2
* Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
3
* Copyright © 2018 Google, Inc.
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:
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
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
25
* Rob Clark <robclark@freedesktop.org>
28
#include "pipe/p_state.h"
29
#include "util/u_memory.h"
30
#include "util/u_prim.h"
31
#include "util/u_string.h"
33
#include "freedreno_resource.h"
34
#include "freedreno_state.h"
36
#include "fd6_context.h"
39
#include "fd6_format.h"
40
#include "fd6_program.h"
47
draw_emit_xfb(struct fd_ringbuffer *ring, struct CP_DRAW_INDX_OFFSET_0 *draw0,
48
const struct pipe_draw_info *info,
49
const struct pipe_draw_indirect_info *indirect)
51
struct fd_stream_output_target *target =
52
fd_stream_output_target(indirect->count_from_stream_output);
53
struct fd_resource *offset = fd_resource(target->offset_buf);
55
/* All known firmware versions do not wait for WFI's with CP_DRAW_AUTO.
56
* Plus, for the common case where the counter buffer is written by
57
* vkCmdEndTransformFeedback, we need to wait for the CP_WAIT_MEM_WRITES to
58
* complete which means we need a WAIT_FOR_ME anyway.
60
OUT_PKT7(ring, CP_WAIT_FOR_ME, 0);
62
OUT_PKT7(ring, CP_DRAW_AUTO, 6);
63
OUT_RING(ring, pack_CP_DRAW_INDX_OFFSET_0(*draw0).value);
64
OUT_RING(ring, info->instance_count);
65
OUT_RELOC(ring, offset->bo, 0, 0, 0);
68
0); /* byte counter offset subtraced from the value read from above */
69
OUT_RING(ring, target->stride);
73
draw_emit_indirect(struct fd_ringbuffer *ring,
74
struct CP_DRAW_INDX_OFFSET_0 *draw0,
75
const struct pipe_draw_info *info,
76
const struct pipe_draw_indirect_info *indirect,
77
unsigned index_offset)
79
struct fd_resource *ind = fd_resource(indirect->buffer);
81
if (info->index_size) {
82
struct pipe_resource *idx = info->index.resource;
83
unsigned max_indices = (idx->width0 - index_offset) / info->index_size;
85
OUT_PKT(ring, CP_DRAW_INDX_INDIRECT, pack_CP_DRAW_INDX_OFFSET_0(*draw0),
86
A5XX_CP_DRAW_INDX_INDIRECT_INDX_BASE(fd_resource(idx)->bo,
88
A5XX_CP_DRAW_INDX_INDIRECT_3(.max_indices = max_indices),
89
A5XX_CP_DRAW_INDX_INDIRECT_INDIRECT(ind->bo, indirect->offset));
91
OUT_PKT(ring, CP_DRAW_INDIRECT, pack_CP_DRAW_INDX_OFFSET_0(*draw0),
92
A5XX_CP_DRAW_INDIRECT_INDIRECT(ind->bo, indirect->offset));
97
draw_emit(struct fd_ringbuffer *ring, struct CP_DRAW_INDX_OFFSET_0 *draw0,
98
const struct pipe_draw_info *info,
99
const struct pipe_draw_start_count_bias *draw, unsigned index_offset)
101
if (info->index_size) {
102
assert(!info->has_user_indices);
104
struct pipe_resource *idx_buffer = info->index.resource;
105
unsigned max_indices =
106
(idx_buffer->width0 - index_offset) / info->index_size;
108
OUT_PKT(ring, CP_DRAW_INDX_OFFSET, pack_CP_DRAW_INDX_OFFSET_0(*draw0),
109
CP_DRAW_INDX_OFFSET_1(.num_instances = info->instance_count),
110
CP_DRAW_INDX_OFFSET_2(.num_indices = draw->count),
111
CP_DRAW_INDX_OFFSET_3(.first_indx = draw->start),
112
A5XX_CP_DRAW_INDX_OFFSET_INDX_BASE(fd_resource(idx_buffer)->bo,
114
A5XX_CP_DRAW_INDX_OFFSET_6(.max_indices = max_indices));
116
OUT_PKT(ring, CP_DRAW_INDX_OFFSET, pack_CP_DRAW_INDX_OFFSET_0(*draw0),
117
CP_DRAW_INDX_OFFSET_1(.num_instances = info->instance_count),
118
CP_DRAW_INDX_OFFSET_2(.num_indices = draw->count));
123
fixup_draw_state(struct fd_context *ctx, struct fd6_emit *emit) assert_dt
125
if (ctx->last.dirty ||
126
(ctx->last.primitive_restart != emit->primitive_restart)) {
127
/* rasterizer state is effected by primitive-restart: */
128
fd_context_dirty(ctx, FD_DIRTY_RASTERIZER);
129
ctx->last.primitive_restart = emit->primitive_restart;
134
fd6_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info,
135
unsigned drawid_offset,
136
const struct pipe_draw_indirect_info *indirect,
137
const struct pipe_draw_start_count_bias *draw,
138
unsigned index_offset) assert_dt
140
struct fd6_context *fd6_ctx = fd6_context(ctx);
141
struct shader_info *gs_info = ir3_get_shader_info(ctx->prog.gs);
142
struct fd6_emit emit = {
146
.drawid_offset = drawid_offset,
147
.indirect = indirect,
154
.rasterflat = ctx->rasterizer->flatshade,
155
.layer_zero = !gs_info || !(gs_info->outputs_written & VARYING_BIT_LAYER),
156
.sample_shading = (ctx->min_samples > 1),
157
.msaa = (ctx->framebuffer.samples > 1),
159
.clip_plane_enable = ctx->rasterizer->clip_plane_enable,
161
.rasterflat = ctx->rasterizer->flatshade,
162
.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable,
163
.sprite_coord_mode = ctx->rasterizer->sprite_coord_mode,
164
.primitive_restart = info->primitive_restart && info->index_size,
165
.patch_vertices = ctx->patch_vertices,
168
if (!(ctx->prog.vs && ctx->prog.fs))
171
if (info->mode == PIPE_PRIM_PATCHES) {
172
emit.key.hs = ctx->prog.hs;
173
emit.key.ds = ctx->prog.ds;
175
if (!(ctx->prog.hs && ctx->prog.ds))
178
struct shader_info *ds_info = ir3_get_shader_info(emit.key.ds);
179
emit.key.key.tessellation = ir3_tess_mode(ds_info->tess._primitive_mode);
180
ctx->gen_dirty |= BIT(FD6_GROUP_PRIMITIVE_PARAMS);
182
struct shader_info *fs_info = ir3_get_shader_info(emit.key.fs);
183
emit.key.key.tcs_store_primid =
184
BITSET_TEST(ds_info->system_values_read, SYSTEM_VALUE_PRIMITIVE_ID) ||
185
(gs_info && BITSET_TEST(gs_info->system_values_read, SYSTEM_VALUE_PRIMITIVE_ID)) ||
186
(fs_info && (fs_info->inputs_read & (1ull << VARYING_SLOT_PRIMITIVE_ID)));
190
emit.key.key.has_gs = true;
191
ctx->gen_dirty |= BIT(FD6_GROUP_PRIMITIVE_PARAMS);
194
if (!(emit.key.hs || emit.key.ds || emit.key.gs || indirect))
195
fd6_vsc_update_sizes(ctx->batch, info, draw);
197
ir3_fixup_shader_state(&ctx->base, &emit.key.key);
199
if (!(ctx->gen_dirty & BIT(FD6_GROUP_PROG))) {
200
emit.prog = fd6_ctx->prog;
202
fd6_ctx->prog = fd6_emit_get_prog(&emit);
205
/* bail if compile failed: */
209
fixup_draw_state(ctx, &emit);
211
/* *after* fixup_shader_state(): */
212
emit.dirty = ctx->dirty;
213
emit.dirty_groups = ctx->gen_dirty;
215
emit.bs = fd6_emit_get_prog(&emit)->bs;
216
emit.vs = fd6_emit_get_prog(&emit)->vs;
217
emit.hs = fd6_emit_get_prog(&emit)->hs;
218
emit.ds = fd6_emit_get_prog(&emit)->ds;
219
emit.gs = fd6_emit_get_prog(&emit)->gs;
220
emit.fs = fd6_emit_get_prog(&emit)->fs;
222
if (emit.vs->need_driver_params || fd6_ctx->has_dp_state)
223
emit.dirty_groups |= BIT(FD6_GROUP_VS_DRIVER_PARAMS);
225
/* If we are doing xfb, we need to emit the xfb state on every draw: */
226
if (emit.prog->stream_output)
227
emit.dirty_groups |= BIT(FD6_GROUP_SO);
229
if (unlikely(ctx->stats_users > 0)) {
230
ctx->stats.vs_regs += ir3_shader_halfregs(emit.vs);
231
ctx->stats.hs_regs += COND(emit.hs, ir3_shader_halfregs(emit.hs));
232
ctx->stats.ds_regs += COND(emit.ds, ir3_shader_halfregs(emit.ds));
233
ctx->stats.gs_regs += COND(emit.gs, ir3_shader_halfregs(emit.gs));
234
ctx->stats.fs_regs += ir3_shader_halfregs(emit.fs);
237
struct fd_ringbuffer *ring = ctx->batch->draw;
239
struct CP_DRAW_INDX_OFFSET_0 draw0 = {
240
.prim_type = ctx->screen->primtypes[info->mode],
241
.vis_cull = USE_VISIBILITY,
242
.gs_enable = !!emit.key.gs,
245
if (indirect && indirect->count_from_stream_output) {
246
draw0.source_select = DI_SRC_SEL_AUTO_XFB;
247
} else if (info->index_size) {
248
draw0.source_select = DI_SRC_SEL_DMA;
249
draw0.index_size = fd4_size2indextype(info->index_size);
251
draw0.source_select = DI_SRC_SEL_AUTO_INDEX;
254
if (info->mode == PIPE_PRIM_PATCHES) {
255
uint32_t factor_stride = ir3_tess_factor_stride(emit.key.key.tessellation);
257
STATIC_ASSERT(IR3_TESS_ISOLINES == TESS_ISOLINES + 1);
258
STATIC_ASSERT(IR3_TESS_TRIANGLES == TESS_TRIANGLES + 1);
259
STATIC_ASSERT(IR3_TESS_QUADS == TESS_QUADS + 1);
260
draw0.patch_type = emit.key.key.tessellation - 1;
262
draw0.prim_type = DI_PT_PATCHES0 + ctx->patch_vertices;
263
draw0.tess_enable = true;
265
/* maximum number of patches that can fit in tess factor/param buffers */
266
uint32_t subdraw_size = MIN2(FD6_TESS_FACTOR_SIZE / factor_stride,
267
FD6_TESS_PARAM_SIZE / (emit.hs->output_size * 4));
268
/* convert from # of patches to draw count */
269
subdraw_size *= ctx->patch_vertices;
271
OUT_PKT7(ring, CP_SET_SUBDRAW_SIZE, 1);
272
OUT_RING(ring, subdraw_size);
274
ctx->batch->tessellation = true;
277
uint32_t index_start = info->index_size ? draw->index_bias : draw->start;
278
if (ctx->last.dirty || (ctx->last.index_start != index_start)) {
279
OUT_PKT4(ring, REG_A6XX_VFD_INDEX_OFFSET, 1);
280
OUT_RING(ring, index_start); /* VFD_INDEX_OFFSET */
281
ctx->last.index_start = index_start;
284
if (ctx->last.dirty || (ctx->last.instance_start != info->start_instance)) {
285
OUT_PKT4(ring, REG_A6XX_VFD_INSTANCE_START_OFFSET, 1);
286
OUT_RING(ring, info->start_instance); /* VFD_INSTANCE_START_OFFSET */
287
ctx->last.instance_start = info->start_instance;
290
uint32_t restart_index =
291
info->primitive_restart ? info->restart_index : 0xffffffff;
292
if (ctx->last.dirty || (ctx->last.restart_index != restart_index)) {
293
OUT_PKT4(ring, REG_A6XX_PC_RESTART_INDEX, 1);
294
OUT_RING(ring, restart_index); /* PC_RESTART_INDEX */
295
ctx->last.restart_index = restart_index;
298
// TODO move fd6_emit_streamout.. I think..
299
if (emit.dirty_groups)
300
fd6_emit_state(ring, &emit);
302
/* for debug after a lock up, write a unique counter value
303
* to scratch7 for each draw, to make it easier to match up
304
* register dumps to cmdstream. The combination of IB
305
* (scratch6) and DRAW is enough to "triangulate" the
306
* particular draw that caused lockup.
308
emit_marker6(ring, 7);
311
if (indirect->count_from_stream_output) {
312
draw_emit_xfb(ring, &draw0, info, indirect);
314
draw_emit_indirect(ring, &draw0, info, indirect, index_offset);
317
draw_emit(ring, &draw0, info, draw, index_offset);
320
emit_marker6(ring, 7);
321
fd_reset_wfi(ctx->batch);
323
if (emit.streamout_mask) {
324
struct fd_ringbuffer *ring = ctx->batch->draw;
326
for (unsigned i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
327
if (emit.streamout_mask & (1 << i)) {
328
fd6_event_write(ctx->batch, ring, FLUSH_SO_0 + i, false);
333
fd_context_all_clean(ctx);
339
fd6_clear_lrz(struct fd_batch *batch, struct fd_resource *zsbuf, double depth) assert_dt
341
struct fd_ringbuffer *ring;
342
struct fd_screen *screen = batch->ctx->screen;
344
ring = fd_batch_get_prologue(batch);
346
emit_marker6(ring, 7);
347
OUT_PKT7(ring, CP_SET_MARKER, 1);
348
OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(RM6_BYPASS));
349
emit_marker6(ring, 7);
353
OUT_REG(ring, A6XX_RB_CCU_CNTL(.color_offset = screen->ccu_offset_bypass));
356
A6XX_HLSQ_INVALIDATE_CMD(.vs_state = true, .hs_state = true,
357
.ds_state = true, .gs_state = true,
358
.fs_state = true, .cs_state = true,
359
.gfx_ibo = true, .cs_ibo = true,
360
.gfx_shared_const = true,
361
.gfx_bindless = 0x1f, .cs_bindless = 0x1f));
363
emit_marker6(ring, 7);
364
OUT_PKT7(ring, CP_SET_MARKER, 1);
365
OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(RM6_BLIT2DSCALE));
366
emit_marker6(ring, 7);
368
OUT_PKT4(ring, REG_A6XX_RB_2D_UNKNOWN_8C01, 1);
371
OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 13);
372
OUT_RING(ring, 0x00000000);
373
OUT_RING(ring, 0x00000000);
374
OUT_RING(ring, 0x00000000);
375
OUT_RING(ring, 0x00000000);
376
OUT_RING(ring, 0x00000000);
377
OUT_RING(ring, 0x00000000);
378
OUT_RING(ring, 0x00000000);
379
OUT_RING(ring, 0x00000000);
380
OUT_RING(ring, 0x00000000);
381
OUT_RING(ring, 0x00000000);
382
OUT_RING(ring, 0x00000000);
383
OUT_RING(ring, 0x00000000);
384
OUT_RING(ring, 0x00000000);
386
OUT_PKT4(ring, REG_A6XX_SP_2D_DST_FORMAT, 1);
387
OUT_RING(ring, 0x0000f410);
389
OUT_PKT4(ring, REG_A6XX_GRAS_2D_BLIT_CNTL, 1);
391
A6XX_GRAS_2D_BLIT_CNTL_COLOR_FORMAT(FMT6_16_UNORM) | 0x4f00080);
393
OUT_PKT4(ring, REG_A6XX_RB_2D_BLIT_CNTL, 1);
394
OUT_RING(ring, A6XX_RB_2D_BLIT_CNTL_COLOR_FORMAT(FMT6_16_UNORM) | 0x4f00080);
396
fd6_event_write(batch, ring, PC_CCU_FLUSH_COLOR_TS, true);
397
fd6_event_write(batch, ring, PC_CCU_INVALIDATE_COLOR, false);
400
OUT_PKT4(ring, REG_A6XX_RB_2D_SRC_SOLID_C0, 4);
401
OUT_RING(ring, fui(depth));
402
OUT_RING(ring, 0x00000000);
403
OUT_RING(ring, 0x00000000);
404
OUT_RING(ring, 0x00000000);
406
OUT_PKT4(ring, REG_A6XX_RB_2D_DST_INFO, 9);
407
OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(FMT6_16_UNORM) |
408
A6XX_RB_2D_DST_INFO_TILE_MODE(TILE6_LINEAR) |
409
A6XX_RB_2D_DST_INFO_COLOR_SWAP(WZYX));
410
OUT_RELOC(ring, zsbuf->lrz, 0, 0, 0);
411
OUT_RING(ring, A6XX_RB_2D_DST_PITCH(zsbuf->lrz_pitch * 2).value);
412
OUT_RING(ring, 0x00000000);
413
OUT_RING(ring, 0x00000000);
414
OUT_RING(ring, 0x00000000);
415
OUT_RING(ring, 0x00000000);
416
OUT_RING(ring, 0x00000000);
418
OUT_REG(ring, A6XX_GRAS_2D_SRC_TL_X(0), A6XX_GRAS_2D_SRC_BR_X(0),
419
A6XX_GRAS_2D_SRC_TL_Y(0), A6XX_GRAS_2D_SRC_BR_Y(0));
421
OUT_PKT4(ring, REG_A6XX_GRAS_2D_DST_TL, 2);
422
OUT_RING(ring, A6XX_GRAS_2D_DST_TL_X(0) | A6XX_GRAS_2D_DST_TL_Y(0));
423
OUT_RING(ring, A6XX_GRAS_2D_DST_BR_X(zsbuf->lrz_width - 1) |
424
A6XX_GRAS_2D_DST_BR_Y(zsbuf->lrz_height - 1));
426
fd6_event_write(batch, ring, 0x3f, false);
430
OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
431
OUT_RING(ring, screen->info->a6xx.magic.RB_UNKNOWN_8E04_blit);
433
OUT_PKT7(ring, CP_BLIT, 1);
434
OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_SCALE));
438
OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
439
OUT_RING(ring, 0x0); /* RB_UNKNOWN_8E04 */
441
fd6_event_write(batch, ring, PC_CCU_FLUSH_COLOR_TS, true);
442
fd6_event_write(batch, ring, PC_CCU_FLUSH_DEPTH_TS, true);
443
fd6_event_write(batch, ring, CACHE_FLUSH_TS, true);
446
fd6_cache_inv(batch, ring);
450
is_z32(enum pipe_format format)
453
case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
454
case PIPE_FORMAT_Z32_UNORM:
455
case PIPE_FORMAT_Z32_FLOAT:
463
fd6_clear(struct fd_context *ctx, unsigned buffers,
464
const union pipe_color_union *color, double depth,
465
unsigned stencil) assert_dt
467
struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
468
const bool has_depth = pfb->zsbuf;
469
unsigned color_buffers = buffers >> 2;
471
/* we need to do multisample clear on 3d pipe, so fallback to u_blitter: */
472
if (pfb->samples > 1)
475
/* If we're clearing after draws, fallback to 3D pipe clears. We could
476
* use blitter clears in the draw batch but then we'd have to patch up the
477
* gmem offsets. This doesn't seem like a useful thing to optimize for
479
if (ctx->batch->num_draws > 0)
482
u_foreach_bit (i, color_buffers)
483
ctx->batch->clear_color[i] = *color;
484
if (buffers & PIPE_CLEAR_DEPTH)
485
ctx->batch->clear_depth = depth;
486
if (buffers & PIPE_CLEAR_STENCIL)
487
ctx->batch->clear_stencil = stencil;
489
ctx->batch->fast_cleared |= buffers;
491
if (has_depth && (buffers & PIPE_CLEAR_DEPTH)) {
492
struct fd_resource *zsbuf = fd_resource(pfb->zsbuf->texture);
493
if (zsbuf->lrz && !is_z32(pfb->zsbuf->format)) {
494
zsbuf->lrz_valid = true;
495
zsbuf->lrz_direction = FD_LRZ_UNKNOWN;
496
fd6_clear_lrz(ctx->batch, zsbuf, depth);
504
fd6_draw_init(struct pipe_context *pctx) disable_thread_safety_analysis
506
struct fd_context *ctx = fd_context(pctx);
507
ctx->draw_vbo = fd6_draw_vbo;
508
ctx->clear = fd6_clear;