diff --git a/‎docs/packages/warp-lang.yaml‎
Lines changed: 1 addition & 0 deletions b/‎docs/packages/warp-lang.yaml‎
Lines changed: 1 addition & 0 deletions
diff --git a/‎patches/warp-lang/1.17.0/0001-Add-riscv64-to-the-build-and-packaging-platform-tabl.patch‎
Lines changed: 132 additions & 0 deletions b/‎patches/warp-lang/1.17.0/0001-Add-riscv64-to-the-build-and-packaging-platform-tabl.patch‎
Lines changed: 132 additions & 0 deletions
diff --git a/‎patches/warp-lang/1.17.0/0002-clang-name-the-rv64gc-lp64d-target-ABI-for-JIT-compi.patch‎
Lines changed: 63 additions & 0 deletions b/‎patches/warp-lang/1.17.0/0002-clang-name-the-rv64gc-lp64d-target-ABI-for-JIT-compi.patch‎
Lines changed: 63 additions & 0 deletions
diff --git a/‎patches/warp-lang/1.17.0/0003-Convert-fp16-in-software-on-RISC-V-targets-without-Z.patch‎
Lines changed: 167 additions & 0 deletions b/‎patches/warp-lang/1.17.0/0003-Convert-fp16-in-software-on-RISC-V-targets-without-Z.patch‎
Lines changed: 167 additions & 0 deletions
@@ -21,3 +21,4 @@ versions:
   gpl-sources:
     filename: gpl-sources.tar
     description: gcc
+- version: 1.17.0
@@ -0,0 +1,132 @@
+From 8b64e2e85baff0f1e96beb7b3ad4d25ac8cbd82e Mon Sep 17 00:00:00 2001
+From: Ludovic Henry <git@ludovic.dev>
+Date: Sun, 20 Sep 2026 12:00:00 +0000
+Subject: [PATCH 1/4] Add riscv64 to the build and packaging platform tables
+
+Upstream-Status: To upstream [no riscv64 CI runner or prebuilt Clang/LLVM SDK upstream yet]
+
+machine_architecture() in warp/_src/build_architecture.py (shared by
+build_lib.py, build_llvm.py, build_dll.py and setup.py) raises
+"Unrecognized machine architecture 'riscv64'", which aborts build_lib.py before
+anything is compiled.
+
+setup.py's `platforms` table is what WarpBDistWheel.get_tag() reads to stamp the
+wheel's platform tag, so riscv64 needs an entry there as well. manylinux_2_39 is
+the flavour of the image these wheels are built in
+(quay.io/pypa/manylinux_2_39_riscv64, Rocky Linux 10, glibc 2.39).
+
+build_llvm.py picks LLVM_TARGETS_TO_BUILD from the same canonical architecture
+string and otherwise falls back to X86, which would build an LLVM unable to emit
+code for the machine it runs on; RISCV is the matching backend name.
+
+Nothing else in the tree is architecture-specific: warp/native carries no SIMD
+intrinsics and the Linux compile flags are a generic -O3 -fPIC --std=c++17.
+---
+ build_llvm.py                   |  4 +++-
+ setup.py                        | 11 ++++++-----
+ warp/_src/build_architecture.py |  5 +++--
+ 3 files changed, 12 insertions(+), 8 deletions(-)
+
+diff --git a/build_llvm.py b/build_llvm.py
+index 46cc828..7ac9c9b 100644
+--- a/build_llvm.py
++++ b/build_llvm.py
+@@ -166,7 +166,7 @@ def build_llvm_clang_from_source_for_arch(args, arch: str, llvm_source: str) ->
+ 
+     Args:
+         args: Command line arguments
+-        arch: Architecture to build for ("aarch64" or "x86_64")
++        arch: Architecture to build for ("aarch64", "riscv64" or "x86_64")
+         llvm_source: Path to the LLVM source code
+     """
+ 
+@@ -224,6 +224,8 @@ def build_llvm_clang_from_source_for_arch(args, arch: str, llvm_source: str) ->
+ 
+     if arch == "aarch64":
+         target_backend = "AArch64"
++    elif arch == "riscv64":
++        target_backend = "RISCV"
+     else:
+         target_backend = "X86"
+ 
+diff --git a/setup.py b/setup.py
+index 327bae0..998d9da 100644
+--- a/setup.py
++++ b/setup.py
+@@ -29,14 +29,14 @@ parser.add_argument(
+     "-P",
+     type=str,
+     default="",
+-    help="Wheel platform: windows-x86_64|windows-aarch64|linux-x86_64|linux-aarch64|macos-aarch64",
++    help="Wheel platform: windows-x86_64|windows-aarch64|linux-x86_64|linux-aarch64|linux-riscv64|macos-aarch64",
+ )
+ parser.add_argument(
+     "--manylinux",
+     "-M",
+     type=str,
+     default="manylinux_2_28",
+-    help="Manylinux flavor for Linux wheels: manylinux_2_28|manylinux_2_34",
++    help="Manylinux flavor for Linux wheels: manylinux_2_28|manylinux_2_34|manylinux_2_39",
+ )
+ args = parser.parse_known_args()[0]
+ 
+@@ -73,6 +73,7 @@ platforms = [
+     Platform("windows", "aarch64", "Windows ARM64", ".dll", "win_arm64"),
+     Platform("linux", "x86_64", "Linux x86-64", ".so", "manylinux_2_28_x86_64"),
+     Platform("linux", "aarch64", "Linux AArch64", ".so", "manylinux_2_34_aarch64"),
++    Platform("linux", "riscv64", "Linux RISC-V 64", ".so", "manylinux_2_39_riscv64"),
+     Platform("macos", "aarch64", "macOS ARM64", ".dylib", "macosx_11_0_arm64"),
+ ]
+ 
+@@ -129,7 +130,7 @@ if args.command == "bdist_wheel":
+         if len(detected_platforms) > 1:
+             print("Libraries for multiple platforms were detected.")
+             print("Run `python -m build --wheel -C--build-option=-P<platform>` to select a specific one.")
+-            print("Available platforms: windows-x86_64, windows-aarch64, linux-x86_64, linux-aarch64, macos-aarch64")
++            print("Available platforms: windows-x86_64, windows-aarch64, linux-x86_64, linux-aarch64, linux-riscv64, macos-aarch64")
+             # Select the libraries corresponding with the this machine's platform
+             for p in platforms:
+                 if p.os == machine_os() and p.arch == machine_architecture():
+@@ -151,8 +152,8 @@ class WarpBDistWheel(bdist_wheel):
+     # setuptools.Command can validate the command line options.
+     user_options: ClassVar[list[tuple[str, str, str]]] = [
+         *bdist_wheel.user_options,
+-        ("platform=", "P", "Wheel platform: windows-x86_64|windows-aarch64|linux-x86_64|linux-aarch64|macos-aarch64"),
+-        ("manylinux=", "M", "Manylinux flavor for Linux wheels: manylinux_2_28|manylinux_2_34"),
++        ("platform=", "P", "Wheel platform: windows-x86_64|windows-aarch64|linux-x86_64|linux-aarch64|linux-riscv64|macos-aarch64"),
++        ("manylinux=", "M", "Manylinux flavor for Linux wheels: manylinux_2_28|manylinux_2_34|manylinux_2_39"),
+     ]
+ 
+     def initialize_options(self):
+diff --git a/warp/_src/build_architecture.py b/warp/_src/build_architecture.py
+index 9df7173..68775d4 100644
+--- a/warp/_src/build_architecture.py
++++ b/warp/_src/build_architecture.py
+@@ -3,7 +3,7 @@
+ 
+ """Detect and normalize architectures used by Warp builds.
+ 
+-This module maps platform-specific x86-64 and ARM64 names to the
++This module maps platform-specific x86-64, ARM64 and RV64 names to the
+ canonical identifiers shared by Warp's build, packaging, and runtime code.
+ """
+ 
+@@ -12,13 +12,14 @@ from __future__ import annotations
+ import platform
+ from typing import Literal
+ 
+-Architecture = Literal["x86_64", "aarch64"]
++Architecture = Literal["x86_64", "aarch64", "riscv64"]
+ 
+ _ARCHITECTURE_ALIASES: dict[str, Architecture] = {
+     "amd64": "x86_64",
+     "x86_64": "x86_64",
+     "arm64": "aarch64",
+     "aarch64": "aarch64",
++    "riscv64": "riscv64",
+ }
+ 
+ 
+-- 
+2.43.0
+
@@ -0,0 +1,63 @@
+From bb045ca51b3f98290f3194192d5a72fde6af1667 Mon Sep 17 00:00:00 2001
+From: Ludovic Henry <git@ludovic.dev>
+Date: Sun, 20 Sep 2026 12:05:00 +0000
+Subject: [PATCH 2/4] clang: name the rv64gc/lp64d target ABI for JIT-compiled
+ kernels
+
+Upstream-Status: To upstream [no riscv64 runner upstream to regression-test it on]
+
+warp-clang.so drives Clang through CompilerInvocation, i.e. at the cc1 level,
+where none of the driver's per-target defaults are applied. On RV64 that leaves
+clang::RISCVTargetInfo with a bare rv64i ISA (initFeatureMap sees no
+-target-feature flags) and, because the invocation names no ABI, an ABIFLen of
+zero -- the soft-float lp64 ABI.
+
+warp.so, crt.cpp and warp-clang.so itself are compiled by the host gcc for
+rv64gc/lp64d, so every call a JIT-compiled kernel makes into them that passes or
+returns a float or a double -- _wp_isfinite(double) and the other crt.cpp shims
+-- would read the value from the wrong register class. The base ISA is wrong for
+the same reason: without M/A/F/D the frontend emits integer-multiply and
+soft-float libcalls for operations the hardware performs natively.
+
+Named explicitly rather than derived from the triple, because
+LLVM_DEFAULT_TARGET_TRIPLE is whatever LLVM was configured with (build_llvm.py
+sets riscv64-pc-linux) and cc1 infers neither a RISC-V ISA nor an ABI from the
+triple's environment field.
+
+Mirrors the x86_64 (+f16c) and aarch64 (+reserve-x28) blocks beside it.
+---
+ warp/native/clang/clang.cpp | 18 ++++++++++++++++++
+ 1 file changed, 18 insertions(+)
+
+diff --git a/warp/native/clang/clang.cpp b/warp/native/clang/clang.cpp
+index a9fc417..d2ca34f 100644
+--- a/warp/native/clang/clang.cpp
++++ b/warp/native/clang/clang.cpp
+@@ -256,6 +256,24 @@ static std::unique_ptr<clang::CompilerInstance> create_compiler(
+         args.push_back("+f16c");
+ #endif
+ 
++#if defined(__riscv) && __riscv_xlen == 64
++        // cc1 defaults RV64 to a bare rv64i ISA and, with no ABI named, to soft-float lp64.
++        // warp.so and crt.cpp are compiled rv64gc/lp64d, so without this every kernel call
++        // into them (_wp_isfinite(double), ...) would pass floats in the wrong registers.
++        args.push_back("-target-abi");
++        args.push_back("lp64d");
++        args.push_back("-target-feature");
++        args.push_back("+m");
++        args.push_back("-target-feature");
++        args.push_back("+a");
++        args.push_back("-target-feature");
++        args.push_back("+f");
++        args.push_back("-target-feature");
++        args.push_back("+d");
++        args.push_back("-target-feature");
++        args.push_back("+c");
++#endif
++
+ #if defined(__aarch64__) || defined(_M_ARM64)
+         if (tiles_in_stack_memory) {
+             // Static memory support is broken on AArch64 CPUs. As a workaround we reserve some stack memory on kernel
+-- 
+2.43.0
+
@@ -0,0 +1,167 @@
+From 7bd5afd2c143e7db533bafc04c17ef4555c45a4b Mon Sep 17 00:00:00 2001
+From: Ludovic Henry <git@ludovic.dev>
+Date: Mon, 21 Sep 2026 10:08:00 +0000
+Subject: [PATCH 3/4] Convert fp16 in software on RISC-V targets without Zfh
+
+Upstream-Status: To upstream [no riscv64 runner upstream to regression-test it on]
+
+float_to_half()/half_to_float() convert through Clang's native _Float16 in
+kernel code. RV64GC has no half-precision hardware, so Clang lowers both
+conversions to the compiler-rt libcalls __extendhfsf2/__truncsfhf2, and
+wp_load_obj() (warp/native/clang/clang.cpp) resolves a JIT-compiled module's
+externals from a curated CRT table that carries no compiler builtins. Every
+module holding an fp16 kernel therefore failed to materialize:
+
+  JIT session error: Symbols not found: [ __extendhfsf2, __truncsfhf2 ]
+  Failed to lookup symbol: Failed to materialize symbols: ...
+  RuntimeError: Failed to find forward kernel '...' for device 'cpu'
+
+and took the module's non-fp16 kernels down with it -- 649 of the 1552 errors
+in a full riscv64 run of warp.tests, spread over test_codegen, test_print,
+test_spatial, test_transform, test_fabricarray and 40 more modules.
+
+x86_64 sidesteps this by adding +f16c to the cc1 invocation, which lowers both
+conversions to vcvtph2ps/vcvtps2ph; aarch64 gets fcvt from armv8-a. RISC-V has
+no such baseline extension -- Zfh is optional, is not part of rv64gc, and
+asking for it would emit fcvt.h.s on hardware that need not implement it -- so
+convert in software instead.
+
+The two routines are bit-for-bit what _Float16 yields where the hardware does
+have half support: round to nearest with ties to even, signalling NaNs quieted,
+NaN payloads preserved. Verified exhaustively against (_Float16) on x86_64
+(+f16c) over all 2^32 float bit patterns and all 2^16 half bit patterns, zero
+differences, so a riscv64 kernel rounds exactly as an x86_64 or aarch64 one
+does rather than picking up the ties-away rounding of the Giesen routine
+warp.cpp uses for the host-side conversions.
+
+Guarded on !__riscv_zfh so a Zfh-enabled target keeps the native path.
+---
+ warp/native/builtin.h | 105 ++++++++++++++++++++++++++++++++++++++++++
+ 1 file changed, 105 insertions(+)
+
+diff --git a/warp/native/builtin.h b/warp/native/builtin.h
+index f0c23b5..b1ad4b7 100644
+--- a/warp/native/builtin.h
++++ b/warp/native/builtin.h
+@@ -367,6 +367,109 @@ CUDA_CALLABLE inline float bfloat16_to_float(wp_bfloat16 x) { return wp_bfloat16
+ 
+ #elif defined(__clang__)
+ 
++#if defined(__riscv) && !defined(__riscv_zfh)
++
++// RISC-V without Zfh has no half-precision hardware, so Clang lowers every _Float16 conversion
++// to the compiler-rt libcalls __extendhfsf2/__truncsfhf2. A JIT-compiled module resolves its
++// externals from the curated CRT table in wp_load_obj() (warp/native/clang/clang.cpp), which
++// carries no compiler builtins, so a module holding any fp16 kernel fails to materialize with
++// "JIT session error: Symbols not found: [ __extendhfsf2, __truncsfhf2 ]".
++//
++// Convert in software instead. Both routines are bit-for-bit what _Float16 yields where the
++// hardware does have half support -- round to nearest with ties to even, signalling NaNs
++// quieted, payloads preserved -- verified exhaustively over all 2^32 floats and all 2^16
++// halves, so kernels round exactly as they do on x86_64 (+f16c) and aarch64.
++CUDA_CALLABLE inline half float_to_half(float x)
++{
++    unsigned int bits;
++    memcpy(&bits, &x, sizeof(bits));
++
++    const unsigned int sign = (bits >> 16) & 0x8000u;
++    const unsigned int magnitude = bits & 0x7fffffffu;
++
++    unsigned int u;
++    if (magnitude >= 0x7f800000u)  // Inf or NaN
++    {
++        const unsigned int mantissa = magnitude & 0x007fffffu;
++        u = sign | 0x7c00u | (mantissa ? ((mantissa >> 13) | 0x0200u) : 0u);  // NaN quieted, Inf kept
++    }
++    else if (magnitude >= 0x47800000u)  // 65536 and above overflows the half range
++    {
++        u = sign | 0x7c00u;
++    }
++    else if (magnitude >= 0x38800000u)  // 2^-14 and above is a normal half
++    {
++        u = sign | (((magnitude >> 23) - 127 + 15) << 10) | ((magnitude >> 13) & 0x03ffu);
++        const unsigned int rest = magnitude & 0x1fffu;  // the bits that do not fit
++        if (rest > 0x1000u || (rest == 0x1000u && (u & 1u)))
++            u += 1;  // round to nearest, ties to even
++    }
++    else
++    {
++        const unsigned int exponent = magnitude >> 23;
++        const unsigned int shift = 126 - exponent;  // a subnormal half's ulp is 2^-24
++        if (exponent == 0 || shift > 24)
++        {
++            u = sign;  // underflows to zero
++        }
++        else
++        {
++            const unsigned int mantissa = (magnitude & 0x007fffffu) | 0x00800000u;  // implicit bit
++            const unsigned int rest = mantissa & ((1u << shift) - 1u);
++            const unsigned int tie = 1u << (shift - 1);
++            u = sign | (mantissa >> shift);
++            if (rest > tie || (rest == tie && (u & 1u)))
++                u += 1;  // round to nearest, ties to even
++        }
++    }
++
++    half h;
++    h.u = static_cast<unsigned short>(u);
++    return h;
++}
++
++CUDA_CALLABLE inline float half_to_float(half h)
++{
++    const unsigned int sign = (static_cast<unsigned int>(h.u) & 0x8000u) << 16;
++    const unsigned int exponent = (static_cast<unsigned int>(h.u) >> 10) & 0x001fu;
++    const unsigned int mantissa = static_cast<unsigned int>(h.u) & 0x03ffu;
++
++    unsigned int bits;
++    if (exponent == 0x1fu)  // Inf or NaN
++    {
++        bits = sign | 0x7f800000u | (mantissa ? ((mantissa << 13) | 0x00400000u) : 0u);
++    }
++    else if (exponent == 0u)  // zero or subnormal
++    {
++        if (mantissa == 0u)
++        {
++            bits = sign;
++        }
++        else
++        {
++            // Renormalize: shift the mantissa up until its leading one leaves the field.
++            unsigned int m = mantissa;
++            unsigned int e = 127 - 15 + 1;
++            while ((m & 0x0400u) == 0u)
++            {
++                m <<= 1;
++                e -= 1;
++            }
++            bits = sign | (e << 23) | ((m & 0x03ffu) << 13);
++        }
++    }
++    else
++    {
++        bits = sign | ((exponent + 127 - 15) << 23) | (mantissa << 13);
++    }
++
++    float val;
++    memcpy(&val, &bits, sizeof(val));
++    return val;
++}
++
++#else
++
+ // _Float16 is Clang's native half-precision floating-point type
+ CUDA_CALLABLE inline half float_to_half(float x)
+ {
+@@ -381,6 +484,8 @@ CUDA_CALLABLE inline float half_to_float(half h)
+     return static_cast<float>(f16);
+ }
+ 
++#endif  // __riscv && !__riscv_zfh
++
+ #ifndef WP_NO_BFLOAT16
+ CUDA_CALLABLE inline wp_bfloat16 float_to_bfloat16(float x)
+ {
+-- 
+2.43.0
+