X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fpanfrost%2Fpandecode%2Fdecode.c;h=148be10fdc47bd5a95397825821d8e3c00bc2c82;hb=be5e30c46be9c6cdfb0a0209f5db118bafcc4c85;hp=2df3cbdc98675e2c5ad2d1f60bc9554f2729777d;hpb=3e47a1181b7c5bc962380553c8c99a49a49f45d9;p=mesa.git diff --git a/src/panfrost/pandecode/decode.c b/src/panfrost/pandecode/decode.c index 2df3cbdc986..148be10fdc4 100644 --- a/src/panfrost/pandecode/decode.c +++ b/src/panfrost/pandecode/decode.c @@ -29,13 +29,17 @@ #include #include #include +#include #include "decode.h" +#include "util/macros.h" #include "util/u_math.h" #include "pan_pretty_print.h" #include "midgard/disassemble.h" #include "bifrost/disassemble.h" +#include "pan_encoder.h" + int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost); #define MEMORY_PROP(obj, p) {\ @@ -54,11 +58,6 @@ int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost); } \ } -#define DYN_MEMORY_PROP(obj, no, p) { \ - if (obj->p) \ - pandecode_prop("%s = %s_%d_p", #p, #p, no); \ -} - /* Semantic logging type. * * Raw: for raw messages to be printed as is. @@ -117,6 +116,44 @@ pandecode_log_cont(const char *format, ...) va_end(ap); } +/* To check for memory safety issues, validates that the given pointer in GPU + * memory is valid, containing at least sz bytes. The goal is to eliminate + * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer + * overruns) by statically validating pointers. + */ + +static void +pandecode_validate_buffer(mali_ptr addr, size_t sz) +{ + if (!addr) { + pandecode_msg("XXX: null pointer deref"); + return; + } + + /* Find a BO */ + + struct pandecode_mapped_memory *bo = + pandecode_find_mapped_gpu_mem_containing(addr); + + if (!bo) { + pandecode_msg("XXX: invalid memory dereference\n"); + return; + } + + /* Bounds check */ + + unsigned offset = addr - bo->gpu_va; + unsigned total = offset + sz; + + if (total > bo->length) { + pandecode_msg("XXX: buffer overrun." + "Chunk of size %d at offset %d in buffer of size %d. " + "Overrun by %d bytes.", + sz, offset, bo->length, total - bo->length); + return; + } +} + struct pandecode_flag_info { u64 flag; const char *name; @@ -304,7 +341,8 @@ pandecode_draw_mode(enum mali_draw_mode mode) DEFINE_CASE(QUAD_STRIP); default: - return "MALI_TRIANGLES /* XXX: Unknown GL mode, check dump */"; + pandecode_msg("XXX: invalid draw mode %X\n", mode); + return ""; } #undef DEFINE_CASE @@ -325,7 +363,8 @@ pandecode_func(enum mali_func mode) DEFINE_CASE(ALWAYS); default: - return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */"; + pandecode_msg("XXX: invalid func %X\n", mode); + return ""; } } #undef DEFINE_CASE @@ -346,7 +385,8 @@ pandecode_alt_func(enum mali_alt_func mode) DEFINE_CASE(ALWAYS); default: - return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */"; + pandecode_msg("XXX: invalid alt func %X\n", mode); + return ""; } } #undef DEFINE_CASE @@ -366,7 +406,8 @@ pandecode_stencil_op(enum mali_stencil_op op) DEFINE_CASE(DECR); default: - return "MALI_STENCIL_KEEP /* XXX: Unknown stencil op, check dump */"; + pandecode_msg("XXX: invalid stencil op %X\n", op); + return ""; } } @@ -382,33 +423,15 @@ static char *pandecode_attr_mode(enum mali_attr_mode mode) DEFINE_CASE(MODULO); DEFINE_CASE(NPOT_DIVIDE); DEFINE_CASE(IMAGE); + DEFINE_CASE(INTERNAL); default: - return "MALI_ATTR_UNUSED /* XXX: Unknown stencil op, check dump */"; + pandecode_msg("XXX: invalid attribute mode %X\n", mode); + return ""; } } #undef DEFINE_CASE -#define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name -static char * -pandecode_channel(enum mali_channel channel) -{ - switch (channel) { - DEFINE_CASE(RED); - DEFINE_CASE(GREEN); - DEFINE_CASE(BLUE); - DEFINE_CASE(ALPHA); - DEFINE_CASE(ZERO); - DEFINE_CASE(ONE); - DEFINE_CASE(RESERVED_0); - DEFINE_CASE(RESERVED_1); - - default: - return "MALI_CHANNEL_ZERO /* XXX: Unknown channel, check dump */"; - } -} -#undef DEFINE_CASE - #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name static char * pandecode_wrap_mode(enum mali_wrap_mode op) @@ -420,36 +443,37 @@ pandecode_wrap_mode(enum mali_wrap_mode op) DEFINE_CASE(MIRRORED_REPEAT); default: - return "MALI_WRAP_REPEAT /* XXX: Unknown wrap mode, check dump */"; + pandecode_msg("XXX: invalid wrap mode %X\n", op); + return ""; } } #undef DEFINE_CASE -#define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name +#define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name static char * -pandecode_texture_type(enum mali_texture_type type) +pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt) { - switch (type) { - DEFINE_CASE(1D); - DEFINE_CASE(2D); - DEFINE_CASE(3D); - DEFINE_CASE(CUBE); + switch (fmt) { + DEFINE_CASE(TILED); + DEFINE_CASE(UNKNOWN); + DEFINE_CASE(LINEAR); + DEFINE_CASE(AFBC); default: - unreachable("Unknown case"); + unreachable("Invalid case"); } } #undef DEFINE_CASE -#define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name +#define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name static char * -pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt) +pandecode_exception_access(enum mali_exception_access access) { - switch (fmt) { - DEFINE_CASE(TILED); - DEFINE_CASE(UNKNOWN); - DEFINE_CASE(LINEAR); - DEFINE_CASE(AFBC); + switch (access) { + DEFINE_CASE(NONE); + DEFINE_CASE(EXECUTE); + DEFINE_CASE(READ); + DEFINE_CASE(WRITE); default: unreachable("Invalid case"); @@ -461,30 +485,107 @@ pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt) * larger FBD */ static void -pandecode_midgard_tiler_descriptor(const struct midgard_tiler_descriptor *t) +pandecode_midgard_tiler_descriptor( + const struct midgard_tiler_descriptor *t, + unsigned width, + unsigned height, + bool is_fragment) { pandecode_log(".tiler = {\n"); pandecode_indent++; - pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask); - pandecode_prop("flags = 0x%" PRIx16, t->flags); - pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size); + if (t->hierarchy_mask == MALI_TILER_DISABLED) + pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED"); + else + pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask); + + /* We know this name from the kernel, but we never see it nonzero */ + + if (t->flags) + pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags); MEMORY_PROP(t, polygon_list); - MEMORY_PROP(t, polygon_list_body); + + /* The body is offset from the base of the polygon list */ + assert(t->polygon_list_body > t->polygon_list); + unsigned body_offset = t->polygon_list_body - t->polygon_list; + + /* It needs to fit inside the reported size */ + assert(t->polygon_list_size >= body_offset); + + /* Check that we fit */ + struct pandecode_mapped_memory *plist = + pandecode_find_mapped_gpu_mem_containing(t->polygon_list); + + assert(t->polygon_list_size <= plist->length); + + /* Now that we've sanity checked, we'll try to calculate the sizes + * ourselves for comparison */ + + unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask); + unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask); + + if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) { + pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n", + ref_header, ref_size); + pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size); + pandecode_msg("body offset %d\n", body_offset); + } + + /* The tiler heap has a start and end specified -- it should be + * identical to what we have in the BO. The exception is if tiling is + * disabled. */ MEMORY_PROP(t, heap_start); + assert(t->heap_end >= t->heap_start); + + struct pandecode_mapped_memory *heap = + pandecode_find_mapped_gpu_mem_containing(t->heap_start); + + unsigned heap_size = t->heap_end - t->heap_start; - if (t->heap_start == t->heap_end) { - /* Print identically to show symmetry for empty tiler heaps */ - MEMORY_PROP(t, heap_start); + /* Tiling is enabled with a special flag */ + unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK; + unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask; + + bool tiling_enabled = hierarchy_mask; + + if (tiling_enabled) { + /* When tiling is enabled, the heap should be a tight fit */ + unsigned heap_offset = t->heap_start - heap->gpu_va; + if ((heap_offset + heap_size) != heap->length) { + pandecode_msg("XXX: heap size %d (expected %d)\n", + heap_size, heap->length - heap_offset); + } + + /* We should also have no other flags */ + if (tiler_flags) + pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags); } else { - /* Points to the end of a buffer */ - char *a = pointer_as_memory_reference(t->heap_end - 1); - pandecode_prop("heap_end = %s + 1", a); - free(a); + /* When tiling is disabled, we should have that flag and no others */ + + if (tiler_flags != MALI_TILER_DISABLED) { + pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n", + tiler_flags); + } + + /* We should also have an empty heap */ + if (heap_size) { + pandecode_msg("XXX: tiler heap size %d given, expected empty\n", + heap_size); + } + + /* Disabled tiling is used only for clear-only jobs, which are + * purely FRAGMENT, so we should never see this for + * non-FRAGMENT descriptors. */ + + if (!is_fragment) + pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n"); } + /* We've never seen weights used in practice, but we know from the + * kernel these fields is there */ + bool nonzero_weights = false; for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) { @@ -506,7 +607,7 @@ pandecode_midgard_tiler_descriptor(const struct midgard_tiler_descriptor *t) } static void -pandecode_sfbd(uint64_t gpu_va, int job_no) +pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment) { struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va); const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va); @@ -564,7 +665,7 @@ pandecode_sfbd(uint64_t gpu_va, int job_no) MEMORY_PROP(s, unknown_address_0); const struct midgard_tiler_descriptor t = s->tiler; - pandecode_midgard_tiler_descriptor(&t); + pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment); pandecode_indent--; pandecode_log("};\n"); @@ -618,14 +719,126 @@ pandecode_compute_fbd(uint64_t gpu_va, int job_no) printf("},\n"); } +/* Extracts the number of components associated with a Mali format */ + +static unsigned +pandecode_format_component_count(enum mali_format fmt) +{ + /* Mask out the format class */ + unsigned top = fmt & 0b11100000; + + switch (top) { + case MALI_FORMAT_SNORM: + case MALI_FORMAT_UINT: + case MALI_FORMAT_UNORM: + case MALI_FORMAT_SINT: + return ((fmt >> 3) & 3) + 1; + default: + /* TODO: Validate */ + return 4; + } +} + +/* Extracts a mask of accessed components from a 12-bit Mali swizzle */ + +static unsigned +pandecode_access_mask_from_channel_swizzle(unsigned swizzle) +{ + unsigned mask = 0; + assert(MALI_CHANNEL_RED == 0); + + for (unsigned c = 0; c < 4; ++c) { + enum mali_channel chan = (swizzle >> (3*c)) & 0x7; + + if (chan <= MALI_CHANNEL_ALPHA) + mask |= (1 << chan); + } + + return mask; +} + +/* Validates that a (format, swizzle) pair is valid, in the sense that the + * swizzle doesn't access any components that are undefined in the format. + * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be + * omitted */ + +static bool +pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle) +{ + unsigned nr_comp = pandecode_format_component_count(fmt); + unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle); + unsigned valid_mask = (1 << nr_comp) - 1; + unsigned invalid_mask = ~valid_mask; + + if (access_mask & invalid_mask) { + pandecode_msg("XXX: invalid components accessed\n"); + return false; + } + + /* Check for the default non-swizzling swizzle so we can suppress + * useless printing for the defaults */ + + unsigned default_swizzles[4] = { + MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9), + MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9), + MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9), + MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9) + }; + + return (swizzle == default_swizzles[nr_comp - 1]); +} + +/* Maps MALI_RGBA32F to rgba32f, etc */ + static void -pandecode_swizzle(unsigned swizzle) +pandecode_format_short(enum mali_format fmt, bool srgb) { - pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)", - pandecode_channel((swizzle >> 0) & 0x7), - pandecode_channel((swizzle >> 3) & 0x7), - pandecode_channel((swizzle >> 6) & 0x7), - pandecode_channel((swizzle >> 9) & 0x7)); + /* We want a type-like format, so cut off the initial MALI_ */ + char *format = pandecode_format(fmt); + format += strlen("MALI_"); + + unsigned len = strlen(format); + char *lower_format = calloc(1, len + 1); + + for (unsigned i = 0; i < len; ++i) + lower_format[i] = tolower(format[i]); + + /* Sanity check sRGB flag is applied to RGB, per the name */ + if (srgb && lower_format[0] != 'r') + pandecode_msg("XXX: sRGB applied to non-colour format\n"); + + /* Just prefix with an s, so you get formats like srgba8_unorm */ + if (srgb) + pandecode_log_cont("s"); + + pandecode_log_cont("%s", lower_format); + free(lower_format); +} + +static void +pandecode_swizzle(unsigned swizzle, enum mali_format format) +{ + /* First, do some validation */ + bool trivial_swizzle = pandecode_validate_format_swizzle( + format, swizzle); + + if (trivial_swizzle) + return; + + /* Next, print the swizzle */ + pandecode_log_cont("."); + + static const char components[] = "rgba01"; + + for (unsigned c = 0; c < 4; ++c) { + enum mali_channel chan = (swizzle >> (3 * c)) & 0x7; + + if (chan >= MALI_CHANNEL_RESERVED_0) { + pandecode_log("XXX: invalid swizzle channel %d\n", chan); + continue; + } + pandecode_log_cont("%c", components[chan]); + } } static void @@ -641,6 +854,11 @@ pandecode_rt_format(struct mali_rt_format format) pandecode_prop("block = %s", pandecode_mfbd_block_format(format.block)); + /* TODO: Map formats so we can check swizzles and print nicely */ + pandecode_log("swizzle"); + pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM); + pandecode_log_cont(",\n"); + pandecode_prop("nr_channels = MALI_POSITIVE(%d)", MALI_NEGATIVE(format.nr_channels)); @@ -648,9 +866,19 @@ pandecode_rt_format(struct mali_rt_format format) pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags); pandecode_log_cont(",\n"); - pandecode_swizzle(format.swizzle); + /* In theory, the no_preload bit can be cleared to enable MFBD preload, + * which is a faster hardware-based alternative to the wallpaper method + * to preserve framebuffer contents across frames. In practice, MFBD + * preload is buggy on Midgard, and so this is a chicken bit. If this + * bit isn't set, most likely something broke unrelated to preload */ - pandecode_prop("unk4 = 0x%" PRIx32, format.unk4); + if (!format.no_preload) { + pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n"); + pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload); + } + + if (format.zero) + pandecode_prop("zero = 0x%" PRIx32, format.zero); pandecode_indent--; pandecode_log("},\n"); @@ -686,18 +914,11 @@ pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_f pandecode_indent--; pandecode_log("},\n"); - } else { - pandecode_log(".chunknown = {\n"); - pandecode_indent++; - - pandecode_prop("unk = 0x%" PRIx64, rt->chunknown.unk); - - char *a = pointer_as_memory_reference(rt->chunknown.pointer); - pandecode_prop("pointer = %s", a); - free(a); - - pandecode_indent--; - pandecode_log("},\n"); + } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) { + pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n", + rt->afbc.metadata, + rt->afbc.stride, + rt->afbc.unk); } MEMORY_PROP(rt, framebuffer); @@ -711,7 +932,7 @@ pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_f } if (rt->zero1 || rt->zero2 || rt->zero3) { - pandecode_msg("render target zeros tripped\n"); + pandecode_msg("XXX: render target zeros tripped\n"); pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1); pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2); pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3); @@ -726,7 +947,7 @@ pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_f } static unsigned -pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) +pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment) { struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va); const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va); @@ -784,16 +1005,23 @@ pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags); pandecode_log_cont(",\n"); - pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil); - pandecode_prop("clear_depth = %f", fb->clear_depth); + if (fb->clear_stencil) + pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil); + + if (fb->clear_depth) + pandecode_prop("clear_depth = %f", fb->clear_depth); + + /* TODO: What is this? Let's not blow up.. */ + if (fb->unknown2 != 0x1F) + pandecode_prop("unknown2 = 0x%x", fb->unknown2); pandecode_prop("unknown2 = 0x%x", fb->unknown2); MEMORY_PROP(fb, scratchpad); const struct midgard_tiler_descriptor t = fb->tiler; - pandecode_midgard_tiler_descriptor(&t); + pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment); if (fb->zero3 || fb->zero4) { - pandecode_msg("framebuffer zeros tripped\n"); + pandecode_msg("XXX: framebuffer zeros tripped\n"); pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3); pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4); } @@ -803,7 +1031,7 @@ pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) gpu_va += sizeof(struct bifrost_framebuffer); - if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && with_render_targets) { + if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) { mem = pandecode_find_mapped_gpu_mem_containing(gpu_va); const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va); @@ -829,7 +1057,7 @@ pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil); if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) { - pandecode_msg("Depth/stencil AFBC zeros tripped\n"); + pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n"); pandecode_prop("zero1 = 0x%" PRIx32, fbx->ds_afbc.zero1); pandecode_prop("padding = 0x%" PRIx64, @@ -857,7 +1085,7 @@ pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) if (fbx->ds_linear.depth_stride_zero || fbx->ds_linear.stencil_stride_zero || fbx->ds_linear.zero1 || fbx->ds_linear.zero2) { - pandecode_msg("Depth/stencil zeros tripped\n"); + pandecode_msg("XXX: Depth/stencil zeros tripped\n"); pandecode_prop("depth_stride_zero = 0x%x", fbx->ds_linear.depth_stride_zero); pandecode_prop("stencil_stride_zero = 0x%x", @@ -873,7 +1101,7 @@ pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) } if (fbx->zero3 || fbx->zero4) { - pandecode_msg("fb_extra zeros tripped\n"); + pandecode_msg("XXX: fb_extra zeros tripped\n"); pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3); pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4); } @@ -884,7 +1112,7 @@ pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets) gpu_va += sizeof(struct bifrost_fb_extra); } - if (with_render_targets) + if (is_fragment) pandecode_render_target(gpu_va, job_no, fb); /* Passback the render target count */ @@ -994,37 +1222,30 @@ pandecode_attributes(const struct pandecode_mapped_memory *mem, { char *prefix = varying ? "varyings" : "attributes"; + if (!addr) { + pandecode_msg("no %s\n", prefix); + return; + } + union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count); - char base[128]; - snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix); + pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no); + pandecode_indent++; for (int i = 0; i < count; ++i) { + pandecode_log("{\n"); + pandecode_indent++; + enum mali_attr_mode mode = attr[i].elements & 7; if (mode == MALI_ATTR_UNUSED) - continue; + pandecode_msg("XXX: unused attribute record\n"); mali_ptr raw_elements = attr[i].elements & ~7; - - /* TODO: Do we maybe want to dump the attribute values - * themselves given the specified format? Or is that too hard? - * */ - char *a = pointer_as_memory_reference(raw_elements); - pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a); + pandecode_prop("elements = (%s) | %s", a, pandecode_attr_mode(mode)); free(a); - } - - pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no); - pandecode_indent++; - - for (int i = 0; i < count; ++i) { - pandecode_log("{\n"); - pandecode_indent++; - unsigned mode = attr[i].elements & 7; - pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode(mode)); pandecode_prop("shift = %d", attr[i].shift); pandecode_prop("extra_flags = %d", attr[i].extra_flags); pandecode_prop("stride = 0x%" PRIx32, attr[i].stride); @@ -1048,7 +1269,7 @@ pandecode_attributes(const struct pandecode_mapped_memory *mem, pandecode_prop("unk = 0x%x", attr[i].unk); pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor); if (attr[i].zero != 0) - pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr[i].zero); + pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero); pandecode_prop("divisor = %d", attr[i].divisor); pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags); pandecode_indent--; @@ -1097,7 +1318,7 @@ pandecode_stencil(const char *name, const struct mali_stencil_test *stencil) const char *dppass = pandecode_stencil_op(stencil->dppass); if (stencil->zero) - pandecode_msg("Stencil zero tripped: %X\n", stencil->zero); + pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero); pandecode_log(".stencil_%s = {\n", name); pandecode_indent++; @@ -1115,7 +1336,7 @@ static void pandecode_blend_equation(const struct mali_blend_equation *blend) { if (blend->zero1) - pandecode_msg("Blend zero tripped: %X\n", blend->zero1); + pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1); pandecode_log(".equation = {\n"); pandecode_indent++; @@ -1213,6 +1434,11 @@ pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no) return shader; } +/* Attributes and varyings have descriptor records, which contain information + * about their format and ordering with the attribute/varying buffers. We'll + * want to validate that the combinations specified are self-consistent. + */ + static int pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix) { @@ -1233,27 +1459,67 @@ pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_p attr_meta = pandecode_fetch_gpu_mem(attr_mem, p, sizeof(*attr_mem)); - pandecode_log("{\n"); - pandecode_indent++; - pandecode_prop("index = %d", attr_meta->index); + /* If the record is discard, it should be zero for everything else */ + + if (attr_meta->format == MALI_VARYING_DISCARD) { + uint64_t zero = + attr_meta->index | + attr_meta->unknown1 | + attr_meta->unknown3 | + attr_meta->src_offset; + + if (zero) + pandecode_msg("XXX: expected empty record for varying discard\n"); + + /* We want to look for a literal 0000 swizzle -- this + * is not encoded with all zeroes, however */ + + enum mali_channel z = MALI_CHANNEL_ZERO; + unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9); + bool good_swizzle = attr_meta->swizzle == zero_swizzle; + + if (!good_swizzle) + pandecode_msg("XXX: expected zero swizzle for discard\n"); + + if (!varying) + pandecode_msg("XXX: cannot discard attribute\n"); + + /* If we're all good, omit the record */ + if (!zero && varying && good_swizzle) { + pandecode_log("/* discarded varying */\n"); + continue; + } + } if (attr_meta->index > max_index) max_index = attr_meta->index; - pandecode_swizzle(attr_meta->swizzle); - pandecode_prop("format = %s", pandecode_format(attr_meta->format)); - pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1); - pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3); - pandecode_prop("src_offset = %d", attr_meta->src_offset); - pandecode_indent--; - pandecode_log("},\n"); + if (attr_meta->unknown1 != 0x2) { + pandecode_msg("XXX: expected unknown1 = 0x2\n"); + pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1); + } + + if (attr_meta->unknown3) { + pandecode_msg("XXX: unexpected unknown3 set\n"); + pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3); + } + pandecode_make_indent(); + pandecode_format_short(attr_meta->format, false); + pandecode_log_cont(" %s_%u", prefix, attr_meta->index); + + if (attr_meta->src_offset) + pandecode_log_cont("[%u]", attr_meta->src_offset); + + pandecode_swizzle(attr_meta->swizzle, attr_meta->format); + + pandecode_log_cont(";\n"); } pandecode_indent--; pandecode_log("};\n"); - return max_index; + return count ? (max_index + 1) : 0; } static void @@ -1291,34 +1557,67 @@ bits(u32 word, u32 lo, u32 hi) } static void -pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no) +pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced) { pandecode_log_cont("{\n"); pandecode_indent++; - pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count); - pandecode_prop("size_y_shift = %d", p->size_y_shift); - pandecode_prop("size_z_shift = %d", p->size_z_shift); - pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift); - pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift); - pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift); - pandecode_prop("workgroups_x_shift_2 = 0x%" PRIx32, p->workgroups_x_shift_2); - /* Decode invocation_count. See the comment before the definition of * invocation_count for an explanation. */ - pandecode_msg("size: (%d, %d, %d)\n", - bits(p->invocation_count, 0, p->size_y_shift) + 1, - bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1, - bits(p->invocation_count, p->size_z_shift, - p->workgroups_x_shift) + 1); - pandecode_msg("workgroups: (%d, %d, %d)\n", - bits(p->invocation_count, p->workgroups_x_shift, - p->workgroups_y_shift) + 1, - bits(p->invocation_count, p->workgroups_y_shift, - p->workgroups_z_shift) + 1, - bits(p->invocation_count, p->workgroups_z_shift, - 32) + 1); + + unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1; + unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1; + unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1; + + unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1; + unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1; + unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1; + + /* Even though we have this decoded, we want to ensure that the + * representation is "unique" so we don't lose anything by printing only + * the final result. More specifically, we need to check that we were + * passed something in canonical form, since the definition per the + * hardware is inherently not unique. How? Well, take the resulting + * decode and pack it ourselves! If it is bit exact with what we + * decoded, we're good to go. */ + + struct mali_vertex_tiler_prefix ref; + panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced); + + bool canonical = + (p->invocation_count == ref.invocation_count) && + (p->size_y_shift == ref.size_y_shift) && + (p->size_z_shift == ref.size_z_shift) && + (p->workgroups_x_shift == ref.workgroups_x_shift) && + (p->workgroups_y_shift == ref.workgroups_y_shift) && + (p->workgroups_z_shift == ref.workgroups_z_shift) && + (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2); + + if (!canonical) { + pandecode_msg("XXX: non-canonical workgroups packing\n"); + pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n", + ref.invocation_count, + ref.size_y_shift, + ref.size_z_shift, + ref.workgroups_x_shift, + ref.workgroups_y_shift, + ref.workgroups_z_shift, + ref.workgroups_x_shift_2); + + pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count); + pandecode_prop("size_y_shift = %d", p->size_y_shift); + pandecode_prop("size_z_shift = %d", p->size_z_shift); + pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift); + pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift); + pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift); + pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2); + } + + /* Regardless, print the decode */ + pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n", + size_x, size_y, size_z, + groups_x, groups_y, groups_z); /* TODO: Decode */ if (p->unknown_draw) @@ -1326,22 +1625,19 @@ pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no) pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3); - pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode)); + if (p->draw_mode != MALI_DRAW_NONE) + pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode)); /* Index count only exists for tiler jobs anyway */ if (p->index_count) pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1); - if (p->negative_start) - pandecode_prop("negative_start = %d", p->negative_start); + if (p->offset_bias_correction) + pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction); - DYN_MEMORY_PROP(p, job_no, indices); - - if (p->zero1) { - pandecode_msg("Zero tripped\n"); - pandecode_prop("zero1 = 0x%" PRIx32, p->zero1); - } + /* TODO: Figure out what this is. It's not zero */ + pandecode_prop("zero1 = 0x%" PRIx32, p->zero1); pandecode_indent--; pandecode_log("},\n"); @@ -1351,38 +1647,8 @@ static void pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no) { struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs); - struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs); - for (int i = 0; i < ubufs_count; i++) { - mali_ptr ptr = ubufs[i].ptr << 2; - struct pandecode_mapped_memory *umem2 = pandecode_find_mapped_gpu_mem_containing(ptr); - uint32_t *PANDECODE_PTR_VAR(ubuf, umem2, ptr); - char name[50]; - snprintf(name, sizeof(name), "ubuf_%d", i); - /* The blob uses ubuf 0 to upload internal stuff and - * uniforms that won't fit/are accessed indirectly, so - * it puts it in the batchbuffer. - */ - pandecode_log("uint32_t %s_%d[] = {\n", name, job_no); - pandecode_indent++; - - for (int j = 0; j <= ubufs[i].size; j++) { - for (int k = 0; k < 4; k++) { - if (k == 0) - pandecode_log("0x%"PRIx32", ", ubuf[4 * j + k]); - else - pandecode_log_cont("0x%"PRIx32", ", ubuf[4 * j + k]); - - } - - pandecode_log_cont("\n"); - } - - pandecode_indent--; - pandecode_log("};\n"); - } - pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n", pubufs, job_no); pandecode_indent++; @@ -1390,10 +1656,18 @@ pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no) for (int i = 0; i < ubufs_count; i++) { pandecode_log("{\n"); pandecode_indent++; - pandecode_prop("size = MALI_POSITIVE(%d)", ubufs[i].size + 1); - pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i, job_no); + + unsigned size = (ubufs[i].size + 1) * 16; + mali_ptr addr = ubufs[i].ptr << 2; + + pandecode_validate_buffer(addr, size); + + char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2); + pandecode_prop("size = %u", size); + pandecode_prop("ptr = (%s) >> 2", ptr); pandecode_indent--; pandecode_log("},\n"); + free(ptr); } pandecode_indent--; @@ -1408,8 +1682,10 @@ pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix) struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad); - if (scratchpad->zero) - pandecode_msg("XXX scratchpad zero tripped"); + if (scratchpad->zero) { + pandecode_msg("XXX: scratchpad zero tripped"); + pandecode_prop("zero = 0x%x\n", scratchpad->zero); + } pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix); pandecode_indent++; @@ -1421,9 +1697,11 @@ pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix) pandecode_log("};\n"); } +static unsigned shader_id = 0; + static void pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type, - bool is_bifrost) + bool is_bifrost, unsigned nr_regs) { struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr); uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr); @@ -1436,17 +1714,177 @@ pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type, printf("\n\n"); + char prefix[512]; + + snprintf(prefix, sizeof(prefix) - 1, "shader%d - %s shader: ", + shader_id++, + (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX"); + if (is_bifrost) { disassemble_bifrost(code, sz, false); } else { - disassemble_midgard(code, sz); + disassemble_midgard(code, sz, true, nr_regs, prefix); } printf("\n\n"); } static void -pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, +pandecode_texture(mali_ptr u, + struct pandecode_mapped_memory *tmem, + unsigned job_no, unsigned tex) +{ + struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u); + + pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex); + pandecode_indent++; + + struct mali_texture_format f = t->format; + + /* See the definiton of enum mali_texture_type */ + + bool is_cube = f.type == MALI_TEX_CUBE; + unsigned dimension = is_cube ? 2 : f.type; + + pandecode_make_indent(); + + /* TODO: Are there others? */ + bool is_zs = f.format == MALI_Z32_UNORM; + + /* Recall Z/S switched the meaning of linear/tiled .. */ + if (is_zs && f.layout == MALI_TEXTURE_LINEAR) + pandecode_msg("XXX: depth/stencil cannot be tiled\n"); + + /* Print the layout. Default is linear; a modifier can denote AFBC or + * u-interleaved/tiled modes */ + + if (f.layout == MALI_TEXTURE_AFBC) + pandecode_log_cont("afbc"); + else if (f.layout == MALI_TEXTURE_TILED) + pandecode_log_cont(is_zs ? "linear" : "tiled"); + else if (f.layout == MALI_TEXTURE_LINEAR) + pandecode_log_cont("linear"); + else + pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout); + + pandecode_swizzle(t->swizzle, f.format); + pandecode_log_cont(" "); + + /* Distinguish cube/2D with modifier */ + + if (is_cube) + pandecode_log_cont("cube "); + + pandecode_format_short(f.format, f.srgb); + pandecode_swizzle(f.swizzle, f.format); + + /* All four width/height/depth/array_size dimensions are present + * regardless of the type of texture, but it is an error to have + * non-zero dimensions for unused dimensions. Verify this. array_size + * can always be set, as can width. */ + + if (t->height && dimension < 2) + pandecode_msg("XXX: nonzero height for <2D texture\n"); + + if (t->depth && dimension < 3) + pandecode_msg("XXX: nonzero depth for <2D texture\n"); + + /* Print only the dimensions that are actually there */ + + pandecode_log_cont(": %d", t->width + 1); + + if (dimension >= 2) + pandecode_log_cont("x%u", t->height + 1); + + if (dimension >= 3) + pandecode_log_cont("x%u", t->depth + 1); + + if (t->array_size) + pandecode_log_cont("[%u]", t->array_size + 1); + + if (t->levels) + pandecode_log_cont(" mip %u", t->levels); + + pandecode_log_cont("\n"); + + if (f.unknown1 | f.zero) { + pandecode_msg("XXX: texture format zero tripped\n"); + pandecode_prop("unknown1 = %" PRId32, f.unknown1); + pandecode_prop("zero = %" PRId32, f.zero); + } + + if (!f.unknown2) { + pandecode_msg("XXX: expected unknown texture bit set\n"); + pandecode_prop("unknown2 = %" PRId32, f.unknown1); + } + + if (t->swizzle_zero) { + pandecode_msg("XXX: swizzle zero tripped\n"); + pandecode_prop("swizzle_zero = %d", t->swizzle_zero); + } + + if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) { + pandecode_msg("XXX: texture zero tripped\n"); + pandecode_prop("unknown3 = %" PRId16, t->unknown3); + pandecode_prop("unknown3A = %" PRId8, t->unknown3A); + pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5); + pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6); + pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7); + } + + pandecode_log(".payload = {\n"); + pandecode_indent++; + + /* A bunch of bitmap pointers follow. + * We work out the correct number, + * based on the mipmap/cubemap + * properties, but dump extra + * possibilities to futureproof */ + + int bitmap_count = MALI_NEGATIVE(t->levels); + + /* Miptree for each face */ + if (f.type == MALI_TEX_CUBE) + bitmap_count *= 6; + + /* Array of textures */ + bitmap_count *= MALI_NEGATIVE(t->array_size); + + /* Stride for each element */ + if (f.manual_stride) + bitmap_count *= 2; + + /* Sanity check the size */ + int max_count = sizeof(t->payload) / sizeof(t->payload[0]); + assert (bitmap_count <= max_count); + + for (int i = 0; i < bitmap_count; ++i) { + /* How we dump depends if this is a stride or a pointer */ + + if (f.manual_stride && (i & 1)) { + /* signed 32-bit snuck in as a 64-bit pointer */ + uint64_t stride_set = t->payload[i]; + uint32_t clamped_stride = stride_set; + int32_t stride = clamped_stride; + assert(stride_set == clamped_stride); + pandecode_log("(mali_ptr) %d /* stride */, \n", stride); + } else { + char *a = pointer_as_memory_reference(t->payload[i]); + pandecode_log("%s, \n", a); + free(a); + } + } + + pandecode_indent--; + pandecode_log("},\n"); + + pandecode_indent--; + pandecode_log("};\n"); +} + +static void +pandecode_vertex_tiler_postfix_pre( + const struct mali_vertex_tiler_postfix *p, int job_no, enum mali_job_type job_type, char *suffix, bool is_bifrost) { @@ -1467,7 +1905,7 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, else if (job_type == JOB_TYPE_COMPUTE) pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no); else - pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no); + pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false); int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0; int texture_count = 0, sampler_count = 0; @@ -1489,7 +1927,7 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, uniform_count = s->bifrost2.uniform_count; uniform_buffer_count = s->bifrost1.uniform_buffer_count; } else { - uniform_count = s->midgard1.uniform_count; + uniform_count = s->midgard1.uniform_buffer_count; uniform_buffer_count = s->midgard1.uniform_buffer_count; } @@ -1500,6 +1938,8 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, pandecode_prop("attribute_count = %" PRId16, s->attribute_count); pandecode_prop("varying_count = %" PRId16, s->varying_count); + unsigned nr_registers = 0; + if (is_bifrost) { pandecode_log(".bifrost1 = {\n"); pandecode_indent++; @@ -1516,6 +1956,7 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count); pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count); pandecode_prop("work_count = %" PRId16, s->midgard1.work_count); + nr_registers = s->midgard1.work_count; pandecode_log(".flags = "); pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags); @@ -1620,12 +2061,12 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, shader = pandecode_midgard_blend_mrt(blend_base, job_no, i); if (shader & ~0xF) - pandecode_shader_disassemble(shader, job_no, job_type, false); + pandecode_shader_disassemble(shader, job_no, job_type, false, 0); } } if (shader_ptr & ~0xF) - pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost); + pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers); } else pandecode_msg("\n"); @@ -1659,81 +2100,44 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix); attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes); - pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index + 1, false); + pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false); } /* Varyings are encoded like attributes but not actually sent; we just * pass a zero buffer with the right stride/size set, (or whatever) * since the GPU will write to it itself */ + if (p->varying_meta) { + varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix); + } + if (p->varyings) { attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings); /* Number of descriptors depends on whether there are * non-internal varyings */ - pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count > 1 ? 4 : 1, true); - } - - if (p->varying_meta) { - pandecode_attribute_meta(job_no, varying_count, p, true, suffix); - } - - bool is_compute = job_type == JOB_TYPE_COMPUTE; - - if (p->uniforms && !is_compute) { - int rows = uniform_count, width = 4; - size_t sz = rows * width * sizeof(float); - - struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms); - pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz); - u32 *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms); - - pandecode_log("u32 uniforms_%d%s[] = {\n", job_no, suffix); - - pandecode_indent++; - - for (int row = 0; row < rows; row++) { - for (int i = 0; i < width; i++) { - u32 v = uniforms[i]; - float f; - memcpy(&f, &v, sizeof(v)); - pandecode_log_cont("%X /* %f */, ", v, f); - } - - pandecode_log_cont("\n"); - - uniforms += width; - } - - pandecode_indent--; - pandecode_log("};\n"); - } else if (p->uniforms) { - int rows = uniform_count * 2; - size_t sz = rows * sizeof(mali_ptr); - - struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms); - pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz); - mali_ptr *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms); - - pandecode_log("mali_ptr uniforms_%d%s[] = {\n", job_no, suffix); - - pandecode_indent++; - - for (int row = 0; row < rows; row++) { - char *a = pointer_as_memory_reference(uniforms[row]); - pandecode_log("%s,\n", a); - free(a); - } - - pandecode_indent--; - pandecode_log("};\n"); - + pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true); } if (p->uniform_buffers) { - pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no); - } + if (uniform_buffer_count) + pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no); + else + pandecode_msg("XXX: UBOs specified but not referenced\n"); + } else if (uniform_buffer_count) + pandecode_msg("XXX: UBOs referenced but not specified\n"); + + /* We don't want to actually dump uniforms, but we do need to validate + * that the counts we were given are sane */ + + if (p->uniforms) { + if (uniform_count) + pandecode_validate_buffer(p->uniforms, uniform_count * 16); + else + pandecode_msg("XXX: Uniforms specified but not referenced"); + } else if (uniform_count) + pandecode_msg("XXX: UBOs referenced but not specified\n"); if (p->texture_trampoline) { struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline); @@ -1755,108 +2159,11 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, pandecode_log("};\n"); /* Now, finally, descend down into the texture descriptor */ - for (int tex = 0; tex < texture_count; ++tex) { + for (unsigned tex = 0; tex < texture_count; ++tex) { mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr)); struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u); - - if (tmem) { - struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u); - - pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex); - pandecode_indent++; - - pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1); - pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1); - pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1); - pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1); - pandecode_prop("unknown3 = %" PRId16, t->unknown3); - pandecode_prop("unknown3A = %" PRId8, t->unknown3A); - pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels); - - struct mali_texture_format f = t->format; - - pandecode_log(".format = {\n"); - pandecode_indent++; - - pandecode_swizzle(f.swizzle); - pandecode_prop("format = %s", pandecode_format(f.format)); - pandecode_prop("type = %s", pandecode_texture_type(f.type)); - pandecode_prop("srgb = %" PRId32, f.srgb); - pandecode_prop("unknown1 = %" PRId32, f.unknown1); - pandecode_prop("usage2 = 0x%" PRIx32, f.usage2); - - pandecode_indent--; - pandecode_log("},\n"); - - pandecode_swizzle(t->swizzle); - - if (t->swizzle_zero) { - /* Shouldn't happen */ - pandecode_msg("Swizzle zero tripped but replay will be fine anyway"); - pandecode_prop("swizzle_zero = %d", t->swizzle_zero); - } - - pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3); - - pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5); - pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6); - pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7); - - pandecode_log(".payload = {\n"); - pandecode_indent++; - - /* A bunch of bitmap pointers follow. - * We work out the correct number, - * based on the mipmap/cubemap - * properties, but dump extra - * possibilities to futureproof */ - - int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels); - bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE; - - /* Miptree for each face */ - if (f.type == MALI_TEX_CUBE) - bitmap_count *= 6; - - /* Array of textures */ - bitmap_count *= MALI_NEGATIVE(t->array_size); - - /* Stride for each element */ - if (manual_stride) - bitmap_count *= 2; - - /* Sanity check the size */ - int max_count = sizeof(t->payload) / sizeof(t->payload[0]); - assert (bitmap_count <= max_count); - - /* Dump more to be safe, but not _that_ much more */ - int safe_count = MIN2(bitmap_count * 2, max_count); - - for (int i = 0; i < safe_count; ++i) { - char *prefix = (i >= bitmap_count) ? "// " : ""; - - /* How we dump depends if this is a stride or a pointer */ - - if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) { - /* signed 32-bit snuck in as a 64-bit pointer */ - uint64_t stride_set = t->payload[i]; - uint32_t clamped_stride = stride_set; - int32_t stride = clamped_stride; - assert(stride_set == clamped_stride); - pandecode_log("%s(mali_ptr) %d /* stride */, \n", prefix, stride); - } else { - char *a = pointer_as_memory_reference(t->payload[i]); - pandecode_log("%s%s, \n", prefix, a); - free(a); - } - } - - pandecode_indent--; - pandecode_log("},\n"); - - pandecode_indent--; - pandecode_log("};\n"); - } + if (tmem) + pandecode_texture(*u, tmem, job_no, tex); } } } @@ -1889,7 +2196,7 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func)); if (s->zero || s->zero2) { - pandecode_msg("Zero tripped\n"); + pandecode_msg("XXX: sampler zero tripped\n"); pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2); } @@ -1911,28 +2218,17 @@ pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p, static void pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost) { - pandecode_log_cont("{\n"); + if (!(p->position_varying || p->occlusion_counter || p->flags)) + return; + + pandecode_log(".postfix = {\n"); pandecode_indent++; MEMORY_PROP(p, position_varying); - DYN_MEMORY_PROP(p, job_no, uniform_buffers); - DYN_MEMORY_PROP(p, job_no, texture_trampoline); - DYN_MEMORY_PROP(p, job_no, sampler_descriptor); - DYN_MEMORY_PROP(p, job_no, uniforms); - DYN_MEMORY_PROP(p, job_no, attributes); - DYN_MEMORY_PROP(p, job_no, attribute_meta); - DYN_MEMORY_PROP(p, job_no, varyings); - DYN_MEMORY_PROP(p, job_no, varying_meta); - DYN_MEMORY_PROP(p, job_no, viewport); - DYN_MEMORY_PROP(p, job_no, occlusion_counter); - - if (is_bifrost) - pandecode_prop("framebuffer = scratchpad_%d_p", job_no); - else - pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, p->framebuffer & MALI_MFBD ? "MALI_MFBD" : "0"); + MEMORY_PROP(p, occlusion_counter); - pandecode_prop("_shader_upper = (shader_meta_%d_p) >> 4", job_no); - pandecode_prop("flags = %d", p->flags); + if (p->flags) + pandecode_prop("flags = %d", p->flags); pandecode_indent--; pandecode_log("},\n"); @@ -1947,7 +2243,7 @@ pandecode_vertex_only_bfr(struct bifrost_vertex_only *v) pandecode_prop("unk2 = 0x%x", v->unk2); if (v->zero0 || v->zero1) { - pandecode_msg("vertex only zero tripped"); + pandecode_msg("XXX: vertex only zero tripped"); pandecode_prop("zero0 = 0x%" PRIx32, v->zero0); pandecode_prop("zero1 = 0x%" PRIx64, v->zero1); } @@ -1967,13 +2263,13 @@ pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no) pandecode_indent++; if (h->zero) { - pandecode_msg("tiler heap zero tripped\n"); + pandecode_msg("XXX: tiler heap zero tripped\n"); pandecode_prop("zero = 0x%x", h->zero); } for (int i = 0; i < 12; i++) { if (h->zeros[i] != 0) { - pandecode_msg("tiler heap zero %d tripped, value %x\n", + pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n", i, h->zeros[i]); } } @@ -2011,7 +2307,7 @@ pandecode_tiler_meta(mali_ptr gpu_va, int job_no) pandecode_indent++; if (t->zero0 || t->zero1) { - pandecode_msg("tiler meta zero tripped"); + pandecode_msg("XXX: tiler meta zero tripped\n"); pandecode_prop("zero0 = 0x%" PRIx64, t->zero0); pandecode_prop("zero1 = 0x%" PRIx64, t->zero1); } @@ -2021,11 +2317,10 @@ pandecode_tiler_meta(mali_ptr gpu_va, int job_no) pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1); pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1); - DYN_MEMORY_PROP(t, job_no, tiler_heap_meta); for (int i = 0; i < 12; i++) { if (t->zeros[i] != 0) { - pandecode_msg("tiler heap zero %d tripped, value %" PRIx64 "\n", + pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n", i, t->zeros[i]); } } @@ -2072,12 +2367,11 @@ pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no) /* TODO: gl_PointSize on Bifrost */ pandecode_primitive_size(t->primitive_size, true); - DYN_MEMORY_PROP(t, job_no, tiler_meta); pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER); if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5 || t->zero6 || t->zero7 || t->zero8) { - pandecode_msg("tiler only zero tripped"); + pandecode_msg("XXX: tiler only zero tripped\n"); pandecode_prop("zero1 = 0x%" PRIx64, t->zero1); pandecode_prop("zero2 = 0x%" PRIx64, t->zero2); pandecode_prop("zero3 = 0x%" PRIx64, t->zero3); @@ -2105,12 +2399,11 @@ pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h, pandecode_indent++; pandecode_log(".prefix = "); - pandecode_vertex_tiler_prefix(&v->prefix, job_no); + pandecode_vertex_tiler_prefix(&v->prefix, job_no, false); pandecode_log(".vertex = "); pandecode_vertex_only_bfr(&v->vertex); - pandecode_log(".postfix = "); pandecode_vertex_tiler_postfix(&v->postfix, job_no, true); pandecode_indent--; @@ -2135,12 +2428,11 @@ pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h, pandecode_indent++; pandecode_log(".prefix = "); - pandecode_vertex_tiler_prefix(&t->prefix, job_no); + pandecode_vertex_tiler_prefix(&t->prefix, job_no, false); pandecode_log(".tiler = "); pandecode_tiler_only_bfr(&t->tiler, job_no); - pandecode_log(".postfix = "); pandecode_vertex_tiler_postfix(&t->postfix, job_no, true); pandecode_indent--; @@ -2166,8 +2458,11 @@ pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h, bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE; pandecode_primitive_size(v->primitive_size, !has_primitive_pointer); + bool instanced = v->instance_shift || v->instance_odd; + bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER); + pandecode_log(".prefix = "); - pandecode_vertex_tiler_prefix(&v->prefix, job_no); + pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics); pandecode_gl_enables(v->gl_enables, h->job_type); @@ -2180,15 +2475,14 @@ pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h, pandecode_padded_vertices(v->instance_shift, v->instance_odd); } - if (v->draw_start) - pandecode_prop("draw_start = %d", v->draw_start); + if (v->offset_start) + pandecode_prop("offset_start = %d", v->offset_start); if (v->zero5) { - pandecode_msg("Zero tripped\n"); + pandecode_msg("XXX: midgard payload zero tripped\n"); pandecode_prop("zero5 = 0x%" PRIx64, v->zero5); } - pandecode_log(".postfix = "); pandecode_vertex_tiler_postfix(&v->postfix, job_no, false); pandecode_indent--; @@ -2215,7 +2509,7 @@ pandecode_fragment_job(const struct pandecode_mapped_memory *mem, * including Gxx). In any event, there's some field shuffling * that we haven't looked into yet. */ - pandecode_sfbd(s->framebuffer & FBD_MASK, job_no); + pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true); fbd_dumped = true; } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) { /* We don't know if Bifrost supports SFBD's at all, since the @@ -2253,10 +2547,15 @@ pandecode_fragment_job(const struct pandecode_mapped_memory *mem, const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD"; + /* TODO: Decode */ + unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD; + if (fbd_dumped) - pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, fbd_type); + pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no, + fbd_type, extra_flags); else - pandecode_prop("framebuffer = %s | %s", pointer_as_memory_reference(p), fbd_type); + pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p), + fbd_type, extra_flags); pandecode_indent--; pandecode_log("};\n"); @@ -2311,11 +2610,11 @@ pandecode_jc(mali_ptr jc_gpu_va, bool bifrost) if (h->job_descriptor_size) pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size); - if (h->exception_status != 0x1) - pandecode_prop("exception_status = %x (source ID: 0x%x access: 0x%x exception: 0x%x)", + if (h->exception_status && h->exception_status != 0x1) + pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)", h->exception_status, (h->exception_status >> 16) & 0xFFFF, - (h->exception_status >> 8) & 0x3, + pandecode_exception_access((h->exception_status >> 8) & 0x3), h->exception_status & 0xFF); if (h->first_incomplete_task)