freedreno/a4xx: use generic clear path

Signed-off-by: Rob Clark <[email protected]>

2 files changed, 4 insertions, 215 deletions

diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_context.c b/src/gallium/drivers/freedreno/a4xx/fd4_context.c
index 8ef715f1b7f..291df2d03fc 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_context.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_context.c
@@ -54,7 +54,7 @@ fd4_context_destroy(struct pipe_context *pctx)
 	fd_context_destroy(pctx);
 }
 
-static const uint8_t primtypes[PIPE_PRIM_MAX] = {
+static const uint8_t primtypes[] = {
 		[PIPE_PRIM_POINTS]         = DI_PT_POINTLIST,
 		[PIPE_PRIM_LINES]          = DI_PT_LINELIST,
 		[PIPE_PRIM_LINE_STRIP]     = DI_PT_LINESTRIP,
@@ -62,6 +62,7 @@ static const uint8_t primtypes[PIPE_PRIM_MAX] = {
 		[PIPE_PRIM_TRIANGLES]      = DI_PT_TRILIST,
 		[PIPE_PRIM_TRIANGLE_STRIP] = DI_PT_TRISTRIP,
 		[PIPE_PRIM_TRIANGLE_FAN]   = DI_PT_TRIFAN,
+		[PIPE_PRIM_MAX]            = DI_PT_RECTLIST,  /* internal clear blits */
 };
 
 struct pipe_context *
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_draw.c b/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
index 200e4f26876..06d16fe934f 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
@@ -131,9 +131,8 @@ fd4_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 			.vclamp_color = ctx->rasterizer->clamp_vertex_color,
 			.fclamp_color = ctx->rasterizer->clamp_fragment_color,
 			.rasterflat = ctx->rasterizer->flatshade,
-			// 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),
+			.half_precision = ctx->in_blit &&
+					fd_half_precision(&ctx->batch->framebuffer),
 			.ucp_enables = ctx->rasterizer->clip_plane_enable,
 			.has_per_samp = (fd4_ctx->fsaturate || fd4_ctx->vsaturate ||
 					fd4_ctx->fastc_srgb || fd4_ctx->vastc_srgb),
@@ -193,220 +192,9 @@ fd4_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 	return true;
 }
 
-/* clear operations ignore viewport state, so we need to reset it
- * based on framebuffer state:
- */
-static void
-reset_viewport(struct fd_ringbuffer *ring, struct pipe_framebuffer_state *pfb)
-{
-	float half_width = pfb->width * 0.5f;
-	float half_height = pfb->height * 0.5f;
-
-	OUT_PKT0(ring, REG_A4XX_GRAS_CL_VPORT_XOFFSET_0, 4);
-	OUT_RING(ring, A4XX_GRAS_CL_VPORT_XOFFSET_0(half_width));
-	OUT_RING(ring, A4XX_GRAS_CL_VPORT_XSCALE_0(half_width));
-	OUT_RING(ring, A4XX_GRAS_CL_VPORT_YOFFSET_0(half_height));
-	OUT_RING(ring, A4XX_GRAS_CL_VPORT_YSCALE_0(-half_height));
-}
-
-/* TODO maybe we should just migrate u_blitter for clear and do it in
- * core (so we get normal draw pass state mgmt and binning).. That should
- * work well enough for a3xx/a4xx (but maybe not a2xx?)
- */
-
-static void
-fd4_clear_binning(struct fd_context *ctx, unsigned dirty)
-{
-	struct fd_ringbuffer *ring = ctx->batch->binning;
-	struct fd4_emit emit = {
-		.debug = &ctx->debug,
-		.vtx  = &ctx->solid_vbuf_state,
-		.prog = &ctx->solid_prog,
-		.key = {
-			.binning_pass = true,
-			.half_precision = true,
-		},
-		.dirty = dirty,
-	};
-
-	fd4_emit_state(ctx, ring, &emit);
-	fd4_emit_vertex_bufs(ring, &emit);
-	reset_viewport(ring, &ctx->batch->framebuffer);
-
-	OUT_PKT0(ring, REG_A4XX_PC_PRIM_VTX_CNTL, 2);
-	OUT_RING(ring, A4XX_PC_PRIM_VTX_CNTL_VAROUT(0) |
-			A4XX_PC_PRIM_VTX_CNTL_PROVOKING_VTX_LAST);
-	OUT_RING(ring, A4XX_PC_PRIM_VTX_CNTL2_POLYMODE_FRONT_PTYPE(PC_DRAW_TRIANGLES) |
-			A4XX_PC_PRIM_VTX_CNTL2_POLYMODE_BACK_PTYPE(PC_DRAW_TRIANGLES));
-
-	OUT_PKT0(ring, REG_A4XX_GRAS_ALPHA_CONTROL, 1);
-	OUT_RING(ring, 0x00000002);
-
-	fd4_draw(ctx->batch, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
-			DI_SRC_SEL_AUTO_INDEX, 2, 1, INDEX_SIZE_IGN, 0, 0, NULL);
-}
-
-static void
-fd4_clear(struct fd_context *ctx, unsigned buffers,
-		const union pipe_color_union *color, double depth, unsigned stencil)
-{
-	struct fd_ringbuffer *ring = ctx->batch->draw;
-	struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
-	unsigned char mrt_comp[A4XX_MAX_RENDER_TARGETS] = {0};
-	unsigned dirty = ctx->dirty;
-	unsigned i;
-	struct fd4_emit emit = {
-		.debug = &ctx->debug,
-		.vtx  = &ctx->solid_vbuf_state,
-		.prog = &ctx->solid_prog,
-		.key = {
-			.half_precision = fd_half_precision(pfb),
-		},
-	};
-
-	dirty &= FD_DIRTY_FRAMEBUFFER | FD_DIRTY_SCISSOR;
-	dirty |= FD_DIRTY_PROG;
-	emit.dirty = dirty;
-
-	fd4_clear_binning(ctx, dirty);
-
-	OUT_PKT0(ring, REG_A4XX_PC_PRIM_VTX_CNTL, 1);
-	OUT_RING(ring, A4XX_PC_PRIM_VTX_CNTL_PROVOKING_VTX_LAST);
-
-	/* emit generic state now: */
-	fd4_emit_state(ctx, ring, &emit);
-	reset_viewport(ring, pfb);
-
-	if (buffers & PIPE_CLEAR_DEPTH) {
-		OUT_PKT0(ring, REG_A4XX_RB_DEPTH_CONTROL, 1);
-		OUT_RING(ring, A4XX_RB_DEPTH_CONTROL_Z_WRITE_ENABLE |
-				A4XX_RB_DEPTH_CONTROL_Z_ENABLE |
-				A4XX_RB_DEPTH_CONTROL_ZFUNC(FUNC_ALWAYS));
-
-		fd_wfi(ctx->batch, ring);
-		OUT_PKT0(ring, REG_A4XX_GRAS_CL_VPORT_ZOFFSET_0, 2);
-		OUT_RING(ring, A4XX_GRAS_CL_VPORT_ZOFFSET_0(0.0));
-		OUT_RING(ring, A4XX_GRAS_CL_VPORT_ZSCALE_0(depth));
-		ctx->dirty |= FD_DIRTY_VIEWPORT;
-	} else {
-		OUT_PKT0(ring, REG_A4XX_RB_DEPTH_CONTROL, 1);
-		OUT_RING(ring, A4XX_RB_DEPTH_CONTROL_ZFUNC(FUNC_NEVER));
-	}
-
-	if (buffers & PIPE_CLEAR_STENCIL) {
-		OUT_PKT0(ring, REG_A4XX_RB_STENCILREFMASK, 2);
-		OUT_RING(ring, A4XX_RB_STENCILREFMASK_STENCILREF(stencil) |
-				A4XX_RB_STENCILREFMASK_STENCILMASK(stencil) |
-				A4XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
-		OUT_RING(ring, A4XX_RB_STENCILREFMASK_STENCILREF(0) |
-				A4XX_RB_STENCILREFMASK_STENCILMASK(0) |
-				0xff000000 | // XXX ???
-				A4XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
-
-		OUT_PKT0(ring, REG_A4XX_RB_STENCIL_CONTROL, 2);
-		OUT_RING(ring, A4XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
-				A4XX_RB_STENCIL_CONTROL_FUNC(FUNC_ALWAYS) |
-				A4XX_RB_STENCIL_CONTROL_FAIL(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZPASS(STENCIL_REPLACE) |
-				A4XX_RB_STENCIL_CONTROL_ZFAIL(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_FUNC_BF(FUNC_NEVER) |
-				A4XX_RB_STENCIL_CONTROL_FAIL_BF(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZPASS_BF(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZFAIL_BF(STENCIL_KEEP));
-		OUT_RING(ring, A4XX_RB_STENCIL_CONTROL2_STENCIL_BUFFER);
-	} else {
-		OUT_PKT0(ring, REG_A4XX_RB_STENCILREFMASK, 2);
-		OUT_RING(ring, A4XX_RB_STENCILREFMASK_STENCILREF(0) |
-				A4XX_RB_STENCILREFMASK_STENCILMASK(0) |
-				A4XX_RB_STENCILREFMASK_STENCILWRITEMASK(0));
-		OUT_RING(ring, A4XX_RB_STENCILREFMASK_BF_STENCILREF(0) |
-				A4XX_RB_STENCILREFMASK_BF_STENCILMASK(0) |
-				A4XX_RB_STENCILREFMASK_BF_STENCILWRITEMASK(0));
-
-		OUT_PKT0(ring, REG_A4XX_RB_STENCIL_CONTROL, 2);
-		OUT_RING(ring, A4XX_RB_STENCIL_CONTROL_FUNC(FUNC_NEVER) |
-				A4XX_RB_STENCIL_CONTROL_FAIL(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZPASS(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZFAIL(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_FUNC_BF(FUNC_NEVER) |
-				A4XX_RB_STENCIL_CONTROL_FAIL_BF(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZPASS_BF(STENCIL_KEEP) |
-				A4XX_RB_STENCIL_CONTROL_ZFAIL_BF(STENCIL_KEEP));
-		OUT_RING(ring, 0x00000000); /* RB_STENCIL_CONTROL2 */
-	}
-
-	if (buffers & PIPE_CLEAR_COLOR) {
-		OUT_PKT0(ring, REG_A4XX_RB_ALPHA_CONTROL, 1);
-		OUT_RING(ring, A4XX_RB_ALPHA_CONTROL_ALPHA_TEST_FUNC(FUNC_NEVER));
-	}
-
-	for (i = 0; i < A4XX_MAX_RENDER_TARGETS; i++) {
-		mrt_comp[i] = (buffers & (PIPE_CLEAR_COLOR0 << i)) ? 0xf : 0x0;
-
-		OUT_PKT0(ring, REG_A4XX_RB_MRT_CONTROL(i), 1);
-		OUT_RING(ring, A4XX_RB_MRT_CONTROL_ROP_CODE(ROP_COPY) |
-				A4XX_RB_MRT_CONTROL_COMPONENT_ENABLE(0xf));
-
-		OUT_PKT0(ring, REG_A4XX_RB_MRT_BLEND_CONTROL(i), 1);
-		OUT_RING(ring, A4XX_RB_MRT_BLEND_CONTROL_RGB_SRC_FACTOR(FACTOR_ONE) |
-				A4XX_RB_MRT_BLEND_CONTROL_RGB_BLEND_OPCODE(BLEND_DST_PLUS_SRC) |
-				A4XX_RB_MRT_BLEND_CONTROL_RGB_DEST_FACTOR(FACTOR_ZERO) |
-				A4XX_RB_MRT_BLEND_CONTROL_ALPHA_SRC_FACTOR(FACTOR_ONE) |
-				A4XX_RB_MRT_BLEND_CONTROL_ALPHA_BLEND_OPCODE(BLEND_DST_PLUS_SRC) |
-				A4XX_RB_MRT_BLEND_CONTROL_ALPHA_DEST_FACTOR(FACTOR_ZERO));
-	}
-
-	OUT_PKT0(ring, REG_A4XX_RB_RENDER_COMPONENTS, 1);
-	OUT_RING(ring, A4XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
-			A4XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
-			A4XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
-			A4XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
-			A4XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
-			A4XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
-			A4XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
-			A4XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
-
-	fd4_emit_vertex_bufs(ring, &emit);
-
-	OUT_PKT0(ring, REG_A4XX_GRAS_ALPHA_CONTROL, 1);
-	OUT_RING(ring, 0x0);          /* XXX GRAS_ALPHA_CONTROL */
-
-	OUT_PKT0(ring, REG_A4XX_GRAS_CLEAR_CNTL, 1);
-	OUT_RING(ring, 0x00000000);
-
-	/* until fastclear works: */
-	fd4_emit_const(ring, SHADER_FRAGMENT, 0, 0, 4, color->ui, NULL);
-
-	OUT_PKT0(ring, REG_A4XX_VFD_INDEX_OFFSET, 2);
-	OUT_RING(ring, 0);            /* VFD_INDEX_OFFSET */
-	OUT_RING(ring, 0);            /* ??? UNKNOWN_2209 */
-
-	OUT_PKT0(ring, REG_A4XX_PC_RESTART_INDEX, 1);
-	OUT_RING(ring, 0xffffffff);   /* PC_RESTART_INDEX */
-
-	OUT_PKT3(ring, CP_UNKNOWN_1A, 1);
-	OUT_RING(ring, 0x00000001);
-
-	fd4_draw(ctx->batch, ring, DI_PT_RECTLIST, USE_VISIBILITY,
-			DI_SRC_SEL_AUTO_INDEX, 2, 1, INDEX_SIZE_IGN, 0, 0, NULL);
-
-	OUT_PKT3(ring, CP_UNKNOWN_1A, 1);
-	OUT_RING(ring, 0x00000000);
-
-	OUT_PKT0(ring, REG_A4XX_GRAS_CLEAR_CNTL, 1);
-	OUT_RING(ring, A4XX_GRAS_CLEAR_CNTL_NOT_FASTCLEAR);
-
-	OUT_PKT0(ring, REG_A4XX_GRAS_SC_CONTROL, 1);
-	OUT_RING(ring, A4XX_GRAS_SC_CONTROL_RENDER_MODE(RB_RENDERING_PASS) |
-			A4XX_GRAS_SC_CONTROL_MSAA_DISABLE |
-			A4XX_GRAS_SC_CONTROL_MSAA_SAMPLES(MSAA_ONE) |
-			A4XX_GRAS_SC_CONTROL_RASTER_MODE(0));
-}
-
 void
 fd4_draw_init(struct pipe_context *pctx)
 {
 	struct fd_context *ctx = fd_context(pctx);
 	ctx->draw_vbo = fd4_draw_vbo;
-	ctx->clear = fd4_clear;
 }