add v_rcp_f32_e64 to remu (#10393)

* tests from the box

* add v_rcp_f32_e64 to remu

* f32::from_bits utils

* v_cndmask_b32 tests

extra/remu/src/helpers.rs

@@ -55,10 +55,7 @@ where
 {
     fn negate(&self, pos: usize, modifier: usize) -> T {
         match (modifier >> pos) & 1 {
-            1 => match self.is_zero() {
+            1 => -*self,
-                true => T::zero(),
-                false => -*self,
-            },
             _ => *self,
         }
     }
@@ -122,7 +119,7 @@ mod tests {
         assert_eq!(0.3_f32.negate(2, 0b100), -0.3_f32);
         assert_eq!(0.3_f32.negate(0, 0b110), 0.3_f32);
         assert_eq!(0.3_f32.negate(1, 0b010), -0.3_f32);
-        assert_eq!(0.0_f32.negate(0, 0b001).to_bits(), 0);
+        assert_eq!(0.0_f32.negate(0, 0b001).to_bits(), (-0.0f32).to_bits());
         assert_eq!((-0.0_f32).negate(0, 0b001).to_bits(), 0);
     }
 

extra/remu/src/thread.rs

@@ -1246,7 +1246,7 @@ impl<'a> Thread<'a> {
                             }
 
                             let ret = match op {
-                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 531 | 537 | 540 | 551 | 567 | 796 => {
+                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 531 | 537 | 540 | 551 | 567 | 796 => {
                                     let s0 = f32::from_bits(s0).negate(0, neg).absolute(0, abs);
                                     let s1 = f32::from_bits(s1).negate(1, neg).absolute(1, abs);
                                     let s2 = f32::from_bits(s2).negate(2, neg).absolute(2, abs);
@@ -1257,6 +1257,7 @@ impl<'a> Thread<'a> {
                                         264 => s0 * s1,
                                         272 => f32::max(s0, s1),
                                         299 => f32::mul_add(s0, s1, f32::from_bits(self.vec_reg[vdst])),
+                                        426 => s0.recip(),
                                         531 => f32::mul_add(s0, s1, s2),
                                         537 => f32::min(f32::min(s0, s1), s2),
                                         540 => f32::max(f32::max(s0, s1), s2),
@@ -3032,7 +3033,7 @@ mod test_vop3 {
 
     #[test]
     fn test_v_cndmask_b32_e64_neg() {
-        [[0.0f32, 0.0], [-0.0f32, 0.0], [1.0f32, -1.0], [-1.0f32, 1.0]].iter().for_each(|[input, ret]| {
+        [[0.0f32, -0.0], [-0.0f32, 0.0], [1.0f32, -1.0], [-1.0f32, 1.0]].iter().for_each(|[input, ret]| {
             let mut thread = _helper_test_thread();
             thread.scalar_reg[0] = false as u32;
             thread.vec_reg[3] = input.to_bits();

extra/remu/test/hwtest.py

@@ -1,6 +1,6 @@
 import numpy as np
 import unittest
-import subprocess, struct
+import subprocess, struct, math
 from typing import cast
 from tinygrad.runtime.ops_amd import AMDProgram, AMDDevice
 from tinygrad import Tensor, dtypes, Device
@@ -95,6 +95,8 @@ def get_output(s:str, n_threads:int=1):
   return test.numpy()
 
 def f16_to_bits(x:float) -> int: return struct.unpack('<H', struct.pack('<e', x))[0]
+def f32_from_bits(x:int) -> float: return struct.unpack('<f', struct.pack('<I', x))[0]
+def f32_to_bits(x:float) -> int: return struct.unpack('<I', struct.pack('<f', x))[0]
 
 @unittest.skipUnless(Device.DEFAULT == "AMD", "tests RDNA3")
 class TestHW(unittest.TestCase):
@@ -168,5 +170,33 @@ class TestHW(unittest.TestCase):
     s_abs_i32(0xffffffff, 0x00000001, scc=1)
     s_abs_i32(0, 0, scc=0)
 
+  def test_v_rcp_f32_neg_vop3(self):
+    def v_neg_rcp_f32(x:float, y:float):
+      out = get_output(f"""
+      v_mov_b32_e32 v1 {f32_to_bits(x)}
+      v_rcp_f32_e64 v1, -v1
+      """)[0]
+      assert out == f32_to_bits(y), f"{f32_from_bits(out)} != {y} / {out} != {f32_to_bits(y)}"
+    v_neg_rcp_f32(math.inf, -0.0)
+    v_neg_rcp_f32(-math.inf, 0.0)
+    v_neg_rcp_f32(0.0, -math.inf)
+    v_neg_rcp_f32(-0.0, math.inf)
+    v_neg_rcp_f32(-2.0, 0.5)
+    v_neg_rcp_f32(2.0, -0.5)
+
+  def test_v_cndmask_b32_neg(self):
+    def v_neg(x:int|float, y:float):
+      out = get_output(f"""
+      v_mov_b32_e32 v1 {f32_to_bits(x)}
+      s_mov_b32_e32 s10 1 // always pick -v1
+      v_cndmask_b32 v1, v1, -v1 s10
+      """)[0]
+      assert out == f32_to_bits(y), f"{f32_from_bits(out)} != {y} / {out} != {f32_to_bits(y)}"
+    v_neg(-0.0, 0.0)
+    v_neg(0.0, -0.0)
+    v_neg(2.0, -2.0)
+    v_neg(math.inf, -math.inf)
+    v_neg(-math.inf, math.inf)
+
 if __name__ == "__main__":
   unittest.main()