broadcom/vc5: Update state setup for V3D 4.1.

7 files changed, 206 insertions, 14 deletions

diff --git a/src/gallium/drivers/vc5/Makefile.sources b/src/gallium/drivers/vc5/Makefile.sources
index 45c4ccd67f6..252431b54b5 100644
--- a/src/gallium/drivers/vc5/Makefile.sources
+++ b/src/gallium/drivers/vc5/Makefile.sources
@@ -18,7 +18,6 @@ C_SOURCES := \
 	vc5_screen.c \
 	vc5_screen.h \
 	vc5_simulator.c \
-	vc5_state.c \
 	vc5_tiling.c \
 	vc5_tiling.h \
 	vc5_uniforms.c \
@@ -29,4 +28,5 @@ VC5_PER_VERSION_SOURCES = \
 	v3dx_simulator.c \
 	vc5_draw.c \
 	vc5_rcl.c \
+	vc5_state.c \
 	$()
diff --git a/src/gallium/drivers/vc5/meson.build b/src/gallium/drivers/vc5/meson.build
index e0dd45802a5..aa74e624bdb 100644
--- a/src/gallium/drivers/vc5/meson.build
+++ b/src/gallium/drivers/vc5/meson.build
@@ -38,7 +38,6 @@ files_libvc5 = files(
   'vc5_screen.h',
   'vc5_simulator.c',
   'vc5_simulator_wrapper.cpp',
-  'vc5_state.c',
   'vc5_tiling.c',
   'vc5_tiling.h',
   'vc5_uniforms.c',
@@ -49,6 +48,7 @@ files_per_version = files(
   'v3dx_simulator.c',
   'vc5_draw.c',
   'vc5_rcl.c',
+  'vc5_state.c',
 )
 
 v3dv3_c_args = []
diff --git a/src/gallium/drivers/vc5/v3dx_context.h b/src/gallium/drivers/vc5/v3dx_context.h
index 445038c76b3..f92712dcf79 100644
--- a/src/gallium/drivers/vc5/v3dx_context.h
+++ b/src/gallium/drivers/vc5/v3dx_context.h
@@ -31,6 +31,7 @@ struct vc5_format;
 
 void v3dX(emit_rcl)(struct vc5_job *job);
 void v3dX(draw_init)(struct pipe_context *pctx);
+void v3dX(state_init)(struct pipe_context *pctx);
 
 void v3dX(simulator_init_regs)(struct v3d_hw *v3d);
 int v3dX(simulator_get_param_ioctl)(struct v3d_hw *v3d,
diff --git a/src/gallium/drivers/vc5/vc5_context.c b/src/gallium/drivers/vc5/vc5_context.c
index f6ae0ad989b..9403f8ffdd4 100644
--- a/src/gallium/drivers/vc5/vc5_context.c
+++ b/src/gallium/drivers/vc5/vc5_context.c
@@ -134,11 +134,13 @@ vc5_context_create(struct pipe_screen *pscreen, void *priv, unsigned flags)
         pctx->flush = vc5_pipe_flush;
         pctx->invalidate_resource = vc5_invalidate_resource;
 
-        if (screen->devinfo.ver >= 41)
+        if (screen->devinfo.ver >= 41) {
                 v3d41_draw_init(pctx);
-        else
+                v3d41_state_init(pctx);
+        } else {
                 v3d33_draw_init(pctx);
-        vc5_state_init(pctx);
+                v3d33_state_init(pctx);
+        }
         vc5_program_init(pctx);
         vc5_query_init(pctx);
         vc5_resource_context_init(pctx);
diff --git a/src/gallium/drivers/vc5/vc5_context.h b/src/gallium/drivers/vc5/vc5_context.h
index 508c894071a..e4c2eb551bd 100644
--- a/src/gallium/drivers/vc5/vc5_context.h
+++ b/src/gallium/drivers/vc5/vc5_context.h
@@ -93,6 +93,8 @@ struct vc5_sampler_view {
         uint8_t swizzle[4];
 
         uint8_t texture_shader_state[32];
+        /* V3D 4.x: Texture state struct. */
+        struct vc5_bo *bo;
 };
 
 struct vc5_sampler_state {
@@ -100,7 +102,10 @@ struct vc5_sampler_state {
         uint32_t p0;
         uint32_t p1;
 
+        /* V3D 3.x: Packed texture state. */
         uint8_t texture_shader_state[32];
+        /* V3D 4.x: Sampler state struct. */
+        struct vc5_bo *bo;
 };
 
 struct vc5_texture_stateobj {
@@ -446,7 +451,6 @@ vc5_sampler_state(struct pipe_sampler_state *psampler)
 
 struct pipe_context *vc5_context_create(struct pipe_screen *pscreen,
                                         void *priv, unsigned flags);
-void vc5_state_init(struct pipe_context *pctx);
 void vc5_program_init(struct pipe_context *pctx);
 void vc5_program_fini(struct pipe_context *pctx);
 void vc5_query_init(struct pipe_context *pctx);
diff --git a/src/gallium/drivers/vc5/vc5_state.c b/src/gallium/drivers/vc5/vc5_state.c
index 79795c765aa..81ae33d7178 100644
--- a/src/gallium/drivers/vc5/vc5_state.c
+++ b/src/gallium/drivers/vc5/vc5_state.c
@@ -31,7 +31,6 @@
 #include "util/u_helpers.h"
 
 #include "vc5_context.h"
-#define V3D_VERSION 33
 #include "broadcom/common/v3d_macros.h"
 #include "broadcom/cle/v3dx_pack.h"
 
@@ -503,6 +502,7 @@ static void *
 vc5_create_sampler_state(struct pipe_context *pctx,
                          const struct pipe_sampler_state *cso)
 {
+        MAYBE_UNUSED struct vc5_context *vc5 = vc5_context(pctx);
         struct vc5_sampler_state *so = CALLOC_STRUCT(vc5_sampler_state);
 
         if (!so)
@@ -514,6 +514,62 @@ vc5_create_sampler_state(struct pipe_context *pctx,
                 (cso->mag_img_filter == PIPE_TEX_MIPFILTER_NEAREST ||
                  cso->min_img_filter == PIPE_TEX_MIPFILTER_NEAREST);
 
+#if V3D_VERSION >= 40
+        so->bo = vc5_bo_alloc(vc5->screen, cl_packet_length(SAMPLER_STATE),
+                              "sampler");
+        void *map = vc5_bo_map(so->bo);
+
+        v3dx_pack(map, SAMPLER_STATE, sampler) {
+                sampler.wrap_i_border = false;
+
+                sampler.wrap_s = translate_wrap(cso->wrap_s, either_nearest);
+                sampler.wrap_t = translate_wrap(cso->wrap_t, either_nearest);
+                sampler.wrap_r = translate_wrap(cso->wrap_r, either_nearest);
+
+                sampler.fixed_bias = cso->lod_bias;
+                sampler.depth_compare_function = cso->compare_func;
+
+                sampler.min_filter_nearest =
+                        cso->min_img_filter == PIPE_TEX_FILTER_NEAREST;
+                sampler.mag_filter_nearest =
+                        cso->mag_img_filter == PIPE_TEX_FILTER_NEAREST;
+                sampler.mip_filter_nearest =
+                        cso->min_mip_filter == PIPE_TEX_MIPFILTER_NEAREST;
+
+                sampler.min_level_of_detail = MIN2(MAX2(0, cso->min_lod),
+                                                   15);
+                sampler.max_level_of_detail = MIN2(cso->max_lod, 15);
+
+                if (cso->max_anisotropy) {
+                        sampler.anisotropy_enable = true;
+
+                        if (cso->max_anisotropy > 8)
+                                sampler.maximum_anisotropy = 3;
+                        else if (cso->max_anisotropy > 4)
+                                sampler.maximum_anisotropy = 2;
+                        else if (cso->max_anisotropy > 2)
+                                sampler.maximum_anisotropy = 1;
+                }
+
+                sampler.border_colour_mode = V3D_BORDER_COLOUR_FOLLOWS;
+                /* XXX: The border colour field is in the TMU blending format
+                 * (32, f16, or i16), and we need to customize it based on
+                 * that.
+                 *
+                 * XXX: for compat alpha formats, we need the alpha field to
+                 * be in the red channel.
+                 */
+                sampler.border_colour_red =
+                        util_float_to_half(cso->border_color.f[0]);
+                sampler.border_colour_green =
+                        util_float_to_half(cso->border_color.f[1]);
+                sampler.border_colour_blue =
+                        util_float_to_half(cso->border_color.f[2]);
+                sampler.border_colour_alpha =
+                        util_float_to_half(cso->border_color.f[3]);
+        }
+
+#else /* V3D_VERSION < 40 */
         v3dx_pack(&so->p0, TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1, p0) {
                 p0.s_wrap_mode = translate_wrap(cso->wrap_s, either_nearest);
                 p0.t_wrap_mode = translate_wrap(cso->wrap_t, either_nearest);
@@ -524,7 +580,7 @@ vc5_create_sampler_state(struct pipe_context *pctx,
                 tex.depth_compare_function = cso->compare_func;
                 tex.fixed_bias = cso->lod_bias;
         }
-
+#endif /* V3D_VERSION < 40 */
         return so;
 }
 
@@ -553,6 +609,37 @@ vc5_sampler_states_bind(struct pipe_context *pctx,
         stage_tex->num_samplers = new_nr;
 }
 
+static void
+vc5_sampler_state_delete(struct pipe_context *pctx,
+                         void *hwcso)
+{
+        struct pipe_sampler_state *psampler = hwcso;
+        struct vc5_sampler_state *sampler = vc5_sampler_state(psampler);
+
+        vc5_bo_unreference(&sampler->bo);
+        free(psampler);
+}
+
+#if V3D_VERSION >= 40
+static uint32_t
+translate_swizzle(unsigned char pipe_swizzle)
+{
+        switch (pipe_swizzle) {
+        case PIPE_SWIZZLE_0:
+                return 0;
+        case PIPE_SWIZZLE_1:
+                return 1;
+        case PIPE_SWIZZLE_X:
+        case PIPE_SWIZZLE_Y:
+        case PIPE_SWIZZLE_Z:
+        case PIPE_SWIZZLE_W:
+                return 2 + pipe_swizzle;
+        default:
+                unreachable("unknown swizzle");
+        }
+}
+#endif
+
 static struct pipe_sampler_view *
 vc5_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
                         const struct pipe_sampler_view *cso)
@@ -589,9 +676,29 @@ vc5_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
 
         int msaa_scale = prsc->nr_samples > 1 ? 2 : 1;
 
+#if V3D_VERSION >= 40
+        so->bo = vc5_bo_alloc(vc5->screen, cl_packet_length(SAMPLER_STATE),
+                              "sampler");
+        void *map = vc5_bo_map(so->bo);
+
+        v3dx_pack(map, TEXTURE_SHADER_STATE, tex) {
+#else /* V3D_VERSION < 40 */
         v3dx_pack(&so->texture_shader_state, TEXTURE_SHADER_STATE, tex) {
+#endif
+
                 tex.image_width = prsc->width0 * msaa_scale;
                 tex.image_height = prsc->height0 * msaa_scale;
+
+#if V3D_VERSION >= 40
+                /* On 4.x, the height of a 1D texture is redefined to be the
+                 * upper 14 bits of the width (which is only usable with txf).
+                 */
+                if (prsc->target == PIPE_TEXTURE_1D ||
+                    prsc->target == PIPE_TEXTURE_1D_ARRAY) {
+                        tex.image_height = tex.image_width >> 14;
+                }
+#endif
+
                 if (prsc->target == PIPE_TEXTURE_3D) {
                         tex.image_depth = prsc->depth0;
                 } else {
@@ -602,6 +709,21 @@ vc5_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
                 tex.srgb = util_format_is_srgb(cso->format);
 
                 tex.base_level = cso->u.tex.first_level;
+#if V3D_VERSION >= 40
+                tex.max_level = cso->u.tex.last_level;
+                /* Note that we don't have a job to reference the texture's sBO
+                 * at state create time, so any time this sampler view is used
+                 * we need to add the texture to the job.
+                 */
+                tex.texture_base_pointer = cl_address(NULL,
+                                                      rsc->bo->offset +
+                                                      rsc->slices[0].offset),
+
+                tex.swizzle_r = translate_swizzle(so->swizzle[0]);
+                tex.swizzle_g = translate_swizzle(so->swizzle[1]);
+                tex.swizzle_b = translate_swizzle(so->swizzle[2]);
+                tex.swizzle_a = translate_swizzle(so->swizzle[3]);
+#endif
                 tex.array_stride_64_byte_aligned = rsc->cube_map_stride / 64;
 
                 if (prsc->nr_samples > 1) {
@@ -661,6 +783,13 @@ vc5_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
                                                VC5_TILING_UIF_NO_XOR);
                 tex.level_0_xor_enable = (rsc->slices[0].tiling ==
                                           VC5_TILING_UIF_XOR);
+
+#if V3D_VERSION >= 40
+                if (tex.uif_xor_disable ||
+                    tex.level_0_is_strictly_uif) {
+                        tex.extended = true;
+                }
+#endif /* V3D_VERSION >= 40 */
         };
 
         return &so->base;
@@ -668,10 +797,13 @@ vc5_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
 
 static void
 vc5_sampler_view_destroy(struct pipe_context *pctx,
-                         struct pipe_sampler_view *view)
+                         struct pipe_sampler_view *psview)
 {
-        pipe_resource_reference(&view->texture, NULL);
-        free(view);
+        struct vc5_sampler_view *sview = vc5_sampler_view(psview);
+
+        vc5_bo_unreference(&sview->bo);
+        pipe_resource_reference(&psview->texture, NULL);
+        free(psview);
 }
 
 static void
@@ -754,7 +886,7 @@ vc5_set_stream_output_targets(struct pipe_context *pctx,
 }
 
 void
-vc5_state_init(struct pipe_context *pctx)
+v3dX(state_init)(struct pipe_context *pctx)
 {
         pctx->set_blend_color = vc5_set_blend_color;
         pctx->set_stencil_ref = vc5_set_stencil_ref;
@@ -785,7 +917,7 @@ vc5_state_init(struct pipe_context *pctx)
         pctx->bind_vertex_elements_state = vc5_vertex_state_bind;
 
         pctx->create_sampler_state = vc5_create_sampler_state;
-        pctx->delete_sampler_state = vc5_generic_cso_state_delete;
+        pctx->delete_sampler_state = vc5_sampler_state_delete;
         pctx->bind_sampler_states = vc5_sampler_states_bind;
 
         pctx->create_sampler_view = vc5_create_sampler_view;
diff --git a/src/gallium/drivers/vc5/vc5_uniforms.c b/src/gallium/drivers/vc5/vc5_uniforms.c
index 676ab1a32c9..faf49dbc359 100644
--- a/src/gallium/drivers/vc5/vc5_uniforms.c
+++ b/src/gallium/drivers/vc5/vc5_uniforms.c
@@ -170,7 +170,7 @@ vc5_upload_ubo(struct vc5_context *vc5,
 }
 
 /**
- *  Writes the P0 (CFG_MODE=1) texture parameter.
+ *  Writes the V3D 3.x P0 (CFG_MODE=1) texture parameter.
  *
  * Some bits of this field are dependent on the type of sample being done by
  * the shader, while other bits are dependent on the sampler state.  We OR the
@@ -189,6 +189,7 @@ write_texture_p0(struct vc5_job *job,
         cl_aligned_u32(uniforms, shader_data | sampler->p0);
 }
 
+/** Writes the V3D 3.x P1 (CFG_MODE=1) texture parameter. */
 static void
 write_texture_p1(struct vc5_job *job,
                  struct vc5_cl_out **uniforms,
@@ -216,6 +217,46 @@ write_texture_p1(struct vc5_job *job,
         cl_aligned_u32(uniforms, p1 | packed | sview->p1);
 }
 
+/** Writes the V3D 4.x TMU configuration parameter 0. */
+static void
+write_tmu_p0(struct vc5_job *job,
+             struct vc5_cl_out **uniforms,
+             struct vc5_texture_stateobj *texstate,
+             uint32_t data)
+{
+        /* Extract the texture unit from the top bits, and the compiler's
+         * packed p0 from the bottom.
+         */
+        uint32_t unit = data >> 24;
+        uint32_t p0 = data & 0x00ffffff;
+
+        struct pipe_sampler_view *psview = texstate->textures[unit];
+        struct vc5_sampler_view *sview = vc5_sampler_view(psview);
+        struct vc5_resource *rsc = vc5_resource(psview->texture);
+
+        cl_aligned_reloc(&job->indirect, uniforms, sview->bo, p0);
+        vc5_job_add_bo(job, rsc->bo);
+}
+
+/** Writes the V3D 4.x TMU configuration parameter 1. */
+static void
+write_tmu_p1(struct vc5_job *job,
+             struct vc5_cl_out **uniforms,
+             struct vc5_texture_stateobj *texstate,
+             uint32_t data)
+{
+        /* Extract the texture unit from the top bits, and the compiler's
+         * packed p1 from the bottom.
+         */
+        uint32_t unit = data >> 24;
+        uint32_t p0 = data & 0x00ffffff;
+
+        struct pipe_sampler_state *psampler = texstate->samplers[unit];
+        struct vc5_sampler_state *sampler = vc5_sampler_state(psampler);
+
+        cl_aligned_reloc(&job->indirect, uniforms, sampler->bo, p0);
+}
+
 struct vc5_cl_reloc
 vc5_write_uniforms(struct vc5_context *vc5, struct vc5_compiled_shader *shader,
                    struct vc5_constbuf_stateobj *cb,
@@ -266,6 +307,16 @@ vc5_write_uniforms(struct vc5_context *vc5, struct vc5_compiled_shader *shader,
                                      vc5->clip.ucp[uinfo->data[i] / 4][uinfo->data[i] % 4]);
                         break;
 
+                case QUNIFORM_TMU_CONFIG_P0:
+                        write_tmu_p0(job, &uniforms, texstate,
+                                         uinfo->data[i]);
+                        break;
+
+                case QUNIFORM_TMU_CONFIG_P1:
+                        write_tmu_p1(job, &uniforms, texstate,
+                                         uinfo->data[i]);
+                        break;
+
                 case QUNIFORM_TEXTURE_CONFIG_P1:
                         write_texture_p1(job, &uniforms, texstate,
                                          uinfo->data[i]);
@@ -388,6 +439,8 @@ vc5_set_shader_uniform_dirty_flags(struct vc5_compiled_shader *shader)
                         dirty |= VC5_DIRTY_CLIP;
                         break;
 
+                case QUNIFORM_TMU_CONFIG_P0:
+                case QUNIFORM_TMU_CONFIG_P1:
                 case QUNIFORM_TEXTURE_CONFIG_P1:
                 case QUNIFORM_TEXTURE_BORDER_COLOR:
                 case QUNIFORM_TEXTURE_FIRST_LEVEL: