886 lines
50 KiB
Plaintext
886 lines
50 KiB
Plaintext
1. store_src_rg coords = src.rg
|
|
2. init_lane_masks CondMask = LoopMask = RetMask = true
|
|
3. copy_constant _0_ok = 0xFFFFFFFF
|
|
4. copy_constant _1_splat_4(0) = 0x40800000 (4.0)
|
|
5. copy_constant _1_splat_4(1) = 0x40800000 (4.0)
|
|
6. copy_constant _1_splat_4(2) = 0x40800000 (4.0)
|
|
7. copy_constant _1_splat_4(3) = 0x40800000 (4.0)
|
|
8. copy_constant _1_splat_4(4) = 0x40800000 (4.0)
|
|
9. copy_constant _1_splat_4(5) = 0x40800000 (4.0)
|
|
10. copy_constant _1_splat_4(6) = 0x40800000 (4.0)
|
|
11. copy_constant _1_splat_4(7) = 0x40800000 (4.0)
|
|
12. copy_constant _1_splat_4(8) = 0x40800000 (4.0)
|
|
13. copy_constant _2_splat_2(0) = 0x40000000 (2.0)
|
|
14. copy_constant _2_splat_2(1) = 0x40000000 (2.0)
|
|
15. copy_constant _2_splat_2(2) = 0x40000000 (2.0)
|
|
16. copy_constant _2_splat_2(3) = 0x40000000 (2.0)
|
|
17. copy_constant _2_splat_2(4) = 0x40000000 (2.0)
|
|
18. copy_constant _2_splat_2(5) = 0x40000000 (2.0)
|
|
19. copy_constant _2_splat_2(6) = 0x40000000 (2.0)
|
|
20. copy_constant _2_splat_2(7) = 0x40000000 (2.0)
|
|
21. copy_constant _2_splat_2(8) = 0x40000000 (2.0)
|
|
22. zero_4_slots_unmasked _3_m(0..3) = 0
|
|
23. zero_4_slots_unmasked _3_m(4..7) = 0
|
|
24. zero_slot_unmasked _3_m(8) = 0
|
|
25. zero_slot_unmasked $0 = 0
|
|
26. copy_constant $1 = 0x40000000 (2.0)
|
|
27. shuffle $0..8 = ($0..8)[1 0 0 0 1 0 0 0 1]
|
|
28. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
29. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
30. copy_slot_unmasked _3_m(8) = $8
|
|
31. copy_4_slots_unmasked $9..12 = _1_splat_4(0..3)
|
|
32. copy_4_slots_unmasked $13..16 = _1_splat_4(4..7)
|
|
33. copy_slot_unmasked $17 = _1_splat_4(8)
|
|
34. add_n_floats $0..8 += $9..17
|
|
35. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
36. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
37. copy_slot_unmasked _3_m(8) = $8
|
|
38. copy_slot_unmasked $0 = _0_ok
|
|
39. copy_4_slots_unmasked $1..4 = _3_m(0..3)
|
|
40. copy_4_slots_unmasked $5..8 = _3_m(4..7)
|
|
41. copy_slot_unmasked $9 = _3_m(8)
|
|
42. copy_constant $10 = 0x40C00000 (6.0)
|
|
43. copy_constant $11 = 0x40800000 (4.0)
|
|
44. copy_constant $12 = 0x40800000 (4.0)
|
|
45. copy_constant $13 = 0x40800000 (4.0)
|
|
46. copy_constant $14 = 0x40C00000 (6.0)
|
|
47. copy_constant $15 = 0x40800000 (4.0)
|
|
48. copy_constant $16 = 0x40800000 (4.0)
|
|
49. copy_constant $17 = 0x40800000 (4.0)
|
|
50. copy_constant $18 = 0x40C00000 (6.0)
|
|
51. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
52. bitwise_and_4_ints $2..5 &= $6..9
|
|
53. bitwise_and_2_ints $2..3 &= $4..5
|
|
54. bitwise_and_int $2 &= $3
|
|
55. bitwise_and_int $1 &= $2
|
|
56. bitwise_and_int $0 &= $1
|
|
57. copy_slot_unmasked _0_ok = $0
|
|
58. zero_slot_unmasked $0 = 0
|
|
59. copy_constant $1 = 0x40000000 (2.0)
|
|
60. shuffle $0..8 = ($0..8)[1 0 0 0 1 0 0 0 1]
|
|
61. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
62. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
63. copy_slot_unmasked _3_m(8) = $8
|
|
64. copy_4_slots_unmasked $9..12 = _1_splat_4(0..3)
|
|
65. copy_4_slots_unmasked $13..16 = _1_splat_4(4..7)
|
|
66. copy_slot_unmasked $17 = _1_splat_4(8)
|
|
67. sub_n_floats $0..8 -= $9..17
|
|
68. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
69. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
70. copy_slot_unmasked _3_m(8) = $8
|
|
71. copy_slot_unmasked $0 = _0_ok
|
|
72. copy_4_slots_unmasked $1..4 = _3_m(0..3)
|
|
73. copy_4_slots_unmasked $5..8 = _3_m(4..7)
|
|
74. copy_slot_unmasked $9 = _3_m(8)
|
|
75. copy_constant $10 = 0xC0000000 (-2.0)
|
|
76. copy_constant $11 = 0xC0800000 (-4.0)
|
|
77. copy_constant $12 = 0xC0800000 (-4.0)
|
|
78. copy_constant $13 = 0xC0800000 (-4.0)
|
|
79. copy_constant $14 = 0xC0000000 (-2.0)
|
|
80. copy_constant $15 = 0xC0800000 (-4.0)
|
|
81. copy_constant $16 = 0xC0800000 (-4.0)
|
|
82. copy_constant $17 = 0xC0800000 (-4.0)
|
|
83. copy_constant $18 = 0xC0000000 (-2.0)
|
|
84. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
85. bitwise_and_4_ints $2..5 &= $6..9
|
|
86. bitwise_and_2_ints $2..3 &= $4..5
|
|
87. bitwise_and_int $2 &= $3
|
|
88. bitwise_and_int $1 &= $2
|
|
89. bitwise_and_int $0 &= $1
|
|
90. copy_slot_unmasked _0_ok = $0
|
|
91. zero_slot_unmasked $0 = 0
|
|
92. copy_constant $1 = 0x40000000 (2.0)
|
|
93. shuffle $0..8 = ($0..8)[1 0 0 0 1 0 0 0 1]
|
|
94. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
95. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
96. copy_slot_unmasked _3_m(8) = $8
|
|
97. copy_4_slots_unmasked $9..12 = _1_splat_4(0..3)
|
|
98. copy_4_slots_unmasked $13..16 = _1_splat_4(4..7)
|
|
99. copy_slot_unmasked $17 = _1_splat_4(8)
|
|
100. div_n_floats $0..8 /= $9..17
|
|
101. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
102. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
103. copy_slot_unmasked _3_m(8) = $8
|
|
104. copy_slot_unmasked $0 = _0_ok
|
|
105. copy_4_slots_unmasked $1..4 = _3_m(0..3)
|
|
106. copy_4_slots_unmasked $5..8 = _3_m(4..7)
|
|
107. copy_slot_unmasked $9 = _3_m(8)
|
|
108. zero_slot_unmasked $10 = 0
|
|
109. copy_constant $11 = 0x3F000000 (0.5)
|
|
110. shuffle $10..18 = ($10..18)[1 0 0 0 1 0 0 0 1]
|
|
111. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
112. bitwise_and_4_ints $2..5 &= $6..9
|
|
113. bitwise_and_2_ints $2..3 &= $4..5
|
|
114. bitwise_and_int $2 &= $3
|
|
115. bitwise_and_int $1 &= $2
|
|
116. bitwise_and_int $0 &= $1
|
|
117. copy_slot_unmasked _0_ok = $0
|
|
118. copy_4_slots_unmasked $0..3 = _1_splat_4(0..3)
|
|
119. copy_4_slots_unmasked $4..7 = _1_splat_4(4..7)
|
|
120. copy_slot_unmasked $8 = _1_splat_4(8)
|
|
121. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
122. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
123. copy_slot_unmasked _3_m(8) = $8
|
|
124. zero_slot_unmasked $9 = 0
|
|
125. copy_constant $10 = 0x40000000 (2.0)
|
|
126. shuffle $9..17 = ($9..17)[1 0 0 0 1 0 0 0 1]
|
|
127. add_n_floats $0..8 += $9..17
|
|
128. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
129. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
130. copy_slot_unmasked _3_m(8) = $8
|
|
131. copy_slot_unmasked $0 = _0_ok
|
|
132. copy_4_slots_unmasked $1..4 = _3_m(0..3)
|
|
133. copy_4_slots_unmasked $5..8 = _3_m(4..7)
|
|
134. copy_slot_unmasked $9 = _3_m(8)
|
|
135. copy_constant $10 = 0x40C00000 (6.0)
|
|
136. copy_constant $11 = 0x40800000 (4.0)
|
|
137. copy_constant $12 = 0x40800000 (4.0)
|
|
138. copy_constant $13 = 0x40800000 (4.0)
|
|
139. copy_constant $14 = 0x40C00000 (6.0)
|
|
140. copy_constant $15 = 0x40800000 (4.0)
|
|
141. copy_constant $16 = 0x40800000 (4.0)
|
|
142. copy_constant $17 = 0x40800000 (4.0)
|
|
143. copy_constant $18 = 0x40C00000 (6.0)
|
|
144. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
145. bitwise_and_4_ints $2..5 &= $6..9
|
|
146. bitwise_and_2_ints $2..3 &= $4..5
|
|
147. bitwise_and_int $2 &= $3
|
|
148. bitwise_and_int $1 &= $2
|
|
149. bitwise_and_int $0 &= $1
|
|
150. copy_slot_unmasked _0_ok = $0
|
|
151. copy_4_slots_unmasked $0..3 = _1_splat_4(0..3)
|
|
152. copy_4_slots_unmasked $4..7 = _1_splat_4(4..7)
|
|
153. copy_slot_unmasked $8 = _1_splat_4(8)
|
|
154. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
155. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
156. copy_slot_unmasked _3_m(8) = $8
|
|
157. zero_slot_unmasked $9 = 0
|
|
158. copy_constant $10 = 0x40000000 (2.0)
|
|
159. shuffle $9..17 = ($9..17)[1 0 0 0 1 0 0 0 1]
|
|
160. sub_n_floats $0..8 -= $9..17
|
|
161. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
162. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
163. copy_slot_unmasked _3_m(8) = $8
|
|
164. copy_slot_unmasked $0 = _0_ok
|
|
165. copy_4_slots_unmasked $1..4 = _3_m(0..3)
|
|
166. copy_4_slots_unmasked $5..8 = _3_m(4..7)
|
|
167. copy_slot_unmasked $9 = _3_m(8)
|
|
168. copy_constant $10 = 0x40000000 (2.0)
|
|
169. copy_constant $11 = 0x40800000 (4.0)
|
|
170. copy_constant $12 = 0x40800000 (4.0)
|
|
171. copy_constant $13 = 0x40800000 (4.0)
|
|
172. copy_constant $14 = 0x40000000 (2.0)
|
|
173. copy_constant $15 = 0x40800000 (4.0)
|
|
174. copy_constant $16 = 0x40800000 (4.0)
|
|
175. copy_constant $17 = 0x40800000 (4.0)
|
|
176. copy_constant $18 = 0x40000000 (2.0)
|
|
177. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
178. bitwise_and_4_ints $2..5 &= $6..9
|
|
179. bitwise_and_2_ints $2..3 &= $4..5
|
|
180. bitwise_and_int $2 &= $3
|
|
181. bitwise_and_int $1 &= $2
|
|
182. bitwise_and_int $0 &= $1
|
|
183. copy_slot_unmasked _0_ok = $0
|
|
184. copy_4_slots_unmasked $0..3 = _1_splat_4(0..3)
|
|
185. copy_4_slots_unmasked $4..7 = _1_splat_4(4..7)
|
|
186. copy_slot_unmasked $8 = _1_splat_4(8)
|
|
187. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
188. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
189. copy_slot_unmasked _3_m(8) = $8
|
|
190. copy_4_slots_unmasked $9..12 = _2_splat_2(0..3)
|
|
191. copy_4_slots_unmasked $13..16 = _2_splat_2(4..7)
|
|
192. copy_slot_unmasked $17 = _2_splat_2(8)
|
|
193. div_n_floats $0..8 /= $9..17
|
|
194. copy_4_slots_unmasked _3_m(0..3) = $0..3
|
|
195. copy_4_slots_unmasked _3_m(4..7) = $4..7
|
|
196. copy_slot_unmasked _3_m(8) = $8
|
|
197. copy_slot_unmasked $0 = _0_ok
|
|
198. copy_4_slots_unmasked $1..4 = _3_m(0..3)
|
|
199. copy_4_slots_unmasked $5..8 = _3_m(4..7)
|
|
200. copy_slot_unmasked $9 = _3_m(8)
|
|
201. copy_constant $10 = 0x40000000 (2.0)
|
|
202. copy_constant $11 = 0x40000000 (2.0)
|
|
203. copy_constant $12 = 0x40000000 (2.0)
|
|
204. copy_constant $13 = 0x40000000 (2.0)
|
|
205. copy_constant $14 = 0x40000000 (2.0)
|
|
206. copy_constant $15 = 0x40000000 (2.0)
|
|
207. copy_constant $16 = 0x40000000 (2.0)
|
|
208. copy_constant $17 = 0x40000000 (2.0)
|
|
209. copy_constant $18 = 0x40000000 (2.0)
|
|
210. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
211. bitwise_and_4_ints $2..5 &= $6..9
|
|
212. bitwise_and_2_ints $2..3 &= $4..5
|
|
213. bitwise_and_int $2 &= $3
|
|
214. bitwise_and_int $1 &= $2
|
|
215. bitwise_and_int $0 &= $1
|
|
216. copy_slot_unmasked _0_ok = $0
|
|
217. copy_constant _4_m(0) = 0x3F800000 (1.0)
|
|
218. copy_constant _4_m(1) = 0x40000000 (2.0)
|
|
219. copy_constant _4_m(2) = 0x40400000 (3.0)
|
|
220. copy_constant _4_m(3) = 0x40800000 (4.0)
|
|
221. copy_constant _4_m(4) = 0x40A00000 (5.0)
|
|
222. copy_constant _4_m(5) = 0x40C00000 (6.0)
|
|
223. copy_constant _4_m(6) = 0x40E00000 (7.0)
|
|
224. copy_constant _4_m(7) = 0x41000000 (8.0)
|
|
225. copy_constant _4_m(8) = 0x41100000 (9.0)
|
|
226. copy_constant _4_m(9) = 0x41200000 (10.0)
|
|
227. copy_constant _4_m(10) = 0x41300000 (11.0)
|
|
228. copy_constant _4_m(11) = 0x41400000 (12.0)
|
|
229. copy_constant _4_m(12) = 0x41500000 (13.0)
|
|
230. copy_constant _4_m(13) = 0x41600000 (14.0)
|
|
231. copy_constant _4_m(14) = 0x41700000 (15.0)
|
|
232. copy_constant _4_m(15) = 0x41800000 (16.0)
|
|
233. copy_4_slots_unmasked $0..3 = _4_m(0..3)
|
|
234. copy_4_slots_unmasked $4..7 = _4_m(4..7)
|
|
235. copy_4_slots_unmasked $8..11 = _4_m(8..11)
|
|
236. copy_4_slots_unmasked $12..15 = _4_m(12..15)
|
|
237. copy_constant $16 = 0x41800000 (16.0)
|
|
238. copy_constant $17 = 0x41700000 (15.0)
|
|
239. copy_constant $18 = 0x41600000 (14.0)
|
|
240. copy_constant $19 = 0x41500000 (13.0)
|
|
241. copy_constant $20 = 0x41400000 (12.0)
|
|
242. copy_constant $21 = 0x41300000 (11.0)
|
|
243. copy_constant $22 = 0x41200000 (10.0)
|
|
244. copy_constant $23 = 0x41100000 (9.0)
|
|
245. copy_constant $24 = 0x41000000 (8.0)
|
|
246. copy_constant $25 = 0x40E00000 (7.0)
|
|
247. copy_constant $26 = 0x40C00000 (6.0)
|
|
248. copy_constant $27 = 0x40A00000 (5.0)
|
|
249. copy_constant $28 = 0x40800000 (4.0)
|
|
250. copy_constant $29 = 0x40400000 (3.0)
|
|
251. copy_constant $30 = 0x40000000 (2.0)
|
|
252. copy_constant $31 = 0x3F800000 (1.0)
|
|
253. add_n_floats $0..15 += $16..31
|
|
254. copy_4_slots_unmasked _4_m(0..3) = $0..3
|
|
255. copy_4_slots_unmasked _4_m(4..7) = $4..7
|
|
256. copy_4_slots_unmasked _4_m(8..11) = $8..11
|
|
257. copy_4_slots_unmasked _4_m(12..15) = $12..15
|
|
258. copy_slot_unmasked $0 = _0_ok
|
|
259. copy_4_slots_unmasked $1..4 = _4_m(0..3)
|
|
260. copy_4_slots_unmasked $5..8 = _4_m(4..7)
|
|
261. copy_4_slots_unmasked $9..12 = _4_m(8..11)
|
|
262. copy_4_slots_unmasked $13..16 = _4_m(12..15)
|
|
263. copy_constant $17 = 0x41880000 (17.0)
|
|
264. copy_constant $18 = 0x41880000 (17.0)
|
|
265. copy_constant $19 = 0x41880000 (17.0)
|
|
266. copy_constant $20 = 0x41880000 (17.0)
|
|
267. copy_constant $21 = 0x41880000 (17.0)
|
|
268. copy_constant $22 = 0x41880000 (17.0)
|
|
269. copy_constant $23 = 0x41880000 (17.0)
|
|
270. copy_constant $24 = 0x41880000 (17.0)
|
|
271. copy_constant $25 = 0x41880000 (17.0)
|
|
272. copy_constant $26 = 0x41880000 (17.0)
|
|
273. copy_constant $27 = 0x41880000 (17.0)
|
|
274. copy_constant $28 = 0x41880000 (17.0)
|
|
275. copy_constant $29 = 0x41880000 (17.0)
|
|
276. copy_constant $30 = 0x41880000 (17.0)
|
|
277. copy_constant $31 = 0x41880000 (17.0)
|
|
278. copy_constant $32 = 0x41880000 (17.0)
|
|
279. cmpeq_n_floats $1..16 = equal($1..16, $17..32)
|
|
280. bitwise_and_4_ints $9..12 &= $13..16
|
|
281. bitwise_and_4_ints $5..8 &= $9..12
|
|
282. bitwise_and_4_ints $1..4 &= $5..8
|
|
283. bitwise_and_2_ints $1..2 &= $3..4
|
|
284. bitwise_and_int $1 &= $2
|
|
285. bitwise_and_int $0 &= $1
|
|
286. copy_slot_unmasked _0_ok = $0
|
|
287. copy_constant _5_m(0) = 0x41200000 (10.0)
|
|
288. copy_constant _5_m(1) = 0x41A00000 (20.0)
|
|
289. copy_constant _5_m(2) = 0x41F00000 (30.0)
|
|
290. copy_constant _5_m(3) = 0x42200000 (40.0)
|
|
291. copy_4_slots_unmasked $0..3 = _5_m
|
|
292. copy_constant $4 = 0x3F800000 (1.0)
|
|
293. copy_constant $5 = 0x40000000 (2.0)
|
|
294. copy_constant $6 = 0x40400000 (3.0)
|
|
295. copy_constant $7 = 0x40800000 (4.0)
|
|
296. sub_4_floats $0..3 -= $4..7
|
|
297. copy_4_slots_unmasked _5_m = $0..3
|
|
298. copy_slot_unmasked $0 = _0_ok
|
|
299. copy_4_slots_unmasked $1..4 = _5_m
|
|
300. copy_constant $5 = 0x41100000 (9.0)
|
|
301. copy_constant $6 = 0x41900000 (18.0)
|
|
302. copy_constant $7 = 0x41D80000 (27.0)
|
|
303. copy_constant $8 = 0x42100000 (36.0)
|
|
304. cmpeq_4_floats $1..4 = equal($1..4, $5..8)
|
|
305. bitwise_and_2_ints $1..2 &= $3..4
|
|
306. bitwise_and_int $1 &= $2
|
|
307. bitwise_and_int $0 &= $1
|
|
308. copy_slot_unmasked _0_ok = $0
|
|
309. copy_constant _6_m(0) = 0x40000000 (2.0)
|
|
310. copy_constant _6_m(1) = 0x40800000 (4.0)
|
|
311. copy_constant _6_m(2) = 0x40C00000 (6.0)
|
|
312. copy_constant _6_m(3) = 0x41000000 (8.0)
|
|
313. copy_4_slots_unmasked $0..3 = _6_m
|
|
314. copy_constant $4 = 0x40000000 (2.0)
|
|
315. copy_constant $5 = 0x40000000 (2.0)
|
|
316. copy_constant $6 = 0x40000000 (2.0)
|
|
317. copy_constant $7 = 0x40800000 (4.0)
|
|
318. div_4_floats $0..3 /= $4..7
|
|
319. copy_4_slots_unmasked _6_m = $0..3
|
|
320. copy_slot_unmasked $0 = _0_ok
|
|
321. copy_4_slots_unmasked $1..4 = _6_m
|
|
322. copy_constant $5 = 0x3F800000 (1.0)
|
|
323. copy_constant $6 = 0x40000000 (2.0)
|
|
324. copy_constant $7 = 0x40400000 (3.0)
|
|
325. copy_constant $8 = 0x40000000 (2.0)
|
|
326. cmpeq_4_floats $1..4 = equal($1..4, $5..8)
|
|
327. bitwise_and_2_ints $1..2 &= $3..4
|
|
328. bitwise_and_int $1 &= $2
|
|
329. bitwise_and_int $0 &= $1
|
|
330. copy_slot_unmasked _0_ok = $0
|
|
331. copy_constant _7_m(0) = 0x3F800000 (1.0)
|
|
332. copy_constant _7_m(1) = 0x40000000 (2.0)
|
|
333. copy_constant _7_m(2) = 0x40E00000 (7.0)
|
|
334. copy_constant _7_m(3) = 0x40800000 (4.0)
|
|
335. copy_4_slots_unmasked $52..55 = _7_m
|
|
336. swizzle_3 $53..55 = ($53..55).yxz
|
|
337. copy_constant $56 = 0x40400000 (3.0)
|
|
338. copy_constant $57 = 0x40A00000 (5.0)
|
|
339. copy_constant $58 = 0x40400000 (3.0)
|
|
340. copy_constant $59 = 0x40000000 (2.0)
|
|
341. copy_2_slots_unmasked $0..1 = $52..53
|
|
342. copy_2_slots_unmasked $2..3 = $56..57
|
|
343. dot_2_floats $0 = dot($0..1, $2..3)
|
|
344. copy_4_slots_unmasked $1..4 = $54..57
|
|
345. dot_2_floats $1 = dot($1..2, $3..4)
|
|
346. copy_2_slots_unmasked $2..3 = $52..53
|
|
347. copy_2_slots_unmasked $4..5 = $58..59
|
|
348. dot_2_floats $2 = dot($2..3, $4..5)
|
|
349. copy_2_slots_unmasked $3..4 = $54..55
|
|
350. copy_2_slots_unmasked $5..6 = $58..59
|
|
351. dot_2_floats $3 = dot($3..4, $5..6)
|
|
352. copy_4_slots_unmasked _7_m = $0..3
|
|
353. copy_slot_unmasked $0 = _0_ok
|
|
354. copy_4_slots_unmasked $1..4 = _7_m
|
|
355. copy_constant $5 = 0x42180000 (38.0)
|
|
356. copy_constant $6 = 0x41D00000 (26.0)
|
|
357. copy_constant $7 = 0x41880000 (17.0)
|
|
358. copy_constant $8 = 0x41600000 (14.0)
|
|
359. cmpeq_4_floats $1..4 = equal($1..4, $5..8)
|
|
360. bitwise_and_2_ints $1..2 &= $3..4
|
|
361. bitwise_and_int $1 &= $2
|
|
362. bitwise_and_int $0 &= $1
|
|
363. copy_slot_unmasked _0_ok = $0
|
|
364. copy_constant _8_m(0) = 0x41200000 (10.0)
|
|
365. copy_constant _8_m(1) = 0x40800000 (4.0)
|
|
366. copy_constant _8_m(2) = 0x40000000 (2.0)
|
|
367. copy_constant _8_m(3) = 0x41A00000 (20.0)
|
|
368. copy_constant _8_m(4) = 0x40A00000 (5.0)
|
|
369. copy_constant _8_m(5) = 0x40400000 (3.0)
|
|
370. copy_constant _8_m(6) = 0x41200000 (10.0)
|
|
371. copy_constant _8_m(7) = 0x40C00000 (6.0)
|
|
372. copy_constant _8_m(8) = 0x40A00000 (5.0)
|
|
373. copy_4_slots_unmasked $52..55 = _8_m(0..3)
|
|
374. copy_4_slots_unmasked $56..59 = _8_m(4..7)
|
|
375. copy_slot_unmasked $60 = _8_m(8)
|
|
376. shuffle $53..60 = ($53..60)[2 5 0 3 6 1 4 7]
|
|
377. copy_constant $61 = 0x40400000 (3.0)
|
|
378. copy_constant $62 = 0x40400000 (3.0)
|
|
379. copy_constant $63 = 0x40800000 (4.0)
|
|
380. copy_constant $64 = 0x40000000 (2.0)
|
|
381. copy_constant $65 = 0x40400000 (3.0)
|
|
382. copy_constant $66 = 0x40800000 (4.0)
|
|
383. copy_constant $67 = 0x40800000 (4.0)
|
|
384. copy_constant $68 = 0x41100000 (9.0)
|
|
385. copy_constant $69 = 0x40000000 (2.0)
|
|
386. copy_3_slots_unmasked $0..2 = $52..54
|
|
387. copy_3_slots_unmasked $3..5 = $61..63
|
|
388. dot_3_floats $0 = dot($0..2, $3..5)
|
|
389. copy_3_slots_unmasked $1..3 = $55..57
|
|
390. copy_3_slots_unmasked $4..6 = $61..63
|
|
391. dot_3_floats $1 = dot($1..3, $4..6)
|
|
392. copy_4_slots_unmasked $2..5 = $58..61
|
|
393. copy_2_slots_unmasked $6..7 = $62..63
|
|
394. dot_3_floats $2 = dot($2..4, $5..7)
|
|
395. copy_3_slots_unmasked $3..5 = $52..54
|
|
396. copy_3_slots_unmasked $6..8 = $64..66
|
|
397. dot_3_floats $3 = dot($3..5, $6..8)
|
|
398. copy_3_slots_unmasked $4..6 = $55..57
|
|
399. copy_3_slots_unmasked $7..9 = $64..66
|
|
400. dot_3_floats $4 = dot($4..6, $7..9)
|
|
401. copy_3_slots_unmasked $5..7 = $58..60
|
|
402. copy_3_slots_unmasked $8..10 = $64..66
|
|
403. dot_3_floats $5 = dot($5..7, $8..10)
|
|
404. copy_3_slots_unmasked $6..8 = $52..54
|
|
405. copy_3_slots_unmasked $9..11 = $67..69
|
|
406. dot_3_floats $6 = dot($6..8, $9..11)
|
|
407. copy_3_slots_unmasked $7..9 = $55..57
|
|
408. copy_3_slots_unmasked $10..12 = $67..69
|
|
409. dot_3_floats $7 = dot($7..9, $10..12)
|
|
410. copy_3_slots_unmasked $8..10 = $58..60
|
|
411. copy_3_slots_unmasked $11..13 = $67..69
|
|
412. dot_3_floats $8 = dot($8..10, $11..13)
|
|
413. copy_4_slots_unmasked _8_m(0..3) = $0..3
|
|
414. copy_4_slots_unmasked _8_m(4..7) = $4..7
|
|
415. copy_slot_unmasked _8_m(8) = $8
|
|
416. copy_slot_unmasked $0 = _0_ok
|
|
417. copy_4_slots_unmasked $1..4 = _8_m(0..3)
|
|
418. copy_4_slots_unmasked $5..8 = _8_m(4..7)
|
|
419. copy_slot_unmasked $9 = _8_m(8)
|
|
420. copy_constant $10 = 0x43020000 (130.0)
|
|
421. copy_constant $11 = 0x424C0000 (51.0)
|
|
422. copy_constant $12 = 0x420C0000 (35.0)
|
|
423. copy_constant $13 = 0x42F00000 (120.0)
|
|
424. copy_constant $14 = 0x423C0000 (47.0)
|
|
425. copy_constant $15 = 0x42040000 (33.0)
|
|
426. copy_constant $16 = 0x43700000 (240.0)
|
|
427. copy_constant $17 = 0x42920000 (73.0)
|
|
428. copy_constant $18 = 0x42340000 (45.0)
|
|
429. cmpeq_n_floats $1..9 = equal($1..9, $10..18)
|
|
430. bitwise_and_4_ints $2..5 &= $6..9
|
|
431. bitwise_and_2_ints $2..3 &= $4..5
|
|
432. bitwise_and_int $2 &= $3
|
|
433. bitwise_and_int $1 &= $2
|
|
434. bitwise_and_int $0 &= $1
|
|
435. copy_slot_unmasked _0_ok = $0
|
|
436. store_condition_mask $52 = CondMask
|
|
437. copy_slot_unmasked $53 = _0_ok
|
|
438. zero_slot_unmasked $0 = 0
|
|
439. merge_condition_mask CondMask = $52 & $53
|
|
440. branch_if_no_active_lanes branch_if_no_active_lanes +438 (label 1 at #878)
|
|
441. copy_constant ok = 0xFFFFFFFF
|
|
442. copy_constant splat_4(0) = 0x40800000 (4.0)
|
|
443. copy_constant splat_4(1) = 0x40800000 (4.0)
|
|
444. copy_constant splat_4(2) = 0x40800000 (4.0)
|
|
445. copy_constant splat_4(3) = 0x40800000 (4.0)
|
|
446. copy_constant splat_4(4) = 0x40800000 (4.0)
|
|
447. copy_constant splat_4(5) = 0x40800000 (4.0)
|
|
448. copy_constant splat_4(6) = 0x40800000 (4.0)
|
|
449. copy_constant splat_4(7) = 0x40800000 (4.0)
|
|
450. copy_constant splat_4(8) = 0x40800000 (4.0)
|
|
451. copy_constant splat_2(0) = 0x40000000 (2.0)
|
|
452. copy_constant splat_2(1) = 0x40000000 (2.0)
|
|
453. copy_constant splat_2(2) = 0x40000000 (2.0)
|
|
454. copy_constant splat_2(3) = 0x40000000 (2.0)
|
|
455. copy_constant splat_2(4) = 0x40000000 (2.0)
|
|
456. copy_constant splat_2(5) = 0x40000000 (2.0)
|
|
457. copy_constant splat_2(6) = 0x40000000 (2.0)
|
|
458. copy_constant splat_2(7) = 0x40000000 (2.0)
|
|
459. copy_constant splat_2(8) = 0x40000000 (2.0)
|
|
460. zero_4_slots_unmasked m(0..3) = 0
|
|
461. zero_4_slots_unmasked m(4..7) = 0
|
|
462. zero_slot_unmasked m(8) = 0
|
|
463. zero_slot_unmasked $1 = 0
|
|
464. copy_constant $2 = 0x40000000 (2.0)
|
|
465. shuffle $1..9 = ($1..9)[1 0 0 0 1 0 0 0 1]
|
|
466. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
467. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
468. copy_slot_masked m(8) = Mask($9)
|
|
469. copy_4_slots_unmasked $10..13 = splat_4(0..3)
|
|
470. copy_4_slots_unmasked $14..17 = splat_4(4..7)
|
|
471. copy_slot_unmasked $18 = splat_4(8)
|
|
472. add_n_floats $1..9 += $10..18
|
|
473. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
474. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
475. copy_slot_masked m(8) = Mask($9)
|
|
476. copy_slot_unmasked $1 = ok
|
|
477. copy_4_slots_unmasked $2..5 = m(0..3)
|
|
478. copy_4_slots_unmasked $6..9 = m(4..7)
|
|
479. copy_slot_unmasked $10 = m(8)
|
|
480. copy_constant $11 = 0x40C00000 (6.0)
|
|
481. copy_constant $12 = 0x40800000 (4.0)
|
|
482. copy_constant $13 = 0x40800000 (4.0)
|
|
483. copy_constant $14 = 0x40800000 (4.0)
|
|
484. copy_constant $15 = 0x40C00000 (6.0)
|
|
485. copy_constant $16 = 0x40800000 (4.0)
|
|
486. copy_constant $17 = 0x40800000 (4.0)
|
|
487. copy_constant $18 = 0x40800000 (4.0)
|
|
488. copy_constant $19 = 0x40C00000 (6.0)
|
|
489. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
490. bitwise_and_4_ints $3..6 &= $7..10
|
|
491. bitwise_and_2_ints $3..4 &= $5..6
|
|
492. bitwise_and_int $3 &= $4
|
|
493. bitwise_and_int $2 &= $3
|
|
494. bitwise_and_int $1 &= $2
|
|
495. copy_slot_masked ok = Mask($1)
|
|
496. zero_slot_unmasked $1 = 0
|
|
497. copy_constant $2 = 0x40000000 (2.0)
|
|
498. shuffle $1..9 = ($1..9)[1 0 0 0 1 0 0 0 1]
|
|
499. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
500. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
501. copy_slot_masked m(8) = Mask($9)
|
|
502. stack_rewind
|
|
503. copy_4_slots_unmasked $10..13 = splat_4(0..3)
|
|
504. copy_4_slots_unmasked $14..17 = splat_4(4..7)
|
|
505. copy_slot_unmasked $18 = splat_4(8)
|
|
506. sub_n_floats $1..9 -= $10..18
|
|
507. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
508. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
509. copy_slot_masked m(8) = Mask($9)
|
|
510. copy_slot_unmasked $1 = ok
|
|
511. copy_4_slots_unmasked $2..5 = m(0..3)
|
|
512. copy_4_slots_unmasked $6..9 = m(4..7)
|
|
513. copy_slot_unmasked $10 = m(8)
|
|
514. copy_constant $11 = 0xC0000000 (-2.0)
|
|
515. copy_constant $12 = 0xC0800000 (-4.0)
|
|
516. copy_constant $13 = 0xC0800000 (-4.0)
|
|
517. copy_constant $14 = 0xC0800000 (-4.0)
|
|
518. copy_constant $15 = 0xC0000000 (-2.0)
|
|
519. copy_constant $16 = 0xC0800000 (-4.0)
|
|
520. copy_constant $17 = 0xC0800000 (-4.0)
|
|
521. copy_constant $18 = 0xC0800000 (-4.0)
|
|
522. copy_constant $19 = 0xC0000000 (-2.0)
|
|
523. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
524. bitwise_and_4_ints $3..6 &= $7..10
|
|
525. bitwise_and_2_ints $3..4 &= $5..6
|
|
526. bitwise_and_int $3 &= $4
|
|
527. bitwise_and_int $2 &= $3
|
|
528. bitwise_and_int $1 &= $2
|
|
529. copy_slot_masked ok = Mask($1)
|
|
530. zero_slot_unmasked $1 = 0
|
|
531. copy_constant $2 = 0x40000000 (2.0)
|
|
532. shuffle $1..9 = ($1..9)[1 0 0 0 1 0 0 0 1]
|
|
533. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
534. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
535. copy_slot_masked m(8) = Mask($9)
|
|
536. copy_4_slots_unmasked $10..13 = splat_4(0..3)
|
|
537. copy_4_slots_unmasked $14..17 = splat_4(4..7)
|
|
538. copy_slot_unmasked $18 = splat_4(8)
|
|
539. div_n_floats $1..9 /= $10..18
|
|
540. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
541. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
542. copy_slot_masked m(8) = Mask($9)
|
|
543. copy_slot_unmasked $1 = ok
|
|
544. copy_4_slots_unmasked $2..5 = m(0..3)
|
|
545. copy_4_slots_unmasked $6..9 = m(4..7)
|
|
546. copy_slot_unmasked $10 = m(8)
|
|
547. zero_slot_unmasked $11 = 0
|
|
548. copy_constant $12 = 0x3F000000 (0.5)
|
|
549. shuffle $11..19 = ($11..19)[1 0 0 0 1 0 0 0 1]
|
|
550. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
551. bitwise_and_4_ints $3..6 &= $7..10
|
|
552. bitwise_and_2_ints $3..4 &= $5..6
|
|
553. bitwise_and_int $3 &= $4
|
|
554. bitwise_and_int $2 &= $3
|
|
555. bitwise_and_int $1 &= $2
|
|
556. copy_slot_masked ok = Mask($1)
|
|
557. copy_4_slots_unmasked $1..4 = splat_4(0..3)
|
|
558. copy_4_slots_unmasked $5..8 = splat_4(4..7)
|
|
559. copy_slot_unmasked $9 = splat_4(8)
|
|
560. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
561. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
562. copy_slot_masked m(8) = Mask($9)
|
|
563. zero_slot_unmasked $10 = 0
|
|
564. copy_constant $11 = 0x40000000 (2.0)
|
|
565. shuffle $10..18 = ($10..18)[1 0 0 0 1 0 0 0 1]
|
|
566. add_n_floats $1..9 += $10..18
|
|
567. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
568. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
569. copy_slot_masked m(8) = Mask($9)
|
|
570. copy_slot_unmasked $1 = ok
|
|
571. copy_4_slots_unmasked $2..5 = m(0..3)
|
|
572. copy_4_slots_unmasked $6..9 = m(4..7)
|
|
573. copy_slot_unmasked $10 = m(8)
|
|
574. copy_constant $11 = 0x40C00000 (6.0)
|
|
575. copy_constant $12 = 0x40800000 (4.0)
|
|
576. copy_constant $13 = 0x40800000 (4.0)
|
|
577. copy_constant $14 = 0x40800000 (4.0)
|
|
578. copy_constant $15 = 0x40C00000 (6.0)
|
|
579. copy_constant $16 = 0x40800000 (4.0)
|
|
580. copy_constant $17 = 0x40800000 (4.0)
|
|
581. copy_constant $18 = 0x40800000 (4.0)
|
|
582. copy_constant $19 = 0x40C00000 (6.0)
|
|
583. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
584. bitwise_and_4_ints $3..6 &= $7..10
|
|
585. bitwise_and_2_ints $3..4 &= $5..6
|
|
586. bitwise_and_int $3 &= $4
|
|
587. bitwise_and_int $2 &= $3
|
|
588. bitwise_and_int $1 &= $2
|
|
589. copy_slot_masked ok = Mask($1)
|
|
590. copy_4_slots_unmasked $1..4 = splat_4(0..3)
|
|
591. copy_4_slots_unmasked $5..8 = splat_4(4..7)
|
|
592. copy_slot_unmasked $9 = splat_4(8)
|
|
593. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
594. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
595. copy_slot_masked m(8) = Mask($9)
|
|
596. zero_slot_unmasked $10 = 0
|
|
597. copy_constant $11 = 0x40000000 (2.0)
|
|
598. shuffle $10..18 = ($10..18)[1 0 0 0 1 0 0 0 1]
|
|
599. sub_n_floats $1..9 -= $10..18
|
|
600. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
601. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
602. copy_slot_masked m(8) = Mask($9)
|
|
603. copy_slot_unmasked $1 = ok
|
|
604. copy_4_slots_unmasked $2..5 = m(0..3)
|
|
605. copy_4_slots_unmasked $6..9 = m(4..7)
|
|
606. copy_slot_unmasked $10 = m(8)
|
|
607. copy_constant $11 = 0x40000000 (2.0)
|
|
608. copy_constant $12 = 0x40800000 (4.0)
|
|
609. copy_constant $13 = 0x40800000 (4.0)
|
|
610. copy_constant $14 = 0x40800000 (4.0)
|
|
611. copy_constant $15 = 0x40000000 (2.0)
|
|
612. copy_constant $16 = 0x40800000 (4.0)
|
|
613. copy_constant $17 = 0x40800000 (4.0)
|
|
614. copy_constant $18 = 0x40800000 (4.0)
|
|
615. copy_constant $19 = 0x40000000 (2.0)
|
|
616. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
617. bitwise_and_4_ints $3..6 &= $7..10
|
|
618. bitwise_and_2_ints $3..4 &= $5..6
|
|
619. bitwise_and_int $3 &= $4
|
|
620. bitwise_and_int $2 &= $3
|
|
621. bitwise_and_int $1 &= $2
|
|
622. copy_slot_masked ok = Mask($1)
|
|
623. copy_4_slots_unmasked $1..4 = splat_4(0..3)
|
|
624. copy_4_slots_unmasked $5..8 = splat_4(4..7)
|
|
625. copy_slot_unmasked $9 = splat_4(8)
|
|
626. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
627. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
628. copy_slot_masked m(8) = Mask($9)
|
|
629. copy_4_slots_unmasked $10..13 = splat_2(0..3)
|
|
630. copy_4_slots_unmasked $14..17 = splat_2(4..7)
|
|
631. copy_slot_unmasked $18 = splat_2(8)
|
|
632. div_n_floats $1..9 /= $10..18
|
|
633. copy_4_slots_masked m(0..3) = Mask($1..4)
|
|
634. copy_4_slots_masked m(4..7) = Mask($5..8)
|
|
635. copy_slot_masked m(8) = Mask($9)
|
|
636. copy_slot_unmasked $1 = ok
|
|
637. copy_4_slots_unmasked $2..5 = m(0..3)
|
|
638. copy_4_slots_unmasked $6..9 = m(4..7)
|
|
639. copy_slot_unmasked $10 = m(8)
|
|
640. copy_constant $11 = 0x40000000 (2.0)
|
|
641. copy_constant $12 = 0x40000000 (2.0)
|
|
642. copy_constant $13 = 0x40000000 (2.0)
|
|
643. copy_constant $14 = 0x40000000 (2.0)
|
|
644. copy_constant $15 = 0x40000000 (2.0)
|
|
645. copy_constant $16 = 0x40000000 (2.0)
|
|
646. copy_constant $17 = 0x40000000 (2.0)
|
|
647. copy_constant $18 = 0x40000000 (2.0)
|
|
648. copy_constant $19 = 0x40000000 (2.0)
|
|
649. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
650. bitwise_and_4_ints $3..6 &= $7..10
|
|
651. bitwise_and_2_ints $3..4 &= $5..6
|
|
652. bitwise_and_int $3 &= $4
|
|
653. bitwise_and_int $2 &= $3
|
|
654. bitwise_and_int $1 &= $2
|
|
655. copy_slot_masked ok = Mask($1)
|
|
656. copy_constant m₁(0) = 0x3F800000 (1.0)
|
|
657. copy_constant m₁(1) = 0x40000000 (2.0)
|
|
658. copy_constant m₁(2) = 0x40400000 (3.0)
|
|
659. copy_constant m₁(3) = 0x40800000 (4.0)
|
|
660. copy_constant m₁(4) = 0x40A00000 (5.0)
|
|
661. copy_constant m₁(5) = 0x40C00000 (6.0)
|
|
662. copy_constant m₁(6) = 0x40E00000 (7.0)
|
|
663. copy_constant m₁(7) = 0x41000000 (8.0)
|
|
664. copy_constant m₁(8) = 0x41100000 (9.0)
|
|
665. copy_constant m₁(9) = 0x41200000 (10.0)
|
|
666. copy_constant m₁(10) = 0x41300000 (11.0)
|
|
667. copy_constant m₁(11) = 0x41400000 (12.0)
|
|
668. copy_constant m₁(12) = 0x41500000 (13.0)
|
|
669. copy_constant m₁(13) = 0x41600000 (14.0)
|
|
670. copy_constant m₁(14) = 0x41700000 (15.0)
|
|
671. copy_constant m₁(15) = 0x41800000 (16.0)
|
|
672. copy_4_slots_unmasked $1..4 = m₁(0..3)
|
|
673. copy_4_slots_unmasked $5..8 = m₁(4..7)
|
|
674. copy_4_slots_unmasked $9..12 = m₁(8..11)
|
|
675. copy_4_slots_unmasked $13..16 = m₁(12..15)
|
|
676. copy_constant $17 = 0x41800000 (16.0)
|
|
677. copy_constant $18 = 0x41700000 (15.0)
|
|
678. copy_constant $19 = 0x41600000 (14.0)
|
|
679. copy_constant $20 = 0x41500000 (13.0)
|
|
680. copy_constant $21 = 0x41400000 (12.0)
|
|
681. copy_constant $22 = 0x41300000 (11.0)
|
|
682. copy_constant $23 = 0x41200000 (10.0)
|
|
683. copy_constant $24 = 0x41100000 (9.0)
|
|
684. copy_constant $25 = 0x41000000 (8.0)
|
|
685. copy_constant $26 = 0x40E00000 (7.0)
|
|
686. copy_constant $27 = 0x40C00000 (6.0)
|
|
687. copy_constant $28 = 0x40A00000 (5.0)
|
|
688. copy_constant $29 = 0x40800000 (4.0)
|
|
689. copy_constant $30 = 0x40400000 (3.0)
|
|
690. copy_constant $31 = 0x40000000 (2.0)
|
|
691. copy_constant $32 = 0x3F800000 (1.0)
|
|
692. add_n_floats $1..16 += $17..32
|
|
693. copy_4_slots_masked m₁(0..3) = Mask($1..4)
|
|
694. copy_4_slots_masked m₁(4..7) = Mask($5..8)
|
|
695. copy_4_slots_masked m₁(8..11) = Mask($9..12)
|
|
696. copy_4_slots_masked m₁(12..15) = Mask($13..16)
|
|
697. copy_slot_unmasked $1 = ok
|
|
698. copy_4_slots_unmasked $2..5 = m₁(0..3)
|
|
699. copy_4_slots_unmasked $6..9 = m₁(4..7)
|
|
700. copy_4_slots_unmasked $10..13 = m₁(8..11)
|
|
701. copy_4_slots_unmasked $14..17 = m₁(12..15)
|
|
702. copy_constant $18 = 0x41880000 (17.0)
|
|
703. copy_constant $19 = 0x41880000 (17.0)
|
|
704. copy_constant $20 = 0x41880000 (17.0)
|
|
705. copy_constant $21 = 0x41880000 (17.0)
|
|
706. copy_constant $22 = 0x41880000 (17.0)
|
|
707. copy_constant $23 = 0x41880000 (17.0)
|
|
708. copy_constant $24 = 0x41880000 (17.0)
|
|
709. copy_constant $25 = 0x41880000 (17.0)
|
|
710. copy_constant $26 = 0x41880000 (17.0)
|
|
711. copy_constant $27 = 0x41880000 (17.0)
|
|
712. copy_constant $28 = 0x41880000 (17.0)
|
|
713. copy_constant $29 = 0x41880000 (17.0)
|
|
714. copy_constant $30 = 0x41880000 (17.0)
|
|
715. copy_constant $31 = 0x41880000 (17.0)
|
|
716. copy_constant $32 = 0x41880000 (17.0)
|
|
717. copy_constant $33 = 0x41880000 (17.0)
|
|
718. cmpeq_n_floats $2..17 = equal($2..17, $18..33)
|
|
719. bitwise_and_4_ints $10..13 &= $14..17
|
|
720. bitwise_and_4_ints $6..9 &= $10..13
|
|
721. bitwise_and_4_ints $2..5 &= $6..9
|
|
722. bitwise_and_2_ints $2..3 &= $4..5
|
|
723. bitwise_and_int $2 &= $3
|
|
724. bitwise_and_int $1 &= $2
|
|
725. copy_slot_masked ok = Mask($1)
|
|
726. copy_constant m₂(0) = 0x41200000 (10.0)
|
|
727. copy_constant m₂(1) = 0x41A00000 (20.0)
|
|
728. copy_constant m₂(2) = 0x41F00000 (30.0)
|
|
729. copy_constant m₂(3) = 0x42200000 (40.0)
|
|
730. copy_4_slots_unmasked $1..4 = m₂
|
|
731. copy_constant $5 = 0x3F800000 (1.0)
|
|
732. copy_constant $6 = 0x40000000 (2.0)
|
|
733. copy_constant $7 = 0x40400000 (3.0)
|
|
734. copy_constant $8 = 0x40800000 (4.0)
|
|
735. sub_4_floats $1..4 -= $5..8
|
|
736. copy_4_slots_masked m₂ = Mask($1..4)
|
|
737. copy_slot_unmasked $1 = ok
|
|
738. copy_4_slots_unmasked $2..5 = m₂
|
|
739. copy_constant $6 = 0x41100000 (9.0)
|
|
740. copy_constant $7 = 0x41900000 (18.0)
|
|
741. copy_constant $8 = 0x41D80000 (27.0)
|
|
742. copy_constant $9 = 0x42100000 (36.0)
|
|
743. cmpeq_4_floats $2..5 = equal($2..5, $6..9)
|
|
744. bitwise_and_2_ints $2..3 &= $4..5
|
|
745. bitwise_and_int $2 &= $3
|
|
746. bitwise_and_int $1 &= $2
|
|
747. copy_slot_masked ok = Mask($1)
|
|
748. copy_constant m₃(0) = 0x40000000 (2.0)
|
|
749. copy_constant m₃(1) = 0x40800000 (4.0)
|
|
750. copy_constant m₃(2) = 0x40C00000 (6.0)
|
|
751. copy_constant m₃(3) = 0x41000000 (8.0)
|
|
752. copy_4_slots_unmasked $1..4 = m₃
|
|
753. copy_constant $5 = 0x40000000 (2.0)
|
|
754. copy_constant $6 = 0x40000000 (2.0)
|
|
755. copy_constant $7 = 0x40000000 (2.0)
|
|
756. copy_constant $8 = 0x40800000 (4.0)
|
|
757. div_4_floats $1..4 /= $5..8
|
|
758. copy_4_slots_masked m₃ = Mask($1..4)
|
|
759. copy_slot_unmasked $1 = ok
|
|
760. copy_4_slots_unmasked $2..5 = m₃
|
|
761. copy_constant $6 = 0x3F800000 (1.0)
|
|
762. copy_constant $7 = 0x40000000 (2.0)
|
|
763. copy_constant $8 = 0x40400000 (3.0)
|
|
764. copy_constant $9 = 0x40000000 (2.0)
|
|
765. cmpeq_4_floats $2..5 = equal($2..5, $6..9)
|
|
766. bitwise_and_2_ints $2..3 &= $4..5
|
|
767. bitwise_and_int $2 &= $3
|
|
768. bitwise_and_int $1 &= $2
|
|
769. copy_slot_masked ok = Mask($1)
|
|
770. copy_constant m₄(0) = 0x3F800000 (1.0)
|
|
771. copy_constant m₄(1) = 0x40000000 (2.0)
|
|
772. copy_constant m₄(2) = 0x40E00000 (7.0)
|
|
773. copy_constant m₄(3) = 0x40800000 (4.0)
|
|
774. copy_4_slots_unmasked $34..37 = m₄
|
|
775. swizzle_3 $35..37 = ($35..37).yxz
|
|
776. copy_constant $38 = 0x40400000 (3.0)
|
|
777. copy_constant $39 = 0x40A00000 (5.0)
|
|
778. copy_constant $40 = 0x40400000 (3.0)
|
|
779. copy_constant $41 = 0x40000000 (2.0)
|
|
780. copy_2_slots_unmasked $1..2 = $34..35
|
|
781. copy_2_slots_unmasked $3..4 = $38..39
|
|
782. dot_2_floats $1 = dot($1..2, $3..4)
|
|
783. copy_4_slots_unmasked $2..5 = $36..39
|
|
784. dot_2_floats $2 = dot($2..3, $4..5)
|
|
785. copy_2_slots_unmasked $3..4 = $34..35
|
|
786. copy_2_slots_unmasked $5..6 = $40..41
|
|
787. dot_2_floats $3 = dot($3..4, $5..6)
|
|
788. copy_2_slots_unmasked $4..5 = $36..37
|
|
789. copy_2_slots_unmasked $6..7 = $40..41
|
|
790. dot_2_floats $4 = dot($4..5, $6..7)
|
|
791. copy_4_slots_masked m₄ = Mask($1..4)
|
|
792. copy_slot_unmasked $1 = ok
|
|
793. copy_4_slots_unmasked $2..5 = m₄
|
|
794. copy_constant $6 = 0x42180000 (38.0)
|
|
795. copy_constant $7 = 0x41D00000 (26.0)
|
|
796. copy_constant $8 = 0x41880000 (17.0)
|
|
797. copy_constant $9 = 0x41600000 (14.0)
|
|
798. cmpeq_4_floats $2..5 = equal($2..5, $6..9)
|
|
799. bitwise_and_2_ints $2..3 &= $4..5
|
|
800. bitwise_and_int $2 &= $3
|
|
801. bitwise_and_int $1 &= $2
|
|
802. copy_slot_masked ok = Mask($1)
|
|
803. copy_constant m₅(0) = 0x41200000 (10.0)
|
|
804. copy_constant m₅(1) = 0x40800000 (4.0)
|
|
805. copy_constant m₅(2) = 0x40000000 (2.0)
|
|
806. copy_constant m₅(3) = 0x41A00000 (20.0)
|
|
807. copy_constant m₅(4) = 0x40A00000 (5.0)
|
|
808. copy_constant m₅(5) = 0x40400000 (3.0)
|
|
809. copy_constant m₅(6) = 0x41200000 (10.0)
|
|
810. copy_constant m₅(7) = 0x40C00000 (6.0)
|
|
811. copy_constant m₅(8) = 0x40A00000 (5.0)
|
|
812. copy_4_slots_unmasked $34..37 = m₅(0..3)
|
|
813. copy_4_slots_unmasked $38..41 = m₅(4..7)
|
|
814. copy_slot_unmasked $42 = m₅(8)
|
|
815. shuffle $35..42 = ($35..42)[2 5 0 3 6 1 4 7]
|
|
816. copy_constant $43 = 0x40400000 (3.0)
|
|
817. copy_constant $44 = 0x40400000 (3.0)
|
|
818. copy_constant $45 = 0x40800000 (4.0)
|
|
819. copy_constant $46 = 0x40000000 (2.0)
|
|
820. copy_constant $47 = 0x40400000 (3.0)
|
|
821. copy_constant $48 = 0x40800000 (4.0)
|
|
822. copy_constant $49 = 0x40800000 (4.0)
|
|
823. copy_constant $50 = 0x41100000 (9.0)
|
|
824. copy_constant $51 = 0x40000000 (2.0)
|
|
825. copy_3_slots_unmasked $1..3 = $34..36
|
|
826. copy_3_slots_unmasked $4..6 = $43..45
|
|
827. dot_3_floats $1 = dot($1..3, $4..6)
|
|
828. copy_3_slots_unmasked $2..4 = $37..39
|
|
829. copy_3_slots_unmasked $5..7 = $43..45
|
|
830. dot_3_floats $2 = dot($2..4, $5..7)
|
|
831. copy_4_slots_unmasked $3..6 = $40..43
|
|
832. copy_2_slots_unmasked $7..8 = $44..45
|
|
833. dot_3_floats $3 = dot($3..5, $6..8)
|
|
834. copy_3_slots_unmasked $4..6 = $34..36
|
|
835. copy_3_slots_unmasked $7..9 = $46..48
|
|
836. dot_3_floats $4 = dot($4..6, $7..9)
|
|
837. copy_3_slots_unmasked $5..7 = $37..39
|
|
838. copy_3_slots_unmasked $8..10 = $46..48
|
|
839. dot_3_floats $5 = dot($5..7, $8..10)
|
|
840. copy_3_slots_unmasked $6..8 = $40..42
|
|
841. copy_3_slots_unmasked $9..11 = $46..48
|
|
842. dot_3_floats $6 = dot($6..8, $9..11)
|
|
843. copy_3_slots_unmasked $7..9 = $34..36
|
|
844. copy_3_slots_unmasked $10..12 = $49..51
|
|
845. dot_3_floats $7 = dot($7..9, $10..12)
|
|
846. copy_3_slots_unmasked $8..10 = $37..39
|
|
847. copy_3_slots_unmasked $11..13 = $49..51
|
|
848. dot_3_floats $8 = dot($8..10, $11..13)
|
|
849. copy_3_slots_unmasked $9..11 = $40..42
|
|
850. copy_3_slots_unmasked $12..14 = $49..51
|
|
851. dot_3_floats $9 = dot($9..11, $12..14)
|
|
852. copy_4_slots_masked m₅(0..3) = Mask($1..4)
|
|
853. copy_4_slots_masked m₅(4..7) = Mask($5..8)
|
|
854. copy_slot_masked m₅(8) = Mask($9)
|
|
855. copy_slot_unmasked $1 = ok
|
|
856. copy_4_slots_unmasked $2..5 = m₅(0..3)
|
|
857. copy_4_slots_unmasked $6..9 = m₅(4..7)
|
|
858. copy_slot_unmasked $10 = m₅(8)
|
|
859. copy_constant $11 = 0x43020000 (130.0)
|
|
860. copy_constant $12 = 0x424C0000 (51.0)
|
|
861. copy_constant $13 = 0x420C0000 (35.0)
|
|
862. copy_constant $14 = 0x42F00000 (120.0)
|
|
863. copy_constant $15 = 0x423C0000 (47.0)
|
|
864. copy_constant $16 = 0x42040000 (33.0)
|
|
865. copy_constant $17 = 0x43700000 (240.0)
|
|
866. copy_constant $18 = 0x42920000 (73.0)
|
|
867. copy_constant $19 = 0x42340000 (45.0)
|
|
868. cmpeq_n_floats $2..10 = equal($2..10, $11..19)
|
|
869. bitwise_and_4_ints $3..6 &= $7..10
|
|
870. bitwise_and_2_ints $3..4 &= $5..6
|
|
871. bitwise_and_int $3 &= $4
|
|
872. bitwise_and_int $2 &= $3
|
|
873. bitwise_and_int $1 &= $2
|
|
874. copy_slot_masked ok = Mask($1)
|
|
875. copy_slot_masked [test_matrix_op_matrix_half].result = Mask($1)
|
|
876. label label 0x00000002
|
|
877. copy_slot_masked $0 = Mask($1)
|
|
878. label label 0x00000001
|
|
879. load_condition_mask CondMask = $52
|
|
880. swizzle_4 $0..3 = ($0..3).xxxx
|
|
881. copy_4_constants $4..7 = colorRed
|
|
882. copy_4_constants $8..11 = colorGreen
|
|
883. mix_4_ints $0..3 = mix($4..7, $8..11, $0..3)
|
|
884. copy_4_slots_unmasked [main].result = $0..3
|
|
885. load_src src.rgba = [main].result
|