From f92de5e81ef9dea7d4bbd53faec55ebcf1939440 Mon Sep 17 00:00:00 2001 From: Emil Muratov Date: Mon, 7 Dec 2020 16:41:33 +0300 Subject: fillrate.md with some test details Signed-off-by: Emil Muratov --- fillrate.md | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+) create mode 100644 fillrate.md diff --git a/fillrate.md b/fillrate.md new file mode 100644 index 0000000..b07e1af --- /dev/null +++ b/fillrate.md @@ -0,0 +1,25 @@ +## Estimating fillrate + +Here are some results of simple tests on filling DMA buffer with data. +Filling DMA buffer requres lot's of memory operations on a bit level rather than doing simple byte/word wide store and copy. And it looks like it's quite a task both for esp32 core and compiler. +I've done this while optimizing loops and bit logic along with testing compiler results. + +So the testbed is: + - Matrix modules: 4 x FM6126A based 64x64 modules chained in 256x64 + +A simple sketch: + - allocating single DMA buffs for 256x64 + - allocating (NUM_LEDS*3) bytes for CRGB buffer + - measuring microseconds for the following calls: + - clearScreen() - full blanking + - fillScreenRGB888() with monochrome/gray colors + - filling some gradient into CRGB buff + - painting CRGB buff into DMA buff with looped drawPixelRGB888() + +|Pattern |Reference|Ref+SPEED|Optimized|Optimized+SPEED| +|--|--|--|--|--| +|fillScreenRGB888()|14570|14570|13400 (8.5% faster)|5520 (164% faster)| +|CRGB buff fill|760|760|760|760| +|updateMatrixDMABuffer(CRGB)|7700|32080|55780 (38% faster)|33500(+4.2% slower)| + +to be continued... \ No newline at end of file -- cgit v1.3.1