nvc0: add hardware ETC2 and ASTC support on GK20A and GM107+

Signed-off-by: Ilia Mirkin <[email protected]>

3 files changed, 64 insertions, 2 deletions

diff --git a/src/gallium/drivers/nouveau/nv50/nv50_formats.c b/src/gallium/drivers/nouveau/nv50/nv50_formats.c
index 717067cf2f7..5b965672357 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_formats.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_formats.c
@@ -23,6 +23,7 @@
 #if NOUVEAU_DRIVER == 0xc0
 # include "nvc0/nvc0_screen.h"
 # include "nvc0/nvc0_3d.xml.h"
+# include "nvc0/gm107_texture.xml.h"
 #else
 # include "nv50/nv50_screen.h"
 # include "nv50/nv50_3d.xml.h"
@@ -65,6 +66,7 @@
 #define SF_A(sz) G80_TIC_0_COMPONENTS_SIZES_##sz
 #define SF_B(sz) G200_TIC_0_COMPONENTS_SIZES_##sz
 #define SF_C(sz) GF100_TIC_0_COMPONENTS_SIZES_##sz
+#define SF_D(sz) GM107_TIC2_0_COMPONENTS_SIZES_##sz
 #define SF(c, pf, sf, r, g, b, a, t0, t1, t2, t3, sz, u)                \
    [PIPE_FORMAT_##pf] = {                                               \
       sf, {                                                             \
@@ -236,6 +238,50 @@ const struct nv50_format nv50_format_table[PIPE_FORMAT_COUNT] =
    F3(C, BPTC_RGB_FLOAT,  NONE, R, G, B, xx, FLOAT, BC6H_SF16, t),
    F3(C, BPTC_RGB_UFLOAT, NONE, R, G, B, xx, FLOAT, BC6H_UF16, t),
 
+#if NOUVEAU_DRIVER == 0xc0
+   F3(D, ETC1_RGB8,       NONE, R,  G,  B, xx, UNORM, ETC2_RGB,     t),
+   F3(D, ETC2_RGB8,       NONE, R,  G,  B, xx, UNORM, ETC2_RGB,     t),
+   F3(D, ETC2_SRGB8,      NONE, R,  G,  B, xx, UNORM, ETC2_RGB,     t),
+   C4(D, ETC2_RGB8A1,     NONE, R,  G,  B,  A, UNORM, ETC2_RGB_PTA, t),
+   C4(D, ETC2_SRGB8A1,    NONE, R,  G,  B,  A, UNORM, ETC2_RGB_PTA, t),
+   C4(D, ETC2_RGBA8,      NONE, R,  G,  B,  A, UNORM, ETC2_RGBA,    t),
+   C4(D, ETC2_SRGBA8,     NONE, R,  G,  B,  A, UNORM, ETC2_RGBA,    t),
+   F1(D, ETC2_R11_UNORM,  NONE, R, xx, xx, xx, UNORM, EAC,          t),
+   F1(D, ETC2_R11_SNORM,  NONE, R, xx, xx, xx, SNORM, EAC,          t),
+   F2(D, ETC2_RG11_UNORM, NONE, R,  G, xx, xx, UNORM, EACX2,        t),
+   F2(D, ETC2_RG11_SNORM, NONE, R,  G, xx, xx, SNORM, EACX2,        t),
+
+   C4(D, ASTC_4x4,        NONE, R, G, B, A, UNORM, ASTC_2D_4X4,   t),
+   C4(D, ASTC_5x4,        NONE, R, G, B, A, UNORM, ASTC_2D_5X4,   t),
+   C4(D, ASTC_5x5,        NONE, R, G, B, A, UNORM, ASTC_2D_5X5,   t),
+   C4(D, ASTC_6x5,        NONE, R, G, B, A, UNORM, ASTC_2D_6X5,   t),
+   C4(D, ASTC_6x6,        NONE, R, G, B, A, UNORM, ASTC_2D_6X6,   t),
+   C4(D, ASTC_8x5,        NONE, R, G, B, A, UNORM, ASTC_2D_8X5,   t),
+   C4(D, ASTC_8x6,        NONE, R, G, B, A, UNORM, ASTC_2D_8X6,   t),
+   C4(D, ASTC_8x8,        NONE, R, G, B, A, UNORM, ASTC_2D_8X8,   t),
+   C4(D, ASTC_10x5,       NONE, R, G, B, A, UNORM, ASTC_2D_10X5,  t),
+   C4(D, ASTC_10x6,       NONE, R, G, B, A, UNORM, ASTC_2D_10X6,  t),
+   C4(D, ASTC_10x8,       NONE, R, G, B, A, UNORM, ASTC_2D_10X8,  t),
+   C4(D, ASTC_10x10,      NONE, R, G, B, A, UNORM, ASTC_2D_10X10, t),
+   C4(D, ASTC_12x10,      NONE, R, G, B, A, UNORM, ASTC_2D_12X10, t),
+   C4(D, ASTC_12x12,      NONE, R, G, B, A, UNORM, ASTC_2D_12X12, t),
+
+   C4(D, ASTC_4x4_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_4X4,   t),
+   C4(D, ASTC_5x4_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_5X4,   t),
+   C4(D, ASTC_5x5_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_5X5,   t),
+   C4(D, ASTC_6x5_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_6X5,   t),
+   C4(D, ASTC_6x6_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_6X6,   t),
+   C4(D, ASTC_8x5_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_8X5,   t),
+   C4(D, ASTC_8x6_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_8X6,   t),
+   C4(D, ASTC_8x8_SRGB,   NONE, R, G, B, A, UNORM, ASTC_2D_8X8,   t),
+   C4(D, ASTC_10x5_SRGB,  NONE, R, G, B, A, UNORM, ASTC_2D_10X5,  t),
+   C4(D, ASTC_10x6_SRGB,  NONE, R, G, B, A, UNORM, ASTC_2D_10X6,  t),
+   C4(D, ASTC_10x8_SRGB,  NONE, R, G, B, A, UNORM, ASTC_2D_10X8,  t),
+   C4(D, ASTC_10x10_SRGB, NONE, R, G, B, A, UNORM, ASTC_2D_10X10, t),
+   C4(D, ASTC_12x10_SRGB, NONE, R, G, B, A, UNORM, ASTC_2D_12X10, t),
+   C4(D, ASTC_12x12_SRGB, NONE, R, G, B, A, UNORM, ASTC_2D_12X12, t),
+#endif
+
    C4(A, R32G32B32A32_FLOAT, RGBA32_FLOAT, R, G, B, A, FLOAT, R32_G32_B32_A32, IB),
    C4(A, R32G32B32A32_UNORM, NONE, R, G, B, A, UNORM, R32_G32_B32_A32, T),
    C4(A, R32G32B32A32_SNORM, NONE, R, G, B, A, SNORM, R32_G32_B32_A32, T),
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
index 14438ced7a3..ec2340ee0c3 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
@@ -45,6 +45,8 @@ nvc0_screen_is_format_supported(struct pipe_screen *pscreen,
                                 unsigned sample_count,
                                 unsigned bindings)
 {
+   const struct util_format_description *desc = util_format_description(format);
+
    if (sample_count > 8)
       return false;
    if (!(0x117 & (1 << sample_count))) /* 0, 1, 2, 4 or 8 */
@@ -65,6 +67,17 @@ nvc0_screen_is_format_supported(struct pipe_screen *pscreen,
           sample_count > 1)
          return false;
 
+   /* Restrict ETC2 and ASTC formats here. These are only supported on GK20A.
+    */
+   if ((desc->layout == UTIL_FORMAT_LAYOUT_ETC ||
+        desc->layout == UTIL_FORMAT_LAYOUT_ASTC) &&
+       /* The claim is that this should work on GM107 but it doesn't. Need to
+        * test further and figure out if it's a nouveau issue or a HW one.
+       nouveau_screen(pscreen)->class_3d < GM107_3D_CLASS &&
+        */
+       nouveau_screen(pscreen)->class_3d != NVEA_3D_CLASS)
+      return false;
+
    /* transfers & shared are always supported */
    bindings &= ~(PIPE_BIND_TRANSFER_READ |
                  PIPE_BIND_TRANSFER_WRITE |
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_tex.c b/src/gallium/drivers/nouveau/nvc0/nvc0_tex.c
index ce6a6dce39c..e04573f8249 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_tex.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_tex.c
@@ -250,6 +250,7 @@ gf100_create_texture_view(struct pipe_context *pipe,
    uint32_t swz[4];
    uint32_t width, height;
    uint32_t depth;
+   uint32_t tex_fmt;
    struct nv50_tic_entry *view;
    struct nv50_miptree *mt;
    bool tex_int;
@@ -275,12 +276,13 @@ gf100_create_texture_view(struct pipe_context *pipe,
    fmt = &nvc0_format_table[view->pipe.format];
 
    tex_int = util_format_is_pure_integer(view->pipe.format);
+   tex_fmt = fmt->tic.format & 0x3f;
 
    swz[0] = nv50_tic_swizzle(fmt, view->pipe.swizzle_r, tex_int);
    swz[1] = nv50_tic_swizzle(fmt, view->pipe.swizzle_g, tex_int);
    swz[2] = nv50_tic_swizzle(fmt, view->pipe.swizzle_b, tex_int);
    swz[3] = nv50_tic_swizzle(fmt, view->pipe.swizzle_a, tex_int);
-   tic[0] = (fmt->tic.format << G80_TIC_0_COMPONENTS_SIZES__SHIFT) |
+   tic[0] = (tex_fmt << G80_TIC_0_COMPONENTS_SIZES__SHIFT) |
             (fmt->tic.type_r << G80_TIC_0_R_DATA_TYPE__SHIFT) |
             (fmt->tic.type_g << G80_TIC_0_G_DATA_TYPE__SHIFT) |
             (fmt->tic.type_b << G80_TIC_0_B_DATA_TYPE__SHIFT) |
@@ -288,7 +290,8 @@ gf100_create_texture_view(struct pipe_context *pipe,
             (swz[0] << G80_TIC_0_X_SOURCE__SHIFT) |
             (swz[1] << G80_TIC_0_Y_SOURCE__SHIFT) |
             (swz[2] << G80_TIC_0_Z_SOURCE__SHIFT) |
-            (swz[3] << G80_TIC_0_W_SOURCE__SHIFT);
+            (swz[3] << G80_TIC_0_W_SOURCE__SHIFT) |
+            ((fmt->tic.format & 0x40) << (GK20A_TIC_0_USE_COMPONENT_SIZES_EXTENDED__SHIFT - 6));
 
    address = mt->base.address;