ARM float32 -> float16 lazarus / 33-macOS - Mac OS X ---------------------------------------------------------------------- ChrisR · 2021-01-17 14:49 Hello all,   Can anyone help me convert this tiny C program that calls the Neon "vcvt_f32_f16" instruction, e.g. converting to FPC assembly?   My OpenGL code uses GL_INT_2_10_10_10_REV datatype for normals which is not supported on the M1. I can use GL_FLOAT (float32), but that is wasteful. Therefore, I would like to use GL_HALF (float16). Included is a simple C program which calls the Neon SIMD instruction. My Pascal uses scalar code but the results appear different. Perhaps this reflects that Arm supports two half-precision (16-bit) floating-point scalar data types:  The IEEE 754-2008 __fp16 data type, defined in the Arm C Language Extensions.  The _Float16 data type, defined in the C11 extension ISO/IEC TS 18661-3:2015 (the attached file "f16.inc" should be renamed "f16.c", extension changed to allow upload). $gcc f16.c -o f16; ./f16 f32: 32.012 64.094 12.096 3.14159274 f16: 0x00000020 0x00000040 0x0000000c 0x00000003 f32: 32 64.125 12.0938 3.14062500 $fpc f16.pas; ./f16    Free Pascal Compiler version 3.3.1 [2021/01/01] for aarch64 Copyright (c) 1993-2020 by Florian Klaempfl and others Target OS: Darwin for AArch64 Compiling f16.pas Assembling f16 Linking f16 200 lines compiled, 0.5 sec f32: 32.0120010375977 64.0940017700195 12.0959997177124 3.14159274101257 f16: 0x5000 0x5402 0x4A0C 0x4248 f32: 32 64.125 12.09375 3.140625 ---------------------------------------------------------------------- Laksen · 2021-01-17 16:16 Something like this might be possible.. It's very ugly but it bypasses the current lack of advanced simd/fp16 handling in the aarch64 backend Code: [Select]function f32tof16Neon(v: single): word;assembler; nostackframe; asm     .long 0x1e23c000 //fcvt h0, s0     .long 0x1ee60000 //fmov w0, h0 end; function f16tof32Neon(v: word): single;assembler; nostackframe; asm     .long 0x1ee70000 // fmov h0, w0     .long 0x1ee24000 // fcvt s0, h0 end; procedure f32tof16Neon(var output: float16x4; constref f32x4: float32x4 ); begin     output.x := f32tof16Neon(f32x4.x);     output.y := f32tof16Neon(f32x4.y);     output.z := f32tof16Neon(f32x4.z);     output.w := f32tof16Neon(f32x4.w); end; procedure f16tof32Neon(var output: float32x4; constref f16x4: float16x4 ); begin     output.x := f16tof32Neon(f16x4.x);     output.y := f16tof32Neon(f16x4.y);     output.z := f16tof32Neon(f16x4.z);     output.w := f16tof32Neon(f16x4.w); end; ---------------------------------------------------------------------- ChrisR · 2021-01-17 16:41 1. Laksen: Thanks for your swift reply. 2. I have attached a copy of Laksen's elegant solution (f16.pas) which using scalar instructions not the Neon SIMD. 3. The text display of the hex codes still does not match my C code, though maybe I am not typecasting correctly in C. I will see how your solution looks with OpenGL. 4. I would be interested if anyone has a SIMD solution, as my input arrays are DTI tractography maps with thousands of values. fpc f16.pas; ./f16    Free Pascal Compiler version 3.3.1 [2021/01/01] for aarch64 Copyright (c) 1993-2020 by Florian Klaempfl and others Target OS: Darwin for AArch64 Compiling f16.pas Assembling f16 Linking f16 59 lines compiled, 0.2 sec f32: 32.0120010375977 64.0940017700195 12.0959997177124 3.14159274101257 f16: 0x5000 0x5402 0x4A0C 0x4248 f32: 32 64.125 12.09375 3.140625 ---------------------------------------------------------------------- ChrisR · 2021-01-17 17:02 Laksen   Your solution works perfectly with OpenGL. Thanks a lot! The conversion is only done once when the mesh is loaded to the GPU, so I can live with the scalar solution (though I would like to see if anyone can provide the NEON simd solution). Thanks again for such a rapid and elegant response! ---------------------------------------------------------------------- Laksen · 2021-01-18 15:44 The reason the hex display doesn't work is because your C version is casting the fp16 value to an integer which simply returns the rounded value converted to an integer (0x00000020 is 32). Try this: Code: [Select]printf("f16: 0x%04x 0x%04x 0x%04x 0x%04x\n", *(uint16_t*)(&f16x4[0]), *(uint16_t*)(&f16x4[1]),*(uint16_t*)(&f16x4[2 ]),*(uint16_t*)(&f16x4[3])); Maybe this will work for the SIMD version: Code: [Select]procedure f32tof16Neon2(var output: float16x4; constref f32x4: float32x4 );assembler; nostackframe; asm ldr q0, [x1] .long 0x0e216800 // fcvtn   v0.4h, v0.4s str d0, [x0] end; It assumes that both pointers have proper alignment but that should be easy to ensure ---------------------------------------------------------------------- ChrisR · 2021-01-18 16:40 Thanks for the help with C. I can confirm your Neon SIMD code works. Both the scalar and SIMD are exceptionally fast. My arrays and element data are aligned, but it is a good caveat to mention. Thanks again for your help. The upcoming release of Surfice will use you code and avoids the major performance penalty of GL_INT_2_10_10_10_REV. You have solved my issue.