radeonsi: implement 32-bit SNORM/UNORM/SSCALED/USCALED vertex formats

Reviewed-by: Nicolai Hähnle <[email protected]>

3 files changed, 90 insertions, 18 deletions

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index ddb6ca1ae17..840b3790f04 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -434,7 +434,11 @@ static void declare_input_vs(
 	}
 
 	fix_fetch = (ctx->shader->key.mono.vs.fix_fetch >> (4 * input_index)) & 0xf;
-	if (fix_fetch) {
+
+	switch (fix_fetch) {
+	case SI_FIX_FETCH_A2_SNORM:
+	case SI_FIX_FETCH_A2_SSCALED:
+	case SI_FIX_FETCH_A2_SINT: {
 		/* The hardware returns an unsigned value; convert it to a
 		 * signed one.
 		 */
@@ -470,6 +474,52 @@ static void declare_input_vs(
 		}
 
 		out[3] = tmp;
+		break;
+	}
+	case SI_FIX_FETCH_RGBA_32_UNORM:
+	case SI_FIX_FETCH_RGBX_32_UNORM:
+		for (chan = 0; chan < 4; chan++) {
+			out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
+						     ctx->i32, "");
+			out[chan] = LLVMBuildUIToFP(gallivm->builder,
+						    out[chan], ctx->f32, "");
+			out[chan] = LLVMBuildFMul(gallivm->builder, out[chan],
+						  LLVMConstReal(ctx->f32, 1.0 / UINT_MAX), "");
+		}
+		/* RGBX UINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
+		if (fix_fetch == SI_FIX_FETCH_RGBX_32_UNORM)
+			out[3] = LLVMConstReal(ctx->f32, 1);
+		break;
+	case SI_FIX_FETCH_RGBA_32_SNORM:
+	case SI_FIX_FETCH_RGBX_32_SNORM:
+		for (chan = 0; chan < 4; chan++) {
+			out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
+						     ctx->i32, "");
+			out[chan] = LLVMBuildSIToFP(gallivm->builder,
+						    out[chan], ctx->f32, "");
+			out[chan] = LLVMBuildFMul(gallivm->builder, out[chan],
+						  LLVMConstReal(ctx->f32, 1.0 / INT_MAX), "");
+		}
+		/* RGBX SINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
+		if (fix_fetch == SI_FIX_FETCH_RGBX_32_SNORM)
+			out[3] = LLVMConstReal(ctx->f32, 1);
+		break;
+	case SI_FIX_FETCH_RGBA_32_USCALED:
+		for (chan = 0; chan < 4; chan++) {
+			out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
+						     ctx->i32, "");
+			out[chan] = LLVMBuildUIToFP(gallivm->builder,
+						    out[chan], ctx->f32, "");
+		}
+		break;
+	case SI_FIX_FETCH_RGBA_32_SSCALED:
+		for (chan = 0; chan < 4; chan++) {
+			out[chan] = LLVMBuildBitCast(gallivm->builder, out[chan],
+						     ctx->i32, "");
+			out[chan] = LLVMBuildSIToFP(gallivm->builder,
+						    out[chan], ctx->f32, "");
+		}
+		break;
 	}
 }
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 89f9628a1f6..5e554d9afe3 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -236,9 +236,15 @@ enum {
 /* For VS shader key fix_fetch. */
 enum {
 	SI_FIX_FETCH_NONE = 0,
-	SI_FIX_FETCH_A2_SNORM = 1,
-	SI_FIX_FETCH_A2_SSCALED = 2,
-	SI_FIX_FETCH_A2_SINT = 3,
+	SI_FIX_FETCH_A2_SNORM,
+	SI_FIX_FETCH_A2_SSCALED,
+	SI_FIX_FETCH_A2_SINT,
+	SI_FIX_FETCH_RGBA_32_UNORM,
+	SI_FIX_FETCH_RGBX_32_UNORM,
+	SI_FIX_FETCH_RGBA_32_SNORM,
+	SI_FIX_FETCH_RGBX_32_SNORM,
+	SI_FIX_FETCH_RGBA_32_USCALED,
+	SI_FIX_FETCH_RGBA_32_SSCALED,
 };
 
 struct si_shader;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index fa78a56f73e..c8d1099421c 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -1746,14 +1746,6 @@ static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
 		}
 		break;
 	case 32:
-		/* From the Southern Islands ISA documentation about MTBUF:
-		 * 'Memory reads of data in memory that is 32 or 64 bits do not
-		 * undergo any format conversion.'
-		 */
-		if (type != UTIL_FORMAT_TYPE_FLOAT &&
-		    !desc->channel[first_non_void].pure_integer)
-			return V_008F0C_BUF_DATA_FORMAT_INVALID;
-
 		switch (desc->nr_channels) {
 		case 1:
 			return V_008F0C_BUF_DATA_FORMAT_32;
@@ -1781,18 +1773,20 @@ static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
 
 	switch (desc->channel[first_non_void].type) {
 	case UTIL_FORMAT_TYPE_SIGNED:
-		if (desc->channel[first_non_void].normalized)
-			return V_008F0C_BUF_NUM_FORMAT_SNORM;
-		else if (desc->channel[first_non_void].pure_integer)
+		if (desc->channel[first_non_void].size >= 32 ||
+		    desc->channel[first_non_void].pure_integer)
 			return V_008F0C_BUF_NUM_FORMAT_SINT;
+		else if (desc->channel[first_non_void].normalized)
+			return V_008F0C_BUF_NUM_FORMAT_SNORM;
 		else
 			return V_008F0C_BUF_NUM_FORMAT_SSCALED;
 		break;
 	case UTIL_FORMAT_TYPE_UNSIGNED:
-		if (desc->channel[first_non_void].normalized)
-			return V_008F0C_BUF_NUM_FORMAT_UNORM;
-		else if (desc->channel[first_non_void].pure_integer)
+		if (desc->channel[first_non_void].size >= 32 ||
+		    desc->channel[first_non_void].pure_integer)
 			return V_008F0C_BUF_NUM_FORMAT_UINT;
+		else if (desc->channel[first_non_void].normalized)
+			return V_008F0C_BUF_NUM_FORMAT_UNORM;
 		else
 			return V_008F0C_BUF_NUM_FORMAT_USCALED;
 		break;
@@ -3342,6 +3336,7 @@ static void *si_create_vertex_elements(struct pipe_context *ctx,
 	v->count = count;
 	for (i = 0; i < count; ++i) {
 		const struct util_format_description *desc;
+		const struct util_format_channel_description *channel;
 		unsigned data_format, num_format;
 		int first_non_void;
 
@@ -3349,6 +3344,7 @@ static void *si_create_vertex_elements(struct pipe_context *ctx,
 		first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
 		data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
 		num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
+		channel = &desc->channel[first_non_void];
 
 		v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
 				   S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
@@ -3370,6 +3366,26 @@ static void *si_create_vertex_elements(struct pipe_context *ctx,
 				/* This isn't actually used in OpenGL. */
 				v->fix_fetch |= (uint64_t)SI_FIX_FETCH_A2_SINT << (4 * i);
 			}
+		} else if (channel->size == 32 && !channel->pure_integer) {
+			if (channel->type == UTIL_FORMAT_TYPE_SIGNED) {
+				if (channel->normalized) {
+					if (desc->swizzle[3] == PIPE_SWIZZLE_1)
+						v->fix_fetch |= (uint64_t)SI_FIX_FETCH_RGBX_32_SNORM << (4 * i);
+					else
+						v->fix_fetch |= (uint64_t)SI_FIX_FETCH_RGBA_32_SNORM << (4 * i);
+				} else {
+					v->fix_fetch |= (uint64_t)SI_FIX_FETCH_RGBA_32_SSCALED << (4 * i);
+				}
+			} else if (channel->type == UTIL_FORMAT_TYPE_UNSIGNED) {
+				if (channel->normalized) {
+					if (desc->swizzle[3] == PIPE_SWIZZLE_1)
+						v->fix_fetch |= (uint64_t)SI_FIX_FETCH_RGBX_32_UNORM << (4 * i);
+					else
+						v->fix_fetch |= (uint64_t)SI_FIX_FETCH_RGBA_32_UNORM << (4 * i);
+				} else {
+					v->fix_fetch |= (uint64_t)SI_FIX_FETCH_RGBA_32_USCALED << (4 * i);
+				}
+			}
 		}
 
 		/* We work around the fact that 8_8_8 and 16_16_16 data formats