ss[0] |= SET_FIELD(lod_bias & 0x1fff, GEN7_SAMPLER_LOD_BIAS);
if (min_filter == BRW_MAPFILTER_ANISOTROPIC)
- ss[0] |= GEN7_SAMPLER_EWA_ANISOTROPIC_ALGORIHTM;
+ ss[0] |= GEN7_SAMPLER_EWA_ANISOTROPIC_ALGORITHM;
ss[1] = SET_FIELD(min_lod, GEN7_SAMPLER_MIN_LOD) |
SET_FIELD(max_lod, GEN7_SAMPLER_MAX_LOD) |
}
}
+/**
+ * Return true if the given wrap mode requires the border color to exist.
+ */
+static bool
+wrap_mode_needs_border_color(unsigned wrap_mode)
+{
+ return wrap_mode == BRW_TEXCOORDMODE_CLAMP_BORDER ||
+ wrap_mode == GEN8_TEXCOORDMODE_HALF_BORDER;
+}
+
/**
* Upload SAMPLER_BORDER_COLOR_STATE.
*/
static void
upload_default_color(struct brw_context *brw,
const struct gl_sampler_object *sampler,
- int unit,
+ mesa_format format, GLenum base_format,
+ bool is_integer_format,
uint32_t *sdc_offset)
{
- struct gl_context *ctx = &brw->ctx;
- struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
- struct gl_texture_object *texObj = texUnit->_Current;
- struct gl_texture_image *firstImage = texObj->Image[0][texObj->BaseLevel];
- float color[4];
+ union gl_color_union color;
- switch (firstImage->_BaseFormat) {
+ switch (base_format) {
case GL_DEPTH_COMPONENT:
/* GL specs that border color for depth textures is taken from the
* R channel, while the hardware uses A. Spam R into all the
* channels for safety.
*/
- color[0] = sampler->BorderColor.f[0];
- color[1] = sampler->BorderColor.f[0];
- color[2] = sampler->BorderColor.f[0];
- color[3] = sampler->BorderColor.f[0];
+ color.ui[0] = sampler->BorderColor.ui[0];
+ color.ui[1] = sampler->BorderColor.ui[0];
+ color.ui[2] = sampler->BorderColor.ui[0];
+ color.ui[3] = sampler->BorderColor.ui[0];
break;
case GL_ALPHA:
- color[0] = 0.0;
- color[1] = 0.0;
- color[2] = 0.0;
- color[3] = sampler->BorderColor.f[3];
+ color.ui[0] = 0u;
+ color.ui[1] = 0u;
+ color.ui[2] = 0u;
+ color.ui[3] = sampler->BorderColor.ui[3];
break;
case GL_INTENSITY:
- color[0] = sampler->BorderColor.f[0];
- color[1] = sampler->BorderColor.f[0];
- color[2] = sampler->BorderColor.f[0];
- color[3] = sampler->BorderColor.f[0];
+ color.ui[0] = sampler->BorderColor.ui[0];
+ color.ui[1] = sampler->BorderColor.ui[0];
+ color.ui[2] = sampler->BorderColor.ui[0];
+ color.ui[3] = sampler->BorderColor.ui[0];
break;
case GL_LUMINANCE:
- color[0] = sampler->BorderColor.f[0];
- color[1] = sampler->BorderColor.f[0];
- color[2] = sampler->BorderColor.f[0];
- color[3] = 1.0;
+ color.ui[0] = sampler->BorderColor.ui[0];
+ color.ui[1] = sampler->BorderColor.ui[0];
+ color.ui[2] = sampler->BorderColor.ui[0];
+ color.ui[3] = float_as_int(1.0);
break;
case GL_LUMINANCE_ALPHA:
- color[0] = sampler->BorderColor.f[0];
- color[1] = sampler->BorderColor.f[0];
- color[2] = sampler->BorderColor.f[0];
- color[3] = sampler->BorderColor.f[3];
+ color.ui[0] = sampler->BorderColor.ui[0];
+ color.ui[1] = sampler->BorderColor.ui[0];
+ color.ui[2] = sampler->BorderColor.ui[0];
+ color.ui[3] = sampler->BorderColor.ui[3];
break;
default:
- color[0] = sampler->BorderColor.f[0];
- color[1] = sampler->BorderColor.f[1];
- color[2] = sampler->BorderColor.f[2];
- color[3] = sampler->BorderColor.f[3];
+ color.ui[0] = sampler->BorderColor.ui[0];
+ color.ui[1] = sampler->BorderColor.ui[1];
+ color.ui[2] = sampler->BorderColor.ui[2];
+ color.ui[3] = sampler->BorderColor.ui[3];
break;
}
* where we've initialized the A channel to 1.0. We also have to set
* the border color alpha to 1.0 in that case.
*/
- if (firstImage->_BaseFormat == GL_RGB)
- color[3] = 1.0;
+ if (base_format == GL_RGB)
+ color.ui[3] = float_as_int(1.0);
if (brw->gen >= 8) {
/* On Broadwell, the border color is represented as four 32-bit floats,
* integers, or unsigned values, interpreted according to the surface
- * format. This matches the sampler->BorderColor union exactly. Since
- * we use floats both here and in the above reswizzling code, we preserve
- * the original bit pattern. So we actually handle all three formats.
+ * format. This matches the sampler->BorderColor union exactly; just
+ * memcpy the values.
*/
- float *sdc = brw_state_batch(brw, AUB_TRACE_SAMPLER_DEFAULT_COLOR,
- 4 * 4, 64, sdc_offset);
- COPY_4FV(sdc, color);
+ uint32_t *sdc = brw_state_batch(brw, AUB_TRACE_SAMPLER_DEFAULT_COLOR,
+ 4 * 4, 64, sdc_offset);
+ memcpy(sdc, color.ui, 4 * 4);
+ } else if (brw->is_haswell && is_integer_format) {
+ /* Haswell's integer border color support is completely insane:
+ * SAMPLER_BORDER_COLOR_STATE is 20 DWords. The first four are
+ * for float colors. The next 12 DWords are MBZ and only exist to
+ * pad it out to a 64 byte cacheline boundary. DWords 16-19 then
+ * contain integer colors; these are only used if SURFACE_STATE
+ * has the "Integer Surface Format" bit set. Even then, the
+ * arrangement of the RGBA data devolves into madness.
+ */
+ uint32_t *sdc = brw_state_batch(brw, AUB_TRACE_SAMPLER_DEFAULT_COLOR,
+ 20 * 4, 512, sdc_offset);
+ memset(sdc, 0, 20 * 4);
+ sdc = &sdc[16];
+
+ int bits_per_channel = _mesa_get_format_bits(format, GL_RED_BITS);
+
+ /* From the Haswell PRM, "Command Reference: Structures", Page 36:
+ * "If any color channel is missing from the surface format,
+ * corresponding border color should be programmed as zero and if
+ * alpha channel is missing, corresponding Alpha border color should
+ * be programmed as 1."
+ */
+ unsigned c[4] = { 0, 0, 0, 1 };
+ for (int i = 0; i < 4; i++) {
+ if (_mesa_format_has_color_component(format, i))
+ c[i] = color.ui[i];
+ }
+
+ switch (bits_per_channel) {
+ case 8:
+ /* Copy RGBA in order. */
+ for (int i = 0; i < 4; i++)
+ ((uint8_t *) sdc)[i] = c[i];
+ break;
+ case 10:
+ /* R10G10B10A2_UINT is treated like a 16-bit format. */
+ case 16:
+ ((uint16_t *) sdc)[0] = c[0]; /* R -> DWord 0, bits 15:0 */
+ ((uint16_t *) sdc)[1] = c[1]; /* G -> DWord 0, bits 31:16 */
+ /* DWord 1 is Reserved/MBZ! */
+ ((uint16_t *) sdc)[4] = c[2]; /* B -> DWord 2, bits 15:0 */
+ ((uint16_t *) sdc)[5] = c[3]; /* A -> DWord 3, bits 31:16 */
+ break;
+ case 32:
+ if (base_format == GL_RG) {
+ /* Careful inspection of the tables reveals that for RG32 formats,
+ * the green channel needs to go where blue normally belongs.
+ */
+ sdc[0] = c[0];
+ sdc[2] = c[1];
+ sdc[3] = 1;
+ } else {
+ /* Copy RGBA in order. */
+ for (int i = 0; i < 4; i++)
+ sdc[i] = c[i];
+ }
+ break;
+ default:
+ assert(!"Invalid number of bits per channel in integer format.");
+ break;
+ }
} else if (brw->gen == 5 || brw->gen == 6) {
struct gen5_sampler_default_color *sdc;
memset(sdc, 0, sizeof(*sdc));
- UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[0], color[0]);
- UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[1], color[1]);
- UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[2], color[2]);
- UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[3], color[3]);
+ UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[0], color.f[0]);
+ UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[1], color.f[1]);
+ UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[2], color.f[2]);
+ UNCLAMPED_FLOAT_TO_UBYTE(sdc->ub[3], color.f[3]);
- UNCLAMPED_FLOAT_TO_USHORT(sdc->us[0], color[0]);
- UNCLAMPED_FLOAT_TO_USHORT(sdc->us[1], color[1]);
- UNCLAMPED_FLOAT_TO_USHORT(sdc->us[2], color[2]);
- UNCLAMPED_FLOAT_TO_USHORT(sdc->us[3], color[3]);
+ UNCLAMPED_FLOAT_TO_USHORT(sdc->us[0], color.f[0]);
+ UNCLAMPED_FLOAT_TO_USHORT(sdc->us[1], color.f[1]);
+ UNCLAMPED_FLOAT_TO_USHORT(sdc->us[2], color.f[2]);
+ UNCLAMPED_FLOAT_TO_USHORT(sdc->us[3], color.f[3]);
- UNCLAMPED_FLOAT_TO_SHORT(sdc->s[0], color[0]);
- UNCLAMPED_FLOAT_TO_SHORT(sdc->s[1], color[1]);
- UNCLAMPED_FLOAT_TO_SHORT(sdc->s[2], color[2]);
- UNCLAMPED_FLOAT_TO_SHORT(sdc->s[3], color[3]);
+ UNCLAMPED_FLOAT_TO_SHORT(sdc->s[0], color.f[0]);
+ UNCLAMPED_FLOAT_TO_SHORT(sdc->s[1], color.f[1]);
+ UNCLAMPED_FLOAT_TO_SHORT(sdc->s[2], color.f[2]);
+ UNCLAMPED_FLOAT_TO_SHORT(sdc->s[3], color.f[3]);
- sdc->hf[0] = _mesa_float_to_half(color[0]);
- sdc->hf[1] = _mesa_float_to_half(color[1]);
- sdc->hf[2] = _mesa_float_to_half(color[2]);
- sdc->hf[3] = _mesa_float_to_half(color[3]);
+ sdc->hf[0] = _mesa_float_to_half(color.f[0]);
+ sdc->hf[1] = _mesa_float_to_half(color.f[1]);
+ sdc->hf[2] = _mesa_float_to_half(color.f[2]);
+ sdc->hf[3] = _mesa_float_to_half(color.f[3]);
sdc->b[0] = sdc->s[0] >> 8;
sdc->b[1] = sdc->s[1] >> 8;
sdc->b[2] = sdc->s[2] >> 8;
sdc->b[3] = sdc->s[3] >> 8;
- sdc->f[0] = color[0];
- sdc->f[1] = color[1];
- sdc->f[2] = color[2];
- sdc->f[3] = color[3];
+ sdc->f[0] = color.f[0];
+ sdc->f[1] = color.f[1];
+ sdc->f[2] = color.f[2];
+ sdc->f[3] = color.f[3];
} else {
float *sdc = brw_state_batch(brw, AUB_TRACE_SAMPLER_DEFAULT_COLOR,
4 * 4, 32, sdc_offset);
- memcpy(sdc, color, 4 * 4);
+ memcpy(sdc, color.f, 4 * 4);
}
}
* Sets the sampler state for a single unit based off of the sampler key
* entry.
*/
-static void
+void
brw_update_sampler_state(struct brw_context *brw,
- int unit,
+ GLenum target, bool tex_cube_map_seamless,
+ GLfloat tex_unit_lod_bias,
+ mesa_format format, GLenum base_format,
+ bool is_integer_format,
+ const struct gl_sampler_object *sampler,
uint32_t *sampler_state,
uint32_t batch_offset_for_sampler_state)
{
- struct gl_context *ctx = &brw->ctx;
- const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
- const struct gl_texture_object *texObj = texUnit->_Current;
- const struct gl_sampler_object *sampler = _mesa_get_samplerobj(ctx, unit);
-
- /* These don't use samplers at all. */
- if (texObj->Target == GL_TEXTURE_BUFFER)
- return;
-
unsigned min_filter, mag_filter, mip_filter;
/* Select min and mip filters. */
/* Enable anisotropic filtering if desired. */
unsigned max_anisotropy = BRW_ANISORATIO_2;
- if (sampler->MaxAnisotropy > 1.0) {
+ if (sampler->MaxAnisotropy > 1.0f) {
min_filter = BRW_MAPFILTER_ANISOTROPIC;
mag_filter = BRW_MAPFILTER_ANISOTROPIC;
- if (sampler->MaxAnisotropy > 2.0) {
+ if (sampler->MaxAnisotropy > 2.0f) {
max_anisotropy =
MIN2((sampler->MaxAnisotropy - 2) / 2, BRW_ANISORATIO_16);
}
unsigned wrap_t = translate_wrap_mode(brw, sampler->WrapT, either_nearest);
unsigned wrap_r = translate_wrap_mode(brw, sampler->WrapR, either_nearest);
- if (texObj->Target == GL_TEXTURE_CUBE_MAP ||
- texObj->Target == GL_TEXTURE_CUBE_MAP_ARRAY) {
+ if (target == GL_TEXTURE_CUBE_MAP ||
+ target == GL_TEXTURE_CUBE_MAP_ARRAY) {
/* Cube maps must use the same wrap mode for all three coordinate
* dimensions. Prior to Haswell, only CUBE and CLAMP are valid.
*/
- if ((ctx->Texture.CubeMapSeamless || sampler->CubeMapSeamless) &&
+ if ((tex_cube_map_seamless || sampler->CubeMapSeamless) &&
(sampler->MinFilter != GL_NEAREST ||
sampler->MagFilter != GL_NEAREST)) {
wrap_s = BRW_TEXCOORDMODE_CUBE;
wrap_t = BRW_TEXCOORDMODE_CLAMP;
wrap_r = BRW_TEXCOORDMODE_CLAMP;
}
- } else if (texObj->Target == GL_TEXTURE_1D) {
+ } else if (target == GL_TEXTURE_1D) {
/* There's a bug in 1D texture sampling - it actually pays
* attention to the wrap_t value, though it should not.
* Override the wrap_t value here to GL_REPEAT to keep
const unsigned min_lod = U_FIXED(CLAMP(sampler->MinLod, 0, 13), lod_bits);
const unsigned max_lod = U_FIXED(CLAMP(sampler->MaxLod, 0, 13), lod_bits);
const int lod_bias =
- S_FIXED(CLAMP(texUnit->LodBias + sampler->LodBias, -16, 15), lod_bits);
+ S_FIXED(CLAMP(tex_unit_lod_bias + sampler->LodBias, -16, 15), lod_bits);
const unsigned base_level = U_FIXED(0, 1);
- uint32_t border_color_offset;
- upload_default_color(brw, sampler, unit, &border_color_offset);
+ /* Upload the border color if necessary. If not, just point it at
+ * offset 0 (the start of the batch) - the color should be ignored,
+ * but that address won't fault in case something reads it anyway.
+ */
+ uint32_t border_color_offset = 0;
+ if (wrap_mode_needs_border_color(wrap_s) ||
+ wrap_mode_needs_border_color(wrap_t) ||
+ wrap_mode_needs_border_color(wrap_r)) {
+ upload_default_color(brw, sampler,
+ format, base_format, is_integer_format,
+ &border_color_offset);
+ }
- const bool non_normalized_coords = texObj->Target == GL_TEXTURE_RECTANGLE;
+ const bool non_normalized_coords = target == GL_TEXTURE_RECTANGLE;
brw_emit_sampler_state(brw,
sampler_state,
border_color_offset);
}
+static void
+update_sampler_state(struct brw_context *brw,
+ int unit,
+ uint32_t *sampler_state,
+ uint32_t batch_offset_for_sampler_state)
+{
+ struct gl_context *ctx = &brw->ctx;
+ const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
+ const struct gl_texture_object *texObj = texUnit->_Current;
+ const struct gl_sampler_object *sampler = _mesa_get_samplerobj(ctx, unit);
+
+ /* These don't use samplers at all. */
+ if (texObj->Target == GL_TEXTURE_BUFFER)
+ return;
+
+ struct gl_texture_image *firstImage = texObj->Image[0][texObj->BaseLevel];
+ brw_update_sampler_state(brw, texObj->Target, ctx->Texture.CubeMapSeamless,
+ texUnit->LodBias,
+ firstImage->TexFormat, firstImage->_BaseFormat,
+ texObj->_IsIntegerFormat,
+ sampler,
+ sampler_state, batch_offset_for_sampler_state);
+}
static void
brw_upload_sampler_state_table(struct brw_context *brw,
if (SamplersUsed & (1 << s)) {
const unsigned unit = prog->SamplerUnits[s];
if (ctx->Texture.Unit[unit]._Current) {
- brw_update_sampler_state(brw, unit, sampler_state,
+ update_sampler_state(brw, unit, sampler_state,
batch_offset_for_sampler_state);
}
}
/* Flag that the sampler state table pointer has changed; later atoms
* will handle it.
*/
- brw->state.dirty.cache |= CACHE_NEW_SAMPLER;
+ brw->ctx.NewDriverState |= BRW_NEW_SAMPLER_STATE_TABLE;
}
}
.mesa = _NEW_TEXTURE,
.brw = BRW_NEW_BATCH |
BRW_NEW_FRAGMENT_PROGRAM,
- .cache = 0
},
.emit = brw_upload_fs_samplers,
};
.mesa = _NEW_TEXTURE,
.brw = BRW_NEW_BATCH |
BRW_NEW_VERTEX_PROGRAM,
- .cache = 0
},
.emit = brw_upload_vs_samplers,
};
.mesa = _NEW_TEXTURE,
.brw = BRW_NEW_BATCH |
BRW_NEW_GEOMETRY_PROGRAM,
- .cache = 0
},
.emit = brw_upload_gs_samplers,
};