[Mesa-dev] [PATCH] st/mesa: add st fp64 support (v7)
From: Dave Airlie
v2 : add double to int/unsigned conversion
v3: handle fp64 consts better
v4: use DRSQ
v4.1: add d2b
v4.2: drop DDIV
v5: split out some prep patches.
v5.1: add some comments.
v5.2: more comments
v6: simplify down the double instruction
generation loop.
v7: Merge Ilia's two cleanup patches.
Signed-off-by: Dave Airlie
---
src/mesa/state_tracker/st_extensions.c | 6 +
src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 578 ++---
2 files changed, 458 insertions(+), 126 deletions(-)
diff --git a/src/mesa/state_tracker/st_extensions.c
b/src/mesa/state_tracker/st_extensions.c
index 9137a50..ce29d07 100644
--- a/src/mesa/state_tracker/st_extensions.c
+++ b/src/mesa/state_tracker/st_extensions.c
@@ -900,4 +900,10 @@ void st_init_extensions(struct pipe_screen *screen,
PIPE_VIDEO_CAP_SUPPORTS_INTERLACED)) {
extensions->NV_vdpau_interop = GL_TRUE;
}
+
+ if (screen->get_shader_param(screen, PIPE_SHADER_VERTEX,
+PIPE_SHADER_CAP_DOUBLES) &&
+ screen->get_shader_param(screen, PIPE_SHADER_FRAGMENT,
+PIPE_SHADER_CAP_DOUBLES))
+ extensions->ARB_gpu_shader_fp64 = GL_TRUE;
}
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index 56502fb..003d280 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -229,7 +229,7 @@ public:
DECLARE_RALLOC_CXX_OPERATORS(glsl_to_tgsi_instruction)
unsigned op;
- st_dst_reg dst[1];
+ st_dst_reg dst[2];
st_src_reg src[4];
/** Pointer to the ir source this tree came from for debugging */
ir_instruction *ir;
@@ -262,16 +262,17 @@ public:
class immediate_storage : public exec_node {
public:
- immediate_storage(gl_constant_value *values, int size, int type)
+ immediate_storage(gl_constant_value *values, int size32, int type)
{
- memcpy(this->values, values, size * sizeof(gl_constant_value));
- this->size = size;
+ memcpy(this->values, values, size32 * sizeof(gl_constant_value));
+ this->size32 = size32;
this->type = type;
}
+ /* doubles are stored across 2 gl_constant_values */
gl_constant_value values[4];
- int size; /**< Number of components (1-4) */
- int type; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
+ int size32; /**< Number of 32-bit components (1-4) */
+ int type; /**< GL_DOUBLE, GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
};
class function_entry : public exec_node {
@@ -334,7 +335,7 @@ public:
variable_storage *find_variable_storage(ir_variable *var);
- int add_constant(gl_register_file file, gl_constant_value values[4],
+ int add_constant(gl_register_file file, gl_constant_value values[8],
int size, int datatype, GLuint *swizzle_out);
function_entry *get_function_signature(ir_function_signature *sig);
@@ -342,6 +343,7 @@ public:
st_src_reg get_temp(const glsl_type *type);
void reladdr_to_temp(ir_instruction *ir, st_src_reg *reg, int *num_reladdr);
+ st_src_reg st_src_reg_for_double(double val);
st_src_reg st_src_reg_for_float(float val);
st_src_reg st_src_reg_for_int(int val);
st_src_reg st_src_reg_for_type(int type, int val);
@@ -397,6 +399,10 @@ public:
st_dst_reg dst, st_src_reg src0);
glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
+ st_dst_reg dst, st_dst_reg dst1,
+ st_src_reg src0);
+
+ glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
st_dst_reg dst, st_src_reg src0, st_src_reg
src1);
glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
@@ -408,6 +414,11 @@ public:
st_src_reg src0, st_src_reg src1,
st_src_reg src2, st_src_reg src3);
+ glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
+ st_dst_reg dst, st_dst_reg dst1,
+ st_src_reg src0, st_src_reg src1,
+ st_src_reg src2, st_src_reg src3);
+
unsigned get_opcode(ir_instruction *ir, unsigned op,
st_dst_reg dst,
st_src_reg src0, st_src_reg src1);
@@ -432,6 +443,7 @@ public:
void emit_scs(ir_instruction *ir, unsigned op,
st_dst_reg dst, const st_src_reg &src);
+
bool try_emit_mad(ir_expression *ir,
int mul_operand);
bool try_emit_mad_for_and_not(ir_expression *ir,
@@ -451,6 +463,8 @@ public:
void copy_propagate(void);
int eliminate_dead_code(void);
+
+ void merge_two_dsts(void);
void merge_registers(void);
void renumber_registers(void);
@@ -464,7 +478,6 @@ public:
static st_src_reg undef_src = st_src_reg(PROGRA
Re: [Mesa-dev] [PATCH] st/mesa: add st fp64 support (v7)
On Thu, Feb 19, 2015 at 6:09 PM, Dave Airlie wrote:
> From: Dave Airlie
>
> v2 : add double to int/unsigned conversion
> v3: handle fp64 consts better
> v4: use DRSQ
> v4.1: add d2b
> v4.2: drop DDIV
>
> v5: split out some prep patches.
> v5.1: add some comments.
> v5.2: more comments
>
> v6: simplify down the double instruction
> generation loop.
>
> v7: Merge Ilia's two cleanup patches.
>
> Signed-off-by: Dave Airlie
> ---
> src/mesa/state_tracker/st_extensions.c | 6 +
> src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 578
> ++---
> 2 files changed, 458 insertions(+), 126 deletions(-)
>
> diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> index 56502fb..003d280 100644
> --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> @@ -464,7 +478,6 @@ public:
> static st_src_reg undef_src = st_src_reg(PROGRAM_UNDEFINED, 0,
> GLSL_TYPE_ERROR);
>
> static st_dst_reg undef_dst = st_dst_reg(PROGRAM_UNDEFINED, SWIZZLE_NOOP,
> GLSL_TYPE_ERROR);
> -
> static st_dst_reg address_reg = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X,
> GLSL_TYPE_FLOAT, 0);
> static st_dst_reg address_reg2 = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X,
> GLSL_TYPE_FLOAT, 1);
> static st_dst_reg sampler_reladdr = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X,
> GLSL_TYPE_FLOAT, 2);
Drop this hunk.
> @@ -597,22 +616,129 @@ glsl_to_tgsi_visitor::emit(ir_instruction *ir,
> unsigned op,
>
> this->instructions.push_tail(inst);
>
> + /*
> +* This section contains the double processing.
> +* GLSL just represents doubles as single channel values,
> +* however most HW and TGSI represent doubles as pairs of register
> channels.
> +*
> +* so we have to fixup destination writemask/index and src
> swizzle/indexes.
> +* dest writemasks need to translate from single channel write mask
> +* to a dual-channel writemask, but also need to modify the index,
> +* if we are touching the Z,W fields in the pre-translated writemask.
> +*
> +* src channels have similiar index modifications along with swizzle
> +* changes to we pick the XY, ZW pairs from the correct index.
> +*
> +* GLSL [0].x -> TGSI [0].xy
> +* GLSL [0].y -> TGSI [0].zw
> +* GLSL [0].z -> TGSI [1].xy
> +* GLSL [0].w -> TGSI [1].zw
> +*/
> + if (inst->dst[0].type == GLSL_TYPE_DOUBLE || inst->dst[1].type ==
> GLSL_TYPE_DOUBLE ||
> + inst->src[0].type == GLSL_TYPE_DOUBLE) {
> + glsl_to_tgsi_instruction *dinst = NULL;
> + int initial_src_swz[4], initial_src_idx[4];
> + int initial_dst_idx[2], initial_dst_writemask[2];
> + /* select the writemask for dst0 or dst1 */
> + unsigned writemask = inst->dst[0].file == PROGRAM_UNDEFINED ?
> inst->dst[1].writemask : inst->dst[0].writemask;
> +
> + /* copy out the writemask, index and swizzles for all src/dsts. */
> + for (j = 0; j < 2; j++) {
> + initial_dst_writemask[j] = inst->dst[j].writemask;
> + initial_dst_idx[j] = inst->dst[j].index;
> + }
> +
> + for (j = 0; j < 4; j++) {
> + initial_src_swz[j] = inst->src[j].swizzle;
> + initial_src_idx[j] = inst->src[j].index;
> + }
> +
> + /*
> + * scan all the components in the dst writemask
> + * generate an instruction for each of them if required.
> + */
> + while (writemask) {
> +
> + int i = u_bit_scan(&writemask);
> +
> + /* first time use previous instruction */
> + if (dinst == NULL) {
> +dinst = inst;
> + } else {
> +/* create a new instructions for subsequent attempts */
> +dinst = new(mem_ctx) glsl_to_tgsi_instruction();
> +*dinst = *inst;
> +dinst->next = NULL;
> +dinst->prev = NULL;
> +this->instructions.push_tail(dinst);
> + }
> +
> + /* modify the destination if we are splitting */
> + for (j = 0; j < 2; j++) {
> +if (dinst->dst[j].type == GLSL_TYPE_DOUBLE) {
> + dinst->dst[j].writemask = (i & 1) ? WRITEMASK_ZW :
> WRITEMASK_XY;
> + dinst->dst[j].index = initial_dst_idx[j];
> + if (i > 1)
> + dinst->dst[j].index++;
> +} else {
> + /* if we aren't writing to a double, just get the bit of the
> initial writemask
> + for this channel */
> + dinst->dst[j].writemask = initial_dst_writemask[j] & (1 << i);
> +}
> + }
> +
> + /* modify the src registers */
> + for (j = 0; j < 4; j++) {
> +int swz = GET_SWZ(initial_src_swz[j], i);
> +
> +if (dinst->src[j].type == GLSL_TYPE_DOUBLE) {
> + dinst->src[j].index = initial_src_idx[j];
> + if (swz > 1)
> + dinst->src[j].index++;
> +
> + if (swz & 1)
> +
