1/*
2 * Copyright (C) 2019 Rob Clark <robclark@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 *    Rob Clark <robclark@freedesktop.org>
25 */
26
27#include "pipe/p_state.h"
28
29#include "freedreno_resource.h"
30
31#include "fd6_compute.h"
32#include "fd6_context.h"
33#include "fd6_emit.h"
34
35struct fd6_compute_stateobj {
36	struct ir3_shader *shader;
37};
38
39
40static void *
41fd6_create_compute_state(struct pipe_context *pctx,
42		const struct pipe_compute_state *cso)
43{
44	struct fd_context *ctx = fd_context(pctx);
45
46	/* req_input_mem will only be non-zero for cl kernels (ie. clover).
47	 * This isn't a perfect test because I guess it is possible (but
48	 * uncommon) for none for the kernel parameters to be a global,
49	 * but ctx->set_global_bindings() can't fail, so this is the next
50	 * best place to fail if we need a newer version of kernel driver:
51	 */
52	if ((cso->req_input_mem > 0) &&
53			fd_device_version(ctx->dev) < FD_VERSION_BO_IOVA) {
54		return NULL;
55	}
56
57	struct ir3_compiler *compiler = ctx->screen->compiler;
58	struct fd6_compute_stateobj *so = CALLOC_STRUCT(fd6_compute_stateobj);
59	so->shader = ir3_shader_create_compute(compiler, cso, &ctx->debug, pctx->screen);
60	return so;
61}
62
63static void
64fd6_delete_compute_state(struct pipe_context *pctx, void *hwcso)
65{
66	struct fd6_compute_stateobj *so = hwcso;
67	ir3_shader_destroy(so->shader);
68	free(so);
69}
70
71/* maybe move to fd6_program? */
72static void
73cs_program_emit(struct fd_ringbuffer *ring, struct ir3_shader_variant *v,
74		const struct pipe_grid_info *info)
75{
76	const struct ir3_info *i = &v->info;
77	enum a3xx_threadsize thrsz = FOUR_QUADS;
78
79	OUT_PKT4(ring, REG_A6XX_HLSQ_UPDATE_CNTL, 1);
80	OUT_RING(ring, 0xff);
81
82	unsigned constlen = align(v->constlen, 4);
83	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_CNTL, 1);
84	OUT_RING(ring, A6XX_HLSQ_CS_CNTL_CONSTLEN(constlen) |
85			A6XX_HLSQ_CS_CNTL_ENABLED);
86
87	OUT_PKT4(ring, REG_A6XX_SP_CS_CONFIG, 2);
88	OUT_RING(ring, A6XX_SP_CS_CONFIG_ENABLED |
89			A6XX_SP_CS_CONFIG_NIBO(v->image_mapping.num_ibo) |
90			A6XX_SP_CS_CONFIG_NTEX(v->num_samp) |
91			A6XX_SP_CS_CONFIG_NSAMP(v->num_samp));    /* SP_VS_CONFIG */
92	OUT_RING(ring, v->instrlen);                      /* SP_VS_INSTRLEN */
93
94	OUT_PKT4(ring, REG_A6XX_SP_CS_CTRL_REG0, 1);
95	OUT_RING(ring, A6XX_SP_CS_CTRL_REG0_THREADSIZE(thrsz) |
96			A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT(i->max_reg + 1) |
97			A6XX_SP_CS_CTRL_REG0_MERGEDREGS |
98			A6XX_SP_CS_CTRL_REG0_BRANCHSTACK(v->branchstack) |
99			COND(v->need_pixlod, A6XX_SP_CS_CTRL_REG0_PIXLODENABLE));
100
101	OUT_PKT4(ring, REG_A6XX_SP_CS_UNKNOWN_A9B1, 1);
102	OUT_RING(ring, 0x41);
103
104	uint32_t local_invocation_id, work_group_id;
105	local_invocation_id = ir3_find_sysval_regid(v, SYSTEM_VALUE_LOCAL_INVOCATION_ID);
106	work_group_id = ir3_find_sysval_regid(v, SYSTEM_VALUE_WORK_GROUP_ID);
107
108	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_CNTL_0, 2);
109	OUT_RING(ring, A6XX_HLSQ_CS_CNTL_0_WGIDCONSTID(work_group_id) |
110		A6XX_HLSQ_CS_CNTL_0_UNK0(regid(63, 0)) |
111		A6XX_HLSQ_CS_CNTL_0_UNK1(regid(63, 0)) |
112		A6XX_HLSQ_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
113	OUT_RING(ring, 0x2fc);             /* HLSQ_CS_UNKNOWN_B998 */
114
115	OUT_PKT4(ring, REG_A6XX_SP_CS_OBJ_START_LO, 2);
116	OUT_RELOC(ring, v->bo, 0, 0, 0);   /* SP_CS_OBJ_START_LO/HI */
117
118	if (v->instrlen > 0)
119		fd6_emit_shader(ring, v);
120}
121
122static void
123fd6_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info)
124{
125	struct fd6_compute_stateobj *so = ctx->compute;
126	struct ir3_shader_key key = {};
127	struct ir3_shader_variant *v;
128	struct fd_ringbuffer *ring = ctx->batch->draw;
129	unsigned i, nglobal = 0;
130
131	fd6_emit_restore(ctx->batch, ring);
132
133	v = ir3_shader_variant(so->shader, key, false, &ctx->debug);
134	if (!v)
135		return;
136
137	if (ctx->dirty_shader[PIPE_SHADER_COMPUTE] & FD_DIRTY_SHADER_PROG)
138		cs_program_emit(ring, v, info);
139
140	fd6_emit_cs_state(ctx, ring, v);
141	ir3_emit_cs_consts(v, ring, ctx, info);
142
143	foreach_bit(i, ctx->global_bindings.enabled_mask)
144		nglobal++;
145
146	if (nglobal > 0) {
147		/* global resources don't otherwise get an OUT_RELOC(), since
148		 * the raw ptr address is emitted in ir3_emit_cs_consts().
149		 * So to make the kernel aware that these buffers are referenced
150		 * by the batch, emit dummy reloc's as part of a no-op packet
151		 * payload:
152		 */
153		OUT_PKT7(ring, CP_NOP, 2 * nglobal);
154		foreach_bit(i, ctx->global_bindings.enabled_mask) {
155			struct pipe_resource *prsc = ctx->global_bindings.buf[i];
156			OUT_RELOCW(ring, fd_resource(prsc)->bo, 0, 0, 0);
157		}
158	}
159
160	OUT_PKT7(ring, CP_SET_MARKER, 1);
161	OUT_RING(ring, A2XX_CP_SET_MARKER_0_MODE(0x8));
162
163	const unsigned *local_size = info->block; // v->shader->nir->info->cs.local_size;
164	const unsigned *num_groups = info->grid;
165	/* for some reason, mesa/st doesn't set info->work_dim, so just assume 3: */
166	const unsigned work_dim = info->work_dim ? info->work_dim : 3;
167	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_NDRANGE_0, 7);
168	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_0_KERNELDIM(work_dim) |
169		A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEX(local_size[0] - 1) |
170		A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEY(local_size[1] - 1) |
171		A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEZ(local_size[2] - 1));
172	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_1_GLOBALSIZE_X(local_size[0] * num_groups[0]));
173	OUT_RING(ring, 0);            /* HLSQ_CS_NDRANGE_2_GLOBALOFF_X */
174	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_3_GLOBALSIZE_Y(local_size[1] * num_groups[1]));
175	OUT_RING(ring, 0);            /* HLSQ_CS_NDRANGE_4_GLOBALOFF_Y */
176	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_5_GLOBALSIZE_Z(local_size[2] * num_groups[2]));
177	OUT_RING(ring, 0);            /* HLSQ_CS_NDRANGE_6_GLOBALOFF_Z */
178
179	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_KERNEL_GROUP_X, 3);
180	OUT_RING(ring, 1);            /* HLSQ_CS_KERNEL_GROUP_X */
181	OUT_RING(ring, 1);            /* HLSQ_CS_KERNEL_GROUP_Y */
182	OUT_RING(ring, 1);            /* HLSQ_CS_KERNEL_GROUP_Z */
183
184	if (info->indirect) {
185		struct fd_resource *rsc = fd_resource(info->indirect);
186
187		OUT_PKT7(ring, CP_EXEC_CS_INDIRECT, 4);
188		OUT_RING(ring, 0x00000000);
189		OUT_RELOC(ring, rsc->bo, info->indirect_offset, 0, 0);  /* ADDR_LO/HI */
190		OUT_RING(ring, A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEX(local_size[0] - 1) |
191				A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEY(local_size[1] - 1) |
192				A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEZ(local_size[2] - 1));
193	} else {
194		OUT_PKT7(ring, CP_EXEC_CS, 4);
195		OUT_RING(ring, 0x00000000);
196		OUT_RING(ring, CP_EXEC_CS_1_NGROUPS_X(info->grid[0]));
197		OUT_RING(ring, CP_EXEC_CS_2_NGROUPS_Y(info->grid[1]));
198		OUT_RING(ring, CP_EXEC_CS_3_NGROUPS_Z(info->grid[2]));
199	}
200
201	OUT_WFI5(ring);
202
203	fd6_cache_flush(ctx->batch, ring);
204}
205
206void
207fd6_compute_init(struct pipe_context *pctx)
208{
209	struct fd_context *ctx = fd_context(pctx);
210	ctx->launch_grid = fd6_launch_grid;
211	pctx->create_compute_state = fd6_create_compute_state;
212	pctx->delete_compute_state = fd6_delete_compute_state;
213}
214