From 0500c9e5145e651566b0204c3b62984919ba2450 Mon Sep 17 00:00:00 2001 From: Boris Brezillon Date: Sat, 14 Sep 2019 08:00:27 +0200 Subject: [PATCH] panfrost: Get rid of pan_drm.c pan_drm.c was only meaningful when we were supporting 2 kernel drivers (mali_kbase, and the drm one). Now that there's now kernel-driver abstraction we're better off moving those functions were they belong: * BO related functions in pan_bo.c * fence related functions + query_gpu_version() in pan_screen.c * submit related functions in pan_job.c While at it, we rename the functions according to the place they're being moved to. Signed-off-by: Boris Brezillon Reviewed-by: Alyssa Rosenzweig --- src/gallium/drivers/panfrost/meson.build | 1 - src/gallium/drivers/panfrost/pan_allocate.c | 2 +- src/gallium/drivers/panfrost/pan_assemble.c | 2 +- src/gallium/drivers/panfrost/pan_blend_cso.c | 2 +- src/gallium/drivers/panfrost/pan_bo.c | 199 +++++++++++++- src/gallium/drivers/panfrost/pan_context.c | 18 +- src/gallium/drivers/panfrost/pan_context.h | 3 - src/gallium/drivers/panfrost/pan_drm.c | 395 --------------------------- src/gallium/drivers/panfrost/pan_job.c | 82 +++++- src/gallium/drivers/panfrost/pan_resource.c | 14 +- src/gallium/drivers/panfrost/pan_screen.c | 78 +++++- src/gallium/drivers/panfrost/pan_screen.h | 30 +- 12 files changed, 382 insertions(+), 444 deletions(-) delete mode 100644 src/gallium/drivers/panfrost/pan_drm.c diff --git a/src/gallium/drivers/panfrost/meson.build b/src/gallium/drivers/panfrost/meson.build index 73c3b54..ca9f6b7 100644 --- a/src/gallium/drivers/panfrost/meson.build +++ b/src/gallium/drivers/panfrost/meson.build @@ -35,7 +35,6 @@ files_panfrost = files( 'pan_bo.c', 'pan_blit.c', 'pan_job.c', - 'pan_drm.c', 'pan_allocate.c', 'pan_assemble.c', 'pan_format.c', diff --git a/src/gallium/drivers/panfrost/pan_allocate.c b/src/gallium/drivers/panfrost/pan_allocate.c index 44af631..e7970c1 100644 --- a/src/gallium/drivers/panfrost/pan_allocate.c +++ b/src/gallium/drivers/panfrost/pan_allocate.c @@ -65,7 +65,7 @@ panfrost_allocate_transient(struct panfrost_batch *batch, size_t sz) TRANSIENT_SLAB_SIZE : ALIGN_POT(sz, 4096); /* We can't reuse the current BO, but we can create a new one. */ - bo = panfrost_drm_create_bo(screen, bo_sz, 0); + bo = panfrost_bo_create(screen, bo_sz, 0); panfrost_batch_add_bo(batch, bo); /* Creating a BO adds a reference, and then the job adds a diff --git a/src/gallium/drivers/panfrost/pan_assemble.c b/src/gallium/drivers/panfrost/pan_assemble.c index b57cd5e..de73cf8 100644 --- a/src/gallium/drivers/panfrost/pan_assemble.c +++ b/src/gallium/drivers/panfrost/pan_assemble.c @@ -82,7 +82,7 @@ panfrost_shader_compile( * I bet someone just thought that would be a cute pun. At least, * that's how I'd do it. */ - state->bo = panfrost_drm_create_bo(screen, size, PAN_ALLOCATE_EXECUTE); + state->bo = panfrost_bo_create(screen, size, PAN_ALLOCATE_EXECUTE); memcpy(state->bo->cpu, dst, size); meta->shader = state->bo->gpu | program.first_tag; diff --git a/src/gallium/drivers/panfrost/pan_blend_cso.c b/src/gallium/drivers/panfrost/pan_blend_cso.c index ab49772..82527a5 100644 --- a/src/gallium/drivers/panfrost/pan_blend_cso.c +++ b/src/gallium/drivers/panfrost/pan_blend_cso.c @@ -272,7 +272,7 @@ panfrost_get_blend_for_context(struct panfrost_context *ctx, unsigned rti) final.shader.first_tag = shader->first_tag; /* Upload the shader */ - final.shader.bo = panfrost_drm_create_bo(screen, shader->size, PAN_ALLOCATE_EXECUTE); + final.shader.bo = panfrost_bo_create(screen, shader->size, PAN_ALLOCATE_EXECUTE); memcpy(final.shader.bo->cpu, shader->buffer, shader->size); /* Pass BO ownership to job */ diff --git a/src/gallium/drivers/panfrost/pan_bo.c b/src/gallium/drivers/panfrost/pan_bo.c index f2f4943..22476f0 100644 --- a/src/gallium/drivers/panfrost/pan_bo.c +++ b/src/gallium/drivers/panfrost/pan_bo.c @@ -23,11 +23,19 @@ * Authors (Collabora): * Alyssa Rosenzweig */ +#include +#include #include #include #include "drm-uapi/panfrost_drm.h" #include "pan_screen.h" +#include "pan_util.h" +#include "pandecode/decode.h" + +#include "os/os_mman.h" + +#include "util/u_inlines.h" #include "util/u_math.h" /* This file implements a userspace BO cache. Allocating and freeing @@ -105,7 +113,7 @@ panfrost_bo_cache_fetch( ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MADVISE, &madv); if (!ret && !madv.retained) { - panfrost_drm_release_bo(screen, entry, false); + panfrost_bo_release(screen, entry, false); continue; } /* Let's go! */ @@ -159,9 +167,196 @@ panfrost_bo_cache_evict_all( list_for_each_entry_safe(struct panfrost_bo, entry, bucket, link) { list_del(&entry->link); - panfrost_drm_release_bo(screen, entry, false); + panfrost_bo_release(screen, entry, false); } } pthread_mutex_unlock(&screen->bo_cache_lock); } +void +panfrost_bo_mmap(struct panfrost_screen *screen, struct panfrost_bo *bo) +{ + struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle }; + int ret; + + if (bo->cpu) + return; + + ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo); + if (ret) { + fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n"); + assert(0); + } + + bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED, + screen->fd, mmap_bo.offset); + if (bo->cpu == MAP_FAILED) { + fprintf(stderr, "mmap failed: %p %m\n", bo->cpu); + assert(0); + } + + /* Record the mmap if we're tracing */ + if (pan_debug & PAN_DBG_TRACE) + pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL); +} + +static void +panfrost_bo_munmap(struct panfrost_screen *screen, struct panfrost_bo *bo) +{ + if (!bo->cpu) + return; + + if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) { + perror("munmap"); + abort(); + } + + bo->cpu = NULL; +} + +struct panfrost_bo * +panfrost_bo_create(struct panfrost_screen *screen, size_t size, + uint32_t flags) +{ + struct panfrost_bo *bo; + + /* Kernel will fail (confusingly) with EPERM otherwise */ + assert(size > 0); + + /* To maximize BO cache usage, don't allocate tiny BOs */ + size = MAX2(size, 4096); + + /* GROWABLE BOs cannot be mmapped */ + if (flags & PAN_ALLOCATE_GROWABLE) + assert(flags & PAN_ALLOCATE_INVISIBLE); + + unsigned translated_flags = 0; + + if (screen->kernel_version->version_major > 1 || + screen->kernel_version->version_minor >= 1) { + if (flags & PAN_ALLOCATE_GROWABLE) + translated_flags |= PANFROST_BO_HEAP; + if (!(flags & PAN_ALLOCATE_EXECUTE)) + translated_flags |= PANFROST_BO_NOEXEC; + } + + struct drm_panfrost_create_bo create_bo = { + .size = size, + .flags = translated_flags, + }; + + /* Before creating a BO, we first want to check the cache */ + + bo = panfrost_bo_cache_fetch(screen, size, flags); + + if (bo == NULL) { + /* Otherwise, the cache misses and we need to allocate a BO fresh from + * the kernel */ + + int ret; + + ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo); + if (ret) { + fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n"); + assert(0); + } + + /* We have a BO allocated from the kernel; fill in the userspace + * version */ + + bo = rzalloc(screen, struct panfrost_bo); + bo->size = create_bo.size; + bo->gpu = create_bo.offset; + bo->gem_handle = create_bo.handle; + bo->flags = flags; + } + + /* Only mmap now if we know we need to. For CPU-invisible buffers, we + * never map since we don't care about their contents; they're purely + * for GPU-internal use. But we do trace them anyway. */ + + if (!(flags & (PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_DELAY_MMAP))) + panfrost_bo_mmap(screen, bo); + else if (flags & PAN_ALLOCATE_INVISIBLE) { + if (pan_debug & PAN_DBG_TRACE) + pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL); + } + + pipe_reference_init(&bo->reference, 1); + return bo; +} + +void +panfrost_bo_release(struct panfrost_screen *screen, struct panfrost_bo *bo, + bool cacheable) +{ + if (!bo) + return; + + struct drm_gem_close gem_close = { .handle = bo->gem_handle }; + int ret; + + /* Rather than freeing the BO now, we'll cache the BO for later + * allocations if we're allowed to */ + + panfrost_bo_munmap(screen, bo); + + if (cacheable) { + bool cached = panfrost_bo_cache_put(screen, bo); + + if (cached) + return; + } + + /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */ + + ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close); + if (ret) { + fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n"); + assert(0); + } + + ralloc_free(bo); +} + +struct panfrost_bo * +panfrost_bo_import(struct panfrost_screen *screen, int fd) +{ + struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo); + struct drm_panfrost_get_bo_offset get_bo_offset = {0,}; + ASSERTED int ret; + unsigned gem_handle; + + ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle); + assert(!ret); + + get_bo_offset.handle = gem_handle; + ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset); + assert(!ret); + + bo->gem_handle = gem_handle; + bo->gpu = (mali_ptr) get_bo_offset.offset; + bo->size = lseek(fd, 0, SEEK_END); + assert(bo->size > 0); + pipe_reference_init(&bo->reference, 1); + + // TODO map and unmap on demand? + panfrost_bo_mmap(screen, bo); + return bo; +} + +int +panfrost_bo_export(struct panfrost_screen *screen, const struct panfrost_bo *bo) +{ + struct drm_prime_handle args = { + .handle = bo->gem_handle, + .flags = DRM_CLOEXEC, + }; + + int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args); + if (ret == -1) + return -1; + + return args.fd; +} + diff --git a/src/gallium/drivers/panfrost/pan_context.c b/src/gallium/drivers/panfrost/pan_context.c index 323a480..cadb462 100644 --- a/src/gallium/drivers/panfrost/pan_context.c +++ b/src/gallium/drivers/panfrost/pan_context.c @@ -2425,9 +2425,9 @@ panfrost_destroy(struct pipe_context *pipe) if (panfrost->blitter_wallpaper) util_blitter_destroy(panfrost->blitter_wallpaper); - panfrost_drm_release_bo(screen, panfrost->scratchpad, false); - panfrost_drm_release_bo(screen, panfrost->tiler_heap, false); - panfrost_drm_release_bo(screen, panfrost->tiler_dummy, false); + panfrost_bo_release(screen, panfrost->scratchpad, false); + panfrost_bo_release(screen, panfrost->tiler_heap, false); + panfrost_bo_release(screen, panfrost->tiler_dummy, false); ralloc_free(pipe); } @@ -2611,11 +2611,11 @@ panfrost_setup_hardware(struct panfrost_context *ctx) struct pipe_context *gallium = (struct pipe_context *) ctx; struct panfrost_screen *screen = pan_screen(gallium->screen); - ctx->scratchpad = panfrost_drm_create_bo(screen, 64 * 4 * 4096, 0); - ctx->tiler_heap = panfrost_drm_create_bo(screen, 4096 * 4096, + ctx->scratchpad = panfrost_bo_create(screen, 64 * 4 * 4096, 0); + ctx->tiler_heap = panfrost_bo_create(screen, 4096 * 4096, PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_GROWABLE); - ctx->tiler_dummy = panfrost_drm_create_bo(screen, 4096, + ctx->tiler_dummy = panfrost_bo_create(screen, 4096, PAN_ALLOCATE_INVISIBLE); assert(ctx->scratchpad && ctx->tiler_heap && ctx->tiler_dummy); } @@ -2698,7 +2698,11 @@ panfrost_create_context(struct pipe_screen *screen, void *priv, unsigned flags) panfrost_blend_context_init(gallium); panfrost_compute_context_init(gallium); - panfrost_drm_init_context(ctx); + ASSERTED int ret; + + ret = drmSyncobjCreate(pscreen->fd, DRM_SYNCOBJ_CREATE_SIGNALED, + &ctx->out_sync); + assert(!ret); panfrost_setup_hardware(ctx); diff --git a/src/gallium/drivers/panfrost/pan_context.h b/src/gallium/drivers/panfrost/pan_context.h index f0578d6..4f5c137 100644 --- a/src/gallium/drivers/panfrost/pan_context.h +++ b/src/gallium/drivers/panfrost/pan_context.h @@ -291,9 +291,6 @@ pan_context(struct pipe_context *pcontext) return (struct panfrost_context *) pcontext; } -struct panfrost_fence * -panfrost_fence_create(struct panfrost_context *ctx); - struct pipe_context * panfrost_create_context(struct pipe_screen *screen, void *priv, unsigned flags); diff --git a/src/gallium/drivers/panfrost/pan_drm.c b/src/gallium/drivers/panfrost/pan_drm.c deleted file mode 100644 index c41701e..0000000 --- a/src/gallium/drivers/panfrost/pan_drm.c +++ /dev/null @@ -1,395 +0,0 @@ -/* - * © Copyright 2019 Collabora, Ltd. - * Copyright 2019 Alyssa Rosenzweig - * - * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the "Software"), - * to deal in the Software without restriction, including without limitation - * the rights to use, copy, modify, merge, publish, distribute, sublicense, - * and/or sell copies of the Software, and to permit persons to whom the - * Software is furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice (including the next - * paragraph) shall be included in all copies or substantial portions of the - * Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL - * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - * - */ - -#include -#include - -#include "drm-uapi/panfrost_drm.h" - -#include "util/u_memory.h" -#include "util/os_time.h" -#include "os/os_mman.h" - -#include "pan_screen.h" -#include "pan_resource.h" -#include "pan_context.h" -#include "pan_util.h" -#include "pandecode/decode.h" - -void -panfrost_drm_mmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo) -{ - struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle }; - int ret; - - if (bo->cpu) - return; - - ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo); - if (ret) { - fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n"); - assert(0); - } - - bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED, - screen->fd, mmap_bo.offset); - if (bo->cpu == MAP_FAILED) { - fprintf(stderr, "mmap failed: %p %m\n", bo->cpu); - assert(0); - } - - /* Record the mmap if we're tracing */ - if (pan_debug & PAN_DBG_TRACE) - pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL); -} - -static void -panfrost_drm_munmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo) -{ - if (!bo->cpu) - return; - - if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) { - perror("munmap"); - abort(); - } - - bo->cpu = NULL; -} - -struct panfrost_bo * -panfrost_drm_create_bo(struct panfrost_screen *screen, size_t size, - uint32_t flags) -{ - struct panfrost_bo *bo; - - /* Kernel will fail (confusingly) with EPERM otherwise */ - assert(size > 0); - - /* To maximize BO cache usage, don't allocate tiny BOs */ - size = MAX2(size, 4096); - - /* GROWABLE BOs cannot be mmapped */ - if (flags & PAN_ALLOCATE_GROWABLE) - assert(flags & PAN_ALLOCATE_INVISIBLE); - - unsigned translated_flags = 0; - - if (screen->kernel_version->version_major > 1 || - screen->kernel_version->version_minor >= 1) { - if (flags & PAN_ALLOCATE_GROWABLE) - translated_flags |= PANFROST_BO_HEAP; - if (!(flags & PAN_ALLOCATE_EXECUTE)) - translated_flags |= PANFROST_BO_NOEXEC; - } - - struct drm_panfrost_create_bo create_bo = { - .size = size, - .flags = translated_flags, - }; - - /* Before creating a BO, we first want to check the cache */ - - bo = panfrost_bo_cache_fetch(screen, size, flags); - - if (bo == NULL) { - /* Otherwise, the cache misses and we need to allocate a BO fresh from - * the kernel */ - - int ret; - - ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo); - if (ret) { - fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n"); - assert(0); - } - - /* We have a BO allocated from the kernel; fill in the userspace - * version */ - - bo = rzalloc(screen, struct panfrost_bo); - bo->size = create_bo.size; - bo->gpu = create_bo.offset; - bo->gem_handle = create_bo.handle; - bo->flags = flags; - } - - /* Only mmap now if we know we need to. For CPU-invisible buffers, we - * never map since we don't care about their contents; they're purely - * for GPU-internal use. But we do trace them anyway. */ - - if (!(flags & (PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_DELAY_MMAP))) - panfrost_drm_mmap_bo(screen, bo); - else if (flags & PAN_ALLOCATE_INVISIBLE) { - if (pan_debug & PAN_DBG_TRACE) - pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL); - } - - pipe_reference_init(&bo->reference, 1); - return bo; -} - -void -panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable) -{ - if (!bo) - return; - - struct drm_gem_close gem_close = { .handle = bo->gem_handle }; - int ret; - - /* Rather than freeing the BO now, we'll cache the BO for later - * allocations if we're allowed to */ - - panfrost_drm_munmap_bo(screen, bo); - - if (cacheable) { - bool cached = panfrost_bo_cache_put(screen, bo); - - if (cached) - return; - } - - /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */ - - ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close); - if (ret) { - fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n"); - assert(0); - } - - ralloc_free(bo); -} - -struct panfrost_bo * -panfrost_drm_import_bo(struct panfrost_screen *screen, int fd) -{ - struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo); - struct drm_panfrost_get_bo_offset get_bo_offset = {0,}; - ASSERTED int ret; - unsigned gem_handle; - - ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle); - assert(!ret); - - get_bo_offset.handle = gem_handle; - ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset); - assert(!ret); - - bo->gem_handle = gem_handle; - bo->gpu = (mali_ptr) get_bo_offset.offset; - bo->size = lseek(fd, 0, SEEK_END); - assert(bo->size > 0); - pipe_reference_init(&bo->reference, 1); - - // TODO map and unmap on demand? - panfrost_drm_mmap_bo(screen, bo); - return bo; -} - -int -panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo) -{ - struct drm_prime_handle args = { - .handle = bo->gem_handle, - .flags = DRM_CLOEXEC, - }; - - int ret = drmIoctl(screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args); - if (ret == -1) - return -1; - - return args.fd; -} - -static int -panfrost_drm_submit_batch(struct panfrost_batch *batch, u64 first_job_desc, - int reqs) -{ - struct panfrost_context *ctx = batch->ctx; - struct pipe_context *gallium = (struct pipe_context *) ctx; - struct panfrost_screen *screen = pan_screen(gallium->screen); - struct drm_panfrost_submit submit = {0,}; - uint32_t *bo_handles; - int ret; - - submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync; - submit.in_sync_count = 1; - - submit.out_sync = ctx->out_sync; - - submit.jc = first_job_desc; - submit.requirements = reqs; - - bo_handles = calloc(batch->bos->entries, sizeof(*bo_handles)); - assert(bo_handles); - - set_foreach(batch->bos, entry) { - struct panfrost_bo *bo = (struct panfrost_bo *)entry->key; - assert(bo->gem_handle > 0); - bo_handles[submit.bo_handle_count++] = bo->gem_handle; - } - - submit.bo_handles = (u64) (uintptr_t) bo_handles; - ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit); - free(bo_handles); - if (ret) { - fprintf(stderr, "Error submitting: %m\n"); - return errno; - } - - /* Trace the job if we're doing that */ - if (pan_debug & PAN_DBG_TRACE) { - /* Wait so we can get errors reported back */ - drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL); - pandecode_jc(submit.jc, FALSE); - } - - return 0; -} - -int -panfrost_drm_submit_vs_fs_batch(struct panfrost_batch *batch) -{ - struct panfrost_context *ctx = batch->ctx; - bool has_draws = batch->last_job.gpu; - int ret = 0; - - panfrost_batch_add_bo(batch, ctx->scratchpad); - panfrost_batch_add_bo(batch, ctx->tiler_heap); - - if (batch->first_job.gpu) { - ret = panfrost_drm_submit_batch(batch, batch->first_job.gpu, 0); - assert(!ret); - } - - if (batch->first_tiler.gpu || batch->clear) { - ret = panfrost_drm_submit_batch(batch, - panfrost_fragment_job(batch, has_draws), - PANFROST_JD_REQ_FS); - assert(!ret); - } - - return ret; -} - -struct panfrost_fence * -panfrost_fence_create(struct panfrost_context *ctx) -{ - struct pipe_context *gallium = (struct pipe_context *) ctx; - struct panfrost_screen *screen = pan_screen(gallium->screen); - struct panfrost_fence *f = calloc(1, sizeof(*f)); - if (!f) - return NULL; - - /* Snapshot the last Panfrost's rendering's out fence. We'd rather have - * another syncobj instead of a sync file, but this is all we get. - * (HandleToFD/FDToHandle just gives you another syncobj ID for the - * same syncobj). - */ - drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd); - if (f->fd == -1) { - fprintf(stderr, "export failed: %m\n"); - free(f); - return NULL; - } - - pipe_reference_init(&f->reference, 1); - - return f; -} - -unsigned -panfrost_drm_query_gpu_version(struct panfrost_screen *screen) -{ - struct drm_panfrost_get_param get_param = {0,}; - ASSERTED int ret; - - get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID; - ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param); - assert(!ret); - - return get_param.value; -} - -int -panfrost_drm_init_context(struct panfrost_context *ctx) -{ - struct pipe_context *gallium = (struct pipe_context *) ctx; - struct panfrost_screen *screen = pan_screen(gallium->screen); - - return drmSyncobjCreate(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED, - &ctx->out_sync); -} - -void -panfrost_drm_fence_reference(struct pipe_screen *screen, - struct pipe_fence_handle **ptr, - struct pipe_fence_handle *fence) -{ - struct panfrost_fence **p = (struct panfrost_fence **)ptr; - struct panfrost_fence *f = (struct panfrost_fence *)fence; - struct panfrost_fence *old = *p; - - if (pipe_reference(&(*p)->reference, &f->reference)) { - close(old->fd); - free(old); - } - *p = f; -} - -boolean -panfrost_drm_fence_finish(struct pipe_screen *pscreen, - struct pipe_context *ctx, - struct pipe_fence_handle *fence, - uint64_t timeout) -{ - struct panfrost_screen *screen = pan_screen(pscreen); - struct panfrost_fence *f = (struct panfrost_fence *)fence; - int ret; - - unsigned syncobj; - ret = drmSyncobjCreate(screen->fd, 0, &syncobj); - if (ret) { - fprintf(stderr, "Failed to create syncobj to wait on: %m\n"); - return false; - } - - ret = drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd); - if (ret) { - fprintf(stderr, "Failed to import fence to syncobj: %m\n"); - return false; - } - - uint64_t abs_timeout = os_time_get_absolute_timeout(timeout); - if (abs_timeout == OS_TIMEOUT_INFINITE) - abs_timeout = INT64_MAX; - - ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL); - - drmSyncobjDestroy(screen->fd, syncobj); - - return ret >= 0; -} diff --git a/src/gallium/drivers/panfrost/pan_job.c b/src/gallium/drivers/panfrost/pan_job.c index ecaf081..7bd0629 100644 --- a/src/gallium/drivers/panfrost/pan_job.c +++ b/src/gallium/drivers/panfrost/pan_job.c @@ -25,11 +25,15 @@ #include +#include "drm-uapi/panfrost_drm.h" + #include "pan_context.h" #include "util/hash_table.h" #include "util/ralloc.h" #include "util/u_format.h" #include "util/u_pack_color.h" +#include "pan_util.h" +#include "pandecode/decode.h" static struct panfrost_batch * panfrost_create_batch(struct panfrost_context *ctx, @@ -153,8 +157,8 @@ panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size) /* Create the BO as invisible, as there's no reason to map */ - batch->polygon_list = panfrost_drm_create_bo(screen, - size, PAN_ALLOCATE_INVISIBLE); + batch->polygon_list = panfrost_bo_create(screen, size, + PAN_ALLOCATE_INVISIBLE); panfrost_batch_add_bo(batch, batch->polygon_list); /* A BO reference has been retained by panfrost_batch_add_bo(), @@ -267,6 +271,78 @@ panfrost_batch_draw_wallpaper(struct panfrost_batch *batch) batch->ctx->wallpaper_batch = NULL; } +static int +panfrost_batch_submit_ioctl(struct panfrost_batch *batch, + mali_ptr first_job_desc, + uint32_t reqs) +{ + struct panfrost_context *ctx = batch->ctx; + struct pipe_context *gallium = (struct pipe_context *) ctx; + struct panfrost_screen *screen = pan_screen(gallium->screen); + struct drm_panfrost_submit submit = {0,}; + uint32_t *bo_handles; + int ret; + + submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync; + submit.in_sync_count = 1; + + submit.out_sync = ctx->out_sync; + + submit.jc = first_job_desc; + submit.requirements = reqs; + + bo_handles = calloc(batch->bos->entries, sizeof(*bo_handles)); + assert(bo_handles); + + set_foreach(batch->bos, entry) { + struct panfrost_bo *bo = (struct panfrost_bo *)entry->key; + assert(bo->gem_handle > 0); + bo_handles[submit.bo_handle_count++] = bo->gem_handle; + } + + submit.bo_handles = (u64) (uintptr_t) bo_handles; + ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit); + free(bo_handles); + if (ret) { + fprintf(stderr, "Error submitting: %m\n"); + return errno; + } + + /* Trace the job if we're doing that */ + if (pan_debug & PAN_DBG_TRACE) { + /* Wait so we can get errors reported back */ + drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL); + pandecode_jc(submit.jc, FALSE); + } + + return 0; +} + +static int +panfrost_batch_submit_jobs(struct panfrost_batch *batch) +{ + struct panfrost_context *ctx = batch->ctx; + bool has_draws = batch->first_job.gpu; + int ret = 0; + + panfrost_batch_add_bo(batch, ctx->scratchpad); + panfrost_batch_add_bo(batch, ctx->tiler_heap); + + if (has_draws) { + ret = panfrost_batch_submit_ioctl(batch, batch->first_job.gpu, 0); + assert(!ret); + } + + if (batch->first_tiler.gpu || batch->clear) { + mali_ptr fragjob = panfrost_fragment_job(batch, has_draws); + + ret = panfrost_batch_submit_ioctl(batch, fragjob, PANFROST_JD_REQ_FS); + assert(!ret); + } + + return ret; +} + void panfrost_batch_submit(struct panfrost_batch *batch) { @@ -284,7 +360,7 @@ panfrost_batch_submit(struct panfrost_batch *batch) panfrost_scoreboard_link_batch(batch); - ret = panfrost_drm_submit_vs_fs_batch(batch); + ret = panfrost_batch_submit_jobs(batch); if (ret) fprintf(stderr, "panfrost_batch_submit failed: %d\n", ret); diff --git a/src/gallium/drivers/panfrost/pan_resource.c b/src/gallium/drivers/panfrost/pan_resource.c index 0be6b2b..47a44bf 100644 --- a/src/gallium/drivers/panfrost/pan_resource.c +++ b/src/gallium/drivers/panfrost/pan_resource.c @@ -81,7 +81,7 @@ panfrost_resource_from_handle(struct pipe_screen *pscreen, pipe_reference_init(&prsc->reference, 1); prsc->screen = pscreen; - rsc->bo = panfrost_drm_import_bo(screen, whandle->handle); + rsc->bo = panfrost_bo_import(screen, whandle->handle); rsc->slices[0].stride = whandle->stride; rsc->slices[0].initialized = true; panfrost_resource_reset_damage(rsc); @@ -133,7 +133,7 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen, return true; } else { - int fd = panfrost_drm_export_bo(screen, rsrc->bo); + int fd = panfrost_bo_export(screen, rsrc->bo); if (fd < 0) return false; @@ -412,7 +412,7 @@ panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_reso /* We create a BO immediately but don't bother mapping, since we don't * care to map e.g. FBOs which the CPU probably won't touch */ - pres->bo = panfrost_drm_create_bo(screen, bo_size, PAN_ALLOCATE_DELAY_MMAP); + pres->bo = panfrost_bo_create(screen, bo_size, PAN_ALLOCATE_DELAY_MMAP); } void @@ -537,7 +537,7 @@ panfrost_bo_unreference(struct pipe_screen *screen, struct panfrost_bo *bo) /* When the reference count goes to zero, we need to cleanup */ if (pipe_reference(&bo->reference, NULL)) - panfrost_drm_release_bo(pan_screen(screen), bo, true); + panfrost_bo_release(pan_screen(screen), bo, true); } static void @@ -582,7 +582,7 @@ panfrost_transfer_map(struct pipe_context *pctx, if (!bo->cpu) { struct panfrost_screen *screen = pan_screen(pctx->screen); - panfrost_drm_mmap_bo(screen, bo); + panfrost_bo_mmap(screen, bo); } /* Check if we're bound for rendering and this is a read pixels. If so, @@ -861,8 +861,8 @@ panfrost_resource_hint_layout( /* If we grew in size, reallocate the BO */ if (new_size > rsrc->bo->size) { - panfrost_drm_release_bo(screen, rsrc->bo, true); - rsrc->bo = panfrost_drm_create_bo(screen, new_size, PAN_ALLOCATE_DELAY_MMAP); + panfrost_bo_release(screen, rsrc->bo, true); + rsrc->bo = panfrost_bo_create(screen, new_size, PAN_ALLOCATE_DELAY_MMAP); } } diff --git a/src/gallium/drivers/panfrost/pan_screen.c b/src/gallium/drivers/panfrost/pan_screen.c index 5dcceac..54ef2ef 100644 --- a/src/gallium/drivers/panfrost/pan_screen.c +++ b/src/gallium/drivers/panfrost/pan_screen.c @@ -41,6 +41,7 @@ #include #include "drm-uapi/drm_fourcc.h" +#include "drm-uapi/panfrost_drm.h" #include "pan_screen.h" #include "pan_resource.h" @@ -568,7 +569,15 @@ panfrost_fence_reference(struct pipe_screen *pscreen, struct pipe_fence_handle **ptr, struct pipe_fence_handle *fence) { - panfrost_drm_fence_reference(pscreen, ptr, fence); + struct panfrost_fence **p = (struct panfrost_fence **)ptr; + struct panfrost_fence *f = (struct panfrost_fence *)fence; + struct panfrost_fence *old = *p; + + if (pipe_reference(&(*p)->reference, &f->reference)) { + close(old->fd); + free(old); + } + *p = f; } static bool @@ -577,7 +586,57 @@ panfrost_fence_finish(struct pipe_screen *pscreen, struct pipe_fence_handle *fence, uint64_t timeout) { - return panfrost_drm_fence_finish(pscreen, ctx, fence, timeout); + struct panfrost_screen *screen = pan_screen(pscreen); + struct panfrost_fence *f = (struct panfrost_fence *)fence; + int ret; + + unsigned syncobj; + ret = drmSyncobjCreate(screen->fd, 0, &syncobj); + if (ret) { + fprintf(stderr, "Failed to create syncobj to wait on: %m\n"); + return false; + } + + ret = drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd); + if (ret) { + fprintf(stderr, "Failed to import fence to syncobj: %m\n"); + return false; + } + + uint64_t abs_timeout = os_time_get_absolute_timeout(timeout); + if (abs_timeout == OS_TIMEOUT_INFINITE) + abs_timeout = INT64_MAX; + + ret = drmSyncobjWait(screen->fd, &syncobj, 1, abs_timeout, 0, NULL); + + drmSyncobjDestroy(screen->fd, syncobj); + + return ret >= 0; +} + +struct panfrost_fence * +panfrost_fence_create(struct panfrost_context *ctx) +{ + struct panfrost_screen *screen = pan_screen(ctx->base.screen); + struct panfrost_fence *f = calloc(1, sizeof(*f)); + if (!f) + return NULL; + + /* Snapshot the last Panfrost's rendering's out fence. We'd rather have + * another syncobj instead of a sync file, but this is all we get. + * (HandleToFD/FDToHandle just gives you another syncobj ID for the + * same syncobj). + */ + drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd); + if (f->fd == -1) { + fprintf(stderr, "export failed: %m\n"); + free(f); + return NULL; + } + + pipe_reference_init(&f->reference, 1); + + return f; } static const void * @@ -588,6 +647,19 @@ panfrost_screen_get_compiler_options(struct pipe_screen *pscreen, return &midgard_nir_options; } +static unsigned +panfrost_query_gpu_version(struct panfrost_screen *screen) +{ + struct drm_panfrost_get_param get_param = {0,}; + ASSERTED int ret; + + get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID; + ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param); + assert(!ret); + + return get_param.value; +} + struct pipe_screen * panfrost_create_screen(int fd, struct renderonly *ro) { @@ -622,7 +694,7 @@ panfrost_create_screen(int fd, struct renderonly *ro) screen->fd = fd; - screen->gpu_id = panfrost_drm_query_gpu_version(screen); + screen->gpu_id = panfrost_query_gpu_version(screen); screen->require_sfbd = screen->gpu_id < 0x0750; /* T760 is the first to support MFBD */ screen->kernel_version = drmGetVersion(fd); diff --git a/src/gallium/drivers/panfrost/pan_screen.h b/src/gallium/drivers/panfrost/pan_screen.h index 2f17e2d..2f0ebd4 100644 --- a/src/gallium/drivers/panfrost/pan_screen.h +++ b/src/gallium/drivers/panfrost/pan_screen.h @@ -120,32 +120,22 @@ pan_screen(struct pipe_screen *p) return (struct panfrost_screen *)p; } +struct panfrost_fence * +panfrost_fence_create(struct panfrost_context *ctx); + struct panfrost_bo * -panfrost_drm_create_bo(struct panfrost_screen *screen, size_t size, +panfrost_bo_create(struct panfrost_screen *screen, size_t size, uint32_t flags); void -panfrost_drm_mmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo); +panfrost_bo_mmap(struct panfrost_screen *screen, struct panfrost_bo *bo); void -panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable); +panfrost_bo_release(struct panfrost_screen *screen, struct panfrost_bo *bo, + bool cacheable); struct panfrost_bo * -panfrost_drm_import_bo(struct panfrost_screen *screen, int fd); -int -panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo); -int -panfrost_drm_submit_vs_fs_batch(struct panfrost_batch *batch); -unsigned -panfrost_drm_query_gpu_version(struct panfrost_screen *screen); +panfrost_bo_import(struct panfrost_screen *screen, int fd); int -panfrost_drm_init_context(struct panfrost_context *ctx); -void -panfrost_drm_fence_reference(struct pipe_screen *screen, - struct pipe_fence_handle **ptr, - struct pipe_fence_handle *fence); -boolean -panfrost_drm_fence_finish(struct pipe_screen *pscreen, - struct pipe_context *ctx, - struct pipe_fence_handle *fence, - uint64_t timeout); +panfrost_bo_export(struct panfrost_screen *screen, const struct panfrost_bo *bo); + struct panfrost_bo * panfrost_bo_cache_fetch( struct panfrost_screen *screen, -- 2.7.4