From 27c5bdbb5cfa088dc978db9f28c1bc636595ee8b Mon Sep 17 00:00:00 2001
From: Jeroen Ubbink <jeroen@blackbyte.nl>
Date: Tue, 14 Jul 2026 09:28:04 +0200
Subject: [PATCH] Fix build with CUDA 13.1 and newer glibc

nvcc's crt/math_functions.h declares rsqrt/rsqrtf without noexcept,
which conflicts with glibc >= 2.39's C23 rsqrt declaration (pulled in
via the _GNU_SOURCE that g++ defines by default), causing a hard
compile error. Compile GPURummage.cu with -Xcompiler -U_GNU_SOURCE to
avoid the conflict.

Also make CUDA and CCAP overridable (?=) instead of hardcoded, and
default CUDA to /usr/local/cuda so the Makefile isn't pinned to one
toolkit version.

Verified: builds cleanly against CUDA 13.1 and successfully mines
matching npub addresses on an RTX 5070 Ti (CCAP=120).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---
 Makefile      |  6 +++---
 docs/BUILD.md | 24 +++++++++++++++++++-----
 2 files changed, 22 insertions(+), 8 deletions(-)

diff --git a/Makefile b/Makefile
index 75b5fda..5d53e79 100644
--- a/Makefile
+++ b/Makefile
@@ -19,8 +19,8 @@ OBJET = $(addprefix $(OBJDIR)/, \
         rummage.o \
 )
 
-CCAP      = 86
-CUDA      = /usr/local/cuda-11.8
+CCAP     ?= 86
+CUDA     ?= /usr/local/cuda
 CXX       = g++
 CXXCUDA   = /usr/bin/g++
 
@@ -42,7 +42,7 @@ NVCC      = $(CUDA)/bin/nvcc
 all: rummage
 
 $(OBJDIR)/GPU/GPURummage.o: $(SRCDIR)/GPU/GPURummage.cu
-	$(NVCC) -allow-unsupported-compiler --compile --compiler-options -fPIC -ccbin $(CXXCUDA) -m64 -O2 -I$(SRCDIR) -I$(CUDA)/include \
+	$(NVCC) -allow-unsupported-compiler --compile --compiler-options -fPIC -Xcompiler -U_GNU_SOURCE -ccbin $(CXXCUDA) -m64 -O2 -I$(SRCDIR) -I$(CUDA)/include \
 	-DNOSTR_BLOCKS_PER_GRID=$(NOSTR_BLOCKS_PER_GRID) \
 	-DNOSTR_THREADS_PER_BLOCK=$(NOSTR_THREADS_PER_BLOCK) \
 	-DKEYS_PER_THREAD_BATCH=$(KEYS_PER_THREAD_BATCH) \
diff --git a/docs/BUILD.md b/docs/BUILD.md
index 7db9fb9..afaa8bb 100644
--- a/docs/BUILD.md
+++ b/docs/BUILD.md
@@ -9,7 +9,7 @@ For known good configurations and performance results, see [PERFORMANCE.md](PERF
 ### Linux (CUDA)
 
 **Required:**
-- CUDA Toolkit (11.x or 12.x)
+- CUDA Toolkit (11.x, 12.x, or 13.x)
 - GMP library (GNU Multiple Precision)
 - g++ compiler
 
@@ -112,6 +112,7 @@ CCAP = 86
 ```
 
 **Common compute capabilities:**
+- RTX 50-series (Blackwell): `120`
 - RTX 40-series (Ada): `89`
 - RTX 30-series (Ampere): `86`
 - RTX 20-series (Turing): `75`
@@ -120,13 +121,17 @@ CCAP = 86
 
 Find your GPU's compute capability at: https://developer.nvidia.com/cuda-gpus
 
+`CCAP` can also be overridden without editing the Makefile, e.g. `make CCAP=120`.
+
 ## CUDA Installation Path
 
-Update the CUDA path if installed in a non-standard location:
+The Makefile defaults to `CUDA = /usr/local/cuda`, which is normally a symlink to
+whichever version you last installed. If you have multiple CUDA versions installed
+and need a specific one, override it on the command line instead of editing the
+Makefile:
 
-**In Makefile, line 21:**
-```make
-CUDA = /usr/local/cuda-11.8
+```bash
+make CUDA=/usr/local/cuda-11.8
 ```
 
 Check your CUDA version:
@@ -134,6 +139,15 @@ Check your CUDA version:
 nvcc --version
 ```
 
+**Newer glibc (2.39+) build error:**
+```
+error: exception specification is incompatible with that of previous function "rsqrt"
+```
+This happens because recent glibc versions declare `rsqrt`/`rsqrtf` (C23 extensions)
+with a different exception specification than CUDA's `math_functions.h`. The Makefile
+already compiles `GPURummage.cu` with `-Xcompiler -U_GNU_SOURCE` to avoid pulling in
+those glibc declarations, which resolves the conflict on CUDA 11.8 through 13.x.
+
 ## GPU Performance Tuning
 
 For optimal performance, tune these parameters in the Makefile: