#include "a2xx/fd2_screen.h"
#include "a3xx/fd3_screen.h"
#include "a4xx/fd4_screen.h"
+#include "a5xx/fd5_screen.h"
+
+#include "ir3/ir3_nir.h"
/* XXX this should go away */
#include "state_tracker/drm_driver.h"
{"glsl120", FD_DBG_GLSL120,"Temporary flag to force GLSL 1.20 (rather than 1.30) on a3xx+"},
{"shaderdb", FD_DBG_SHADERDB, "Enable shaderdb output"},
{"flush", FD_DBG_FLUSH, "Force flush after every draw"},
+ {"deqp", FD_DBG_DEQP, "Enable dEQP hacks"},
+ {"nir", FD_DBG_NIR, "Prefer NIR as native IR"},
+ {"inorder", FD_DBG_INORDER,"Disable reordering for draws/blits"},
+ {"bstat", FD_DBG_BSTAT, "Print batch stats at context destroy"},
+ {"nogrow", FD_DBG_NOGROW, "Disable \"growable\" cmdstream buffers, even if kernel supports it"},
DEBUG_NAMED_VALUE_END
};
static uint64_t
fd_screen_get_timestamp(struct pipe_screen *pscreen)
{
- int64_t cpu_time = os_time_get() * 1000;
- return cpu_time + fd_screen(pscreen)->cpu_gpu_time_delta;
+ struct fd_screen *screen = fd_screen(pscreen);
+
+ if (screen->has_timestamp) {
+ uint64_t n;
+ fd_pipe_get_param(screen->pipe, FD_TIMESTAMP, &n);
+ debug_assert(screen->max_freq > 0);
+ return n * 1000000000 / screen->max_freq;
+ } else {
+ int64_t cpu_time = os_time_get() * 1000;
+ return cpu_time + screen->cpu_gpu_time_delta;
+ }
+
}
static void
if (screen->dev)
fd_device_del(screen->dev);
+ fd_bc_fini(&screen->batch_cache);
+
+ slab_destroy_parent(&screen->transfer_pool);
+
+ mtx_destroy(&screen->lock);
+
+ ralloc_free(screen->compiler);
+
free(screen);
}
case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
- case PIPE_CAP_USER_CONSTANT_BUFFERS:
case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT:
- case PIPE_CAP_VERTEXID_NOBASE:
+ case PIPE_CAP_STRING_MARKER:
+ case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
return 1;
+ case PIPE_CAP_VERTEXID_NOBASE:
+ return is_a3xx(screen) || is_a4xx(screen);
+
+ case PIPE_CAP_USER_CONSTANT_BUFFERS:
+ return is_a4xx(screen) ? 0 : 1;
+
case PIPE_CAP_SHADER_STENCIL_EXPORT:
case PIPE_CAP_TGSI_TEXCOORD:
case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
case PIPE_CAP_TEXTURE_BARRIER:
case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
case PIPE_CAP_COMPUTE:
+ case PIPE_CAP_QUERY_MEMORY_INFO:
+ case PIPE_CAP_PCI_GROUP:
+ case PIPE_CAP_PCI_BUS:
+ case PIPE_CAP_PCI_DEVICE:
+ case PIPE_CAP_PCI_FUNCTION:
return 0;
case PIPE_CAP_SM3:
case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
case PIPE_CAP_DEPTH_CLIP_DISABLE:
case PIPE_CAP_CLIP_HALFZ:
- return is_a3xx(screen) || is_a4xx(screen);
+ return is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen);
+ case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
+ return 0;
case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
if (is_a3xx(screen)) return 16;
if (is_a4xx(screen)) return 32;
+ if (is_a5xx(screen)) return 32;
return 0;
case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
/* We could possibly emulate more by pretending 2d/rect textures and
*/
if (is_a3xx(screen)) return 8192;
if (is_a4xx(screen)) return 16384;
+ if (is_a5xx(screen)) return 16384;
return 0;
case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
case PIPE_CAP_START_INSTANCE:
case PIPE_CAP_SAMPLER_VIEW_TARGET:
case PIPE_CAP_TEXTURE_QUERY_LOD:
- return is_a4xx(screen);
+ return is_a4xx(screen) || is_a5xx(screen);
case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
- return 256;
+ return 64;
case PIPE_CAP_GLSL_FEATURE_LEVEL:
if (glsl120)
case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
- case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
- case PIPE_CAP_VERTEX_COLOR_CLAMPED:
case PIPE_CAP_USER_VERTEX_BUFFERS:
- case PIPE_CAP_USER_INDEX_BUFFERS:
case PIPE_CAP_QUERY_PIPELINE_STATISTICS:
case PIPE_CAP_TEXTURE_BORDER_COLOR_QUIRK:
case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT:
case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION:
case PIPE_CAP_DRAW_INDIRECT:
+ case PIPE_CAP_MULTI_DRAW_INDIRECT:
+ case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE:
case PIPE_CAP_POLYGON_OFFSET_CLAMP:
case PIPE_CAP_MULTISAMPLE_Z_RESOLVE:
case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
case PIPE_CAP_DEPTH_BOUNDS_TEST:
case PIPE_CAP_TGSI_TXQS:
+ /* TODO if we need this, do it in nir/ir3 backend to avoid breaking precompile: */
case PIPE_CAP_FORCE_PERSAMPLE_INTERP:
- case PIPE_CAP_SHAREABLE_SHADERS:
case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS:
case PIPE_CAP_CLEAR_TEXTURE:
case PIPE_CAP_DRAW_PARAMETERS:
case PIPE_CAP_TGSI_PACK_HALF_FLOAT:
+ case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
+ case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
+ case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
+ case PIPE_CAP_INVALIDATE_BUFFER:
+ case PIPE_CAP_GENERATE_MIPMAP:
+ case PIPE_CAP_SURFACE_REINTERPRET_BLOCKS:
+ case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT:
+ case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR:
+ case PIPE_CAP_CULL_DISTANCE:
+ case PIPE_CAP_PRIMITIVE_RESTART_FOR_PATCHES:
+ case PIPE_CAP_TGSI_VOTE:
+ case PIPE_CAP_MAX_WINDOW_RECTANGLES:
+ case PIPE_CAP_POLYGON_OFFSET_UNITS_UNSCALED:
+ case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS:
+ case PIPE_CAP_TGSI_ARRAY_COMPONENTS:
+ case PIPE_CAP_TGSI_CAN_READ_OUTPUTS:
+ case PIPE_CAP_TGSI_FS_FBFETCH:
+ case PIPE_CAP_TGSI_MUL_ZERO_WINS:
+ case PIPE_CAP_DOUBLES:
+ case PIPE_CAP_INT64:
+ case PIPE_CAP_INT64_DIVMOD:
+ case PIPE_CAP_TGSI_TEX_TXF_LZ:
+ case PIPE_CAP_TGSI_CLOCK:
+ case PIPE_CAP_POLYGON_MODE_FILL_RECTANGLE:
+ case PIPE_CAP_SPARSE_BUFFER_PAGE_SIZE:
+ case PIPE_CAP_TGSI_BALLOT:
+ case PIPE_CAP_TGSI_TES_LAYER_VIEWPORT:
return 0;
case PIPE_CAP_MAX_VIEWPORTS:
return 1;
+ case PIPE_CAP_SHAREABLE_SHADERS:
+ case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY:
+ /* manage the variants for these ourself, to avoid breaking precompile: */
+ case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
+ case PIPE_CAP_VERTEX_COLOR_CLAMPED:
+ if (is_ir3(screen))
+ return 1;
+ return 0;
+
/* Stream output. */
case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
if (is_ir3(screen))
return PIPE_MAX_SO_BUFFERS;
return 0;
case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
+ case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
if (is_ir3(screen))
return 1;
return 0;
return 11;
case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
- return (is_a3xx(screen) || is_a4xx(screen)) ? 256 : 0;
+ return (is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen)) ? 256 : 0;
/* Render targets. */
case PIPE_CAP_MAX_RENDER_TARGETS:
return is_a3xx(screen) ? 1 : 0;
/* Queries. */
- case PIPE_CAP_QUERY_TIME_ELAPSED:
- case PIPE_CAP_QUERY_TIMESTAMP:
+ case PIPE_CAP_QUERY_BUFFER_OBJECT:
return 0;
case PIPE_CAP_OCCLUSION_QUERY:
- return is_a3xx(screen) || is_a4xx(screen);
+ return is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen);
+ case PIPE_CAP_QUERY_TIMESTAMP:
+ case PIPE_CAP_QUERY_TIME_ELAPSED:
+ /* only a4xx, requires new enough kernel so we know max_freq: */
+ return (screen->max_freq > 0) && is_a4xx(screen);
case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
case PIPE_CAP_MIN_TEXEL_OFFSET:
return 10;
case PIPE_CAP_UMA:
return 1;
+ case PIPE_CAP_NATIVE_FENCE_FD:
+ return fd_device_version(screen->dev) >= FD_VERSION_FENCE_FD;
}
debug_printf("unknown param %d\n", param);
return 0;
switch (param) {
case PIPE_CAPF_MAX_LINE_WIDTH:
case PIPE_CAPF_MAX_LINE_WIDTH_AA:
+ /* NOTE: actual value is 127.0f, but this is working around a deqp
+ * bug.. dEQP-GLES3.functional.rasterization.primitives.lines_wide
+ * uses too small of a render target size, and gets confused when
+ * the lines start going offscreen.
+ *
+ * See: https://code.google.com/p/android/issues/detail?id=206513
+ */
+ if (fd_mesa_debug & FD_DBG_DEQP)
+ return 48.0f;
+ return 127.0f;
case PIPE_CAPF_MAX_POINT_WIDTH:
case PIPE_CAPF_MAX_POINT_WIDTH_AA:
return 4092.0f;
}
static int
-fd_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
+fd_screen_get_shader_param(struct pipe_screen *pscreen,
+ enum pipe_shader_type shader,
enum pipe_shader_cap param)
{
struct fd_screen *screen = fd_screen(pscreen);
* split between VS and FS. Use lower limit of 256 to
* avoid getting into impossible situations:
*/
- return ((is_a3xx(screen) || is_a4xx(screen)) ? 4096 : 64) * sizeof(float[4]);
+ return ((is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen)) ? 4096 : 64) * sizeof(float[4]);
case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
return is_ir3(screen) ? 16 : 1;
- case PIPE_SHADER_CAP_MAX_PREDS:
- return 0; /* nothing uses this */
case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
return 1;
case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
+ /* Technically this should be the same as for TEMP/CONST, since
+ * everything is just normal registers. This is just temporary
+ * hack until load_input/store_output handle arrays in a similar
+ * way as load_var/store_var..
+ */
+ return 0;
case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
- return 1;
+ /* a2xx compiler doesn't handle indirect: */
+ return is_ir3(screen) ? 1 : 0;
case PIPE_SHADER_CAP_SUBROUTINES:
- case PIPE_SHADER_CAP_DOUBLES:
case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
- case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
+ case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
return 0;
case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
return 1;
case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
return 16;
case PIPE_SHADER_CAP_PREFERRED_IR:
+ if ((fd_mesa_debug & FD_DBG_NIR) && is_ir3(screen))
+ return PIPE_SHADER_IR_NIR;
return PIPE_SHADER_IR_TGSI;
+ case PIPE_SHADER_CAP_SUPPORTED_IRS:
+ return 0;
case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
return 32;
+ case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
+ case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
+ case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD:
+ case PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS:
+ return 0;
}
debug_printf("unknown shader param %d\n", param);
return 0;
}
+static const void *
+fd_get_compiler_options(struct pipe_screen *pscreen,
+ enum pipe_shader_ir ir, unsigned shader)
+{
+ struct fd_screen *screen = fd_screen(pscreen);
+
+ if (is_ir3(screen))
+ return ir3_get_compiler_options();
+
+ return NULL;
+}
+
boolean
fd_screen_bo_get_handle(struct pipe_screen *pscreen,
struct fd_bo *bo,
struct fd_bo *
fd_screen_bo_from_handle(struct pipe_screen *pscreen,
- struct winsys_handle *whandle,
- unsigned *out_stride)
+ struct winsys_handle *whandle)
{
struct fd_screen *screen = fd_screen(pscreen);
struct fd_bo *bo;
return NULL;
}
- *out_stride = whandle->stride;
-
return bo;
}
}
screen->device_id = val;
+ if (fd_pipe_get_param(screen->pipe, FD_MAX_FREQ, &val)) {
+ DBG("could not get gpu freq");
+ /* this limits what performance related queries are
+ * supported but is not fatal
+ */
+ screen->max_freq = 0;
+ } else {
+ screen->max_freq = val;
+ if (fd_pipe_get_param(screen->pipe, FD_TIMESTAMP, &val) == 0)
+ screen->has_timestamp = true;
+ }
+
if (fd_pipe_get_param(screen->pipe, FD_GPU_ID, &val)) {
DBG("could not get gpu-id");
goto fail;
fd3_screen_init(pscreen);
break;
case 420:
+ case 430:
fd4_screen_init(pscreen);
break;
+ case 530:
+ fd5_screen_init(pscreen);
+ break;
default:
debug_printf("unsupported GPU: a%03d\n", screen->gpu_id);
goto fail;
}
+ if (screen->gpu_id >= 500) {
+ screen->gmem_alignw = 64;
+ screen->gmem_alignh = 32;
+ } else {
+ screen->gmem_alignw = 32;
+ screen->gmem_alignh = 32;
+ }
+
+ /* NOTE: don't enable reordering on a2xx, since completely untested.
+ * Also, don't enable if we have too old of a kernel to support
+ * growable cmdstream buffers, since memory requirement for cmdstream
+ * buffers would be too much otherwise.
+ */
+ if ((screen->gpu_id >= 300) && (fd_device_version(dev) >= FD_VERSION_UNLIMITED_CMDS))
+ screen->reorder = !(fd_mesa_debug & FD_DBG_INORDER);
+
+ fd_bc_init(&screen->batch_cache);
+
+ (void) mtx_init(&screen->lock, mtx_plain);
+
pscreen->destroy = fd_screen_destroy;
pscreen->get_param = fd_screen_get_param;
pscreen->get_paramf = fd_screen_get_paramf;
pscreen->get_shader_param = fd_screen_get_shader_param;
+ pscreen->get_compiler_options = fd_get_compiler_options;
fd_resource_screen_init(pscreen);
fd_query_screen_init(pscreen);
pscreen->get_timestamp = fd_screen_get_timestamp;
- pscreen->fence_reference = fd_screen_fence_ref;
- pscreen->fence_finish = fd_screen_fence_finish;
+ pscreen->fence_reference = fd_fence_ref;
+ pscreen->fence_finish = fd_fence_finish;
+ pscreen->fence_get_fd = fd_fence_get_fd;
+
+ slab_create_parent(&screen->transfer_pool, sizeof(struct fd_transfer), 16);
util_format_s3tc_init();