* shader: emit float min/max as host instructions on AMD GPUs
V_MIN_F32, V_MAX_F32, V_MIN3_F32 and V_MAX3_F32 were emulated with a
classify-and-select sequence of about a dozen instructions each, to give
the guest's NaN and signed-zero results on any host. AMD GPUs execute
GLSL.std.450 NMin/NMax with those same rules, so emit them directly when
the device is AMD and supports VK_KHR_shader_float_controls2. Each
instruction is decorated FPFastMathMode None, since
SignedZeroInfNanPreserve does not cover extended instructions. MED3 keeps
its exact emulation. KYTY_EMULATED_FLOAT_MIN_MAX=1 keeps the emulation.
In SAROS the two heaviest compute shaders get 37% and 59% faster on an
RX 9070 XT (RADV), about 4 ms of GPU time per frame.
Adds VectorMinMaxF32RuntimeEdges (runtime operands, so nothing is folded)
and a --float-min-max-only mode that runs it and
VectorMinMaxF32NanAndSignedZeroEdges with the emulation and, where
supported, with host instructions.
Developed with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* shader: gate host float min/max on RADV and check the emitted instructions
* shader: drop float controls 2 from host float min/max
* shader: drop the float min/max env override and atomic flag
* shader: pass host capabilities to the SPIR-V emitter as a profile built once per device
* shader: enable host float min/max on AMD GPUs
---------
Co-authored-by: arda <ardaeska@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: nmzik <Nmzik@mail.ru>
A float inline constant (240-248) read by a 64-bit operand is the f64 bit
pattern for integer B64 opcodes too: AMD's compilers emit
v_lshrrev_b64 v[4:5], v0, 1.0 for 0x3ff0000000000000 >> lane. ReadU32Pair
returned {f32 bits, 0}, so only F64-typed reads had the right value.
Based on TheCruZ's 284deab28 (closed#1219).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
V_PK_MIN_F16 and V_PK_MAX_F16 return the quieted NaN for a signaling NaN
operand and the other operand for a quiet one (per half, as V_MIN_F16 and
V_MAX_F16 do). The signaling test masked the half with 0x7c00 and checked
for a nonzero payload, so a quiet NaN with payload bits such as 0x7e01 was
taken as signaling and returned. Mask with 0x7e00 so the quiet bit must be
clear.
Two more vectors in PackedMinMaxF16NanAndSignedZeroEdges cover it.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The 16-bit encodings of the float inline constants are fixed: the exact
halves of 0.5, 1.0, 2.0 and 4.0 and their negations, and 0x3118 for
1/(2*pi). Read16LaneBits converted the 32-bit value at run time instead,
through PackHalf2x16, which the AMD driver rounds toward zero, so
1/(2*pi) became 0x3117. Use the encodings directly.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* shader: support DS_CMPST_RTN_B32
Assassin's Creed Valhalla (PPSA01490, #281) stops on DS opcode 0x30 in a
compute shader. DS compares DATA and stores DATA2, the opposite of
BUFFER_ATOMIC_CMPSWAP, so the 32-bit atomic path swaps the two for this
opcode, as the 64-bit form does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* shader: reject unsupported GDS compare-store variant
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: nmzik <Nmzik@mail.ru>
nmzik