src/gallium/drivers/panfrost/pan_instancing.c

   1 /*
   2  * Copyright (C) 2018-2019 Alyssa Rosenzweig
   3  * Copyright (C) 2019 Collabora, Ltd.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  */
  25
  26 #include "pan_context.h"
  27
  28 /* See mali_job for notes on how this works. But basically, for small vertex
  29  * counts, we have a lookup table, and for large vertex counts, we look at the
  30  * high bits as a heuristic. This has to match exactly how the hardware
  31  * calculates this (which is why the algorithm is so weird) or else instancing
  32  * will break. */
  33
  34 /* Given an odd number (of the form 2k + 1), compute k */
  35 #define ODD(odd) ((odd - 1) >> 1)
  36
  37 /* Given the shift/odd pair, recover the original padded integer */
  38
  39 unsigned
  40 pan_expand_shift_odd(struct pan_shift_odd o)
  41 {
  42         unsigned odd = 2*o.odd + 1;
  43         unsigned shift = 1 << o.shift;
  44         return odd * shift;
  45 }
  46
  47 static inline struct pan_shift_odd
  48 pan_factored(unsigned pot, unsigned odd)
  49 {
  50         struct pan_shift_odd out;
  51
  52         assert(util_is_power_of_two_or_zero(pot));
  53         assert(odd & 1);
  54
  55         /* Odd is of the form (2k + 1) = (k << 1) + 1 = (k << 1) | 1.
  56          *
  57          * So (odd >> 1) = ((k << 1) | 1) >> 1 = ((k << 1) >> 1) | (1 >> 1)
  58          *  = k | 0 = k */
  59
  60         out.odd = (odd >> 1);
  61
  62         /* POT is the form (1 << shift) */
  63         out.shift = __builtin_ctz(pot);
  64
  65         return out;
  66 }
  67
  68
  69 /* For small vertices. Second argument is whether the primitive takes a
  70  * power-of-two argument, which determines how rounding works. True for POINTS
  71  * and LINES, false for TRIANGLES. Presumably true for QUADS but you'd be crazy
  72  * to try instanced quads on ES class hardware <3 */
  73
  74 static struct {
  75         unsigned pot;
  76         unsigned odd;
  77 } small_lut[] = {
  78         {  0, 1 },
  79         {  1, 1 },
  80         {  2, 1 },
  81         {  1, 3 },
  82         {  4, 1 },
  83         {  1, 5 },
  84         {  2, 3 },
  85         {  1, 7 },
  86         {  8, 1 },
  87         {  1, 9 },
  88         {  2, 5 },
  89         {  4, 3 }, /* 11 */
  90         {  4, 3 },
  91         {  2, 7 }, /* 13 */
  92         {  2, 7 },
  93         { 16, 1 }, /* 15 */
  94         { 16, 1 },
  95         {  2, 9 },
  96         {  4, 5 }, /* 20 */
  97         {  4, 5 }
  98 };
  99
 100 static struct pan_shift_odd
 101 panfrost_small_padded_vertex_count(unsigned idx)
 102 {
 103         return pan_factored(
 104                        small_lut[idx].pot,
 105                        small_lut[idx].odd);
 106 }
 107
 108 static struct pan_shift_odd
 109 panfrost_large_padded_vertex_count(uint32_t vertex_count)
 110 {
 111         struct pan_shift_odd out = { 0 };
 112
 113         /* First, we have to find the highest set one */
 114         unsigned highest = 32 - __builtin_clz(vertex_count);
 115
 116         /* Using that, we mask out the highest 4-bits */
 117         unsigned n = highest - 4;
 118         unsigned nibble = (vertex_count >> n) & 0xF;
 119
 120         /* Great, we have the nibble. Now we can just try possibilities. Note
 121          * that we don't care about the bottom most bit in most cases, and we
 122          * know the top bit must be 1 */
 123
 124         unsigned middle_two = (nibble >> 1) & 0x3;
 125
 126         switch (middle_two) {
 127         case 0b00:
 128                 if (nibble & 1)
 129                         return pan_factored(1 << n, 9);
 130                 else
 131                         return pan_factored(1 << (n + 1), 5);
 132         case 0b01:
 133                 return pan_factored(1 << (n + 2), 3);
 134         case 0b10:
 135                 return pan_factored(1 << (n + 1), 7);
 136         case 0b11:
 137                 return pan_factored(1 << (n + 4), 1);
 138         default:
 139                 unreachable("Invalid two bits");
 140         }
 141
 142         return out;
 143 }
 144
 145 struct pan_shift_odd
 146 panfrost_padded_vertex_count(
 147         unsigned vertex_count,
 148         bool pot)
 149 {
 150         assert(vertex_count > 0);
 151
 152         if (vertex_count < 20) {
 153                 /* Add an off-by-one if it won't align naturally (quirk of the hardware) */
 154                 //if (!pot)
 155                 //      vertex_count++;
 156
 157                 return panfrost_small_padded_vertex_count(vertex_count);
 158         } else
 159                 return panfrost_large_padded_vertex_count(vertex_count);
 160 }
 161
 162 /* The much, much more irritating case -- instancing is enabled. See
 163  * panfrost_job.h for notes on how this works */
 164
 165 static unsigned
 166 panfrost_vertex_instanced(
 167         struct panfrost_job *batch,
 168         struct panfrost_resource *rsrc,
 169         unsigned divisor,
 170         union mali_attr *attrs,
 171         mali_ptr addr,
 172         unsigned vertex_count,
 173         unsigned instance_count)
 174 {
 175         /* First, grab the padded vertex count */
 176
 177         struct pan_shift_odd o = {
 178                 .shift = batch->ctx->payloads[PIPE_SHADER_FRAGMENT].instance_shift,
 179                 .odd = batch->ctx->payloads[PIPE_SHADER_FRAGMENT].instance_odd,
 180         };
 181
 182         unsigned padded_count = batch->ctx->padded_count;
 183
 184         /* Depending if there is an instance divisor or not, packing varies.
 185          * When there is a divisor, the hardware-level divisor is actually the
 186          * product of the instance divisor and the padded count */
 187
 188         unsigned hw_divisor = padded_count * divisor;
 189
 190         if (divisor == 0) {
 191                 /* Per-vertex attributes use the MODULO mode. First, compute
 192                  * the modulus */
 193
 194                 attrs->elements |= MALI_ATTR_MODULO;
 195                 attrs->shift = o.shift;
 196                 attrs->extra_flags = o.odd;
 197
 198                 return 1;
 199         } else if (util_is_power_of_two_or_zero(hw_divisor)) {
 200                 /* If there is a divisor but the hardware divisor works out to
 201                  * a power of two (not terribly exceptional), we can use an
 202                  * easy path (just shifting) */
 203
 204                 attrs->elements |= MALI_ATTR_POT_DIVIDE;
 205                 attrs->shift = __builtin_ctz(hw_divisor);
 206
 207                 return 1;
 208         } else {
 209                 /* We have a NPOT divisor. Here's the fun one (multipling by
 210                  * the inverse and shifting) */
 211
 212                 /* floor(log2(d)) */
 213                 unsigned shift = util_logbase2(hw_divisor);
 214
 215                 /* m = ceil(2^(32 + shift) / d) */
 216                 uint64_t shift_hi = 32 + shift;
 217                 uint64_t t = 1ll << shift_hi;
 218                 double t_f = t;
 219                 double hw_divisor_d = hw_divisor;
 220                 double m_f = ceil(t_f / hw_divisor_d);
 221                 unsigned m = m_f;
 222
 223                 /* Default case */
 224                 unsigned magic_divisor = m, extra_flags = 0;
 225
 226                 /* e = 2^(shift + 32) % d */
 227                 uint64_t e = t % hw_divisor;
 228
 229                 /* Apply round-down algorithm? e <= 2^shift?. XXX: The blob
 230                  * seems to use a different condition */
 231                 if (e <= (1ll << shift)) {
 232                         magic_divisor = m - 1;
 233                         extra_flags = 1;
 234                 }
 235
 236                 /* Top flag implicitly set */
 237                 assert(magic_divisor & (1 << 31));
 238                 magic_divisor &= ~(1 << 31);
 239
 240                 /* Upload to two different slots */
 241
 242                 attrs[0].elements |= MALI_ATTR_NPOT_DIVIDE;
 243                 attrs[0].shift = shift;
 244                 attrs[0].extra_flags = extra_flags;
 245
 246                 attrs[1].unk = 0x20;
 247                 attrs[1].magic_divisor = magic_divisor;
 248                 attrs[1].zero = 0;
 249                 attrs[1].divisor = divisor;
 250
 251                 return 2;
 252         }
 253 }
 254
 255 void
 256 panfrost_emit_vertex_data(struct panfrost_job *batch)
 257 {
 258         struct panfrost_context *ctx = batch->ctx;
 259         struct panfrost_vertex_state *so = ctx->vertex;
 260
 261         /* Staged mali_attr, and index into them. i =/= k, depending on the
 262          * vertex buffer mask and instancing. Twice as much room is allocated,
 263          * for a worst case of NPOT_DIVIDEs which take up extra slot */
 264         union mali_attr attrs[PIPE_MAX_ATTRIBS * 2];
 265         unsigned k = 0;
 266
 267         unsigned vertex_count = ctx->vertex_count;
 268         unsigned instanced_count = ctx->instance_count;
 269
 270         for (unsigned i = 0; i < so->num_elements; ++i) {
 271                 /* We map a mali_attr to be 1:1 with the mali_attr_meta, which
 272                  * means duplicating some vertex buffers (who cares? aside from
 273                  * maybe some caching implications but I somehow doubt that
 274                  * matters) */
 275
 276                 struct pipe_vertex_element *elem = &so->pipe[i];
 277                 unsigned vbi = elem->vertex_buffer_index;
 278
 279                 /* The exception to 1:1 mapping is that we can have multiple
 280                  * entries (NPOT divisors), so we fixup anyways */
 281
 282                 so->hw[i].index = k;
 283
 284                 if (!(ctx->vb_mask & (1 << vbi))) continue;
 285
 286                 struct pipe_vertex_buffer *buf = &ctx->vertex_buffers[vbi];
 287                 struct panfrost_resource *rsrc = (struct panfrost_resource *) (buf->buffer.resource);
 288
 289                 if (!rsrc) continue;
 290
 291                 /* Align to 64 bytes by masking off the lower bits. This
 292                  * will be adjusted back when we fixup the src_offset in
 293                  * mali_attr_meta */
 294
 295                 mali_ptr raw_addr = rsrc->bo->gpu + buf->buffer_offset;
 296                 mali_ptr addr = raw_addr & ~63;
 297                 unsigned chopped_addr = raw_addr - addr;
 298
 299                 /* Add a dependency of the batch on the vertex buffer */
 300                 panfrost_job_add_bo(batch, rsrc->bo);
 301
 302                 /* Set common fields */
 303                 attrs[k].elements = addr;
 304                 attrs[k].stride = buf->stride;
 305
 306                 /* Since we advanced the base pointer, we shrink the buffer
 307                  * size */
 308                 attrs[k].size = rsrc->base.width0 - buf->buffer_offset;
 309
 310                 /* We need to add the extra size we masked off (for
 311                  * correctness) so the data doesn't get clamped away */
 312                 attrs[k].size += chopped_addr;
 313
 314                 /* For non-instancing make sure we initialize */
 315                 attrs[k].shift = attrs[k].extra_flags = 0;
 316
 317                 /* Instancing uses a dramatically different code path than
 318                  * linear, so dispatch for the actual emission now that the
 319                  * common code is finished */
 320
 321                 unsigned divisor = elem->instance_divisor;
 322
 323                 if (divisor && instanced_count == 1) {
 324                         /* Silly corner case where there's a divisor(=1) but
 325                          * there's no legitimate instancing. So we want *every*
 326                          * attribute to be the same. So set stride to zero so
 327                          * we don't go anywhere. */
 328
 329                         attrs[k].size = attrs[k].stride + chopped_addr;
 330                         attrs[k].stride = 0;
 331                         attrs[k++].elements |= MALI_ATTR_LINEAR;
 332                 } else if (instanced_count <= 1) {
 333                         /* Normal, non-instanced attributes */
 334                         attrs[k++].elements |= MALI_ATTR_LINEAR;
 335                 } else {
 336                         k += panfrost_vertex_instanced(
 337                                      batch, rsrc, divisor, &attrs[k], addr, vertex_count, instanced_count);
 338                 }
 339         }
 340
 341         /* Upload whatever we emitted and go */
 342
 343         ctx->payloads[PIPE_SHADER_VERTEX].postfix.attributes =
 344                 panfrost_upload_transient(ctx, attrs, k * sizeof(union mali_attr));
 345 }
 346
 347