panfrost: Use packs for attributes
[mesa.git] / src / gallium / drivers / panfrost / pan_cmdstream.c
index 8f6142f7b4a76415eda630638ce97e7d4bc1120f..1069dab5c1e9200fff802f7287e4baa77688ddd6 100644 (file)
@@ -350,183 +350,101 @@ panfrost_shader_meta_init(struct panfrost_context *ctx,
 }
 
 static unsigned
-panfrost_translate_compare_func(enum pipe_compare_func in)
+translate_tex_wrap(enum pipe_tex_wrap w)
 {
-        switch (in) {
-        case PIPE_FUNC_NEVER:
-                return MALI_FUNC_NEVER;
-
-        case PIPE_FUNC_LESS:
-                return MALI_FUNC_LESS;
-
-        case PIPE_FUNC_EQUAL:
-                return MALI_FUNC_EQUAL;
-
-        case PIPE_FUNC_LEQUAL:
-                return MALI_FUNC_LEQUAL;
-
-        case PIPE_FUNC_GREATER:
-                return MALI_FUNC_GREATER;
-
-        case PIPE_FUNC_NOTEQUAL:
-                return MALI_FUNC_NOT_EQUAL;
-
-        case PIPE_FUNC_GEQUAL:
-                return MALI_FUNC_GEQUAL;
-
-        case PIPE_FUNC_ALWAYS:
-                return MALI_FUNC_ALWAYS;
-
-        default:
-                unreachable("Invalid func");
+        switch (w) {
+        case PIPE_TEX_WRAP_REPEAT: return MALI_WRAP_MODE_REPEAT;
+        case PIPE_TEX_WRAP_CLAMP: return MALI_WRAP_MODE_CLAMP;
+        case PIPE_TEX_WRAP_CLAMP_TO_EDGE: return MALI_WRAP_MODE_CLAMP_TO_EDGE;
+        case PIPE_TEX_WRAP_CLAMP_TO_BORDER: return MALI_WRAP_MODE_CLAMP_TO_BORDER;
+        case PIPE_TEX_WRAP_MIRROR_REPEAT: return MALI_WRAP_MODE_MIRRORED_REPEAT;
+        case PIPE_TEX_WRAP_MIRROR_CLAMP: return MALI_WRAP_MODE_MIRRORED_CLAMP;
+        case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_EDGE;
+        case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_BORDER;
+        default: unreachable("Invalid wrap");
         }
 }
 
-static unsigned
-panfrost_translate_stencil_op(enum pipe_stencil_op in)
-{
-        switch (in) {
-        case PIPE_STENCIL_OP_KEEP:
-                return MALI_STENCIL_OP_KEEP;
-
-        case PIPE_STENCIL_OP_ZERO:
-                return MALI_STENCIL_OP_ZERO;
-
-        case PIPE_STENCIL_OP_REPLACE:
-               return MALI_STENCIL_OP_REPLACE;
-
-        case PIPE_STENCIL_OP_INCR:
-                return MALI_STENCIL_OP_INCR_SAT;
-
-        case PIPE_STENCIL_OP_DECR:
-                return MALI_STENCIL_OP_DECR_SAT;
-
-        case PIPE_STENCIL_OP_INCR_WRAP:
-                return MALI_STENCIL_OP_INCR_WRAP;
+/* The hardware compares in the wrong order order, so we have to flip before
+ * encoding. Yes, really. */
 
-        case PIPE_STENCIL_OP_DECR_WRAP:
-                return MALI_STENCIL_OP_DECR_WRAP;
+static enum mali_func
+panfrost_sampler_compare_func(const struct pipe_sampler_state *cso)
+{
+        if (!cso->compare_mode)
+                return MALI_FUNC_NEVER;
 
-        case PIPE_STENCIL_OP_INVERT:
-                return MALI_STENCIL_OP_INVERT;
+        enum mali_func f = panfrost_translate_compare_func(cso->compare_func);
+        return panfrost_flip_compare_func(f);
+}
 
-        default:
-                unreachable("Invalid stencil op");
+static enum mali_mipmap_mode
+pan_pipe_to_mipmode(enum pipe_tex_mipfilter f)
+{
+        switch (f) {
+        case PIPE_TEX_MIPFILTER_NEAREST: return MALI_MIPMAP_MODE_NEAREST;
+        case PIPE_TEX_MIPFILTER_LINEAR: return MALI_MIPMAP_MODE_TRILINEAR;
+        case PIPE_TEX_MIPFILTER_NONE: return MALI_MIPMAP_MODE_NONE;
+        default: unreachable("Invalid");
         }
 }
 
-static unsigned
-translate_tex_wrap(enum pipe_tex_wrap w)
+void panfrost_sampler_desc_init(const struct pipe_sampler_state *cso,
+                                struct mali_midgard_sampler_packed *hw)
 {
-        switch (w) {
-        case PIPE_TEX_WRAP_REPEAT:
-                return MALI_WRAP_MODE_REPEAT;
-
-        case PIPE_TEX_WRAP_CLAMP:
-                return MALI_WRAP_MODE_CLAMP;
+        pan_pack(hw, MIDGARD_SAMPLER, cfg) {
+                cfg.magnify_nearest = cso->mag_img_filter == PIPE_TEX_FILTER_NEAREST;
+                cfg.minify_nearest = cso->min_img_filter == PIPE_TEX_FILTER_NEAREST;
+                cfg.mipmap_mode = (cso->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR) ?
+                        MALI_MIPMAP_MODE_TRILINEAR : MALI_MIPMAP_MODE_NEAREST;
+                cfg.normalized_coordinates = cso->normalized_coords;
 
-        case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
-                return MALI_WRAP_MODE_CLAMP_TO_EDGE;
+                cfg.lod_bias = FIXED_16(cso->lod_bias, true);
 
-        case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
-                return MALI_WRAP_MODE_CLAMP_TO_BORDER;
+                cfg.minimum_lod = FIXED_16(cso->min_lod, false);
 
-        case PIPE_TEX_WRAP_MIRROR_REPEAT:
-                return MALI_WRAP_MODE_MIRRORED_REPEAT;
+                /* If necessary, we disable mipmapping in the sampler descriptor by
+                 * clamping the LOD as tight as possible (from 0 to epsilon,
+                 * essentially -- remember these are fixed point numbers, so
+                 * epsilon=1/256) */
 
-        case PIPE_TEX_WRAP_MIRROR_CLAMP:
-                return MALI_WRAP_MODE_MIRRORED_CLAMP;
+                cfg.maximum_lod = (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) ?
+                        cfg.minimum_lod + 1 :
+                        FIXED_16(cso->max_lod, false);
 
-        case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
-                return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_EDGE;
+                cfg.wrap_mode_s = translate_tex_wrap(cso->wrap_s);
+                cfg.wrap_mode_t = translate_tex_wrap(cso->wrap_t);
+                cfg.wrap_mode_r = translate_tex_wrap(cso->wrap_r);
 
-        case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
-                return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_BORDER;
+                cfg.compare_function = panfrost_sampler_compare_func(cso);
+                cfg.seamless_cube_map = cso->seamless_cube_map;
 
-        default:
-                unreachable("Invalid wrap");
+                cfg.border_color_r = cso->border_color.f[0];
+                cfg.border_color_g = cso->border_color.f[1];
+                cfg.border_color_b = cso->border_color.f[2];
+                cfg.border_color_a = cso->border_color.f[3];
         }
 }
 
-void panfrost_sampler_desc_init(const struct pipe_sampler_state *cso,
-                                struct mali_sampler_descriptor *hw)
-{
-        unsigned func = panfrost_translate_compare_func(cso->compare_func);
-        bool min_nearest = cso->min_img_filter == PIPE_TEX_FILTER_NEAREST;
-        bool mag_nearest = cso->mag_img_filter == PIPE_TEX_FILTER_NEAREST;
-        bool mip_linear  = cso->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR;
-        unsigned min_filter = min_nearest ? MALI_SAMP_MIN_NEAREST : 0;
-        unsigned mag_filter = mag_nearest ? MALI_SAMP_MAG_NEAREST : 0;
-        unsigned mip_filter = mip_linear  ?
-                              (MALI_SAMP_MIP_LINEAR_1 | MALI_SAMP_MIP_LINEAR_2) : 0;
-        unsigned normalized = cso->normalized_coords ? MALI_SAMP_NORM_COORDS : 0;
-
-        *hw = (struct mali_sampler_descriptor) {
-                .filter_mode = min_filter | mag_filter | mip_filter |
-                               normalized,
-                .wrap_s = translate_tex_wrap(cso->wrap_s),
-                .wrap_t = translate_tex_wrap(cso->wrap_t),
-                .wrap_r = translate_tex_wrap(cso->wrap_r),
-                .compare_func = cso->compare_mode ?
-                        panfrost_flip_compare_func(func) :
-                        MALI_FUNC_NEVER,
-                .border_color = {
-                        cso->border_color.f[0],
-                        cso->border_color.f[1],
-                        cso->border_color.f[2],
-                        cso->border_color.f[3]
-                },
-                .min_lod = FIXED_16(cso->min_lod, false), /* clamp at 0 */
-                .max_lod = FIXED_16(cso->max_lod, false),
-                .lod_bias = FIXED_16(cso->lod_bias, true), /* can be negative */
-                .seamless_cube_map = cso->seamless_cube_map,
-        };
-
-        /* If necessary, we disable mipmapping in the sampler descriptor by
-         * clamping the LOD as tight as possible (from 0 to epsilon,
-         * essentially -- remember these are fixed point numbers, so
-         * epsilon=1/256) */
-
-        if (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
-                hw->max_lod = hw->min_lod + 1;
-}
-
 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state *cso,
-                                        struct bifrost_sampler_descriptor *hw)
+                                        struct mali_bifrost_sampler_packed *hw)
 {
-        *hw = (struct bifrost_sampler_descriptor) {
-                .unk1 = 0x1,
-                .wrap_s = translate_tex_wrap(cso->wrap_s),
-                .wrap_t = translate_tex_wrap(cso->wrap_t),
-                .wrap_r = translate_tex_wrap(cso->wrap_r),
-                .unk8 = 0x8,
-                .min_filter = cso->min_img_filter == PIPE_TEX_FILTER_NEAREST,
-                .norm_coords = cso->normalized_coords,
-                .mip_filter = cso->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR,
-                .mag_filter = cso->mag_img_filter == PIPE_TEX_FILTER_LINEAR,
-                .min_lod = FIXED_16(cso->min_lod, false), /* clamp at 0 */
-                .max_lod = FIXED_16(cso->max_lod, false),
-        };
-
-        /* If necessary, we disable mipmapping in the sampler descriptor by
-         * clamping the LOD as tight as possible (from 0 to epsilon,
-         * essentially -- remember these are fixed point numbers, so
-         * epsilon=1/256) */
-
-        if (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
-                hw->max_lod = hw->min_lod + 1;
-}
-
-static void
-panfrost_make_stencil_state(const struct pipe_stencil_state *in,
-                            void *out)
-{
-        pan_pack(out, STENCIL, cfg) {
-                cfg.mask = in->valuemask;
-                cfg.compare_function = panfrost_translate_compare_func(in->func);
-                cfg.stencil_fail = panfrost_translate_stencil_op(in->fail_op);
-                cfg.depth_fail = panfrost_translate_stencil_op(in->zfail_op);
-                cfg.depth_pass = panfrost_translate_stencil_op(in->zpass_op);
+        pan_pack(hw, BIFROST_SAMPLER, cfg) {
+                cfg.magnify_linear = cso->mag_img_filter == PIPE_TEX_FILTER_LINEAR;
+                cfg.minify_linear = cso->min_img_filter == PIPE_TEX_FILTER_LINEAR;
+                cfg.mipmap_mode = pan_pipe_to_mipmode(cso->min_mip_filter);
+                cfg.normalized_coordinates = cso->normalized_coords;
+
+                cfg.lod_bias = FIXED_16(cso->lod_bias, true);
+                cfg.minimum_lod = FIXED_16(cso->min_lod, false);
+                cfg.maximum_lod = FIXED_16(cso->max_lod, false);
+
+                cfg.wrap_mode_s = translate_tex_wrap(cso->wrap_s);
+                cfg.wrap_mode_t = translate_tex_wrap(cso->wrap_t);
+                cfg.wrap_mode_r = translate_tex_wrap(cso->wrap_r);
+
+                cfg.compare_function = panfrost_sampler_compare_func(cso);
+                cfg.seamless_cube_map = cso->seamless_cube_map;
         }
 }
 
@@ -578,57 +496,37 @@ static void
 panfrost_frag_meta_zsa_update(struct panfrost_context *ctx,
                               struct mali_shader_meta *fragmeta)
 {
-        const struct pipe_depth_stencil_alpha_state *zsa = ctx->depth_stencil;
+        const struct panfrost_zsa_state *so = ctx->depth_stencil;
         int zfunc = PIPE_FUNC_ALWAYS;
 
-        if (!zsa) {
-                struct pipe_stencil_state default_stencil = {
-                        .enabled = 0,
-                        .func = PIPE_FUNC_ALWAYS,
-                        .fail_op = PIPE_STENCIL_OP_KEEP,
-                        .zfail_op = PIPE_STENCIL_OP_KEEP,
-                        .zpass_op = PIPE_STENCIL_OP_KEEP,
-                        .writemask = 0xFF,
-                        .valuemask = 0xFF
-                };
-
-                panfrost_make_stencil_state(&default_stencil,
-                                            &fragmeta->stencil_front);
-                fragmeta->stencil_mask_front = default_stencil.writemask;
-                fragmeta->stencil_back = fragmeta->stencil_front;
-                fragmeta->stencil_mask_back = default_stencil.writemask;
+        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,
-                        zsa->stencil[0].enabled);
-                panfrost_make_stencil_state(&zsa->stencil[0],
-                                            &fragmeta->stencil_front);
-                fragmeta->stencil_mask_front = zsa->stencil[0].writemask;
+                        so->base.stencil[0].enabled);
+
+                fragmeta->stencil_mask_front = so->stencil_mask_front;
+                fragmeta->stencil_mask_back = so->stencil_mask_back;
 
-                /* Bottom 8-bits of stencil state is the stencil ref, ref is no
-                 * more than 8-bits. Be extra careful. */
-                fragmeta->stencil_front.opaque[0] |= ctx->stencil_ref.ref_value[0];
+                /* 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 back-stencil is not enabled, use the front values */
 
-                if (zsa->stencil[1].enabled) {
-                        panfrost_make_stencil_state(&zsa->stencil[1],
-                                                    &fragmeta->stencil_back);
-                        fragmeta->stencil_mask_back = zsa->stencil[1].writemask;
-                        fragmeta->stencil_back.opaque[0] |= ctx->stencil_ref.ref_value[1];
-                } else {
+                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;
-                        fragmeta->stencil_mask_back = fragmeta->stencil_mask_front;
-                }
 
-                if (zsa->depth.enabled)
-                        zfunc = zsa->depth.func;
+                if (so->base.depth.enabled)
+                        zfunc = so->base.depth.func;
 
                 /* Depth state (TODO: Refactor) */
 
                 SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_WRITEMASK,
-                        zsa->depth.writemask);
+                        so->base.depth.writemask);
         }
 
         fragmeta->unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
@@ -840,7 +738,7 @@ panfrost_frag_shader_meta_init(struct panfrost_context *ctx,
         fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
 
         bool msaa = ctx->rasterizer && ctx->rasterizer->base.multisample;
-        fragmeta->coverage_mask = (msaa ? ctx->sample_mask : ~0) & 0xF;
+        fragmeta->coverage_mask = msaa ? ctx->sample_mask : ~0;
 
         fragmeta->unknown2_3 = MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS) | 0x10;
         fragmeta->unknown2_4 = 0x4e0;
@@ -878,7 +776,7 @@ panfrost_frag_shader_meta_init(struct panfrost_context *ctx,
                  * depends on if depth/stencil is used for the draw or not.
                  * Just one of depth OR stencil is enough to trigger this. */
 
-                const struct pipe_depth_stencil_alpha_state *zsa = ctx->depth_stencil;
+                const struct pipe_depth_stencil_alpha_state *zsa = &ctx->depth_stencil->base;
                 bool zs_enabled = fs->writes_depth || fs->writes_stencil;
 
                 if (zsa) {
@@ -1371,16 +1269,20 @@ panfrost_emit_texture_descriptors(struct panfrost_batch *batch,
                 return;
 
         if (device->quirks & IS_BIFROST) {
-                struct bifrost_texture_descriptor *descriptors;
+                struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool,
+                                MALI_BIFROST_TEXTURE_LENGTH *
+                                ctx->sampler_view_count[stage]);
 
-                descriptors = malloc(sizeof(struct bifrost_texture_descriptor) *
-                                     ctx->sampler_view_count[stage]);
+                struct mali_bifrost_texture_packed *out =
+                        (struct mali_bifrost_texture_packed *) T.cpu;
 
                 for (int i = 0; i < ctx->sampler_view_count[stage]; ++i) {
                         struct panfrost_sampler_view *view = ctx->sampler_views[stage][i];
                         struct pipe_sampler_view *pview = &view->base;
                         struct panfrost_resource *rsrc = pan_resource(pview->texture);
+
                         panfrost_update_sampler_view(view, &ctx->base);
+                        out[i] = view->bifrost_descriptor;
 
                         /* Add the BOs to the job so they are retained until the job is done. */
 
@@ -1391,16 +1293,9 @@ panfrost_emit_texture_descriptors(struct panfrost_batch *batch,
                         panfrost_batch_add_bo(batch, view->bo,
                                               PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
                                               panfrost_bo_access_for_stage(stage));
-
-                        memcpy(&descriptors[i], view->bifrost_descriptor, sizeof(*view->bifrost_descriptor));
                 }
 
-                postfix->textures = panfrost_pool_upload(&batch->pool,
-                                                              descriptors,
-                                                              sizeof(struct bifrost_texture_descriptor) *
-                                                                      ctx->sampler_view_count[stage]);
-
-                free(descriptors);
+                postfix->textures = T.gpu;
         } else {
                 uint64_t trampolines[PIPE_MAX_SHADER_SAMPLER_VIEWS];
 
@@ -1425,51 +1320,21 @@ panfrost_emit_sampler_descriptors(struct panfrost_batch *batch,
                                   struct mali_vertex_tiler_postfix *postfix)
 {
         struct panfrost_context *ctx = batch->ctx;
-        struct panfrost_device *device = pan_device(ctx->base.screen);
 
         if (!ctx->sampler_count[stage])
                 return;
 
-        if (device->quirks & IS_BIFROST) {
-                size_t desc_size = sizeof(struct bifrost_sampler_descriptor);
-                size_t transfer_size = desc_size * ctx->sampler_count[stage];
-                struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool,
-                                                                                transfer_size);
-                struct bifrost_sampler_descriptor *desc = (struct bifrost_sampler_descriptor *)transfer.cpu;
+        size_t desc_size = MALI_BIFROST_SAMPLER_LENGTH;
+        assert(MALI_BIFROST_SAMPLER_LENGTH == MALI_MIDGARD_SAMPLER_LENGTH);
 
-                for (int i = 0; i < ctx->sampler_count[stage]; ++i)
-                        desc[i] = ctx->samplers[stage][i]->bifrost_hw;
+        size_t sz = desc_size * ctx->sampler_count[stage];
+        struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool, sz);
+        struct mali_midgard_sampler_packed *out = (struct mali_midgard_sampler_packed *) T.cpu;
 
-                postfix->sampler_descriptor = transfer.gpu;
-        } else {
-                size_t desc_size = sizeof(struct mali_sampler_descriptor);
-                size_t transfer_size = desc_size * ctx->sampler_count[stage];
-                struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool,
-                                                                                transfer_size);
-                struct mali_sampler_descriptor *desc = (struct mali_sampler_descriptor *)transfer.cpu;
+        for (unsigned i = 0; i < ctx->sampler_count[stage]; ++i)
+                out[i] = ctx->samplers[stage][i]->hw;
 
-                for (int i = 0; i < ctx->sampler_count[stage]; ++i)
-                        desc[i] = ctx->samplers[stage][i]->midgard_hw;
-
-                postfix->sampler_descriptor = transfer.gpu;
-        }
-}
-
-void
-panfrost_emit_vertex_attr_meta(struct panfrost_batch *batch,
-                               struct mali_vertex_tiler_postfix *vertex_postfix)
-{
-        struct panfrost_context *ctx = batch->ctx;
-
-        if (!ctx->vertex)
-                return;
-
-        struct panfrost_vertex_state *so = ctx->vertex;
-
-        panfrost_vertex_state_upd_attr_offs(ctx, vertex_postfix);
-        vertex_postfix->attribute_meta = panfrost_pool_upload(&batch->pool, so->hw,
-                                                               sizeof(*so->hw) *
-                                                               PAN_MAX_ATTRIBUTE);
+        postfix->sampler_descriptor = T.gpu;
 }
 
 void
@@ -1483,6 +1348,7 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch,
          * 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];
+        unsigned attrib_to_buffer[PIPE_MAX_ATTRIBS] = { 0 };
         unsigned k = 0;
 
         for (unsigned i = 0; i < so->num_elements; ++i) {
@@ -1493,11 +1359,7 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch,
 
                 struct pipe_vertex_element *elem = &so->pipe[i];
                 unsigned vbi = elem->vertex_buffer_index;
-
-                /* The exception to 1:1 mapping is that we can have multiple
-                 * entries (NPOT divisors), so we fixup anyways */
-
-                so->hw[i].index = k;
+                attrib_to_buffer[i] = k;
 
                 if (!(ctx->vb_mask & (1 << vbi)))
                         continue;
@@ -1569,15 +1431,68 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch,
 
         /* Add special gl_VertexID/gl_InstanceID buffers */
 
+        struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool,
+                        MALI_ATTRIBUTE_LENGTH * (PAN_INSTANCE_ID + 1));
+
+        struct mali_attribute_packed *out =
+                (struct mali_attribute_packed *) T.cpu;
+
         panfrost_vertex_id(ctx->padded_count, &attrs[k]);
-        so->hw[PAN_VERTEX_ID].index = k++;
+
+        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, &attrs[k]);
-        so->hw[PAN_INSTANCE_ID].index = k++;
 
-        /* Upload whatever we emitted and go */
+        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:
+         *
+         *      base' = base & ~63 = base - (base & 63)
+         *      offset' = offset + (base & 63)
+         *
+         * Since base' + offset' = base + offset, these are equivalent
+         * addressing modes and now base is 64 aligned.
+         */
+
+        unsigned start = vertex_postfix->offset_start;
+
+        for (unsigned i = 0; i < so->num_elements; ++i) {
+                unsigned vbi = so->pipe[i].vertex_buffer_index;
+                struct pipe_vertex_buffer *buf = &ctx->vertex_buffers[vbi];
+
+                /* Adjust by the masked off bits of the offset. Make sure we
+                 * read src_offset from so->hw (which is not GPU visible)
+                 * rather than target (which is) due to caching effects */
+
+                unsigned src_offset = so->pipe[i].src_offset;
+
+                /* BOs aligned to 4k so guaranteed aligned to 64 */
+                src_offset += (buf->buffer_offset & 63);
+
+                /* Also, somewhat obscurely per-instance data needs to be
+                 * offset in response to a delayed start in an indexed draw */
+
+                if (so->pipe[i].instance_divisor && ctx->instance_count > 1 && start)
+                        src_offset -= buf->stride * start;
+
+                pan_pack(out + i, ATTRIBUTE, cfg) {
+                        cfg.buffer_index = attrib_to_buffer[i];
+                        cfg.format = so->formats[i];
+                        cfg.offset = src_offset;
+                }
+        }
+
 
         vertex_postfix->attributes = panfrost_pool_upload(&batch->pool, attrs,
                                                            k * sizeof(*attrs));
+
+        vertex_postfix->attribute_meta = T.gpu;
 }
 
 static mali_ptr
@@ -1775,14 +1690,14 @@ pan_emit_vary(unsigned present, enum pan_special_varying buf,
                 unsigned offset)
 {
         unsigned nr_channels = MALI_EXTRACT_CHANNELS(format);
+        unsigned swizzle = quirks & HAS_SWIZZLES ?
+                        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,
-                .swizzle = quirks & HAS_SWIZZLES ?
-                        panfrost_get_default_swizzle(nr_channels) :
-                        panfrost_bifrost_swizzle(nr_channels),
-                .format = format,
+                .format = (format << 12) | swizzle,
                 .src_offset = offset
         };
 
@@ -1836,6 +1751,10 @@ pan_emit_vary_xfb(unsigned present,
                 enum mali_format format,
                 struct pipe_stream_output o)
 {
+        unsigned swizzle = quirks & HAS_SWIZZLES ?
+                        panfrost_get_default_swizzle(o.num_components) :
+                        panfrost_bifrost_swizzle(o.num_components);
+
         /* Otherwise construct a record for it */
         struct mali_attr_meta meta = {
                 /* XFB buffers come after everything else */
@@ -1844,13 +1763,8 @@ pan_emit_vary_xfb(unsigned present,
                 /* As usual unknown bit */
                 .unknown1 = quirks & IS_BIFROST ? 0x0 : 0x2,
 
-                /* Override swizzle with number of channels */
-                .swizzle = quirks & HAS_SWIZZLES ?
-                        panfrost_get_default_swizzle(o.num_components) :
-                        panfrost_bifrost_swizzle(o.num_components),
-
                 /* Override number of channels and precision to highp */
-                .format = pan_xfb_format(format, o.num_components),
+                .format = (pan_xfb_format(format, o.num_components) << 12) | swizzle,
 
                 /* Apply given offsets together */
                 .src_offset = (o.dst_offset * 4) /* dwords */