1/*
2 * Copyright (C) 2019 Collabora, Ltd.
3 * Copyright (C) 2019 Red Hat Inc.
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 * Authors (Collabora):
25 *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
26 *
27 */
28
29#include "pan_context.h"
30#include "pan_bo.h"
31#include "pan_shader.h"
32#include "util/u_memory.h"
33#include "nir_serialize.h"
34
35/* Compute CSOs are tracked like graphics shader CSOs, but are
36 * considerably simpler. We do not implement multiple
37 * variants/keying. So the CSO create function just goes ahead and
38 * compiles the thing. */
39
40static void *
41panfrost_create_compute_state(
42        struct pipe_context *pctx,
43        const struct pipe_compute_state *cso)
44{
45        struct panfrost_context *ctx = pan_context(pctx);
46        struct panfrost_screen *screen = pan_screen(pctx->screen);
47
48        struct panfrost_shader_variants *so = CALLOC_STRUCT(panfrost_shader_variants);
49        so->req_input_mem = cso->req_input_mem;
50
51        struct panfrost_shader_state *v = calloc(1, sizeof(*v));
52        so->variants = v;
53
54        so->variant_count = 1;
55        so->active_variant = 0;
56
57        nir_shader *deserialized = NULL;
58
59        if (cso->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
60                struct blob_reader reader;
61                const struct pipe_binary_program_header *hdr = cso->prog;
62
63                blob_reader_init(&reader, hdr->blob, hdr->num_bytes);
64
65                const struct nir_shader_compiler_options *options =
66                        screen->vtbl.get_compiler_options();
67
68                deserialized = nir_deserialize(NULL, options, &reader);
69        } else {
70                assert(cso->ir_type == PIPE_SHADER_IR_NIR && "TGSI kernels unsupported");
71        }
72
73        panfrost_shader_compile(pctx->screen, &ctx->shaders, &ctx->descs,
74                                deserialized ?: cso->prog, v);
75
76        /* There are no variants so we won't need the NIR again */
77        ralloc_free(deserialized);
78
79        return so;
80}
81
82static void
83panfrost_bind_compute_state(struct pipe_context *pipe, void *cso)
84{
85        struct panfrost_context *ctx = pan_context(pipe);
86        ctx->shader[PIPE_SHADER_COMPUTE] = cso;
87}
88
89static void
90panfrost_delete_compute_state(struct pipe_context *pipe, void *cso)
91{
92        struct panfrost_shader_variants *so =
93                (struct panfrost_shader_variants *)cso;
94
95        free(so->variants);
96        free(cso);
97}
98
99static void
100panfrost_set_compute_resources(struct pipe_context *pctx,
101                         unsigned start, unsigned count,
102                         struct pipe_surface **resources)
103{
104        /* TODO */
105}
106
107static void
108panfrost_set_global_binding(struct pipe_context *pctx,
109                      unsigned first, unsigned count,
110                      struct pipe_resource **resources,
111                      uint32_t **handles)
112{
113        if (!resources)
114                return;
115
116        struct panfrost_context *ctx = pan_context(pctx);
117        struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
118
119        for (unsigned i = first; i < first + count; ++i) {
120                struct panfrost_resource *rsrc = pan_resource(resources[i]);
121                panfrost_batch_write_rsrc(batch, rsrc, PIPE_SHADER_COMPUTE);
122
123                util_range_add(&rsrc->base, &rsrc->valid_buffer_range,
124                                0, rsrc->base.width0);
125
126                /* The handle points to uint32_t, but space is allocated for 64
127                 * bits. We need to respect the offset passed in. This interface
128                 * is so bad.
129                 */
130                mali_ptr addr = 0;
131                static_assert(sizeof(addr) == 8, "size out of sync");
132
133                memcpy(&addr, handles[i], sizeof(addr));
134                addr += rsrc->image.data.bo->ptr.gpu;
135
136                memcpy(handles[i], &addr, sizeof(addr));
137        }
138}
139
140static void
141panfrost_memory_barrier(struct pipe_context *pctx, unsigned flags)
142{
143        /* TODO: Be smart and only flush the minimum needed, maybe emitting a
144         * cache flush job if that would help */
145        panfrost_flush_all_batches(pan_context(pctx), "Memory barrier");
146}
147
148void
149panfrost_compute_context_init(struct pipe_context *pctx)
150{
151        pctx->create_compute_state = panfrost_create_compute_state;
152        pctx->bind_compute_state = panfrost_bind_compute_state;
153        pctx->delete_compute_state = panfrost_delete_compute_state;
154
155        pctx->set_compute_resources = panfrost_set_compute_resources;
156        pctx->set_global_binding = panfrost_set_global_binding;
157
158        pctx->memory_barrier = panfrost_memory_barrier;
159}
160