(no commit message)
[libreriscv.git] / simple_v_extension / vector_ops.mdwn
1 [[!tag standards]]
2
3 # Vector Operations Extension to SV
4
5 This extension defines vector operations that would otherwise take several cycles to complete in software. With 3D priorities being to compute as many pixels per clock as possible, the normal RISC rules (reduce opcode count and make heavy use of macro op fusion) do not necessarily apply.
6
7 This extension is usually dependent on SV SUBVL being implemented. When SUBVL is set to define the length of a subvector the operations in this extension interpret the elements as a single vector.
8
9 Normally in SV all operations are scalar and independent, and the operations on them may inherently be independently parallelised, with the result being a vector of length exactly equal to the input vectors.
10
11 In this extension, the subvector itself is typically the unit, although some operations will work on scalars or standard vectors as well, or the result is a scalar that is dependent on all elements within the vector arguments.
12
13 However given that some of the parameters are vectors (with and without SUBVL set), and some are scalars (where SUBVL will not apply), some clear rules need to be defined as to how the operations work.
14
15 Examples which can require SUBVL include cross product and may in future involve complex numbers.
16
17 ## CORDIC
18
19 6 opcode options (fmt3):
20
21 * CORDIC.lin.rot vd, vs, beta
22 * CORDIC.cir.rot vd, vs, beta
23 * CORDIC.hyp.rot vd, vs, beta
24 * CORDIC.lin.vec vd, vs, beta
25 * CORDIC.cir.vec vd, vs, beta
26 * CORDIC.hyp.vec vd, vs, beta
27
28
29 | Instr | result | src1 | src2 | SUBVL | VL | Notes |
30 | ------------------ | ------ | ---- | ---- | ----- | -- | ------ |
31 | CORDIC.x.t vd, vs1, rs2 | vec2 | vec2 | scal | 2 | any | src2 ignores SUBVL |
32
33 SUBVL must be set to 2 and applies to vd and vs. SUBVL is *ignored* on beta. vd and vs must be marked as vectors.
34
35 VL may be applied. beta as a scalar is ok (applies across all vectors vd and vs). Predication is also ok (single predication) sourced from vd. Use of swizzle is also ok.
36
37 Non vector args vd, vs are reserved encodings.
38
39 CORDIC is an extremely general-purpose algorithm useful for a huge number
40 of diverse purposes. In its full form it does however require quite a
41 few parameters, one of which is a vector, making it awkward to include in
42 a standard "scalar" ISA. Additionally the coordinates can be set to circular,
43 linear or hyperbolic, producing three different modes, and the algorithm
44 may also be run in either "vector" mode or "rotation" mode. See [[discussion]]
45
46 CORDIC can also be used for performing DCT. See
47 <https://arxiv.org/abs/1606.02424>
48
49 CORDIC has several RADIX-4 papers for efficient pipelining. Each stage requires its own ROM tables which can get costly. Two combinatorial blocks may be chained together to double the RADIX and halve the pipeline depth, at the cost of doubling the latency.
50
51 Also, to get good accuracy, particularly at the limits of CORDIC input range, requires double the bitwidth of the output in internal computations. This similar to how MUL requires double the bitwidth to compute.
52
53 Links:
54
55 * <http://www.myhdl.org/docs/examples/sinecomp/>
56 * <https://www.atlantis-press.com/proceedings/jcis2006/232>
57
58 ## Vector cross product
59
60 * VCROSS vd, vs1, vs1
61
62 Result is the cross product of x and y.
63
64 SUBVL must be set to 3, and all regs must be vectors. VL nonzero produces multiple results in vd.
65
66 The resulting components are, in order:
67
68 x[1] * y[2] - y[1] * x[2]
69 x[2] * y[0] - y[2] * x[0]
70 x[0] * y[1] - y[0] * x[1]
71
72 All the operands must be vectors of 3 components of a floating-point type.
73
74 Pseudocode:
75
76 vec3 a, b; // elements in order a.x, a.y, a.z
77 // compute a cross b:
78 vec3 t1 = a.yzx; // produce vector [a.y, a.z, a.x]
79 vec3 t2 = b.zxy;
80 vec3 t3 = a.zxy;
81 vec3 t4 = b.yzx;
82 vec3 p = t3 * t4;
83 vec3 cross = t1 * t2 - p;
84
85 Assembler:
86
87 fswizzlei,2130 F4, F1
88 fswizzlei,1320 F5, F1
89 fswizzlei,2130 F6, F2
90 fswizzlei,1320 F7, F2
91 fmul F8, F5, F6
92 fmulsub F3, F4, F7, F8
93
94 ## Vector dot product
95
96
97 * VDOT rd, vs1, vs2
98
99 Computes the dot product of two vectors. Internal accuracy must be
100 greater than the input vectors and the result.
101
102 There are two possible argument options:
103
104 * SUBVL=2,3,4 vs1 and vs2 set as vectors, multiple results are generated. When VL is set, only the first (unpredicated) SUBVector is used to create a result, if rd is scalar (standard behaviour for single predication). Otherwise, if rd is a vector, multiple scalar results are calculated (i.e. SUBVL is always ignored for rd). Swizzling may be applied.
105 * When rd=scalar, SUBVL=1 and vs1=vec, vs2=vec, one scalar result is generated from the entire src vectors. Predication is allowed on the src vectors.
106
107
108 | Instr | result | src1 | src2 | SUBVL | VL |
109 | ------------------ | ------ | ---- | ---- | ----- | -- |
110 | VDOT rd, vs1, vs2 | scal | vec | vec | 2-4 | any |
111 | VDOT rd, vs1, vs2 | scal | vec | vec | 1 | any |
112
113 Pseudocode in python:
114
115 from operator import mul
116 sum(map(mul, A, B))
117
118 Pseudocode in c:
119
120 double dot_product(float v[], float u[], int n)
121 {
122 double result = 0.0;
123 for (int i = 0; i < n; i++)
124 result += v[i] * u[i];
125 return result;
126 }
127
128 ## Vector Normalisation (not included)
129
130 Vector normalisation may be performed through dot product, recip square root and multiplication:
131
132 fdot F3, F1, F1 # vector dot with self
133 rcpsqrta F3, F3
134 fscale,0 F2, F3, F1
135
136 Or it may be performed through VLEN (Vector length) and division.
137
138 ## Vector length
139
140 * rd=scalar, vs1=vec (SUBVL=1)
141 * rd=scalar, vs1=vec (SUBVL=2,3,4) only 1 (predication rules apply)
142 * rd=vec, SUBVL ignored; vs1=vec, SUBVL=2,3,4
143 * rd=vec, SUBVL ignored; vs1=vec, SUBVL=1: reserved encoding.
144
145 * VLEN rd, vs1
146
147 The scalar length of a vector:
148
149 sqrt(x[0]^2 + x[1]^2 + ...).
150
151 One option is for this to be a macro op fusion sequence, with inverse-sqrt also being a second macro op sequence suitable for normalisation.
152
153 ## Vector distance
154
155 * VDIST rd, vs1, vs2
156
157 The scalar distance between two vectors. Subtracts one vector from the
158 other and returns length:
159
160 length(v0 - v1)
161
162 ## Vector LERP
163
164 * VLERP vd, vs1, rs2 # SUBVL=2: vs1.v0 vs1.v1
165
166 | Instr | result | src1 | src2 | SUBVL | VL |
167 | ------------------ | ------ | ---- | ---- | ----- | -- |
168 | VLERP vd, vs1, rs2 | vec2 | vec2 | scal | 2 | any |
169
170 Known as **fmix** in GLSL.
171
172 <https://en.m.wikipedia.org/wiki/Linear_interpolation>
173
174 Pseudocode:
175
176 // Imprecise method, which does not guarantee v = v1 when t = 1,
177 // due to floating-point arithmetic error.
178 // This form may be used when the hardware has a native fused
179 // multiply-add instruction.
180 float lerp(float v0, float v1, float t) {
181 return v0 + t * (v1 - v0);
182 }
183
184 // Precise method, which guarantees v = v1 when t = 1.
185 float lerp(float v0, float v1, float t) {
186 return (1 - t) * v0 + t * v1;
187 }
188
189 ## Vector SLERP
190
191 * VSLERP vd, vs1, vs2, rs3
192
193 Not recommended as it is not commonly used and has several trigonometric
194 functions, although CORDIC in vector rotate circular mode is designed for this purpose. Also a costly 4 arg operation.
195
196 <https://en.m.wikipedia.org/wiki/Slerp>
197
198 Pseudocode:
199
200 Quaternion slerp(Quaternion v0, Quaternion v1, double t) {
201 // Only unit quaternions are valid rotations.
202 // Normalize to avoid undefined behavior.
203 v0.normalize();
204 v1.normalize();
205
206 // Compute the cosine of the angle between the two vectors.
207 double dot = dot_product(v0, v1);
208
209 // If the dot product is negative, slerp won't take
210 // the shorter path. Note that v1 and -v1 are equivalent when
211 // the negation is applied to all four components. Fix by
212 // reversing one quaternion.
213 if (dot < 0.0f) {
214 v1 = -v1;
215 dot = -dot;
216 }
217
218 const double DOT_THRESHOLD = 0.9995;
219 if (dot > DOT_THRESHOLD) {
220 // If the inputs are too close for comfort, linearly interpolate
221 // and normalize the result.
222
223 Quaternion result = v0 + t*(v1 - v0);
224 result.normalize();
225 return result;
226 }
227
228 // Since dot is in range [0, DOT_THRESHOLD], acos is safe
229 double theta_0 = acos(dot); // theta_0 = angle between input vectors
230 double theta = theta_0*t; // theta = angle between v0 and result
231 double sin_theta = sin(theta); // compute this value only once
232 double sin_theta_0 = sin(theta_0); // compute this value only once
233
234 double s0 = cos(theta) - dot * sin_theta / sin_theta_0; // == sin(theta_0 - theta) / sin(theta_0)
235 double s1 = sin_theta / sin_theta_0;
236
237 return (s0 * v0) + (s1 * v1);
238 }
239
240 However this algorithm does not involve transcendentals except in
241 the computation of the tables: <https://en.wikipedia.org/wiki/CORDIC#Rotation_mode>
242
243 function v = cordic(beta,n)
244 % This function computes v = [cos(beta), sin(beta)] (beta in radians)
245 % using n iterations. Increasing n will increase the precision.
246
247 if beta < -pi/2 || beta > pi/2
248 if beta < 0
249 v = cordic(beta + pi, n);
250 else
251 v = cordic(beta - pi, n);
252 end
253 v = -v; % flip the sign for second or third quadrant
254 return
255 end
256
257 % Initialization of tables of constants used by CORDIC
258 % need a table of arctangents of negative powers of two, in radians:
259 % angles = atan(2.^-(0:27));
260 angles = [ ...
261 0.78539816339745 0.46364760900081
262 0.24497866312686 0.12435499454676 ...
263 0.06241880999596 0.03123983343027
264 0.01562372862048 0.00781234106010 ...
265 0.00390623013197 0.00195312251648
266 0.00097656218956 0.00048828121119 ...
267 0.00024414062015 0.00012207031189
268 0.00006103515617 0.00003051757812 ...
269 0.00001525878906 0.00000762939453
270 0.00000381469727 0.00000190734863 ...
271 0.00000095367432 0.00000047683716
272 0.00000023841858 0.00000011920929 ...
273 0.00000005960464 0.00000002980232
274 0.00000001490116 0.00000000745058 ];
275 % and a table of products of reciprocal lengths of vectors [1, 2^-2j]:
276 % Kvalues = cumprod(1./abs(1 + 1j*2.^(-(0:23))))
277 Kvalues = [ ...
278 0.70710678118655 0.63245553203368
279 0.61357199107790 0.60883391251775 ...
280 0.60764825625617 0.60735177014130
281 0.60727764409353 0.60725911229889 ...
282 0.60725447933256 0.60725332108988
283 0.60725303152913 0.60725295913894 ...
284 0.60725294104140 0.60725293651701
285 0.60725293538591 0.60725293510314 ...
286 0.60725293503245 0.60725293501477
287 0.60725293501035 0.60725293500925 ...
288 0.60725293500897 0.60725293500890
289 0.60725293500889 0.60725293500888 ];
290 Kn = Kvalues(min(n, length(Kvalues)));
291
292 % Initialize loop variables:
293 v = [1;0]; % start with 2-vector cosine and sine of zero
294 poweroftwo = 1;
295 angle = angles(1);
296
297 % Iterations
298 for j = 0:n-1;
299 if beta < 0
300 sigma = -1;
301 else
302 sigma = 1;
303 end
304 factor = sigma * poweroftwo;
305 % Note the matrix multiplication can be done using scaling by
306 % powers of two and addition subtraction
307 R = [1, -factor; factor, 1];
308 v = R * v; % 2-by-2 matrix multiply
309 beta = beta - sigma * angle; % update the remaining angle
310 poweroftwo = poweroftwo / 2;
311 % update the angle from table, or eventually by just dividing by two
312 if j+2 > length(angles)
313 angle = angle / 2;
314 else
315 angle = angles(j+2);
316 end
317 end
318
319 % Adjust length of output vector to be [cos(beta), sin(beta)]:
320 v = v * Kn;
321 return
322
323 endfunction
324
325 2x2 matrix multiply can be done with shifts and adds:
326
327 x = v[0] - sigma * (v[1] * 2^(-j));
328 y = sigma * (v[0] * 2^(-j)) + v[1];
329 v = [x; y];
330
331 The technique is outlined in a paper as being applicable to 3D:
332 <https://www.atlantis-press.com/proceedings/jcis2006/232>
333
334 # Expensive 3-operand OP32 operations
335
336 3-operand operations are extremely expensive in terms of OP32 encoding space. A potential idea is to embed 3 RVC register formats across two out of three 5-bit fields rs1/rs2/rd
337
338 Another is to overwrite one of the src registers.
339
340 # Opcode Table
341
342 TODO
343
344 # Links
345
346 * <http://lists.libre-riscv.org/pipermail/libre-riscv-dev/2019-September/002736.html>
347 * <http://lists.libre-riscv.org/pipermail/libre-riscv-dev/2019-September/002733.html>
348 * <http://bugs.libre-riscv.org/show_bug.cgi?id=142>
349
350 Research Papers
351
352 * <https://www.researchgate.net/publication/2938554_PLX_FP_An_Efficient_Floating-Point_Instruction_Set_for_3D_Graphics>