twice util: Avoid computing hash twice in string_to_uint_map src/compiler/glsl/ir_variable_refcount.cpp | 7 +-- src/compiler/glsl/opt_constant_propagation.cpp | 8 +--- src/compiler/glsl/opt_constant_variable.cpp | 6 -- src/compiler/glsl/opt_copy_propagation_elements.cpp | 7 +-- src/compiler/nir/nir_instr_set.c| 7 +-- src/util/string_to_uint_map.h | 9 ++--- 6 files changed, 30 insertions(+), 14 deletions(-) ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 0/6] Prehash all the things
For the series: Tested-by: Dieter Nützel on radeonsi/RX580 Unigine_Heaven-4.0, Unigine_Valley-1.0, Unigine_Superposition-1.0, LS2015 (Wine-staging), Mesa-demos (objviewer) Dieter Am 22.05.2017 20:55, schrieb Thomas Helland: While this doesn't prehash all the things, it does switch quite a lot of places from doing a search and then a subsequent insert to first hash the key, and then use this hash when searching / inserting. While our new pointer hashing function remedied much of our overhead hashing pointers, there is still more to gain here. This cuts executed instructions / task-clock by about 0.5% on a shader-db run on my i965 running machine. While that's not a lot, it is still a nice little improvement on the way to less overhead. The changes should also be fairly trivial, so it's not much of a burden. Thomas Helland (6): glsl: Prehash in refcount hash table to reduce hashing nir: Prehash in instr_set to avoid hashing twice glsl: Prehash in constant propagation glsl: Prehash in constant variable pass to avoid hashing twice glsl: Prehash to avoid computing the hash twice util: Avoid computing hash twice in string_to_uint_map src/compiler/glsl/ir_variable_refcount.cpp | 7 +-- src/compiler/glsl/opt_constant_propagation.cpp | 8 +--- src/compiler/glsl/opt_constant_variable.cpp | 6 -- src/compiler/glsl/opt_copy_propagation_elements.cpp | 7 +-- src/compiler/nir/nir_instr_set.c| 7 +-- src/util/string_to_uint_map.h | 9 ++--- 6 files changed, 30 insertions(+), 14 deletions(-) ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
