/* this is same for a2xx/a3xx, so split into helper: */
void
fd_draw_emit(struct fd_context *ctx, struct fd_ringbuffer *ring,
enum pc_di_vis_cull_mode vismode,
const struct pipe_draw_info *info)
{
struct pipe_index_buffer *idx = &ctx->indexbuf;
struct fd_bo *idx_bo = NULL;
enum pc_di_index_size idx_type = INDEX_SIZE_IGN;
enum pc_di_src_sel src_sel;
uint32_t idx_size, idx_offset;
if (info->indexed) {
assert(!idx->user_buffer);
idx_bo = fd_resource(idx->buffer)->bo;
idx_type = size2indextype(idx->index_size);
idx_size = idx->index_size * info->count;
idx_offset = idx->offset;
src_sel = DI_SRC_SEL_DMA;
} else {
idx_bo = NULL;
idx_type = INDEX_SIZE_IGN;
idx_size = 0;
idx_offset = 0;
src_sel = DI_SRC_SEL_AUTO_INDEX;
}
fd_draw(ctx, ring, ctx->primtypes[info->mode], vismode, src_sel,
info->count, idx_type, idx_size, idx_offset, idx_bo);
}
static void
emit_gmem2mem_surf(struct fd_batch *batch, uint32_t base,
struct pipe_surface *psurf)
{
struct fd_ringbuffer *ring = batch->gmem;
struct fd_resource *rsc = fd_resource(psurf->texture);
uint32_t swap = fmt2swap(psurf->format);
struct fd_resource_slice *slice =
fd_resource_slice(rsc, psurf->u.tex.level);
uint32_t offset =
fd_resource_offset(rsc, psurf->u.tex.level, psurf->u.tex.first_layer);
assert((slice->pitch & 31) == 0);
assert((offset & 0xfff) == 0);
if (!rsc->valid)
return;
OUT_PKT3(ring, CP_SET_CONSTANT, 2);
OUT_RING(ring, CP_REG(REG_A2XX_RB_COLOR_INFO));
OUT_RING(ring, A2XX_RB_COLOR_INFO_SWAP(swap) |
A2XX_RB_COLOR_INFO_BASE(base) |
A2XX_RB_COLOR_INFO_FORMAT(fd2_pipe2color(psurf->format)));
OUT_PKT3(ring, CP_SET_CONSTANT, 5);
OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_CONTROL));
OUT_RING(ring, 0x00000000); /* RB_COPY_CONTROL */
OUT_RELOCW(ring, rsc->bo, offset, 0, 0); /* RB_COPY_DEST_BASE */
OUT_RING(ring, slice->pitch >> 5); /* RB_COPY_DEST_PITCH */
OUT_RING(ring, /* RB_COPY_DEST_INFO */
A2XX_RB_COPY_DEST_INFO_FORMAT(fd2_pipe2color(psurf->format)) |
A2XX_RB_COPY_DEST_INFO_LINEAR |
A2XX_RB_COPY_DEST_INFO_SWAP(swap) |
A2XX_RB_COPY_DEST_INFO_WRITE_RED |
A2XX_RB_COPY_DEST_INFO_WRITE_GREEN |
A2XX_RB_COPY_DEST_INFO_WRITE_BLUE |
A2XX_RB_COPY_DEST_INFO_WRITE_ALPHA);
if (!is_a20x(batch->ctx->screen)) {
OUT_WFI (ring);
OUT_PKT3(ring, CP_SET_CONSTANT, 3);
OUT_RING(ring, CP_REG(REG_A2XX_VGT_MAX_VTX_INDX));
OUT_RING(ring, 3); /* VGT_MAX_VTX_INDX */
OUT_RING(ring, 0); /* VGT_MIN_VTX_INDX */
}
fd_draw(batch, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
DI_SRC_SEL_AUTO_INDEX, 3, 0, INDEX_SIZE_IGN, 0, 0, NULL);
}
static void
emit_gmem2mem_surf(struct fd_context *ctx, uint32_t base,
struct pipe_surface *psurf)
{
struct fd_ringbuffer *ring = ctx->ring;
struct fd_resource *rsc = fd_resource(psurf->texture);
uint32_t swap = fmt2swap(psurf->format);
OUT_PKT3(ring, CP_SET_CONSTANT, 2);
OUT_RING(ring, CP_REG(REG_A2XX_RB_COLOR_INFO));
OUT_RING(ring, A2XX_RB_COLOR_INFO_SWAP(swap) |
A2XX_RB_COLOR_INFO_BASE(base) |
A2XX_RB_COLOR_INFO_FORMAT(fd2_pipe2color(psurf->format)));
OUT_PKT3(ring, CP_SET_CONSTANT, 5);
OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_CONTROL));
OUT_RING(ring, 0x00000000); /* RB_COPY_CONTROL */
OUT_RELOCW(ring, rsc->bo, 0, 0, 0); /* RB_COPY_DEST_BASE */
OUT_RING(ring, rsc->slices[0].pitch >> 5); /* RB_COPY_DEST_PITCH */
OUT_RING(ring, /* RB_COPY_DEST_INFO */
A2XX_RB_COPY_DEST_INFO_FORMAT(fd2_pipe2color(psurf->format)) |
A2XX_RB_COPY_DEST_INFO_LINEAR |
A2XX_RB_COPY_DEST_INFO_SWAP(swap) |
A2XX_RB_COPY_DEST_INFO_WRITE_RED |
A2XX_RB_COPY_DEST_INFO_WRITE_GREEN |
A2XX_RB_COPY_DEST_INFO_WRITE_BLUE |
A2XX_RB_COPY_DEST_INFO_WRITE_ALPHA);
OUT_WFI (ring);
OUT_PKT3(ring, CP_SET_CONSTANT, 3);
OUT_RING(ring, CP_REG(REG_A2XX_VGT_MAX_VTX_INDX));
OUT_RING(ring, 3); /* VGT_MAX_VTX_INDX */
OUT_RING(ring, 0); /* VGT_MIN_VTX_INDX */
fd_draw(ctx, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
DI_SRC_SEL_AUTO_INDEX, 3, INDEX_SIZE_IGN, 0, 0, NULL);
}
static void
fd3_draw(struct fd_context *ctx, const struct pipe_draw_info *info)
{
unsigned dirty = ctx->dirty;
struct fd3_shader_key key = {
/* do binning pass first: */
.binning_pass = true,
.color_two_side = ctx->rasterizer ? ctx->rasterizer->light_twoside : false,
// TODO set .half_precision based on render target format,
// ie. float16 and smaller use half, float32 use full..
.half_precision = !!(fd_mesa_debug & FD_DBG_FRAGHALF),
};
draw_impl(ctx, info, ctx->binning_ring,
dirty & ~(FD_DIRTY_BLEND), key);
/* and now regular (non-binning) pass: */
key.binning_pass = false;
draw_impl(ctx, info, ctx->ring, dirty, key);
}
/* binning pass cmds for a clear:
* NOTE: newer blob drivers don't use binning for clear, which is probably
* preferable since it is low vtx count. However that doesn't seem to
* actually work for me. Not sure if it is depending on support for
* clear pass (rather than using solid-fill shader), or something else
* that newer blob is doing differently. Once that is figured out, we
* can remove fd3_clear_binning().
*/
static void
fd3_clear_binning(struct fd_context *ctx, unsigned dirty)
{
struct fd3_context *fd3_ctx = fd3_context(ctx);
struct fd_ringbuffer *ring = ctx->binning_ring;
struct fd3_shader_key key = {
.binning_pass = true,
.half_precision = true,
};
fd3_emit_state(ctx, ring, &ctx->solid_prog, dirty, key);
fd3_emit_vertex_bufs(ring, fd3_shader_variant(ctx->solid_prog.vp, key),
(struct fd3_vertex_buf[]) {{
.prsc = fd3_ctx->solid_vbuf,
.stride = 12,
.format = PIPE_FORMAT_R32G32B32_FLOAT,
}}, 1);
OUT_PKT0(ring, REG_A3XX_PC_PRIM_VTX_CNTL, 1);
OUT_RING(ring, A3XX_PC_PRIM_VTX_CNTL_STRIDE_IN_VPC(0) |
A3XX_PC_PRIM_VTX_CNTL_POLYMODE_FRONT_PTYPE(PC_DRAW_TRIANGLES) |
A3XX_PC_PRIM_VTX_CNTL_POLYMODE_BACK_PTYPE(PC_DRAW_TRIANGLES) |
A3XX_PC_PRIM_VTX_CNTL_PROVOKING_VTX_LAST);
OUT_PKT0(ring, REG_A3XX_VFD_INDEX_MIN, 4);
OUT_RING(ring, 0); /* VFD_INDEX_MIN */
OUT_RING(ring, 2); /* VFD_INDEX_MAX */
OUT_RING(ring, 0); /* VFD_INSTANCEID_OFFSET */
OUT_RING(ring, 0); /* VFD_INDEX_OFFSET */
OUT_PKT0(ring, REG_A3XX_PC_RESTART_INDEX, 1);
OUT_RING(ring, 0xffffffff); /* PC_RESTART_INDEX */
OUT_PKT3(ring, CP_EVENT_WRITE, 1);
OUT_RING(ring, PERFCOUNTER_STOP);
fd_draw(ctx, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
DI_SRC_SEL_AUTO_INDEX, 2, INDEX_SIZE_IGN, 0, 0, NULL);
}
static void
fd3_clear(struct fd_context *ctx, unsigned buffers,
const union pipe_color_union *color, double depth, unsigned stencil)
{
struct fd3_context *fd3_ctx = fd3_context(ctx);
struct fd_ringbuffer *ring = ctx->ring;
unsigned dirty = ctx->dirty;
unsigned ce, i;
struct fd3_shader_key key = {
.half_precision = true,
};
dirty &= FD_DIRTY_VIEWPORT | FD_DIRTY_FRAMEBUFFER | FD_DIRTY_SCISSOR;
dirty |= FD_DIRTY_PROG;
fd3_clear_binning(ctx, dirty);
/* emit generic state now: */
fd3_emit_state(ctx, ring, &ctx->solid_prog, dirty, key);
OUT_PKT0(ring, REG_A3XX_RB_BLEND_ALPHA, 1);
OUT_RING(ring, A3XX_RB_BLEND_ALPHA_UINT(0xff) |
A3XX_RB_BLEND_ALPHA_FLOAT(1.0));
OUT_PKT0(ring, REG_A3XX_RB_RENDER_CONTROL, 1);
OUT_RINGP(ring, A3XX_RB_RENDER_CONTROL_ALPHA_TEST_FUNC(FUNC_NEVER),
&fd3_ctx->rbrc_patches);
if (buffers & PIPE_CLEAR_DEPTH) {
OUT_PKT0(ring, REG_A3XX_RB_DEPTH_CONTROL, 1);
OUT_RING(ring, A3XX_RB_DEPTH_CONTROL_Z_WRITE_ENABLE |
A3XX_RB_DEPTH_CONTROL_Z_ENABLE |
A3XX_RB_DEPTH_CONTROL_ZFUNC(FUNC_ALWAYS));
OUT_PKT0(ring, REG_A3XX_GRAS_CL_VPORT_ZOFFSET, 2);
OUT_RING(ring, A3XX_GRAS_CL_VPORT_ZOFFSET(0.0));
OUT_RING(ring, A3XX_GRAS_CL_VPORT_ZSCALE(depth));
ctx->dirty |= FD_DIRTY_VIEWPORT;
} else {
OUT_PKT0(ring, REG_A3XX_RB_DEPTH_CONTROL, 1);
OUT_RING(ring, A3XX_RB_DEPTH_CONTROL_ZFUNC(FUNC_NEVER));
}
if (buffers & PIPE_CLEAR_STENCIL) {
OUT_PKT0(ring, REG_A3XX_RB_STENCILREFMASK, 2);
OUT_RING(ring, A3XX_RB_STENCILREFMASK_STENCILREF(stencil) |
A3XX_RB_STENCILREFMASK_STENCILMASK(stencil) |
A3XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
OUT_RING(ring, A3XX_RB_STENCILREFMASK_STENCILREF(0) |
A3XX_RB_STENCILREFMASK_STENCILMASK(0) |
0xff000000 | // XXX ???
A3XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
OUT_PKT0(ring, REG_A3XX_RB_STENCIL_CONTROL, 1);
OUT_RING(ring, A3XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
A3XX_RB_STENCIL_CONTROL_FUNC(FUNC_ALWAYS) |
A3XX_RB_STENCIL_CONTROL_FAIL(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZPASS(STENCIL_REPLACE) |
A3XX_RB_STENCIL_CONTROL_ZFAIL(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_FUNC_BF(FUNC_NEVER) |
A3XX_RB_STENCIL_CONTROL_FAIL_BF(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZPASS_BF(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZFAIL_BF(STENCIL_KEEP));
} else {
OUT_PKT0(ring, REG_A3XX_RB_STENCILREFMASK, 2);
OUT_RING(ring, A3XX_RB_STENCILREFMASK_STENCILREF(0) |
A3XX_RB_STENCILREFMASK_STENCILMASK(0) |
A3XX_RB_STENCILREFMASK_STENCILWRITEMASK(0));
OUT_RING(ring, A3XX_RB_STENCILREFMASK_BF_STENCILREF(0) |
A3XX_RB_STENCILREFMASK_BF_STENCILMASK(0) |
A3XX_RB_STENCILREFMASK_BF_STENCILWRITEMASK(0));
OUT_PKT0(ring, REG_A3XX_RB_STENCIL_CONTROL, 1);
OUT_RING(ring, A3XX_RB_STENCIL_CONTROL_FUNC(FUNC_NEVER) |
A3XX_RB_STENCIL_CONTROL_FAIL(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZPASS(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZFAIL(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_FUNC_BF(FUNC_NEVER) |
A3XX_RB_STENCIL_CONTROL_FAIL_BF(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZPASS_BF(STENCIL_KEEP) |
A3XX_RB_STENCIL_CONTROL_ZFAIL_BF(STENCIL_KEEP));
}
if (buffers & PIPE_CLEAR_COLOR) {
ce = 0xf;
} else {
ce = 0x0;
}
for (i = 0; i < 4; i++) {
OUT_PKT0(ring, REG_A3XX_RB_MRT_CONTROL(i), 1);
OUT_RING(ring, A3XX_RB_MRT_CONTROL_ROP_CODE(ROP_COPY) |
A3XX_RB_MRT_CONTROL_DITHER_MODE(DITHER_ALWAYS) |
A3XX_RB_MRT_CONTROL_COMPONENT_ENABLE(ce));
OUT_PKT0(ring, REG_A3XX_RB_MRT_BLEND_CONTROL(i), 1);
OUT_RING(ring, A3XX_RB_MRT_BLEND_CONTROL_RGB_SRC_FACTOR(FACTOR_ONE) |
A3XX_RB_MRT_BLEND_CONTROL_RGB_BLEND_OPCODE(BLEND_DST_PLUS_SRC) |
A3XX_RB_MRT_BLEND_CONTROL_RGB_DEST_FACTOR(FACTOR_ZERO) |
A3XX_RB_MRT_BLEND_CONTROL_ALPHA_SRC_FACTOR(FACTOR_ONE) |
A3XX_RB_MRT_BLEND_CONTROL_ALPHA_BLEND_OPCODE(BLEND_DST_PLUS_SRC) |
A3XX_RB_MRT_BLEND_CONTROL_ALPHA_DEST_FACTOR(FACTOR_ZERO) |
A3XX_RB_MRT_BLEND_CONTROL_CLAMP_ENABLE);
}
OUT_PKT0(ring, REG_A3XX_GRAS_SU_MODE_CONTROL, 1);
OUT_RING(ring, A3XX_GRAS_SU_MODE_CONTROL_LINEHALFWIDTH(0));
fd3_emit_vertex_bufs(ring, fd3_shader_variant(ctx->solid_prog.vp, key),
(struct fd3_vertex_buf[]) {{
.prsc = fd3_ctx->solid_vbuf,
.stride = 12,
.format = PIPE_FORMAT_R32G32B32_FLOAT,
}}, 1);
fd_wfi(ctx, ring);
fd3_emit_constant(ring, SB_FRAG_SHADER, 0, 0, 4, color->ui, NULL);
OUT_PKT0(ring, REG_A3XX_PC_PRIM_VTX_CNTL, 1);
OUT_RING(ring, A3XX_PC_PRIM_VTX_CNTL_STRIDE_IN_VPC(0) |
A3XX_PC_PRIM_VTX_CNTL_POLYMODE_FRONT_PTYPE(PC_DRAW_TRIANGLES) |
A3XX_PC_PRIM_VTX_CNTL_POLYMODE_BACK_PTYPE(PC_DRAW_TRIANGLES) |
A3XX_PC_PRIM_VTX_CNTL_PROVOKING_VTX_LAST);
OUT_PKT0(ring, REG_A3XX_VFD_INDEX_MIN, 4);
OUT_RING(ring, 0); /* VFD_INDEX_MIN */
OUT_RING(ring, 2); /* VFD_INDEX_MAX */
OUT_RING(ring, 0); /* VFD_INSTANCEID_OFFSET */
OUT_RING(ring, 0); /* VFD_INDEX_OFFSET */
OUT_PKT0(ring, REG_A3XX_PC_RESTART_INDEX, 1);
OUT_RING(ring, 0xffffffff); /* PC_RESTART_INDEX */
OUT_PKT3(ring, CP_EVENT_WRITE, 1);
OUT_RING(ring, PERFCOUNTER_STOP);
fd_draw(ctx, ring, DI_PT_RECTLIST, USE_VISIBILITY,
DI_SRC_SEL_AUTO_INDEX, 2, INDEX_SIZE_IGN, 0, 0, NULL);
}
void
fd3_draw_init(struct pipe_context *pctx)
{
struct fd_context *ctx = fd_context(pctx);
ctx->draw = fd3_draw;
ctx->clear = fd3_clear;
}