#include <freedreno_ringbuffer.h>
#include "pipe/p_format.h"
+#include "pipe/p_state.h"
#include "util/u_debug.h"
+#include "util/u_math.h"
+#include "util/u_half.h"
+#include "util/u_dynarray.h"
+#include "util/u_pack_color.h"
-#include "freedreno_pm4.h"
-#include "freedreno_a2xx_reg.h"
+#include "disasm.h"
+#include "adreno_common.xml.h"
+#include "adreno_pm4.xml.h"
-enum sq_surfaceformat fd_pipe2surface(enum pipe_format format);
-enum rb_colorformatx fd_pipe2color(enum pipe_format format);
-enum rb_depth_format fd_pipe2depth(enum pipe_format format);
+enum adreno_rb_depth_format fd_pipe2depth(enum pipe_format format);
enum pc_di_index_size fd_pipe2index(enum pipe_format format);
-uint32_t fd_tex_swiz(enum pipe_format format, unsigned swizzle_r,
- unsigned swizzle_g, unsigned swizzle_b, unsigned swizzle_a);
+enum adreno_rb_blend_factor fd_blend_factor(unsigned factor);
+enum adreno_pa_su_sc_draw fd_polygon_mode(unsigned mode);
+enum adreno_stencil_op fd_stencil_op(unsigned op);
+#define A3XX_MAX_MIP_LEVELS 14
+/* TBD if it is same on a2xx, but for now: */
+#define MAX_MIP_LEVELS A3XX_MAX_MIP_LEVELS
+
+#define A2XX_MAX_RENDER_TARGETS 1
+#define A3XX_MAX_RENDER_TARGETS 4
+#define A4XX_MAX_RENDER_TARGETS 8
+
+#define MAX_RENDER_TARGETS A4XX_MAX_RENDER_TARGETS
+
+#define FD_DBG_MSGS 0x0001
+#define FD_DBG_DISASM 0x0002
+#define FD_DBG_DCLEAR 0x0004
+#define FD_DBG_DDRAW 0x0008
+#define FD_DBG_NOSCIS 0x0010
+#define FD_DBG_DIRECT 0x0020
+#define FD_DBG_NOBYPASS 0x0040
+#define FD_DBG_FRAGHALF 0x0080
+#define FD_DBG_NOBIN 0x0100
+#define FD_DBG_OPTMSGS 0x0200
+#define FD_DBG_GLSL120 0x0400
+#define FD_DBG_SHADERDB 0x0800
+#define FD_DBG_FLUSH 0x1000
-#define FD_DBG_MSGS 0x1
-#define FD_DBG_DISASM 0x2
extern int fd_mesa_debug;
+extern bool fd_binning_enabled;
#define DBG(fmt, ...) \
do { if (fd_mesa_debug & FD_DBG_MSGS) \
debug_printf("%s:%d: "fmt "\n", \
__FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
-#define ALIGN(v,a) (((v) + (a) - 1) & ~((a) - 1))
-#define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
+/* for conditionally setting boolean flag(s): */
+#define COND(bool, val) ((bool) ? (val) : 0)
+
+#define CP_REG(reg) ((0x4 << 16) | ((unsigned int)((reg) - (0x2000))))
+
+static inline uint32_t DRAW(enum pc_di_primtype prim_type,
+ enum pc_di_src_sel source_select, enum pc_di_index_size index_size,
+ enum pc_di_vis_cull_mode vis_cull_mode,
+ uint8_t instances)
+{
+ return (prim_type << 0) |
+ (source_select << 6) |
+ ((index_size & 1) << 11) |
+ ((index_size >> 1) << 13) |
+ (vis_cull_mode << 9) |
+ (1 << 14) |
+ (instances << 24);
+}
+
+/* for tracking cmdstream positions that need to be patched: */
+struct fd_cs_patch {
+ uint32_t *cs;
+ uint32_t val;
+};
+#define fd_patch_num_elements(buf) ((buf)->size / sizeof(struct fd_cs_patch))
+#define fd_patch_element(buf, i) util_dynarray_element(buf, struct fd_cs_patch, i)
+
+static inline enum pipe_format
+pipe_surface_format(struct pipe_surface *psurf)
+{
+ if (!psurf)
+ return PIPE_FORMAT_NONE;
+ return psurf->format;
+}
+
+static inline bool
+fd_surface_half_precision(const struct pipe_surface *psurf)
+{
+ enum pipe_format format;
+
+ if (!psurf)
+ return true;
+
+ format = psurf->format;
+ /* colors are provided in consts, which go through cov.f32f16, which will
+ * break these values
+ */
+ if (util_format_is_pure_integer(format))
+ return false;
-#define min(a, b) (((a) < (b)) ? (a) : (b))
-#define max(a, b) (((a) > (b)) ? (a) : (b))
+ /* avoid losing precision on 32-bit float formats */
+ if (util_format_is_float(format) &&
+ util_format_get_component_bits(format, UTIL_FORMAT_COLORSPACE_RGB, 0) == 32)
+ return false;
+ return true;
+}
+
+static inline unsigned
+fd_sampler_first_level(const struct pipe_sampler_view *view)
+{
+ if (view->target == PIPE_BUFFER)
+ return 0;
+ return view->u.tex.first_level;
+}
+
+static inline unsigned
+fd_sampler_last_level(const struct pipe_sampler_view *view)
+{
+ if (view->target == PIPE_BUFFER)
+ return 0;
+ return view->u.tex.last_level;
+}
+
+static inline bool
+fd_half_precision(struct pipe_framebuffer_state *pfb)
+{
+ unsigned i;
+
+ for (i = 0; i < pfb->nr_cbufs; i++)
+ if (!fd_surface_half_precision(pfb->cbufs[i]))
+ return false;
+
+ return true;
+}
#define LOG_DWORDS 0
+static inline void emit_marker(struct fd_ringbuffer *ring, int scratch_idx);
static inline void
OUT_RING(struct fd_ringbuffer *ring, uint32_t data)
*(ring->cur++) = data;
}
+/* like OUT_RING() but appends a cmdstream patch point to 'buf' */
+static inline void
+OUT_RINGP(struct fd_ringbuffer *ring, uint32_t data,
+ struct util_dynarray *buf)
+{
+ if (LOG_DWORDS) {
+ DBG("ring[%p]: OUT_RINGP %04x: %08x", ring,
+ (uint32_t)(ring->cur - ring->last_start), data);
+ }
+ util_dynarray_append(buf, struct fd_cs_patch, ((struct fd_cs_patch){
+ .cs = ring->cur++,
+ .val = data,
+ }));
+}
+
static inline void
OUT_RELOC(struct fd_ringbuffer *ring, struct fd_bo *bo,
- uint32_t offset, uint32_t or)
+ uint32_t offset, uint32_t or, int32_t shift)
{
if (LOG_DWORDS) {
- DBG("ring[%p]: OUT_RELOC %04x: %p+%u", ring,
- (uint32_t)(ring->cur - ring->last_start), bo, offset);
+ DBG("ring[%p]: OUT_RELOC %04x: %p+%u << %d", ring,
+ (uint32_t)(ring->cur - ring->last_start), bo, offset, shift);
}
- fd_ringbuffer_emit_reloc(ring, bo, offset, or);
+ fd_ringbuffer_reloc(ring, &(struct fd_reloc){
+ .bo = bo,
+ .flags = FD_RELOC_READ,
+ .offset = offset,
+ .or = or,
+ .shift = shift,
+ });
+}
+
+static inline void
+OUT_RELOCW(struct fd_ringbuffer *ring, struct fd_bo *bo,
+ uint32_t offset, uint32_t or, int32_t shift)
+{
+ if (LOG_DWORDS) {
+ DBG("ring[%p]: OUT_RELOCW %04x: %p+%u << %d", ring,
+ (uint32_t)(ring->cur - ring->last_start), bo, offset, shift);
+ }
+ fd_ringbuffer_reloc(ring, &(struct fd_reloc){
+ .bo = bo,
+ .flags = FD_RELOC_READ | FD_RELOC_WRITE,
+ .offset = offset,
+ .or = or,
+ .shift = shift,
+ });
}
static inline void BEGIN_RING(struct fd_ringbuffer *ring, uint32_t ndwords)
OUT_RING(ring, CP_TYPE3_PKT | ((cnt-1) << 16) | ((opcode & 0xFF) << 8));
}
+static inline void
+OUT_WFI(struct fd_ringbuffer *ring)
+{
+ OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
+ OUT_RING(ring, 0x00000000);
+}
+
static inline void
OUT_IB(struct fd_ringbuffer *ring, struct fd_ringmarker *start,
struct fd_ringmarker *end)
{
+ uint32_t dwords = fd_ringmarker_dwords(start, end);
+
+ assert(dwords > 0);
+
+ /* for debug after a lock up, write a unique counter value
+ * to scratch6 for each IB, to make it easier to match up
+ * register dumps to cmdstream. The combination of IB and
+ * DRAW (scratch7) is enough to "triangulate" the particular
+ * draw that caused lockup.
+ */
+ emit_marker(ring, 6);
+
OUT_PKT3(ring, CP_INDIRECT_BUFFER_PFD, 2);
- fd_ringbuffer_emit_reloc_ring(ring, start);
- OUT_RING(ring, fd_ringmarker_dwords(start, end));
+ fd_ringbuffer_emit_reloc_ring(ring, start, end);
+ OUT_RING(ring, dwords);
+
+ emit_marker(ring, 6);
+}
+
+/* CP_SCRATCH_REG4 is used to hold base address for query results: */
+#define HW_QUERY_BASE_REG REG_AXXX_CP_SCRATCH_REG4
+
+static inline void
+emit_marker(struct fd_ringbuffer *ring, int scratch_idx)
+{
+ extern unsigned marker_cnt;
+ unsigned reg = REG_AXXX_CP_SCRATCH_REG0 + scratch_idx;
+ assert(reg != HW_QUERY_BASE_REG);
+ if (reg == HW_QUERY_BASE_REG)
+ return;
+ OUT_PKT0(ring, reg, 1);
+ OUT_RING(ring, ++marker_cnt);
+}
+
+/* helper to get numeric value from environment variable.. mostly
+ * just leaving this here because it is helpful to brute-force figure
+ * out unknown formats, etc, which blob driver does not support:
+ */
+static inline uint32_t env2u(const char *envvar)
+{
+ char *str = getenv(envvar);
+ if (str)
+ return strtoul(str, NULL, 0);
+ return 0;
+}
+
+static inline uint32_t
+pack_rgba(enum pipe_format format, const float *rgba)
+{
+ union util_color uc;
+ util_pack_color(rgba, format, &uc);
+ return uc.ui[0];
}
#endif /* FREEDRENO_UTIL_H_ */