freedreno/a4xx: better workaround for astc+srgb

This *seems* like a hw bug, and maybe only applies to certain a4xx
variants/revisions.  But setting the SRGB bit in sampler view state
(texconst0) causes invalid alpha for ASTC textures.  Work around this
setting up a second texture state and using that to sample alpha
separately.

This way, srgb->linear conversion happens in hw *prior* to
interpolation.

This fixes 546 dEQP tests: dEQP-GLES3.functional.texture.*astc*srgb*

Signed-off-by: Rob Clark <robclark@freedesktop.org>

10 files changed, 195 insertions, 22 deletions

diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_context.h b/src/gallium/drivers/freedreno/a4xx/fd4_context.h
index 8996de932b8..9467fc529bf 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_context.h
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_context.h
@@ -85,6 +85,9 @@ struct fd4_context {
 	 */
 	uint16_t fsaturate_s, fsaturate_t, fsaturate_r;
 
+	/* bitmask of samplers which need astc srgb workaround: */
+	uint16_t vastc_srgb, fastc_srgb;
+
 	/* some state changes require a different shader variant.  Keep
 	 * track of this so we know when we need to re-emit shader state
 	 * due to variant change.  See fixup_shader_state()
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_draw.c b/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
index e874d223187..68e1f53e2dd 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_draw.c
@@ -93,12 +93,14 @@ fixup_shader_state(struct fd_context *ctx, struct ir3_shader_key *key)
 		if (last_key->has_per_samp || key->has_per_samp) {
 			if ((last_key->vsaturate_s != key->vsaturate_s) ||
 					(last_key->vsaturate_t != key->vsaturate_t) ||
-					(last_key->vsaturate_r != key->vsaturate_r))
+					(last_key->vsaturate_r != key->vsaturate_r) ||
+					(last_key->vastc_srgb != key->vastc_srgb))
 				ctx->prog.dirty |= FD_SHADER_DIRTY_VP;
 
 			if ((last_key->fsaturate_s != key->fsaturate_s) ||
 					(last_key->fsaturate_t != key->fsaturate_t) ||
-					(last_key->fsaturate_r != key->fsaturate_r))
+					(last_key->fsaturate_r != key->fsaturate_r) ||
+					(last_key->fastc_srgb != key->fastc_srgb))
 				ctx->prog.dirty |= FD_SHADER_DIRTY_FP;
 		}
 
@@ -132,13 +134,16 @@ fd4_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
 			// ie. float16 and smaller use half, float32 use full..
 			.half_precision = !!(fd_mesa_debug & FD_DBG_FRAGHALF),
 			.ucp_enables = ctx->rasterizer->clip_plane_enable,
-			.has_per_samp = (fd4_ctx->fsaturate || fd4_ctx->vsaturate),
+			.has_per_samp = (fd4_ctx->fsaturate || fd4_ctx->vsaturate ||
+					fd4_ctx->fastc_srgb || fd4_ctx->vastc_srgb),
 			.vsaturate_s = fd4_ctx->vsaturate_s,
 			.vsaturate_t = fd4_ctx->vsaturate_t,
 			.vsaturate_r = fd4_ctx->vsaturate_r,
 			.fsaturate_s = fd4_ctx->fsaturate_s,
 			.fsaturate_t = fd4_ctx->fsaturate_t,
 			.fsaturate_r = fd4_ctx->fsaturate_r,
+			.vastc_srgb = fd4_ctx->vastc_srgb,
+			.fastc_srgb = fd4_ctx->fastc_srgb,
 		},
 		.rasterflat = ctx->rasterizer->flatshade,
 		.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable,
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_emit.c b/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
index ba5d48909fa..27614f07de5 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_emit.c
@@ -123,7 +123,8 @@ fd4_emit_const_bo(struct fd_ringbuffer *ring, enum shader_t type, boolean write,
 
 static void
 emit_textures(struct fd_context *ctx, struct fd_ringbuffer *ring,
-		enum adreno_state_block sb, struct fd_texture_stateobj *tex)
+		enum adreno_state_block sb, struct fd_texture_stateobj *tex,
+		const struct ir3_shader_variant *v)
 {
 	static const uint32_t bcolor_reg[] = {
 			[SB_VERT_TEX] = REG_A4XX_TPL1_TP_VS_BORDER_COLOR_BASE_ADDR,
@@ -174,12 +175,14 @@ emit_textures(struct fd_context *ctx, struct fd_ringbuffer *ring,
 	}
 
 	if (tex->num_textures > 0) {
+		unsigned num_textures = tex->num_textures + v->astc_srgb.count;
+
 		/* emit texture state: */
-		OUT_PKT3(ring, CP_LOAD_STATE, 2 + (8 * tex->num_textures));
+		OUT_PKT3(ring, CP_LOAD_STATE, 2 + (8 * num_textures));
 		OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(0) |
 				CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 				CP_LOAD_STATE_0_STATE_BLOCK(sb) |
-				CP_LOAD_STATE_0_NUM_UNIT(tex->num_textures));
+				CP_LOAD_STATE_0_NUM_UNIT(num_textures));
 		OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS) |
 				CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 		for (i = 0; i < tex->num_textures; i++) {
@@ -202,6 +205,34 @@ emit_textures(struct fd_context *ctx, struct fd_ringbuffer *ring,
 			OUT_RING(ring, 0x00000000);
 			OUT_RING(ring, 0x00000000);
 		}
+
+		for (i = 0; i < v->astc_srgb.count; i++) {
+			static const struct fd4_pipe_sampler_view dummy_view = {};
+			const struct fd4_pipe_sampler_view *view;
+			unsigned idx = v->astc_srgb.orig_idx[i];
+
+			view = tex->textures[idx] ?
+					fd4_pipe_sampler_view(tex->textures[idx]) :
+					&dummy_view;
+
+			debug_assert(view->texconst0 & A4XX_TEX_CONST_0_SRGB);
+
+			OUT_RING(ring, view->texconst0 & ~A4XX_TEX_CONST_0_SRGB);
+			OUT_RING(ring, view->texconst1);
+			OUT_RING(ring, view->texconst2);
+			OUT_RING(ring, view->texconst3);
+			if (view->base.texture) {
+				struct fd_resource *rsc = fd_resource(view->base.texture);
+				OUT_RELOC(ring, rsc->bo, view->offset, view->texconst4, 0);
+			} else {
+				OUT_RING(ring, 0x00000000);
+			}
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+			OUT_RING(ring, 0x00000000);
+		}
+	} else {
+		debug_assert(v->astc_srgb.count == 0);
 	}
 
 	OUT_PKT0(ring, bcolor_reg[sb], 1);
@@ -681,14 +712,14 @@ fd4_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 
 	if (dirty & FD_DIRTY_VERTTEX) {
 		if (vp->has_samp)
-			emit_textures(ctx, ring, SB_VERT_TEX, &ctx->verttex);
+			emit_textures(ctx, ring, SB_VERT_TEX, &ctx->verttex, vp);
 		else
 			dirty &= ~FD_DIRTY_VERTTEX;
 	}
 
 	if (dirty & FD_DIRTY_FRAGTEX) {
 		if (fp->has_samp)
-			emit_textures(ctx, ring, SB_FRAG_TEX, &ctx->fragtex);
+			emit_textures(ctx, ring, SB_FRAG_TEX, &ctx->fragtex, fp);
 		else
 			dirty &= ~FD_DIRTY_FRAGTEX;
 	}
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_texture.c b/src/gallium/drivers/freedreno/a4xx/fd4_texture.c
index 38348580e21..6d9ecb7da24 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_texture.c
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_texture.c
@@ -209,6 +209,13 @@ tex_type(unsigned target)
 	}
 }
 
+static bool
+use_astc_srgb_workaround(struct pipe_context *pctx, enum pipe_format format)
+{
+	return (fd_screen(pctx->screen)->gpu_id == 420) &&
+		(util_format_description(format)->layout == UTIL_FORMAT_LAYOUT_ASTC);
+}
+
 static struct pipe_sampler_view *
 fd4_sampler_view_create(struct pipe_context *pctx, struct pipe_resource *prsc,
 		const struct pipe_sampler_view *cso)
@@ -233,8 +240,11 @@ fd4_sampler_view_create(struct pipe_context *pctx, struct pipe_resource *prsc,
 		fd4_tex_swiz(cso->format, cso->swizzle_r, cso->swizzle_g,
 				cso->swizzle_b, cso->swizzle_a);
 
-	if (util_format_is_srgb(cso->format))
+	if (util_format_is_srgb(cso->format)) {
+		if (use_astc_srgb_workaround(pctx, cso->format))
+			so->astc_srgb = true;
 		so->texconst0 |= A4XX_TEX_CONST_0_SRGB;
+	}
 
 	if (cso->target == PIPE_BUFFER) {
 		unsigned elements = cso->u.buf.last_element -
@@ -296,11 +306,39 @@ fd4_sampler_view_create(struct pipe_context *pctx, struct pipe_resource *prsc,
 	return &so->base;
 }
 
+static void
+fd4_set_sampler_views(struct pipe_context *pctx, unsigned shader,
+		unsigned start, unsigned nr,
+		struct pipe_sampler_view **views)
+{
+	struct fd_context *ctx = fd_context(pctx);
+	struct fd4_context *fd4_ctx = fd4_context(ctx);
+	uint16_t astc_srgb = 0;
+	unsigned i;
+
+	for (i = 0; i < nr; i++) {
+		if (views[i]) {
+			struct fd4_pipe_sampler_view *view =
+					fd4_pipe_sampler_view(views[i]);
+			if (view->astc_srgb)
+				astc_srgb |= (1 << i);
+		}
+	}
+
+	fd_set_sampler_views(pctx, shader, start, nr, views);
+
+	if (shader == PIPE_SHADER_FRAGMENT) {
+		fd4_ctx->fastc_srgb = astc_srgb;
+	} else if (shader == PIPE_SHADER_VERTEX) {
+		fd4_ctx->vastc_srgb = astc_srgb;
+	}
+}
+
 void
 fd4_texture_init(struct pipe_context *pctx)
 {
 	pctx->create_sampler_state = fd4_sampler_state_create;
 	pctx->bind_sampler_states = fd4_sampler_states_bind;
 	pctx->create_sampler_view = fd4_sampler_view_create;
-	pctx->set_sampler_views = fd_set_sampler_views;
+	pctx->set_sampler_views = fd4_set_sampler_views;
 }
diff --git a/src/gallium/drivers/freedreno/a4xx/fd4_texture.h b/src/gallium/drivers/freedreno/a4xx/fd4_texture.h
index 6ca34ade60d..21ceadd8c63 100644
--- a/src/gallium/drivers/freedreno/a4xx/fd4_texture.h
+++ b/src/gallium/drivers/freedreno/a4xx/fd4_texture.h
@@ -53,6 +53,7 @@ struct fd4_pipe_sampler_view {
 	struct pipe_sampler_view base;
 	uint32_t texconst0, texconst1, texconst2, texconst3, texconst4;
 	uint32_t offset;
+	bool astc_srgb;
 };
 
 static inline struct fd4_pipe_sampler_view *
diff --git a/src/gallium/drivers/freedreno/ir3/ir3.h b/src/gallium/drivers/freedreno/ir3/ir3.h
index f68275e568c..a40d3aa3b40 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3.h
+++ b/src/gallium/drivers/freedreno/ir3/ir3.h
@@ -377,6 +377,12 @@ struct ir3 {
 	unsigned keeps_count, keeps_sz;
 	struct ir3_instruction **keeps;
 
+	/* Track texture sample instructions which need texture state
+	 * patched in (for astc-srgb workaround):
+	 */
+	unsigned astc_srgb_count, astc_srgb_sz;
+	struct ir3_instruction **astc_srgb;
+
 	/* List of blocks: */
 	struct list_head block_list;
 
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_cmdline.c b/src/gallium/drivers/freedreno/ir3/ir3_cmdline.c
index 027673afe1c..b8b9e4a0518 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_cmdline.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_cmdline.c
@@ -94,6 +94,7 @@ static void print_usage(void)
 	printf("    --saturate-s MASK - bitmask of samplers to saturate S coord\n");
 	printf("    --saturate-t MASK - bitmask of samplers to saturate T coord\n");
 	printf("    --saturate-r MASK - bitmask of samplers to saturate R coord\n");
+	printf("    --astc-srgb MASK  - bitmask of samplers to enable astc-srgb workaround\n");
 	printf("    --stream-out      - enable stream-out (aka transform feedback)\n");
 	printf("    --ucp MASK        - bitmask of enabled user-clip-planes\n");
 	printf("    --gpu GPU_ID      - specify gpu-id (default 320)\n");
@@ -174,6 +175,13 @@ int main(int argc, char **argv)
 			continue;
 		}
 
+		if (!strcmp(argv[n], "--astc-srgb")) {
+			debug_printf(" %s %s", argv[n], argv[n+1]);
+			key.vastc_srgb = key.fastc_srgb = strtol(argv[n+1], NULL, 0);
+			n += 2;
+			continue;
+		}
+
 		if (!strcmp(argv[n], "--stream-out")) {
 			struct pipe_stream_output_info *so = &s.stream_output;
 			debug_printf(" %s", argv[n]);
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
index 940ca7744a2..abdb1c27c91 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
@@ -108,8 +108,10 @@ struct ir3_compile {
 	 */
 	bool array_index_add_half;
 
-	/* for looking up which system value is which */
-	unsigned sysval_semantics[8];
+	/* on a4xx, bitmask of samplers which need astc+srgb workaround: */
+	unsigned astc_srgb;
+
+	unsigned max_texture_index;
 
 	/* set if we encounter something we can't handle yet, so we
 	 * can bail cleanly and fallback to TGSI compiler f/e
@@ -134,6 +136,12 @@ compile_init(struct ir3_compiler *compiler,
 		ctx->levels_add_one = false;
 		ctx->unminify_coords = false;
 		ctx->array_index_add_half = true;
+
+		if (so->type == SHADER_VERTEX)
+			ctx->astc_srgb = so->key.vastc_srgb;
+		else if (so->type == SHADER_FRAGMENT)
+			ctx->astc_srgb = so->key.fastc_srgb;
+
 	} else {
 		/* no special handling for "flat" */
 		ctx->flat_bypass = false;
@@ -620,14 +628,14 @@ create_driver_param(struct ir3_compile *ctx, enum ir3_driver_param dp)
  */
 static void
 split_dest(struct ir3_block *block, struct ir3_instruction **dst,
-		struct ir3_instruction *src, unsigned n)
+		struct ir3_instruction *src, unsigned base, unsigned n)
 {
 	struct ir3_instruction *prev = NULL;
 	for (int i = 0, j = 0; i < n; i++) {
 		struct ir3_instruction *split = ir3_instr_create(block, OPC_META_FO);
 		ir3_reg_create(split, 0, IR3_REG_SSA);
 		ir3_reg_create(split, 0, IR3_REG_SSA)->instr = src;
-		split->fo.off = i;
+		split->fo.off = i + base;
 
 		if (prev) {
 			split->cp.left = prev;
@@ -637,7 +645,7 @@ split_dest(struct ir3_block *block, struct ir3_instruction **dst,
 		}
 		prev = split;
 
-		if (src->regs[0]->wrmask & (1 << i))
+		if (src->regs[0]->wrmask & (1 << (i + base)))
 			dst[j++] = split;
 	}
 }
@@ -1543,12 +1551,35 @@ emit_tex(struct ir3_compile *ctx, nir_tex_instr *tex)
 	if (opc == OPC_GETLOD)
 		type = TYPE_U32;
 
-	sam = ir3_SAM(b, opc, type, TGSI_WRITEMASK_XYZW,
-			flags, tex->texture_index, tex->texture_index,
-			create_collect(b, src0, nsrc0),
-			create_collect(b, src1, nsrc1));
+	unsigned tex_idx = tex->texture_index;
+
+	ctx->max_texture_index = MAX2(ctx->max_texture_index, tex_idx);
+
+	struct ir3_instruction *col0 = create_collect(b, src0, nsrc0);
+	struct ir3_instruction *col1 = create_collect(b, src1, nsrc1);
+
+	sam = ir3_SAM(b, opc, type, TGSI_WRITEMASK_XYZW, flags,
+			tex_idx, tex_idx, col0, col1);
+
+	if ((ctx->astc_srgb & (1 << tex_idx)) && !nir_tex_instr_is_query(tex)) {
+		/* only need first 3 components: */
+		sam->regs[0]->wrmask = 0x7;
+		split_dest(b, dst, sam, 0, 3);
 
-	split_dest(b, dst, sam, 4);
+		/* we need to sample the alpha separately with a non-ASTC
+		 * texture state:
+		 */
+		sam = ir3_SAM(b, opc, type, TGSI_WRITEMASK_W, flags,
+				tex_idx, tex_idx, col0, col1);
+
+		array_insert(ctx->ir->astc_srgb, sam);
+
+		/* fixup .w component: */
+		split_dest(b, &dst[3], sam, 3, 1);
+	} else {
+		/* normal (non-workaround) case: */
+		split_dest(b, dst, sam, 0, 4);
+	}
 
 	/* GETLOD returns results in 4.8 fixed point */
 	if (opc == OPC_GETLOD) {
@@ -1576,7 +1607,7 @@ emit_tex_query_levels(struct ir3_compile *ctx, nir_tex_instr *tex)
 	/* even though there is only one component, since it ends
 	 * up in .z rather than .x, we need a split_dest()
 	 */
-	split_dest(b, dst, sam, 3);
+	split_dest(b, dst, sam, 0, 3);
 
 	/* The # of levels comes from getinfo.z. We need to add 1 to it, since
 	 * the value in TEX_CONST_0 is zero-based.
@@ -1610,7 +1641,7 @@ emit_tex_txs(struct ir3_compile *ctx, nir_tex_instr *tex)
 	sam = ir3_SAM(b, OPC_GETSIZE, TYPE_U32, TGSI_WRITEMASK_XYZW, flags,
 			tex->texture_index, tex->texture_index, lod, NULL);
 
-	split_dest(b, dst, sam, 4);
+	split_dest(b, dst, sam, 0, 4);
 
 	/* Array size actually ends up in .w rather than .z. This doesn't
 	 * matter for miplevel 0, but for higher mips the value in z is
@@ -2268,6 +2299,40 @@ fixup_frag_inputs(struct ir3_compile *ctx)
 	ir->inputs = inputs;
 }
 
+/* Fixup tex sampler state for astc/srgb workaround instructions.  We
+ * need to assign the tex state indexes for these after we know the
+ * max tex index.
+ */
+static void
+fixup_astc_srgb(struct ir3_compile *ctx)
+{
+	struct ir3_shader_variant *so = ctx->so;
+	/* indexed by original tex idx, value is newly assigned alpha sampler
+	 * state tex idx.  Zero is invalid since there is at least one sampler
+	 * if we get here.
+	 */
+	unsigned alt_tex_state[16] = {0};
+	unsigned tex_idx = ctx->max_texture_index + 1;
+	unsigned idx = 0;
+
+	so->astc_srgb.base = tex_idx;
+
+	for (unsigned i = 0; i < ctx->ir->astc_srgb_count; i++) {
+		struct ir3_instruction *sam = ctx->ir->astc_srgb[i];
+
+		compile_assert(ctx, sam->cat5.tex < ARRAY_SIZE(alt_tex_state));
+
+		if (alt_tex_state[sam->cat5.tex] == 0) {
+			/* assign new alternate/alpha tex state slot: */
+			alt_tex_state[sam->cat5.tex] = tex_idx++;
+			so->astc_srgb.orig_idx[idx++] = sam->cat5.tex;
+			so->astc_srgb.count++;
+		}
+
+		sam->cat5.tex = alt_tex_state[sam->cat5.tex];
+	}
+}
+
 int
 ir3_compile_shader_nir(struct ir3_compiler *compiler,
 		struct ir3_shader_variant *so)
@@ -2433,6 +2498,9 @@ ir3_compile_shader_nir(struct ir3_compiler *compiler,
 		so->inputs[i].compmask = compmask;
 	}
 
+	if (ctx->astc_srgb)
+		fixup_astc_srgb(ctx);
+
 	/* We need to do legalize after (for frag shader's) the "bary.f"
 	 * offsets (inloc) have been assigned.
 	 */
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_shader.c b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
index c05b52e7a5e..435a565e61e 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_shader.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
@@ -223,6 +223,7 @@ ir3_shader_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 			key.vsaturate_s = 0;
 			key.vsaturate_t = 0;
 			key.vsaturate_r = 0;
+			key.vastc_srgb = 0;
 		}
 		break;
 	case SHADER_VERTEX:
@@ -233,6 +234,7 @@ ir3_shader_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 			key.fsaturate_s = 0;
 			key.fsaturate_t = 0;
 			key.fsaturate_r = 0;
+			key.fastc_srgb = 0;
 		}
 		break;
 	}
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_shader.h b/src/gallium/drivers/freedreno/ir3/ir3_shader.h
index c89dc29ff08..e81e80d328f 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_shader.h
+++ b/src/gallium/drivers/freedreno/ir3/ir3_shader.h
@@ -104,6 +104,9 @@ struct ir3_shader_key {
 	 * shader:
 	 */
 	uint16_t fsaturate_s, fsaturate_t, fsaturate_r;
+
+	/* bitmask of samplers which need astc srgb workaround: */
+	uint16_t vastc_srgb, fastc_srgb;
 };
 
 static inline bool
@@ -222,6 +225,14 @@ struct ir3_shader_variant {
 		uint32_t val[4];
 	} immediates[64];
 
+	/* for astc srgb workaround, the number/base of additional
+	 * alpha tex states we need, and index of original tex states
+	 */
+	struct {
+		unsigned base, count;
+		unsigned orig_idx[16];
+	} astc_srgb;
+
 	/* shader variants form a linked list: */
 	struct ir3_shader_variant *next;