ac/surface/gfx6: compute FMASK together with the color surface

instead of invoking FMASK computation separately.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

5 files changed, 149 insertions, 100 deletions

diff --git a/src/amd/common/ac_surface.c b/src/amd/common/ac_surface.c
index b2af1f70b69..b7423e5b520 100644
--- a/src/amd/common/ac_surface.c
+++ b/src/amd/common/ac_surface.c
@@ -227,8 +227,16 @@ ADDR_HANDLE amdgpu_addr_create(const struct radeon_info *info,
 	return addrCreateOutput.hLib;
 }
 
-static int surf_config_sanity(const struct ac_surf_config *config)
+static int surf_config_sanity(const struct ac_surf_config *config,
+			      unsigned flags)
 {
+	/* FMASK is allocated together with the color surface and can't be
+	 * allocated separately.
+	 */
+	assert(!(flags & RADEON_SURF_FMASK));
+	if (flags & RADEON_SURF_FMASK)
+		return -EINVAL;
+
 	/* all dimension must be at least 1 ! */
 	if (!config->info.width || !config->info.height || !config->info.depth ||
 	    !config->info.array_size || !config->info.levels)
@@ -445,7 +453,6 @@ static bool get_display_flag(const struct ac_surf_config *config,
 	unsigned bpe = surf->bpe;
 
 	if (surf->flags & RADEON_SURF_SCANOUT &&
-	    !(surf->flags & RADEON_SURF_FMASK) &&
 	    config->info.samples <= 1 &&
 	    surf->blk_w <= 2 && surf->blk_h == 1) {
 		/* subsampled */
@@ -556,9 +563,8 @@ static int gfx6_compute_surface(ADDR_HANDLE addrlib,
 
 	compressed = surf->blk_w == 4 && surf->blk_h == 4;
 
-	/* MSAA and FMASK require 2D tiling. */
-	if (config->info.samples > 1 ||
-	    (surf->flags & RADEON_SURF_FMASK))
+	/* MSAA requires 2D tiling. */
+	if (config->info.samples > 1)
 		mode = RADEON_SURF_MODE_2D;
 
 	/* DB doesn't support linear layouts. */
@@ -607,7 +613,7 @@ static int gfx6_compute_surface(ADDR_HANDLE addrlib,
 	/* Set the micro tile type. */
 	if (surf->flags & RADEON_SURF_SCANOUT)
 		AddrSurfInfoIn.tileType = ADDR_DISPLAYABLE;
-	else if (surf->flags & (RADEON_SURF_Z_OR_SBUFFER | RADEON_SURF_FMASK))
+	else if (surf->flags & RADEON_SURF_Z_OR_SBUFFER)
 		AddrSurfInfoIn.tileType = ADDR_DEPTH_SAMPLE_ORDER;
 	else
 		AddrSurfInfoIn.tileType = ADDR_NON_DISPLAYABLE;
@@ -615,7 +621,6 @@ static int gfx6_compute_surface(ADDR_HANDLE addrlib,
 	AddrSurfInfoIn.flags.color = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
 	AddrSurfInfoIn.flags.depth = (surf->flags & RADEON_SURF_ZBUFFER) != 0;
 	AddrSurfInfoIn.flags.cube = config->is_cube;
-	AddrSurfInfoIn.flags.fmask = (surf->flags & RADEON_SURF_FMASK) != 0;
 	AddrSurfInfoIn.flags.display = get_display_flag(config, surf);
 	AddrSurfInfoIn.flags.pow2Pad = config->info.levels > 1;
 	AddrSurfInfoIn.flags.tcCompatible = (surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE) != 0;
@@ -680,8 +685,6 @@ static int gfx6_compute_surface(ADDR_HANDLE addrlib,
 	if (AddrSurfInfoIn.tileMode >= ADDR_TM_2D_TILED_THIN1 &&
 	    surf->u.legacy.bankw && surf->u.legacy.bankh &&
 	    surf->u.legacy.mtilea && surf->u.legacy.tile_split) {
-		assert(!(surf->flags & RADEON_SURF_FMASK));
-
 		/* If any of these parameters are incorrect, the calculation
 		 * will fail. */
 		AddrTileInfoIn.banks = surf->u.legacy.num_banks;
@@ -828,6 +831,67 @@ static int gfx6_compute_surface(ADDR_HANDLE addrlib,
 		}
 	}
 
+	/* Compute FMASK. */
+	if (config->info.samples >= 2 && AddrSurfInfoIn.flags.color) {
+		ADDR_COMPUTE_FMASK_INFO_INPUT fin = {0};
+		ADDR_COMPUTE_FMASK_INFO_OUTPUT fout = {0};
+		ADDR_TILEINFO fmask_tile_info = {};
+
+		fin.size = sizeof(fin);
+		fout.size = sizeof(fout);
+
+		fin.tileMode = AddrSurfInfoOut.tileMode;
+		fin.pitch = AddrSurfInfoOut.pitch;
+		fin.height = config->info.height;
+		fin.numSlices = AddrSurfInfoIn.numSlices;
+		fin.numSamples = AddrSurfInfoIn.numSamples;
+		fin.numFrags = AddrSurfInfoIn.numFrags;
+		fin.tileIndex = AddrSurfInfoOut.tileIndex;
+		fout.pTileInfo = &fmask_tile_info;
+
+		r = AddrComputeFmaskInfo(addrlib, &fin, &fout);
+		if (r)
+			return r;
+
+		surf->u.legacy.fmask.size = fout.fmaskBytes;
+		surf->u.legacy.fmask.alignment = fout.baseAlign;
+		surf->u.legacy.fmask.tile_swizzle = 0;
+
+		surf->u.legacy.fmask.slice_tile_max =
+			(fout.pitch * fout.height) / 64;
+		if (surf->u.legacy.fmask.slice_tile_max)
+		    surf->u.legacy.fmask.slice_tile_max -= 1;
+
+		surf->u.legacy.fmask.tiling_index = fout.tileIndex;
+		surf->u.legacy.fmask.bankh = fout.pTileInfo->bankHeight;
+		surf->u.legacy.fmask.pitch_in_pixels = fout.pitch;
+
+		/* Compute tile swizzle for FMASK. */
+		if (config->info.fmask_surf_index &&
+		    !(surf->flags & RADEON_SURF_SHAREABLE)) {
+			ADDR_COMPUTE_BASE_SWIZZLE_INPUT xin = {0};
+			ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT xout = {0};
+
+			xin.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_INPUT);
+			xout.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT);
+
+			/* This counter starts from 1 instead of 0. */
+			xin.surfIndex = p_atomic_inc_return(config->info.fmask_surf_index);
+			xin.tileIndex = fout.tileIndex;
+			xin.macroModeIndex = fout.macroModeIndex;
+			xin.pTileInfo = fout.pTileInfo;
+			xin.tileMode = fin.tileMode;
+
+			int r = AddrComputeBaseSwizzle(addrlib, &xin, &xout);
+			if (r != ADDR_OK)
+				return r;
+
+			assert(xout.tileSwizzle <=
+			       u_bit_consecutive(0, sizeof(surf->tile_swizzle) * 8));
+			surf->u.legacy.fmask.tile_swizzle = xout.tileSwizzle;
+		}
+	}
+
 	/* Recalculate the whole DCC miptree size including disabled levels.
 	 * This is what addrlib does, but calling addrlib would be a lot more
 	 * complicated.
@@ -1197,8 +1261,6 @@ static int gfx9_compute_surface(ADDR_HANDLE addrlib,
 	ADDR2_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn = {0};
 	int r;
 
-	assert(!(surf->flags & RADEON_SURF_FMASK));
-
 	AddrSurfInfoIn.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT);
 
 	compressed = surf->blk_w == 4 && surf->blk_h == 4;
@@ -1422,7 +1484,7 @@ int ac_compute_surface(ADDR_HANDLE addrlib, const struct radeon_info *info,
 {
 	int r;
 
-	r = surf_config_sanity(config);
+	r = surf_config_sanity(config, surf->flags);
 	if (r)
 		return r;
 
diff --git a/src/amd/common/ac_surface.h b/src/amd/common/ac_surface.h
index 37df859e6de..d0249684ad2 100644
--- a/src/amd/common/ac_surface.h
+++ b/src/amd/common/ac_surface.h
@@ -79,6 +79,16 @@ struct legacy_surf_level {
     enum radeon_surf_mode       mode:2;
 };
 
+struct legacy_surf_fmask {
+    uint64_t size;
+    unsigned alignment;
+    unsigned tile_swizzle;
+    unsigned slice_tile_max; /* max 4M */
+    uint8_t tiling_index;    /* max 31 */
+    uint8_t bankh;           /* max 8 */
+    uint16_t pitch_in_pixels;
+};
+
 struct legacy_surf_layout {
     unsigned                    bankw:4;  /* max 8 */
     unsigned                    bankh:4;  /* max 8 */
@@ -101,6 +111,7 @@ struct legacy_surf_layout {
     struct legacy_surf_level    stencil_level[RADEON_SURF_MAX_LEVELS];
     uint8_t                     tiling_index[RADEON_SURF_MAX_LEVELS];
     uint8_t                     stencil_tiling_index[RADEON_SURF_MAX_LEVELS];
+    struct legacy_surf_fmask    fmask;
 };
 
 /* Same as addrlib - AddrResourceType. */
@@ -222,7 +233,7 @@ struct ac_surf_info {
 	uint8_t num_channels; /* heuristic for displayability */
 	uint16_t array_size;
 	uint32_t *surf_index; /* Set a monotonic counter for tile swizzling. */
-	uint32_t *fmask_surf_index; /* GFX9+ */
+	uint32_t *fmask_surf_index;
 };
 
 struct ac_surf_config {
diff --git a/src/amd/vulkan/radv_image.c b/src/amd/vulkan/radv_image.c
index 07618853049..ec99197fde9 100644
--- a/src/amd/vulkan/radv_image.c
+++ b/src/amd/vulkan/radv_image.c
@@ -733,58 +733,20 @@ radv_image_get_fmask_info(struct radv_device *device,
 			  unsigned nr_samples,
 			  struct radv_fmask_info *out)
 {
-	/* FMASK is allocated like an ordinary texture. */
-	struct radeon_surf fmask = {};
-	struct ac_surf_info info = image->info;
-	memset(out, 0, sizeof(*out));
-
 	if (device->physical_device->rad_info.chip_class >= GFX9) {
 		out->alignment = image->surface.u.gfx9.fmask_alignment;
 		out->size = image->surface.u.gfx9.fmask_size;
+		out->tile_swizzle = image->surface.u.gfx9.fmask_tile_swizzle;
 		return;
 	}
 
-	fmask.blk_w = image->surface.blk_w;
-	fmask.blk_h = image->surface.blk_h;
-	info.samples = 1;
-	fmask.flags = image->surface.flags | RADEON_SURF_FMASK;
-
-	if (!image->shareable) {
-		info.fmask_surf_index = &device->fmask_mrt_offset_counter;
-		info.surf_index = &device->fmask_mrt_offset_counter;
-	}
-
-	/* Force 2D tiling if it wasn't set. This may occur when creating
-	 * FMASK for MSAA resolve on R6xx. On R6xx, the single-sample
-	 * destination buffer must have an FMASK too. */
-	fmask.flags = RADEON_SURF_CLR(fmask.flags, MODE);
-	fmask.flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE);
-
-	switch (nr_samples) {
-	case 2:
-	case 4:
-		fmask.bpe = 1;
-		break;
-	case 8:
-		fmask.bpe = 4;
-		break;
-	default:
-		return;
-	}
-
-	device->ws->surface_init(device->ws, &info, &fmask);
-	assert(fmask.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
-
-	out->slice_tile_max = (fmask.u.legacy.level[0].nblk_x * fmask.u.legacy.level[0].nblk_y) / 64;
-	if (out->slice_tile_max)
-		out->slice_tile_max -= 1;
-
-	out->tile_mode_index = fmask.u.legacy.tiling_index[0];
-	out->pitch_in_pixels = fmask.u.legacy.level[0].nblk_x;
-	out->bank_height = fmask.u.legacy.bankh;
-	out->tile_swizzle = fmask.tile_swizzle;
-	out->alignment = MAX2(256, fmask.surf_alignment);
-	out->size = fmask.surf_size;
+	out->slice_tile_max = image->surface.u.legacy.fmask.slice_tile_max;
+	out->tile_mode_index = image->surface.u.legacy.fmask.tiling_index;
+	out->pitch_in_pixels = image->surface.u.legacy.fmask.pitch_in_pixels;
+	out->bank_height = image->surface.u.legacy.fmask.bankh;
+	out->tile_swizzle = image->surface.u.legacy.fmask.tile_swizzle;
+	out->alignment = image->surface.u.legacy.fmask.alignment;
+	out->size = image->surface.u.legacy.fmask.size;
 
 	assert(!out->tile_swizzle || !image->shareable);
 }
diff --git a/src/gallium/drivers/radeonsi/si_texture.c b/src/gallium/drivers/radeonsi/si_texture.c
index 17f87a3039b..e072fbd0b4d 100644
--- a/src/gallium/drivers/radeonsi/si_texture.c
+++ b/src/gallium/drivers/radeonsi/si_texture.c
@@ -851,13 +851,6 @@ void si_texture_get_fmask_info(struct si_screen *sscreen,
 			       unsigned nr_samples,
 			       struct r600_fmask_info *out)
 {
-	/* FMASK is allocated like an ordinary texture. */
-	struct pipe_resource templ = rtex->buffer.b.b;
-	struct radeon_surf fmask = {};
-	unsigned flags, bpe;
-
-	memset(out, 0, sizeof(*out));
-
 	if (sscreen->info.chip_class >= GFX9) {
 		out->alignment = rtex->surface.u.gfx9.fmask_alignment;
 		out->size = rtex->surface.u.gfx9.fmask_size;
@@ -865,40 +858,13 @@ void si_texture_get_fmask_info(struct si_screen *sscreen,
 		return;
 	}
 
-	templ.nr_samples = 1;
-	flags = rtex->surface.flags | RADEON_SURF_FMASK;
-
-	switch (nr_samples) {
-	case 2:
-	case 4:
-		bpe = 1;
-		break;
-	case 8:
-		bpe = 4;
-		break;
-	default:
-		PRINT_ERR("Invalid sample count for FMASK allocation.\n");
-		return;
-	}
-
-	if (sscreen->ws->surface_init(sscreen->ws, &templ, flags, bpe,
-				      RADEON_SURF_MODE_2D, &fmask)) {
-		PRINT_ERR("Got error in surface_init while allocating FMASK.\n");
-		return;
-	}
-
-	assert(fmask.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
-
-	out->slice_tile_max = (fmask.u.legacy.level[0].nblk_x * fmask.u.legacy.level[0].nblk_y) / 64;
-	if (out->slice_tile_max)
-		out->slice_tile_max -= 1;
-
-	out->tile_mode_index = fmask.u.legacy.tiling_index[0];
-	out->pitch_in_pixels = fmask.u.legacy.level[0].nblk_x;
-	out->bank_height = fmask.u.legacy.bankh;
-	out->tile_swizzle = fmask.tile_swizzle;
-	out->alignment = MAX2(256, fmask.surf_alignment);
-	out->size = fmask.surf_size;
+	out->slice_tile_max = rtex->surface.u.legacy.fmask.slice_tile_max;
+	out->tile_mode_index = rtex->surface.u.legacy.fmask.tiling_index;
+	out->pitch_in_pixels = rtex->surface.u.legacy.fmask.pitch_in_pixels;
+	out->bank_height = rtex->surface.u.legacy.fmask.bankh;
+	out->tile_swizzle = rtex->surface.u.legacy.fmask.tile_swizzle;
+	out->alignment = rtex->surface.u.legacy.fmask.alignment;
+	out->size = rtex->surface.u.legacy.fmask.size;
 }
 
 static void si_texture_allocate_fmask(struct si_screen *sscreen,
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_surface.c b/src/gallium/winsys/radeon/drm/radeon_drm_surface.c
index 77fb7757c20..58114888966 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_surface.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_surface.c
@@ -243,6 +243,54 @@ static int radeon_winsys_surface_init(struct radeon_winsys *rws,
         return r;
 
     surf_drm_to_winsys(ws, surf_ws, &surf_drm);
+
+    /* Compute FMASK. */
+    if (ws->gen == DRV_SI &&
+        tex->nr_samples >= 2 &&
+        !(flags & (RADEON_SURF_Z_OR_SBUFFER | RADEON_SURF_FMASK))) {
+        /* FMASK is allocated like an ordinary texture. */
+        struct pipe_resource templ = *tex;
+        struct radeon_surf fmask = {};
+        unsigned fmask_flags, bpe;
+
+        templ.nr_samples = 1;
+        fmask_flags = flags | RADEON_SURF_FMASK;
+
+        switch (tex->nr_samples) {
+        case 2:
+        case 4:
+            bpe = 1;
+            break;
+        case 8:
+            bpe = 4;
+            break;
+        default:
+            fprintf(stderr, "radeon: Invalid sample count for FMASK allocation.\n");
+            return -1;
+        }
+
+        if (radeon_winsys_surface_init(rws, &templ, fmask_flags, bpe,
+                                       RADEON_SURF_MODE_2D, &fmask)) {
+            fprintf(stderr, "Got error in surface_init while allocating FMASK.\n");
+            return -1;
+        }
+
+        assert(fmask.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
+
+        surf_ws->u.legacy.fmask.size = fmask.surf_size;
+        surf_ws->u.legacy.fmask.alignment = MAX2(256, fmask.surf_alignment);
+        surf_ws->u.legacy.fmask.tile_swizzle = fmask.tile_swizzle;
+
+        surf_ws->u.legacy.fmask.slice_tile_max =
+            (fmask.u.legacy.level[0].nblk_x * fmask.u.legacy.level[0].nblk_y) / 64;
+        if (surf_ws->u.legacy.fmask.slice_tile_max)
+            surf_ws->u.legacy.fmask.slice_tile_max -= 1;
+
+        surf_ws->u.legacy.fmask.tiling_index = fmask.u.legacy.tiling_index[0];
+        surf_ws->u.legacy.fmask.bankh = fmask.u.legacy.bankh;
+        surf_ws->u.legacy.fmask.pitch_in_pixels = fmask.u.legacy.level[0].nblk_x;
+    }
+
     return 0;
 }