X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fpanfrost%2Fpan_cmdstream.c;h=bc9158b03af6b026b79709eac2de53cef317bc8e;hb=cfef6de429d40785a306ef2a2db7092c4ce0388a;hp=457c02cbbc13653b6c02613bb70d4d4157e962ee;hpb=e9dc22512202f793f89d6cb86a780f81c2b88cd5;p=mesa.git diff --git a/src/gallium/drivers/panfrost/pan_cmdstream.c b/src/gallium/drivers/panfrost/pan_cmdstream.c index 457c02cbbc1..bc9158b03af 100644 --- a/src/gallium/drivers/panfrost/pan_cmdstream.c +++ b/src/gallium/drivers/panfrost/pan_cmdstream.c @@ -58,13 +58,21 @@ panfrost_vt_emit_shared_memory(struct panfrost_context *ctx, struct panfrost_device *dev = pan_device(ctx->base.screen); struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx); - unsigned shift = panfrost_get_stack_shift(batch->stack_size); struct mali_shared_memory shared = { - .stack_shift = shift, - .scratchpad = panfrost_batch_get_scratchpad(batch, shift, dev->thread_tls_alloc, dev->core_count)->gpu, .shared_workgroup_count = ~0, }; - postfix->shared_memory = panfrost_pool_upload(&batch->pool, &shared, sizeof(shared)); + + if (batch->stack_size) { + struct panfrost_bo *stack = + panfrost_batch_get_scratchpad(batch, batch->stack_size, + dev->thread_tls_alloc, + dev->core_count); + + shared.stack_shift = panfrost_get_stack_shift(batch->stack_size); + shared.scratchpad = stack->gpu; + } + + postfix->shared_memory = panfrost_pool_upload_aligned(&batch->pool, &shared, sizeof(shared), 64); } static void @@ -76,21 +84,19 @@ panfrost_vt_attach_framebuffer(struct panfrost_context *ctx, } static void -panfrost_vt_update_rasterizer(struct panfrost_context *ctx, +panfrost_vt_update_rasterizer(struct panfrost_rasterizer *rasterizer, struct mali_vertex_tiler_prefix *prefix, struct mali_vertex_tiler_postfix *postfix) { - struct panfrost_rasterizer *rasterizer = ctx->rasterizer; - postfix->gl_enables |= 0x7; SET_BIT(postfix->gl_enables, MALI_FRONT_CCW_TOP, - rasterizer && rasterizer->base.front_ccw); + rasterizer->base.front_ccw); SET_BIT(postfix->gl_enables, MALI_CULL_FACE_FRONT, - rasterizer && (rasterizer->base.cull_face & PIPE_FACE_FRONT)); + (rasterizer->base.cull_face & PIPE_FACE_FRONT)); SET_BIT(postfix->gl_enables, MALI_CULL_FACE_BACK, - rasterizer && (rasterizer->base.cull_face & PIPE_FACE_BACK)); + (rasterizer->base.cull_face & PIPE_FACE_BACK)); SET_BIT(prefix->unknown_draw, MALI_DRAW_FLATSHADE_FIRST, - rasterizer && rasterizer->base.flatshade_first); + rasterizer->base.flatshade_first); } void @@ -101,11 +107,7 @@ panfrost_vt_update_primitive_size(struct panfrost_context *ctx, struct panfrost_rasterizer *rasterizer = ctx->rasterizer; if (!panfrost_writes_point_size(ctx)) { - bool points = prefix->draw_mode == MALI_DRAW_MODE_POINTS; - float val = 0.0f; - - if (rasterizer) - val = points ? + float val = (prefix->draw_mode == MALI_DRAW_MODE_POINTS) ? rasterizer->base.point_size : rasterizer->base.line_width; @@ -153,7 +155,7 @@ panfrost_vt_init(struct panfrost_context *ctx, if (stage == PIPE_SHADER_FRAGMENT) { panfrost_vt_update_occlusion_query(ctx, postfix); - panfrost_vt_update_rasterizer(ctx, prefix, postfix); + panfrost_vt_update_rasterizer(ctx->rasterizer, prefix, postfix); } } @@ -215,9 +217,13 @@ panfrost_get_index_buffer_bounded(struct panfrost_context *ctx, } else { /* Otherwise, we need to upload to transient memory */ const uint8_t *ibuf8 = (const uint8_t *) info->index.user; - out = panfrost_pool_upload(&batch->pool, ibuf8 + offset, - info->count * - info->index_size); + struct panfrost_transfer T = + panfrost_pool_alloc_aligned(&batch->pool, + info->count * info->index_size, + info->index_size); + + memcpy(T.cpu, ibuf8 + offset, info->count * info->index_size); + out = T.gpu; } if (needs_indices) { @@ -345,7 +351,7 @@ panfrost_shader_meta_init(struct panfrost_context *ctx, meta->midgard1.flags_lo = 0x20; meta->midgard1.uniform_buffer_count = panfrost_ubo_count(ctx, st); - SET_BIT(meta->midgard1.flags_hi, MALI_WRITES_GLOBAL, ss->writes_global); + SET_BIT(meta->midgard1.flags_lo, MALI_WRITES_GLOBAL, ss->writes_global); } } @@ -452,18 +458,6 @@ static void panfrost_frag_meta_rasterizer_update(struct panfrost_context *ctx, struct mali_shader_meta *fragmeta) { - if (!ctx->rasterizer) { - SET_BIT(fragmeta->unknown2_4, MALI_NO_MSAA, true); - SET_BIT(fragmeta->unknown2_3, MALI_HAS_MSAA, false); - fragmeta->depth_units = 0.0f; - fragmeta->depth_factor = 0.0f; - SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_A, false); - SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_B, false); - SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_CLIP_NEAR, true); - SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_CLIP_FAR, true); - return; - } - struct pipe_rasterizer_state *rast = &ctx->rasterizer->base; bool msaa = rast->multisample; @@ -497,40 +491,29 @@ panfrost_frag_meta_zsa_update(struct panfrost_context *ctx, struct mali_shader_meta *fragmeta) { const struct panfrost_zsa_state *so = ctx->depth_stencil; - int zfunc = PIPE_FUNC_ALWAYS; - - if (!so) { - /* If stenciling is disabled, the state is irrelevant */ - SET_BIT(fragmeta->unknown2_4, MALI_STENCIL_TEST, false); - SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_WRITEMASK, false); - } else { - SET_BIT(fragmeta->unknown2_4, MALI_STENCIL_TEST, - so->base.stencil[0].enabled); - fragmeta->stencil_mask_front = so->stencil_mask_front; - fragmeta->stencil_mask_back = so->stencil_mask_back; + SET_BIT(fragmeta->unknown2_4, MALI_STENCIL_TEST, + so->base.stencil[0].enabled); - /* Bottom bits for stencil ref, exactly one word */ - fragmeta->stencil_front.opaque[0] = so->stencil_front.opaque[0] | ctx->stencil_ref.ref_value[0]; + fragmeta->stencil_mask_front = so->stencil_mask_front; + fragmeta->stencil_mask_back = so->stencil_mask_back; - /* If back-stencil is not enabled, use the front values */ - - if (so->base.stencil[1].enabled) - fragmeta->stencil_back.opaque[0] = so->stencil_back.opaque[0] | ctx->stencil_ref.ref_value[1]; - else - fragmeta->stencil_back = fragmeta->stencil_front; + /* Bottom bits for stencil ref, exactly one word */ + fragmeta->stencil_front.opaque[0] = so->stencil_front.opaque[0] | ctx->stencil_ref.ref_value[0]; - if (so->base.depth.enabled) - zfunc = so->base.depth.func; + /* If back-stencil is not enabled, use the front values */ - /* Depth state (TODO: Refactor) */ + if (so->base.stencil[1].enabled) + fragmeta->stencil_back.opaque[0] = so->stencil_back.opaque[0] | ctx->stencil_ref.ref_value[1]; + else + fragmeta->stencil_back = fragmeta->stencil_front; - SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_WRITEMASK, - so->base.depth.writemask); - } + SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_WRITEMASK, + so->base.depth.writemask); fragmeta->unknown2_3 &= ~MALI_DEPTH_FUNC_MASK; - fragmeta->unknown2_3 |= MALI_DEPTH_FUNC(panfrost_translate_compare_func(zfunc)); + fragmeta->unknown2_3 |= MALI_DEPTH_FUNC(panfrost_translate_compare_func( + so->base.depth.enabled ? so->base.depth.func : PIPE_FUNC_ALWAYS)); } static bool @@ -557,7 +540,8 @@ panfrost_fs_required( static void panfrost_frag_meta_blend_update(struct panfrost_context *ctx, struct mali_shader_meta *fragmeta, - void *rts) + void *rts, + struct panfrost_blend_final *blend) { struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx); const struct panfrost_device *dev = pan_device(ctx->base.screen); @@ -572,15 +556,7 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, ctx->blend->base.alpha_to_coverage); /* Get blending setup */ - unsigned rt_count = MAX2(ctx->pipe_framebuffer.nr_cbufs, 1); - - struct panfrost_blend_final blend[PIPE_MAX_COLOR_BUFS]; - unsigned shader_offset = 0; - struct panfrost_bo *shader_bo = NULL; - - for (unsigned c = 0; c < rt_count; ++c) - blend[c] = panfrost_get_blend_for_context(ctx, c, &shader_bo, - &shader_offset); + unsigned rt_count = ctx->pipe_framebuffer.nr_cbufs; /* Disable shader execution if we can */ if (dev->quirks & MIDGARD_SHADERLESS @@ -616,7 +592,7 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, fragmeta->blend.shader = 0; - for (signed rt = (rt_count - 1); rt >= 0; --rt) { + for (signed rt = ((signed) rt_count - 1); rt >= 0; --rt) { if (!blend[rt].is_shader) continue; @@ -658,16 +634,22 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, /* Additional blend descriptor tacked on for jobs using MFBD */ + struct bifrost_blend_rt *brts = rts; + struct midgard_blend_rt *mrts = rts; + + /* Disable blending for depth-only on Bifrost */ + + if (rt_count == 0 && dev->quirks & IS_BIFROST) + brts[0].unk2 = 0x3; + for (unsigned i = 0; i < rt_count; ++i) { unsigned flags = 0; - if (ctx->pipe_framebuffer.nr_cbufs > i && !blend[i].no_colour) { + if (!blend[i].no_colour) { flags = 0x200; batch->draws |= (PIPE_CLEAR_COLOR0 << i); - bool is_srgb = (ctx->pipe_framebuffer.nr_cbufs > i) && - (ctx->pipe_framebuffer.cbufs[i]) && - util_format_is_srgb(ctx->pipe_framebuffer.cbufs[i]->format); + bool is_srgb = util_format_is_srgb(ctx->pipe_framebuffer.cbufs[i]->format); SET_BIT(flags, MALI_BLEND_MRT_SHADER, blend[i].is_shader); SET_BIT(flags, MALI_BLEND_LOAD_TIB, !blend[i].no_blending); @@ -676,8 +658,6 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, } if (dev->quirks & IS_BIFROST) { - struct bifrost_blend_rt *brts = rts; - brts[i].flags = flags; if (blend[i].is_shader) { @@ -689,7 +669,7 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, (fs->bo->gpu & (0xffffffffull << 32))); brts[i].shader = blend[i].shader.gpu; brts[i].unk2 = 0x0; - } else if (ctx->pipe_framebuffer.nr_cbufs > i) { + } else { enum pipe_format format = ctx->pipe_framebuffer.cbufs[i]->format; const struct util_format_description *format_desc; format_desc = util_format_description(format); @@ -707,14 +687,9 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, * blending. */ brts[i].unk2 = blend[i].no_blending ? 0x19 : 0x1a; - brts[i].shader_type = fs->blend_types[i]; - } else { - /* Dummy attachment for depth-only */ - brts[i].unk2 = 0x3; brts[i].shader_type = fs->blend_types[i]; } } else { - struct midgard_blend_rt *mrts = rts; mrts[i].flags = flags; if (blend[i].is_shader) { @@ -730,14 +705,15 @@ panfrost_frag_meta_blend_update(struct panfrost_context *ctx, static void panfrost_frag_shader_meta_init(struct panfrost_context *ctx, struct mali_shader_meta *fragmeta, - void *rts) + void *rts, + struct panfrost_blend_final *blend) { const struct panfrost_device *dev = pan_device(ctx->base.screen); struct panfrost_shader_state *fs; fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT); - bool msaa = ctx->rasterizer && ctx->rasterizer->base.multisample; + bool msaa = ctx->rasterizer->base.multisample; fragmeta->coverage_mask = msaa ? ctx->sample_mask : ~0; fragmeta->unknown2_3 = MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS) | 0x10; @@ -777,12 +753,10 @@ panfrost_frag_shader_meta_init(struct panfrost_context *ctx, * Just one of depth OR stencil is enough to trigger this. */ const struct pipe_depth_stencil_alpha_state *zsa = &ctx->depth_stencil->base; - bool zs_enabled = fs->writes_depth || fs->writes_stencil; - - if (zsa) { - zs_enabled |= (zsa->depth.enabled && zsa->depth.func != PIPE_FUNC_ALWAYS); - zs_enabled |= zsa->stencil[0].enabled; - } + bool zs_enabled = + fs->writes_depth || fs->writes_stencil || + (zsa->depth.enabled && zsa->depth.func != PIPE_FUNC_ALWAYS) || + zsa->stencil[0].enabled; SET_BIT(fragmeta->midgard1.flags_lo, MALI_READS_TILEBUFFER, fs->outputs_read || (!zs_enabled && fs->can_discard)); @@ -791,7 +765,7 @@ panfrost_frag_shader_meta_init(struct panfrost_context *ctx, panfrost_frag_meta_rasterizer_update(ctx, fragmeta); panfrost_frag_meta_zsa_update(ctx, fragmeta); - panfrost_frag_meta_blend_update(ctx, fragmeta, rts); + panfrost_frag_meta_blend_update(ctx, fragmeta, rts, blend); } void @@ -839,9 +813,14 @@ panfrost_emit_shader_meta(struct panfrost_batch *batch, if (rt_size) rts = rzalloc_size(ctx, rt_size * rt_count); - panfrost_frag_shader_meta_init(ctx, &meta, rts); + struct panfrost_blend_final blend[PIPE_MAX_COLOR_BUFS]; + + for (unsigned c = 0; c < ctx->pipe_framebuffer.nr_cbufs; ++c) + blend[c] = panfrost_get_blend_for_context(ctx, c); + + panfrost_frag_shader_meta_init(ctx, &meta, rts, blend); - xfer = panfrost_pool_alloc(&batch->pool, desc_size); + xfer = panfrost_pool_alloc_aligned(&batch->pool, desc_size, sizeof(meta)); memcpy(xfer.cpu, &meta, sizeof(meta)); memcpy(xfer.cpu + sizeof(meta), rts, rt_size * rt_count); @@ -886,7 +865,7 @@ panfrost_emit_viewport(struct panfrost_batch *batch, unsigned miny = MIN2(fb->height, vp_miny); unsigned maxy = MIN2(fb->height, vp_maxy); - if (ss && rast && rast->scissor) { + if (ss && rast->scissor) { minx = MAX2(ss->minx, minx); miny = MAX2(ss->miny, miny); maxx = MIN2(ss->maxx, maxx); @@ -928,10 +907,10 @@ panfrost_map_constant_buffer_gpu(struct panfrost_batch *batch, * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */ return rsrc->bo->gpu + cb->buffer_offset; } else if (cb->user_buffer) { - return panfrost_pool_upload(&batch->pool, + return panfrost_pool_upload_aligned(&batch->pool, cb->user_buffer + cb->buffer_offset, - cb->buffer_size); + cb->buffer_size, 16); } else { unreachable("No constant buffer"); } @@ -1133,8 +1112,8 @@ panfrost_emit_const_buf(struct panfrost_batch *batch, size_t sys_size = sizeof(float) * 4 * ss->sysval_count; size_t uniform_size = has_uniforms ? (buf->cb[0].buffer_size) : 0; size_t size = sys_size + uniform_size; - struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool, - size); + struct panfrost_transfer transfer = + panfrost_pool_alloc_aligned(&batch->pool, size, 16); /* Upload sysvals requested by the shader */ panfrost_upload_sysvals(batch, transfer.cpu, ss, stage); @@ -1152,7 +1131,10 @@ panfrost_emit_const_buf(struct panfrost_batch *batch, assert(ubo_count >= 1); size_t sz = MALI_UNIFORM_BUFFER_LENGTH * ubo_count; - struct panfrost_transfer ubos = panfrost_pool_alloc(&batch->pool, sz); + struct panfrost_transfer ubos = + panfrost_pool_alloc_aligned(&batch->pool, sz, + MALI_UNIFORM_BUFFER_LENGTH); + uint64_t *ubo_ptr = (uint64_t *) ubos.cpu; /* Upload uniforms as a UBO */ @@ -1197,28 +1179,30 @@ panfrost_emit_shared_memory(struct panfrost_batch *batch, struct midgard_payload_vertex_tiler *vtp) { struct panfrost_context *ctx = batch->ctx; + struct panfrost_device *dev = pan_device(ctx->base.screen); struct panfrost_shader_variants *all = ctx->shader[PIPE_SHADER_COMPUTE]; struct panfrost_shader_state *ss = &all->variants[all->active_variant]; unsigned single_size = util_next_power_of_two(MAX2(ss->shared_size, 128)); - unsigned shared_size = single_size * info->grid[0] * info->grid[1] * - info->grid[2] * 4; + + unsigned log2_instances = + util_logbase2_ceil(info->grid[0]) + + util_logbase2_ceil(info->grid[1]) + + util_logbase2_ceil(info->grid[2]); + + unsigned shared_size = single_size * (1 << log2_instances) * dev->core_count; struct panfrost_bo *bo = panfrost_batch_get_shared_memory(batch, shared_size, 1); struct mali_shared_memory shared = { .shared_memory = bo->gpu, - .shared_workgroup_count = - util_logbase2_ceil(info->grid[0]) + - util_logbase2_ceil(info->grid[1]) + - util_logbase2_ceil(info->grid[2]), - .shared_unk1 = 0x2, - .shared_shift = util_logbase2(single_size) - 1 + .shared_workgroup_count = log2_instances, + .shared_shift = util_logbase2(single_size) + 1 }; - vtp->postfix.shared_memory = panfrost_pool_upload(&batch->pool, &shared, - sizeof(shared)); + vtp->postfix.shared_memory = panfrost_pool_upload_aligned(&batch->pool, &shared, + sizeof(shared), 64); } static mali_ptr @@ -1269,9 +1253,10 @@ panfrost_emit_texture_descriptors(struct panfrost_batch *batch, return; if (device->quirks & IS_BIFROST) { - struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool, + struct panfrost_transfer T = panfrost_pool_alloc_aligned(&batch->pool, MALI_BIFROST_TEXTURE_LENGTH * - ctx->sampler_view_count[stage]); + ctx->sampler_view_count[stage], + MALI_BIFROST_TEXTURE_LENGTH); struct mali_bifrost_texture_packed *out = (struct mali_bifrost_texture_packed *) T.cpu; @@ -1307,10 +1292,11 @@ panfrost_emit_texture_descriptors(struct panfrost_batch *batch, trampolines[i] = panfrost_get_tex_desc(batch, stage, view); } - postfix->textures = panfrost_pool_upload(&batch->pool, + postfix->textures = panfrost_pool_upload_aligned(&batch->pool, trampolines, sizeof(uint64_t) * - ctx->sampler_view_count[stage]); + ctx->sampler_view_count[stage], + sizeof(uint64_t)); } } @@ -1328,7 +1314,7 @@ panfrost_emit_sampler_descriptors(struct panfrost_batch *batch, assert(MALI_BIFROST_SAMPLER_LENGTH == MALI_MIDGARD_SAMPLER_LENGTH); size_t sz = desc_size * ctx->sampler_count[stage]; - struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool, sz); + struct panfrost_transfer T = panfrost_pool_alloc_aligned(&batch->pool, sz, desc_size); struct mali_midgard_sampler_packed *out = (struct mali_midgard_sampler_packed *) T.cpu; for (unsigned i = 0; i < ctx->sampler_count[stage]; ++i) @@ -1343,19 +1329,35 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch, { struct panfrost_context *ctx = batch->ctx; struct panfrost_vertex_state *so = ctx->vertex; + struct panfrost_shader_state *vs = panfrost_get_shader_state(ctx, PIPE_SHADER_VERTEX); unsigned instance_shift = vertex_postfix->instance_shift; unsigned instance_odd = vertex_postfix->instance_odd; - /* Staged mali_attr, and index into them. i =/= k, depending on the - * vertex buffer mask and instancing. Twice as much room is allocated, - * for a worst case of NPOT_DIVIDEs which take up extra slot */ - union mali_attr attrs[PIPE_MAX_ATTRIBS * 2]; + /* Worst case: everything is NPOT, which is only possible if instancing + * is enabled. Otherwise single record is gauranteed */ + bool could_npot = instance_shift || instance_odd; + + struct panfrost_transfer S = panfrost_pool_alloc_aligned(&batch->pool, + MALI_ATTRIBUTE_BUFFER_LENGTH * vs->attribute_count * + (could_npot ? 2 : 1), + MALI_ATTRIBUTE_BUFFER_LENGTH * 2); + + struct panfrost_transfer T = panfrost_pool_alloc_aligned(&batch->pool, + MALI_ATTRIBUTE_LENGTH * vs->attribute_count, + MALI_ATTRIBUTE_LENGTH); + + struct mali_attribute_buffer_packed *bufs = + (struct mali_attribute_buffer_packed *) S.cpu; + + struct mali_attribute_packed *out = + (struct mali_attribute_packed *) T.cpu; + unsigned attrib_to_buffer[PIPE_MAX_ATTRIBS] = { 0 }; unsigned k = 0; for (unsigned i = 0; i < so->num_elements; ++i) { - /* We map a mali_attr to be 1:1 with the mali_attr_meta, which + /* We map buffers 1:1 with the attributes, which * means duplicating some vertex buffers (who cares? aside from * maybe some caching implications but I somehow doubt that * matters) */ @@ -1374,81 +1376,96 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch, if (!rsrc) continue; - /* Align to 64 bytes by masking off the lower bits. This - * will be adjusted back when we fixup the src_offset in - * mali_attr_meta */ - - mali_ptr raw_addr = rsrc->bo->gpu + buf->buffer_offset; - mali_ptr addr = raw_addr & ~63; - unsigned chopped_addr = raw_addr - addr; - /* Add a dependency of the batch on the vertex buffer */ panfrost_batch_add_bo(batch, rsrc->bo, PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ | PAN_BO_ACCESS_VERTEX_TILER); - /* Set common fields */ - attrs[k].elements = addr; - attrs[k].stride = buf->stride; + /* Mask off lower bits, see offset fixup below */ + mali_ptr raw_addr = rsrc->bo->gpu + buf->buffer_offset; + mali_ptr addr = raw_addr & ~63; /* Since we advanced the base pointer, we shrink the buffer - * size */ - attrs[k].size = rsrc->base.width0 - buf->buffer_offset; + * size, but add the offset we subtracted */ + unsigned size = rsrc->base.width0 + (raw_addr - addr) + - buf->buffer_offset; - /* We need to add the extra size we masked off (for - * correctness) so the data doesn't get clamped away */ - attrs[k].size += chopped_addr; + /* When there is a divisor, the hardware-level divisor is + * the product of the instance divisor and the padded count */ + unsigned divisor = elem->instance_divisor; + unsigned hw_divisor = ctx->padded_count * divisor; + unsigned stride = buf->stride; - /* For non-instancing make sure we initialize */ - attrs[k].shift = attrs[k].extra_flags = 0; + /* If there's a divisor(=1) but no instancing, we want every + * attribute to be the same */ - /* Instancing uses a dramatically different code path than - * linear, so dispatch for the actual emission now that the - * common code is finished */ + if (divisor && ctx->instance_count == 1) + stride = 0; - unsigned divisor = elem->instance_divisor; + if (!divisor || ctx->instance_count <= 1) { + pan_pack(bufs + k, ATTRIBUTE_BUFFER, cfg) { + if (ctx->instance_count > 1) + cfg.type = MALI_ATTRIBUTE_TYPE_1D_MODULUS; + + cfg.pointer = addr; + cfg.stride = stride; + cfg.size = size; + cfg.divisor_r = instance_shift; + cfg.divisor_p = instance_odd; + } + } else if (util_is_power_of_two_or_zero(hw_divisor)) { + pan_pack(bufs + k, ATTRIBUTE_BUFFER, cfg) { + cfg.type = MALI_ATTRIBUTE_TYPE_1D_POT_DIVISOR; + cfg.pointer = addr; + cfg.stride = stride; + cfg.size = size; + cfg.divisor_r = __builtin_ctz(hw_divisor); + } - if (divisor && ctx->instance_count == 1) { - /* Silly corner case where there's a divisor(=1) but - * there's no legitimate instancing. So we want *every* - * attribute to be the same. So set stride to zero so - * we don't go anywhere. */ - - attrs[k].size = attrs[k].stride + chopped_addr; - attrs[k].stride = 0; - attrs[k++].elements |= MALI_ATTR_LINEAR; - } else if (ctx->instance_count <= 1) { - /* Normal, non-instanced attributes */ - attrs[k++].elements |= MALI_ATTR_LINEAR; } else { - k += panfrost_vertex_instanced(ctx->padded_count, - instance_shift, - instance_odd, - divisor, &attrs[k]); - } - } + unsigned shift = 0, extra_flags = 0; - /* Add special gl_VertexID/gl_InstanceID buffers */ + unsigned magic_divisor = + panfrost_compute_magic_divisor(hw_divisor, &shift, &extra_flags); - struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool, - MALI_ATTRIBUTE_LENGTH * (PAN_INSTANCE_ID + 1)); + pan_pack(bufs + k, ATTRIBUTE_BUFFER, cfg) { + cfg.type = MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR; + cfg.pointer = addr; + cfg.stride = stride; + cfg.size = size; - struct mali_attribute_packed *out = - (struct mali_attribute_packed *) T.cpu; + cfg.divisor_r = shift; + cfg.divisor_e = extra_flags; + } + + pan_pack(bufs + k + 1, ATTRIBUTE_BUFFER_CONTINUATION_NPOT, cfg) { + cfg.divisor_numerator = magic_divisor; + cfg.divisor = divisor; + } - panfrost_vertex_id(ctx->padded_count, &attrs[k]); + ++k; + } - pan_pack(out + PAN_VERTEX_ID, ATTRIBUTE, cfg) { - cfg.buffer_index = k++; - cfg.format = so->formats[PAN_VERTEX_ID]; + ++k; } - panfrost_instance_id(ctx->padded_count, &attrs[k]); + /* Add special gl_VertexID/gl_InstanceID buffers */ + + if (unlikely(vs->attribute_count >= PAN_VERTEX_ID)) { + panfrost_vertex_id(ctx->padded_count, &bufs[k], ctx->instance_count > 1); + + pan_pack(out + PAN_VERTEX_ID, ATTRIBUTE, cfg) { + cfg.buffer_index = k++; + cfg.format = so->formats[PAN_VERTEX_ID]; + } + + panfrost_instance_id(ctx->padded_count, &bufs[k], ctx->instance_count > 1); - pan_pack(out + PAN_INSTANCE_ID, ATTRIBUTE, cfg) { - cfg.buffer_index = k++; - cfg.format = so->formats[PAN_INSTANCE_ID]; + pan_pack(out + PAN_INSTANCE_ID, ATTRIBUTE, cfg) { + cfg.buffer_index = k++; + cfg.format = so->formats[PAN_INSTANCE_ID]; + } } /* Attribute addresses require 64-byte alignment, so let: @@ -1488,28 +1505,25 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch, } } - - vertex_postfix->attributes = panfrost_pool_upload(&batch->pool, attrs, - k * sizeof(*attrs)); - + vertex_postfix->attributes = S.gpu; vertex_postfix->attribute_meta = T.gpu; } static mali_ptr -panfrost_emit_varyings(struct panfrost_batch *batch, union mali_attr *slot, - unsigned stride, unsigned count) +panfrost_emit_varyings(struct panfrost_batch *batch, + struct mali_attribute_buffer_packed *slot, + unsigned stride, unsigned count) { - /* Fill out the descriptor */ - slot->stride = stride; - slot->size = stride * count; - slot->shift = slot->extra_flags = 0; - - struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool, - slot->size); + unsigned size = stride * count; + mali_ptr ptr = panfrost_pool_alloc_aligned(&batch->invisible_pool, size, 64).gpu; - slot->elements = transfer.gpu | MALI_ATTR_LINEAR; + pan_pack(slot, ATTRIBUTE_BUFFER, cfg) { + cfg.stride = stride; + cfg.size = size; + cfg.pointer = ptr; + } - return transfer.gpu; + return ptr; } static unsigned @@ -1520,16 +1534,14 @@ panfrost_streamout_offset(unsigned stride, unsigned offset, } static void -panfrost_emit_streamout(struct panfrost_batch *batch, union mali_attr *slot, - unsigned stride, unsigned offset, unsigned count, +panfrost_emit_streamout(struct panfrost_batch *batch, + struct mali_attribute_buffer_packed *slot, + unsigned stride_words, unsigned offset, unsigned count, struct pipe_stream_output_target *target) { - /* Fill out the descriptor */ - slot->stride = stride * 4; - slot->shift = slot->extra_flags = 0; - + unsigned stride = stride_words * 4; unsigned max_size = target->buffer_size; - unsigned expected_size = slot->stride * count; + unsigned expected_size = stride * count; /* Grab the BO and bind it to the batch */ struct panfrost_bo *bo = pan_resource(target->buffer)->bo; @@ -1544,9 +1556,13 @@ panfrost_emit_streamout(struct panfrost_batch *batch, union mali_attr *slot, PAN_BO_ACCESS_FRAGMENT); /* We will have an offset applied to get alignment */ - mali_ptr addr = bo->gpu + target->buffer_offset + (offset * slot->stride); - slot->elements = (addr & ~63) | MALI_ATTR_LINEAR; - slot->size = MIN2(max_size, expected_size) + (addr & 63); + mali_ptr addr = bo->gpu + target->buffer_offset + (offset * stride); + + pan_pack(slot, ATTRIBUTE_BUFFER, cfg) { + cfg.pointer = (addr & ~63); + cfg.stride = stride; + cfg.size = MIN2(max_size, expected_size) + (addr & 63); + } } static bool @@ -1684,8 +1700,9 @@ pan_varying_present( /* Emitters for varying records */ -static struct mali_attr_meta -pan_emit_vary(unsigned present, enum pan_special_varying buf, +static void +pan_emit_vary(struct mali_attribute_packed *out, + unsigned present, enum pan_special_varying buf, unsigned quirks, enum mali_format format, unsigned offset) { @@ -1694,22 +1711,21 @@ pan_emit_vary(unsigned present, enum pan_special_varying buf, panfrost_get_default_swizzle(nr_channels) : panfrost_bifrost_swizzle(nr_channels); - struct mali_attr_meta meta = { - .index = pan_varying_index(present, buf), - .unknown1 = quirks & IS_BIFROST ? 0x0 : 0x2, - .format = (format << 12) | swizzle, - .src_offset = offset - }; - - return meta; + pan_pack(out, ATTRIBUTE, cfg) { + cfg.buffer_index = pan_varying_index(present, buf); + cfg.unknown = quirks & IS_BIFROST ? 0x0 : 0x1; + cfg.format = (format << 12) | swizzle; + cfg.offset = offset; + } } /* General varying that is unused */ -static struct mali_attr_meta -pan_emit_vary_only(unsigned present, unsigned quirks) +static void +pan_emit_vary_only(struct mali_attribute_packed *out, + unsigned present, unsigned quirks) { - return pan_emit_vary(present, 0, quirks, MALI_VARYING_DISCARD, 0); + pan_emit_vary(out, present, 0, quirks, MALI_VARYING_DISCARD, 0); } /* Special records */ @@ -1722,12 +1738,13 @@ static const enum mali_format pan_varying_formats[PAN_VARY_MAX] = { [PAN_VARY_FRAGCOORD] = MALI_RGBA32F }; -static struct mali_attr_meta -pan_emit_vary_special(unsigned present, enum pan_special_varying buf, +static void +pan_emit_vary_special(struct mali_attribute_packed *out, + unsigned present, enum pan_special_varying buf, unsigned quirks) { assert(buf < PAN_VARY_MAX); - return pan_emit_vary(present, buf, quirks, pan_varying_formats[buf], 0); + pan_emit_vary(out, present, buf, quirks, pan_varying_formats[buf], 0); } static enum mali_format @@ -1743,8 +1760,9 @@ pan_xfb_format(enum mali_format format, unsigned nr) * a bitfield) 32-bit, smaller than a 64-bit pointer, so may as well pass by * value. */ -static struct mali_attr_meta -pan_emit_vary_xfb(unsigned present, +static void +pan_emit_vary_xfb(struct mali_attribute_packed *out, + unsigned present, unsigned max_xfb, unsigned *streamout_offsets, unsigned quirks, @@ -1755,23 +1773,18 @@ pan_emit_vary_xfb(unsigned present, panfrost_get_default_swizzle(o.num_components) : panfrost_bifrost_swizzle(o.num_components); - /* Otherwise construct a record for it */ - struct mali_attr_meta meta = { + pan_pack(out, ATTRIBUTE, cfg) { /* XFB buffers come after everything else */ - .index = pan_xfb_base(present) + o.output_buffer, - - /* As usual unknown bit */ - .unknown1 = quirks & IS_BIFROST ? 0x0 : 0x2, + cfg.buffer_index = pan_xfb_base(present) + o.output_buffer; + cfg.unknown = quirks & IS_BIFROST ? 0x0 : 0x1; /* Override number of channels and precision to highp */ - .format = (pan_xfb_format(format, o.num_components) << 12) | swizzle, + cfg.format = (pan_xfb_format(format, o.num_components) << 12) | swizzle; /* Apply given offsets together */ - .src_offset = (o.dst_offset * 4) /* dwords */ - + streamout_offsets[o.output_buffer] - }; - - return meta; + cfg.offset = (o.dst_offset * 4) /* dwords */ + + streamout_offsets[o.output_buffer]; + } } /* Determine if we should capture a varying for XFB. This requires actually @@ -1789,51 +1802,21 @@ panfrost_xfb_captured(struct panfrost_shader_state *xfb, return o->output_buffer < max_xfb; } -/* Higher-level wrapper around all of the above, classifying a varying into one - * of the above types */ - -static struct mali_attr_meta -panfrost_emit_varying( - struct panfrost_shader_state *stage, +static void +pan_emit_general_varying(struct mali_attribute_packed *out, struct panfrost_shader_state *other, struct panfrost_shader_state *xfb, + gl_varying_slot loc, + enum mali_format format, unsigned present, - unsigned max_xfb, - unsigned *streamout_offsets, unsigned quirks, unsigned *gen_offsets, enum mali_format *gen_formats, unsigned *gen_stride, unsigned idx, - bool should_alloc, - bool is_fragment) + bool should_alloc) { - gl_varying_slot loc = stage->varyings_loc[idx]; - enum mali_format format = stage->varyings[idx]; - - /* Override format to match linkage */ - if (!should_alloc && gen_formats[idx]) - format = gen_formats[idx]; - - if (has_point_coord(stage->point_sprite_mask, loc)) { - return pan_emit_vary_special(present, PAN_VARY_PNTCOORD, quirks); - } else if (panfrost_xfb_captured(xfb, loc, max_xfb)) { - struct pipe_stream_output *o = pan_get_so(&xfb->stream_output, loc); - return pan_emit_vary_xfb(present, max_xfb, streamout_offsets, quirks, format, *o); - } else if (loc == VARYING_SLOT_POS) { - if (is_fragment) - return pan_emit_vary_special(present, PAN_VARY_FRAGCOORD, quirks); - else - return pan_emit_vary_special(present, PAN_VARY_POSITION, quirks); - } else if (loc == VARYING_SLOT_PSIZ) { - return pan_emit_vary_special(present, PAN_VARY_PSIZ, quirks); - } else if (loc == VARYING_SLOT_PNTC) { - return pan_emit_vary_special(present, PAN_VARY_PNTCOORD, quirks); - } else if (loc == VARYING_SLOT_FACE) { - return pan_emit_vary_special(present, PAN_VARY_FACE, quirks); - } - - /* We've exhausted special cases, so it's otherwise a general varying. Check if we're linked */ + /* Check if we're linked */ signed other_idx = -1; for (unsigned j = 0; j < other->varying_count; ++j) { @@ -1843,8 +1826,10 @@ panfrost_emit_varying( } } - if (other_idx < 0) - return pan_emit_vary_only(present, quirks); + if (other_idx < 0) { + pan_emit_vary_only(out, present, quirks); + return; + } unsigned offset = gen_offsets[other_idx]; @@ -1876,25 +1861,72 @@ panfrost_emit_varying( *gen_stride += size; } - return pan_emit_vary(present, PAN_VARY_GENERAL, - quirks, format, offset); + pan_emit_vary(out, present, PAN_VARY_GENERAL, quirks, format, offset); } +/* Higher-level wrapper around all of the above, classifying a varying into one + * of the above types */ + static void -pan_emit_special_input(union mali_attr *varyings, +panfrost_emit_varying( + struct mali_attribute_packed *out, + struct panfrost_shader_state *stage, + struct panfrost_shader_state *other, + struct panfrost_shader_state *xfb, + unsigned present, + unsigned max_xfb, + unsigned *streamout_offsets, + unsigned quirks, + unsigned *gen_offsets, + enum mali_format *gen_formats, + unsigned *gen_stride, + unsigned idx, + bool should_alloc, + bool is_fragment) +{ + gl_varying_slot loc = stage->varyings_loc[idx]; + enum mali_format format = stage->varyings[idx]; + + /* Override format to match linkage */ + if (!should_alloc && gen_formats[idx]) + format = gen_formats[idx]; + + if (has_point_coord(stage->point_sprite_mask, loc)) { + pan_emit_vary_special(out, present, PAN_VARY_PNTCOORD, quirks); + } else if (panfrost_xfb_captured(xfb, loc, max_xfb)) { + struct pipe_stream_output *o = pan_get_so(&xfb->stream_output, loc); + pan_emit_vary_xfb(out, present, max_xfb, streamout_offsets, quirks, format, *o); + } else if (loc == VARYING_SLOT_POS) { + if (is_fragment) + pan_emit_vary_special(out, present, PAN_VARY_FRAGCOORD, quirks); + else + pan_emit_vary_special(out, present, PAN_VARY_POSITION, quirks); + } else if (loc == VARYING_SLOT_PSIZ) { + pan_emit_vary_special(out, present, PAN_VARY_PSIZ, quirks); + } else if (loc == VARYING_SLOT_PNTC) { + pan_emit_vary_special(out, present, PAN_VARY_PNTCOORD, quirks); + } else if (loc == VARYING_SLOT_FACE) { + pan_emit_vary_special(out, present, PAN_VARY_FACE, quirks); + } else { + pan_emit_general_varying(out, other, xfb, loc, format, present, + quirks, gen_offsets, gen_formats, gen_stride, + idx, should_alloc); + } +} + +static void +pan_emit_special_input(struct mali_attribute_buffer_packed *out, unsigned present, enum pan_special_varying v, - mali_ptr addr) + unsigned special) { if (present & (1 << v)) { - /* Ensure we write exactly once for performance and with fields - * zeroed appropriately to avoid flakes */ - - union mali_attr s = { - .elements = addr - }; + unsigned idx = pan_varying_index(present, v); - varyings[pan_varying_index(present, v)] = s; + pan_pack(out + idx, ATTRIBUTE_BUFFER, cfg) { + cfg.special = special; + cfg.type = 0; + } } } @@ -1915,12 +1947,11 @@ panfrost_emit_varying_descriptor(struct panfrost_batch *batch, vs = panfrost_get_shader_state(ctx, PIPE_SHADER_VERTEX); fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT); - vs_size = sizeof(struct mali_attr_meta) * vs->varying_count; - fs_size = sizeof(struct mali_attr_meta) * fs->varying_count; + vs_size = MALI_ATTRIBUTE_LENGTH * vs->varying_count; + fs_size = MALI_ATTRIBUTE_LENGTH * fs->varying_count; - struct panfrost_transfer trans = panfrost_pool_alloc(&batch->pool, - vs_size + - fs_size); + struct panfrost_transfer trans = panfrost_pool_alloc_aligned( + &batch->pool, vs_size + fs_size, MALI_ATTRIBUTE_LENGTH); struct pipe_stream_output_info *so = &vs->stream_output; unsigned present = pan_varying_present(vs, fs, dev->quirks); @@ -1948,27 +1979,29 @@ panfrost_emit_varying_descriptor(struct panfrost_batch *batch, ctx->streamout.targets[i]); } - struct mali_attr_meta *ovs = (struct mali_attr_meta *)trans.cpu; - struct mali_attr_meta *ofs = ovs + vs->varying_count; + struct mali_attribute_packed *ovs = (struct mali_attribute_packed *)trans.cpu; + struct mali_attribute_packed *ofs = ovs + vs->varying_count; for (unsigned i = 0; i < vs->varying_count; i++) { - ovs[i] = panfrost_emit_varying(vs, fs, vs, present, + panfrost_emit_varying(ovs + i, vs, fs, vs, present, ctx->streamout.num_targets, streamout_offsets, dev->quirks, gen_offsets, gen_formats, &gen_stride, i, true, false); } for (unsigned i = 0; i < fs->varying_count; i++) { - ofs[i] = panfrost_emit_varying(fs, vs, vs, present, + panfrost_emit_varying(ofs + i, fs, vs, vs, present, ctx->streamout.num_targets, streamout_offsets, dev->quirks, gen_offsets, gen_formats, &gen_stride, i, false, true); } unsigned xfb_base = pan_xfb_base(present); - struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool, - sizeof(union mali_attr) * (xfb_base + ctx->streamout.num_targets)); - union mali_attr *varyings = (union mali_attr *) T.cpu; + struct panfrost_transfer T = panfrost_pool_alloc_aligned(&batch->pool, + MALI_ATTRIBUTE_BUFFER_LENGTH * (xfb_base + ctx->streamout.num_targets), + MALI_ATTRIBUTE_BUFFER_LENGTH * 2); + struct mali_attribute_buffer_packed *varyings = + (struct mali_attribute_buffer_packed *) T.cpu; /* Emit the stream out buffers */ @@ -1998,9 +2031,9 @@ panfrost_emit_varying_descriptor(struct panfrost_batch *batch, 2, vertex_count); } - pan_emit_special_input(varyings, present, PAN_VARY_PNTCOORD, MALI_VARYING_POINT_COORD); - pan_emit_special_input(varyings, present, PAN_VARY_FACE, MALI_VARYING_FRONT_FACING); - pan_emit_special_input(varyings, present, PAN_VARY_FRAGCOORD, MALI_VARYING_FRAG_COORD); + pan_emit_special_input(varyings, present, PAN_VARY_PNTCOORD, MALI_ATTRIBUTE_SPECIAL_POINT_COORD); + pan_emit_special_input(varyings, present, PAN_VARY_FACE, MALI_ATTRIBUTE_SPECIAL_FRONT_FACING); + pan_emit_special_input(varyings, present, PAN_VARY_FRAGCOORD, MALI_ATTRIBUTE_SPECIAL_FRAG_COORD); vertex_postfix->varyings = T.gpu; tiler_postfix->varyings = T.gpu; @@ -2064,13 +2097,10 @@ panfrost_emit_vertex_tiler_jobs(struct panfrost_batch *batch, /* If rasterizer discard is enable, only submit the vertex */ - bool rasterizer_discard = ctx->rasterizer && - ctx->rasterizer->base.rasterizer_discard; - unsigned vertex = panfrost_new_job(&batch->pool, &batch->scoreboard, MALI_JOB_TYPE_VERTEX, false, 0, vp, vp_size, false); - if (rasterizer_discard) + if (ctx->rasterizer->base.rasterizer_discard) return; panfrost_new_job(&batch->pool, &batch->scoreboard, MALI_JOB_TYPE_TILER, false, vertex, tp, tp_size, @@ -2132,5 +2162,5 @@ panfrost_emit_sample_locations(struct panfrost_batch *batch) 0, 0, }; - return panfrost_pool_upload(&batch->pool, locations, 96 * sizeof(uint16_t)); + return panfrost_pool_upload_aligned(&batch->pool, locations, 96 * sizeof(uint16_t), 64); }